[dev.cc] all: merge master (5f1efe7) into dev.cc

Conflicts:
	src/cmd/dist/build.go

Change-Id: I98a4b5e010bee91507b85bb8efd9c74e1a1f649c

268 files changed, 14391 insertions, 13130 deletions

diff --git a/src/cmd/5a/a.y b/src/cmd/5a/a.y
index 429f7437c6..10e9f6feea 100644
--- a/src/cmd/5a/a.y
+++ b/src/cmd/5a/a.y
@@ -460,20 +460,32 @@ fcon:
 reglist:
 	spreg
 	{
-		$$ = 1 << $1;
+		if($1 < REG_R0 || $1 > REG_R15)
+			yyerror("invalid register in reglist");
+
+		$$ = 1 << ($1&15);
 	}
 |	spreg '-' spreg
 	{
 		int i;
+
+		if($1 < REG_R0 || $1 > REG_R15)
+			yyerror("invalid register in reglist");
+		if($3 < REG_R0 || $3 > REG_R15)
+			yyerror("invalid register in reglist");
+
 		$$=0;
 		for(i=$1; i<=$3; i++)
-			$$ |= 1<<i;
+			$$ |= 1<<(i&15);
 		for(i=$3; i<=$1; i++)
-			$$ |= 1<<i;
+			$$ |= 1<<(i&15);
 	}
 |	spreg comma reglist
 	{
-		$$ = (1<<$1) | $3;
+		if($1 < REG_R0 || $1 > REG_R15)
+			yyerror("invalid register in reglist");
+
+		$$ = (1<<($1&15)) | $3;
 	}
 
 gen:
diff --git a/src/cmd/5a/y.tab.c b/src/cmd/5a/y.tab.c
index d9af383d78..416af9a321 100644
--- a/src/cmd/5a/y.tab.c
+++ b/src/cmd/5a/y.tab.c
@@ -580,14 +580,14 @@ static const yytype_uint16 yyrline[] =
      220,   232,   237,   249,   260,   267,   274,   278,   282,   286,
      293,   315,   323,   332,   339,   348,   359,   365,   368,   372,
      377,   378,   381,   387,   398,   405,   412,   419,   427,   433,
-     438,   444,   447,   453,   461,   465,   474,   480,   481,   482,
-     483,   488,   494,   500,   506,   507,   510,   511,   519,   528,
-     529,   538,   539,   545,   548,   549,   550,   552,   560,   568,
-     577,   583,   589,   595,   603,   609,   617,   618,   622,   630,
-     631,   637,   638,   646,   647,   650,   656,   664,   672,   680,
-     690,   693,   697,   703,   704,   705,   708,   709,   713,   717,
-     721,   725,   731,   734,   740,   741,   745,   749,   753,   757,
-     761,   765,   769,   773,   777
+     438,   444,   447,   453,   461,   468,   483,   492,   493,   494,
+     495,   500,   506,   512,   518,   519,   522,   523,   531,   540,
+     541,   550,   551,   557,   560,   561,   562,   564,   572,   580,
+     589,   595,   601,   607,   615,   621,   629,   630,   634,   642,
+     643,   649,   650,   658,   659,   662,   668,   676,   684,   692,
+     702,   705,   709,   715,   716,   717,   720,   721,   725,   729,
+     733,   737,   743,   746,   752,   753,   757,   761,   765,   769,
+     773,   777,   781,   785,   789
 };
 #endif
 
@@ -2223,31 +2223,43 @@ yyreduce:
   case 64:
 #line 462 "a.y"
     {
-		(yyval.lval) = 1 << (yyvsp[(1) - (1)].lval);
+		if((yyvsp[(1) - (1)].lval) < REG_R0 || (yyvsp[(1) - (1)].lval) > REG_R15)
+			yyerror("invalid register in reglist");
+
+		(yyval.lval) = 1 << ((yyvsp[(1) - (1)].lval)&15);
 	}
     break;
 
   case 65:
-#line 466 "a.y"
+#line 469 "a.y"
     {
 		int i;
+
+		if((yyvsp[(1) - (3)].lval) < REG_R0 || (yyvsp[(1) - (3)].lval) > REG_R15)
+			yyerror("invalid register in reglist");
+		if((yyvsp[(3) - (3)].lval) < REG_R0 || (yyvsp[(3) - (3)].lval) > REG_R15)
+			yyerror("invalid register in reglist");
+
 		(yyval.lval)=0;
 		for(i=(yyvsp[(1) - (3)].lval); i<=(yyvsp[(3) - (3)].lval); i++)
-			(yyval.lval) |= 1<<i;
+			(yyval.lval) |= 1<<(i&15);
 		for(i=(yyvsp[(3) - (3)].lval); i<=(yyvsp[(1) - (3)].lval); i++)
-			(yyval.lval) |= 1<<i;
+			(yyval.lval) |= 1<<(i&15);
 	}
     break;
 
   case 66:
-#line 475 "a.y"
+#line 484 "a.y"
     {
-		(yyval.lval) = (1<<(yyvsp[(1) - (3)].lval)) | (yyvsp[(3) - (3)].lval);
+		if((yyvsp[(1) - (3)].lval) < REG_R0 || (yyvsp[(1) - (3)].lval) > REG_R15)
+			yyerror("invalid register in reglist");
+
+		(yyval.lval) = (1<<((yyvsp[(1) - (3)].lval)&15)) | (yyvsp[(3) - (3)].lval);
 	}
     break;
 
   case 70:
-#line 484 "a.y"
+#line 496 "a.y"
     {
 		(yyval.addr) = (yyvsp[(1) - (4)].addr);
 		(yyval.addr).reg = (yyvsp[(3) - (4)].lval);
@@ -2255,7 +2267,7 @@ yyreduce:
     break;
 
   case 71:
-#line 489 "a.y"
+#line 501 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_REG;
@@ -2264,7 +2276,7 @@ yyreduce:
     break;
 
   case 72:
-#line 495 "a.y"
+#line 507 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_REG;
@@ -2273,7 +2285,7 @@ yyreduce:
     break;
 
   case 73:
-#line 501 "a.y"
+#line 513 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_MEM;
@@ -2282,7 +2294,7 @@ yyreduce:
     break;
 
   case 77:
-#line 512 "a.y"
+#line 524 "a.y"
     {
 		(yyval.addr) = (yyvsp[(1) - (1)].addr);
 		if((yyvsp[(1) - (1)].addr).name != NAME_EXTERN && (yyvsp[(1) - (1)].addr).name != NAME_STATIC) {
@@ -2291,7 +2303,7 @@ yyreduce:
     break;
 
   case 78:
-#line 520 "a.y"
+#line 532 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_MEM;
@@ -2301,7 +2313,7 @@ yyreduce:
     break;
 
   case 80:
-#line 530 "a.y"
+#line 542 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_MEM;
@@ -2311,7 +2323,7 @@ yyreduce:
     break;
 
   case 82:
-#line 540 "a.y"
+#line 552 "a.y"
     {
 		(yyval.addr) = (yyvsp[(1) - (4)].addr);
 		(yyval.addr).type = TYPE_MEM;
@@ -2320,7 +2332,7 @@ yyreduce:
     break;
 
   case 87:
-#line 553 "a.y"
+#line 565 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_CONST;
@@ -2329,7 +2341,7 @@ yyreduce:
     break;
 
   case 88:
-#line 561 "a.y"
+#line 573 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_REG;
@@ -2338,7 +2350,7 @@ yyreduce:
     break;
 
   case 89:
-#line 569 "a.y"
+#line 581 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_REGREG;
@@ -2348,7 +2360,7 @@ yyreduce:
     break;
 
   case 90:
-#line 578 "a.y"
+#line 590 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_SHIFT;
@@ -2357,7 +2369,7 @@ yyreduce:
     break;
 
   case 91:
-#line 584 "a.y"
+#line 596 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_SHIFT;
@@ -2366,7 +2378,7 @@ yyreduce:
     break;
 
   case 92:
-#line 590 "a.y"
+#line 602 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_SHIFT;
@@ -2375,7 +2387,7 @@ yyreduce:
     break;
 
   case 93:
-#line 596 "a.y"
+#line 608 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_SHIFT;
@@ -2384,7 +2396,7 @@ yyreduce:
     break;
 
   case 94:
-#line 604 "a.y"
+#line 616 "a.y"
     {
 		if((yyval.lval) < REG_R0 || (yyval.lval) > REG_R15)
 			print("register value out of range in shift\n");
@@ -2393,7 +2405,7 @@ yyreduce:
     break;
 
   case 95:
-#line 610 "a.y"
+#line 622 "a.y"
     {
 		if((yyval.lval) < 0 || (yyval.lval) >= 32)
 			print("shift value out of range\n");
@@ -2402,14 +2414,14 @@ yyreduce:
     break;
 
   case 97:
-#line 619 "a.y"
+#line 631 "a.y"
     {
 		(yyval.lval) = REGPC;
 	}
     break;
 
   case 98:
-#line 623 "a.y"
+#line 635 "a.y"
     {
 		if((yyvsp[(3) - (4)].lval) < 0 || (yyvsp[(3) - (4)].lval) >= NREG)
 			print("register value out of range in R(...)\n");
@@ -2418,14 +2430,14 @@ yyreduce:
     break;
 
   case 100:
-#line 632 "a.y"
+#line 644 "a.y"
     {
 		(yyval.lval) = REGSP;
 	}
     break;
 
   case 102:
-#line 639 "a.y"
+#line 651 "a.y"
     {
 		if((yyvsp[(3) - (4)].lval) < 0 || (yyvsp[(3) - (4)].lval) >= NREG)
 			print("register value out of range in C(...)\n");
@@ -2434,7 +2446,7 @@ yyreduce:
     break;
 
   case 105:
-#line 651 "a.y"
+#line 663 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_REG;
@@ -2443,7 +2455,7 @@ yyreduce:
     break;
 
   case 106:
-#line 657 "a.y"
+#line 669 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_REG;
@@ -2452,7 +2464,7 @@ yyreduce:
     break;
 
   case 107:
-#line 665 "a.y"
+#line 677 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_MEM;
@@ -2463,7 +2475,7 @@ yyreduce:
     break;
 
   case 108:
-#line 673 "a.y"
+#line 685 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_MEM;
@@ -2474,7 +2486,7 @@ yyreduce:
     break;
 
   case 109:
-#line 681 "a.y"
+#line 693 "a.y"
     {
 		(yyval.addr) = nullgen;
 		(yyval.addr).type = TYPE_MEM;
@@ -2485,140 +2497,140 @@ yyreduce:
     break;
 
   case 110:
-#line 690 "a.y"
+#line 702 "a.y"
     {
 		(yyval.lval) = 0;
 	}
     break;
 
   case 111:
-#line 694 "a.y"
+#line 706 "a.y"
     {
 		(yyval.lval) = (yyvsp[(2) - (2)].lval);
 	}
     break;
 
   case 112:
-#line 698 "a.y"
+#line 710 "a.y"
     {
 		(yyval.lval) = -(yyvsp[(2) - (2)].lval);
 	}
     break;
 
   case 117:
-#line 710 "a.y"
+#line 722 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (1)].sym)->value;
 	}
     break;
 
   case 118:
-#line 714 "a.y"
+#line 726 "a.y"
     {
 		(yyval.lval) = -(yyvsp[(2) - (2)].lval);
 	}
     break;
 
   case 119:
-#line 718 "a.y"
+#line 730 "a.y"
     {
 		(yyval.lval) = (yyvsp[(2) - (2)].lval);
 	}
     break;
 
   case 120:
-#line 722 "a.y"
+#line 734 "a.y"
     {
 		(yyval.lval) = ~(yyvsp[(2) - (2)].lval);
 	}
     break;
 
   case 121:
-#line 726 "a.y"
+#line 738 "a.y"
     {
 		(yyval.lval) = (yyvsp[(2) - (3)].lval);
 	}
     break;
 
   case 122:
-#line 731 "a.y"
+#line 743 "a.y"
     {
 		(yyval.lval) = 0;
 	}
     break;
 
   case 123:
-#line 735 "a.y"
+#line 747 "a.y"
     {
 		(yyval.lval) = (yyvsp[(2) - (2)].lval);
 	}
     break;
 
   case 125:
-#line 742 "a.y"
+#line 754 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (3)].lval) + (yyvsp[(3) - (3)].lval);
 	}
     break;
 
   case 126:
-#line 746 "a.y"
+#line 758 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (3)].lval) - (yyvsp[(3) - (3)].lval);
 	}
     break;
 
   case 127:
-#line 750 "a.y"
+#line 762 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (3)].lval) * (yyvsp[(3) - (3)].lval);
 	}
     break;
 
   case 128:
-#line 754 "a.y"
+#line 766 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (3)].lval) / (yyvsp[(3) - (3)].lval);
 	}
     break;
 
   case 129:
-#line 758 "a.y"
+#line 770 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (3)].lval) % (yyvsp[(3) - (3)].lval);
 	}
     break;
 
   case 130:
-#line 762 "a.y"
+#line 774 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (4)].lval) << (yyvsp[(4) - (4)].lval);
 	}
     break;
 
   case 131:
-#line 766 "a.y"
+#line 778 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (4)].lval) >> (yyvsp[(4) - (4)].lval);
 	}
     break;
 
   case 132:
-#line 770 "a.y"
+#line 782 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (3)].lval) & (yyvsp[(3) - (3)].lval);
 	}
     break;
 
   case 133:
-#line 774 "a.y"
+#line 786 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (3)].lval) ^ (yyvsp[(3) - (3)].lval);
 	}
     break;
 
   case 134:
-#line 778 "a.y"
+#line 790 "a.y"
     {
 		(yyval.lval) = (yyvsp[(1) - (3)].lval) | (yyvsp[(3) - (3)].lval);
 	}
@@ -2626,7 +2638,7 @@ yyreduce:
 
 
 /* Line 1267 of yacc.c.  */
-#line 2630 "y.tab.c"
+#line 2642 "y.tab.c"
       default: break;
     }
   YY_SYMBOL_PRINT ("-> $$ =", yyr1[yyn], &yyval, &yyloc);
diff --git a/src/cmd/5g/cgen.c b/src/cmd/5g/cgen.c
index 6acf6dfdf0..354e0cbfd6 100644
--- a/src/cmd/5g/cgen.c
+++ b/src/cmd/5g/cgen.c
@@ -1148,27 +1148,20 @@ bgen(Node *n, int true, int likely, Prog *to)
 		goto ret;
 
 	case OANDAND:
-		if(!true)
-			goto caseor;
-
-	caseand:
-		p1 = gbranch(AB, T, 0);
-		p2 = gbranch(AB, T, 0);
-		patch(p1, pc);
-		bgen(n->left, !true, -likely, p2);
-		bgen(n->right, !true, -likely, p2);
-		p1 = gbranch(AB, T, 0);
-		patch(p1, to);
-		patch(p2, pc);
-		goto ret;
-
 	case OOROR:
-		if(!true)
-			goto caseand;
-
-	caseor:
-		bgen(n->left, true, likely, to);
-		bgen(n->right, true, likely, to);
+		if((n->op == OANDAND) == true) {
+			p1 = gbranch(AJMP, T, 0);
+			p2 = gbranch(AJMP, T, 0);
+			patch(p1, pc);
+			bgen(n->left, !true, -likely, p2);
+			bgen(n->right, !true, -likely, p2);
+			p1 = gbranch(AJMP, T, 0);
+			patch(p1, to);
+			patch(p2, pc);
+		} else {
+			bgen(n->left, true, likely, to);
+			bgen(n->right, true, likely, to);
+		}
 		goto ret;
 
 	case OEQ:
@@ -1626,6 +1619,8 @@ cadable(Node *n)
 /*
  * copy a composite value by moving its individual components.
  * Slices, strings and interfaces are supported.
+ * Small structs or arrays with elements of basic type are
+ * also supported.
  * nr is N when assigning a zero value.
  * return 1 if can do, 0 if cant.
  */
@@ -1680,7 +1675,7 @@ componentgen(Node *nr, Node *nl)
 
 	nodl = *nl;
 	if(!cadable(nl)) {
-		if(nr == N || !cadable(nr))
+		if(nr != N && !cadable(nr))
 			goto no;
 		igen(nl, &nodl, N);
 		freel = 1;
@@ -1700,7 +1695,6 @@ componentgen(Node *nr, Node *nl)
 		freer = 1;
 	}
 
-	
 	// nl and nr are 'cadable' which basically means they are names (variables) now.
 	// If they are the same variable, don't generate any code, because the
 	// VARDEF we generate will mark the old value as dead incorrectly.
diff --git a/src/cmd/5g/galign.c b/src/cmd/5g/galign.c
index 85ad347833..c4d74f0a71 100644
--- a/src/cmd/5g/galign.c
+++ b/src/cmd/5g/galign.c
@@ -42,61 +42,46 @@ betypeinit(void)
 void
 main(int argc, char **argv)
 {
-	arch.thechar = thechar;
-	arch.thestring = thestring;
-	arch.thelinkarch = thelinkarch;
-	arch.typedefs = typedefs;
-	arch.MAXWIDTH = MAXWIDTH;
-	arch.afunclit = afunclit;
-	arch.anyregalloc = anyregalloc;
-	arch.betypeinit = betypeinit;
-	arch.bgen = bgen;
-	arch.cgen = cgen;
-	arch.cgen_asop = cgen_asop;
-	arch.cgen_call = cgen_call;
-	arch.cgen_callinter = cgen_callinter;
-	arch.cgen_ret = cgen_ret;
-	arch.clearfat = clearfat;
-	arch.clearp = clearp;
-	arch.defframe = defframe;
-	arch.dgostringptr = dgostringptr;
-	arch.dgostrlitptr = dgostrlitptr;
-	arch.dsname = dsname;
-	arch.dsymptr = dsymptr;
-	arch.dumpdata = dumpdata;
-	arch.dumpit = dumpit;
-	arch.excise = excise;
-	arch.expandchecks = expandchecks;
-	arch.fixautoused = fixautoused;
-	arch.gclean = gclean;
-	arch.gdata = gdata;
-	arch.gdatacomplex = gdatacomplex;
-	arch.gdatastring = gdatastring;
-	arch.ggloblnod = ggloblnod;
-	arch.ggloblsym = ggloblsym;
-	arch.ginit = ginit;
-	arch.gins = gins;
-	arch.ginscall = ginscall;
-	arch.gjmp = gjmp;
-	arch.gtrack = gtrack;
-	arch.gused = gused;
-	arch.igen = igen;
-	arch.isfat = isfat;
-	arch.linkarchinit = linkarchinit;
-	arch.markautoused = markautoused;
-	arch.naddr = naddr;
-	arch.newplist = newplist;
-	arch.nodarg = nodarg;
-	arch.patch = patch;
-	arch.proginfo = proginfo;
-	arch.regalloc = regalloc;
-	arch.regfree = regfree;
-	arch.regopt = regopt;
-	arch.regtyp = regtyp;
-	arch.sameaddr = sameaddr;
-	arch.smallindir = smallindir;
-	arch.stackaddr = stackaddr;
-	arch.unpatch = unpatch;
+	thearch.thechar = thechar;
+	thearch.thestring = thestring;
+	thearch.thelinkarch = thelinkarch;
+	thearch.typedefs = typedefs;
+	thearch.REGSP = REGSP;
+	thearch.REGCTXT = REGCTXT;
+	thearch.MAXWIDTH = MAXWIDTH;
+	thearch.anyregalloc = anyregalloc;
+	thearch.betypeinit = betypeinit;
+	thearch.bgen = bgen;
+	thearch.cgen = cgen;
+	thearch.cgen_call = cgen_call;
+	thearch.cgen_callinter = cgen_callinter;
+	thearch.cgen_ret = cgen_ret;
+	thearch.clearfat = clearfat;
+	thearch.defframe = defframe;
+	thearch.excise = excise;
+	thearch.expandchecks = expandchecks;
+	thearch.gclean = gclean;
+	thearch.ginit = ginit;
+	thearch.gins = gins;
+	thearch.ginscall = ginscall;
+	thearch.igen = igen;
+	thearch.linkarchinit = linkarchinit;
+	thearch.peep = peep;
+	thearch.proginfo = proginfo;
+	thearch.regalloc = regalloc;
+	thearch.regfree = regfree;
+	thearch.regtyp = regtyp;
+	thearch.sameaddr = sameaddr;
+	thearch.smallindir = smallindir;
+	thearch.stackaddr = stackaddr;
+	thearch.excludedregs = excludedregs;
+	thearch.RtoB = RtoB;
+	thearch.FtoB = RtoB;
+	thearch.BtoR = BtoR;
+	thearch.BtoF = BtoF;
+	thearch.optoas = optoas;
+	thearch.doregbits = doregbits;
+	thearch.regnames = regnames;
 	
 	gcmain(argc, argv);
 }
diff --git a/src/cmd/5g/gg.h b/src/cmd/5g/gg.h
index 8a75311d73..b12c7e2561 100644
--- a/src/cmd/5g/gg.h
+++ b/src/cmd/5g/gg.h
@@ -17,10 +17,7 @@ enum
 	REGALLOC_FMAX = FREGEXT,
 };
 
-EXTERN	int32	dynloc;
 EXTERN	uchar	reg[REGALLOC_FMAX+1];
-EXTERN	int32	pcloc;		// instruction counter
-EXTERN	Strlit	emptystring;
 extern	long	unmappedzero;
 
 /*
@@ -118,7 +115,6 @@ int anyregalloc(void);
 void betypeinit(void);
 void bgen(Node*, int, int, Prog*);
 void cgen(Node*, Node*);
-void cgen_asop(Node*);
 void cgen_call(Node*, int);
 void cgen_callinter(Node*, Node*, int);
 void cgen_ret(Node*);
@@ -163,3 +159,19 @@ int sameaddr(Addr*, Addr*);
 int smallindir(Addr*, Addr*);
 int stackaddr(Addr*);
 Prog* unpatch(Prog*);
+
+/*
+ * reg.c
+ */
+uint64 excludedregs(void);
+uint64 RtoB(int);
+uint64 FtoB(int);
+int BtoR(uint64);
+int BtoF(uint64);
+uint64 doregbits(int);
+char** regnames(int*);
+
+/*
+ * peep.c
+ */
+void peep(Prog*);
diff --git a/src/cmd/5g/ggen.c b/src/cmd/5g/ggen.c
index b7c621be32..62b9beadb0 100644
--- a/src/cmd/5g/ggen.c
+++ b/src/cmd/5g/ggen.c
@@ -7,7 +7,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
 static Prog* appendpp(Prog*, int, int, int, int32, int, int, int32);
 static Prog *zerorange(Prog *p, vlong frame, vlong lo, vlong hi, uint32 *r0);
@@ -117,52 +117,6 @@ appendpp(Prog *p, int as, int ftype, int freg, int32 foffset, int ttype, int tre
 	return q;	
 }
 
-// Sweep the prog list to mark any used nodes.
-void
-markautoused(Prog* p)
-{
-	for (; p; p = p->link) {
-		if (p->as == ATYPE || p->as == AVARDEF || p->as == AVARKILL)
-			continue;
-
-		if (p->from.node)
-			((Node*)(p->from.node))->used = 1;
-
-		if (p->to.node)
-			((Node*)(p->to.node))->used = 1;
-	}
-}
-
-// Fixup instructions after allocauto (formerly compactframe) has moved all autos around.
-void
-fixautoused(Prog* p)
-{
-	Prog **lp;
-
-	for (lp=&p; (p=*lp) != P; ) {
-		if (p->as == ATYPE && p->from.node && p->from.name == NAME_AUTO && !((Node*)(p->from.node))->used) {
-			*lp = p->link;
-			continue;
-		}
-		if ((p->as == AVARDEF || p->as == AVARKILL) && p->to.node && !((Node*)(p->to.node))->used) {
-			// Cannot remove VARDEF instruction, because - unlike TYPE handled above -
-			// VARDEFs are interspersed with other code, and a jump might be using the
-			// VARDEF as a target. Replace with a no-op instead. A later pass will remove
-			// the no-ops.
-			nopout(p);
-			continue;
-		}
-
-		if (p->from.name == NAME_AUTO && p->from.node)
-			p->from.offset += ((Node*)(p->from.node))->stkdelta;
-
-		if (p->to.name == NAME_AUTO && p->to.node)
-			p->to.offset += ((Node*)(p->to.node))->stkdelta;
-
-		lp = &p->link;
-	}
-}
-
 /*
  * generate:
  *	call f
@@ -176,7 +130,7 @@ void
 ginscall(Node *f, int proc)
 {
 	Prog *p;
-	Node n1, r, r1, con;
+	Node r, r1, con;
 	int32 extra;
 
 	if(f->type != T) {
@@ -238,10 +192,7 @@ ginscall(Node *f, int proc)
 		p->to.reg = REGSP;
 		p->to.offset = 4;
 
-		memset(&n1, 0, sizeof n1);
-		n1.op = OADDR;
-		n1.left = f;
-		gins(AMOVW, &n1, &r);
+		gins(AMOVW, f, &r);
 		p = gins(AMOVW, &r, N);
 		p->to.type = TYPE_MEM;
 		p->to.reg = REGSP;
@@ -487,147 +438,6 @@ cgen_ret(Node *n)
 }
 
 /*
- * generate += *= etc.
- */
-void
-cgen_asop(Node *n)
-{
-	Node n1, n2, n3, n4;
-	Node *nl, *nr;
-	Prog *p1;
-	Addr addr;
-	int a, w;
-
-	nl = n->left;
-	nr = n->right;
-
-	if(nr->ullman >= UINF && nl->ullman >= UINF) {
-		tempname(&n1, nr->type);
-		cgen(nr, &n1);
-		n2 = *n;
-		n2.right = &n1;
-		cgen_asop(&n2);
-		goto ret;
-	}
-
-	if(!isint[nl->type->etype])
-		goto hard;
-	if(!isint[nr->type->etype])
-		goto hard;
-	if(is64(nl->type) || is64(nr->type))
-		goto hard64;
-
-	switch(n->etype) {
-	case OADD:
-	case OSUB:
-	case OXOR:
-	case OAND:
-	case OOR:
-		a = optoas(n->etype, nl->type);
-		if(nl->addable) {
-			if(smallintconst(nr))
-				n3 = *nr;
-			else {
-				regalloc(&n3, nr->type, N);
-				cgen(nr, &n3);
-			}
-			regalloc(&n2, nl->type, N);
-			cgen(nl, &n2);
-			gins(a, &n3, &n2);
-			cgen(&n2, nl);
-			regfree(&n2);
-			if(n3.op != OLITERAL)
-				regfree(&n3);
-			goto ret;
-		}
-		if(nr->ullman < UINF)
-		if(sudoaddable(a, nl, &addr, &w)) {
-			w = optoas(OAS, nl->type);
-			regalloc(&n2, nl->type, N);
-			p1 = gins(w, N, &n2);
-			p1->from = addr;
-			regalloc(&n3, nr->type, N);
-			cgen(nr, &n3);
-			gins(a, &n3, &n2);
-			p1 = gins(w, &n2, N);
-			p1->to = addr;
-			regfree(&n2);
-			regfree(&n3);
-			sudoclean();
-			goto ret;
-		}
-	}
-
-hard:
-	n2.op = 0;
-	n1.op = 0;
-	if(nr->op == OLITERAL) {
-		// don't allocate a register for literals.
-	} else if(nr->ullman >= nl->ullman || nl->addable) {
-		regalloc(&n2, nr->type, N);
-		cgen(nr, &n2);
-		nr = &n2;
-	} else {
-		tempname(&n2, nr->type);
-		cgen(nr, &n2);
-		nr = &n2;
-	}
-	if(!nl->addable) {
-		igen(nl, &n1, N);
-		nl = &n1;
-	}
-
-	n3 = *n;
-	n3.left = nl;
-	n3.right = nr;
-	n3.op = n->etype;
-
-	regalloc(&n4, nl->type, N);
-	cgen(&n3, &n4);
-	gmove(&n4, nl);
-
-	if(n1.op)
-		regfree(&n1);
-	if(n2.op == OREGISTER)
-		regfree(&n2);
-	regfree(&n4);
-	goto ret;
-
-hard64:
-	if(nr->ullman > nl->ullman) {
-		tempname(&n2, nr->type);
-		cgen(nr, &n2);
-		igen(nl, &n1, N);
-	} else {
-		igen(nl, &n1, N);
-		tempname(&n2, nr->type);
-		cgen(nr, &n2);
-	}
-
-	n3 = *n;
-	n3.left = &n1;
-	n3.right = &n2;
-	n3.op = n->etype;
-
-	cgen(&n3, &n1);
-
-ret:
-	;
-}
-
-int
-samereg(Node *a, Node *b)
-{
-	if(a->op != OREGISTER)
-		return 0;
-	if(b->op != OREGISTER)
-		return 0;
-	if(a->val.u.reg != b->val.u.reg)
-		return 0;
-	return 1;
-}
-
-/*
  * generate high multiply
  *  res = (nl * nr) >> wordsize
  */
diff --git a/src/cmd/5g/gobj.c b/src/cmd/5g/gobj.c
deleted file mode 100644
index 13d06efe55..0000000000
--- a/src/cmd/5g/gobj.c
+++ /dev/null
@@ -1,267 +0,0 @@
-// Derived from Inferno utils/5c/swt.c
-// http://code.google.com/p/inferno-os/source/browse/utils/5c/swt.c
-//
-//	Copyright © 1994-1999 Lucent Technologies Inc.  All rights reserved.
-//	Portions Copyright © 1995-1997 C H Forsyth (forsyth@terzarima.net)
-//	Portions Copyright © 1997-1999 Vita Nuova Limited
-//	Portions Copyright © 2000-2007 Vita Nuova Holdings Limited (www.vitanuova.com)
-//	Portions Copyright © 2004,2006 Bruce Ellis
-//	Portions Copyright © 2005-2007 C H Forsyth (forsyth@terzarima.net)
-//	Revisions Copyright © 2000-2007 Lucent Technologies Inc. and others
-//	Portions Copyright © 2009 The Go Authors.  All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-//
-// The above copyright notice and this permission notice shall be included in
-// all copies or substantial portions of the Software.
-//
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
-// THE SOFTWARE.
-
-#include <u.h>
-#include <libc.h>
-#include "gg.h"
-
-int
-dsname(Sym *sym, int off, char *t, int n)
-{
-	Prog *p;
-
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.etype = TINT32;
-	p->from.offset = off;
-	p->from.reg = 0;
-	p->from.sym = linksym(sym);
-	
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = n;
-	
-	p->to.type = TYPE_SCONST;
-	p->to.name = NAME_NONE;
-	p->to.reg = 0;
-	p->to.offset = 0;
-	memmove(p->to.u.sval, t, n);
-	return off + n;
-}
-
-/*
- * make a refer to the data s, s+len
- * emitting DATA if needed.
- */
-void
-datastring(char *s, int len, Addr *a)
-{
-	Sym *sym;
-	
-	sym = stringsym(s, len);
-	a->type = TYPE_MEM;
-	a->name = NAME_EXTERN;
-	a->etype = TINT32;
-	a->offset = widthptr+4;  // skip header
-	a->reg = 0;
-	a->sym = linksym(sym);
-	a->node = sym->def;
-}
-
-/*
- * make a refer to the string sval,
- * emitting DATA if needed.
- */
-void
-datagostring(Strlit *sval, Addr *a)
-{
-	Sym *sym;
-	
-	sym = stringsym(sval->s, sval->len);
-	a->type = TYPE_MEM;
-	a->name = NAME_EXTERN;
-	a->etype = TSTRING;
-	a->offset = 0;  // header
-	a->reg = 0;
-	a->sym = linksym(sym);
-	a->node = sym->def;
-}
-
-void
-gdata(Node *nam, Node *nr, int wid)
-{
-	Prog *p;
-	vlong v;
-
-	if(nr->op == OLITERAL) {
-		switch(nr->val.ctype) {
-		case CTCPLX:
-			gdatacomplex(nam, nr->val.u.cval);
-			return;
-		case CTSTR:
-			gdatastring(nam, nr->val.u.sval);
-			return;
-		}
-	}
-
-	if(wid == 8 && is64(nr->type)) {
-		v = mpgetfix(nr->val.u.xval);
-		p = gins(ADATA, nam, nodintconst(v));
-		p->from3.type = TYPE_CONST;
-		p->from3.offset = 4;
-		p = gins(ADATA, nam, nodintconst(v>>32));
-		p->from3.type = TYPE_CONST;
-		p->from3.offset = 4;
-		p->from.offset += 4;
-		return;
-	}
-	p = gins(ADATA, nam, nr);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = wid;
-}
-
-void
-gdatacomplex(Node *nam, Mpcplx *cval)
-{
-	Prog *p;
-	int w;
-
-	w = cplxsubtype(nam->type->etype);
-	w = types[w]->width;
-
-	p = gins(ADATA, nam, N);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = w;
-	p->to.type = TYPE_FCONST;
-	p->to.u.dval = mpgetflt(&cval->real);
-
-	p = gins(ADATA, nam, N);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = w;
-	p->from.offset += w;
-	p->to.type = TYPE_FCONST;
-	p->to.u.dval = mpgetflt(&cval->imag);
-}
-
-void
-gdatastring(Node *nam, Strlit *sval)
-{
-	Prog *p;
-	Node nod1;
-
-	p = gins(ADATA, nam, N);
-	datastring(sval->s, sval->len, &p->to);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = types[tptr]->width;
-	p->to.type = TYPE_CONST;
-	p->to.etype = TINT32;
-//print("%P\n", p);
-
-	nodconst(&nod1, types[TINT32], sval->len);
-	p = gins(ADATA, nam, &nod1);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = types[TINT32]->width;
-	p->from.offset += types[tptr]->width;
-}
-
-int
-dstringptr(Sym *s, int off, char *str)
-{
-	Prog *p;
-
-	off = rnd(off, widthptr);
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-
-	datastring(str, strlen(str)+1, &p->to);
-	p->to.type = TYPE_CONST;
-	p->to.etype = TINT32;
-	off += widthptr;
-
-	return off;
-}
-
-int
-dgostrlitptr(Sym *s, int off, Strlit *lit)
-{
-	Prog *p;
-
-	if(lit == nil)
-		return duintptr(s, off, 0);
-
-	off = rnd(off, widthptr);
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-	datagostring(lit, &p->to);
-	p->to.type = TYPE_CONST;
-	p->to.etype = TINT32;
-	off += widthptr;
-
-	return off;
-}
-
-int
-dgostringptr(Sym *s, int off, char *str)
-{
-	int n;
-	Strlit *lit;
-
-	if(str == nil)
-		return duintptr(s, off, 0);
-
-	n = strlen(str);
-	lit = mal(sizeof *lit + n);
-	strcpy(lit->s, str);
-	lit->len = n;
-	return dgostrlitptr(s, off, lit);
-}
-
-int
-dsymptr(Sym *s, int off, Sym *x, int xoff)
-{
-	Prog *p;
-
-	off = rnd(off, widthptr);
-
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-	p->to.type = TYPE_ADDR;
-	p->to.name = NAME_EXTERN;
-	p->to.sym = linksym(x);
-	p->to.offset = xoff;
-	off += widthptr;
-
-	return off;
-}
-
-void
-nopout(Prog *p)
-{
-	p->as = ANOP;
-	p->scond = zprog.scond;
-	p->from = zprog.from;
-	p->to = zprog.to;
-	p->reg = zprog.reg;
-}
diff --git a/src/cmd/5g/gsubr.c b/src/cmd/5g/gsubr.c
index ef5a509147..bdb70270a8 100644
--- a/src/cmd/5g/gsubr.c
+++ b/src/cmd/5g/gsubr.c
@@ -38,229 +38,6 @@
 // At the same time, can raise StackBig in ../../runtime/stack.h.
 long unmappedzero = 4096;
 
-void
-clearp(Prog *p)
-{
-	p->as = AEND;
-	p->reg = 0;
-	p->scond = C_SCOND_NONE;
-	p->from.type = TYPE_NONE;
-	p->from.name = NAME_NONE;
-	p->from.reg = 0;
-	p->to.type = TYPE_NONE;
-	p->to.name = NAME_NONE;
-	p->to.reg = 0;
-	p->pc = pcloc;
-	pcloc++;
-}
-
-static int ddumped;
-static Prog *dfirst;
-static Prog *dpc;
-
-/*
- * generate and return proc with p->as = as,
- * linked into program.  pc is next instruction.
- */
-Prog*
-prog(int as)
-{
-	Prog *p;
-
-	if(as == ADATA || as == AGLOBL) {
-		if(ddumped)
-			fatal("already dumped data");
-		if(dpc == nil) {
-			dpc = mal(sizeof(*dpc));
-			dfirst = dpc;
-		}
-		p = dpc;
-		dpc = mal(sizeof(*dpc));
-		p->link = dpc;
-		p->reg = 0;  // used for flags
-	} else {
-		p = pc;
-		pc = mal(sizeof(*pc));
-		clearp(pc);
-		p->link = pc;
-	}
-
-	if(lineno == 0) {
-		if(debug['K'])
-			warn("prog: line 0");
-	}
-
-	p->as = as;
-	p->lineno = lineno;
-	return p;
-}
-
-void
-dumpdata(void)
-{
-	ddumped = 1;
-	if(dfirst == nil)
-		return;
-	newplist();
-	*pc = *dfirst;
-	pc = dpc;
-	clearp(pc);
-}
-
-/*
- * generate a branch.
- * t is ignored.
- * likely values are for branch prediction:
- *	-1 unlikely
- *	0 no opinion
- *	+1 likely
- */
-Prog*
-gbranch(int as, Type *t, int likely)
-{
-	Prog *p;
-
-	USED(t);
-	USED(likely);  // TODO: record this for linker
-
-	p = prog(as);
-	p->to.type = TYPE_BRANCH;
-	p->to.u.branch = P;
-	return p;
-}
-
-/*
- * patch previous branch to jump to to.
- */
-void
-patch(Prog *p, Prog *to)
-{
-	if(p->to.type != TYPE_BRANCH)
-		fatal("patch: not a branch");
-	p->to.u.branch = to;
-	p->to.offset = to->pc;
-}
-
-Prog*
-unpatch(Prog *p)
-{
-	Prog *q;
-
-	if(p->to.type != TYPE_BRANCH)
-		fatal("unpatch: not a branch");
-	q = p->to.u.branch;
-	p->to.u.branch = P;
-	p->to.offset = 0;
-	return q;
-}
-
-/*
- * start a new Prog list.
- */
-Plist*
-newplist(void)
-{
-	Plist *pl;
-
-	pl = linknewplist(ctxt);
-
-	pc = mal(sizeof(*pc));
-	clearp(pc);
-	pl->firstpc = pc;
-
-	return pl;
-}
-
-void
-gused(Node *n)
-{
-	gins(ANOP, n, N);	// used
-}
-
-Prog*
-gjmp(Prog *to)
-{
-	Prog *p;
-
-	p = gbranch(AB, T, 0);
-	if(to != P)
-		patch(p, to);
-	return p;
-}
-
-void
-ggloblnod(Node *nam)
-{
-	Prog *p;
-
-	p = gins(AGLOBL, nam, N);
-	p->lineno = nam->lineno;
-	p->from.sym->gotype = linksym(ngotype(nam));
-	p->to.sym = nil;
-	p->to.type = TYPE_CONST;
-	p->to.offset = nam->type->width;
-	if(nam->readonly)
-		p->from3.offset = RODATA;
-	if(nam->type != T && !haspointers(nam->type))
-		p->from3.offset |= NOPTR;
-}
-
-void
-ggloblsym(Sym *s, int32 width, int8 flags)
-{
-	Prog *p;
-
-	p = gins(AGLOBL, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->to.type = TYPE_CONST;
-	p->to.name = NAME_NONE;
-	p->to.offset = width;
-	p->from3.offset = flags;
-}
-
-void
-gtrack(Sym *s)
-{
-	Prog *p;
-	
-	p = gins(AUSEFIELD, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-}
-
-int
-isfat(Type *t)
-{
-	if(t != T)
-	switch(t->etype) {
-	case TSTRUCT:
-	case TARRAY:
-	case TSTRING:
-	case TINTER:	// maybe remove later
-		return 1;
-	}
-	return 0;
-}
-
-/*
- * naddr of func generates code for address of func.
- * if using opcode that can take address implicitly,
- * call afunclit to fix up the argument.
- * also fix up direct register references to be TYPE_MEM.
- */
-void
-afunclit(Addr *a, Node *n)
-{
-	if(a->type == TYPE_ADDR && a->name == NAME_EXTERN || a->type == TYPE_REG) {
-		a->type = TYPE_MEM;
-		if(n->op == ONAME)
-			a->sym = linksym(n->sym);
-	}
-}
-
 static	int	resvd[] =
 {
 	9,     // reserved for m
@@ -292,7 +69,7 @@ gclean(void)
 			yyerror("reg %R left allocated\n", i);
 }
 
-int32
+int
 anyregalloc(void)
 {
 	int i, j;
@@ -427,101 +204,6 @@ regfree(Node *n)
 }
 
 /*
- * initialize n to be register r of type t.
- */
-void
-nodreg(Node *n, Type *t, int r)
-{
-	if(t == T)
-		fatal("nodreg: t nil");
-
-	memset(n, 0, sizeof(*n));
-	n->op = OREGISTER;
-	n->addable = 1;
-	ullmancalc(n);
-	n->val.u.reg = r;
-	n->type = t;
-}
-
-/*
- * initialize n to be indirect of register r; n is type t.
- */
-void
-nodindreg(Node *n, Type *t, int r)
-{
-	nodreg(n, t, r);
-	n->op = OINDREG;
-}
-
-Node*
-nodarg(Type *t, int fp)
-{
-	Node *n;
-	NodeList *l;
-	Type *first;
-	Iter savet;
-
-	// entire argument struct, not just one arg
-	if(t->etype == TSTRUCT && t->funarg) {
-		n = nod(ONAME, N, N);
-		n->sym = lookup(".args");
-		n->type = t;
-		first = structfirst(&savet, &t);
-		if(first == nil)
-			fatal("nodarg: bad struct");
-		if(first->width == BADWIDTH)
-			fatal("nodarg: offset not computed for %T", t);
-		n->xoffset = first->width;
-		n->addable = 1;
-		goto fp;
-	}
-
-	if(t->etype != TFIELD)
-		fatal("nodarg: not field %T", t);
-
-	if(fp == 1) {
-		for(l=curfn->dcl; l; l=l->next) {
-			n = l->n;
-			if((n->class == PPARAM || n->class == PPARAMOUT) && !isblanksym(t->sym) && n->sym == t->sym)
-				return n;
-		}
-	}
-
-	n = nod(ONAME, N, N);
-	n->type = t->type;
-	n->sym = t->sym;
-	if(t->width == BADWIDTH)
-		fatal("nodarg: offset not computed for %T", t);
-	n->xoffset = t->width;
-	n->addable = 1;
-	n->orig = t->nname;
-
-fp:
-	// Rewrite argument named _ to __,
-	// or else the assignment to _ will be
-	// discarded during code generation.
-	if(isblank(n))
-		n->sym = lookup("__");
-
-	switch(fp) {
-	default:
-		fatal("nodarg %T %d", t, fp);
-
-	case 0:		// output arg for calling another function
-		n->op = OINDREG;
-		n->val.u.reg = REGSP;
-		n->xoffset += 4;
-		break;
-
-	case 1:		// input arg to current function
-		n->class = PPARAM;
-		break;
-	}
-	n->typecheck = 1;
-	return n;
-}
-
-/*
  * return constant i node.
  * overwritten by next call, but useful in calls to gins.
  */
@@ -536,22 +218,6 @@ ncon(uint32 i)
 	return &n;
 }
 
-/*
- * Is this node a memory operand?
- */
-int
-ismem(Node *n)
-{
-	switch(n->op) {
-	case OINDREG:
-	case ONAME:
-	case OPARAM:
-	case OCLOSUREVAR:
-		return 1;
-	}
-	return 0;
-}
-
 Node sclean[10];
 int nsclean;
 
@@ -573,22 +239,25 @@ split64(Node *n, Node *lo, Node *hi)
 	nsclean++;
 	switch(n->op) {
 	default:
-		if(!dotaddable(n, &n1)) {
-			igen(n, &n1, N);
-			sclean[nsclean-1] = n1;
-		}
-		n = &n1;
-		goto common;
-	case ONAME:
-		if(n->class == PPARAMREF) {
-			cgen(n->heapaddr, &n1);
-			sclean[nsclean-1] = n1;
-			// fall through.
+		switch(n->op) {
+		default:
+			if(!dotaddable(n, &n1)) {
+				igen(n, &n1, N);
+				sclean[nsclean-1] = n1;
+			}
 			n = &n1;
+			break;
+		case ONAME:
+			if(n->class == PPARAMREF) {
+				cgen(n->heapaddr, &n1);
+				sclean[nsclean-1] = n1;
+				n = &n1;
+			}
+			break;
+		case OINDREG:
+			// nothing
+			break;
 		}
-		goto common;
-	case OINDREG:
-	common:
 		*lo = *n;
 		*hi = *n;
 		lo->type = types[TUINT32];
@@ -1182,228 +851,6 @@ gregshift(int as, Node *lhs, int32 stype, Node *reg, Node *rhs)
 }
 
 /*
- * generate code to compute n;
- * make a refer to result.
- */
-void
-naddr(Node *n, Addr *a, int canemitcode)
-{
-	Sym *s;
-
-	a->type = TYPE_NONE;
-	a->name = NAME_NONE;
-	a->reg = 0;
-	a->gotype = nil;
-	a->node = N;
-	a->etype = 0;
-	if(n == N)
-		return;
-
-	if(n->type != T && n->type->etype != TIDEAL) {
-		dowidth(n->type);
-		a->width = n->type->width;
-	}
-
-	switch(n->op) {
-	default:
-		fatal("naddr: bad %O %D", n->op, a);
-		break;
-
-	case OREGISTER:
-		a->type = TYPE_REG;
-		a->reg = n->val.u.reg;
-		a->sym = nil;
-		break;
-
-	case OINDEX:
-	case OIND:
-		fatal("naddr: OINDEX");
-//		naddr(n->left, a);
-//		if(a->type >= D_AX && a->type <= D_DI)
-//			a->type += D_INDIR;
-//		else
-//		if(a->type == TYPE_ADDR)
-//			a->type = TYPE_NONE+D_INDIR;
-//		else
-//		if(a->type == TYPE_ADDR) {
-//			a->type = a->index;
-//			a->index = TYPE_NONE;
-//		} else
-//			goto bad;
-//		if(n->op == OINDEX) {
-//			a->index = idx.reg;
-//			a->scale = n->scale;
-//		}
-//		break;
-
-	case OINDREG:
-		a->type = TYPE_MEM;
-		a->reg = n->val.u.reg;
-		a->sym = linksym(n->sym);
-		a->offset = n->xoffset;
-		break;
-
-	case OPARAM:
-		// n->left is PHEAP ONAME for stack parameter.
-		// compute address of actual parameter on stack.
-		a->etype = simtype[n->left->type->etype];
-		a->width = n->left->type->width;
-		a->offset = n->xoffset;
-		a->sym = linksym(n->left->sym);
-		a->type = TYPE_MEM;
-		a->name = NAME_PARAM;
-		a->node = n->left->orig;
-		break;
-	
-	case OCLOSUREVAR:
-		if(!curfn->needctxt)
-			fatal("closurevar without needctxt");
-		a->type = TYPE_MEM;
-		a->reg = REG_R7;
-		a->offset = n->xoffset;
-		a->sym = nil;
-		break;		
-
-	case OCFUNC:
-		naddr(n->left, a, canemitcode);
-		a->sym = linksym(n->left->sym);
-		break;
-
-	case ONAME:
-		a->etype = 0;
-		a->width = 0;
-		a->reg = 0;
-		if(n->type != T) {
-			a->etype = simtype[n->type->etype];
-			a->width = n->type->width;
-		}
-		a->offset = n->xoffset;
-		s = n->sym;
-		a->node = n->orig;
-		//if(a->node >= (Node*)&n)
-		//	fatal("stack node");
-		if(s == S)
-			s = lookup(".noname");
-		if(n->method) {
-			if(n->type != T)
-			if(n->type->sym != S)
-			if(n->type->sym->pkg != nil)
-				s = pkglookup(s->name, n->type->sym->pkg);
-		}
-
-		a->type = TYPE_MEM;
-		switch(n->class) {
-		default:
-			fatal("naddr: ONAME class %S %d\n", n->sym, n->class);
-		case PEXTERN:
-			a->name = NAME_EXTERN;
-			break;
-		case PAUTO:
-			a->name = NAME_AUTO;
-			break;
-		case PPARAM:
-		case PPARAMOUT:
-			a->name = NAME_PARAM;
-			break;
-		case PFUNC:
-			a->name = NAME_EXTERN;
-			a->type = TYPE_ADDR;
-			s = funcsym(s);
-			break;
-		}
-		a->sym = linksym(s);
-		break;
-
-	case OLITERAL:
-		switch(n->val.ctype) {
-		default:
-			fatal("naddr: const %lT", n->type);
-			break;
-		case CTFLT:
-			a->type = TYPE_FCONST;
-			a->u.dval = mpgetflt(n->val.u.fval);
-			break;
-		case CTINT:
-		case CTRUNE:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = mpgetfix(n->val.u.xval);
-			break;
-		case CTSTR:
-			datagostring(n->val.u.sval, a);
-			break;
-		case CTBOOL:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = n->val.u.bval;
-			break;
-		case CTNIL:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = 0;
-			break;
-		}
-		break;
-
-	case OITAB:
-		// itable of interface value
-		naddr(n->left, a, canemitcode);
-		a->etype = simtype[tptr];
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// len(nil)
-		a->width = widthptr;
-		break;
-
-	case OSPTR:
-		// pointer in a string or slice
-		naddr(n->left, a, canemitcode);
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// ptr(nil)
-		a->etype = simtype[tptr];
-		a->offset += Array_array;
-		a->width = widthptr;
-		break;
-
-	case OLEN:
-		// len of string or slice
-		naddr(n->left, a, canemitcode);
-		a->etype = TINT32;
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// len(nil)
-		a->offset += Array_nel;
-		break;
-
-	case OCAP:
-		// cap of string or slice
-		naddr(n->left, a, canemitcode);
-		a->etype = TINT32;
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// cap(nil)
-		a->offset += Array_cap;
-		break;
-
-	case OADDR:
-		naddr(n->left, a, canemitcode);
-		a->etype = tptr;
-		switch(a->type) {
-		case TYPE_MEM:
-			a->type = TYPE_ADDR;
-			break;
-
-		case TYPE_REG:
-		case TYPE_ADDR:
-			break;
-		
-		default:
-			fatal("naddr: OADDR %d\n", a->type);
-		}
-	}
-	
-	if(a->width < 0)
-		fatal("naddr: bad width for %N -> %D", n, a);
-}
-
-/*
  * return Axxx for Oxxx on type t.
  */
 int
@@ -1812,6 +1259,8 @@ sudoaddable(int as, Node *n, Addr *a, int *w)
 	if(n->type == T)
 		return 0;
 
+	memset(a, 0, sizeof *a);
+
 	switch(n->op) {
 	case OLITERAL:
 		if(!isconst(n, CTINT))
diff --git a/src/cmd/5g/opt.h b/src/cmd/5g/opt.h
deleted file mode 100644
index 524607419d..0000000000
--- a/src/cmd/5g/opt.h
+++ /dev/null
@@ -1,179 +0,0 @@
-// Inferno utils/5c/gc.h
-// http://code.google.com/p/inferno-os/source/browse/utils/5c/gc.h
-//
-//	Copyright © 1994-1999 Lucent Technologies Inc.  All rights reserved.
-//	Portions Copyright © 1995-1997 C H Forsyth (forsyth@terzarima.net)
-//	Portions Copyright © 1997-1999 Vita Nuova Limited
-//	Portions Copyright © 2000-2007 Vita Nuova Holdings Limited (www.vitanuova.com)
-//	Portions Copyright © 2004,2006 Bruce Ellis
-//	Portions Copyright © 2005-2007 C H Forsyth (forsyth@terzarima.net)
-//	Revisions Copyright © 2000-2007 Lucent Technologies Inc. and others
-//	Portions Copyright © 2009 The Go Authors.  All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-//
-// The above copyright notice and this permission notice shall be included in
-// all copies or substantial portions of the Software.
-//
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
-// THE SOFTWARE.
-
-
-#define	Z	N
-#define	Adr	Addr
-
-#define	D_HI	TYPE_NONE
-#define	D_LO	TYPE_NONE
-
-#define	BLOAD(r)	band(bnot(r->refbehind), r->refahead)
-#define	BSTORE(r)	band(bnot(r->calbehind), r->calahead)
-#define	LOAD(r)		(~r->refbehind.b[z] & r->refahead.b[z])
-#define	STORE(r)	(~r->calbehind.b[z] & r->calahead.b[z])
-
-#define	CLOAD	5
-#define	CREF	5
-#define	CINF	1000
-#define	LOOP	3
-
-typedef	struct	Reg	Reg;
-typedef	struct	Rgn	Rgn;
-
-/*c2go
-extern Node *Z;
-enum
-{
-	D_HI = TYPE_NONE,
-	D_LO = TYPE_NONE,
-	CLOAD = 5,
-	CREF = 5,
-	CINF = 1000,
-	LOOP = 3,
-};
-
-uint32 BLOAD(Reg*);
-uint32 BSTORE(Reg*);
-uint64 LOAD(Reg*);
-uint64 STORE(Reg*);
-*/
-
-// A Reg is a wrapper around a single Prog (one instruction) that holds
-// register optimization information while the optimizer runs.
-// r->prog is the instruction.
-// r->prog->opt points back to r.
-struct	Reg
-{
-	Flow	f;
-
-	Bits	set;  		// regopt variables written by this instruction.
-	Bits	use1; 		// regopt variables read by prog->from.
-	Bits	use2; 		// regopt variables read by prog->to.
-
-	// refahead/refbehind are the regopt variables whose current
-	// value may be used in the following/preceding instructions
-	// up to a CALL (or the value is clobbered).
-	Bits	refbehind;
-	Bits	refahead;
-	// calahead/calbehind are similar, but for variables in
-	// instructions that are reachable after hitting at least one
-	// CALL.
-	Bits	calbehind;
-	Bits	calahead;
-	Bits	regdiff;
-	Bits	act;
-
-	int32	regu;		// register used bitmap
-};
-#define	R	((Reg*)0)
-/*c2go extern Reg *R; */
-
-#define	NRGN	600
-/*c2go enum { NRGN = 600 }; */
-
-// A Rgn represents a single regopt variable over a region of code
-// where a register could potentially be dedicated to that variable.
-// The code encompassed by a Rgn is defined by the flow graph,
-// starting at enter, flood-filling forward while varno is refahead
-// and backward while varno is refbehind, and following branches.  A
-// single variable may be represented by multiple disjoint Rgns and
-// each Rgn may choose a different register for that variable.
-// Registers are allocated to regions greedily in order of descending
-// cost.
-struct	Rgn
-{
-	Reg*	enter;
-	short	cost;
-	short	varno;
-	short	regno;
-};
-
-EXTERN	Reg	zreg;
-EXTERN	Reg*	freer;
-EXTERN	Reg**	rpo2r;
-EXTERN	Rgn	region[NRGN];
-EXTERN	Rgn*	rgp;
-EXTERN	int	nregion;
-EXTERN	int	nvar;
-EXTERN	int32	regbits;
-EXTERN	Bits	externs;
-EXTERN	Bits	params;
-EXTERN	Bits	consts;
-EXTERN	Bits	addrs;
-EXTERN	Bits	ivar;
-EXTERN	Bits	ovar;
-EXTERN	int	change;
-EXTERN	int32	maxnr;
-EXTERN	int32*	idom;
-
-EXTERN	struct
-{
-	int32	ncvtreg;
-	int32	nspill;
-	int32	nreload;
-	int32	ndelmov;
-	int32	nvar;
-	int32	naddr;
-} ostats;
-
-/*
- * reg.c
- */
-Reg*	rega(void);
-int	rcmp(const void*, const void*);
-void	regopt(Prog*);
-void	addmove(Reg*, int, int, int);
-Bits	mkvar(Reg *r, Adr *a);
-void	prop(Reg*, Bits, Bits);
-void	synch(Reg*, Bits);
-uint32	allreg(uint32, Rgn*);
-void	paint1(Reg*, int);
-uint32	paint2(Reg*, int, int);
-void	paint3(Reg*, int, uint32, int);
-void	addreg(Adr*, int);
-void	dumpit(char *str, Flow *r0, int);
-
-/*
- * peep.c
- */
-void	peep(Prog*);
-void	excise(Flow*);
-int	copyu(Prog*, Adr*, Adr*);
-
-uint32	RtoB(int);
-uint32	FtoB(int);
-int	BtoR(uint32);
-int	BtoF(uint32);
-
-/*
- * prog.c
- */
-void proginfo(ProgInfo*, Prog*);
diff --git a/src/cmd/5g/peep.c b/src/cmd/5g/peep.c
index 1a4df8d622..7fe3e1071f 100644
--- a/src/cmd/5g/peep.c
+++ b/src/cmd/5g/peep.c
@@ -32,7 +32,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
 static int	xtramodes(Graph*, Flow*, Adr*);
 static int	shortprop(Flow *r);
@@ -47,6 +47,7 @@ static Flow*	findpre(Flow *r, Adr *v);
 static int	copyau1(Prog *p, Adr *v);
 static int	isdconst(Addr *a);
 static int	isfloatreg(Addr*);
+static int	copyu(Prog *p, Adr *v, Adr *s);
 
 static uint32	gactive;
 
@@ -64,7 +65,7 @@ peep(Prog *firstp)
 	Prog *p;
 	int t;
 
-	g = flowstart(firstp, sizeof(Flow));
+	g = flowstart(firstp, 0);
 	if(g == nil)
 		return;
 	gactive = 0;
@@ -121,15 +122,14 @@ loop1:
 			}
 			break;
 
-#ifdef NOTDEF
-XXX
+			/*
 			if(p->scond == C_SCOND_NONE)
 			if(regtyp(&p->to))
 			if(isdconst(&p->from)) {
 				constprop(&p->from, &p->to, r->s1);
 			}
 			break;
-#endif
+			*/
 		}
 	}
 	if(t)
@@ -561,8 +561,6 @@ gotit:
  * AXXX (x<<y),a,b
  * ..
  */
-#define FAIL(msg) { if(debug['P']) print("\t%s; FAILURE\n", msg); return 0; }
-/*c2go void FAIL(char*); */
 
 int
 shiftprop(Flow *r)
@@ -573,8 +571,11 @@ shiftprop(Flow *r)
 	Adr a;
 
 	p = r->prog;
-	if(p->to.type != TYPE_REG)
-		FAIL("BOTCH: result not reg");
+	if(p->to.type != TYPE_REG) {
+		if(debug['P'])
+			print("\tBOTCH: result not reg; FAILURE\n");
+		return 0;
+	}
 	n = p->to.reg;
 	a = zprog.from;
 	if(p->reg != 0 && p->reg != p->to.reg) {
@@ -587,28 +588,43 @@ shiftprop(Flow *r)
 	for(;;) {
 		/* find first use of shift result; abort if shift operands or result are changed */
 		r1 = uniqs(r1);
-		if(r1 == nil)
-			FAIL("branch");
-		if(uniqp(r1) == nil)
-			FAIL("merge");
+		if(r1 == nil) {
+			if(debug['P'])
+				print("\tbranch; FAILURE\n");
+			return 0;
+		}
+		if(uniqp(r1) == nil) {
+			if(debug['P'])
+				print("\tmerge; FAILURE\n");
+			return 0;
+		}
 		p1 = r1->prog;
 		if(debug['P'])
 			print("\n%P", p1);
 		switch(copyu(p1, &p->to, nil)) {
 		case 0:	/* not used or set */
 			if((p->from.type == TYPE_REG && copyu(p1, &p->from, nil) > 1) ||
-			   (a.type == TYPE_REG && copyu(p1, &a, nil) > 1))
-				FAIL("args modified");
+			   (a.type == TYPE_REG && copyu(p1, &a, nil) > 1)) {
+				if(debug['P'])
+					print("\targs modified; FAILURE\n");
+				return 0;
+			}
 			continue;
-		case 3:	/* set, not used */
-			FAIL("BOTCH: noref");
+		case 3:	/* set, not used */ {
+			if(debug['P'])
+				print("\tBOTCH: noref; FAILURE\n");
+			return 0;
+		}
 		}
 		break;
 	}
 	/* check whether substitution can be done */
 	switch(p1->as) {
 	default:
-		FAIL("non-dpi");
+		if(debug['P'])
+			print("\tnon-dpi; FAILURE\n");
+		return 0;
+
 	case AAND:
 	case AEOR:
 	case AADD:
@@ -619,8 +635,11 @@ shiftprop(Flow *r)
 	case ARSB:
 	case ARSC:
 		if(p1->reg == n || (p1->reg == 0 && p1->to.type == TYPE_REG && p1->to.reg == n)) {
-			if(p1->from.type != TYPE_REG)
-				FAIL("can't swap");
+			if(p1->from.type != TYPE_REG) {
+				if(debug['P'])
+					print("\tcan't swap; FAILURE\n");
+				return 0;
+			}
 			p1->reg = p1->from.reg;
 			p1->from.reg = n;
 			switch(p1->as) {
@@ -644,15 +663,27 @@ shiftprop(Flow *r)
 	case ATST:
 	case ACMP:
 	case ACMN:
-		if(p1->reg == n)
-			FAIL("can't swap");
-		if(p1->reg == 0 && p1->to.reg == n)
-			FAIL("shift result used twice");
+		if(p1->reg == n) {
+			if(debug['P'])
+				print("\tcan't swap; FAILURE\n");
+			return 0;
+		}
+		if(p1->reg == 0 && p1->to.reg == n) {
+			if(debug['P'])
+				print("\tshift result used twice; FAILURE\n");
+			return 0;
+		}
 //	case AMVN:
-		if(p1->from.type == TYPE_SHIFT)
-			FAIL("shift result used in shift");
-		if(p1->from.type != TYPE_REG || p1->from.reg != n)
-			FAIL("BOTCH: where is it used?");
+		if(p1->from.type == TYPE_SHIFT) {
+			if(debug['P'])
+				print("\tshift result used in shift; FAILURE\n");
+			return 0;
+		}
+		if(p1->from.type != TYPE_REG || p1->from.reg != n) {
+			if(debug['P'])
+				print("\tBOTCH: where is it used?; FAILURE\n");
+			return 0;
+		}
 		break;
 	}
 	/* check whether shift result is used subsequently */
@@ -660,8 +691,11 @@ shiftprop(Flow *r)
 	if(p1->to.reg != n)
 	for (;;) {
 		r1 = uniqs(r1);
-		if(r1 == nil)
-			FAIL("inconclusive");
+		if(r1 == nil) {
+			if(debug['P'])
+				print("\tinconclusive; FAILURE\n");
+			return 0;
+		}
 		p1 = r1->prog;
 		if(debug['P'])
 			print("\n%P", p1);
@@ -671,7 +705,9 @@ shiftprop(Flow *r)
 		case 3: /* set, not used */
 			break;
 		default:/* used */
-			FAIL("reused");
+			if(debug['P'])
+				print("\treused; FAILURE\n");
+			return 0;
 		}
 		break;
 	}
@@ -941,7 +977,7 @@ xtramodes(Graph *g, Flow *r, Adr *a)
  * 4 if set and used
  * 0 otherwise (not touched)
  */
-int
+static int
 copyu(Prog *p, Adr *v, Adr *s)
 {
 	switch(p->as) {
@@ -1572,3 +1608,12 @@ smallindir(Addr *a, Addr *reg)
 		a->reg == reg->reg &&
 		0 <= a->offset && a->offset < 4096;
 }
+
+void
+excise(Flow *r)
+{
+	Prog *p;
+
+	p = r->prog;
+	nopout(p);
+}
diff --git a/src/cmd/5g/prog.c b/src/cmd/5g/prog.c
index a77f2336e9..9d5adefe69 100644
--- a/src/cmd/5g/prog.c
+++ b/src/cmd/5g/prog.c
@@ -5,7 +5,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
 enum
 {
@@ -148,4 +148,13 @@ proginfo(ProgInfo *info, Prog *p)
 	
 	if(((p->scond & C_SCOND) != C_SCOND_NONE) && (info->flags & RightWrite))
 		info->flags |= RightRead;
+	
+	switch(p->as) {
+	case ADIV:
+	case ADIVU:
+	case AMOD:
+	case AMODU:
+		info->regset |= RtoB(REG_R12);
+		break;
+	}
 }
diff --git a/src/cmd/5g/reg.c b/src/cmd/5g/reg.c
index 93090ebe42..1216e01bd5 100644
--- a/src/cmd/5g/reg.c
+++ b/src/cmd/5g/reg.c
@@ -32,66 +32,11 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
-#define	NREGVAR	32
-#define	REGBITS	((uint64)0xffffffffull)
-/*c2go enum {
+enum {
 	NREGVAR = 32,
-	REGBITS = 0xffffffff,
 };
-*/
-
-	void	addsplits(void);
-static	Reg*	firstr;
-static	int	first	= 1;
-
-int
-rcmp(const void *a1, const void *a2)
-{
-	Rgn *p1, *p2;
-	int c1, c2;
-
-	p1 = (Rgn*)a1;
-	p2 = (Rgn*)a2;
-	c1 = p2->cost;
-	c2 = p1->cost;
-	if(c1 -= c2)
-		return c1;
-	return p2->varno - p1->varno;
-}
-
-void
-excise(Flow *r)
-{
-	Prog *p;
-
-	p = r->prog;
-	nopout(p);
-}
-
-static void
-setaddrs(Bits bit)
-{
-	int i, n;
-	Var *v;
-	Node *node;
-
-	while(bany(&bit)) {
-		// convert each bit to a variable
-		i = bnum(bit);
-		node = var[i].node;
-		n = var[i].name;
-		biclr(&bit, i);
-
-		// disable all pieces of that variable
-		for(i=0; i<nvar; i++) {
-			v = var+i;
-			if(v->node == node && v->name == n)
-				v->addr = 2;
-		}
-	}
-}
 
 static char* regname[] = {
 	".R0",
@@ -128,1188 +73,26 @@ static char* regname[] = {
 	".F15",
 };
 
-static Node* regnodes[NREGVAR];
-
-static void walkvardef(Node *n, Reg *r, int active);
-
-void
-regopt(Prog *firstp)
-{
-	Reg *r, *r1;
-	Prog *p;
-	Graph *g;
-	int i, z, active;
-	uint32 vreg;
-	Bits bit;
-	ProgInfo info;
-
-	if(first) {
-		fmtinstall('Q', Qconv);
-		first = 0;
-	}
-
-	mergetemp(firstp);
-
-	/*
-	 * control flow is more complicated in generated go code
-	 * than in generated c code.  define pseudo-variables for
-	 * registers, so we have complete register usage information.
-	 */
-	nvar = NREGVAR;
-	memset(var, 0, NREGVAR*sizeof var[0]);
-	for(i=0; i<NREGVAR; i++) {
-		if(regnodes[i] == N)
-			regnodes[i] = newname(lookup(regname[i]));
-		var[i].node = regnodes[i];
-	}
-
-	regbits = RtoB(REGSP)|RtoB(REGLINK)|RtoB(REGPC);
-	for(z=0; z<BITS; z++) {
-		externs.b[z] = 0;
-		params.b[z] = 0;
-		consts.b[z] = 0;
-		addrs.b[z] = 0;
-		ivar.b[z] = 0;
-		ovar.b[z] = 0;
-	}
-
-	/*
-	 * pass 1
-	 * build aux data structure
-	 * allocate pcs
-	 * find use and set of variables
-	 */
-	g = flowstart(firstp, sizeof(Reg));
-	if(g == nil) {
-		for(i=0; i<nvar; i++)
-			var[i].node->opt = nil;
-		return;
-	}
-
-	firstr = (Reg*)g->start;
-
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		p = r->f.prog;
-		if(p->as == AVARDEF || p->as == AVARKILL)
-			continue;
-		proginfo(&info, p);
-
-		// Avoid making variables for direct-called functions.
-		if(p->as == ABL && p->to.name == NAME_EXTERN)
-			continue;
-
-		bit = mkvar(r, &p->from);
-		if(info.flags & LeftRead)
-			for(z=0; z<BITS; z++)
-				r->use1.b[z] |= bit.b[z];
-		if(info.flags & LeftAddr)
-			setaddrs(bit);
-
-		if(info.flags & RegRead)	
-			r->use1.b[0] |= RtoB(p->reg);
-
-		if(info.flags & (RightAddr | RightRead | RightWrite)) {
-			bit = mkvar(r, &p->to);
-			if(info.flags & RightAddr)
-				setaddrs(bit);
-			if(info.flags & RightRead)
-				for(z=0; z<BITS; z++)
-					r->use2.b[z] |= bit.b[z];
-			if(info.flags & RightWrite)
-				for(z=0; z<BITS; z++)
-					r->set.b[z] |= bit.b[z];
-		}
-
-		/* the mod/div runtime routines smash R12 */
-		if(p->as == ADIV || p->as == ADIVU || p->as == AMOD || p->as == AMODU)
-			r->set.b[0] |= RtoB(REG_R12);
-	}
-	if(firstr == R)
-		return;
-
-	for(i=0; i<nvar; i++) {
-		Var *v = var+i;
-		if(v->addr) {
-			bit = blsh(i);
-			for(z=0; z<BITS; z++)
-				addrs.b[z] |= bit.b[z];
-		}
-
-		if(debug['R'] && debug['v'])
-			print("bit=%2d addr=%d et=%-6E w=%-2d s=%N + %lld\n",
-				i, v->addr, v->etype, v->width, v->node, v->offset);
-	}
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass1", &firstr->f, 1);
-
-	/*
-	 * pass 2
-	 * find looping structure
-	 */
-	flowrpo(g);
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass2", &firstr->f, 1);
-
-	/*
-	 * pass 2.5
-	 * iterate propagating fat vardef covering forward
-	 * r->act records vars with a VARDEF since the last CALL.
-	 * (r->act will be reused in pass 5 for something else,
-	 * but we'll be done with it by then.)
-	 */
-	active = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		r->f.active = 0;
-		r->act = zbits;
-	}
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		p = r->f.prog;
-		if(p->as == AVARDEF && isfat(((Node*)(p->to.node))->type) && ((Node*)(p->to.node))->opt != nil) {
-			active++;
-			walkvardef(p->to.node, r, active);
-		}
-	}
-
-	/*
-	 * pass 3
-	 * iterate propagating usage
-	 * 	back until flow graph is complete
-	 */
-loop1:
-	change = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->f.active = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		if(r->f.prog->as == ARET)
-			prop(r, zbits, zbits);
-loop11:
-	/* pick up unreachable code */
-	i = 0;
-	for(r = firstr; r != R; r = r1) {
-		r1 = (Reg*)r->f.link;
-		if(r1 && r1->f.active && !r->f.active) {
-			prop(r, zbits, zbits);
-			i = 1;
-		}
-	}
-	if(i)
-		goto loop11;
-	if(change)
-		goto loop1;
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass3", &firstr->f, 1);
-
-
-	/*
-	 * pass 4
-	 * iterate propagating register/variable synchrony
-	 * 	forward until graph is complete
-	 */
-loop2:
-	change = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->f.active = 0;
-	synch(firstr, zbits);
-	if(change)
-		goto loop2;
-
-	addsplits();
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass4", &firstr->f, 1);
-
-	if(debug['R'] > 1) {
-		print("\nprop structure:\n");
-		for(r = firstr; r != R; r = (Reg*)r->f.link) {
-			print("%d:%P", r->f.loop, r->f.prog);
-			for(z=0; z<BITS; z++) {
-				bit.b[z] = r->set.b[z] |
-					r->refahead.b[z] | r->calahead.b[z] |
-					r->refbehind.b[z] | r->calbehind.b[z] |
-					r->use1.b[z] | r->use2.b[z];
-				bit.b[z] &= ~addrs.b[z];
-			}
-
-			if(bany(&bit)) {
-				print("\t");
-				if(bany(&r->use1))
-					print(" u1=%Q", r->use1);
-				if(bany(&r->use2))
-					print(" u2=%Q", r->use2);
-				if(bany(&r->set))
-					print(" st=%Q", r->set);
-				if(bany(&r->refahead))
-					print(" ra=%Q", r->refahead);
-				if(bany(&r->calahead))
-					print(" ca=%Q", r->calahead);
-				if(bany(&r->refbehind))
-					print(" rb=%Q", r->refbehind);
-				if(bany(&r->calbehind))
-					print(" cb=%Q", r->calbehind);
-			}
-			print("\n");
-		}
-	}
-
-	/*
-	 * pass 4.5
-	 * move register pseudo-variables into regu.
-	 */
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		r->regu = (r->refbehind.b[0] | r->set.b[0]) & REGBITS;
-
-		r->set.b[0] &= ~REGBITS;
-		r->use1.b[0] &= ~REGBITS;
-		r->use2.b[0] &= ~REGBITS;
-		r->refbehind.b[0] &= ~REGBITS;
-		r->refahead.b[0] &= ~REGBITS;
-		r->calbehind.b[0] &= ~REGBITS;
-		r->calahead.b[0] &= ~REGBITS;
-		r->regdiff.b[0] &= ~REGBITS;
-		r->act.b[0] &= ~REGBITS;
-	}
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass4.5", &firstr->f, 1);
-
-	/*
-	 * pass 5
-	 * isolate regions
-	 * calculate costs (paint1)
-	 */
-	r = firstr;
-	if(r) {
-		for(z=0; z<BITS; z++)
-			bit.b[z] = (r->refahead.b[z] | r->calahead.b[z]) &
-			  ~(externs.b[z] | params.b[z] | addrs.b[z] | consts.b[z]);
-		if(bany(&bit) && !r->f.refset) {
-			// should never happen - all variables are preset
-			if(debug['w'])
-				print("%L: used and not set: %Q\n", r->f.prog->lineno, bit);
-			r->f.refset = 1;
-		}
-	}
-
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->act = zbits;
-	rgp = region;
-	nregion = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		for(z=0; z<BITS; z++)
-			bit.b[z] = r->set.b[z] &
-			  ~(r->refahead.b[z] | r->calahead.b[z] | addrs.b[z]);
-		if(bany(&bit) && !r->f.refset) {
-			if(debug['w'])
-				print("%L: set and not used: %Q\n", r->f.prog->lineno, bit);
-			r->f.refset = 1;
-			excise(&r->f);
-		}
-		for(z=0; z<BITS; z++)
-			bit.b[z] = LOAD(r) & ~(r->act.b[z] | addrs.b[z]);
-		while(bany(&bit)) {
-			i = bnum(bit);
-			rgp->enter = r;
-			rgp->varno = i;
-			change = 0;
-			if(debug['R'] > 1)
-				print("\n");
-			paint1(r, i);
-			biclr(&bit, i);
-			if(change <= 0) {
-				if(debug['R'])
-					print("%L $%d: %Q\n",
-						r->f.prog->lineno, change, blsh(i));
-				continue;
-			}
-			rgp->cost = change;
-			nregion++;
-			if(nregion >= NRGN) {
-				if(debug['R'] > 1)
-					print("too many regions\n");
-				goto brk;
-			}
-			rgp++;
-		}
-	}
-brk:
-	qsort(region, nregion, sizeof(region[0]), rcmp);
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass5", &firstr->f, 1);
-
-	/*
-	 * pass 6
-	 * determine used registers (paint2)
-	 * replace code (paint3)
-	 */
-	rgp = region;
-	if(debug['R'] && debug['v'])
-		print("\nregisterizing\n");
-	for(i=0; i<nregion; i++) {
-		if(debug['R'] && debug['v'])
-			print("region %d: cost %d varno %d enter %lld\n", i, rgp->cost, rgp->varno, rgp->enter->f.prog->pc);
-		bit = blsh(rgp->varno);
-		vreg = paint2(rgp->enter, rgp->varno, 0);
-		vreg = allreg(vreg, rgp);
-		if(debug['R']) {
-			print("%L $%d %R: %Q\n",
-				rgp->enter->f.prog->lineno,
-				rgp->cost,
-				rgp->regno,
-				bit);
-		}
-		if(rgp->regno != 0)
-			paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
-		rgp++;
-	}
-
-	/*
-	 * free aux structures. peep allocates new ones.
-	 */
-	for(i=0; i<nvar; i++)
-		var[i].node->opt = nil;
-	flowend(g);
-	firstr = R;
-
-	if(debug['R'] && debug['v']) {
-		// Rebuild flow graph, since we inserted instructions
-		g = flowstart(firstp, sizeof(Reg));
-		firstr = (Reg*)g->start;
-		dumpit("pass6", &firstr->f, 1);
-		flowend(g);
-		firstr = R;
-	}
-
-	/*
-	 * pass 7
-	 * peep-hole on basic block
-	 */
-	if(!debug['R'] || debug['P']) {
-		peep(firstp);
-	}
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass7", &firstr->f, 1);
-
-	/*
-	 * last pass
-	 * eliminate nops
-	 * free aux structures
-	 * adjust the stack pointer
-	 *	MOVW.W 	R1,-12(R13)			<<- start
-	 *	MOVW   	R0,R1
-	 *	MOVW   	R1,8(R13)
-	 *	MOVW   	$0,R1
-	 *	MOVW   	R1,4(R13)
-	 *	BL     	,runtime.newproc+0(SB)
-	 *	MOVW   	&ft+-32(SP),R7			<<- adjust
-	 *	MOVW   	&j+-40(SP),R6			<<- adjust
-	 *	MOVW   	autotmp_0003+-24(SP),R5		<<- adjust
-	 *	MOVW   	$12(R13),R13			<<- finish
-	 */
-	vreg = 0;
-	for(p = firstp; p != P; p = p->link) {
-		while(p->link != P && p->link->as == ANOP)
-			p->link = p->link->link;
-		if(p->to.type == TYPE_BRANCH)
-			while(p->to.u.branch != P && p->to.u.branch->as == ANOP)
-				p->to.u.branch = p->to.u.branch->link;
-		if(p->as == AMOVW && p->to.reg == 13) {
-			if(p->scond & C_WBIT) {
-				vreg = -p->to.offset;		// in adjust region
-//				print("%P adjusting %d\n", p, vreg);
-				continue;
-			}
-			if(p->from.type == TYPE_CONST && p->to.type == TYPE_REG) {
-				if(p->from.offset != vreg)
-					print("in and out different\n");
-//				print("%P finish %d\n", p, vreg);
-				vreg = 0;	// done adjust region
-				continue;
-			}
-
-//			print("%P %d %d from type\n", p, p->from.type, TYPE_CONST);
-//			print("%P %d %d to type\n\n", p, p->to.type, TYPE_REG);
-		}
-
-		if(p->as == AMOVW && vreg != 0) {
-			if(p->from.sym != nil)
-			if(p->from.name == NAME_AUTO || p->from.name == NAME_PARAM) {
-				p->from.offset += vreg;
-//				print("%P adjusting from %d %d\n", p, vreg, p->from.type);
-			}
-			if(p->to.sym != nil)
-			if(p->to.name == NAME_AUTO || p->to.name == NAME_PARAM) {
-				p->to.offset += vreg;
-//				print("%P adjusting to %d %d\n", p, vreg, p->from.type);
-			}
-		}
-	}
-}
-
-static void
-walkvardef(Node *n, Reg *r, int active)
-{
-	Reg *r1, *r2;
-	int bn;
-	Var *v;
-	
-	for(r1=r; r1!=R; r1=(Reg*)r1->f.s1) {
-		if(r1->f.active == active)
-			break;
-		r1->f.active = active;
-		if(r1->f.prog->as == AVARKILL && r1->f.prog->to.node == n)
-			break;
-		for(v=n->opt; v!=nil; v=v->nextinnode) {
-			bn = v - var;
-			biset(&r1->act, bn);
-		}
-		if(r1->f.prog->as == ABL)
-			break;
-	}
-
-	for(r2=r; r2!=r1; r2=(Reg*)r2->f.s1)
-		if(r2->f.s2 != nil)
-			walkvardef(n, (Reg*)r2->f.s2, active);
-}
-
-void
-addsplits(void)
-{
-	Reg *r, *r1;
-	int z, i;
-	Bits bit;
-
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		if(r->f.loop > 1)
-			continue;
-		if(r->f.prog->as == ABL)
-			continue;
-		if(r->f.prog->as == ADUFFZERO)
-			continue;
-		if(r->f.prog->as == ADUFFCOPY)
-			continue;
-		for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link) {
-			if(r1->f.loop <= 1)
-				continue;
-			for(z=0; z<BITS; z++)
-				bit.b[z] = r1->calbehind.b[z] &
-					(r->refahead.b[z] | r->use1.b[z] | r->use2.b[z]) &
-					~(r->calahead.b[z] & addrs.b[z]);
-			while(bany(&bit)) {
-				i = bnum(bit);
-				biclr(&bit, i);
-			}
-		}
-	}
-}
-
-/*
- * add mov b,rn
- * just after r
- */
-void
-addmove(Reg *r, int bn, int rn, int f)
-{
-	Prog *p, *p1, *p2;
-	Adr *a;
-	Var *v;
-
-	p1 = mal(sizeof(*p1));
-	*p1 = zprog;
-	p = r->f.prog;
-	
-	// If there's a stack fixup coming (after BL newproc or BL deferproc),
-	// delay the load until after the fixup.
-	p2 = p->link;
-	if(p2 && p2->as == AMOVW && p2->from.type == TYPE_ADDR && p2->from.reg == REGSP && p2->to.reg == REGSP && p2->to.type == TYPE_REG)
-		p = p2;
-
-	p1->link = p->link;
-	p->link = p1;
-	p1->lineno = p->lineno;
-
-	v = var + bn;
-
-	a = &p1->to;
-	a->name = v->name;
-	a->node = v->node;
-	a->sym = linksym(v->node->sym);
-	a->offset = v->offset;
-	a->etype = v->etype;
-	a->type = TYPE_MEM;
-	if(a->etype == TARRAY)
-		a->type = TYPE_ADDR;
-	else if(a->sym == nil)
-		a->type = TYPE_CONST;
-
-	if(v->addr)
-		fatal("addmove: shouldn't be doing this %A\n", a);
-
-	switch(v->etype) {
-	default:
-		print("What is this %E\n", v->etype);
-
-	case TINT8:
-		p1->as = AMOVBS;
-		break;
-	case TBOOL:
-	case TUINT8:
-//print("movbu %E %d %S\n", v->etype, bn, v->sym);
-		p1->as = AMOVBU;
-		break;
-	case TINT16:
-		p1->as = AMOVHS;
-		break;
-	case TUINT16:
-		p1->as = AMOVHU;
-		break;
-	case TINT32:
-	case TUINT32:
-	case TPTR32:
-		p1->as = AMOVW;
-		break;
-	case TFLOAT32:
-		p1->as = AMOVF;
-		break;
-	case TFLOAT64:
-		p1->as = AMOVD;
-		break;
-	}
-
-	p1->from.type = TYPE_REG;
-	p1->from.reg = rn;
-	if(!f) {
-		p1->from = *a;
-		*a = zprog.from;
-		a->type = TYPE_REG;
-		a->reg = rn;
-		if(v->etype == TUINT8 || v->etype == TBOOL)
-			p1->as = AMOVBU;
-		if(v->etype == TUINT16)
-			p1->as = AMOVHU;
-	}
-	if(debug['R'])
-		print("%P\t.a%P\n", p, p1);
-}
-
-static int
-overlap(int32 o1, int w1, int32 o2, int w2)
-{
-	int32 t1, t2;
-
-	t1 = o1+w1;
-	t2 = o2+w2;
-
-	if(!(t1 > o2 && t2 > o1))
-		return 0;
-
-	return 1;
-}
-
-Bits
-mkvar(Reg *r, Adr *a)
-{
-	Var *v;
-	int i, t, n, et, z, w, flag;
-	int32 o;
-	Bits bit;
-	Node *node;
-
-	// mark registers used
-	t = a->type;
-
-	flag = 0;
-	switch(t) {
-	default:
-		print("type %d %d %D\n", t, a->name, a);
-		goto none;
-
-	case TYPE_NONE:
-	case TYPE_FCONST:
-	case TYPE_BRANCH:
-		break;
-
-
-	case TYPE_REGREG:
-	case TYPE_REGREG2:
-		bit = zbits;
-		if(a->offset != 0)
-			bit.b[0] |= RtoB(a->offset);
-		if(a->reg != 0)
-			bit.b[0] |= RtoB(a->reg);
-		return bit;
-	
-	case TYPE_CONST:
-		if(a->reg != 0)
-			fatal("found CONST instead of ADDR: %D", a);
-		break;
-
-	case TYPE_ADDR:
-	case TYPE_REG:
-	case TYPE_SHIFT:
-		if(a->reg != 0) {
-			bit = zbits;
-			bit.b[0] = RtoB(a->reg);
-			return bit;
-		}
-		break;
-
-	case TYPE_MEM:
-		if(a->reg != 0) {
-			if(a == &r->f.prog->from)
-				r->use1.b[0] |= RtoB(a->reg);
-			else
-				r->use2.b[0] |= RtoB(a->reg);
-			if(r->f.prog->scond & (C_PBIT|C_WBIT))
-				r->set.b[0] |= RtoB(a->reg);
-		}
-		break;
-	}
-
-	switch(a->name) {
-	default:
-		goto none;
-
-	case NAME_EXTERN:
-	case NAME_STATIC:
-	case NAME_AUTO:
-	case NAME_PARAM:
-		n = a->name;
-		break;
-	}
-
-	node = a->node;
-	if(node == N || node->op != ONAME || node->orig == N)
-		goto none;
-	node = node->orig;
-	if(node->orig != node)
-		fatal("%D: bad node", a);
-	if(node->sym == S || node->sym->name[0] == '.')
-		goto none;
-	et = a->etype;
-	o = a->offset;
-	w = a->width;
-	if(w < 0)
-		fatal("bad width %d for %D", w, a);
-
-	for(i=0; i<nvar; i++) {
-		v = var+i;
-		if(v->node == node && v->name == n) {
-			if(v->offset == o)
-			if(v->etype == et)
-			if(v->width == w)
-				if(!flag)
-					return blsh(i);
-
-			// if they overlap, disable both
-			if(overlap(v->offset, v->width, o, w)) {
-				v->addr = 1;
-				flag = 1;
-			}
-		}
-	}
-
-	switch(et) {
-	case 0:
-	case TFUNC:
-		goto none;
-	}
-
-	if(nvar >= NVAR) {
-		if(debug['w'] > 1 && node)
-			fatal("variable not optimized: %D", a);
-		
-		// If we're not tracking a word in a variable, mark the rest as
-		// having its address taken, so that we keep the whole thing
-		// live at all calls. otherwise we might optimize away part of
-		// a variable but not all of it.
-		for(i=0; i<nvar; i++) {
-			v = var+i;
-			if(v->node == node)
-				v->addr = 1;
-		}
-		goto none;
-	}
-
-	i = nvar;
-	nvar++;
-//print("var %d %E %D %S\n", i, et, a, s);
-	v = var+i;
-	v->offset = o;
-	v->name = n;
-	v->etype = et;
-	v->width = w;
-	v->addr = flag;		// funny punning
-	v->node = node;
-	
-	// node->opt is the head of a linked list
-	// of Vars within the given Node, so that
-	// we can start at a Var and find all the other
-	// Vars in the same Go variable.
-	v->nextinnode = node->opt;
-	node->opt = v;
-	
-	bit = blsh(i);
-	if(n == NAME_EXTERN || n == NAME_STATIC)
-		for(z=0; z<BITS; z++)
-			externs.b[z] |= bit.b[z];
-	if(n == NAME_PARAM)
-		for(z=0; z<BITS; z++)
-			params.b[z] |= bit.b[z];
-
-	if(node->class == PPARAM)
-		for(z=0; z<BITS; z++)
-			ivar.b[z] |= bit.b[z];
-	if(node->class == PPARAMOUT)
-		for(z=0; z<BITS; z++)
-			ovar.b[z] |= bit.b[z];
-
-	// Treat values with their address taken as live at calls,
-	// because the garbage collector's liveness analysis in ../gc/plive.c does.
-	// These must be consistent or else we will elide stores and the garbage
-	// collector will see uninitialized data.
-	// The typical case where our own analysis is out of sync is when the
-	// node appears to have its address taken but that code doesn't actually
-	// get generated and therefore doesn't show up as an address being
-	// taken when we analyze the instruction stream.
-	// One instance of this case is when a closure uses the same name as
-	// an outer variable for one of its own variables declared with :=.
-	// The parser flags the outer variable as possibly shared, and therefore
-	// sets addrtaken, even though it ends up not being actually shared.
-	// If we were better about _ elision, _ = &x would suffice too.
-	// The broader := in a closure problem is mentioned in a comment in
-	// closure.c:/^typecheckclosure and dcl.c:/^oldname.
-	if(node->addrtaken)
-		v->addr = 1;
-
-	// Disable registerization for globals, because:
-	// (1) we might panic at any time and we want the recovery code
-	// to see the latest values (issue 1304).
-	// (2) we don't know what pointers might point at them and we want
-	// loads via those pointers to see updated values and vice versa (issue 7995).
-	//
-	// Disable registerization for results if using defer, because the deferred func
-	// might recover and return, causing the current values to be used.
-	if(node->class == PEXTERN || (hasdefer && node->class == PPARAMOUT))
-		v->addr = 1;
-
-	if(debug['R'])
-		print("bit=%2d et=%2E w=%d+%d %#N %D flag=%d\n", i, et, o, w, node, a, v->addr);
-
-	return bit;
-
-none:
-	return zbits;
-}
-
-void
-prop(Reg *r, Bits ref, Bits cal)
+char**
+regnames(int *n)
 {
-	Reg *r1, *r2;
-	int z, i, j;
-	Var *v, *v1;
-
-	for(r1 = r; r1 != R; r1 = (Reg*)r1->f.p1) {
-		for(z=0; z<BITS; z++) {
-			ref.b[z] |= r1->refahead.b[z];
-			if(ref.b[z] != r1->refahead.b[z]) {
-				r1->refahead.b[z] = ref.b[z];
-				change++;
-			}
-			cal.b[z] |= r1->calahead.b[z];
-			if(cal.b[z] != r1->calahead.b[z]) {
-				r1->calahead.b[z] = cal.b[z];
-				change++;
-			}
-		}
-		switch(r1->f.prog->as) {
-		case ABL:
-			if(noreturn(r1->f.prog))
-				break;
-
-			// Mark all input variables (ivar) as used, because that's what the
-			// liveness bitmaps say. The liveness bitmaps say that so that a
-			// panic will not show stale values in the parameter dump.
-			// Mark variables with a recent VARDEF (r1->act) as used,
-			// so that the optimizer flushes initializations to memory,
-			// so that if a garbage collection happens during this CALL,
-			// the collector will see initialized memory. Again this is to
-			// match what the liveness bitmaps say.
-			for(z=0; z<BITS; z++) {
-				cal.b[z] |= ref.b[z] | externs.b[z] | ivar.b[z] | r1->act.b[z];
-				ref.b[z] = 0;
-			}
-			
-			// cal.b is the current approximation of what's live across the call.
-			// Every bit in cal.b is a single stack word. For each such word,
-			// find all the other tracked stack words in the same Go variable
-			// (struct/slice/string/interface) and mark them live too.
-			// This is necessary because the liveness analysis for the garbage
-			// collector works at variable granularity, not at word granularity.
-			// It is fundamental for slice/string/interface: the garbage collector
-			// needs the whole value, not just some of the words, in order to
-			// interpret the other bits correctly. Specifically, slice needs a consistent
-			// ptr and cap, string needs a consistent ptr and len, and interface
-			// needs a consistent type word and data word.
-			for(z=0; z<BITS; z++) {
-				if(cal.b[z] == 0)
-					continue;
-				for(i=0; i<64; i++) {
-					if(z*64+i >= nvar || ((cal.b[z]>>i)&1) == 0)
-						continue;
-					v = var+z*64+i;
-					if(v->node->opt == nil) // v represents fixed register, not Go variable
-						continue;
-
-					// v->node->opt is the head of a linked list of Vars
-					// corresponding to tracked words from the Go variable v->node.
-					// Walk the list and set all the bits.
-					// For a large struct this could end up being quadratic:
-					// after the first setting, the outer loop (for z, i) would see a 1 bit
-					// for all of the remaining words in the struct, and for each such
-					// word would go through and turn on all the bits again.
-					// To avoid the quadratic behavior, we only turn on the bits if
-					// v is the head of the list or if the head's bit is not yet turned on.
-					// This will set the bits at most twice, keeping the overall loop linear.
-					v1 = v->node->opt;
-					j = v1 - var;
-					if(v == v1 || !btest(&cal, j)) {
-						for(; v1 != nil; v1 = v1->nextinnode) {
-							j = v1 - var;
-							biset(&cal, j);
-						}
-					}
-				}
-			}
-			break;
-
-		case ATEXT:
-			for(z=0; z<BITS; z++) {
-				cal.b[z] = 0;
-				ref.b[z] = 0;
-			}
-			break;
-
-		case ARET:
-			for(z=0; z<BITS; z++) {
-				cal.b[z] = externs.b[z] | ovar.b[z];
-				ref.b[z] = 0;
-			}
-			break;
-		}
-		for(z=0; z<BITS; z++) {
-			ref.b[z] = (ref.b[z] & ~r1->set.b[z]) |
-				r1->use1.b[z] | r1->use2.b[z];
-			cal.b[z] &= ~(r1->set.b[z] | r1->use1.b[z] | r1->use2.b[z]);
-			r1->refbehind.b[z] = ref.b[z];
-			r1->calbehind.b[z] = cal.b[z];
-		}
-		if(r1->f.active)
-			break;
-		r1->f.active = 1;
-	}
-	for(; r != r1; r = (Reg*)r->f.p1)
-		for(r2 = (Reg*)r->f.p2; r2 != R; r2 = (Reg*)r2->f.p2link)
-			prop(r2, r->refbehind, r->calbehind);
+	*n = NREGVAR;
+	return regname;
 }
 
-void
-synch(Reg *r, Bits dif)
+uint64
+excludedregs(void)
 {
-	Reg *r1;
-	int z;
-
-	for(r1 = r; r1 != R; r1 = (Reg*)r1->f.s1) {
-		for(z=0; z<BITS; z++) {
-			dif.b[z] = (dif.b[z] &
-				~(~r1->refbehind.b[z] & r1->refahead.b[z])) |
-					r1->set.b[z] | r1->regdiff.b[z];
-			if(dif.b[z] != r1->regdiff.b[z]) {
-				r1->regdiff.b[z] = dif.b[z];
-				change++;
-			}
-		}
-		if(r1->f.active)
-			break;
-		r1->f.active = 1;
-		for(z=0; z<BITS; z++)
-			dif.b[z] &= ~(~r1->calbehind.b[z] & r1->calahead.b[z]);
-		if(r1->f.s2 != nil)
-			synch((Reg*)r1->f.s2, dif);
-	}
+	return RtoB(REGSP)|RtoB(REGLINK)|RtoB(REGPC);
 }
 
-uint32
-allreg(uint32 b, Rgn *r)
+uint64
+doregbits(int r)
 {
-	Var *v;
-	int i;
-
-	v = var + r->varno;
-	r->regno = 0;
-	switch(v->etype) {
-
-	default:
-		fatal("unknown etype %d/%E", bitno(b), v->etype);
-		break;
-
-	case TINT8:
-	case TUINT8:
-	case TINT16:
-	case TUINT16:
-	case TINT32:
-	case TUINT32:
-	case TINT:
-	case TUINT:
-	case TUINTPTR:
-	case TBOOL:
-	case TPTR32:
-		i = BtoR(~b);
-		if(i && r->cost >= 0) {
-			r->regno = i;
-			return RtoB(i);
-		}
-		break;
-
-	case TFLOAT32:
-	case TFLOAT64:
-		i = BtoF(~b);
-		if(i && r->cost >= 0) {
-			r->regno = i;
-			return RtoB(i);
-		}
-		break;
-
-	case TINT64:
-	case TUINT64:
-	case TPTR64:
-	case TINTER:
-	case TSTRUCT:
-	case TARRAY:
-		break;
-	}
+	USED(r);
 	return 0;
 }
 
-void
-paint1(Reg *r, int bn)
-{
-	Reg *r1;
-	Prog *p;
-	int z;
-	uint64 bb;
-
-	z = bn/64;
-	bb = 1LL<<(bn%64);
-	if(r->act.b[z] & bb)
-		return;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(r1->act.b[z] & bb)
-			break;
-		r = r1;
-	}
-
-	if(LOAD(r) & ~(r->set.b[z] & ~(r->use1.b[z]|r->use2.b[z])) & bb) {
-		change -= CLOAD * r->f.loop;
-		if(debug['R'] > 1)
-			print("%d%P\td %Q $%d\n", r->f.loop,
-				r->f.prog, blsh(bn), change);
-	}
-	for(;;) {
-		r->act.b[z] |= bb;
-		p = r->f.prog;
-
-
-		if(r->f.prog->as != ANOP) { // don't give credit for NOPs
-			if(r->use1.b[z] & bb) {
-				change += CREF * r->f.loop;
-				if(debug['R'] > 1)
-					print("%d%P\tu1 %Q $%d\n", r->f.loop,
-						p, blsh(bn), change);
-			}
-			if((r->use2.b[z]|r->set.b[z]) & bb) {
-				change += CREF * r->f.loop;
-				if(debug['R'] > 1)
-					print("%d%P\tu2 %Q $%d\n", r->f.loop,
-						p, blsh(bn), change);
-			}
-		}
-
-		if(STORE(r) & r->regdiff.b[z] & bb) {
-			change -= CLOAD * r->f.loop;
-			if(debug['R'] > 1)
-				print("%d%P\tst %Q $%d\n", r->f.loop,
-					p, blsh(bn), change);
-		}
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					paint1(r1, bn);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				paint1(r1, bn);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(r->act.b[z] & bb)
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-}
-
-uint32
-paint2(Reg *r, int bn, int depth)
-{
-	Reg *r1;
-	int z;
-	uint64 bb, vreg;
-
-	z = bn/64;
-	bb = 1LL << (bn%64);
-	vreg = regbits;
-	if(!(r->act.b[z] & bb))
-		return vreg;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(!(r1->act.b[z] & bb))
-			break;
-		r = r1;
-	}
-	for(;;) {
-		if(debug['R'] && debug['v'])
-			print("  paint2 %d %P\n", depth, r->f.prog);
-
-		r->act.b[z] &= ~bb;
-
-		vreg |= r->regu;
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					vreg |= paint2(r1, bn, depth+1);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				vreg |= paint2(r1, bn, depth+1);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(!(r->act.b[z] & bb))
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-	return vreg;
-}
-
-void
-paint3(Reg *r, int bn, uint32 rb, int rn)
-{
-	Reg *r1;
-	Prog *p;
-	int z;
-	uint64 bb;
-
-	z = bn/64;
-	bb = 1LL << (bn%64);
-	if(r->act.b[z] & bb)
-		return;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(r1->act.b[z] & bb)
-			break;
-		r = r1;
-	}
-
-	if(LOAD(r) & ~(r->set.b[z] & ~(r->use1.b[z]|r->use2.b[z])) & bb)
-		addmove(r, bn, rn, 0);
-
-	for(;;) {
-		r->act.b[z] |= bb;
-		p = r->f.prog;
-
-		if(r->use1.b[z] & bb) {
-			if(debug['R'])
-				print("%P", p);
-			addreg(&p->from, rn);
-			if(debug['R'])
-				print("\t.c%P\n", p);
-		}
-		if((r->use2.b[z]|r->set.b[z]) & bb) {
-			if(debug['R'])
-				print("%P", p);
-			addreg(&p->to, rn);
-			if(debug['R'])
-				print("\t.c%P\n", p);
-		}
-
-		if(STORE(r) & r->regdiff.b[z] & bb)
-			addmove(r, bn, rn, 1);
-		r->regu |= rb;
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					paint3(r1, bn, rb, rn);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				paint3(r1, bn, rb, rn);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(r->act.b[z] & bb)
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-}
-
-void
-addreg(Adr *a, int rn)
-{
-	a->sym = nil;
-	a->node = nil;
-	a->name = NAME_NONE;
-	a->type = TYPE_REG;
-	a->reg = rn;
-}
-
 /*
  *	bit	reg
  *	0	R0
@@ -1324,26 +107,26 @@ addreg(Adr *a, int rn)
  *	...	...
  *	31	F15
  */
-uint32
+uint64
 RtoB(int r)
 {
 	if(REG_R0 <= r && r <= REG_R15) {
 		if(r >= REGTMP-2 && r != REG_R12)	// excluded R9 and R10 for m and g, but not R12
 			return 0;
-		return 1L << (r - REG_R0);
+		return 1ULL << (r - REG_R0);
 	}
 	
 	if(REG_F0 <= r && r <= REG_F15) {
 		if(r < REG_F2 || r > REG_F0+NFREG-1)
 			return 0;
-		return 1L << ((r - REG_F0) + 16);
+		return 1ULL << ((r - REG_F0) + 16);
 	}
 	
 	return 0;
 }
 
 int
-BtoR(uint32 b)
+BtoR(uint64 b)
 {
 	// TODO Allow R0 and R1, but be careful with a 0 return
 	// TODO Allow R9. Only R10 is reserved now (just g, not m).
@@ -1354,88 +137,10 @@ BtoR(uint32 b)
 }
 
 int
-BtoF(uint32 b)
+BtoF(uint64 b)
 {
 	b &= 0xfffc0000L;
 	if(b == 0)
 		return 0;
 	return bitno(b) - 16 + REG_F0;
 }
-
-void
-dumpone(Flow *f, int isreg)
-{
-	int z;
-	Bits bit;
-	Reg *r;
-
-	print("%d:%P", f->loop, f->prog);
-	if(isreg) {
-		r = (Reg*)f;
-		for(z=0; z<BITS; z++)
-			bit.b[z] =
-				r->set.b[z] |
-				r->use1.b[z] |
-				r->use2.b[z] |
-				r->refbehind.b[z] |
-				r->refahead.b[z] |
-				r->calbehind.b[z] |
-				r->calahead.b[z] |
-				r->regdiff.b[z] |
-				r->act.b[z] |
-					0;
-		if(bany(&bit)) {
-			print("\t");
-			if(bany(&r->set))
-				print(" s:%Q", r->set);
-			if(bany(&r->use1))
-				print(" u1:%Q", r->use1);
-			if(bany(&r->use2))
-				print(" u2:%Q", r->use2);
-			if(bany(&r->refbehind))
-				print(" rb:%Q ", r->refbehind);
-			if(bany(&r->refahead))
-				print(" ra:%Q ", r->refahead);
-			if(bany(&r->calbehind))
-				print(" cb:%Q ", r->calbehind);
-			if(bany(&r->calahead))
-				print(" ca:%Q ", r->calahead);
-			if(bany(&r->regdiff))
-				print(" d:%Q ", r->regdiff);
-			if(bany(&r->act))
-				print(" a:%Q ", r->act);
-		}
-	}
-	print("\n");
-}
-
-void
-dumpit(char *str, Flow *r0, int isreg)
-{
-	Flow *r, *r1;
-
-	print("\n%s\n", str);
-	for(r = r0; r != nil; r = r->link) {
-		dumpone(r, isreg);
-		r1 = r->p2;
-		if(r1 != nil) {
-			print("	pred:");
-			for(; r1 != nil; r1 = r1->p2link)
-				print(" %.4ud", (int)r1->prog->pc);
-			if(r->p1 != nil)
-				print(" (and %.4ud)", (int)r->p1->prog->pc);
-			else
-				print(" (only)");
-			print("\n");
-		}
-		// Print successors if it's not just the next one
-		if(r->s1 != r->link || r->s2 != nil) {
-			print("	succ:");
-			if(r->s1 != nil)
-				print(" %.4ud", (int)r->s1->prog->pc);
-			if(r->s2 != nil)
-				print(" %.4ud", (int)r->s2->prog->pc);
-			print("\n");
-		}
-	}
-}
diff --git a/src/cmd/5l/5.out.h b/src/cmd/5l/5.out.h
index fa1dcc34e8..a266a6b87c 100644
--- a/src/cmd/5l/5.out.h
+++ b/src/cmd/5l/5.out.h
@@ -91,6 +91,7 @@ enum
 	REGG = REGEXT-0,
 	REGM = REGEXT-1,
 
+	REGCTXT = REG_R7,
 	REGTMP = REG_R11,
 	REGSP = REG_R13,
 	REGLINK = REG_R14,
diff --git a/src/cmd/5l/asm.c b/src/cmd/5l/asm.c
index 5993079126..ed4d68a1fb 100644
--- a/src/cmd/5l/asm.c
+++ b/src/cmd/5l/asm.c
@@ -33,16 +33,9 @@
 #include	"l.h"
 #include	"../ld/lib.h"
 #include	"../ld/elf.h"
+#include	"../ld/macho.h"
 #include	"../ld/dwarf.h"
 
-
-char linuxdynld[] = "/lib/ld-linux.so.3"; // 2 for OABI, 3 for EABI
-char freebsddynld[] = "/usr/libexec/ld-elf.so.1";
-char openbsddynld[] = "XXX";
-char netbsddynld[] = "/libexec/ld.elf_so";
-char dragonflydynld[] = "XXX";
-char solarisdynld[] = "XXX";
-
 static int
 needlib(char *name)
 {
@@ -63,8 +56,6 @@ needlib(char *name)
 	return 0;
 }
 
-int	nelfsym = 1;
-
 static void	addpltsym(Link*, LSym*);
 static void	addgotsym(Link*, LSym*);
 static void	addgotsyminternal(Link*, LSym*);
@@ -127,11 +118,11 @@ adddynrel(LSym *s, Reloc *r)
 			addgotsym(ctxt, targ);
 		}
 		r->type = R_CONST;	// write r->add during relocsym
-		r->sym = S;
+		r->sym = nil;
 		r->add += targ->got;
 		return;
 
-	case 256 + R_ARM_GOT_PREL: // GOT(S) + A - P
+	case 256 + R_ARM_GOT_PREL: // GOT(nil) + A - nil
 		if(targ->type != SDYNIMPORT) {
 			addgotsyminternal(ctxt, targ);
 		} else {
@@ -178,7 +169,7 @@ adddynrel(LSym *s, Reloc *r)
 			// R_ARM_V4BX is ABS relocation, so this symbol is a dummy symbol, ignore it
 			r->sym->type = 0;
 		}
-		r->sym = S;
+		r->sym = nil;
 		return;
 
 	case 256 + R_ARM_PC24:
@@ -210,9 +201,9 @@ adddynrel(LSym *s, Reloc *r)
 			adddynsym(ctxt, targ);
 			rel = linklookup(ctxt, ".rel", 0);
 			addaddrplus(ctxt, rel, s, r->off);
-			adduint32(ctxt, rel, ELF32_R_INFO(targ->dynid, R_ARM_GLOB_DAT)); // we need a S + A dynmic reloc
+			adduint32(ctxt, rel, ELF32_R_INFO(targ->dynid, R_ARM_GLOB_DAT)); // we need a nil + A dynmic reloc
 			r->type = R_CONST;	// write r->add during relocsym
-			r->sym = S;
+			r->sym = nil;
 			return;
 		}
 		break;
@@ -227,7 +218,7 @@ elfreloc1(Reloc *r, vlong sectoff)
 {
 	int32 elfsym;
 	
-	LPUT(sectoff);
+	thearch.lput(sectoff);
 
 	elfsym = r->xsym->elfsym;
 	switch(r->type) {
@@ -236,14 +227,14 @@ elfreloc1(Reloc *r, vlong sectoff)
 
 	case R_ADDR:
 		if(r->siz == 4)
-			LPUT(R_ARM_ABS32 | elfsym<<8);
+			thearch.lput(R_ARM_ABS32 | elfsym<<8);
 		else
 			return -1;
 		break;
 
 	case R_PCREL:
 		if(r->siz == 4)
-			LPUT(R_ARM_REL32 | elfsym<<8);
+			thearch.lput(R_ARM_REL32 | elfsym<<8);
 		else
 			return -1;
 		break;
@@ -251,9 +242,9 @@ elfreloc1(Reloc *r, vlong sectoff)
 	case R_CALLARM:
 		if(r->siz == 4) {
 			if((r->add & 0xff000000) == 0xeb000000) // BL
-				LPUT(R_ARM_CALL | elfsym<<8);
+				thearch.lput(R_ARM_CALL | elfsym<<8);
 			else
-				LPUT(R_ARM_JUMP24 | elfsym<<8);
+				thearch.lput(R_ARM_JUMP24 | elfsym<<8);
 		} else
 			return -1;
 		break;
@@ -261,9 +252,9 @@ elfreloc1(Reloc *r, vlong sectoff)
 	case R_TLS:
 		if(r->siz == 4) {
 			if(flag_shared)
-				LPUT(R_ARM_TLS_IE32 | elfsym<<8);
+				thearch.lput(R_ARM_TLS_IE32 | elfsym<<8);
 			else
-				LPUT(R_ARM_TLS_LE32 | elfsym<<8);
+				thearch.lput(R_ARM_TLS_LE32 | elfsym<<8);
 		} else
 			return -1;
 		break;
@@ -301,10 +292,58 @@ elfsetupplt(void)
 int
 machoreloc1(Reloc *r, vlong sectoff)
 {
-	USED(r);
-	USED(sectoff);
+	uint32 v;
+	LSym *rs;
 
-	return -1;
+	rs = r->xsym;
+
+	if(rs->type == SHOSTOBJ || r->type == R_CALLARM) {
+		if(rs->dynid < 0) {
+			diag("reloc %d to non-macho symbol %s type=%d", r->type, rs->name, rs->type);
+			return -1;
+		}
+		v = rs->dynid;
+		v |= 1<<27; // external relocation
+	} else {
+		v = rs->sect->extnum;
+		if(v == 0) {
+			diag("reloc %d to symbol %s in non-macho section %s type=%d", r->type, rs->name, rs->sect->name, rs->type);
+			return -1;
+		}
+	}
+
+	switch(r->type) {
+	default:
+		return -1;
+	case R_ADDR:
+		v |= MACHO_GENERIC_RELOC_VANILLA<<28;
+		break;
+	case R_CALLARM:
+		v |= 1<<24; // pc-relative bit
+		v |= MACHO_ARM_RELOC_BR24<<28;
+		break;
+	}
+
+	switch(r->siz) {
+	default:
+		return -1;
+	case 1:
+		v |= 0<<25;
+		break;
+	case 2:
+		v |= 1<<25;
+		break;
+	case 4:
+		v |= 2<<25;
+		break;
+	case 8:
+		v |= 3<<25;
+		break;
+	}
+
+	thearch.lput(sectoff);
+	thearch.lput(v);
+	return 0;
 }
 
 
@@ -333,6 +372,14 @@ archreloc(Reloc *r, LSym *s, vlong *val)
 				diag("missing section for %s", rs->name);
 			r->xsym = rs;
 
+			// ld64 for arm seems to want the symbol table to contain offset
+			// into the section rather than pseudo virtual address that contains
+			// the section load address.
+			// we need to compensate that by removing the instruction's address
+			// from addend.
+			if(HEADTYPE == Hdarwin)
+				r->xadd -= symaddr(s) + r->off;
+
 			*val = braddoff((0xff000000U & (uint32)r->add), 
 							(0xffffff & (uint32)(r->xadd / 4)));
 			return 0;
@@ -539,6 +586,8 @@ adddynlib(char *lib)
 		if(s->size == 0)
 			addstring(s, "");
 		elfwritedynent(linklookup(ctxt, ".dynamic", 0), DT_NEEDED, addstring(s, lib));
+	} else if(HEADTYPE == Hdarwin) {
+		machoadddynlib(lib);
 	} else {
 		diag("adddynlib: unsupported binary format");
 	}
@@ -547,7 +596,7 @@ adddynlib(char *lib)
 void
 asmb(void)
 {
-	uint32 symo;
+	uint32 symo, dwarfoff, machlink;
 	Section *sect;
 	LSym *sym;
 	int i;
@@ -583,6 +632,22 @@ asmb(void)
 	cseek(segdata.fileoff);
 	datblk(segdata.vaddr, segdata.filelen);
 
+	machlink = 0;
+	if(HEADTYPE == Hdarwin) {
+		if(debug['v'])
+			Bprint(&bso, "%5.2f dwarf\n", cputime());
+
+		if(!debug['w']) { // TODO(minux): enable DWARF Support
+			dwarfoff = rnd(HEADR+segtext.len, INITRND) + rnd(segdata.filelen, INITRND);
+			cseek(dwarfoff);
+
+			segdwarf.fileoff = cpos();
+			dwarfemitdebugsections();
+			segdwarf.filelen = cpos() - segdwarf.fileoff;
+		}
+		machlink = domacholink();
+	}
+
 	/* output symbol table */
 	symsize = 0;
 	lcsize = 0;
@@ -599,6 +664,9 @@ asmb(void)
 		case Hplan9:
 			symo = segdata.fileoff+segdata.filelen;
 			break;
+		case Hdarwin:
+			symo = rnd(HEADR+segtext.filelen, INITRND)+rnd(segdata.filelen, INITRND)+machlink;
+			break;
 		ElfSym:
 			symo = segdata.fileoff+segdata.filelen;
 			symo = rnd(symo, INITRND);
@@ -635,6 +703,10 @@ asmb(void)
 				cflush();
 			}
 			break;
+		case Hdarwin:
+			if(linkmode == LinkExternal)
+				machoemitreloc();
+			break;
 		}
 	}
 
@@ -646,14 +718,14 @@ asmb(void)
 	switch(HEADTYPE) {
 	default:
 	case Hplan9:	/* plan 9 */
-		LPUT(0x647);			/* magic */
-		LPUT(segtext.filelen);			/* sizes */
-		LPUT(segdata.filelen);
-		LPUT(segdata.len - segdata.filelen);
-		LPUT(symsize);			/* nsyms */
-		LPUT(entryvalue());		/* va of entry */
-		LPUT(0L);
-		LPUT(lcsize);
+		thearch.lput(0x647);			/* magic */
+		thearch.lput(segtext.filelen);			/* sizes */
+		thearch.lput(segdata.filelen);
+		thearch.lput(segdata.len - segdata.filelen);
+		thearch.lput(symsize);			/* nsyms */
+		thearch.lput(entryvalue());		/* va of entry */
+		thearch.lput(0L);
+		thearch.lput(lcsize);
 		break;
 	case Hlinux:
 	case Hfreebsd:
@@ -662,6 +734,9 @@ asmb(void)
 	case Hnacl:
 		asmbelf(symo);
 		break;
+	case Hdarwin:
+		asmbmacho();
+		break;
 	}
 	cflush();
 	if(debug['c']){
@@ -673,18 +748,3 @@ asmb(void)
 		print("total=%lld\n", segtext.filelen+segdata.len+symsize+lcsize);
 	}
 }
-
-int32
-rnd(int32 v, int32 r)
-{
-	int32 c;
-
-	if(r <= 0)
-		return v;
-	v += r - 1;
-	c = v % r;
-	if(c < 0)
-		c += r;
-	v -= c;
-	return v;
-}
diff --git a/src/cmd/5l/l.h b/src/cmd/5l/l.h
index f9cdc5bd7e..b32b89bb06 100644
--- a/src/cmd/5l/l.h
+++ b/src/cmd/5l/l.h
@@ -32,7 +32,6 @@
 #include	<libc.h>
 #include	<bio.h>
 #include	<link.h>
-#include	"5.out.h"
 
 enum
 {
@@ -41,43 +40,14 @@ enum
 	IntSize = 4,
 	RegSize = 4,
 	MaxAlign = 8,	// max data alignment
-	FuncAlign = 4  // single-instruction alignment
+	FuncAlign = 4,  // single-instruction alignment
+	MINLC	= 4,
 };
 
 #ifndef	EXTERN
 #define	EXTERN	extern
 #endif
 
-#define	P		((Prog*)0)
-#define	S		((LSym*)0)
-
-enum
-{
-/* mark flags */
-	FOLL		= 1<<0,
-	LABEL		= 1<<1,
-	LEAF		= 1<<2,
-
-	MINLC	= 4,
-};
-
-EXTERN	int32	autosize;
-EXTERN	LSym*	datap;
-EXTERN	int	debug[128];
-EXTERN	char*	noname;
-EXTERN	Prog*	lastp;
-EXTERN	int32	lcsize;
-EXTERN	char	literal[32];
-EXTERN	int	nerrors;
-EXTERN	int32	instoffset;
-EXTERN	char*	rpath;
-EXTERN	uint32	stroffset;
-EXTERN	int32	symsize;
-EXTERN	int	armsize;
-
-#pragma	varargck	type	"I"	uint32*
-
-int	Iconv(Fmt *fp);
 void	adddynlib(char *lib);
 void	adddynrel(LSym *s, Reloc *r);
 void	adddynrela(LSym *rel, LSym *s, Reloc *r);
@@ -89,13 +59,6 @@ int	elfreloc1(Reloc *r, vlong sectoff);
 void	elfsetupplt(void);
 void	listinit(void);
 int	machoreloc1(Reloc *r, vlong sectoff);
-void	main(int argc, char *argv[]);
-int32	rnd(int32 v, int32 r);
-
-/* Native is little-endian */
-#define	LPUT(a)	lputl(a)
-#define	WPUT(a)	wputl(a)
-#define	VPUT(a)	abort()
 
 /* Used by ../ld/dwarf.c */
 enum
diff --git a/src/cmd/5l/list.c b/src/cmd/5l/list.c
index 875fc3e6b2..d4cc8acd7c 100644
--- a/src/cmd/5l/list.c
+++ b/src/cmd/5l/list.c
@@ -37,34 +37,4 @@ void
 listinit(void)
 {
 	listinit5();
-	fmtinstall('I', Iconv);
-}
-
-int
-Iconv(Fmt *fp)
-{
-	int i, n;
-	uint32 *p;
-	char *s;
-	Fmt fmt;
-	
-	n = fp->prec;
-	fp->prec = 0;
-	if(!(fp->flags&FmtPrec) || n < 0)
-		return fmtstrcpy(fp, "%I");
-	fp->flags &= ~FmtPrec;
-	p = va_arg(fp->args, uint32*);
-
-	// format into temporary buffer and
-	// call fmtstrcpy to handle padding.
-	fmtstrinit(&fmt);
-	for(i=0; i<n/4; i++) {
-		if(i > 0)
-			fmtprint(&fmt, " ");
-		fmtprint(&fmt, "%.8ux", *p++);
-	}
-	s = fmtstrflush(&fmt);
-	fmtstrcpy(fp, s);
-	free(s);
-	return 0;
 }
diff --git a/src/cmd/5l/obj.c b/src/cmd/5l/obj.c
index c6f60ee7c8..a7bd0f45ea 100644
--- a/src/cmd/5l/obj.c
+++ b/src/cmd/5l/obj.c
@@ -33,15 +33,55 @@
 #include	"l.h"
 #include	"../ld/lib.h"
 #include	"../ld/elf.h"
+#include	"../ld/macho.h"
 #include	"../ld/dwarf.h"
 #include	<ar.h>
 
-char *thestring = "arm";
-LinkArch *thelinkarch = &linkarm;
+void
+main(int argc, char **argv)
+{
+	linkarchinit();
+	ldmain(argc, argv);
+}
+
 
 void
 linkarchinit(void)
 {
+	thestring = "arm";
+	thelinkarch = &linkarm;
+
+	thearch.thechar = thechar;
+	thearch.ptrsize = thelinkarch->ptrsize;
+	thearch.intsize = thelinkarch->ptrsize;
+	thearch.regsize = thelinkarch->regsize;
+	thearch.funcalign = FuncAlign;
+	thearch.maxalign = MaxAlign;
+	thearch.minlc = MINLC;
+	thearch.dwarfregsp = DWARFREGSP;
+
+	thearch.adddynlib = adddynlib;
+	thearch.adddynrel = adddynrel;
+	thearch.adddynsym = adddynsym;
+	thearch.archinit = archinit;
+	thearch.archreloc = archreloc;
+	thearch.archrelocvariant = archrelocvariant;
+	thearch.asmb = asmb;
+	thearch.elfreloc1 = elfreloc1;
+	thearch.elfsetupplt = elfsetupplt;
+	thearch.gentext = gentext;
+	thearch.listinit = listinit;
+	thearch.machoreloc1 = machoreloc1;
+	thearch.lput = lputl;
+	thearch.wput = wputl;
+	thearch.vput = vputl;
+
+	thearch.linuxdynld = "/lib/ld-linux.so.3"; // 2 for OABI, 3 for EABI
+	thearch.freebsddynld = "/usr/libexec/ld-elf.so.1";
+	thearch.openbsddynld = "XXX";
+	thearch.netbsddynld = "/libexec/ld.elf_so";
+	thearch.dragonflydynld = "XXX";
+	thearch.solarisdynld = "XXX";
 }
 
 void
@@ -64,6 +104,7 @@ archinit(void)
 	case Hlinux:
 	case Hfreebsd:
 	case Hnacl:
+	case Hdarwin:
 		break;
 	}
 
@@ -104,6 +145,17 @@ archinit(void)
 		if(INITRND == -1)
 			INITRND = 0x10000;
 		break;
+	case Hdarwin:   /* apple MACH */
+		debug['w'] = 1; // disable DWARF generataion
+		machoinit();
+		HEADR = INITIAL_MACHO_HEADR;
+		if(INITTEXT == -1)
+			INITTEXT = 4096+HEADR;
+		if(INITDAT == -1)
+			INITDAT = 0;
+		if(INITRND == -1)
+			INITRND = 4096;
+		break;
 	}
 	if(INITDAT != 0 && INITRND != 0)
 		print("warning: -D0x%ux is ignored because of -R0x%ux\n",
diff --git a/src/cmd/6g/cgen.c b/src/cmd/6g/cgen.c
index d1645cc568..77ab2d2167 100644
--- a/src/cmd/6g/cgen.c
+++ b/src/cmd/6g/cgen.c
@@ -790,7 +790,7 @@ agenr(Node *n, Node *a, Node *res)
 void
 agen(Node *n, Node *res)
 {
-	Node *nl, *nr;
+	Node *nl;
 	Node n1, n2;
 
 	if(debug['g']) {
@@ -827,8 +827,6 @@ agen(Node *n, Node *res)
 	}
 
 	nl = n->left;
-	nr = n->right;
-	USED(nr);
 
 	switch(n->op) {
 	default:
@@ -1065,17 +1063,7 @@ bgen(Node *n, int true, int likely, Prog *to)
 
 	switch(n->op) {
 	default:
-	def:
-		regalloc(&n1, n->type, N);
-		cgen(n, &n1);
-		nodconst(&n2, n->type, 0);
-		gins(optoas(OCMP, n->type), &n1, &n2);
-		a = AJNE;
-		if(!true)
-			a = AJEQ;
-		patch(gbranch(a, n->type, likely), to);
-		regfree(&n1);
-		goto ret;
+		goto def;
 
 	case OLITERAL:
 		// need to ask if it is bool?
@@ -1095,27 +1083,20 @@ bgen(Node *n, int true, int likely, Prog *to)
 		goto ret;
 
 	case OANDAND:
-		if(!true)
-			goto caseor;
-
-	caseand:
-		p1 = gbranch(AJMP, T, 0);
-		p2 = gbranch(AJMP, T, 0);
-		patch(p1, pc);
-		bgen(n->left, !true, -likely, p2);
-		bgen(n->right, !true, -likely, p2);
-		p1 = gbranch(AJMP, T, 0);
-		patch(p1, to);
-		patch(p2, pc);
-		goto ret;
-
 	case OOROR:
-		if(!true)
-			goto caseand;
-
-	caseor:
-		bgen(n->left, true, likely, to);
-		bgen(n->right, true, likely, to);
+		if((n->op == OANDAND) == true) {
+			p1 = gbranch(AJMP, T, 0);
+			p2 = gbranch(AJMP, T, 0);
+			patch(p1, pc);
+			bgen(n->left, !true, -likely, p2);
+			bgen(n->right, !true, -likely, p2);
+			p1 = gbranch(AJMP, T, 0);
+			patch(p1, to);
+			patch(p2, pc);
+		} else {
+			bgen(n->left, true, likely, to);
+			bgen(n->right, true, likely, to);
+		}
 		goto ret;
 
 	case OEQ:
@@ -1273,6 +1254,18 @@ bgen(Node *n, int true, int likely, Prog *to)
 	}
 	goto ret;
 
+def:
+	regalloc(&n1, n->type, N);
+	cgen(n, &n1);
+	nodconst(&n2, n->type, 0);
+	gins(optoas(OCMP, n->type), &n1, &n2);
+	a = AJNE;
+	if(!true)
+		a = AJEQ;
+	patch(gbranch(a, n->type, likely), to);
+	regfree(&n1);
+	goto ret;
+
 ret:
 	;
 }
@@ -1390,22 +1383,25 @@ sgen(Node *n, Node *ns, int64 w)
 		return;
 	}
 
+	nodreg(&noddi, types[tptr], REG_DI);
+	nodreg(&nodsi, types[tptr], REG_SI);
+
 	if(n->ullman >= ns->ullman) {
-		agenr(n, &nodr, N);
+		agenr(n, &nodr, &nodsi);
 		if(ns->op == ONAME)
 			gvardef(ns);
-		agenr(ns, &nodl, N);
+		agenr(ns, &nodl, &noddi);
 	} else {
 		if(ns->op == ONAME)
 			gvardef(ns);
-		agenr(ns, &nodl, N);
-		agenr(n, &nodr, N);
+		agenr(ns, &nodl, &noddi);
+		agenr(n, &nodr, &nodsi);
 	}
 	
-	nodreg(&noddi, types[tptr], REG_DI);
-	nodreg(&nodsi, types[tptr], REG_SI);
-	gmove(&nodl, &noddi);
-	gmove(&nodr, &nodsi);
+	if(nodl.val.u.reg != REG_DI)
+		gmove(&nodl, &noddi);
+	if(nodr.val.u.reg != REG_SI)
+		gmove(&nodr, &nodsi);
 	regfree(&nodl);
 	regfree(&nodr);
 
@@ -1454,6 +1450,18 @@ sgen(Node *n, Node *ns, int64 w)
 			p->to.sym = linksym(pkglookup("duffcopy", runtimepkg));
 			// 14 and 128 = magic constants: see ../../runtime/asm_amd64.s
 			p->to.offset = 14*(128-q);
+		} else if(!nacl && c == 0) {
+			// We don't need the MOVSQ side-effect of updating SI and DI,
+			// and issuing a sequence of MOVQs directly is faster.
+			nodsi.op = OINDREG;
+			noddi.op = OINDREG;
+			while(q > 0) {
+				gmove(&nodsi, &cx); // MOVQ x+(SI),CX
+				gmove(&cx, &noddi); // MOVQ CX,x+(DI)
+				nodsi.xoffset += 8;
+				noddi.xoffset += 8;
+				q--;
+			}
 		} else
 		while(q > 0) {
 			gins(AMOVSQ, N, N);	// MOVQ *(SI)+,*(DI)+
@@ -1468,24 +1476,29 @@ sgen(Node *n, Node *ns, int64 w)
 		} else if(w < 8 || c <= 4) {
 			nodsi.op = OINDREG;
 			noddi.op = OINDREG;
+			cx.type = types[TINT32];
 			nodsi.type = types[TINT32];
 			noddi.type = types[TINT32];
 			if(c > 4) {
 				nodsi.xoffset = 0;
 				noddi.xoffset = 0;
-				gmove(&nodsi, &noddi);
+				gmove(&nodsi, &cx);
+				gmove(&cx, &noddi);
 			}
 			nodsi.xoffset = c-4;
 			noddi.xoffset = c-4;
-			gmove(&nodsi, &noddi);
+			gmove(&nodsi, &cx);
+			gmove(&cx, &noddi);
 		} else {
 			nodsi.op = OINDREG;
 			noddi.op = OINDREG;
+			cx.type = types[TINT64];
 			nodsi.type = types[TINT64];
 			noddi.type = types[TINT64];
 			nodsi.xoffset = c-8;
 			noddi.xoffset = c-8;
-			gmove(&nodsi, &noddi);
+			gmove(&nodsi, &cx);
+			gmove(&cx, &noddi);
 		}
 	}
 
@@ -1519,7 +1532,7 @@ cadable(Node *n)
 int
 componentgen(Node *nr, Node *nl)
 {
-	Node nodl, nodr;
+	Node nodl, nodr, tmp;
 	Type *t;
 	int freel, freer;
 	vlong fldcount;
@@ -1567,7 +1580,7 @@ componentgen(Node *nr, Node *nl)
 
 	nodl = *nl;
 	if(!cadable(nl)) {
-		if(nr == N || !cadable(nr))
+		if(nr != N && !cadable(nr))
 			goto no;
 		igen(nl, &nodl, N);
 		freel = 1;
@@ -1579,6 +1592,12 @@ componentgen(Node *nr, Node *nl)
 			igen(nr, &nodr, N);
 			freer = 1;
 		}
+	} else {
+		// When zeroing, prepare a register containing zero.
+		nodconst(&tmp, nl->type, 0);
+		regalloc(&nodr, types[TUINT], N);
+		gmove(&tmp, &nodr);
+		freer = 1;
 	}
 	
 	// nl and nr are 'cadable' which basically means they are names (variables) now.
@@ -1615,8 +1634,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_nel-Array_array;
@@ -1625,8 +1643,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_nel-Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_cap-Array_nel;
@@ -1635,8 +1652,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_cap-Array_nel;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		goto yes;
@@ -1650,8 +1666,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_nel-Array_array;
@@ -1660,8 +1675,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_nel-Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		goto yes;
@@ -1675,8 +1689,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_nel-Array_array;
@@ -1685,8 +1698,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_nel-Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		goto yes;
diff --git a/src/cmd/6g/galign.c b/src/cmd/6g/galign.c
index 83833e189b..fa9998b0ee 100644
--- a/src/cmd/6g/galign.c
+++ b/src/cmd/6g/galign.c
@@ -15,9 +15,9 @@ linkarchinit(void)
 {
 	if(strcmp(getgoarch(), "amd64p32") == 0) {
 		thelinkarch = &linkamd64p32;
-		arch.thelinkarch = thelinkarch;
+		thearch.thelinkarch = thelinkarch;
 		thestring = "amd64p32";
-		arch.thestring = "amd64p32";
+		thearch.thestring = "amd64p32";
 	}
 }
 
@@ -65,61 +65,46 @@ betypeinit(void)
 void
 main(int argc, char **argv)
 {
-	arch.thechar = thechar;
-	arch.thestring = thestring;
-	arch.thelinkarch = thelinkarch;
-	arch.typedefs = typedefs;
-	arch.MAXWIDTH = MAXWIDTH;
-	arch.afunclit = afunclit;
-	arch.anyregalloc = anyregalloc;
-	arch.betypeinit = betypeinit;
-	arch.bgen = bgen;
-	arch.cgen = cgen;
-	arch.cgen_asop = cgen_asop;
-	arch.cgen_call = cgen_call;
-	arch.cgen_callinter = cgen_callinter;
-	arch.cgen_ret = cgen_ret;
-	arch.clearfat = clearfat;
-	arch.clearp = clearp;
-	arch.defframe = defframe;
-	arch.dgostringptr = dgostringptr;
-	arch.dgostrlitptr = dgostrlitptr;
-	arch.dsname = dsname;
-	arch.dsymptr = dsymptr;
-	arch.dumpdata = dumpdata;
-	arch.dumpit = dumpit;
-	arch.excise = excise;
-	arch.expandchecks = expandchecks;
-	arch.fixautoused = fixautoused;
-	arch.gclean = gclean;
-	arch.gdata = gdata;
-	arch.gdatacomplex = gdatacomplex;
-	arch.gdatastring = gdatastring;
-	arch.ggloblnod = ggloblnod;
-	arch.ggloblsym = ggloblsym;
-	arch.ginit = ginit;
-	arch.gins = gins;
-	arch.ginscall = ginscall;
-	arch.gjmp = gjmp;
-	arch.gtrack = gtrack;
-	arch.gused = gused;
-	arch.igen = igen;
-	arch.isfat = isfat;
-	arch.linkarchinit = linkarchinit;
-	arch.markautoused = markautoused;
-	arch.naddr = naddr;
-	arch.newplist = newplist;
-	arch.nodarg = nodarg;
-	arch.patch = patch;
-	arch.proginfo = proginfo;
-	arch.regalloc = regalloc;
-	arch.regfree = regfree;
-	arch.regopt = regopt;
-	arch.regtyp = regtyp;
-	arch.sameaddr = sameaddr;
-	arch.smallindir = smallindir;
-	arch.stackaddr = stackaddr;
-	arch.unpatch = unpatch;
+	thearch.thechar = thechar;
+	thearch.thestring = thestring;
+	thearch.thelinkarch = thelinkarch;
+	thearch.typedefs = typedefs;
+	thearch.REGSP = REGSP;
+	thearch.REGCTXT = REGCTXT;
+	thearch.MAXWIDTH = MAXWIDTH;
+	thearch.anyregalloc = anyregalloc;
+	thearch.betypeinit = betypeinit;
+	thearch.bgen = bgen;
+	thearch.cgen = cgen;
+	thearch.cgen_call = cgen_call;
+	thearch.cgen_callinter = cgen_callinter;
+	thearch.cgen_ret = cgen_ret;
+	thearch.clearfat = clearfat;
+	thearch.defframe = defframe;
+	thearch.excise = excise;
+	thearch.expandchecks = expandchecks;
+	thearch.gclean = gclean;
+	thearch.ginit = ginit;
+	thearch.gins = gins;
+	thearch.ginscall = ginscall;
+	thearch.igen = igen;
+	thearch.linkarchinit = linkarchinit;
+	thearch.peep = peep;
+	thearch.proginfo = proginfo;
+	thearch.regalloc = regalloc;
+	thearch.regfree = regfree;
+	thearch.regtyp = regtyp;
+	thearch.sameaddr = sameaddr;
+	thearch.smallindir = smallindir;
+	thearch.stackaddr = stackaddr;
+	thearch.excludedregs = excludedregs;
+	thearch.RtoB = RtoB;
+	thearch.FtoB = FtoB;
+	thearch.BtoR = BtoR;
+	thearch.BtoF = BtoF;
+	thearch.optoas = optoas;
+	thearch.doregbits = doregbits;
+	thearch.regnames = regnames;
 	
 	gcmain(argc, argv);
 }
diff --git a/src/cmd/6g/gg.h b/src/cmd/6g/gg.h
index 7a09f673ef..a6dfad9c8e 100644
--- a/src/cmd/6g/gg.h
+++ b/src/cmd/6g/gg.h
@@ -9,10 +9,7 @@
 #include "../gc/go.h"
 #include "../6l/6.out.h"
 
-EXTERN	int32	dynloc;
 EXTERN	uchar	reg[MAXREG];
-EXTERN	int32	pcloc;		// instruction counter
-EXTERN	Strlit	emptystring;
 EXTERN	Node*	panicdiv;
 extern	vlong	unmappedzero;
 extern	int	addptr;
@@ -117,7 +114,6 @@ int anyregalloc(void);
 void betypeinit(void);
 void bgen(Node*, int, int, Prog*);
 void cgen(Node*, Node*);
-void cgen_asop(Node*);
 void cgen_call(Node*, int);
 void cgen_callinter(Node*, Node*, int);
 void cgen_ret(Node*);
@@ -163,3 +159,18 @@ int smallindir(Addr*, Addr*);
 int stackaddr(Addr*);
 Prog* unpatch(Prog*);
 
+/*
+ * reg.c
+ */
+uint64 excludedregs(void);
+uint64 RtoB(int);
+uint64 FtoB(int);
+int BtoR(uint64);
+int BtoF(uint64);
+uint64 doregbits(int);
+char** regnames(int*);
+
+/*
+ * peep.c
+ */
+void peep(Prog*);
diff --git a/src/cmd/6g/ggen.c b/src/cmd/6g/ggen.c
index 9f4b4d4ae3..72104589a3 100644
--- a/src/cmd/6g/ggen.c
+++ b/src/cmd/6g/ggen.c
@@ -7,7 +7,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
 static Prog *appendpp(Prog*, int, int, int, vlong, int, int, vlong);
 static Prog *zerorange(Prog *p, vlong frame, vlong lo, vlong hi, uint32 *ax);
@@ -115,52 +115,6 @@ appendpp(Prog *p, int as, int ftype, int freg, vlong foffset, int ttype, int tre
 	return q;	
 }
 
-// Sweep the prog list to mark any used nodes.
-void
-markautoused(Prog* p)
-{
-	for (; p; p = p->link) {
-		if (p->as == ATYPE || p->as == AVARDEF || p->as == AVARKILL)
-			continue;
-
-		if (p->from.node)
-			((Node*)(p->from.node))->used = 1;
-
-		if (p->to.node)
-			((Node*)(p->to.node))->used = 1;
-	}
-}
-
-// Fixup instructions after allocauto (formerly compactframe) has moved all autos around.
-void
-fixautoused(Prog *p)
-{
-	Prog **lp;
-
-	for (lp=&p; (p=*lp) != P; ) {
-		if (p->as == ATYPE && p->from.node && p->from.name == NAME_AUTO && !((Node*)(p->from.node))->used) {
-			*lp = p->link;
-			continue;
-		}
-		if ((p->as == AVARDEF || p->as == AVARKILL) && p->to.node && !((Node*)(p->to.node))->used) {
-			// Cannot remove VARDEF instruction, because - unlike TYPE handled above -
-			// VARDEFs are interspersed with other code, and a jump might be using the
-			// VARDEF as a target. Replace with a no-op instead. A later pass will remove
-			// the no-ops.
-			nopout(p);
-			continue;
-		}
-		if (p->from.name == NAME_AUTO && p->from.node)
-			p->from.offset += ((Node*)(p->from.node))->stkdelta;
-
-		if (p->to.name == NAME_AUTO && p->to.node)
-			p->to.offset += ((Node*)(p->to.node))->stkdelta;
-
-		lp = &p->link;
-	}
-}
-
-
 /*
  * generate:
  *	call f
@@ -478,159 +432,6 @@ cgen_ret(Node *n)
 }
 
 /*
- * generate += *= etc.
- */
-void
-cgen_asop(Node *n)
-{
-	Node n1, n2, n3, n4;
-	Node *nl, *nr;
-	Prog *p1;
-	Addr addr;
-	int a;
-
-	nl = n->left;
-	nr = n->right;
-
-	if(nr->ullman >= UINF && nl->ullman >= UINF) {
-		tempname(&n1, nr->type);
-		cgen(nr, &n1);
-		n2 = *n;
-		n2.right = &n1;
-		cgen_asop(&n2);
-		goto ret;
-	}
-
-	if(!isint[nl->type->etype])
-		goto hard;
-	if(!isint[nr->type->etype])
-		goto hard;
-
-	switch(n->etype) {
-	case OADD:
-		if(smallintconst(nr))
-		if(mpgetfix(nr->val.u.xval) == 1) {
-			a = optoas(OINC, nl->type);
-			if(nl->addable) {
-				gins(a, N, nl);
-				goto ret;
-			}
-			if(sudoaddable(a, nl, &addr)) {
-				p1 = gins(a, N, N);
-				p1->to = addr;
-				sudoclean();
-				goto ret;
-			}
-		}
-		break;
-
-	case OSUB:
-		if(smallintconst(nr))
-		if(mpgetfix(nr->val.u.xval) == 1) {
-			a = optoas(ODEC, nl->type);
-			if(nl->addable) {
-				gins(a, N, nl);
-				goto ret;
-			}
-			if(sudoaddable(a, nl, &addr)) {
-				p1 = gins(a, N, N);
-				p1->to = addr;
-				sudoclean();
-				goto ret;
-			}
-		}
-		break;
-	}
-
-	switch(n->etype) {
-	case OADD:
-	case OSUB:
-	case OXOR:
-	case OAND:
-	case OOR:
-		a = optoas(n->etype, nl->type);
-		if(nl->addable) {
-			if(smallintconst(nr)) {
-				gins(a, nr, nl);
-				goto ret;
-			}
-			regalloc(&n2, nr->type, N);
-			cgen(nr, &n2);
-			gins(a, &n2, nl);
-			regfree(&n2);
-			goto ret;
-		}
-		if(nr->ullman < UINF)
-		if(sudoaddable(a, nl, &addr)) {
-			if(smallintconst(nr)) {
-				p1 = gins(a, nr, N);
-				p1->to = addr;
-				sudoclean();
-				goto ret;
-			}
-			regalloc(&n2, nr->type, N);
-			cgen(nr, &n2);
-			p1 = gins(a, &n2, N);
-			p1->to = addr;
-			regfree(&n2);
-			sudoclean();
-			goto ret;
-		}
-	}
-
-hard:
-	n2.op = 0;
-	n1.op = 0;
-	if(nr->op == OLITERAL) {
-		// don't allocate a register for literals.
-	} else if(nr->ullman >= nl->ullman || nl->addable) {
-		regalloc(&n2, nr->type, N);
-		cgen(nr, &n2);
-		nr = &n2;
-	} else {
-		tempname(&n2, nr->type);
-		cgen(nr, &n2);
-		nr = &n2;
-	}
-	if(!nl->addable) {
-		igen(nl, &n1, N);
-		nl = &n1;
-	}
-
-	n3 = *n;
-	n3.left = nl;
-	n3.right = nr;
-	n3.op = n->etype;
-
-	regalloc(&n4, nl->type, N);
-	cgen(&n3, &n4);
-	gmove(&n4, nl);
-
-	if(n1.op)
-		regfree(&n1);
-	if(n2.op == OREGISTER)
-		regfree(&n2);
-	regfree(&n4);
-
-ret:
-	;
-}
-
-int
-samereg(Node *a, Node *b)
-{
-	if(a == N || b == N)
-		return 0;
-	if(a->op != OREGISTER)
-		return 0;
-	if(b->op != OREGISTER)
-		return 0;
-	if(a->val.u.reg != b->val.u.reg)
-		return 0;
-	return 1;
-}
-
-/*
  * generate division.
  * generates one of:
  *	res = nl / nr
diff --git a/src/cmd/6g/gobj.c b/src/cmd/6g/gobj.c
deleted file mode 100644
index 33ee4d65c4..0000000000
--- a/src/cmd/6g/gobj.c
+++ /dev/null
@@ -1,244 +0,0 @@
-// Derived from Inferno utils/6c/swt.c
-// http://code.google.com/p/inferno-os/source/browse/utils/6c/swt.c
-//
-//	Copyright © 1994-1999 Lucent Technologies Inc.  All rights reserved.
-//	Portions Copyright © 1995-1997 C H Forsyth (forsyth@terzarima.net)
-//	Portions Copyright © 1997-1999 Vita Nuova Limited
-//	Portions Copyright © 2000-2007 Vita Nuova Holdings Limited (www.vitanuova.com)
-//	Portions Copyright © 2004,2006 Bruce Ellis
-//	Portions Copyright © 2005-2007 C H Forsyth (forsyth@terzarima.net)
-//	Revisions Copyright © 2000-2007 Lucent Technologies Inc. and others
-//	Portions Copyright © 2009 The Go Authors.  All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-//
-// The above copyright notice and this permission notice shall be included in
-// all copies or substantial portions of the Software.
-//
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
-// THE SOFTWARE.
-
-#include <u.h>
-#include <libc.h>
-#include "gg.h"
-
-int
-dsname(Sym *s, int off, char *t, int n)
-{
-	Prog *p;
-
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.offset = off;
-	p->from.sym = linksym(s);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = n;
-	
-	p->to.type = TYPE_SCONST;
-	memmove(p->to.u.sval, t, n);
-	return off + n;
-}
-
-/*
- * make a refer to the data s, s+len
- * emitting DATA if needed.
- */
-void
-datastring(char *s, int len, Addr *a)
-{
-	Sym *sym;
-	
-	sym = stringsym(s, len);
-	a->type = TYPE_MEM;
-	a->name = NAME_EXTERN;
-	a->sym = linksym(sym);
-	a->node = sym->def;
-	a->offset = widthptr+widthint;  // skip header
-	a->etype = simtype[TINT];
-}
-
-/*
- * make a refer to the string sval,
- * emitting DATA if needed.
- */
-void
-datagostring(Strlit *sval, Addr *a)
-{
-	Sym *sym;
-
-	sym = stringsym(sval->s, sval->len);
-	a->type = TYPE_MEM;
-	a->name = NAME_EXTERN;
-	a->sym = linksym(sym);
-	a->node = sym->def;
-	a->offset = 0;  // header
-	a->etype = TSTRING;
-}
-
-void
-gdata(Node *nam, Node *nr, int wid)
-{
-	Prog *p;
-
-	if(nr->op == OLITERAL) {
-		switch(nr->val.ctype) {
-		case CTCPLX:
-			gdatacomplex(nam, nr->val.u.cval);
-			return;
-		case CTSTR:
-			gdatastring(nam, nr->val.u.sval);
-			return;
-		}
-	}
-	p = gins(ADATA, nam, nr);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = wid;
-}
-
-void
-gdatacomplex(Node *nam, Mpcplx *cval)
-{
-	Prog *p;
-	int w;
-
-	w = cplxsubtype(nam->type->etype);
-	w = types[w]->width;
-
-	p = gins(ADATA, nam, N);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = w;
-	p->to.type = TYPE_FCONST;
-	p->to.u.dval = mpgetflt(&cval->real);
-
-	p = gins(ADATA, nam, N);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = w;
-	p->from.offset += w;
-	p->to.type = TYPE_FCONST;
-	p->to.u.dval = mpgetflt(&cval->imag);
-}
-
-void
-gdatastring(Node *nam, Strlit *sval)
-{
-	Prog *p;
-	Node nod1;
-
-	p = gins(ADATA, nam, N);
-	datastring(sval->s, sval->len, &p->to);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = types[tptr]->width;
-	p->to.type = TYPE_ADDR;
-//print("%P\n", p);
-
-	nodconst(&nod1, types[TINT], sval->len);
-	p = gins(ADATA, nam, &nod1);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthint;
-	p->from.offset += widthptr;
-}
-
-int
-dstringptr(Sym *s, int off, char *str)
-{
-	Prog *p;
-
-	off = rnd(off, widthptr);
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-
-	datastring(str, strlen(str)+1, &p->to);
-	p->to.type = TYPE_ADDR;
-	p->to.etype = simtype[TINT];
-	off += widthptr;
-
-	return off;
-}
-
-int
-dgostrlitptr(Sym *s, int off, Strlit *lit)
-{
-	Prog *p;
-
-	if(lit == nil)
-		return duintptr(s, off, 0);
-
-	off = rnd(off, widthptr);
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-	datagostring(lit, &p->to);
-	p->to.type = TYPE_ADDR;
-	p->to.etype = simtype[TINT];
-	off += widthptr;
-
-	return off;
-}
-
-int
-dgostringptr(Sym *s, int off, char *str)
-{
-	int n;
-	Strlit *lit;
-
-	if(str == nil)
-		return duintptr(s, off, 0);
-
-	n = strlen(str);
-	lit = mal(sizeof *lit + n);
-	strcpy(lit->s, str);
-	lit->len = n;
-	return dgostrlitptr(s, off, lit);
-}
-
-int
-dsymptr(Sym *s, int off, Sym *x, int xoff)
-{
-	Prog *p;
-
-	off = rnd(off, widthptr);
-
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-	p->to.type = TYPE_ADDR;
-	p->to.name = NAME_EXTERN;
-	p->to.sym = linksym(x);
-	p->to.offset = xoff;
-	off += widthptr;
-
-	return off;
-}
-
-void
-nopout(Prog *p)
-{
-	p->as = ANOP;
-	p->from = zprog.from;
-	p->to = zprog.to;
-}
-
diff --git a/src/cmd/6g/gsubr.c b/src/cmd/6g/gsubr.c
index 697016c6f3..55d3425081 100644
--- a/src/cmd/6g/gsubr.c
+++ b/src/cmd/6g/gsubr.c
@@ -37,225 +37,6 @@
 // the text segment up higher in 6l for all GOOS.
 // At the same time, can raise StackBig in ../../runtime/stack.h.
 vlong unmappedzero = 4096;
-
-void
-clearp(Prog *p)
-{
-	p->as = AEND;
-	p->from.type = TYPE_NONE;
-	p->from.index = TYPE_NONE;
-	p->to.type = TYPE_NONE;
-	p->to.index = TYPE_NONE;
-	p->pc = pcloc;
-	pcloc++;
-}
-
-static int ddumped;
-static Prog *dfirst;
-static Prog *dpc;
-
-/*
- * generate and return proc with p->as = as,
- * linked into program. pc is next instruction.
- */
-Prog*
-prog(int as)
-{
-	Prog *p;
-
-	if(as == ADATA || as == AGLOBL) {
-		if(ddumped)
-			fatal("already dumped data");
-		if(dpc == nil) {
-			dpc = mal(sizeof(*dpc));
-			dfirst = dpc;
-		}
-		p = dpc;
-		dpc = mal(sizeof(*dpc));
-		p->link = dpc;
-	} else {
-		p = pc;
-		pc = mal(sizeof(*pc));
-		clearp(pc);
-		p->link = pc;
-	}
-
-	if(lineno == 0) {
-		if(debug['K'])
-			warn("prog: line 0");
-	}
-
-	p->as = as;
-	p->lineno = lineno;
-	return p;
-}
-
-void
-dumpdata(void)
-{
-	ddumped = 1;
-	if(dfirst == nil)
-		return;
-	newplist();
-	*pc = *dfirst;
-	pc = dpc;
-	clearp(pc);
-}
-
-/*
- * generate a branch.
- * t is ignored.
- * likely values are for branch prediction:
- *	-1 unlikely
- *	0 no opinion
- *	+1 likely
- */
-Prog*
-gbranch(int as, Type *t, int likely)
-{
-	Prog *p;
-	
-	USED(t);
-
-	p = prog(as);
-	p->to.type = TYPE_BRANCH;
-	p->to.u.branch = P;
-	if(as != AJMP && likely != 0) {
-		p->from.type = TYPE_CONST;
-		p->from.offset = likely > 0;
-	}
-	return p;
-}
-
-/*
- * patch previous branch to jump to to.
- */
-void
-patch(Prog *p, Prog *to)
-{
-	if(p->to.type != TYPE_BRANCH)
-		fatal("patch: not a branch");
-	p->to.u.branch = to;
-	p->to.offset = to->pc;
-}
-
-Prog*
-unpatch(Prog *p)
-{
-	Prog *q;
-
-	if(p->to.type != TYPE_BRANCH)
-		fatal("unpatch: not a branch");
-	q = p->to.u.branch;
-	p->to.u.branch = P;
-	p->to.offset = 0;
-	return q;
-}
-
-/*
- * start a new Prog list.
- */
-Plist*
-newplist(void)
-{
-	Plist *pl;
-
-	pl = linknewplist(ctxt);
-
-	pc = mal(sizeof(*pc));
-	clearp(pc);
-	pl->firstpc = pc;
-
-	return pl;
-}
-
-void
-gused(Node *n)
-{
-	gins(ANOP, n, N);	// used
-}
-
-Prog*
-gjmp(Prog *to)
-{
-	Prog *p;
-
-	p = gbranch(AJMP, T, 0);
-	if(to != P)
-		patch(p, to);
-	return p;
-}
-
-void
-ggloblnod(Node *nam)
-{
-	Prog *p;
-
-	p = gins(AGLOBL, nam, N);
-	p->lineno = nam->lineno;
-	p->from.sym->gotype = linksym(ngotype(nam));
-	p->to.sym = nil;
-	p->to.type = TYPE_CONST;
-	p->to.offset = nam->type->width;
-	if(nam->readonly)
-		p->from3.offset = RODATA;
-	if(nam->type != T && !haspointers(nam->type))
-		p->from3.offset |= NOPTR;
-}
-
-void
-gtrack(Sym *s)
-{
-	Prog *p;
-	
-	p = gins(AUSEFIELD, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-}
-
-void
-ggloblsym(Sym *s, int32 width, int8 flags)
-{
-	Prog *p;
-
-	p = gins(AGLOBL, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->to.type = TYPE_CONST;
-	p->to.offset = width;
-	p->from3.offset = flags;
-}
-
-int
-isfat(Type *t)
-{
-	if(t != T)
-	switch(t->etype) {
-	case TSTRUCT:
-	case TARRAY:
-	case TSTRING:
-	case TINTER:	// maybe remove later
-		return 1;
-	}
-	return 0;
-}
-
-/*
- * naddr of func generates code for address of func.
- * if using opcode that can take address implicitly,
- * call afunclit to fix up the argument.
- */
-void
-afunclit(Addr *a, Node *n)
-{
-	if(a->type == TYPE_ADDR && a->name == NAME_EXTERN) {
-		a->type = TYPE_MEM;
-		a->sym = linksym(n->sym);
-	}
-}
-
 static	int	resvd[] =
 {
 	REG_DI,	// for movstring
@@ -314,7 +95,7 @@ gclean(void)
 			yyerror("reg %R left allocated\n", i);
 }
 
-int32
+int
 anyregalloc(void)
 {
 	int i, j;
@@ -422,105 +203,6 @@ regfree(Node *n)
 }
 
 /*
- * initialize n to be register r of type t.
- */
-void
-nodreg(Node *n, Type *t, int r)
-{
-	if(t == T)
-		fatal("nodreg: t nil");
-
-	memset(n, 0, sizeof(*n));
-	n->op = OREGISTER;
-	n->addable = 1;
-	ullmancalc(n);
-	n->val.u.reg = r;
-	n->type = t;
-}
-
-/*
- * initialize n to be indirect of register r; n is type t.
- */
-void
-nodindreg(Node *n, Type *t, int r)
-{
-	nodreg(n, t, r);
-	n->op = OINDREG;
-}
-
-Node*
-nodarg(Type *t, int fp)
-{
-	Node *n;
-	NodeList *l;
-	Type *first;
-	Iter savet;
-
-	// entire argument struct, not just one arg
-	if(t->etype == TSTRUCT && t->funarg) {
-		n = nod(ONAME, N, N);
-		n->sym = lookup(".args");
-		n->type = t;
-		first = structfirst(&savet, &t);
-		if(first == nil)
-			fatal("nodarg: bad struct");
-		if(first->width == BADWIDTH)
-			fatal("nodarg: offset not computed for %T", t);
-		n->xoffset = first->width;
-		n->addable = 1;
-		goto fp;
-	}
-
-	if(t->etype != TFIELD)
-		fatal("nodarg: not field %T", t);
-	
-	if(fp == 1) {
-		for(l=curfn->dcl; l; l=l->next) {
-			n = l->n;
-			if((n->class == PPARAM || n->class == PPARAMOUT) && !isblanksym(t->sym) && n->sym == t->sym)
-				return n;
-		}
-	}
-
-	n = nod(ONAME, N, N);
-	n->type = t->type;
-	n->sym = t->sym;
-	
-	if(t->width == BADWIDTH)
-		fatal("nodarg: offset not computed for %T", t);
-	n->xoffset = t->width;
-	n->addable = 1;
-	n->orig = t->nname;
-
-fp:
-	// Rewrite argument named _ to __,
-	// or else the assignment to _ will be
-	// discarded during code generation.
-	if(isblank(n))
-		n->sym = lookup("__");
-
-	switch(fp) {
-	case 0:		// output arg
-		n->op = OINDREG;
-		n->val.u.reg = REG_SP;
-		break;
-
-	case 1:		// input arg
-		n->class = PPARAM;
-		break;
-
-	case 2:		// offset output arg
-fatal("shouldn't be used");
-		n->op = OINDREG;
-		n->val.u.reg = REG_SP;
-		n->xoffset += types[tptr]->width;
-		break;
-	}
-	n->typecheck = 1;
-	return n;
-}
-
-/*
  * generate
  *	as $c, reg
  */
@@ -577,27 +259,6 @@ ginscon(int as, vlong c, Node *n2)
 /*c2go int CASE(int, int); */
 
 /*
- * Is this node a memory operand?
- */
-int
-ismem(Node *n)
-{
-	switch(n->op) {
-	case OITAB:
-	case OSPTR:
-	case OLEN:
-	case OCAP:
-	case OINDREG:
-	case ONAME:
-	case OPARAM:
-	case OCLOSUREVAR:
-	case OADDR:
-		return 1;
-	}
-	return 0;
-}
-
-/*
  * set up nodes representing 2^63
  */
 Node bigi;
@@ -1112,213 +773,6 @@ fixlargeoffset(Node *n)
 }
 
 /*
- * generate code to compute n;
- * make a refer to result.
- */
-void
-naddr(Node *n, Addr *a, int canemitcode)
-{
-	Sym *s;
-
-	a->scale = 0;
-	a->reg = REG_NONE;
-	a->index = REG_NONE;
-	a->type = TYPE_NONE;
-	a->name = NAME_NONE;
-	a->gotype = nil;
-	a->node = N;
-	a->width = 0;
-	if(n == N)
-		return;
-
-	if(n->type != T && n->type->etype != TIDEAL) {
-		dowidth(n->type);
-		a->width = n->type->width;
-	}
-
-	switch(n->op) {
-	default:
-		fatal("naddr: bad %O %D", n->op, a);
-		break;
-
-	case OREGISTER:
-		a->type = TYPE_REG;
-		a->reg = n->val.u.reg;
-		a->sym = nil;
-		break;
-
-	case OINDREG:
-		a->type = TYPE_MEM;
-		a->reg = n->val.u.reg;
-		a->sym = linksym(n->sym);
-		a->offset = n->xoffset;
-		if(a->offset != (int32)a->offset)
-			yyerror("offset %lld too large for OINDREG", a->offset);
-		break;
-
-	case OPARAM:
-		// n->left is PHEAP ONAME for stack parameter.
-		// compute address of actual parameter on stack.
-		a->etype = simtype[n->left->type->etype];
-		a->width = n->left->type->width;
-		a->offset = n->xoffset;
-		a->sym = linksym(n->left->sym);
-		a->type = TYPE_MEM;
-		a->name = NAME_PARAM;
-		a->node = n->left->orig;
-		break;
-	
-	case OCLOSUREVAR:
-		if(!curfn->needctxt)
-			fatal("closurevar without needctxt");
-		a->type = TYPE_MEM;
-		a->reg = REG_DX;
-		a->sym = nil;
-		a->offset = n->xoffset;
-		break;
-	
-	case OCFUNC:
-		naddr(n->left, a, canemitcode);
-		a->sym = linksym(n->left->sym);
-		break;
-
-	case ONAME:
-		a->etype = 0;
-		if(n->type != T)
-			a->etype = simtype[n->type->etype];
-		a->offset = n->xoffset;
-		s = n->sym;
-		a->node = n->orig;
-		//if(a->node >= (Node*)&n)
-		//	fatal("stack node");
-		if(s == S)
-			s = lookup(".noname");
-		if(n->method) {
-			if(n->type != T)
-			if(n->type->sym != S)
-			if(n->type->sym->pkg != nil)
-				s = pkglookup(s->name, n->type->sym->pkg);
-		}
-
-		a->type = TYPE_MEM;
-		switch(n->class) {
-		default:
-			fatal("naddr: ONAME class %S %d\n", n->sym, n->class);
-		case PEXTERN:
-			a->name = NAME_EXTERN;
-			break;
-		case PAUTO:
-			a->name = NAME_AUTO;
-			break;
-		case PPARAM:
-		case PPARAMOUT:
-			a->name = NAME_PARAM;
-			break;
-		case PFUNC:
-			a->name = NAME_EXTERN;
-			a->type = TYPE_ADDR;
-			a->width = widthptr;
-			s = funcsym(s);			
-			break;
-		}
-		a->sym = linksym(s);
-		break;
-
-	case OLITERAL:
-		switch(n->val.ctype) {
-		default:
-			fatal("naddr: const %lT", n->type);
-			break;
-		case CTFLT:
-			a->type = TYPE_FCONST;
-			a->u.dval = mpgetflt(n->val.u.fval);
-			break;
-		case CTINT:
-		case CTRUNE:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = mpgetfix(n->val.u.xval);
-			break;
-		case CTSTR:
-			datagostring(n->val.u.sval, a);
-			break;
-		case CTBOOL:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = n->val.u.bval;
-			break;
-		case CTNIL:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = 0;
-			break;
-		}
-		break;
-
-	case OADDR:
-		naddr(n->left, a, canemitcode);
-		a->width = widthptr;
-		if(a->type != TYPE_MEM)
-			fatal("naddr: OADDR %D", a);
-		a->type = TYPE_ADDR;
-		break;
-	
-	case OITAB:
-		// itable of interface value
-		naddr(n->left, a, canemitcode);
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;  // itab(nil)
-		a->etype = tptr;
-		a->width = widthptr;
-		break;
-
-	case OSPTR:
-		// pointer in a string or slice
-		naddr(n->left, a, canemitcode);
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// ptr(nil)
-		a->etype = simtype[tptr];
-		a->offset += Array_array;
-		a->width = widthptr;
-		break;
-
-	case OLEN:
-		// len of string or slice
-		naddr(n->left, a, canemitcode);
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// len(nil)
-		a->etype = simtype[TUINT];
-		a->offset += Array_nel;
-		a->width = widthint;
-		break;
-
-	case OCAP:
-		// cap of string or slice
-		naddr(n->left, a, canemitcode);
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// cap(nil)
-		a->etype = simtype[TUINT];
-		a->offset += Array_cap;
-		a->width = widthint;
-		break;
-
-//	case OADD:
-//		if(n->right->op == OLITERAL) {
-//			v = n->right->vconst;
-//			naddr(n->left, a, canemitcode);
-//		} else
-//		if(n->left->op == OLITERAL) {
-//			v = n->left->vconst;
-//			naddr(n->right, a, canemitcode);
-//		} else
-//			goto bad;
-//		a->offset += v;
-//		break;
-
-	}
-}
-
-/*
  * return Axxx for Oxxx on type t.
  */
 int
@@ -1966,6 +1420,8 @@ sudoaddable(int as, Node *n, Addr *a)
 	if(n->type == T)
 		return 0;
 
+	memset(a, 0, sizeof *a);
+
 	switch(n->op) {
 	case OLITERAL:
 		if(!isconst(n, CTINT))
diff --git a/src/cmd/6g/peep.c b/src/cmd/6g/peep.c
index e05a06087f..261cb6e0a1 100644
--- a/src/cmd/6g/peep.c
+++ b/src/cmd/6g/peep.c
@@ -31,7 +31,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
 static void	conprop(Flow *r);
 static void	elimshortmov(Graph *g);
@@ -44,9 +44,15 @@ static int	copy1(Adr*, Adr*, Flow*, int);
 static int	copyas(Adr*, Adr*);
 static int	copyau(Adr*, Adr*);
 static int	copysub(Adr*, Adr*, Adr*, int);
+static int	copyu(Prog*, Adr*, Adr*);
 
 static uint32	gactive;
 
+enum
+{
+	exregoffset = REG_R15,
+};
+
 // do we need the carry bit
 static int
 needc(Prog *p)
@@ -91,7 +97,7 @@ peep(Prog *firstp)
 	Prog *p, *p1;
 	int t;
 
-	g = flowstart(firstp, sizeof(Flow));
+	g = flowstart(firstp, 0);
 	if(g == nil)
 		return;
 	gactive = 0;
@@ -737,7 +743,7 @@ copy1(Adr *v1, Adr *v2, Flow *r, int f)
  * 4 if set and used
  * 0 otherwise (not touched)
  */
-int
+static int
 copyu(Prog *p, Adr *v, Adr *s)
 {
 	ProgInfo info;
@@ -761,7 +767,7 @@ copyu(Prog *p, Adr *v, Adr *s)
 	case ACALL:
 		if(REGEXT && v->type == TYPE_REG && v->reg <= REGEXT && v->reg > exregoffset)
 			return 2;
-		if(REGARG >= 0 && v->type == TYPE_REG && v->reg == (uchar)REGARG)
+		if(REGARG >= 0 && v->type == TYPE_REG && v->reg == REGARG)
 			return 2;
 		if(v->type == p->from.type && v->reg == p->from.reg)
 			return 2;
@@ -776,7 +782,7 @@ copyu(Prog *p, Adr *v, Adr *s)
 		return 3;
 
 	case ATEXT:
-		if(REGARG >= 0 && v->type == TYPE_REG && v->reg == (uchar)REGARG)
+		if(REGARG >= 0 && v->type == TYPE_REG && v->reg == REGARG)
 			return 3;
 		return 0;
 	}
diff --git a/src/cmd/6g/prog.c b/src/cmd/6g/prog.c
index 32d5256f8c..79b7911e5b 100644
--- a/src/cmd/6g/prog.c
+++ b/src/cmd/6g/prog.c
@@ -5,7 +5,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
 // Matches real RtoB but can be used in global initializer.
 #define RtoB(r) (1<<((r)-REG_AX))
diff --git a/src/cmd/6g/reg.c b/src/cmd/6g/reg.c
index 7db44245f1..e01f265a13 100644
--- a/src/cmd/6g/reg.c
+++ b/src/cmd/6g/reg.c
@@ -31,56 +31,11 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
-#define	NREGVAR	32	/* 16 general + 16 floating */
-#define	REGBITS	((uint64)0xffffffffull)
-/*c2go enum {
+enum {
 	NREGVAR = 32,
-	REGBITS = 0xffffffff,
 };
-*/
-
-static	Reg*	firstr;
-static	int	first	= 1;
-
-int
-rcmp(const void *a1, const void *a2)
-{
-	Rgn *p1, *p2;
-	int c1, c2;
-
-	p1 = (Rgn*)a1;
-	p2 = (Rgn*)a2;
-	c1 = p2->cost;
-	c2 = p1->cost;
-	if(c1 -= c2)
-		return c1;
-	return p2->varno - p1->varno;
-}
-
-static void
-setaddrs(Bits bit)
-{
-	int i, n;
-	Var *v;
-	Node *node;
-
-	while(bany(&bit)) {
-		// convert each bit to a variable
-		i = bnum(bit);
-		node = var[i].node;
-		n = var[i].name;
-		biclr(&bit, i);
-
-		// disable all pieces of that variable
-		for(i=0; i<nvar; i++) {
-			v = var+i;
-			if(v->node == node && v->name == n)
-				v->addr = 2;
-		}
-	}
-}
 
 static char* regname[] = {
 	".AX",
@@ -117,471 +72,23 @@ static char* regname[] = {
 	".X15",
 };
 
-static Node* regnodes[NREGVAR];
-
-static void walkvardef(Node *n, Reg *r, int active);
-
-void
-regopt(Prog *firstp)
-{
-	Reg *r, *r1;
-	Prog *p;
-	Graph *g;
-	ProgInfo info;
-	int i, z, active;
-	uint32 vreg;
-	Bits bit;
-
-	if(first) {
-		fmtinstall('Q', Qconv);
-		exregoffset = REG_R15;
-		first = 0;
-	}
-
-	mergetemp(firstp);
-
-	/*
-	 * control flow is more complicated in generated go code
-	 * than in generated c code.  define pseudo-variables for
-	 * registers, so we have complete register usage information.
-	 */
-	nvar = NREGVAR;
-	memset(var, 0, NREGVAR*sizeof var[0]);
-	for(i=0; i<NREGVAR; i++) {
-		if(regnodes[i] == N)
-			regnodes[i] = newname(lookup(regname[i]));
-		var[i].node = regnodes[i];
-	}
-
-	regbits = RtoB(REG_SP);
-	for(z=0; z<BITS; z++) {
-		externs.b[z] = 0;
-		params.b[z] = 0;
-		consts.b[z] = 0;
-		addrs.b[z] = 0;
-		ivar.b[z] = 0;
-		ovar.b[z] = 0;
-	}
-
-	/*
-	 * pass 1
-	 * build aux data structure
-	 * allocate pcs
-	 * find use and set of variables
-	 */
-	g = flowstart(firstp, sizeof(Reg));
-	if(g == nil) {
-		for(i=0; i<nvar; i++)
-			var[i].node->opt = nil;
-		return;
-	}
-
-	firstr = (Reg*)g->start;
-
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		p = r->f.prog;
-		if(p->as == AVARDEF || p->as == AVARKILL)
-			continue;
-		proginfo(&info, p);
-
-		// Avoid making variables for direct-called functions.
-		if(p->as == ACALL && p->to.type == TYPE_MEM && p->to.name == NAME_EXTERN)
-			continue;
-
-		r->use1.b[0] |= info.reguse | info.regindex;
-		r->set.b[0] |= info.regset;
-
-		bit = mkvar(r, &p->from);
-		if(bany(&bit)) {
-			if(info.flags & LeftAddr)
-				setaddrs(bit);
-			if(info.flags & LeftRead)
-				for(z=0; z<BITS; z++)
-					r->use1.b[z] |= bit.b[z];
-			if(info.flags & LeftWrite)
-				for(z=0; z<BITS; z++)
-					r->set.b[z] |= bit.b[z];
-		}
-
-		bit = mkvar(r, &p->to);
-		if(bany(&bit)) {	
-			if(info.flags & RightAddr)
-				setaddrs(bit);
-			if(info.flags & RightRead)
-				for(z=0; z<BITS; z++)
-					r->use2.b[z] |= bit.b[z];
-			if(info.flags & RightWrite)
-				for(z=0; z<BITS; z++)
-					r->set.b[z] |= bit.b[z];
-		}
-	}
-
-	for(i=0; i<nvar; i++) {
-		Var *v = var+i;
-		if(v->addr) {
-			bit = blsh(i);
-			for(z=0; z<BITS; z++)
-				addrs.b[z] |= bit.b[z];
-		}
-
-		if(debug['R'] && debug['v'])
-			print("bit=%2d addr=%d et=%-6E w=%-2d s=%N + %lld\n",
-				i, v->addr, v->etype, v->width, v->node, v->offset);
-	}
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass1", &firstr->f, 1);
-
-	/*
-	 * pass 2
-	 * find looping structure
-	 */
-	flowrpo(g);
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass2", &firstr->f, 1);
-
-	/*
-	 * pass 2.5
-	 * iterate propagating fat vardef covering forward
-	 * r->act records vars with a VARDEF since the last CALL.
-	 * (r->act will be reused in pass 5 for something else,
-	 * but we'll be done with it by then.)
-	 */
-	active = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		r->f.active = 0;
-		r->act = zbits;
-	}
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		p = r->f.prog;
-		if(p->as == AVARDEF && isfat(((Node*)(p->to.node))->type) && ((Node*)(p->to.node))->opt != nil) {
-			active++;
-			walkvardef(p->to.node, r, active);
-		}
-	}
-
-	/*
-	 * pass 3
-	 * iterate propagating usage
-	 * 	back until flow graph is complete
-	 */
-loop1:
-	change = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->f.active = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		if(r->f.prog->as == ARET)
-			prop(r, zbits, zbits);
-loop11:
-	/* pick up unreachable code */
-	i = 0;
-	for(r = firstr; r != R; r = r1) {
-		r1 = (Reg*)r->f.link;
-		if(r1 && r1->f.active && !r->f.active) {
-			prop(r, zbits, zbits);
-			i = 1;
-		}
-	}
-	if(i)
-		goto loop11;
-	if(change)
-		goto loop1;
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass3", &firstr->f, 1);
-
-	/*
-	 * pass 4
-	 * iterate propagating register/variable synchrony
-	 * 	forward until graph is complete
-	 */
-loop2:
-	change = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->f.active = 0;
-	synch(firstr, zbits);
-	if(change)
-		goto loop2;
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass4", &firstr->f, 1);
-
-	/*
-	 * pass 4.5
-	 * move register pseudo-variables into regu.
-	 */
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		r->regu = (r->refbehind.b[0] | r->set.b[0]) & REGBITS;
-
-		r->set.b[0] &= ~REGBITS;
-		r->use1.b[0] &= ~REGBITS;
-		r->use2.b[0] &= ~REGBITS;
-		r->refbehind.b[0] &= ~REGBITS;
-		r->refahead.b[0] &= ~REGBITS;
-		r->calbehind.b[0] &= ~REGBITS;
-		r->calahead.b[0] &= ~REGBITS;
-		r->regdiff.b[0] &= ~REGBITS;
-		r->act.b[0] &= ~REGBITS;
-	}
-
-	/*
-	 * pass 5
-	 * isolate regions
-	 * calculate costs (paint1)
-	 */
-	r = firstr;
-	if(r) {
-		for(z=0; z<BITS; z++)
-			bit.b[z] = (r->refahead.b[z] | r->calahead.b[z]) &
-			  ~(externs.b[z] | params.b[z] | addrs.b[z] | consts.b[z]);
-		if(bany(&bit) && !r->f.refset) {
-			// should never happen - all variables are preset
-			if(debug['w'])
-				print("%L: used and not set: %Q\n", r->f.prog->lineno, bit);
-			r->f.refset = 1;
-		}
-	}
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->act = zbits;
-	rgp = region;
-	nregion = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		for(z=0; z<BITS; z++)
-			bit.b[z] = r->set.b[z] &
-			  ~(r->refahead.b[z] | r->calahead.b[z] | addrs.b[z]);
-		if(bany(&bit) && !r->f.refset) {
-			if(debug['w'])
-				print("%L: set and not used: %Q\n", r->f.prog->lineno, bit);
-			r->f.refset = 1;
-			excise(&r->f);
-		}
-		for(z=0; z<BITS; z++)
-			bit.b[z] = LOAD(r) & ~(r->act.b[z] | addrs.b[z]);
-		while(bany(&bit)) {
-			i = bnum(bit);
-			rgp->enter = r;
-			rgp->varno = i;
-			change = 0;
-			paint1(r, i);
-			biclr(&bit, i);
-			if(change <= 0)
-				continue;
-			rgp->cost = change;
-			nregion++;
-			if(nregion >= NRGN) {
-				if(debug['R'] && debug['v'])
-					print("too many regions\n");
-				goto brk;
-			}
-			rgp++;
-		}
-	}
-brk:
-	qsort(region, nregion, sizeof(region[0]), rcmp);
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass5", &firstr->f, 1);
-
-	/*
-	 * pass 6
-	 * determine used registers (paint2)
-	 * replace code (paint3)
-	 */
-	rgp = region;
-	if(debug['R'] && debug['v'])
-		print("\nregisterizing\n");
-	for(i=0; i<nregion; i++) {
-		if(debug['R'] && debug['v'])
-			print("region %d: cost %d varno %d enter %lld\n", i, rgp->cost, rgp->varno, rgp->enter->f.prog->pc);
-		bit = blsh(rgp->varno);
-		vreg = paint2(rgp->enter, rgp->varno, 0);
-		vreg = allreg(vreg, rgp);
-		if(rgp->regno != 0) {
-			if(debug['R'] && debug['v']) {
-				Var *v;
-
-				v = var + rgp->varno;
-				print("registerize %N+%lld (bit=%2d et=%2E) in %R\n",
-						v->node, v->offset, rgp->varno, v->etype, rgp->regno);
-			}
-			paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
-		}
-		rgp++;
-	}
-
-	/*
-	 * free aux structures. peep allocates new ones.
-	 */
-	for(i=0; i<nvar; i++)
-		var[i].node->opt = nil;
-	flowend(g);
-	firstr = R;
-
-	if(debug['R'] && debug['v']) {
-		// Rebuild flow graph, since we inserted instructions
-		g = flowstart(firstp, sizeof(Reg));
-		firstr = (Reg*)g->start;
-		dumpit("pass6", &firstr->f, 1);
-		flowend(g);
-		firstr = R;
-	}
-
-	/*
-	 * pass 7
-	 * peep-hole on basic block
-	 */
-	if(!debug['R'] || debug['P'])
-		peep(firstp);
-
-	/*
-	 * eliminate nops
-	 */
-	for(p=firstp; p!=P; p=p->link) {
-		while(p->link != P && p->link->as == ANOP)
-			p->link = p->link->link;
-		if(p->to.type == TYPE_BRANCH)
-			while(p->to.u.branch != P && p->to.u.branch->as == ANOP)
-				p->to.u.branch = p->to.u.branch->link;
-	}
-
-	if(debug['R']) {
-		if(ostats.ncvtreg ||
-		   ostats.nspill ||
-		   ostats.nreload ||
-		   ostats.ndelmov ||
-		   ostats.nvar ||
-		   ostats.naddr ||
-		   0)
-			print("\nstats\n");
-
-		if(ostats.ncvtreg)
-			print("	%4d cvtreg\n", ostats.ncvtreg);
-		if(ostats.nspill)
-			print("	%4d spill\n", ostats.nspill);
-		if(ostats.nreload)
-			print("	%4d reload\n", ostats.nreload);
-		if(ostats.ndelmov)
-			print("	%4d delmov\n", ostats.ndelmov);
-		if(ostats.nvar)
-			print("	%4d var\n", ostats.nvar);
-		if(ostats.naddr)
-			print("	%4d addr\n", ostats.naddr);
-
-		memset(&ostats, 0, sizeof(ostats));
-	}
-}
-
-static void
-walkvardef(Node *n, Reg *r, int active)
+char**
+regnames(int *n)
 {
-	Reg *r1, *r2;
-	int bn;
-	Var *v;
-	
-	for(r1=r; r1!=R; r1=(Reg*)r1->f.s1) {
-		if(r1->f.active == active)
-			break;
-		r1->f.active = active;
-		if(r1->f.prog->as == AVARKILL && r1->f.prog->to.node == n)
-			break;
-		for(v=n->opt; v!=nil; v=v->nextinnode) {
-			bn = v - var;
-			biset(&r1->act, bn);
-		}
-		if(r1->f.prog->as == ACALL)
-			break;
-	}
-
-	for(r2=r; r2!=r1; r2=(Reg*)r2->f.s1)
-		if(r2->f.s2 != nil)
-			walkvardef(n, (Reg*)r2->f.s2, active);
+	*n = NREGVAR;
+	return regname;
 }
 
-/*
- * add mov b,rn
- * just after r
- */
-void
-addmove(Reg *r, int bn, int rn, int f)
+uint64
+excludedregs(void)
 {
-	Prog *p, *p1;
-	Adr *a;
-	Var *v;
-
-	p1 = mal(sizeof(*p1));
-	clearp(p1);
-	p1->pc = 9999;
-
-	p = r->f.prog;
-	p1->link = p->link;
-	p->link = p1;
-	p1->lineno = p->lineno;
-
-	v = var + bn;
-
-	a = &p1->to;
-	a->offset = v->offset;
-	a->etype = v->etype;
-	a->type = TYPE_MEM;
-	a->name = v->name;
-	a->node = v->node;
-	a->sym = linksym(v->node->sym);
-
-	// need to clean this up with wptr and
-	// some of the defaults
-	p1->as = AMOVL;
-	switch(simtype[(uchar)v->etype]) {
-	default:
-		fatal("unknown type %E", v->etype);
-	case TINT8:
-	case TUINT8:
-	case TBOOL:
-		p1->as = AMOVB;
-		break;
-	case TINT16:
-	case TUINT16:
-		p1->as = AMOVW;
-		break;
-	case TINT64:
-	case TUINT64:
-	case TPTR64:
-		p1->as = AMOVQ;
-		break;
-	case TFLOAT32:
-		p1->as = AMOVSS;
-		break;
-	case TFLOAT64:
-		p1->as = AMOVSD;
-		break;
-	case TINT32:
-	case TUINT32:
-	case TPTR32:
-		break;
-	}
-
-	p1->from.type = TYPE_REG;
-	p1->from.reg = rn;
-	p1->from.name = NAME_NONE;
-	if(!f) {
-		p1->from = *a;
-		*a = zprog.from;
-		a->type = TYPE_REG;
-		a->reg = rn;
-		if(v->etype == TUINT8)
-			p1->as = AMOVB;
-		if(v->etype == TUINT16)
-			p1->as = AMOVW;
-	}
-	if(debug['R'] && debug['v'])
-		print("%P ===add=== %P\n", p, p1);
-	ostats.nspill++;
+	return RtoB(REG_SP);
 }
 
-uint32
+uint64
 doregbits(int r)
 {
-	uint32 b;
+	uint64 b;
 
 	b = 0;
 	if(r >= REG_AX && r <= REG_R15)
@@ -598,592 +105,19 @@ doregbits(int r)
 	return b;
 }
 
-static int
-overlap(int64 o1, int w1, int64 o2, int w2)
-{
-	int64 t1, t2;
-
-	t1 = o1+w1;
-	t2 = o2+w2;
-
-	if(!(t1 > o2 && t2 > o1))
-		return 0;
-
-	return 1;
-}
-
-Bits
-mkvar(Reg *r, Adr *a)
-{
-	Var *v;
-	int i, n, et, z, flag;
-	int64 w;
-	uint32 regu;
-	int64 o;
-	Bits bit;
-	Node *node;
-
-	/*
-	 * mark registers used
-	 */
-	if(a->type == TYPE_NONE)
-		goto none;
-
-	if(r != R)
-		r->use1.b[0] |= doregbits(a->index);
-
-	switch(a->type) {
-	default:
-		regu = doregbits(a->reg);
-		if(regu == 0)
-			goto none;
-		bit = zbits;
-		bit.b[0] = regu;
-		return bit;
-
-	case TYPE_ADDR:
-		a->type = TYPE_MEM;
-		bit = mkvar(r, a);
-		setaddrs(bit);
-		a->type = TYPE_ADDR;
-		ostats.naddr++;
-		goto none;
-
-	case TYPE_MEM:
-		switch(a->name) {
-		default:
-			goto none;
-		case NAME_EXTERN:
-		case NAME_STATIC:
-		case NAME_PARAM:
-		case NAME_AUTO:
-			n = a->name;
-			break;
-		}
-	}
-
-	node = a->node;
-	if(node == N || node->op != ONAME || node->orig == N)
-		goto none;
-	node = node->orig;
-	if(node->orig != node)
-		fatal("%D: bad node", a);
-	if(node->sym == S || node->sym->name[0] == '.')
-		goto none;
-	et = a->etype;
-	o = a->offset;
-	w = a->width;
-	if(w < 0)
-		fatal("bad width %lld for %D", w, a);
-
-	flag = 0;
-	for(i=0; i<nvar; i++) {
-		v = var+i;
-		if(v->node == node && v->name == n) {
-			if(v->offset == o)
-			if(v->etype == et)
-			if(v->width == w)
-				return blsh(i);
-
-			// if they overlaps, disable both
-			if(overlap(v->offset, v->width, o, w)) {
-//				print("disable overlap %s %d %d %d %d, %E != %E\n", s->name, v->offset, v->width, o, w, v->etype, et);
-				v->addr = 1;
-				flag = 1;
-			}
-		}
-	}
-	switch(et) {
-	case 0:
-	case TFUNC:
-		goto none;
-	}
-
-	if(nvar >= NVAR) {
-		if(debug['w'] > 1 && node != N)
-			fatal("variable not optimized: %#N", node);
-		
-		// If we're not tracking a word in a variable, mark the rest as
-		// having its address taken, so that we keep the whole thing
-		// live at all calls. otherwise we might optimize away part of
-		// a variable but not all of it.
-		for(i=0; i<nvar; i++) {
-			v = var+i;
-			if(v->node == node)
-				v->addr = 1;
-		}
-		goto none;
-	}
-
-	i = nvar;
-	nvar++;
-	v = var+i;
-	v->offset = o;
-	v->name = n;
-	v->etype = et;
-	v->width = w;
-	v->addr = flag;		// funny punning
-	v->node = node;
-	
-	// node->opt is the head of a linked list
-	// of Vars within the given Node, so that
-	// we can start at a Var and find all the other
-	// Vars in the same Go variable.
-	v->nextinnode = node->opt;
-	node->opt = v;
-
-	bit = blsh(i);
-	if(n == NAME_EXTERN || n == NAME_STATIC)
-		for(z=0; z<BITS; z++)
-			externs.b[z] |= bit.b[z];
-	if(n == NAME_PARAM)
-		for(z=0; z<BITS; z++)
-			params.b[z] |= bit.b[z];
-
-	if(node->class == PPARAM)
-		for(z=0; z<BITS; z++)
-			ivar.b[z] |= bit.b[z];
-	if(node->class == PPARAMOUT)
-		for(z=0; z<BITS; z++)
-			ovar.b[z] |= bit.b[z];
-
-	// Treat values with their address taken as live at calls,
-	// because the garbage collector's liveness analysis in ../gc/plive.c does.
-	// These must be consistent or else we will elide stores and the garbage
-	// collector will see uninitialized data.
-	// The typical case where our own analysis is out of sync is when the
-	// node appears to have its address taken but that code doesn't actually
-	// get generated and therefore doesn't show up as an address being
-	// taken when we analyze the instruction stream.
-	// One instance of this case is when a closure uses the same name as
-	// an outer variable for one of its own variables declared with :=.
-	// The parser flags the outer variable as possibly shared, and therefore
-	// sets addrtaken, even though it ends up not being actually shared.
-	// If we were better about _ elision, _ = &x would suffice too.
-	// The broader := in a closure problem is mentioned in a comment in
-	// closure.c:/^typecheckclosure and dcl.c:/^oldname.
-	if(node->addrtaken)
-		v->addr = 1;
-
-	// Disable registerization for globals, because:
-	// (1) we might panic at any time and we want the recovery code
-	// to see the latest values (issue 1304).
-	// (2) we don't know what pointers might point at them and we want
-	// loads via those pointers to see updated values and vice versa (issue 7995).
-	//
-	// Disable registerization for results if using defer, because the deferred func
-	// might recover and return, causing the current values to be used.
-	if(node->class == PEXTERN || (hasdefer && node->class == PPARAMOUT))
-		v->addr = 1;
-
-	if(debug['R'])
-		print("bit=%2d et=%2E w=%lld+%lld %#N %D flag=%d\n", i, et, o, w, node, a, v->addr);
-	ostats.nvar++;
-
-	return bit;
-
-none:
-	return zbits;
-}
-
-void
-prop(Reg *r, Bits ref, Bits cal)
-{
-	Reg *r1, *r2;
-	int z, i, j;
-	Var *v, *v1;
-
-	for(r1 = r; r1 != R; r1 = (Reg*)r1->f.p1) {
-		for(z=0; z<BITS; z++) {
-			ref.b[z] |= r1->refahead.b[z];
-			if(ref.b[z] != r1->refahead.b[z]) {
-				r1->refahead.b[z] = ref.b[z];
-				change++;
-			}
-			cal.b[z] |= r1->calahead.b[z];
-			if(cal.b[z] != r1->calahead.b[z]) {
-				r1->calahead.b[z] = cal.b[z];
-				change++;
-			}
-		}
-		switch(r1->f.prog->as) {
-		case ACALL:
-			if(noreturn(r1->f.prog))
-				break;
-
-			// Mark all input variables (ivar) as used, because that's what the
-			// liveness bitmaps say. The liveness bitmaps say that so that a
-			// panic will not show stale values in the parameter dump.
-			// Mark variables with a recent VARDEF (r1->act) as used,
-			// so that the optimizer flushes initializations to memory,
-			// so that if a garbage collection happens during this CALL,
-			// the collector will see initialized memory. Again this is to
-			// match what the liveness bitmaps say.
-			for(z=0; z<BITS; z++) {
-				cal.b[z] |= ref.b[z] | externs.b[z] | ivar.b[z] | r1->act.b[z];
-				ref.b[z] = 0;
-			}
-			
-			// cal.b is the current approximation of what's live across the call.
-			// Every bit in cal.b is a single stack word. For each such word,
-			// find all the other tracked stack words in the same Go variable
-			// (struct/slice/string/interface) and mark them live too.
-			// This is necessary because the liveness analysis for the garbage
-			// collector works at variable granularity, not at word granularity.
-			// It is fundamental for slice/string/interface: the garbage collector
-			// needs the whole value, not just some of the words, in order to
-			// interpret the other bits correctly. Specifically, slice needs a consistent
-			// ptr and cap, string needs a consistent ptr and len, and interface
-			// needs a consistent type word and data word.
-			for(z=0; z<BITS; z++) {
-				if(cal.b[z] == 0)
-					continue;
-				for(i=0; i<64; i++) {
-					if(z*64+i >= nvar || ((cal.b[z]>>i)&1) == 0)
-						continue;
-					v = var+z*64+i;
-					if(v->node->opt == nil) // v represents fixed register, not Go variable
-						continue;
-
-					// v->node->opt is the head of a linked list of Vars
-					// corresponding to tracked words from the Go variable v->node.
-					// Walk the list and set all the bits.
-					// For a large struct this could end up being quadratic:
-					// after the first setting, the outer loop (for z, i) would see a 1 bit
-					// for all of the remaining words in the struct, and for each such
-					// word would go through and turn on all the bits again.
-					// To avoid the quadratic behavior, we only turn on the bits if
-					// v is the head of the list or if the head's bit is not yet turned on.
-					// This will set the bits at most twice, keeping the overall loop linear.
-					v1 = v->node->opt;
-					j = v1 - var;
-					if(v == v1 || !btest(&cal, j)) {
-						for(; v1 != nil; v1 = v1->nextinnode) {
-							j = v1 - var;
-							biset(&cal, j);
-						}
-					}
-				}
-			}
-			break;
-
-		case ATEXT:
-			for(z=0; z<BITS; z++) {
-				cal.b[z] = 0;
-				ref.b[z] = 0;
-			}
-			break;
-
-		case ARET:
-			for(z=0; z<BITS; z++) {
-				cal.b[z] = externs.b[z] | ovar.b[z];
-				ref.b[z] = 0;
-			}
-			break;
-		}
-		for(z=0; z<BITS; z++) {
-			ref.b[z] = (ref.b[z] & ~r1->set.b[z]) |
-				r1->use1.b[z] | r1->use2.b[z];
-			cal.b[z] &= ~(r1->set.b[z] | r1->use1.b[z] | r1->use2.b[z]);
-			r1->refbehind.b[z] = ref.b[z];
-			r1->calbehind.b[z] = cal.b[z];
-		}
-		if(r1->f.active)
-			break;
-		r1->f.active = 1;
-	}
-	for(; r != r1; r = (Reg*)r->f.p1)
-		for(r2 = (Reg*)r->f.p2; r2 != R; r2 = (Reg*)r2->f.p2link)
-			prop(r2, r->refbehind, r->calbehind);
-}
-
-void
-synch(Reg *r, Bits dif)
-{
-	Reg *r1;
-	int z;
-
-	for(r1 = r; r1 != R; r1 = (Reg*)r1->f.s1) {
-		for(z=0; z<BITS; z++) {
-			dif.b[z] = (dif.b[z] &
-				~(~r1->refbehind.b[z] & r1->refahead.b[z])) |
-					r1->set.b[z] | r1->regdiff.b[z];
-			if(dif.b[z] != r1->regdiff.b[z]) {
-				r1->regdiff.b[z] = dif.b[z];
-				change++;
-			}
-		}
-		if(r1->f.active)
-			break;
-		r1->f.active = 1;
-		for(z=0; z<BITS; z++)
-			dif.b[z] &= ~(~r1->calbehind.b[z] & r1->calahead.b[z]);
-		if(r1->f.s2 != nil)
-			synch((Reg*)r1->f.s2, dif);
-	}
-}
-
-uint32
-allreg(uint32 b, Rgn *r)
-{
-	Var *v;
-	int i;
-
-	v = var + r->varno;
-	r->regno = 0;
-	switch(v->etype) {
-
-	default:
-		fatal("unknown etype %d/%E", bitno(b), v->etype);
-		break;
-
-	case TINT8:
-	case TUINT8:
-	case TINT16:
-	case TUINT16:
-	case TINT32:
-	case TUINT32:
-	case TINT64:
-	case TUINT64:
-	case TINT:
-	case TUINT:
-	case TUINTPTR:
-	case TBOOL:
-	case TPTR32:
-	case TPTR64:
-		i = BtoR(~b);
-		if(i && r->cost > 0) {
-			r->regno = i;
-			return RtoB(i);
-		}
-		break;
-
-	case TFLOAT32:
-	case TFLOAT64:
-		i = BtoF(~b);
-		if(i && r->cost > 0) {
-			r->regno = i;
-			return FtoB(i);
-		}
-		break;
-	}
-	return 0;
-}
-
-void
-paint1(Reg *r, int bn)
-{
-	Reg *r1;
-	int z;
-	uint64 bb;
-
-	z = bn/64;
-	bb = 1LL<<(bn%64);
-	if(r->act.b[z] & bb)
-		return;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(r1->act.b[z] & bb)
-			break;
-		r = r1;
-	}
-
-	if(LOAD(r) & ~(r->set.b[z]&~(r->use1.b[z]|r->use2.b[z])) & bb) {
-		change -= CLOAD * r->f.loop;
-	}
-	for(;;) {
-		r->act.b[z] |= bb;
-
-		if(r->f.prog->as != ANOP) { // don't give credit for NOPs
-			if(r->use1.b[z] & bb)
-				change += CREF * r->f.loop;
-			if((r->use2.b[z]|r->set.b[z]) & bb)
-				change += CREF * r->f.loop;
-		}
-
-		if(STORE(r) & r->regdiff.b[z] & bb) {
-			change -= CLOAD * r->f.loop;
-		}
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					paint1(r1, bn);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				paint1(r1, bn);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(r->act.b[z] & bb)
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-}
-
-uint32
-paint2(Reg *r, int bn, int depth)
-{
-	Reg *r1;
-	int z;
-	uint64 bb, vreg;
-
-	z = bn/64;
-	bb = 1LL << (bn%64);
-	vreg = regbits;
-	if(!(r->act.b[z] & bb))
-		return vreg;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(!(r1->act.b[z] & bb))
-			break;
-		r = r1;
-	}
-	for(;;) {
-		if(debug['R'] && debug['v'])
-			print("  paint2 %d %P\n", depth, r->f.prog);
-
-		r->act.b[z] &= ~bb;
-
-		vreg |= r->regu;
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					vreg |= paint2(r1, bn, depth+1);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				vreg |= paint2(r1, bn, depth+1);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(!(r->act.b[z] & bb))
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-
-	return vreg;
-}
-
-void
-paint3(Reg *r, int bn, uint32 rb, int rn)
-{
-	Reg *r1;
-	Prog *p;
-	int z;
-	uint64 bb;
-
-	z = bn/64;
-	bb = 1LL << (bn%64);
-	if(r->act.b[z] & bb)
-		return;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(r1->act.b[z] & bb)
-			break;
-		r = r1;
-	}
-
-	if(LOAD(r) & ~(r->set.b[z] & ~(r->use1.b[z]|r->use2.b[z])) & bb)
-		addmove(r, bn, rn, 0);
-	for(;;) {
-		r->act.b[z] |= bb;
-		p = r->f.prog;
-
-		if(r->use1.b[z] & bb) {
-			if(debug['R'] && debug['v'])
-				print("%P", p);
-			addreg(&p->from, rn);
-			if(debug['R'] && debug['v'])
-				print(" ===change== %P\n", p);
-		}
-		if((r->use2.b[z]|r->set.b[z]) & bb) {
-			if(debug['R'] && debug['v'])
-				print("%P", p);
-			addreg(&p->to, rn);
-			if(debug['R'] && debug['v'])
-				print(" ===change== %P\n", p);
-		}
-
-		if(STORE(r) & r->regdiff.b[z] & bb)
-			addmove(r, bn, rn, 1);
-		r->regu |= rb;
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					paint3(r1, bn, rb, rn);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				paint3(r1, bn, rb, rn);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(r->act.b[z] & bb)
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-}
-
-void
-addreg(Adr *a, int rn)
-{
-	a->sym = nil;
-	a->node = nil;
-	a->offset = 0;
-	a->type = TYPE_REG;
-	a->reg = rn;
-	a->name = 0;
-
-	ostats.ncvtreg++;
-}
-
-uint32
+uint64
 RtoB(int r)
 {
 
 	if(r < REG_AX || r > REG_R15)
 		return 0;
-	return 1L << (r-REG_AX);
+	return 1ULL << (r-REG_AX);
 }
 
 int
-BtoR(uint32 b)
+BtoR(uint64 b)
 {
-	b &= 0xffffL;
+	b &= 0xffffULL;
 	if(nacl)
 		b &= ~((1<<(REG_BP-REG_AX)) | (1<<(REG_R15-REG_AX)));
 	else if(framepointer_enabled)
@@ -1200,16 +134,16 @@ BtoR(uint32 b)
  *	...
  *	31	X15
  */
-uint32
+uint64
 FtoB(int f)
 {
 	if(f < REG_X0 || f > REG_X15)
 		return 0;
-	return 1L << (f - REG_X0 + 16);
+	return 1ULL << (f - REG_X0 + 16);
 }
 
 int
-BtoF(uint32 b)
+BtoF(uint64 b)
 {
 
 	b &= 0xFFFF0000L;
@@ -1217,77 +151,3 @@ BtoF(uint32 b)
 		return 0;
 	return bitno(b) - 16 + REG_X0;
 }
-
-void
-dumpone(Flow *f, int isreg)
-{
-	int z;
-	Bits bit;
-	Reg *r;
-
-	print("%d:%P", f->loop, f->prog);
-	if(isreg) {	
-		r = (Reg*)f;
-		for(z=0; z<BITS; z++)
-			bit.b[z] =
-				r->set.b[z] |
-				r->use1.b[z] |
-				r->use2.b[z] |
-				r->refbehind.b[z] |
-				r->refahead.b[z] |
-				r->calbehind.b[z] |
-				r->calahead.b[z] |
-				r->regdiff.b[z] |
-				r->act.b[z] |
-					0;
-		if(bany(&bit)) {
-			print("\t");
-			if(bany(&r->set))
-				print(" s:%Q", r->set);
-			if(bany(&r->use1))
-				print(" u1:%Q", r->use1);
-			if(bany(&r->use2))
-				print(" u2:%Q", r->use2);
-			if(bany(&r->refbehind))
-				print(" rb:%Q ", r->refbehind);
-			if(bany(&r->refahead))
-				print(" ra:%Q ", r->refahead);
-			if(bany(&r->calbehind))
-				print(" cb:%Q ", r->calbehind);
-			if(bany(&r->calahead))
-				print(" ca:%Q ", r->calahead);
-			if(bany(&r->regdiff))
-				print(" d:%Q ", r->regdiff);
-			if(bany(&r->act))
-				print(" a:%Q ", r->act);
-		}
-	}
-	print("\n");
-}
-
-void
-dumpit(char *str, Flow *r0, int isreg)
-{
-	Flow *r, *r1;
-
-	print("\n%s\n", str);
-	for(r = r0; r != nil; r = r->link) {
-		dumpone(r, isreg);
-		r1 = r->p2;
-		if(r1 != nil) {
-			print("	pred:");
-			for(; r1 != nil; r1 = r1->p2link)
-				print(" %.4ud", (int)r1->prog->pc);
-			print("\n");
-		}
-		// Print successors if it's not just the next one
-		if(r->s1 != r->link || r->s2 != nil) {
-			print("	succ:");
-			if(r->s1 != nil)
-				print(" %.4ud", (int)r->s1->prog->pc);
-			if(r->s2 != nil)
-				print(" %.4ud", (int)r->s2->prog->pc);
-			print("\n");
-		}
-	}
-}
diff --git a/src/cmd/6l/6.out.h b/src/cmd/6l/6.out.h
index 01d9bb7535..e25c6efec8 100644
--- a/src/cmd/6l/6.out.h
+++ b/src/cmd/6l/6.out.h
@@ -834,6 +834,7 @@ enum
 	FREGRET		= REG_X0,
 	REGSP		= REG_SP,
 	REGTMP		= REG_DI,
+	REGCTXT		= REG_DX,
 	REGEXT		= REG_R15,	/* compiler allocates external registers R15 down */
 	FREGMIN		= REG_X0+5,	/* first register variable */
 	FREGEXT		= REG_X0+15,	/* first external register */
diff --git a/src/cmd/6l/asm.c b/src/cmd/6l/asm.c
index 032c1eea09..a983373fa0 100644
--- a/src/cmd/6l/asm.c
+++ b/src/cmd/6l/asm.c
@@ -39,13 +39,6 @@
 
 #define PADDR(a)	((uint32)(a) & ~0x80000000)
 
-char linuxdynld[] = "/lib64/ld-linux-x86-64.so.2";
-char freebsddynld[] = "/libexec/ld-elf.so.1";
-char openbsddynld[] = "/usr/libexec/ld.so";
-char netbsddynld[] = "/libexec/ld.elf_so";
-char dragonflydynld[] = "/usr/libexec/ld-elf.so.2";
-char solarisdynld[] = "/lib/amd64/ld.so.1";
-
 char	zeroes[32];
 
 static int
@@ -68,8 +61,6 @@ needlib(char *name)
 	return 0;
 }
 
-int nelfsym = 1;
-
 static void addpltsym(LSym*);
 static void addgotsym(LSym*);
 
@@ -236,7 +227,7 @@ adddynrel(LSym *s, Reloc *r)
 			r->type = 256;	// ignore during relocsym
 			return;
 		}
-		if(HEADTYPE == Hdarwin && s->size == PtrSize && r->off == 0) {
+		if(HEADTYPE == Hdarwin && s->size == thearch.ptrsize && r->off == 0) {
 			// Mach-O relocations are a royal pain to lay out.
 			// They use a compact stateful bytecode representation
 			// that is too much bother to deal with.
@@ -271,7 +262,7 @@ elfreloc1(Reloc *r, vlong sectoff)
 {
 	int32 elfsym;
 
-	VPUT(sectoff);
+	thearch.vput(sectoff);
 
 	elfsym = r->xsym->elfsym;
 	switch(r->type) {
@@ -280,16 +271,16 @@ elfreloc1(Reloc *r, vlong sectoff)
 
 	case R_ADDR:
 		if(r->siz == 4)
-			VPUT(R_X86_64_32 | (uint64)elfsym<<32);
+			thearch.vput(R_X86_64_32 | (uint64)elfsym<<32);
 		else if(r->siz == 8)
-			VPUT(R_X86_64_64 | (uint64)elfsym<<32);
+			thearch.vput(R_X86_64_64 | (uint64)elfsym<<32);
 		else
 			return -1;
 		break;
 
 	case R_TLS_LE:
 		if(r->siz == 4)
-			VPUT(R_X86_64_TPOFF32 | (uint64)elfsym<<32);
+			thearch.vput(R_X86_64_TPOFF32 | (uint64)elfsym<<32);
 		else
 			return -1;
 		break;
@@ -297,16 +288,16 @@ elfreloc1(Reloc *r, vlong sectoff)
 	case R_CALL:
 		if(r->siz == 4) {
 			if(r->xsym->type == SDYNIMPORT)
-				VPUT(R_X86_64_GOTPCREL | (uint64)elfsym<<32);
+				thearch.vput(R_X86_64_GOTPCREL | (uint64)elfsym<<32);
 			else
-				VPUT(R_X86_64_PC32 | (uint64)elfsym<<32);
+				thearch.vput(R_X86_64_PC32 | (uint64)elfsym<<32);
 		} else
 			return -1;
 		break;
 
 	case R_PCREL:
 		if(r->siz == 4) {
-			VPUT(R_X86_64_PC32 | (uint64)elfsym<<32);
+			thearch.vput(R_X86_64_PC32 | (uint64)elfsym<<32);
 		} else
 			return -1;
 		break;
@@ -314,15 +305,15 @@ elfreloc1(Reloc *r, vlong sectoff)
 	case R_TLS:
 		if(r->siz == 4) {
 			if(flag_shared)
-				VPUT(R_X86_64_GOTTPOFF | (uint64)elfsym<<32);
+				thearch.vput(R_X86_64_GOTTPOFF | (uint64)elfsym<<32);
 			else
-				VPUT(R_X86_64_TPOFF32 | (uint64)elfsym<<32);
+				thearch.vput(R_X86_64_TPOFF32 | (uint64)elfsym<<32);
 		} else
 			return -1;
 		break;		
 	}
 
-	VPUT(r->xadd);
+	thearch.vput(r->xadd);
 	return 0;
 }
 
@@ -382,8 +373,8 @@ machoreloc1(Reloc *r, vlong sectoff)
 		break;
 	}
 
-	LPUT(sectoff);
-	LPUT(v);
+	thearch.lput(sectoff);
+	thearch.lput(v);
 	return 0;
 }
 
@@ -798,18 +789,3 @@ asmb(void)
 	}
 	cflush();
 }
-
-vlong
-rnd(vlong v, vlong r)
-{
-	vlong c;
-
-	if(r <= 0)
-		return v;
-	v += r - 1;
-	c = v % r;
-	if(c < 0)
-		c += r;
-	v -= c;
-	return v;
-}
diff --git a/src/cmd/6l/l.h b/src/cmd/6l/l.h
index 24eaa453dd..acc97d93d4 100644
--- a/src/cmd/6l/l.h
+++ b/src/cmd/6l/l.h
@@ -32,7 +32,6 @@
 #include	<libc.h>
 #include	<bio.h>
 #include	<link.h>
-#include	"6.out.h"
 
 #ifndef	EXTERN
 #define	EXTERN	extern
@@ -42,49 +41,14 @@ enum
 {
 	thechar = '6',
 	MaxAlign = 32,	// max data alignment
-	
-	// Loop alignment constants:
-	// want to align loop entry to LoopAlign-byte boundary,
-	// and willing to insert at most MaxLoopPad bytes of NOP to do so.
-	// We define a loop entry as the target of a backward jump.
-	//
-	// gcc uses MaxLoopPad = 10 for its 'generic x86-64' config,
-	// and it aligns all jump targets, not just backward jump targets.
-	//
-	// As of 6/1/2012, the effect of setting MaxLoopPad = 10 here
-	// is very slight but negative, so the alignment is disabled by
-	// setting MaxLoopPad = 0. The code is here for reference and
-	// for future experiments.
-	// 
-	LoopAlign = 16,
-	MaxLoopPad = 0,
-
 	FuncAlign = 16
 };
 
-EXTERN	int	PtrSize;
-EXTERN	int	IntSize;
-EXTERN	int	RegSize;
-
-#define	P		((Prog*)0)
-#define	S		((LSym*)0)
 enum
 {
 	MINLC		= 1,
 };
 
-#pragma	varargck	type	"I"	uchar*
-
-EXTERN	LSym*	datap;
-EXTERN	int	debug[128];
-EXTERN	char	literal[32];
-EXTERN	int32	lcsize;
-EXTERN	char*	rpath;
-EXTERN	int32	spsize;
-EXTERN	LSym*	symlist;
-EXTERN	int32	symsize;
-
-int	Iconv(Fmt *fp);
 void	adddynlib(char *lib);
 void	adddynrel(LSym *s, Reloc *r);
 void	adddynrela(LSym *rela, LSym *s, Reloc *r);
@@ -96,12 +60,6 @@ int	elfreloc1(Reloc *r, vlong sectoff);
 void	elfsetupplt(void);
 void	listinit(void);
 int	machoreloc1(Reloc *r, vlong sectoff);
-vlong	rnd(vlong v, vlong r);
-
-/* Native is little-endian */
-#define	LPUT(a)	lputl(a)
-#define	WPUT(a)	wputl(a)
-#define	VPUT(a)	vputl(a)
 
 /* Used by ../ld/dwarf.c */
 enum
diff --git a/src/cmd/6l/list.c b/src/cmd/6l/list.c
index d960fcc915..6c4ea79f91 100644
--- a/src/cmd/6l/list.c
+++ b/src/cmd/6l/list.c
@@ -37,31 +37,4 @@ void
 listinit(void)
 {
 	listinit6();
-	fmtinstall('I', Iconv);
-}
-
-int
-Iconv(Fmt *fp)
-{
-	int i, n;
-	uchar *p;
-	char *s;
-	Fmt fmt;
-	
-	n = fp->prec;
-	fp->prec = 0;
-	if(!(fp->flags&FmtPrec) || n < 0)
-		return fmtstrcpy(fp, "%I");
-	fp->flags &= ~FmtPrec;
-	p = va_arg(fp->args, uchar*);
-
-	// format into temporary buffer and
-	// call fmtstrcpy to handle padding.
-	fmtstrinit(&fmt);
-	for(i=0; i<n; i++)
-		fmtprint(&fmt, "%.2ux", *p++);
-	s = fmtstrflush(&fmt);
-	fmtstrcpy(fp, s);
-	free(s);
-	return 0;
 }
diff --git a/src/cmd/6l/obj.c b/src/cmd/6l/obj.c
index 3b8e8f4d7a..0a2b2cb59e 100644
--- a/src/cmd/6l/obj.c
+++ b/src/cmd/6l/obj.c
@@ -38,17 +38,52 @@
 #include	"../ld/pe.h"
 #include	<ar.h>
 
-char*	thestring 	= "amd64";
-LinkArch*	thelinkarch = &linkamd64;
+void
+main(int argc, char **argv)
+{
+	linkarchinit();
+	ldmain(argc, argv);
+}
 
 void
 linkarchinit(void)
 {
+	thestring = "amd64";
+	thelinkarch = &linkamd64;
 	if(strcmp(getgoarch(), "amd64p32") == 0)
 		thelinkarch = &linkamd64p32;
-	PtrSize = thelinkarch->ptrsize;
-	IntSize = PtrSize;
-	RegSize = thelinkarch->regsize;
+
+	thearch.thechar = thechar;
+	thearch.ptrsize = thelinkarch->ptrsize;
+	thearch.intsize = thelinkarch->ptrsize;
+	thearch.regsize = thelinkarch->regsize;
+	thearch.funcalign = FuncAlign;
+	thearch.maxalign = MaxAlign;
+	thearch.minlc = MINLC;
+	thearch.dwarfregsp = DWARFREGSP;
+
+	thearch.adddynlib = adddynlib;
+	thearch.adddynrel = adddynrel;
+	thearch.adddynsym = adddynsym;
+	thearch.archinit = archinit;
+	thearch.archreloc = archreloc;
+	thearch.archrelocvariant = archrelocvariant;
+	thearch.asmb = asmb;
+	thearch.elfreloc1 = elfreloc1;
+	thearch.elfsetupplt = elfsetupplt;
+	thearch.gentext = gentext;
+	thearch.listinit = listinit;
+	thearch.machoreloc1 = machoreloc1;
+	thearch.lput = lputl;
+	thearch.wput = wputl;
+	thearch.vput = vputl;
+	
+	thearch.linuxdynld = "/lib64/ld-linux-x86-64.so.2";
+	thearch.freebsddynld = "/libexec/ld-elf.so.1";
+	thearch.openbsddynld = "/usr/libexec/ld.so";
+	thearch.netbsddynld = "/libexec/ld.elf_so";
+	thearch.dragonflydynld = "/usr/libexec/ld-elf.so.2";
+	thearch.solarisdynld = "/lib/amd64/ld.so.1";
 }
 
 void
diff --git a/src/cmd/8g/cgen.c b/src/cmd/8g/cgen.c
index eabf52ae81..f06927c905 100644
--- a/src/cmd/8g/cgen.c
+++ b/src/cmd/8g/cgen.c
@@ -956,17 +956,7 @@ bgen(Node *n, int true, int likely, Prog *to)
 
 	switch(n->op) {
 	default:
-	def:
-		regalloc(&n1, n->type, N);
-		cgen(n, &n1);
-		nodconst(&n2, n->type, 0);
-		gins(optoas(OCMP, n->type), &n1, &n2);
-		a = AJNE;
-		if(!true)
-			a = AJEQ;
-		patch(gbranch(a, n->type, likely), to);
-		regfree(&n1);
-		return;
+		goto def;
 
 	case OLITERAL:
 		// need to ask if it is bool?
@@ -986,27 +976,20 @@ bgen(Node *n, int true, int likely, Prog *to)
 		return;
 
 	case OANDAND:
-		if(!true)
-			goto caseor;
-
-	caseand:
-		p1 = gbranch(AJMP, T, 0);
-		p2 = gbranch(AJMP, T, 0);
-		patch(p1, pc);
-		bgen(n->left, !true, -likely, p2);
-		bgen(n->right, !true, -likely, p2);
-		p1 = gbranch(AJMP, T, 0);
-		patch(p1, to);
-		patch(p2, pc);
-		return;
-
 	case OOROR:
-		if(!true)
-			goto caseand;
-
-	caseor:
-		bgen(n->left, true, likely, to);
-		bgen(n->right, true, likely, to);
+		if((n->op == OANDAND) == true) {
+			p1 = gbranch(AJMP, T, 0);
+			p2 = gbranch(AJMP, T, 0);
+			patch(p1, pc);
+			bgen(n->left, !true, -likely, p2);
+			bgen(n->right, !true, -likely, p2);
+			p1 = gbranch(AJMP, T, 0);
+			patch(p1, to);
+			patch(p2, pc);
+		} else {
+			bgen(n->left, true, likely, to);
+			bgen(n->right, true, likely, to);
+		}
 		return;
 
 	case OEQ:
@@ -1150,6 +1133,19 @@ cmp:
 		regfree(nr);
 		break;
 	}
+	return;
+
+def:
+	regalloc(&n1, n->type, N);
+	cgen(n, &n1);
+	nodconst(&n2, n->type, 0);
+	gins(optoas(OCMP, n->type), &n1, &n2);
+	a = AJNE;
+	if(!true)
+		a = AJEQ;
+	patch(gbranch(a, n->type, likely), to);
+	regfree(&n1);
+	return;
 }
 
 /*
@@ -1213,7 +1209,7 @@ stkof(Node *n)
 void
 sgen(Node *n, Node *res, int64 w)
 {
-	Node dst, src, tdst, tsrc;
+	Node dst, src, tdst, tsrc, cx;
 	int32 c, q, odst, osrc;
 	NodeList *l;
 	Prog *p;
@@ -1329,6 +1325,19 @@ sgen(Node *n, Node *res, int64 w)
 			p->to.sym = linksym(pkglookup("duffcopy", runtimepkg));
 			// 10 and 128 = magic constants: see ../../runtime/asm_386.s
 			p->to.offset = 10*(128-q);
+		} else if(!nacl && c == 0) {
+			nodreg(&cx, types[TINT32], REG_CX);
+			// We don't need the MOVSL side-effect of updating SI and DI,
+			// and issuing a sequence of MOVLs directly is faster.
+			src.op = OINDREG;
+			dst.op = OINDREG;
+			while(q > 0) {
+				gmove(&src, &cx); // MOVL x+(SI),CX
+				gmove(&cx, &dst); // MOVL CX,x+(DI)
+				src.xoffset += 4;
+				dst.xoffset += 4;
+				q--;
+			}
 		} else
 		while(q > 0) {
 			gins(AMOVSL, N, N);	// MOVL *(SI)+,*(DI)+
@@ -1360,13 +1369,15 @@ cadable(Node *n)
 /*
  * copy a composite value by moving its individual components.
  * Slices, strings and interfaces are supported.
+ * Small structs or arrays with elements of basic type are
+ * also supported.
  * nr is N when assigning a zero value.
  * return 1 if can do, 0 if can't.
  */
 int
 componentgen(Node *nr, Node *nl)
 {
-	Node nodl, nodr;
+	Node nodl, nodr, tmp;
 	Type *t;
 	int freel, freer;
 	vlong fldcount;
@@ -1414,7 +1425,7 @@ componentgen(Node *nr, Node *nl)
 
 	nodl = *nl;
 	if(!cadable(nl)) {
-		if(nr == N || !cadable(nr))
+		if(nr != N && !cadable(nr))
 			goto no;
 		igen(nl, &nodl, N);
 		freel = 1;
@@ -1426,6 +1437,12 @@ componentgen(Node *nr, Node *nl)
 			igen(nr, &nodr, N);
 			freer = 1;
 		}
+	} else {
+		// When zeroing, prepare a register containing zero.
+		nodconst(&tmp, nl->type, 0);
+		regalloc(&nodr, types[TUINT], N);
+		gmove(&tmp, &nodr);
+		freer = 1;
 	}
 	
 	// nl and nr are 'cadable' which basically means they are names (variables) now.
@@ -1462,8 +1479,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_nel-Array_array;
@@ -1472,8 +1488,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_nel-Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_cap-Array_nel;
@@ -1482,8 +1497,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_cap-Array_nel;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		goto yes;
@@ -1497,8 +1511,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_nel-Array_array;
@@ -1507,8 +1520,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_nel-Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		goto yes;
@@ -1522,8 +1534,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_nel-Array_array;
@@ -1532,8 +1543,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_nel-Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		goto yes;
diff --git a/src/cmd/8g/galign.c b/src/cmd/8g/galign.c
index 83c323516d..33951adfd2 100644
--- a/src/cmd/8g/galign.c
+++ b/src/cmd/8g/galign.c
@@ -42,61 +42,46 @@ betypeinit(void)
 void
 main(int argc, char **argv)
 {
-	arch.thechar = thechar;
-	arch.thestring = thestring;
-	arch.thelinkarch = thelinkarch;
-	arch.typedefs = typedefs;
-	arch.MAXWIDTH = MAXWIDTH;
-	arch.afunclit = afunclit;
-	arch.anyregalloc = anyregalloc;
-	arch.betypeinit = betypeinit;
-	arch.bgen = bgen;
-	arch.cgen = cgen;
-	arch.cgen_asop = cgen_asop;
-	arch.cgen_call = cgen_call;
-	arch.cgen_callinter = cgen_callinter;
-	arch.cgen_ret = cgen_ret;
-	arch.clearfat = clearfat;
-	arch.clearp = clearp;
-	arch.defframe = defframe;
-	arch.dgostringptr = dgostringptr;
-	arch.dgostrlitptr = dgostrlitptr;
-	arch.dsname = dsname;
-	arch.dsymptr = dsymptr;
-	arch.dumpdata = dumpdata;
-	arch.dumpit = dumpit;
-	arch.excise = excise;
-	arch.expandchecks = expandchecks;
-	arch.fixautoused = fixautoused;
-	arch.gclean = gclean;
-	arch.gdata = gdata;
-	arch.gdatacomplex = gdatacomplex;
-	arch.gdatastring = gdatastring;
-	arch.ggloblnod = ggloblnod;
-	arch.ggloblsym = ggloblsym;
-	arch.ginit = ginit;
-	arch.gins = gins;
-	arch.ginscall = ginscall;
-	arch.gjmp = gjmp;
-	arch.gtrack = gtrack;
-	arch.gused = gused;
-	arch.igen = igen;
-	arch.isfat = isfat;
-	arch.linkarchinit = linkarchinit;
-	arch.markautoused = markautoused;
-	arch.naddr = naddr;
-	arch.newplist = newplist;
-	arch.nodarg = nodarg;
-	arch.patch = patch;
-	arch.proginfo = proginfo;
-	arch.regalloc = regalloc;
-	arch.regfree = regfree;
-	arch.regopt = regopt;
-	arch.regtyp = regtyp;
-	arch.sameaddr = sameaddr;
-	arch.smallindir = smallindir;
-	arch.stackaddr = stackaddr;
-	arch.unpatch = unpatch;
+	thearch.thechar = thechar;
+	thearch.thestring = thestring;
+	thearch.thelinkarch = thelinkarch;
+	thearch.typedefs = typedefs;
+	thearch.REGSP = REGSP;
+	thearch.REGCTXT = REGCTXT;
+	thearch.MAXWIDTH = MAXWIDTH;
+	thearch.anyregalloc = anyregalloc;
+	thearch.betypeinit = betypeinit;
+	thearch.bgen = bgen;
+	thearch.cgen = cgen;
+	thearch.cgen_call = cgen_call;
+	thearch.cgen_callinter = cgen_callinter;
+	thearch.cgen_ret = cgen_ret;
+	thearch.clearfat = clearfat;
+	thearch.defframe = defframe;
+	thearch.excise = excise;
+	thearch.expandchecks = expandchecks;
+	thearch.gclean = gclean;
+	thearch.ginit = ginit;
+	thearch.gins = gins;
+	thearch.ginscall = ginscall;
+	thearch.igen = igen;
+	thearch.linkarchinit = linkarchinit;
+	thearch.peep = peep;
+	thearch.proginfo = proginfo;
+	thearch.regalloc = regalloc;
+	thearch.regfree = regfree;
+	thearch.regtyp = regtyp;
+	thearch.sameaddr = sameaddr;
+	thearch.smallindir = smallindir;
+	thearch.stackaddr = stackaddr;
+	thearch.excludedregs = excludedregs;
+	thearch.RtoB = RtoB;
+	thearch.FtoB = FtoB;
+	thearch.BtoR = BtoR;
+	thearch.BtoF = BtoF;
+	thearch.optoas = optoas;
+	thearch.doregbits = doregbits;
+	thearch.regnames = regnames;
 	
 	gcmain(argc, argv);
 }
diff --git a/src/cmd/8g/gg.h b/src/cmd/8g/gg.h
index b2b1178a52..872d946592 100644
--- a/src/cmd/8g/gg.h
+++ b/src/cmd/8g/gg.h
@@ -17,10 +17,7 @@ enum
 	Fpop2 = 1<<2,
 };
 
-EXTERN	int32	dynloc;
 EXTERN	uchar	reg[MAXREG];
-EXTERN	int32	pcloc;		// instruction counter
-EXTERN	Strlit	emptystring;
 EXTERN	Node*	panicdiv;
 extern	uint32	unmappedzero;
 
@@ -130,7 +127,6 @@ int anyregalloc(void);
 void betypeinit(void);
 void bgen(Node*, int, int, Prog*);
 void cgen(Node*, Node*);
-void cgen_asop(Node*);
 void cgen_call(Node*, int);
 void cgen_callinter(Node*, Node*, int);
 void cgen_ret(Node*);
@@ -175,3 +171,19 @@ int sameaddr(Addr*, Addr*);
 int smallindir(Addr*, Addr*);
 int stackaddr(Addr*);
 Prog* unpatch(Prog*);
+
+/*
+ * reg.c
+ */
+uint64 excludedregs(void);
+uint64 RtoB(int);
+uint64 FtoB(int);
+int BtoR(uint64);
+int BtoF(uint64);
+uint64 doregbits(int);
+char** regnames(int*);
+
+/*
+ * peep.c
+ */
+void peep(Prog*);
diff --git a/src/cmd/8g/ggen.c b/src/cmd/8g/ggen.c
index 2fd4a601ff..8188348282 100644
--- a/src/cmd/8g/ggen.c
+++ b/src/cmd/8g/ggen.c
@@ -7,7 +7,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
 static Prog *appendpp(Prog*, int, int, int, vlong, int, int, vlong);
 static Prog *zerorange(Prog *p, vlong frame, vlong lo, vlong hi, uint32 *ax);
@@ -106,52 +106,6 @@ appendpp(Prog *p, int as, int ftype, int freg, vlong foffset, int ttype, int tre
 	return q;	
 }
 
-// Sweep the prog list to mark any used nodes.
-void
-markautoused(Prog* p)
-{
-	for (; p; p = p->link) {
-		if (p->as == ATYPE || p->as == AVARDEF || p->as == AVARKILL)
-			continue;
-
-		if (p->from.node)
-			((Node*)(p->from.node))->used = 1;
-
-		if (p->to.node)
-			((Node*)(p->to.node))->used = 1;
-	}
-}
-
-// Fixup instructions after allocauto (formerly compactframe) has moved all autos around.
-void
-fixautoused(Prog* p)
-{
-	Prog **lp;
-
-	for (lp=&p; (p=*lp) != P; ) {
-		if (p->as == ATYPE && p->from.node && p->from.name == NAME_AUTO && !((Node*)(p->from.node))->used) {
-			*lp = p->link;
-			continue;
-		}
-		if ((p->as == AVARDEF || p->as == AVARKILL) && p->to.node && !((Node*)(p->to.node))->used) {
-			// Cannot remove VARDEF instruction, because - unlike TYPE handled above -
-			// VARDEFs are interspersed with other code, and a jump might be using the
-			// VARDEF as a target. Replace with a no-op instead. A later pass will remove
-			// the no-ops.
-			nopout(p);
-			continue;
-		}
-
-		if (p->from.type == TYPE_MEM && p->from.name == NAME_AUTO && p->from.node)
-			p->from.offset += ((Node*)(p->from.node))->stkdelta;
-
-		if (p->to.type == TYPE_MEM && p->to.name == NAME_AUTO && p->to.node)
-			p->to.offset += ((Node*)(p->to.node))->stkdelta;
-
-		lp = &p->link;
-	}
-}
-
 void
 clearfat(Node *nl)
 {
@@ -525,154 +479,6 @@ cgen_ret(Node *n)
 }
 
 /*
- * generate += *= etc.
- */
-void
-cgen_asop(Node *n)
-{
-	Node n1, n2, n3, n4;
-	Node *nl, *nr;
-	Prog *p1;
-	Addr addr;
-	int a;
-
-	nl = n->left;
-	nr = n->right;
-
-	if(nr->ullman >= UINF && nl->ullman >= UINF) {
-		tempname(&n1, nr->type);
-		cgen(nr, &n1);
-		n2 = *n;
-		n2.right = &n1;
-		cgen_asop(&n2);
-		goto ret;
-	}
-
-	if(!isint[nl->type->etype])
-		goto hard;
-	if(!isint[nr->type->etype])
-		goto hard;
-	if(is64(nl->type) || is64(nr->type))
-		goto hard;
-
-	switch(n->etype) {
-	case OADD:
-		if(smallintconst(nr))
-		if(mpgetfix(nr->val.u.xval) == 1) {
-			a = optoas(OINC, nl->type);
-			if(nl->addable) {
-				gins(a, N, nl);
-				goto ret;
-			}
-			if(sudoaddable(a, nl, &addr)) {
-				p1 = gins(a, N, N);
-				p1->to = addr;
-				sudoclean();
-				goto ret;
-			}
-		}
-		break;
-
-	case OSUB:
-		if(smallintconst(nr))
-		if(mpgetfix(nr->val.u.xval) == 1) {
-			a = optoas(ODEC, nl->type);
-			if(nl->addable) {
-				gins(a, N, nl);
-				goto ret;
-			}
-			if(sudoaddable(a, nl, &addr)) {
-				p1 = gins(a, N, N);
-				p1->to = addr;
-				sudoclean();
-				goto ret;
-			}
-		}
-		break;
-	}
-
-	switch(n->etype) {
-	case OADD:
-	case OSUB:
-	case OXOR:
-	case OAND:
-	case OOR:
-		a = optoas(n->etype, nl->type);
-		if(nl->addable) {
-			if(smallintconst(nr)) {
-				gins(a, nr, nl);
-				goto ret;
-			}
-			regalloc(&n2, nr->type, N);
-			cgen(nr, &n2);
-			gins(a, &n2, nl);
-			regfree(&n2);
-			goto ret;
-		}
-		if(nr->ullman < UINF)
-		if(sudoaddable(a, nl, &addr)) {
-			if(smallintconst(nr)) {
-				p1 = gins(a, nr, N);
-				p1->to = addr;
-				sudoclean();
-				goto ret;
-			}
-			regalloc(&n2, nr->type, N);
-			cgen(nr, &n2);
-			p1 = gins(a, &n2, N);
-			p1->to = addr;
-			regfree(&n2);
-			sudoclean();
-			goto ret;
-		}
-	}
-
-hard:
-	n2.op = 0;
-	n1.op = 0;
-	if(nr->ullman >= nl->ullman || nl->addable) {
-		mgen(nr, &n2, N);
-		nr = &n2;
-	} else {
-		tempname(&n2, nr->type);
-		cgen(nr, &n2);
-		nr = &n2;
-	}
-	if(!nl->addable) {
-		igen(nl, &n1, N);
-		nl = &n1;
-	}
-
-	n3 = *n;
-	n3.left = nl;
-	n3.right = nr;
-	n3.op = n->etype;
-
-	mgen(&n3, &n4, N);
-	gmove(&n4, nl);
-
-	if(n1.op)
-		regfree(&n1);
-	mfree(&n2);
-	mfree(&n4);
-
-ret:
-	;
-}
-
-int
-samereg(Node *a, Node *b)
-{
-	if(a->op != OREGISTER)
-		return 0;
-	if(b->op != OREGISTER)
-		return 0;
-	if(a->val.u.reg != b->val.u.reg)
-		return 0;
-	return 1;
-}
-
-/*
  * generate division.
  * caller must set:
  *	ax = allocated AX register
diff --git a/src/cmd/8g/gobj.c b/src/cmd/8g/gobj.c
deleted file mode 100644
index b59d2268dd..0000000000
--- a/src/cmd/8g/gobj.c
+++ /dev/null
@@ -1,257 +0,0 @@
-// Derived from Inferno utils/8c/swt.c
-// http://code.google.com/p/inferno-os/source/browse/utils/8c/swt.c
-//
-//	Copyright © 1994-1999 Lucent Technologies Inc.  All rights reserved.
-//	Portions Copyright © 1995-1997 C H Forsyth (forsyth@terzarima.net)
-//	Portions Copyright © 1997-1999 Vita Nuova Limited
-//	Portions Copyright © 2000-2007 Vita Nuova Holdings Limited (www.vitanuova.com)
-//	Portions Copyright © 2004,2006 Bruce Ellis
-//	Portions Copyright © 2005-2007 C H Forsyth (forsyth@terzarima.net)
-//	Revisions Copyright © 2000-2007 Lucent Technologies Inc. and others
-//	Portions Copyright © 2009 The Go Authors.  All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-//
-// The above copyright notice and this permission notice shall be included in
-// all copies or substantial portions of the Software.
-//
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
-// THE SOFTWARE.
-
-#include <u.h>
-#include <libc.h>
-#include "gg.h"
-
-int
-dsname(Sym *s, int off, char *t, int n)
-{
-	Prog *p;
-
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = n;
-	p->from.sym = linksym(s);
-	
-	p->to.type = TYPE_SCONST;
-	memmove(p->to.u.sval, t, n);
-	return off + n;
-}
-
-/*
- * make a refer to the data s, s+len
- * emitting DATA if needed.
- */
-void
-datastring(char *s, int len, Addr *a)
-{
-	Sym *sym;
-	
-	sym = stringsym(s, len);
-	a->type = TYPE_MEM;
-	a->name = NAME_EXTERN;
-	a->sym = linksym(sym);
-	a->node = sym->def;
-	a->offset = widthptr+4;  // skip header
-	a->etype = TINT32;
-}
-
-/*
- * make a refer to the string sval,
- * emitting DATA if needed.
- */
-void
-datagostring(Strlit *sval, Addr *a)
-{
-	Sym *sym;
-	
-	sym = stringsym(sval->s, sval->len);
-	a->type = TYPE_MEM;
-	a->name = NAME_EXTERN;
-	a->sym = linksym(sym);
-	a->node = sym->def;
-	a->offset = 0;  // header
-	a->etype = TSTRING;
-}
-
-void
-gdata(Node *nam, Node *nr, int wid)
-{
-	Prog *p;
-	vlong v;
-
-	if(nr->op == OLITERAL) {
-		switch(nr->val.ctype) {
-		case CTCPLX:
-			gdatacomplex(nam, nr->val.u.cval);
-			return;
-		case CTSTR:
-			gdatastring(nam, nr->val.u.sval);
-			return;
-		}
-	}
-
-	if(wid == 8 && is64(nr->type)) {
-		v = mpgetfix(nr->val.u.xval);
-		p = gins(ADATA, nam, nodintconst(v));
-		p->from3.type = TYPE_CONST;
-		p->from3.offset = 4;
-		p = gins(ADATA, nam, nodintconst(v>>32));
-		p->from3.type = TYPE_CONST;
-		p->from3.offset = 4;
-		p->from.offset += 4;
-		return;
-	}
-	p = gins(ADATA, nam, nr);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = wid;
-}
-
-void
-gdatacomplex(Node *nam, Mpcplx *cval)
-{
-	Prog *p;
-	int w;
-
-	w = cplxsubtype(nam->type->etype);
-	w = types[w]->width;
-
-	p = gins(ADATA, nam, N);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = w;
-	p->to.type = TYPE_FCONST;
-	p->to.u.dval = mpgetflt(&cval->real);
-
-	p = gins(ADATA, nam, N);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = w;
-	p->from.offset += w;
-	p->to.type = TYPE_FCONST;
-	p->to.u.dval = mpgetflt(&cval->imag);
-}
-
-void
-gdatastring(Node *nam, Strlit *sval)
-{
-	Prog *p;
-	Node nod1;
-
-	p = gins(ADATA, nam, N);
-	datastring(sval->s, sval->len, &p->to);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = types[tptr]->width;
-	p->to.type = TYPE_ADDR;
-//print("%P\n", p);
-
-	nodconst(&nod1, types[TINT32], sval->len);
-	p = gins(ADATA, nam, &nod1);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = types[TINT32]->width;
-	p->from.offset += types[tptr]->width;
-}
-
-int
-dstringptr(Sym *s, int off, char *str)
-{
-	Prog *p;
-
-	off = rnd(off, widthptr);
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-
-	datastring(str, strlen(str)+1, &p->to);
-	p->to.type = TYPE_ADDR;
-	p->to.etype = TINT32;
-	off += widthptr;
-
-	return off;
-}
-
-int
-dgostrlitptr(Sym *s, int off, Strlit *lit)
-{
-	Prog *p;
-
-	if(lit == nil)
-		return duintptr(s, off, 0);
-
-	off = rnd(off, widthptr);
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-	datagostring(lit, &p->to);
-	p->to.type = TYPE_ADDR;
-	p->to.etype = TINT32;
-	off += widthptr;
-
-	return off;
-}
-
-int
-dgostringptr(Sym *s, int off, char *str)
-{
-	int n;
-	Strlit *lit;
-
-	if(str == nil)
-		return duintptr(s, off, 0);
-
-	n = strlen(str);
-	lit = mal(sizeof *lit + n);
-	strcpy(lit->s, str);
-	lit->len = n;
-	return dgostrlitptr(s, off, lit);
-}
-
-int
-dsymptr(Sym *s, int off, Sym *x, int xoff)
-{
-	Prog *p;
-
-	off = rnd(off, widthptr);
-
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-	p->to.type = TYPE_ADDR;
-	p->to.name = NAME_EXTERN;
-	p->to.sym = linksym(x);
-	p->to.offset = xoff;
-	off += widthptr;
-
-	return off;
-}
-
-void
-nopout(Prog *p)
-{
-	p->as = ANOP;
-	p->from = zprog.from;
-	p->to = zprog.to;
-}
-
diff --git a/src/cmd/8g/gsubr.c b/src/cmd/8g/gsubr.c
index dd75e9f092..b82f7622ef 100644
--- a/src/cmd/8g/gsubr.c
+++ b/src/cmd/8g/gsubr.c
@@ -41,225 +41,6 @@ uint32 unmappedzero = 4096;
 #define	CASE(a,b)	(((a)<<16)|((b)<<0))
 /*c2go int CASE(int, int);*/
 
-void
-clearp(Prog *p)
-{
-	p->as = AEND;
-	p->from.type = TYPE_NONE;
-	p->from.index = TYPE_NONE;
-	p->to.type = TYPE_NONE;
-	p->to.index = TYPE_NONE;
-	p->pc = pcloc;
-	pcloc++;
-}
-
-static int ddumped;
-static Prog *dfirst;
-static Prog *dpc;
-
-/*
- * generate and return proc with p->as = as,
- * linked into program.  pc is next instruction.
- */
-Prog*
-prog(int as)
-{
-	Prog *p;
-
-	if(as == ADATA || as == AGLOBL) {
-		if(ddumped)
-			fatal("already dumped data");
-		if(dpc == nil) {
-			dpc = mal(sizeof(*dpc));
-			dfirst = dpc;
-		}
-		p = dpc;
-		dpc = mal(sizeof(*dpc));
-		p->link = dpc;
-	} else {
-		p = pc;
-		pc = mal(sizeof(*pc));
-		clearp(pc);
-		p->link = pc;
-	}
-
-	if(lineno == 0) {
-		if(debug['K'])
-			warn("prog: line 0");
-	}
-
-	p->as = as;
-	p->lineno = lineno;
-	return p;
-}
-
-void
-dumpdata(void)
-{
-	ddumped = 1;
-	if(dfirst == nil)
-		return;
-	newplist();
-	*pc = *dfirst;
-	pc = dpc;
-	clearp(pc);
-}
-
-/*
- * generate a branch.
- * t is ignored.
- * likely values are for branch prediction:
- *	-1 unlikely
- *	0 no opinion
- *	+1 likely
- */
-Prog*
-gbranch(int as, Type *t, int likely)
-{
-	Prog *p;
-
-	USED(t);
-	p = prog(as);
-	p->to.type = TYPE_BRANCH;
-	p->to.u.branch = P;
-	if(likely != 0) {
-		p->from.type = TYPE_CONST;
-		p->from.offset = likely > 0;
-	}
-	return p;
-}
-
-/*
- * patch previous branch to jump to to.
- */
-void
-patch(Prog *p, Prog *to)
-{
-	if(p->to.type != TYPE_BRANCH)
-		fatal("patch: not a branch");
-	p->to.u.branch = to;
-	p->to.offset = to->pc;
-}
-
-Prog*
-unpatch(Prog *p)
-{
-	Prog *q;
-
-	if(p->to.type != TYPE_BRANCH)
-		fatal("unpatch: not a branch");
-	q = p->to.u.branch;
-	p->to.u.branch = P;
-	p->to.offset = 0;
-	return q;
-}
-
-/*
- * start a new Prog list.
- */
-Plist*
-newplist(void)
-{
-	Plist *pl;
-
-	pl = linknewplist(ctxt);
-	
-	pc = mal(sizeof(*pc));
-	clearp(pc);
-	pl->firstpc = pc;
-
-	return pl;
-}
-
-void
-gused(Node *n)
-{
-	gins(ANOP, n, N);	// used
-}
-
-Prog*
-gjmp(Prog *to)
-{
-	Prog *p;
-
-	p = gbranch(AJMP, T, 0);
-	if(to != P)
-		patch(p, to);
-	return p;
-}
-
-void
-ggloblnod(Node *nam)
-{
-	Prog *p;
-
-	p = gins(AGLOBL, nam, N);
-	p->lineno = nam->lineno;
-	p->from.sym->gotype = linksym(ngotype(nam));
-	p->to.sym = nil;
-	p->to.type = TYPE_CONST;
-	p->to.offset = nam->type->width;
-	if(nam->readonly)
-		p->from3.offset = RODATA;
-	if(nam->type != T && !haspointers(nam->type))
-		p->from3.offset |= NOPTR;
-}
-
-void
-ggloblsym(Sym *s, int32 width, int8 flags)
-{
-	Prog *p;
-
-	p = gins(AGLOBL, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.index = REG_NONE;
-	p->from.sym = linksym(s);
-	p->to.type = TYPE_CONST;
-	p->to.index = REG_NONE;
-	p->to.offset = width;
-	p->from3.offset = flags;
-}
-
-void
-gtrack(Sym *s)
-{
-	Prog *p;
-	
-	p = gins(AUSEFIELD, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-}
-
-int
-isfat(Type *t)
-{
-	if(t != T)
-	switch(t->etype) {
-	case TSTRUCT:
-	case TARRAY:
-	case TSTRING:
-	case TINTER:	// maybe remove later
-		return 1;
-	}
-	return 0;
-}
-
-/*
- * naddr of func generates code for address of func.
- * if using opcode that can take address implicitly,
- * call afunclit to fix up the argument.
- */
-void
-afunclit(Addr *a, Node *n)
-{
-	if(a->type == TYPE_ADDR && a->name == NAME_EXTERN) {
-		a->type = TYPE_MEM;
-		a->sym = linksym(n->sym);
-	}
-}
-
 /*
  * return Axxx for Oxxx on type t.
  */
@@ -406,6 +187,14 @@ optoas(int op, Type *t)
 	case CASE(OAS, TPTR32):
 		a = AMOVL;
 		break;
+	
+	case CASE(OAS, TFLOAT32):
+		a = AMOVSS;
+		break;
+	
+	case CASE(OAS, TFLOAT64):
+		a = AMOVSD;
+		break;
 
 	case CASE(OADD, TINT8):
 	case CASE(OADD, TUINT8):
@@ -878,7 +667,7 @@ gclean(void)
 			yyerror("reg %R left allocated\n", i);
 }
 
-int32
+int
 anyregalloc(void)
 {
 	int i, j;
@@ -935,9 +724,9 @@ regalloc(Node *n, Type *t, Node *o)
 			if(reg[i] == 0)
 				goto out;
 
-		fprint(2, "registers allocated at\n");
+		print("registers allocated at\n");
 		for(i=REG_AX; i<=REG_DI; i++)
-			fprint(2, "\t%R\t%#lux\n", i, regpc[i]);
+			print("\t%R\t%#lux\n", i, regpc[i]);
 		fatal("out of fixed registers");
 		goto err;
 
@@ -955,9 +744,9 @@ regalloc(Node *n, Type *t, Node *o)
 		for(i=REG_X0; i<=REG_X7; i++)
 			if(reg[i] == 0)
 				goto out;
-		fprint(2, "registers allocated at\n");
+		print("registers allocated at\n");
 		for(i=REG_X0; i<=REG_X7; i++)
-			fprint(2, "\t%R\t%#lux\n", i, regpc[i]);
+			print("\t%R\t%#lux\n", i, regpc[i]);
 		fatal("out of floating registers");
 	}
 	yyerror("regalloc: unknown type %T", t);
@@ -1002,105 +791,6 @@ regfree(Node *n)
 }
 
 /*
- * initialize n to be register r of type t.
- */
-void
-nodreg(Node *n, Type *t, int r)
-{
-	if(t == T)
-		fatal("nodreg: t nil");
-
-	memset(n, 0, sizeof(*n));
-	n->op = OREGISTER;
-	n->addable = 1;
-	ullmancalc(n);
-	n->val.u.reg = r;
-	n->type = t;
-}
-
-/*
- * initialize n to be indirect of register r; n is type t.
- */
-void
-nodindreg(Node *n, Type *t, int r)
-{
-	nodreg(n, t, r);
-	n->op = OINDREG;
-}
-
-Node*
-nodarg(Type *t, int fp)
-{
-	Node *n;
-	NodeList *l;
-	Type *first;
-	Iter savet;
-
-	// entire argument struct, not just one arg
-	switch(t->etype) {
-	default:
-		fatal("nodarg %T", t);
-
-	case TSTRUCT:
-		if(!t->funarg)
-			fatal("nodarg: TSTRUCT but not funarg");
-		n = nod(ONAME, N, N);
-		n->sym = lookup(".args");
-		n->type = t;
-		first = structfirst(&savet, &t);
-		if(first == nil)
-			fatal("nodarg: bad struct");
-		if(first->width == BADWIDTH)
-			fatal("nodarg: offset not computed for %T", t);
-		n->xoffset = first->width;
-		n->addable = 1;
-		break;
-
-	case TFIELD:
-		if(fp == 1 && t->sym != S && !isblanksym(t->sym)) {
-			for(l=curfn->dcl; l; l=l->next) {
-				n = l->n;
-				if((n->class == PPARAM || n->class == PPARAMOUT) && n->sym == t->sym)
-					return n;
-			}
-		}
-
-		n = nod(ONAME, N, N);
-		n->type = t->type;
-		n->sym = t->sym;
-		if(t->width == BADWIDTH)
-			fatal("nodarg: offset not computed for %T", t);
-		n->xoffset = t->width;
-		n->addable = 1;
-		n->orig = t->nname;
-		break;
-	}
-	
-	// Rewrite argument named _ to __,
-	// or else the assignment to _ will be
-	// discarded during code generation.
-	if(isblank(n))
-		n->sym = lookup("__");
-
-	switch(fp) {
-	default:
-		fatal("nodarg %T %d", t, fp);
-
-	case 0:		// output arg
-		n->op = OINDREG;
-		n->val.u.reg = REG_SP;
-		break;
-
-	case 1:		// input arg
-		n->class = PPARAM;
-		break;
-	}
-
-	n->typecheck = 1;
-	return n;
-}
-
-/*
  * generate
  *	as $c, reg
  */
@@ -1142,26 +832,6 @@ ncon(uint32 i)
 	return &n;
 }
 
-/*
- * Is this node a memory operand?
- */
-int
-ismem(Node *n)
-{
-	switch(n->op) {
-	case OITAB:
-	case OSPTR:
-	case OLEN:
-	case OCAP:
-	case OINDREG:
-	case ONAME:
-	case OPARAM:
-	case OCLOSUREVAR:
-		return 1;
-	}
-	return 0;
-}
-
 Node sclean[10];
 int nsclean;
 
@@ -1183,22 +853,25 @@ split64(Node *n, Node *lo, Node *hi)
 	nsclean++;
 	switch(n->op) {
 	default:
-		if(!dotaddable(n, &n1)) {
-			igen(n, &n1, N);
-			sclean[nsclean-1] = n1;
-		}
-		n = &n1;
-		goto common;
-	case ONAME:
-		if(n->class == PPARAMREF) {
-			cgen(n->heapaddr, &n1);
-			sclean[nsclean-1] = n1;
-			// fall through.
+		switch(n->op) {
+		default:
+			if(!dotaddable(n, &n1)) {
+				igen(n, &n1, N);
+				sclean[nsclean-1] = n1;
+			}
 			n = &n1;
+			break;
+		case ONAME:
+			if(n->class == PPARAMREF) {
+				cgen(n->heapaddr, &n1);
+				sclean[nsclean-1] = n1;
+				n = &n1;
+			}
+			break;
+		case OINDREG:
+			// nothing
+			break;
 		}
-		goto common;
-	case OINDREG:
-	common:
 		*lo = *n;
 		*hi = *n;
 		lo->type = types[TUINT32];
@@ -1408,8 +1081,8 @@ gmove(Node *f, Node *t)
 			gins(AMOVL, &flo, &tlo);
 			gins(AMOVL, &fhi, &thi);
 		} else {
-			nodreg(&r1, t->type, REG_AX);
-			nodreg(&r2, t->type, REG_DX);
+			nodreg(&r1, types[TUINT32], REG_AX);
+			nodreg(&r2, types[TUINT32], REG_DX);
 			gins(AMOVL, &flo, &r1);
 			gins(AMOVL, &fhi, &r2);
 			gins(AMOVL, &r1, &tlo);
@@ -1796,7 +1469,7 @@ floatmove_387(Node *f, Node *t)
 		gmove(f, &t1);
 		switch(tt) {
 		default:
-			fatal("gmove %T", t);
+			fatal("gmove %N", t);
 		case TINT8:
 			gins(ACMPL, &t1, ncon(-0x80));
 			p1 = gbranch(optoas(OLT, types[TINT32]), T, -1);
@@ -2165,209 +1838,6 @@ gins(int as, Node *f, Node *t)
 	return p;
 }
 
-/*
- * generate code to compute n;
- * make a refer to result.
- */
-void
-naddr(Node *n, Addr *a, int canemitcode)
-{
-	Sym *s;
-
-	a->scale = 0;
-	a->reg = REG_NONE;
-	a->index = REG_NONE;
-	a->type = TYPE_NONE;
-	a->name = NAME_NONE;
-	a->gotype = nil;
-	a->node = N;
-	if(n == N)
-		return;
-
-	switch(n->op) {
-	default:
-		fatal("naddr: bad %O %D", n->op, a);
-		break;
-
-	case OREGISTER:
-		a->type = TYPE_REG;
-		a->reg = n->val.u.reg;
-		a->sym = nil;
-		break;
-
-	case OINDREG:
-		a->type = TYPE_MEM;
-		a->reg = n->val.u.reg;
-		a->sym = linksym(n->sym);
-		a->offset = n->xoffset;
-		break;
-
-	case OPARAM:
-		// n->left is PHEAP ONAME for stack parameter.
-		// compute address of actual parameter on stack.
-		a->etype = n->left->type->etype;
-		a->width = n->left->type->width;
-		a->offset = n->xoffset;
-		a->sym = linksym(n->left->sym);
-		a->type = TYPE_MEM;
-		a->name = NAME_PARAM;
-		a->node = n->left->orig;
-		break;
-
-	case OCLOSUREVAR:
-		if(!curfn->needctxt)
-			fatal("closurevar without needctxt");
-		a->type = TYPE_MEM;
-		a->reg = REG_DX;
-		a->offset = n->xoffset;
-		a->sym = nil;
-		break;
-
-	case OCFUNC:
-		naddr(n->left, a, canemitcode);
-		a->sym = linksym(n->left->sym);
-		break;
-
-	case ONAME:
-		a->etype = 0;
-		a->width = 0;
-		if(n->type != T) {
-			a->etype = simtype[n->type->etype];
-			dowidth(n->type);
-			a->width = n->type->width;
-		}
-		a->offset = n->xoffset;
-		s = n->sym;
-		a->node = n->orig;
-		//if(a->node >= (Node*)&n)
-		//	fatal("stack node");
-		if(s == S)
-			s = lookup(".noname");
-		if(n->method) {
-			if(n->type != T)
-			if(n->type->sym != S)
-			if(n->type->sym->pkg != nil)
-				s = pkglookup(s->name, n->type->sym->pkg);
-		}
-
-		switch(n->class) {
-		default:
-			fatal("naddr: ONAME class %S %d\n", n->sym, n->class);
-		case PEXTERN:
-			a->type = TYPE_MEM;
-			a->name = NAME_EXTERN;
-			break;
-		case PAUTO:
-			a->type = TYPE_MEM;
-			a->name = NAME_AUTO;
-			break;
-		case PPARAM:
-		case PPARAMOUT:
-			a->type = TYPE_MEM;
-			a->name = NAME_PARAM;
-			break;
-		case PFUNC:
-			a->type = TYPE_ADDR;
-			a->name = NAME_EXTERN;
-			s = funcsym(s);
-			break;
-		}
-		a->sym = linksym(s);
-		break;
-
-	case OLITERAL:
-		switch(n->val.ctype) {
-		default:
-			fatal("naddr: const %lT", n->type);
-			break;
-		case CTFLT:
-			a->type = TYPE_FCONST;
-			a->u.dval = mpgetflt(n->val.u.fval);
-			break;
-		case CTINT:
-		case CTRUNE:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = mpgetfix(n->val.u.xval);
-			break;
-		case CTSTR:
-			datagostring(n->val.u.sval, a);
-			break;
-		case CTBOOL:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = n->val.u.bval;
-			break;
-		case CTNIL:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = 0;
-			break;
-		}
-		break;
-
-	case OADDR:
-		naddr(n->left, a, canemitcode);
-		if(a->type != TYPE_MEM)
-			fatal("naddr: OADDR %D", a);
-		a->type = TYPE_ADDR;
-		break;
-	
-	case OITAB:
-		// itable of interface value
-		naddr(n->left, a, canemitcode);
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// len(nil)
-		a->etype = tptr;
-		a->width = widthptr;
-		break;
-
-	case OSPTR:
-		// pointer in a string or slice
-		naddr(n->left, a, canemitcode);
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// ptr(nil)
-		a->etype = simtype[tptr];
-		a->offset += Array_array;
-		a->width = widthptr;
-		break;
-
-	case OLEN:
-		// len of string or slice
-		naddr(n->left, a, canemitcode);
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// len(nil)
-		a->etype = TUINT32;
-		a->offset += Array_nel;
-		a->width = 4;
-		break;
-
-	case OCAP:
-		// cap of string or slice
-		naddr(n->left, a, canemitcode);
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// cap(nil)
-		a->etype = TUINT32;
-		a->offset += Array_cap;
-		a->width = 4;
-		break;
-
-//	case OADD:
-//		if(n->right->op == OLITERAL) {
-//			v = n->right->vconst;
-//			naddr(n->left, a, canemitcode);
-//		} else
-//		if(n->left->op == OLITERAL) {
-//			v = n->left->vconst;
-//			naddr(n->right, a, canemitcode);
-//		} else
-//			goto bad;
-//		a->offset += v;
-//		break;
-
-	}
-}
-
 int
 dotaddable(Node *n, Node *n1)
 {
@@ -2398,7 +1868,7 @@ sudoaddable(int as, Node *n, Addr *a)
 {
 	USED(as);
 	USED(n);
-	USED(a);
 
+	memset(a, 0, sizeof *a);
 	return 0;
 }
diff --git a/src/cmd/8g/opt.h b/src/cmd/8g/opt.h
deleted file mode 100644
index 8378d5d456..0000000000
--- a/src/cmd/8g/opt.h
+++ /dev/null
@@ -1,192 +0,0 @@
-// Derived from Inferno utils/6c/gc.h
-// http://code.google.com/p/inferno-os/source/browse/utils/6c/gc.h
-//
-//	Copyright © 1994-1999 Lucent Technologies Inc.  All rights reserved.
-//	Portions Copyright © 1995-1997 C H Forsyth (forsyth@terzarima.net)
-//	Portions Copyright © 1997-1999 Vita Nuova Limited
-//	Portions Copyright © 2000-2007 Vita Nuova Holdings Limited (www.vitanuova.com)
-//	Portions Copyright © 2004,2006 Bruce Ellis
-//	Portions Copyright © 2005-2007 C H Forsyth (forsyth@terzarima.net)
-//	Revisions Copyright © 2000-2007 Lucent Technologies Inc. and others
-//	Portions Copyright © 2009 The Go Authors.  All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-//
-// The above copyright notice and this permission notice shall be included in
-// all copies or substantial portions of the Software.
-//
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
-// THE SOFTWARE.
-
-
-#define	Z	N
-#define	Adr	Addr
-
-#define	BLOAD(r)	band(bnot(r->refbehind), r->refahead)
-#define	BSTORE(r)	band(bnot(r->calbehind), r->calahead)
-#define	LOAD(r)		(~r->refbehind.b[z] & r->refahead.b[z])
-#define	STORE(r)	(~r->calbehind.b[z] & r->calahead.b[z])
-
-#define	CLOAD	5
-#define	CREF	5
-#define	CINF	1000
-#define	LOOP	3
-
-typedef	struct	Reg	Reg;
-typedef	struct	Rgn	Rgn;
-
-/*c2go
-extern Node *Z;
-enum
-{
-	CLOAD = 5,
-	CREF = 5,
-	CINF = 1000,
-	LOOP = 3,
-};
-
-uint32 BLOAD(Reg*);
-uint32 BSTORE(Reg*);
-uint64 LOAD(Reg*);
-uint64 STORE(Reg*);
-*/
-
-// A Reg is a wrapper around a single Prog (one instruction) that holds
-// register optimization information while the optimizer runs.
-// r->prog is the instruction.
-// r->prog->opt points back to r.
-struct	Reg
-{
-	Flow	f;
-
-	Bits	set;  		// regopt variables written by this instruction.
-	Bits	use1; 		// regopt variables read by prog->from.
-	Bits	use2; 		// regopt variables read by prog->to.
-
-	// refahead/refbehind are the regopt variables whose current
-	// value may be used in the following/preceding instructions
-	// up to a CALL (or the value is clobbered).
-	Bits	refbehind;
-	Bits	refahead;
-	// calahead/calbehind are similar, but for variables in
-	// instructions that are reachable after hitting at least one
-	// CALL.
-	Bits	calbehind;
-	Bits	calahead;
-	Bits	regdiff;
-	Bits	act;
-
-	int32	regu;		// register used bitmap
-	int32	rpo;		// reverse post ordering
-	int32	active;
-
-	uint16	loop;		// x5 for every loop
-	uchar	refset;		// diagnostic generated
-
-	Reg*	p1;     	// predecessors of this instruction: p1,
-	Reg*	p2;     	// and then p2 linked though p2link.
-	Reg*	p2link;
-	Reg*	s1;     	// successors of this instruction (at most two: s1 and s2).
-	Reg*	s2;
-	Reg*	link;   	// next instruction in function code
-	Prog*	prog;   	// actual instruction
-};
-#define	R	((Reg*)0)
-/*c2go extern Reg *R; */
-
-#define	NRGN	600
-/*c2go enum { NRGN = 600 }; */
-
-// A Rgn represents a single regopt variable over a region of code
-// where a register could potentially be dedicated to that variable.
-// The code encompassed by a Rgn is defined by the flow graph,
-// starting at enter, flood-filling forward while varno is refahead
-// and backward while varno is refbehind, and following branches.  A
-// single variable may be represented by multiple disjoint Rgns and
-// each Rgn may choose a different register for that variable.
-// Registers are allocated to regions greedily in order of descending
-// cost.
-struct	Rgn
-{
-	Reg*	enter;
-	short	cost;
-	short	varno;
-	short	regno;
-};
-
-EXTERN	int32	exregoffset;		// not set
-EXTERN	int32	exfregoffset;		// not set
-EXTERN	Reg	zreg;
-EXTERN	Reg*	freer;
-EXTERN	Reg**	rpo2r;
-EXTERN	Rgn	region[NRGN];
-EXTERN	Rgn*	rgp;
-EXTERN	int	nregion;
-EXTERN	int	nvar;
-EXTERN	int32	regbits;
-EXTERN	int32	exregbits;
-EXTERN	Bits	externs;
-EXTERN	Bits	params;
-EXTERN	Bits	consts;
-EXTERN	Bits	addrs;
-EXTERN	Bits	ivar;
-EXTERN	Bits	ovar;
-EXTERN	int	change;
-EXTERN	int32	maxnr;
-EXTERN	int32*	idom;
-
-EXTERN	struct
-{
-	int32	ncvtreg;
-	int32	nspill;
-	int32	nreload;
-	int32	ndelmov;
-	int32	nvar;
-	int32	naddr;
-} ostats;
-
-/*
- * reg.c
- */
-Reg*	rega(void);
-int	rcmp(const void*, const void*);
-void	regopt(Prog*);
-void	addmove(Reg*, int, int, int);
-Bits	mkvar(Reg*, Adr*);
-void	prop(Reg*, Bits, Bits);
-void	loopit(Reg*, int32);
-void	synch(Reg*, Bits);
-uint32	allreg(uint32, Rgn*);
-void	paint1(Reg*, int);
-uint32	paint2(Reg*, int, int);
-void	paint3(Reg*, int, uint32, int);
-void	addreg(Adr*, int);
-void	dumpone(Flow*, int);
-void	dumpit(char*, Flow*, int);
-
-/*
- * peep.c
- */
-void	peep(Prog*);
-void	excise(Flow*);
-int	copyu(Prog*, Adr*, Adr*);
-
-uint32	RtoB(int);
-uint32	FtoB(int);
-int	BtoR(uint32);
-int	BtoF(uint32);
-
-/*
- * prog.c
- */
-void proginfo(ProgInfo*, Prog*);
diff --git a/src/cmd/8g/peep.c b/src/cmd/8g/peep.c
index 6c0865a7e8..712c8fe11b 100644
--- a/src/cmd/8g/peep.c
+++ b/src/cmd/8g/peep.c
@@ -31,10 +31,11 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
 enum {
 	REGEXT = 0,
+	exregoffset = REG_DI,
 };
 
 static void	conprop(Flow *r);
@@ -45,6 +46,7 @@ static int	copy1(Adr*, Adr*, Flow*, int);
 static int	copyas(Adr*, Adr*);
 static int	copyau(Adr*, Adr*);
 static int	copysub(Adr*, Adr*, Adr*, int);
+static int	copyu(Prog*, Adr*, Adr*);
 
 static uint32	gactive;
 
@@ -92,7 +94,7 @@ peep(Prog *firstp)
 	Prog *p, *p1;
 	int t;
 
-	g = flowstart(firstp, sizeof(Flow));
+	g = flowstart(firstp, 0);
 	if(g == nil)
 		return;
 	gactive = 0;
@@ -535,7 +537,7 @@ copy1(Adr *v1, Adr *v2, Flow *r, int f)
  * 4 if set and used
  * 0 otherwise (not touched)
  */
-int
+static int
 copyu(Prog *p, Adr *v, Adr *s)
 {
 	ProgInfo info;
@@ -559,7 +561,7 @@ copyu(Prog *p, Adr *v, Adr *s)
 	case ACALL:
 		if(REGEXT && v->type == TYPE_REG && v->reg <= REGEXT && v->reg > exregoffset)
 			return 2;
-		if(REGARG >= 0 && v->type == TYPE_REG && v->reg == (uchar)REGARG)
+		if(REGARG >= 0 && v->type == TYPE_REG && v->reg == REGARG)
 			return 2;
 		if(v->type == p->from.type && v->reg == p->from.reg)
 			return 2;
@@ -574,7 +576,7 @@ copyu(Prog *p, Adr *v, Adr *s)
 		return 3;
 
 	case ATEXT:
-		if(REGARG >= 0 && v->type == TYPE_REG && v->reg == (uchar)REGARG)
+		if(REGARG >= 0 && v->type == TYPE_REG && v->reg == REGARG)
 			return 3;
 		return 0;
 	}
diff --git a/src/cmd/8g/prog.c b/src/cmd/8g/prog.c
index 8a7371b5c4..e77a026a93 100644
--- a/src/cmd/8g/prog.c
+++ b/src/cmd/8g/prog.c
@@ -5,7 +5,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
 // Matches real RtoB but can be used in global initializer.
 #define RtoB(r) (1<<((r)-REG_AX))
diff --git a/src/cmd/8g/reg.c b/src/cmd/8g/reg.c
index 7d2de53549..0470bdf7b5 100644
--- a/src/cmd/8g/reg.c
+++ b/src/cmd/8g/reg.c
@@ -31,523 +31,34 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
-#define	NREGVAR	16	/* 8 integer + 8 floating */
-#define	REGBITS	((uint64)0xffffull)
-/*c2go enum {
-	NREGVAR = 16,
-	REGBITS = (1<<NREGVAR) - 1,
+enum {
+	NREGVAR = 16,	/* 8 integer + 8 floating */
 };
-*/
-
-static	Reg*	firstr;
-static	int	first	= 1;
-
-int
-rcmp(const void *a1, const void *a2)
-{
-	Rgn *p1, *p2;
-	int c1, c2;
-
-	p1 = (Rgn*)a1;
-	p2 = (Rgn*)a2;
-	c1 = p2->cost;
-	c2 = p1->cost;
-	if(c1 -= c2)
-		return c1;
-	return p2->varno - p1->varno;
-}
-
-static void
-setaddrs(Bits bit)
-{
-	int i, n;
-	Var *v;
-	Node *node;
-
-	while(bany(&bit)) {
-		// convert each bit to a variable
-		i = bnum(bit);
-		node = var[i].node;
-		n = var[i].name;
-		biclr(&bit, i);
-
-		// disable all pieces of that variable
-		for(i=0; i<nvar; i++) {
-			v = var+i;
-			if(v->node == node && v->name == n)
-				v->addr = 2;
-		}
-	}
-}
 
 static char* regname[] = {
 	".ax", ".cx", ".dx", ".bx", ".sp", ".bp", ".si", ".di",
 	".x0", ".x1", ".x2", ".x3", ".x4", ".x5", ".x6", ".x7",
 };
 
-static Node* regnodes[NREGVAR];
-
-static void walkvardef(Node *n, Reg *r, int active);
-
-void
-regopt(Prog *firstp)
+char**
+regnames(int *n)
 {
-	Reg *r, *r1;
-	Prog *p;
-	Graph *g;
-	ProgInfo info;
-	int i, z, active;
-	uint32 vreg;
-	Bits bit;
-
-	if(first) {
-		fmtinstall('Q', Qconv);
-		exregoffset = REG_DI;	// no externals
-		first = 0;
-	}
-
-	mergetemp(firstp);
-
-	/*
-	 * control flow is more complicated in generated go code
-	 * than in generated c code.  define pseudo-variables for
-	 * registers, so we have complete register usage information.
-	 */
-	nvar = NREGVAR;
-	memset(var, 0, NREGVAR*sizeof var[0]);
-	for(i=0; i<NREGVAR; i++) {
-		if(regnodes[i] == N)
-			regnodes[i] = newname(lookup(regname[i]));
-		var[i].node = regnodes[i];
-	}
-
-	regbits = RtoB(REG_SP);
-	for(z=0; z<BITS; z++) {
-		externs.b[z] = 0;
-		params.b[z] = 0;
-		consts.b[z] = 0;
-		addrs.b[z] = 0;
-		ivar.b[z] = 0;
-		ovar.b[z] = 0;
-	}
-
-	/*
-	 * pass 1
-	 * build aux data structure
-	 * allocate pcs
-	 * find use and set of variables
-	 */
-	g = flowstart(firstp, sizeof(Reg));
-	if(g == nil) {
-		for(i=0; i<nvar; i++)
-			var[i].node->opt = nil;
-		return;
-	}
-
-	firstr = (Reg*)g->start;
-
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		p = r->f.prog;
-		if(p->as == AVARDEF || p->as == AVARKILL)
-			continue;
-		proginfo(&info, p);
-
-		// Avoid making variables for direct-called functions.
-		if(p->as == ACALL && p->to.type == TYPE_MEM && p->to.name == NAME_EXTERN)
-			continue;
-
-		r->use1.b[0] |= info.reguse | info.regindex;
-		r->set.b[0] |= info.regset;
-
-		bit = mkvar(r, &p->from);
-		if(bany(&bit)) {
-			if(info.flags & LeftAddr)
-				setaddrs(bit);
-			if(info.flags & LeftRead)
-				for(z=0; z<BITS; z++)
-					r->use1.b[z] |= bit.b[z];
-			if(info.flags & LeftWrite)
-				for(z=0; z<BITS; z++)
-					r->set.b[z] |= bit.b[z];
-		}
-
-		bit = mkvar(r, &p->to);
-		if(bany(&bit)) {	
-			if(info.flags & RightAddr)
-				setaddrs(bit);
-			if(info.flags & RightRead)
-				for(z=0; z<BITS; z++)
-					r->use2.b[z] |= bit.b[z];
-			if(info.flags & RightWrite)
-				for(z=0; z<BITS; z++)
-					r->set.b[z] |= bit.b[z];
-		}
-	}
-	if(firstr == R)
-		return;
-
-	for(i=0; i<nvar; i++) {
-		Var *v = var+i;
-		if(v->addr) {
-			bit = blsh(i);
-			for(z=0; z<BITS; z++)
-				addrs.b[z] |= bit.b[z];
-		}
-
-		if(debug['R'] && debug['v'])
-			print("bit=%2d addr=%d et=%-6E w=%-2d s=%N + %lld\n",
-				i, v->addr, v->etype, v->width, v->node, v->offset);
-	}
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass1", &firstr->f, 1);
-
-	/*
-	 * pass 2
-	 * find looping structure
-	 */
-	flowrpo(g);
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass2", &firstr->f, 1);
-
-	/*
-	 * pass 2.5
-	 * iterate propagating fat vardef covering forward
-	 * r->act records vars with a VARDEF since the last CALL.
-	 * (r->act will be reused in pass 5 for something else,
-	 * but we'll be done with it by then.)
-	 */
-	active = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		r->f.active = 0;
-		r->act = zbits;
-	}
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		p = r->f.prog;
-		if(p->as == AVARDEF && isfat(((Node*)(p->to.node))->type) && ((Node*)(p->to.node))->opt != nil) {
-			active++;
-			walkvardef(p->to.node, r, active);
-		}
-	}
-
-	/*
-	 * pass 3
-	 * iterate propagating usage
-	 * 	back until flow graph is complete
-	 */
-loop1:
-	change = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->f.active = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		if(r->f.prog->as == ARET)
-			prop(r, zbits, zbits);
-loop11:
-	/* pick up unreachable code */
-	i = 0;
-	for(r = firstr; r != R; r = r1) {
-		r1 = (Reg*)r->f.link;
-		if(r1 && r1->f.active && !r->f.active) {
-			prop(r, zbits, zbits);
-			i = 1;
-		}
-	}
-	if(i)
-		goto loop11;
-	if(change)
-		goto loop1;
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass3", &firstr->f, 1);
-
-	/*
-	 * pass 4
-	 * iterate propagating register/variable synchrony
-	 * 	forward until graph is complete
-	 */
-loop2:
-	change = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->f.active = 0;
-	synch(firstr, zbits);
-	if(change)
-		goto loop2;
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass4", &firstr->f, 1);
-
-	/*
-	 * pass 4.5
-	 * move register pseudo-variables into regu.
-	 */
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		r->regu = (r->refbehind.b[0] | r->set.b[0]) & REGBITS;
-
-		r->set.b[0] &= ~REGBITS;
-		r->use1.b[0] &= ~REGBITS;
-		r->use2.b[0] &= ~REGBITS;
-		r->refbehind.b[0] &= ~REGBITS;
-		r->refahead.b[0] &= ~REGBITS;
-		r->calbehind.b[0] &= ~REGBITS;
-		r->calahead.b[0] &= ~REGBITS;
-		r->regdiff.b[0] &= ~REGBITS;
-		r->act.b[0] &= ~REGBITS;
-	}
-
-	/*
-	 * pass 5
-	 * isolate regions
-	 * calculate costs (paint1)
-	 */
-	r = firstr;
-	if(r) {
-		for(z=0; z<BITS; z++)
-			bit.b[z] = (r->refahead.b[z] | r->calahead.b[z]) &
-			  ~(externs.b[z] | params.b[z] | addrs.b[z] | consts.b[z]);
-		if(bany(&bit) && !r->f.refset) {
-			// should never happen - all variables are preset
-			if(debug['w'])
-				print("%L: used and not set: %Q\n", r->f.prog->lineno, bit);
-			r->f.refset = 1;
-		}
-	}
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->act = zbits;
-	rgp = region;
-	nregion = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		for(z=0; z<BITS; z++)
-			bit.b[z] = r->set.b[z] &
-			  ~(r->refahead.b[z] | r->calahead.b[z] | addrs.b[z]);
-		if(bany(&bit) && !r->f.refset) {
-			if(debug['w'])
-				print("%L: set and not used: %Q\n", r->f.prog->lineno, bit);
-			r->f.refset = 1;
-			excise(&r->f);
-		}
-		for(z=0; z<BITS; z++)
-			bit.b[z] = LOAD(r) & ~(r->act.b[z] | addrs.b[z]);
-		while(bany(&bit)) {
-			i = bnum(bit);
-			rgp->enter = r;
-			rgp->varno = i;
-			change = 0;
-			paint1(r, i);
-			biclr(&bit, i);
-			if(change <= 0)
-				continue;
-			rgp->cost = change;
-			nregion++;
-			if(nregion >= NRGN) {
-				if(debug['R'] && debug['v'])
-					print("too many regions\n");
-				goto brk;
-			}
-			rgp++;
-		}
-	}
-brk:
-	qsort(region, nregion, sizeof(region[0]), rcmp);
-
-	/*
-	 * pass 6
-	 * determine used registers (paint2)
-	 * replace code (paint3)
-	 */
-	rgp = region;
-	if(debug['R'] && debug['v'])
-		print("\nregisterizing\n");
-	for(i=0; i<nregion; i++) {
-		if(debug['R'] && debug['v'])
-			print("region %d: cost %d varno %d enter %lld\n", i, rgp->cost, rgp->varno, rgp->enter->f.prog->pc);
-		bit = blsh(rgp->varno);
-		vreg = paint2(rgp->enter, rgp->varno, 0);
-		vreg = allreg(vreg, rgp);
-		if(rgp->regno != 0)
-			paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
-		rgp++;
-	}
-
-	/*
-	 * free aux structures. peep allocates new ones.
-	 */
-	for(i=0; i<nvar; i++)
-		var[i].node->opt = nil;
-	flowend(g);
-	firstr = R;
-
-	if(debug['R'] && debug['v']) {
-		// Rebuild flow graph, since we inserted instructions
-		g = flowstart(firstp, sizeof(Reg));
-		firstr = (Reg*)g->start;
-		dumpit("pass6", &firstr->f, 1);
-		flowend(g);
-		firstr = R;
-	}
-
-	/*
-	 * pass 7
-	 * peep-hole on basic block
-	 */
-	if(!debug['R'] || debug['P'])
-		peep(firstp);
-
-	/*
-	 * eliminate nops
-	 */
-	for(p=firstp; p!=P; p=p->link) {
-		while(p->link != P && p->link->as == ANOP)
-			p->link = p->link->link;
-		if(p->to.type == TYPE_BRANCH)
-			while(p->to.u.branch != P && p->to.u.branch->as == ANOP)
-				p->to.u.branch = p->to.u.branch->link;
-	}
-
-	if(!use_sse)
-	for(p=firstp; p!=P; p=p->link) {
-		if(p->from.reg >= REG_X0 && p->from.reg <= REG_X7)
-			fatal("invalid use of %R with GO386=387: %P", p->from.reg, p);
-		if(p->to.reg >= REG_X0 && p->to.reg <= REG_X7)
-			fatal("invalid use of %R with GO386=387: %P", p->to.reg, p);
-	}
-
-	if(debug['R']) {
-		if(ostats.ncvtreg ||
-		   ostats.nspill ||
-		   ostats.nreload ||
-		   ostats.ndelmov ||
-		   ostats.nvar ||
-		   ostats.naddr ||
-		   0)
-			print("\nstats\n");
-
-		if(ostats.ncvtreg)
-			print("	%4d cvtreg\n", ostats.ncvtreg);
-		if(ostats.nspill)
-			print("	%4d spill\n", ostats.nspill);
-		if(ostats.nreload)
-			print("	%4d reload\n", ostats.nreload);
-		if(ostats.ndelmov)
-			print("	%4d delmov\n", ostats.ndelmov);
-		if(ostats.nvar)
-			print("	%4d var\n", ostats.nvar);
-		if(ostats.naddr)
-			print("	%4d addr\n", ostats.naddr);
-
-		memset(&ostats, 0, sizeof(ostats));
-	}
+	*n = NREGVAR;
+	return regname;
 }
 
-static void
-walkvardef(Node *n, Reg *r, int active)
+uint64
+excludedregs(void)
 {
-	Reg *r1, *r2;
-	int bn;
-	Var *v;
-	
-	for(r1=r; r1!=R; r1=(Reg*)r1->f.s1) {
-		if(r1->f.active == active)
-			break;
-		r1->f.active = active;
-		if(r1->f.prog->as == AVARKILL && r1->f.prog->to.node == n)
-			break;
-		for(v=n->opt; v!=nil; v=v->nextinnode) {
-			bn = v - var;
-			biset(&r1->act, bn);
-		}
-		if(r1->f.prog->as == ACALL)
-			break;
-	}
-
-	for(r2=r; r2!=r1; r2=(Reg*)r2->f.s1)
-		if(r2->f.s2 != nil)
-			walkvardef(n, (Reg*)r2->f.s2, active);
+	return RtoB(REG_SP);
 }
 
-/*
- * add mov b,rn
- * just after r
- */
-void
-addmove(Reg *r, int bn, int rn, int f)
-{
-	Prog *p, *p1;
-	Adr *a;
-	Var *v;
-
-	p1 = mal(sizeof(*p1));
-	clearp(p1);
-	p1->pc = 9999;
-
-	p = r->f.prog;
-	p1->link = p->link;
-	p->link = p1;
-	p1->lineno = p->lineno;
-
-	v = var + bn;
-
-	a = &p1->to;
-	a->offset = v->offset;
-	a->etype = v->etype;
-	a->type = TYPE_MEM;
-	a->name = v->name;
-	a->node = v->node;
-	a->sym = linksym(v->node->sym);
-
-	// need to clean this up with wptr and
-	// some of the defaults
-	p1->as = AMOVL;
-	switch(v->etype) {
-	default:
-		fatal("unknown type %E", v->etype);
-	case TINT8:
-	case TUINT8:
-	case TBOOL:
-		p1->as = AMOVB;
-		break;
-	case TINT16:
-	case TUINT16:
-		p1->as = AMOVW;
-		break;
-	case TFLOAT32:
-		p1->as = AMOVSS;
-		break;
-	case TFLOAT64:
-		p1->as = AMOVSD;
-		break;
-	case TINT:
-	case TUINT:
-	case TINT32:
-	case TUINT32:
-	case TPTR32:
-		break;
-	}
-
-	p1->from.type = TYPE_REG;
-	p1->from.reg = rn;
-	p1->from.name = 0;
-	if(!f) {
-		p1->from = *a;
-		*a = zprog.from;
-		a->type = TYPE_REG;
-		a->reg = rn;
-		if(v->etype == TUINT8)
-			p1->as = AMOVB;
-		if(v->etype == TUINT16)
-			p1->as = AMOVW;
-	}
-	if(debug['R'] && debug['v'])
-		print("%P ===add=== %P\n", p, p1);
-	ostats.nspill++;
-}
-
-uint32
+uint64
 doregbits(int r)
 {
-	uint32 b;
+	uint64 b;
 
 	b = 0;
 	if(r >= REG_AX && r <= REG_DI)
@@ -564,605 +75,17 @@ doregbits(int r)
 	return b;
 }
 
-static int
-overlap(int32 o1, int w1, int32 o2, int w2)
-{
-	int32 t1, t2;
-
-	t1 = o1+w1;
-	t2 = o2+w2;
-
-	if(!(t1 > o2 && t2 > o1))
-		return 0;
-
-	return 1;
-}
-
-Bits
-mkvar(Reg *r, Adr *a)
-{
-	Var *v;
-	int i, n, et, z, w, flag, regu;
-	int32 o;
-	Bits bit;
-	Node *node;
-
-	/*
-	 * mark registers used
-	 */
-	if(a->type == TYPE_NONE)
-		goto none;
-
-	if(r != R)
-		r->use1.b[0] |= doregbits(a->index);
-
-	switch(a->type) {
-	default:
-		regu = doregbits(a->reg);
-		if(regu == 0)
-			goto none;
-		bit = zbits;
-		bit.b[0] = regu;
-		return bit;
-
-	case TYPE_ADDR:
-		a->type = TYPE_MEM;
-		bit = mkvar(r, a);
-		setaddrs(bit);
-		a->type = TYPE_ADDR;
-		ostats.naddr++;
-		goto none;
-
-	case TYPE_MEM:
-		switch(a->name) {
-		default:
-			goto none;
-		case NAME_EXTERN:
-		case NAME_STATIC:
-		case NAME_PARAM:
-		case NAME_AUTO:
-			n = a->name;
-			break;
-		}
-	}
-
-	node = a->node;
-	if(node == N || node->op != ONAME || node->orig == N)
-		goto none;
-	node = node->orig;
-	if(node->orig != node)
-		fatal("%D: bad node", a);
-	if(node->sym == S || node->sym->name[0] == '.')
-		goto none;
-	et = a->etype;
-	o = a->offset;
-	w = a->width;
-	if(w < 0)
-		fatal("bad width %d for %D", w, a);
-
-	flag = 0;
-	for(i=0; i<nvar; i++) {
-		v = var+i;
-		if(v->node == node && v->name == n) {
-			if(v->offset == o)
-			if(v->etype == et)
-			if(v->width == w)
-				return blsh(i);
-
-			// if they overlap, disable both
-			if(overlap(v->offset, v->width, o, w)) {
-				if(debug['R'])
-					print("disable %s\n", node->sym->name);
-				v->addr = 1;
-				flag = 1;
-			}
-		}
-	}
-
-	switch(et) {
-	case 0:
-	case TFUNC:
-		goto none;
-	}
-
-	if(nvar >= NVAR) {
-		if(debug['w'] > 1 && node != N)
-			fatal("variable not optimized: %D", a);
-		
-		// If we're not tracking a word in a variable, mark the rest as
-		// having its address taken, so that we keep the whole thing
-		// live at all calls. otherwise we might optimize away part of
-		// a variable but not all of it.
-		for(i=0; i<nvar; i++) {
-			v = var+i;
-			if(v->node == node)
-				v->addr = 1;
-		}
-		goto none;
-	}
-
-	i = nvar;
-	nvar++;
-	v = var+i;
-	v->offset = o;
-	v->name = n;
-	v->etype = et;
-	v->width = w;
-	v->addr = flag;		// funny punning
-	v->node = node;
-	
-	// node->opt is the head of a linked list
-	// of Vars within the given Node, so that
-	// we can start at a Var and find all the other
-	// Vars in the same Go variable.
-	v->nextinnode = node->opt;
-	node->opt = v;
-
-	bit = blsh(i);
-	if(n == NAME_EXTERN || n == NAME_STATIC)
-		for(z=0; z<BITS; z++)
-			externs.b[z] |= bit.b[z];
-	if(n == NAME_PARAM)
-		for(z=0; z<BITS; z++)
-			params.b[z] |= bit.b[z];
-		
-	if(node->class == PPARAM)
-		for(z=0; z<BITS; z++)
-			ivar.b[z] |= bit.b[z];
-	if(node->class == PPARAMOUT)
-		for(z=0; z<BITS; z++)
-			ovar.b[z] |= bit.b[z];
-
-	// Treat values with their address taken as live at calls,
-	// because the garbage collector's liveness analysis in ../gc/plive.c does.
-	// These must be consistent or else we will elide stores and the garbage
-	// collector will see uninitialized data.
-	// The typical case where our own analysis is out of sync is when the
-	// node appears to have its address taken but that code doesn't actually
-	// get generated and therefore doesn't show up as an address being
-	// taken when we analyze the instruction stream.
-	// One instance of this case is when a closure uses the same name as
-	// an outer variable for one of its own variables declared with :=.
-	// The parser flags the outer variable as possibly shared, and therefore
-	// sets addrtaken, even though it ends up not being actually shared.
-	// If we were better about _ elision, _ = &x would suffice too.
-	// The broader := in a closure problem is mentioned in a comment in
-	// closure.c:/^typecheckclosure and dcl.c:/^oldname.
-	if(node->addrtaken)
-		v->addr = 1;
-
-	// Disable registerization for globals, because:
-	// (1) we might panic at any time and we want the recovery code
-	// to see the latest values (issue 1304).
-	// (2) we don't know what pointers might point at them and we want
-	// loads via those pointers to see updated values and vice versa (issue 7995).
-	//
-	// Disable registerization for results if using defer, because the deferred func
-	// might recover and return, causing the current values to be used.
-	if(node->class == PEXTERN || (hasdefer && node->class == PPARAMOUT))
-		v->addr = 1;
-
-	if(debug['R'])
-		print("bit=%2d et=%2E w=%d+%d %#N %D flag=%d\n", i, et, o, w, node, a, v->addr);
-	ostats.nvar++;
-
-	return bit;
-
-none:
-	return zbits;
-}
-
-void
-prop(Reg *r, Bits ref, Bits cal)
-{
-	Reg *r1, *r2;
-	int z, i, j;
-	Var *v, *v1;
-
-	for(r1 = r; r1 != R; r1 = (Reg*)r1->f.p1) {
-		for(z=0; z<BITS; z++) {
-			ref.b[z] |= r1->refahead.b[z];
-			if(ref.b[z] != r1->refahead.b[z]) {
-				r1->refahead.b[z] = ref.b[z];
-				change++;
-			}
-			cal.b[z] |= r1->calahead.b[z];
-			if(cal.b[z] != r1->calahead.b[z]) {
-				r1->calahead.b[z] = cal.b[z];
-				change++;
-			}
-		}
-		switch(r1->f.prog->as) {
-		case ACALL:
-			if(noreturn(r1->f.prog))
-				break;
-
-			// Mark all input variables (ivar) as used, because that's what the
-			// liveness bitmaps say. The liveness bitmaps say that so that a
-			// panic will not show stale values in the parameter dump.
-			// Mark variables with a recent VARDEF (r1->act) as used,
-			// so that the optimizer flushes initializations to memory,
-			// so that if a garbage collection happens during this CALL,
-			// the collector will see initialized memory. Again this is to
-			// match what the liveness bitmaps say.
-			for(z=0; z<BITS; z++) {
-				cal.b[z] |= ref.b[z] | externs.b[z] | ivar.b[z] | r1->act.b[z];
-				ref.b[z] = 0;
-			}
-			
-			// cal.b is the current approximation of what's live across the call.
-			// Every bit in cal.b is a single stack word. For each such word,
-			// find all the other tracked stack words in the same Go variable
-			// (struct/slice/string/interface) and mark them live too.
-			// This is necessary because the liveness analysis for the garbage
-			// collector works at variable granularity, not at word granularity.
-			// It is fundamental for slice/string/interface: the garbage collector
-			// needs the whole value, not just some of the words, in order to
-			// interpret the other bits correctly. Specifically, slice needs a consistent
-			// ptr and cap, string needs a consistent ptr and len, and interface
-			// needs a consistent type word and data word.
-			for(z=0; z<BITS; z++) {
-				if(cal.b[z] == 0)
-					continue;
-				for(i=0; i<64; i++) {
-					if(z*64+i >= nvar || ((cal.b[z]>>i)&1) == 0)
-						continue;
-					v = var+z*64+i;
-					if(v->node->opt == nil) // v represents fixed register, not Go variable
-						continue;
-
-					// v->node->opt is the head of a linked list of Vars
-					// corresponding to tracked words from the Go variable v->node.
-					// Walk the list and set all the bits.
-					// For a large struct this could end up being quadratic:
-					// after the first setting, the outer loop (for z, i) would see a 1 bit
-					// for all of the remaining words in the struct, and for each such
-					// word would go through and turn on all the bits again.
-					// To avoid the quadratic behavior, we only turn on the bits if
-					// v is the head of the list or if the head's bit is not yet turned on.
-					// This will set the bits at most twice, keeping the overall loop linear.
-					v1 = v->node->opt;
-					j = v1 - var;
-					if(v == v1 || !btest(&cal, j)) {
-						for(; v1 != nil; v1 = v1->nextinnode) {
-							j = v1 - var;
-							biset(&cal, j);
-						}
-					}
-				}
-			}
-			break;
-
-		case ATEXT:
-			for(z=0; z<BITS; z++) {
-				cal.b[z] = 0;
-				ref.b[z] = 0;
-			}
-			break;
-
-		case ARET:
-			for(z=0; z<BITS; z++) {
-				cal.b[z] = externs.b[z] | ovar.b[z];
-				ref.b[z] = 0;
-			}
-			break;
-		}
-		for(z=0; z<BITS; z++) {
-			ref.b[z] = (ref.b[z] & ~r1->set.b[z]) |
-				r1->use1.b[z] | r1->use2.b[z];
-			cal.b[z] &= ~(r1->set.b[z] | r1->use1.b[z] | r1->use2.b[z]);
-			r1->refbehind.b[z] = ref.b[z];
-			r1->calbehind.b[z] = cal.b[z];
-		}
-		if(r1->f.active)
-			break;
-		r1->f.active = 1;
-	}
-	for(; r != r1; r = (Reg*)r->f.p1)
-		for(r2 = (Reg*)r->f.p2; r2 != R; r2 = (Reg*)r2->f.p2link)
-			prop(r2, r->refbehind, r->calbehind);
-}
-
-void
-synch(Reg *r, Bits dif)
-{
-	Reg *r1;
-	int z;
-
-	for(r1 = r; r1 != R; r1 = (Reg*)r1->f.s1) {
-		for(z=0; z<BITS; z++) {
-			dif.b[z] = (dif.b[z] &
-				~(~r1->refbehind.b[z] & r1->refahead.b[z])) |
-					r1->set.b[z] | r1->regdiff.b[z];
-			if(dif.b[z] != r1->regdiff.b[z]) {
-				r1->regdiff.b[z] = dif.b[z];
-				change++;
-			}
-		}
-		if(r1->f.active)
-			break;
-		r1->f.active = 1;
-		for(z=0; z<BITS; z++)
-			dif.b[z] &= ~(~r1->calbehind.b[z] & r1->calahead.b[z]);
-		if((Reg*)r1->f.s2 != R)
-			synch((Reg*)r1->f.s2, dif);
-	}
-}
-
-uint32
-allreg(uint32 b, Rgn *r)
-{
-	Var *v;
-	int i;
-
-	v = var + r->varno;
-	r->regno = 0;
-	switch(v->etype) {
-
-	default:
-		fatal("unknown etype %d/%E", bitno(b), v->etype);
-		break;
-
-	case TINT8:
-	case TUINT8:
-	case TINT16:
-	case TUINT16:
-	case TINT32:
-	case TUINT32:
-	case TINT64:
-	case TINT:
-	case TUINT:
-	case TUINTPTR:
-	case TBOOL:
-	case TPTR32:
-		i = BtoR(~b);
-		if(i && r->cost > 0) {
-			r->regno = i;
-			return RtoB(i);
-		}
-		break;
-
-	case TFLOAT32:
-	case TFLOAT64:
-		if(!use_sse)
-			break;
-		i = BtoF(~b);
-		if(i && r->cost > 0) {
-			r->regno = i;
-			return FtoB(i);
-		}
-		break;
-	}
-	return 0;
-}
-
-void
-paint1(Reg *r, int bn)
-{
-	Reg *r1;
-	Prog *p;
-	int z;
-	uint64 bb, rbz;
-
-	z = bn/64;
-	bb = 1LL<<(bn%64);
-	if(r->act.b[z] & bb)
-		return;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(r1->act.b[z] & bb)
-			break;
-		r = r1;
-	}
-
-	rbz = ~(r->set.b[z]&~(r->use1.b[z]|r->use2.b[z]));
-	if(LOAD(r) & rbz & bb) {
-		change -= CLOAD * r->f.loop;
-	}
-	for(;;) {
-		r->act.b[z] |= bb;
-		p = r->f.prog;
-
-		if(r->f.prog->as != ANOP) { // don't give credit for NOPs
-			if(r->use1.b[z] & bb) {
-				change += CREF * r->f.loop;
-				if(p->as == AFMOVL || p->as == AFMOVW)
-					if(BtoR(bb) != REG_F0)
-						change = -CINF;
-			}
-			if((r->use2.b[z]|r->set.b[z]) & bb) {
-				change += CREF * r->f.loop;
-				if(p->as == AFMOVL || p->as == AFMOVW)
-					if(BtoR(bb) != REG_F0)
-						change = -CINF;
-			}
-		}
-
-		if(STORE(r) & r->regdiff.b[z] & bb) {
-			change -= CLOAD * r->f.loop;
-			if(p->as == AFMOVL || p->as == AFMOVW)
-				if(BtoR(bb) != REG_F0)
-					change = -CINF;
-		}
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					paint1(r1, bn);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				paint1(r1, bn);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(r->act.b[z] & bb)
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-}
-
-uint32
-paint2(Reg *r, int bn, int depth)
-{
-	Reg *r1;
-	int z;
-	uint64 bb, vreg;
-
-	z = bn/64;
-	bb = 1LL << (bn%64);
-	vreg = regbits;
-	if(!(r->act.b[z] & bb))
-		return vreg;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(!(r1->act.b[z] & bb))
-			break;
-		r = r1;
-	}
-	for(;;) {
-		if(debug['R'] && debug['v'])
-			print("  paint2 %d %P\n", depth, r->f.prog);
-
-		r->act.b[z] &= ~bb;
-
-		vreg |= r->regu;
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					vreg |= paint2(r1, bn, depth+1);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				vreg |= paint2(r1, bn, depth+1);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(!(r->act.b[z] & bb))
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-
-	return vreg;
-}
-
-void
-paint3(Reg *r, int bn, uint32 rb, int rn)
-{
-	Reg *r1;
-	Prog *p;
-	int z;
-	uint64 bb, rbz;
-
-	z = bn/64;
-	bb = 1LL << (bn%64);
-	if(r->act.b[z] & bb)
-		return;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(r1->act.b[z] & bb)
-			break;
-		r = r1;
-	}
-
-	rbz = ~(r->set.b[z] & ~(r->use1.b[z]|r->use2.b[z]));
-	if(LOAD(r) & rbz & bb)
-		addmove(r, bn, rn, 0);
-	for(;;) {
-		r->act.b[z] |= bb;
-		p = r->f.prog;
-
-		if(r->use1.b[z] & bb) {
-			if(debug['R'] && debug['v'])
-				print("%P", p);
-			addreg(&p->from, rn);
-			if(debug['R'] && debug['v'])
-				print(" ===change== %P\n", p);
-		}
-		if((r->use2.b[z]|r->set.b[z]) & bb) {
-			if(debug['R'] && debug['v'])
-				print("%P", p);
-			addreg(&p->to, rn);
-			if(debug['R'] && debug['v'])
-				print(" ===change== %P\n", p);
-		}
-
-		if(STORE(r) & r->regdiff.b[z] & bb)
-			addmove(r, bn, rn, 1);
-		r->regu |= rb;
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					paint3(r1, bn, rb, rn);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				paint3(r1, bn, rb, rn);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(r->act.b[z] & bb)
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-}
-
-void
-addreg(Adr *a, int rn)
-{
-	a->sym = nil;
-	a->node = nil;
-	a->offset = 0;
-	a->type = TYPE_REG;
-	a->reg = rn;
-	a->name = 0;
-
-	ostats.ncvtreg++;
-}
-
-uint32
+uint64
 RtoB(int r)
 {
 
 	if(r < REG_AX || r > REG_DI)
 		return 0;
-	return 1L << (r-REG_AX);
+	return 1ULL << (r-REG_AX);
 }
 
 int
-BtoR(uint32 b)
+BtoR(uint64 b)
 {
 
 	b &= 0xffL;
@@ -1171,93 +94,19 @@ BtoR(uint32 b)
 	return bitno(b) + REG_AX;
 }
 
-uint32
+uint64
 FtoB(int f)
 {
 	if(f < REG_X0 || f > REG_X7)
 		return 0;
-	return 1L << (f - REG_X0 + 8);
+	return 1ULL << (f - REG_X0 + 8);
 }
 
 int
-BtoF(uint32 b)
+BtoF(uint64 b)
 {
 	b &= 0xFF00L;
 	if(b == 0)
 		return 0;
 	return bitno(b) - 8 + REG_X0;
 }
-
-void
-dumpone(Flow *f, int isreg)
-{
-	int z;
-	Bits bit;
-	Reg *r;
-
-	print("%d:%P", f->loop, f->prog);
-	if(isreg) {
-		r = (Reg*)f;
-		for(z=0; z<BITS; z++)
-			bit.b[z] =
-				r->set.b[z] |
-				r->use1.b[z] |
-				r->use2.b[z] |
-				r->refbehind.b[z] |
-				r->refahead.b[z] |
-				r->calbehind.b[z] |
-				r->calahead.b[z] |
-				r->regdiff.b[z] |
-				r->act.b[z] |
-					0;
-		if(bany(&bit)) {
-			print("\t");
-			if(bany(&r->set))
-				print(" s:%Q", r->set);
-			if(bany(&r->use1))
-				print(" u1:%Q", r->use1);
-			if(bany(&r->use2))
-				print(" u2:%Q", r->use2);
-			if(bany(&r->refbehind))
-				print(" rb:%Q ", r->refbehind);
-			if(bany(&r->refahead))
-				print(" ra:%Q ", r->refahead);
-			if(bany(&r->calbehind))
-				print(" cb:%Q ", r->calbehind);
-			if(bany(&r->calahead))
-				print(" ca:%Q ", r->calahead);
-			if(bany(&r->regdiff))
-				print(" d:%Q ", r->regdiff);
-			if(bany(&r->act))
-				print(" a:%Q ", r->act);
-		}
-	}
-	print("\n");
-}
-
-void
-dumpit(char *str, Flow *r0, int isreg)
-{
-	Flow *r, *r1;
-
-	print("\n%s\n", str);
-	for(r = r0; r != nil; r = r->link) {
-		dumpone(r, isreg);
-		r1 = r->p2;
-		if(r1 != nil) {
-			print("	pred:");
-			for(; r1 != nil; r1 = r1->p2link)
-				print(" %.4ud", (int)r1->prog->pc);
-			print("\n");
-		}
-		// Print successors if it's not just the next one
-		if(r->s1 != r->link || r->s2 != nil) {
-			print("	succ:");
-			if(r->s1 != nil)
-				print(" %.4ud", (int)r->s1->prog->pc);
-			if(r->s2 != nil)
-				print(" %.4ud", (int)r->s2->prog->pc);
-			print("\n");
-		}
-	}
-}
diff --git a/src/cmd/8l/8.out.h b/src/cmd/8l/8.out.h
index 28d7dab7fb..273a8ad523 100644
--- a/src/cmd/8l/8.out.h
+++ b/src/cmd/8l/8.out.h
@@ -634,6 +634,7 @@ enum
 	FREGRET		= REG_F0,
 	REGSP		= REG_SP,
 	REGTMP		= REG_DI,
+	REGCTXT		= REG_DX,
 };
 
 /*
diff --git a/src/cmd/8l/asm.c b/src/cmd/8l/asm.c
index 44d1ecc035..2da1651c8b 100644
--- a/src/cmd/8l/asm.c
+++ b/src/cmd/8l/asm.c
@@ -37,13 +37,6 @@
 #include	"../ld/macho.h"
 #include	"../ld/pe.h"
 
-char linuxdynld[] = "/lib/ld-linux.so.2";
-char freebsddynld[] = "/usr/libexec/ld-elf.so.1";
-char openbsddynld[] = "/usr/libexec/ld.so";
-char netbsddynld[] = "/usr/libexec/ld.elf_so";
-char dragonflydynld[] = "/usr/libexec/ld-elf.so.2";
-char solarisdynld[] = "/lib/ld.so.1";
-
 static int
 needlib(char *name)
 {
@@ -64,8 +57,6 @@ needlib(char *name)
 	return 0;
 }
 
-int	nelfsym = 1;
-
 static void	addpltsym(Link*, LSym*);
 static void	addgotsym(Link*, LSym*);
 
@@ -141,7 +132,7 @@ adddynrel(LSym *s, Reloc *r)
 		}
 		addgotsym(ctxt, targ);
 		r->type = R_CONST;	// write r->add during relocsym
-		r->sym = S;
+		r->sym = nil;
 		r->add += targ->got;
 		return;
 	
@@ -218,7 +209,7 @@ adddynrel(LSym *s, Reloc *r)
 			addaddrplus(ctxt, rel, s, r->off);
 			adduint32(ctxt, rel, ELF32_R_INFO(targ->dynid, R_386_32));
 			r->type = R_CONST;	// write r->add during relocsym
-			r->sym = S;
+			r->sym = nil;
 			return;
 		}
 		if(HEADTYPE == Hdarwin && s->size == PtrSize && r->off == 0) {
@@ -256,7 +247,7 @@ elfreloc1(Reloc *r, vlong sectoff)
 {
 	int32 elfsym;
 
-	LPUT(sectoff);
+	thearch.lput(sectoff);
 
 	elfsym = r->xsym->elfsym;
 	switch(r->type) {
@@ -265,7 +256,7 @@ elfreloc1(Reloc *r, vlong sectoff)
 
 	case R_ADDR:
 		if(r->siz == 4)
-			LPUT(R_386_32 | elfsym<<8);
+			thearch.lput(R_386_32 | elfsym<<8);
 		else
 			return -1;
 		break;
@@ -273,7 +264,7 @@ elfreloc1(Reloc *r, vlong sectoff)
 	case R_CALL:
 	case R_PCREL:
 		if(r->siz == 4)
-			LPUT(R_386_PC32 | elfsym<<8);
+			thearch.lput(R_386_PC32 | elfsym<<8);
 		else
 			return -1;
 		break;
@@ -281,7 +272,7 @@ elfreloc1(Reloc *r, vlong sectoff)
 	case R_TLS_LE:
 	case R_TLS_IE:
 		if(r->siz == 4)
-			LPUT(R_386_TLS_LE | elfsym<<8);
+			thearch.lput(R_386_TLS_LE | elfsym<<8);
 		else
 			return -1;
 	}
@@ -342,8 +333,8 @@ machoreloc1(Reloc *r, vlong sectoff)
 		break;
 	}
 
-	LPUT(sectoff);
-	LPUT(v);
+	thearch.lput(sectoff);
+	thearch.lput(v);
 	return 0;
 }
 
@@ -715,29 +706,3 @@ asmb(void)
 	}
 	cflush();
 }
-
-void
-s8put(char *n)
-{
-	char name[8];
-	int i;
-
-	strncpy(name, n, sizeof(name));
-	for(i=0; i<sizeof(name); i++)
-		cput(name[i]);
-}
-
-int32
-rnd(int32 v, int32 r)
-{
-	int32 c;
-
-	if(r <= 0)
-		return v;
-	v += r - 1;
-	c = v % r;
-	if(c < 0)
-		c += r;
-	v -= c;
-	return v;
-}
diff --git a/src/cmd/8l/l.h b/src/cmd/8l/l.h
index 3c84d1b21f..e844bc4180 100644
--- a/src/cmd/8l/l.h
+++ b/src/cmd/8l/l.h
@@ -32,7 +32,6 @@
 #include	<libc.h>
 #include	<bio.h>
 #include	<link.h>
-#include	"8.out.h"
 
 #ifndef	EXTERN
 #define	EXTERN	extern
@@ -45,30 +44,10 @@ enum
 	IntSize = 4,
 	RegSize = 4,
 	MaxAlign = 32,	// max data alignment
-	FuncAlign = 16
-};
-
-#define	P		((Prog*)0)
-#define	S		((LSym*)0)
-
-enum
-{
+	FuncAlign = 16,
 	MINLC		= 1,
 };
 
-#pragma	varargck	type	"I"	uchar*
-
-EXTERN	LSym*	datap;
-EXTERN	int	debug[128];
-EXTERN	char	literal[32];
-EXTERN	Prog*	firstp;
-EXTERN	int32	lcsize;
-EXTERN	char*	rpath;
-EXTERN	int32	spsize;
-EXTERN	LSym*	symlist;
-EXTERN	int32	symsize;
-
-int	Iconv(Fmt *fp);
 void	adddynlib(char *lib);
 void	adddynrel(LSym *s, Reloc *r);
 void	adddynrela(LSym *rela, LSym *s, Reloc *r);
@@ -80,14 +59,6 @@ int	elfreloc1(Reloc *r, vlong sectoff);
 void	elfsetupplt(void);
 void	listinit(void);
 int	machoreloc1(Reloc *r, vlong sectoff);
-int32	rnd(int32 v, int32 r);
-void	s8put(char *n);
-char*	xsymname(LSym *s);
-
-/* Native is little-endian */
-#define	LPUT(a)	lputl(a)
-#define	WPUT(a)	wputl(a)
-#define	VPUT(a)	vputl(a)
 
 /* Used by ../ld/dwarf.c */
 enum
diff --git a/src/cmd/8l/list.c b/src/cmd/8l/list.c
index 0a75340603..4cd8f08d34 100644
--- a/src/cmd/8l/list.c
+++ b/src/cmd/8l/list.c
@@ -37,31 +37,4 @@ void
 listinit(void)
 {
 	listinit8();
-	fmtinstall('I', Iconv);
-}
-
-int
-Iconv(Fmt *fp)
-{
-	int i, n;
-	uchar *p;
-	char *s;
-	Fmt fmt;
-	
-	n = fp->prec;
-	fp->prec = 0;
-	if(!(fp->flags&FmtPrec) || n < 0)
-		return fmtstrcpy(fp, "%I");
-	fp->flags &= ~FmtPrec;
-	p = va_arg(fp->args, uchar*);
-
-	// format into temporary buffer and
-	// call fmtstrcpy to handle padding.
-	fmtstrinit(&fmt);
-	for(i=0; i<n; i++)
-		fmtprint(&fmt, "%.2ux", *p++);
-	s = fmtstrflush(&fmt);
-	fmtstrcpy(fp, s);
-	free(s);
-	return 0;
 }
diff --git a/src/cmd/8l/obj.c b/src/cmd/8l/obj.c
index 1b65c5eb9e..41900d6543 100644
--- a/src/cmd/8l/obj.c
+++ b/src/cmd/8l/obj.c
@@ -38,12 +38,50 @@
 #include	"../ld/pe.h"
 #include	<ar.h>
 
-char*	thestring 	= "386";
-LinkArch*	thelinkarch = &link386;
+void
+main(int argc, char **argv)
+{
+	linkarchinit();
+	ldmain(argc, argv);
+}
 
 void
 linkarchinit(void)
 {
+	thestring 	= "386";
+	thelinkarch = &link386;
+
+	thearch.thechar = thechar;
+	thearch.ptrsize = thelinkarch->ptrsize;
+	thearch.intsize = thelinkarch->ptrsize;
+	thearch.regsize = thelinkarch->regsize;
+	thearch.funcalign = FuncAlign;
+	thearch.maxalign = MaxAlign;
+	thearch.minlc = MINLC;
+	thearch.dwarfregsp = DWARFREGSP;
+
+	thearch.adddynlib = adddynlib;
+	thearch.adddynrel = adddynrel;
+	thearch.adddynsym = adddynsym;
+	thearch.archinit = archinit;
+	thearch.archreloc = archreloc;
+	thearch.archrelocvariant = archrelocvariant;
+	thearch.asmb = asmb;
+	thearch.elfreloc1 = elfreloc1;
+	thearch.elfsetupplt = elfsetupplt;
+	thearch.gentext = gentext;
+	thearch.listinit = listinit;
+	thearch.machoreloc1 = machoreloc1;
+	thearch.lput = lputl;
+	thearch.wput = wputl;
+	thearch.vput = vputl;
+
+	thearch.linuxdynld = "/lib/ld-linux.so.2";
+	thearch.freebsddynld = "/usr/libexec/ld-elf.so.1";
+	thearch.openbsddynld = "/usr/libexec/ld.so";
+	thearch.netbsddynld = "/usr/libexec/ld.elf_so";
+	thearch.dragonflydynld = "/usr/libexec/ld-elf.so.2";
+	thearch.solarisdynld = "/lib/ld.so.1";
 }
 
 void
diff --git a/src/cmd/9g/cgen.c b/src/cmd/9g/cgen.c
index 706658de0e..009ea1ed7a 100644
--- a/src/cmd/9g/cgen.c
+++ b/src/cmd/9g/cgen.c
@@ -805,7 +805,7 @@ ginsadd(int as, vlong off, Node *dst)
 void
 agen(Node *n, Node *res)
 {
-	Node *nl, *nr;
+	Node *nl;
 	Node n1, n2, n3;
 
 	if(debug['g']) {
@@ -848,8 +848,6 @@ agen(Node *n, Node *res)
 	}
 
 	nl = n->left;
-	nr = n->right;
-	USED(nr);
 
 	switch(n->op) {
 	default:
@@ -1109,27 +1107,20 @@ bgen(Node *n, int true, int likely, Prog *to)
 		goto ret;
 
 	case OANDAND:
-		if(!true)
-			goto caseor;
-
-	caseand:
-		p1 = gbranch(ABR, T, 0);
-		p2 = gbranch(ABR, T, 0);
-		patch(p1, pc);
-		bgen(n->left, !true, -likely, p2);
-		bgen(n->right, !true, -likely, p2);
-		p1 = gbranch(ABR, T, 0);
-		patch(p1, to);
-		patch(p2, pc);
-		goto ret;
-
 	case OOROR:
-		if(!true)
-			goto caseand;
-
-	caseor:
-		bgen(n->left, true, likely, to);
-		bgen(n->right, true, likely, to);
+		if((n->op == OANDAND) == true) {
+			p1 = gbranch(AJMP, T, 0);
+			p2 = gbranch(AJMP, T, 0);
+			patch(p1, pc);
+			bgen(n->left, !true, -likely, p2);
+			bgen(n->right, !true, -likely, p2);
+			p1 = gbranch(AJMP, T, 0);
+			patch(p1, to);
+			patch(p2, pc);
+		} else {
+			bgen(n->left, true, likely, to);
+			bgen(n->right, true, likely, to);
+		}
 		goto ret;
 
 	case OEQ:
@@ -1554,7 +1545,7 @@ cadable(Node *n)
 int
 componentgen(Node *nr, Node *nl)
 {
-	Node nodl, nodr;
+	Node nodl, nodr, tmp;
 	Type *t;
 	int freel, freer;
 	vlong fldcount;
@@ -1602,7 +1593,7 @@ componentgen(Node *nr, Node *nl)
 
 	nodl = *nl;
 	if(!cadable(nl)) {
-		if(nr == N || !cadable(nr))
+		if(nr != N && !cadable(nr))
 			goto no;
 		igen(nl, &nodl, N);
 		freel = 1;
@@ -1614,6 +1605,12 @@ componentgen(Node *nr, Node *nl)
 			igen(nr, &nodr, N);
 			freer = 1;
 		}
+	} else {
+		// When zeroing, prepare a register containing zero.
+		nodconst(&tmp, nl->type, 0);
+		regalloc(&nodr, types[TUINT], N);
+		gmove(&tmp, &nodr);
+		freer = 1;
 	}
 	
 	// nl and nr are 'cadable' which basically means they are names (variables) now.
@@ -1650,8 +1647,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_nel-Array_array;
@@ -1660,8 +1656,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_nel-Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_cap-Array_nel;
@@ -1670,8 +1665,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_cap-Array_nel;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		goto yes;
@@ -1685,8 +1679,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_nel-Array_array;
@@ -1695,8 +1688,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_nel-Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		goto yes;
@@ -1710,8 +1702,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		nodl.xoffset += Array_nel-Array_array;
@@ -1720,8 +1711,7 @@ componentgen(Node *nr, Node *nl)
 		if(nr != N) {
 			nodr.xoffset += Array_nel-Array_array;
 			nodr.type = nodl.type;
-		} else
-			nodconst(&nodr, nodl.type, 0);
+		}
 		gmove(&nodr, &nodl);
 
 		goto yes;
diff --git a/src/cmd/9g/galign.c b/src/cmd/9g/galign.c
index df4291e90e..74856e2af0 100644
--- a/src/cmd/9g/galign.c
+++ b/src/cmd/9g/galign.c
@@ -14,12 +14,12 @@ void
 linkarchinit(void)
 {
 	thestring = getgoarch();
-	arch.thestring = thestring;
+	thearch.thestring = thestring;
 	if(strcmp(thestring, "ppc64le") == 0)
 		thelinkarch = &linkppc64le;
 	else
 		thelinkarch = &linkppc64;
-	arch.thelinkarch = thelinkarch;
+	thearch.thelinkarch = thelinkarch;
 }
 
 vlong MAXWIDTH = 1LL<<50;
@@ -49,61 +49,46 @@ betypeinit(void)
 void
 main(int argc, char **argv)
 {
-	arch.thechar = thechar;
-	arch.thestring = thestring;
-	arch.thelinkarch = thelinkarch;
-	arch.typedefs = typedefs;
-	arch.MAXWIDTH = MAXWIDTH;
-	arch.afunclit = afunclit;
-	arch.anyregalloc = anyregalloc;
-	arch.betypeinit = betypeinit;
-	arch.bgen = bgen;
-	arch.cgen = cgen;
-	arch.cgen_asop = cgen_asop;
-	arch.cgen_call = cgen_call;
-	arch.cgen_callinter = cgen_callinter;
-	arch.cgen_ret = cgen_ret;
-	arch.clearfat = clearfat;
-	arch.clearp = clearp;
-	arch.defframe = defframe;
-	arch.dgostringptr = dgostringptr;
-	arch.dgostrlitptr = dgostrlitptr;
-	arch.dsname = dsname;
-	arch.dsymptr = dsymptr;
-	arch.dumpdata = dumpdata;
-	arch.dumpit = dumpit;
-	arch.excise = excise;
-	arch.expandchecks = expandchecks;
-	arch.fixautoused = fixautoused;
-	arch.gclean = gclean;
-	arch.gdata = gdata;
-	arch.gdatacomplex = gdatacomplex;
-	arch.gdatastring = gdatastring;
-	arch.ggloblnod = ggloblnod;
-	arch.ggloblsym = ggloblsym;
-	arch.ginit = ginit;
-	arch.gins = gins;
-	arch.ginscall = ginscall;
-	arch.gjmp = gjmp;
-	arch.gtrack = gtrack;
-	arch.gused = gused;
-	arch.igen = igen;
-	arch.isfat = isfat;
-	arch.linkarchinit = linkarchinit;
-	arch.markautoused = markautoused;
-	arch.naddr = naddr;
-	arch.newplist = newplist;
-	arch.nodarg = nodarg;
-	arch.patch = patch;
-	arch.proginfo = proginfo;
-	arch.regalloc = regalloc;
-	arch.regfree = regfree;
-	arch.regopt = regopt;
-	arch.regtyp = regtyp;
-	arch.sameaddr = sameaddr;
-	arch.smallindir = smallindir;
-	arch.stackaddr = stackaddr;
-	arch.unpatch = unpatch;
+	thearch.thechar = thechar;
+	thearch.thestring = thestring;
+	thearch.thelinkarch = thelinkarch;
+	thearch.typedefs = typedefs;
+	thearch.REGSP = REGSP;
+	thearch.REGCTXT = REGCTXT;
+	thearch.MAXWIDTH = MAXWIDTH;
+	thearch.anyregalloc = anyregalloc;
+	thearch.betypeinit = betypeinit;
+	thearch.bgen = bgen;
+	thearch.cgen = cgen;
+	thearch.cgen_call = cgen_call;
+	thearch.cgen_callinter = cgen_callinter;
+	thearch.cgen_ret = cgen_ret;
+	thearch.clearfat = clearfat;
+	thearch.defframe = defframe;
+	thearch.excise = excise;
+	thearch.expandchecks = expandchecks;
+	thearch.gclean = gclean;
+	thearch.ginit = ginit;
+	thearch.gins = gins;
+	thearch.ginscall = ginscall;
+	thearch.igen = igen;
+	thearch.linkarchinit = linkarchinit;
+	thearch.peep = peep;
+	thearch.proginfo = proginfo;
+	thearch.regalloc = regalloc;
+	thearch.regfree = regfree;
+	thearch.regtyp = regtyp;
+	thearch.sameaddr = sameaddr;
+	thearch.smallindir = smallindir;
+	thearch.stackaddr = stackaddr;
+	thearch.excludedregs = excludedregs;
+	thearch.RtoB = RtoB;
+	thearch.FtoB = RtoB;
+	thearch.BtoR = BtoR;
+	thearch.BtoF = BtoF;
+	thearch.optoas = optoas;
+	thearch.doregbits = doregbits;
+	thearch.regnames = regnames;
 	
 	gcmain(argc, argv);
 }
diff --git a/src/cmd/9g/gg.h b/src/cmd/9g/gg.h
index 367a858e86..938675937e 100644
--- a/src/cmd/9g/gg.h
+++ b/src/cmd/9g/gg.h
@@ -9,13 +9,7 @@
 #include "../gc/go.h"
 #include "../9l/9.out.h"
 
-// TODO(minux): Remove when no longer used.
-#define noimpl sysfatal("%s not implemented (%s:%d).", __func__, __FILE__, __LINE__)
-
-EXTERN	int32	dynloc;
 EXTERN	uchar	reg[NREG+NFREG];
-EXTERN	int32	pcloc;		// instruction counter
-EXTERN	Strlit	emptystring;
 EXTERN	Node*	panicdiv;
 extern	vlong	unmappedzero;
 
@@ -112,7 +106,6 @@ int anyregalloc(void);
 void betypeinit(void);
 void bgen(Node*, int, int, Prog*);
 void cgen(Node*, Node*);
-void cgen_asop(Node*);
 void cgen_call(Node*, int);
 void cgen_callinter(Node*, Node*, int);
 void cgen_ret(Node*);
@@ -158,3 +151,19 @@ int smallindir(Addr*, Addr*);
 int stackaddr(Addr*);
 Prog* unpatch(Prog*);
 
+
+/*
+ * reg.c
+ */
+uint64 excludedregs(void);
+uint64 RtoB(int);
+uint64 FtoB(int);
+int BtoR(uint64);
+int BtoF(uint64);
+uint64 doregbits(int);
+char** regnames(int*);
+
+/*
+ * peep.c
+ */
+void peep(Prog*);
diff --git a/src/cmd/9g/ggen.c b/src/cmd/9g/ggen.c
index 3865051018..7b34282685 100644
--- a/src/cmd/9g/ggen.c
+++ b/src/cmd/9g/ggen.c
@@ -7,7 +7,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
 static Prog *appendpp(Prog *p, int as, int ftype, int freg, vlong foffset, int ttype, int treg, vlong toffset);
 static Prog *zerorange(Prog *p, vlong frame, vlong lo, vlong hi);
@@ -113,51 +113,6 @@ appendpp(Prog *p, int as, int ftype, int freg, vlong foffset, int ttype, int tre
 	return q;
 }
 
-// Sweep the prog list to mark any used nodes.
-void
-markautoused(Prog *p)
-{
-	for (; p; p = p->link) {
-		if (p->as == ATYPE || p->as == AVARDEF || p->as == AVARKILL)
-			continue;
-
-		if (p->from.node)
-			((Node*)(p->from.node))->used = 1;
-
-		if (p->to.node)
-			((Node*)(p->to.node))->used = 1;
-	}
-}
-
-// Fixup instructions after allocauto (formerly compactframe) has moved all autos around.
-void
-fixautoused(Prog *p)
-{
-	Prog **lp;
-
-	for (lp=&p; (p=*lp) != P; ) {
-		if (p->as == ATYPE && p->from.node && p->from.name == NAME_AUTO && !((Node*)(p->from.node))->used) {
-			*lp = p->link;
-			continue;
-		}
-		if ((p->as == AVARDEF || p->as == AVARKILL) && p->to.node && !((Node*)(p->to.node))->used) {
-			// Cannot remove VARDEF instruction, because - unlike TYPE handled above -
-			// VARDEFs are interspersed with other code, and a jump might be using the
-			// VARDEF as a target. Replace with a no-op instead. A later pass will remove
-			// the no-ops.
-			nopout(p);
-			continue;
-		}
-		if (p->from.name == NAME_AUTO && p->from.node)
-			p->from.offset += ((Node*)(p->from.node))->stkdelta;
-
-		if (p->to.name == NAME_AUTO && p->to.node)
-			p->to.offset += ((Node*)(p->to.node))->stkdelta;
-
-		lp = &p->link;
-	}
-}
-
 /*
  * generate: BL reg, f
  * where both reg and f are registers.
@@ -226,7 +181,7 @@ ginscall(Node *f, int proc)
 				gins(AUNDEF, N, N);
 			break;
 		}
-		nodreg(&reg, types[tptr], REGENV);
+		nodreg(&reg, types[tptr], REGCTXT);
 		nodreg(&r1, types[tptr], REG_R3);
 		gmove(f, &reg);
 		reg.op = OINDREG;
@@ -487,27 +442,6 @@ cgen_ret(Node *n)
 	}
 }
 
-void
-cgen_asop(Node *n)
-{
-	USED(n);
-	fatal("cgen_asop"); // no longer used
-}
-
-int
-samereg(Node *a, Node *b)
-{
-	if(a == N || b == N)
-		return 0;
-	if(a->op != OREGISTER)
-		return 0;
-	if(b->op != OREGISTER)
-		return 0;
-	if(a->val.u.reg != b->val.u.reg)
-		return 0;
-	return 1;
-}
-
 /*
  * generate division.
  * generates one of:
diff --git a/src/cmd/9g/gobj.c b/src/cmd/9g/gobj.c
deleted file mode 100644
index 35e4f458f1..0000000000
--- a/src/cmd/9g/gobj.c
+++ /dev/null
@@ -1,250 +0,0 @@
-// Derived from Inferno utils/6c/swt.c
-// http://code.google.com/p/inferno-os/source/browse/utils/6c/swt.c
-//
-//	Copyright © 1994-1999 Lucent Technologies Inc.  All rights reserved.
-//	Portions Copyright © 1995-1997 C H Forsyth (forsyth@terzarima.net)
-//	Portions Copyright © 1997-1999 Vita Nuova Limited
-//	Portions Copyright © 2000-2007 Vita Nuova Holdings Limited (www.vitanuova.com)
-//	Portions Copyright © 2004,2006 Bruce Ellis
-//	Portions Copyright © 2005-2007 C H Forsyth (forsyth@terzarima.net)
-//	Revisions Copyright © 2000-2007 Lucent Technologies Inc. and others
-//	Portions Copyright © 2009 The Go Authors.  All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-//
-// The above copyright notice and this permission notice shall be included in
-// all copies or substantial portions of the Software.
-//
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
-// THE SOFTWARE.
-
-#include <u.h>
-#include <libc.h>
-#include "gg.h"
-
-int
-dsname(Sym *s, int off, char *t, int n)
-{
-	Prog *p;
-
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.offset = off;
-	p->from.sym = linksym(s);
-
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = n;
-	
-	p->to.type = TYPE_SCONST;
-	p->to.name = NAME_NONE;
-	p->to.offset = 0;
-	memmove(p->to.u.sval, t, n);
-	return off + n;
-}
-
-/*
- * make a refer to the data s, s+len
- * emitting DATA if needed.
- */
-void
-datastring(char *s, int len, Addr *a)
-{
-	Sym *sym;
-	
-	sym = stringsym(s, len);
-	a->type = TYPE_MEM;
-	a->name = NAME_EXTERN;
-	a->etype = simtype[TINT];
-	a->offset = widthptr+widthint;  // skip header
-	a->reg = 0;
-	a->sym = linksym(sym);
-	a->node = sym->def;
-}
-
-/*
- * make a refer to the string sval,
- * emitting DATA if needed.
- */
-void
-datagostring(Strlit *sval, Addr *a)
-{
-	Sym *sym;
-
-	sym = stringsym(sval->s, sval->len);
-	a->type = TYPE_MEM;
-	a->name = NAME_EXTERN;
-	a->sym = linksym(sym);
-	a->reg = 0;
-	a->node = sym->def;
-	a->offset = 0;  // header
-	a->etype = TSTRING;
-}
-
-void
-gdata(Node *nam, Node *nr, int wid)
-{
-	Prog *p;
-
-	if(nr->op == OLITERAL) {
-		switch(nr->val.ctype) {
-		case CTCPLX:
-			gdatacomplex(nam, nr->val.u.cval);
-			return;
-		case CTSTR:
-			gdatastring(nam, nr->val.u.sval);
-			return;
-		}
-	}
-	p = gins(ADATA, nam, nr);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = wid;
-}
-
-void
-gdatacomplex(Node *nam, Mpcplx *cval)
-{
-	Prog *p;
-	int w;
-
-	w = cplxsubtype(nam->type->etype);
-	w = types[w]->width;
-
-	p = gins(ADATA, nam, N);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = w;
-	p->to.type = TYPE_FCONST;
-	p->to.u.dval = mpgetflt(&cval->real);
-
-	p = gins(ADATA, nam, N);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = w;
-	p->from.offset += w;
-	p->to.type = TYPE_FCONST;
-	p->to.u.dval = mpgetflt(&cval->imag);
-}
-
-void
-gdatastring(Node *nam, Strlit *sval)
-{
-	Prog *p;
-	Node nod1;
-
-	p = gins(ADATA, nam, N);
-	datastring(sval->s, sval->len, &p->to);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = types[tptr]->width;
-	p->to.type = TYPE_ADDR;
-	p->to.etype = simtype[tptr];
-
-	nodconst(&nod1, types[TINT], sval->len);
-	p = gins(ADATA, nam, &nod1);
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthint;
-	p->from.offset += widthptr;
-}
-
-int
-dstringptr(Sym *s, int off, char *str)
-{
-	Prog *p;
-
-	off = rnd(off, widthptr);
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-
-	datastring(str, strlen(str)+1, &p->to);
-	p->to.type = TYPE_CONST;
-	p->to.etype = simtype[TINT];
-	off += widthptr;
-
-	return off;
-}
-
-int
-dgostrlitptr(Sym *s, int off, Strlit *lit)
-{
-	Prog *p;
-
-	if(lit == nil)
-		return duintptr(s, off, 0);
-
-	off = rnd(off, widthptr);
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-	datagostring(lit, &p->to);
-	p->to.type = TYPE_ADDR;
-	p->to.etype = simtype[TINT];
-	off += widthptr;
-
-	return off;
-}
-
-int
-dgostringptr(Sym *s, int off, char *str)
-{
-	int n;
-	Strlit *lit;
-
-	if(str == nil)
-		return duintptr(s, off, 0);
-
-	n = strlen(str);
-	lit = mal(sizeof *lit + n);
-	strcpy(lit->s, str);
-	lit->len = n;
-	return dgostrlitptr(s, off, lit);
-}
-
-int
-dsymptr(Sym *s, int off, Sym *x, int xoff)
-{
-	Prog *p;
-
-	off = rnd(off, widthptr);
-
-	p = gins(ADATA, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->from.offset = off;
-	p->from3.type = TYPE_CONST;
-	p->from3.offset = widthptr;
-	p->to.type = TYPE_ADDR;
-	p->to.name = NAME_EXTERN;
-	p->to.sym = linksym(x);
-	p->to.offset = xoff;
-	off += widthptr;
-
-	return off;
-}
-
-void
-nopout(Prog *p)
-{
-	p->as = ANOP;
-	p->from = zprog.from;
-	p->from3 = zprog.from3;
-	p->reg = zprog.reg;
-	p->to = zprog.to;
-}
diff --git a/src/cmd/9g/gsubr.c b/src/cmd/9g/gsubr.c
index 49f184d51e..1fb1511361 100644
--- a/src/cmd/9g/gsubr.c
+++ b/src/cmd/9g/gsubr.c
@@ -38,227 +38,6 @@
 // At the same time, can raise StackBig in ../../runtime/stack.h.
 vlong unmappedzero = 4096;
 
-void
-clearp(Prog *p)
-{
-	*p = zprog;
-	p->as = AEND;
-	p->pc = pcloc;
-	pcloc++;
-}
-
-static int ddumped;
-static Prog *dfirst;
-static Prog *dpc;
-
-/*
- * generate and return proc with p->as = as,
- * linked into program. pc is next instruction.
- */
-Prog*
-prog(int as)
-{
-	Prog *p;
-
-	if(as == ADATA || as == AGLOBL) {
-		if(ddumped)
-			fatal("already dumped data");
-		if(dpc == nil) {
-			dpc = mal(sizeof(*dpc));
-			dfirst = dpc;
-		}
-		p = dpc;
-		dpc = mal(sizeof(*dpc));
-		p->link = dpc;
-		p->reg = 0; // used for flags
-	} else {
-		p = pc;
-		pc = mal(sizeof(*pc));
-		clearp(pc);
-		p->link = pc;
-	}
-
-	if(lineno == 0) {
-		if(debug['K'])
-			warn("prog: line 0");
-	}
-
-	p->as = as;
-	p->lineno = lineno;
-	return p;
-}
-
-void
-dumpdata(void)
-{
-	ddumped = 1;
-	if(dfirst == nil)
-		return;
-	newplist();
-	*pc = *dfirst;
-	pc = dpc;
-	clearp(pc);
-}
-
-/*
- * generate a branch.
- * t is ignored.
- * likely values are for branch prediction:
- *	-1 unlikely
- *	0 no opinion
- *	+1 likely
- */
-Prog*
-gbranch(int as, Type *t, int likely)
-{
-	Prog *p;
-	
-	USED(t);
-
-	p = prog(as);
-	p->to.type = TYPE_BRANCH;
-	p->to.u.branch = P;
-	// TODO(minux): Enable this code.
-	// Note: liblink used Bcc CR0, label form, so we need another way
-	// to set likely/unlikely flag. Also note the y bit is not exactly
-	// likely/unlikely bit.
-	if(0 && as != ABR && likely != 0) {
-		p->from.type = TYPE_CONST;
-		p->from.offset = likely > 0;
-	}
-	return p;
-}
-
-/*
- * patch previous branch to jump to to.
- */
-void
-patch(Prog *p, Prog *to)
-{
-	if(p->to.type != TYPE_BRANCH)
-		fatal("patch: not a branch");
-	p->to.u.branch = to;
-	p->to.offset = to->pc;
-}
-
-Prog*
-unpatch(Prog *p)
-{
-	Prog *q;
-
-	if(p->to.type != TYPE_BRANCH)
-		fatal("unpatch: not a branch");
-	q = p->to.u.branch;
-	p->to.u.branch = P;
-	p->to.offset = 0;
-	return q;
-}
-
-/*
- * start a new Prog list.
- */
-Plist*
-newplist(void)
-{
-	Plist *pl;
-
-	pl = linknewplist(ctxt);
-
-	pc = mal(sizeof(*pc));
-	clearp(pc);
-	pl->firstpc = pc;
-
-	return pl;
-}
-
-void
-gused(Node *n)
-{
-	gins(ANOP, n, N);	// used
-}
-
-Prog*
-gjmp(Prog *to)
-{
-	Prog *p;
-
-	p = gbranch(ABR, T, 0);
-	if(to != P)
-		patch(p, to);
-	return p;
-}
-
-void
-ggloblnod(Node *nam)
-{
-	Prog *p;
-
-	p = gins(AGLOBL, nam, N);
-	p->lineno = nam->lineno;
-	p->from.sym->gotype = linksym(ngotype(nam));
-	p->to.sym = nil;
-	p->to.type = TYPE_CONST;
-	p->to.offset = nam->type->width;
-	if(nam->readonly)
-		p->from3.offset = RODATA;
-	if(nam->type != T && !haspointers(nam->type))
-		p->from3.offset |= NOPTR;
-}
-
-void
-gtrack(Sym *s)
-{
-	Prog *p;
-	
-	p = gins(AUSEFIELD, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-}
-
-void
-ggloblsym(Sym *s, int32 width, int8 flags)
-{
-	Prog *p;
-
-	p = gins(AGLOBL, N, N);
-	p->from.type = TYPE_MEM;
-	p->from.name = NAME_EXTERN;
-	p->from.sym = linksym(s);
-	p->to.type = TYPE_CONST;
-	p->to.name = NAME_NONE;
-	p->to.offset = width;
-	p->from3.offset = flags;
-}
-
-int
-isfat(Type *t)
-{
-	if(t != T)
-	switch(t->etype) {
-	case TSTRUCT:
-	case TARRAY:
-	case TSTRING:
-	case TINTER:	// maybe remove later
-		return 1;
-	}
-	return 0;
-}
-
-/*
- * naddr of func generates code for address of func.
- * if using opcode that can take address implicitly,
- * call afunclit to fix up the argument.
- */
-void
-afunclit(Addr *a, Node *n)
-{
-	if(a->type == TYPE_ADDR && a->name == NAME_EXTERN) {
-		a->type = TYPE_MEM;
-		a->sym = linksym(n->sym);
-	}
-}
-
 static	int	resvd[] =
 {
 	REGZERO,
@@ -308,7 +87,7 @@ gclean(void)
 			yyerror("reg %R left allocated, %p\n", i+REG_R0, regpc[i]);
 }
 
-int32
+int
 anyregalloc(void)
 {
 	int i, j;
@@ -432,109 +211,6 @@ regfree(Node *n)
 }
 
 /*
- * initialize n to be register r of type t.
- */
-void
-nodreg(Node *n, Type *t, int r)
-{
-	if(t == T)
-		fatal("nodreg: t nil");
-
-	memset(n, 0, sizeof(*n));
-	n->op = OREGISTER;
-	n->addable = 1;
-	ullmancalc(n);
-	n->val.u.reg = r;
-	n->type = t;
-}
-
-/*
- * initialize n to be indirect of register r; n is type t.
- */
-void
-nodindreg(Node *n, Type *t, int r)
-{
-	nodreg(n, t, r);
-	n->op = OINDREG;
-}
-
-Node*
-nodarg(Type *t, int fp)
-{
-	Node *n;
-	NodeList *l;
-	Type *first;
-	Iter savet;
-
-	// entire argument struct, not just one arg
-	if(t->etype == TSTRUCT && t->funarg) {
-		n = nod(ONAME, N, N);
-		n->sym = lookup(".args");
-		n->type = t;
-		first = structfirst(&savet, &t);
-		if(first == nil)
-			fatal("nodarg: bad struct");
-		if(first->width == BADWIDTH)
-			fatal("nodarg: offset not computed for %T", t);
-		n->xoffset = first->width;
-		n->addable = 1;
-		goto fp;
-	}
-
-	if(t->etype != TFIELD)
-		fatal("nodarg: not field %T", t);
-	
-	if(fp == 1) {
-		for(l=curfn->dcl; l; l=l->next) {
-			n = l->n;
-			if((n->class == PPARAM || n->class == PPARAMOUT) && !isblanksym(t->sym) && n->sym == t->sym)
-				return n;
-		}
-	}
-
-	n = nod(ONAME, N, N);
-	n->type = t->type;
-	n->sym = t->sym;
-	
-	if(t->width == BADWIDTH)
-		fatal("nodarg: offset not computed for %T", t);
-	n->xoffset = t->width;
-	n->addable = 1;
-	n->orig = t->nname;
-
-fp:
-	// Rewrite argument named _ to __,
-	// or else the assignment to _ will be
-	// discarded during code generation.
-	if(isblank(n))
-		n->sym = lookup("__");
-
-	switch(fp) {
-	default:
-		fatal("nodarg %T %d", t, fp);
-
-	case 0:		// output arg for calling another function
-		n->op = OINDREG;
-		n->val.u.reg = REGSP;
-		n->xoffset += 8;
-		break;
-
-	case 1:		// input arg to current function
-		n->class = PPARAM;
-		break;
-
-	case 2:		// offset output arg
-fatal("shouldn't be used");
-		n->op = OINDREG;
-		n->val.u.reg = REGSP;
-		n->xoffset += types[tptr]->width;
-		break;
-	}
-	n->typecheck = 1;
-	return n;
-}
-
-/*
  * generate
  *	as $c, n
  */
@@ -595,27 +271,6 @@ ginscon2(int as, Node *n2, vlong c)
 /*c2go int CASE(int, int); */
 
 /*
- * Is this node a memory operand?
- */
-int
-ismem(Node *n)
-{
-	switch(n->op) {
-	case OITAB:
-	case OSPTR:
-	case OLEN:
-	case OCAP:
-	case OINDREG:
-	case ONAME:
-	case OPARAM:
-	case OCLOSUREVAR:
-	case OADDR:
-		return 1;
-	}
-	return 0;
-}
-
-/*
  * set up nodes representing 2^63
  */
 Node bigi;
@@ -1077,8 +732,7 @@ fixlargeoffset(Node *n)
 	if(n->xoffset != (int32)n->xoffset) {
 		// TODO(minux): offset too large, move into R31 and add to R31 instead.
 		// this is used only in test/fixedbugs/issue6036.go.
-		print("offset too large: %N\n", n);
-		noimpl;
+		fatal("offset too large: %N", n);
 		a = *n;
 		a.op = OREGISTER;
 		a.type = types[tptr];
@@ -1090,208 +744,6 @@ fixlargeoffset(Node *n)
 }
 
 /*
- * generate code to compute n;
- * make a refer to result.
- */
-void
-naddr(Node *n, Addr *a, int canemitcode)
-{
-	Sym *s;
-
-	a->type = TYPE_NONE;
-	a->name = NAME_NONE;
-	a->reg = 0;
-	a->gotype = nil;
-	a->node = N;
-	a->etype = 0;
-	a->width = 0;
-	if(n == N)
-		return;
-
-	if(n->type != T && n->type->etype != TIDEAL) {
-		dowidth(n->type);
-		a->width = n->type->width;
-	}
-
-	switch(n->op) {
-	default:
-		fatal("naddr: bad %O %D", n->op, a);
-		break;
-
-	case ONAME:
-		a->etype = 0;
-		a->reg = 0;
-		if(n->type != T)
-			a->etype = simtype[n->type->etype];
-		a->offset = n->xoffset;
-		s = n->sym;
-		a->node = n->orig;
-		//if(a->node >= (Node*)&n)
-		//	fatal("stack node");
-		if(s == S)
-			s = lookup(".noname");
-		if(n->method) {
-			if(n->type != T)
-			if(n->type->sym != S)
-			if(n->type->sym->pkg != nil)
-				s = pkglookup(s->name, n->type->sym->pkg);
-		}
-
-		a->type = TYPE_MEM;
-		switch(n->class) {
-		default:
-			fatal("naddr: ONAME class %S %d\n", n->sym, n->class);
-		case PEXTERN:
-			a->name = NAME_EXTERN;
-			break;
-		case PAUTO:
-			a->name = NAME_AUTO;
-			break;
-		case PPARAM:
-		case PPARAMOUT:
-			a->name = NAME_PARAM;
-			break;
-		case PFUNC:
-			a->name = NAME_EXTERN;
-			a->type = TYPE_ADDR;
-			a->width = widthptr;
-			s = funcsym(s);
-			break;
-		}
-		a->sym = linksym(s);
-		break;
-
-	case OLITERAL:
-		switch(n->val.ctype) {
-		default:
-			fatal("naddr: const %lT", n->type);
-			break;
-		case CTFLT:
-			a->type = TYPE_FCONST;
-			a->u.dval = mpgetflt(n->val.u.fval);
-			break;
-		case CTINT:
-		case CTRUNE:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = mpgetfix(n->val.u.xval);
-			break;
-		case CTSTR:
-			datagostring(n->val.u.sval, a);
-			break;
-		case CTBOOL:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = n->val.u.bval;
-			break;
-		case CTNIL:
-			a->sym = nil;
-			a->type = TYPE_CONST;
-			a->offset = 0;
-			break;
-		}
-		break;
-
-	case OREGISTER:
-		a->type = TYPE_REG;
-		a->reg = n->val.u.reg;
-		a->sym = nil;
-		break;
-
-	case OINDREG:
-		a->type = TYPE_MEM;
-		a->reg = n->val.u.reg;
-		a->sym = linksym(n->sym);
-		a->offset = n->xoffset;
-		if(a->offset != (int32)a->offset)
-			yyerror("offset %lld too large for OINDREG", a->offset);
-		break;
-
-	case OPARAM:
-		// n->left is PHEAP ONAME for stack parameter.
-		// compute address of actual parameter on stack.
-		a->etype = simtype[n->left->type->etype];
-		a->width = n->left->type->width;
-		a->offset = n->xoffset;
-		a->sym = linksym(n->left->sym);
-		a->type = TYPE_MEM;
-		a->name = NAME_PARAM;
-		a->node = n->left->orig;
-		break;
-
-	case OCLOSUREVAR:
-		if(!curfn->needctxt)
-			fatal("closurevar without needctxt");
-		a->type = TYPE_MEM;
-		a->reg = REGENV;
-		a->offset = n->xoffset;
-		a->sym = nil;
-		break;
-
-	case OCFUNC:
-		naddr(n->left, a, canemitcode);
-		a->sym = linksym(n->left->sym);
-		break;
-
-	case OITAB:
-		// itable of interface value
-		naddr(n->left, a, canemitcode);
-		a->etype = simtype[tptr];
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// itab(nil)
-		a->width = widthptr;
-		break;
-
-	case OSPTR:
-		// pointer in a string or slice
-		naddr(n->left, a, canemitcode);
-		a->etype = simtype[tptr];
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// ptr(nil)
-		a->offset += Array_array;
-		a->width = widthptr;
-		break;
-
-	case OLEN:
-		// len of string or slice
-		naddr(n->left, a, canemitcode);
-		a->etype = simtype[TINT];
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// len(nil)
-		a->offset += Array_nel;
-		a->width = widthint;
-		break;
-
-	case OCAP:
-		// cap of string or slice
-		naddr(n->left, a, canemitcode);
-		a->etype = simtype[TINT];
-		if(a->type == TYPE_CONST && a->offset == 0)
-			break;	// cap(nil)
-		a->offset += Array_cap;
-		a->width = widthint;
-		break;
-
-	case OADDR:
-		naddr(n->left, a, canemitcode);
-		a->etype = tptr;
-		switch(a->type) {
-		case TYPE_MEM:
-			a->type = TYPE_ADDR;
-			break;
-
-		case TYPE_REG:
-		case TYPE_CONST:
-			break;
-
-		default:
-			fatal("naddr: OADDR %d\n", a->type);
-		}
-		break;
-	}
-}
-
-/*
  * return Axxx for Oxxx on type t.
  */
 int
@@ -1700,6 +1152,7 @@ int
 sudoaddable(int as, Node *n, Addr *a)
 {
 	// TODO(minux)
-	USED(as); USED(n); USED(a);
+	USED(as); USED(n);
+	memset(a, 0, sizeof *a);
 	return 0;
 }
diff --git a/src/cmd/9g/opt.h b/src/cmd/9g/opt.h
index 25b6703279..79a34fb1f0 100644
--- a/src/cmd/9g/opt.h
+++ b/src/cmd/9g/opt.h
@@ -1,175 +1,6 @@
-// Derived from Inferno utils/6c/gc.h
-// http://code.google.com/p/inferno-os/source/browse/utils/6c/gc.h
-//
-//	Copyright © 1994-1999 Lucent Technologies Inc.  All rights reserved.
-//	Portions Copyright © 1995-1997 C H Forsyth (forsyth@terzarima.net)
-//	Portions Copyright © 1997-1999 Vita Nuova Limited
-//	Portions Copyright © 2000-2007 Vita Nuova Holdings Limited (www.vitanuova.com)
-//	Portions Copyright © 2004,2006 Bruce Ellis
-//	Portions Copyright © 2005-2007 C H Forsyth (forsyth@terzarima.net)
-//	Revisions Copyright © 2000-2007 Lucent Technologies Inc. and others
-//	Portions Copyright © 2009 The Go Authors.  All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-//
-// The above copyright notice and this permission notice shall be included in
-// all copies or substantial portions of the Software.
-//
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
-// THE SOFTWARE.
-
-
-#define	Z	N
-#define	Adr	Addr
-
-#define	BLOAD(r)	band(bnot(r->refbehind), r->refahead)
-#define	BSTORE(r)	band(bnot(r->calbehind), r->calahead)
-#define	LOAD(r)		(~r->refbehind.b[z] & r->refahead.b[z])
-#define	STORE(r)	(~r->calbehind.b[z] & r->calahead.b[z])
-
-#define	CLOAD	5
-#define	CREF	5
-#define	CINF	1000
-#define	LOOP	3
-
-typedef	struct	Reg	Reg;
-typedef	struct	Rgn	Rgn;
-
-/*c2go
-extern Node *Z;
-enum
-{
-	CLOAD = 5,
-	CREF = 5,
-	CINF = 1000,
-	LOOP = 3,
-};
-
-uint32 BLOAD(Reg*);
-uint32 BSTORE(Reg*);
-uint32 LOAD(Reg*);
-uint32 STORE(Reg*);
-*/
-
-// A Reg is a wrapper around a single Prog (one instruction) that holds
-// register optimization information while the optimizer runs.
-// r->prog is the instruction.
-// r->prog->opt points back to r.
-struct	Reg
-{
-	Flow	f;
-
-	Bits	set;  		// regopt variables written by this instruction.
-	Bits	use1; 		// regopt variables read by prog->from.
-	Bits	use2; 		// regopt variables read by prog->to.
-
-	// refahead/refbehind are the regopt variables whose current
-	// value may be used in the following/preceding instructions
-	// up to a CALL (or the value is clobbered).
-	Bits	refbehind;
-	Bits	refahead;
-	// calahead/calbehind are similar, but for variables in
-	// instructions that are reachable after hitting at least one
-	// CALL.
-	Bits	calbehind;
-	Bits	calahead;
-	Bits	regdiff;
-	Bits	act;
-
-	uint64	regu;		// register used bitmap
-};
-#define	R	((Reg*)0)
-/*c2go extern Reg *R; */
-
-#define	NRGN	600
-/*c2go enum { NRGN = 600 }; */
-
-// A Rgn represents a single regopt variable over a region of code
-// where a register could potentially be dedicated to that variable.
-// The code encompassed by a Rgn is defined by the flow graph,
-// starting at enter, flood-filling forward while varno is refahead
-// and backward while varno is refbehind, and following branches.  A
-// single variable may be represented by multiple disjoint Rgns and
-// each Rgn may choose a different register for that variable.
-// Registers are allocated to regions greedily in order of descending
-// cost.
-struct	Rgn
-{
-	Reg*	enter;
-	short	cost;
-	short	varno;
-	short	regno;
-};
-
-EXTERN	int32	exregoffset;		// not set
-EXTERN	int32	exfregoffset;		// not set
-EXTERN	Reg	zreg;
-EXTERN	Rgn	region[NRGN];
-EXTERN	Rgn*	rgp;
-EXTERN	int	nregion;
-EXTERN	int	nvar;
-EXTERN	int32	regbits;
-EXTERN	int32	exregbits;		// TODO(austin) not used; remove
-EXTERN	Bits	externs;
-EXTERN	Bits	params;
-EXTERN	Bits	consts;
-EXTERN	Bits	addrs;
-EXTERN	Bits	ivar;
-EXTERN	Bits	ovar;
-EXTERN	int	change;
-EXTERN	int32	maxnr;
-
-EXTERN	struct
-{
-	int32	ncvtreg;
-	int32	nspill;
-	int32	ndelmov;
-	int32	nvar;
-} ostats;
-
-/*
- * reg.c
- */
-int	rcmp(const void*, const void*);
-void	regopt(Prog*);
-void	addmove(Reg*, int, int, int);
-Bits	mkvar(Reg*, Adr*);
-void	prop(Reg*, Bits, Bits);
-void	synch(Reg*, Bits);
-uint64	allreg(uint64, Rgn*);
-void	paint1(Reg*, int);
-uint64	paint2(Reg*, int, int);
-void	paint3(Reg*, int, uint64, int);
-void	addreg(Adr*, int);
-void	dumpone(Flow*, int);
-void	dumpit(char*, Flow*, int);
-
-/*
- * peep.c
- */
-void	peep(Prog*);
-void	excise(Flow*);
-int	copyu(Prog*, Adr*, Adr*);
-
-uint64	RtoB(int);
-uint64	FtoB(int);
-int	BtoR(uint64);
-int	BtoF(uint64);
-
-/*
- * prog.c
- */
-void proginfo(ProgInfo*, Prog*);
+// Copyright 2014 The Go Authors.  All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
 
 // Many Power ISA arithmetic and logical instructions come in four
 // standard variants.  These bits let us map between variants.
diff --git a/src/cmd/9g/peep.c b/src/cmd/9g/peep.c
index 95ff0b4d58..cf96d5dcda 100644
--- a/src/cmd/9g/peep.c
+++ b/src/cmd/9g/peep.c
@@ -31,6 +31,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
+#include "../gc/popt.h"
 #include "opt.h"
 
 static int	regzer(Addr *a);
@@ -42,6 +43,7 @@ static int	copyau(Addr*, Addr*);
 static int	copysub(Addr*, Addr*, Addr*, int);
 static int	copysub1(Prog*, Addr*, Addr*, int);
 static int	copyau1(Prog *p, Addr *v);
+static int	copyu(Prog *p, Addr *v, Addr *s);
 
 static uint32	gactive;
 
@@ -53,7 +55,7 @@ peep(Prog *firstp)
 	Prog *p, *p1;
 	int t;
 
-	g = flowstart(firstp, sizeof(Flow));
+	g = flowstart(firstp, 0);
 	if(g == nil)
 		return;
 	gactive = 0;
@@ -568,12 +570,12 @@ copy1(Addr *v1, Addr *v2, Flow *r, int f)
 //	4 if v is set in one address and used in another (so addresses
 // 	  can be rewritten independently)
 //	0 otherwise (not touched)
-int
+static int
 copyu(Prog *p, Addr *v, Addr *s)
 {
 	if(p->from3.type != TYPE_NONE)
 		// 9g never generates a from3
-		print("copyu: from3 (%D) not implemented\n", p->from3);
+		print("copyu: from3 (%D) not implemented\n", &p->from3);
 
 	switch(p->as) {
 
diff --git a/src/cmd/9g/prog.c b/src/cmd/9g/prog.c
index 1775993a97..561249c358 100644
--- a/src/cmd/9g/prog.c
+++ b/src/cmd/9g/prog.c
@@ -5,6 +5,7 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
+#include "../gc/popt.h"
 #include "opt.h"
 
 enum {
diff --git a/src/cmd/9g/reg.c b/src/cmd/9g/reg.c
index a7ee07e547..84e1747e8d 100644
--- a/src/cmd/9g/reg.c
+++ b/src/cmd/9g/reg.c
@@ -31,56 +31,12 @@
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
-#include "opt.h"
+#include "../gc/popt.h"
 
-#define	NREGVAR	64	/* 32 general + 32 floating */
-#define	REGBITS	((uint64)0xffffffffffffffffull)
-/*c2go enum {
-	NREGVAR = 64,
-	REGBITS = 0xffffffffffffffff,
+enum {
+	NREGVAR = 64,	/* 32 general + 32 floating */
 };
-*/
 
-static	Reg*	firstr;
-static	int	first	= 1;
-
-int
-rcmp(const void *a1, const void *a2)
-{
-	Rgn *p1, *p2;
-	int c1, c2;
-
-	p1 = (Rgn*)a1;
-	p2 = (Rgn*)a2;
-	c1 = p2->cost;
-	c2 = p1->cost;
-	if(c1 -= c2)
-		return c1;
-	return p2->varno - p1->varno;
-}
-
-static void
-setaddrs(Bits bit)
-{
-	int i, n;
-	Var *v;
-	Node *node;
-
-	while(bany(&bit)) {
-		// convert each bit to a variable
-		i = bnum(bit);
-		node = var[i].node;
-		n = var[i].name;
-		biclr(&bit, i);
-
-		// disable all pieces of that variable
-		for(i=0; i<nvar; i++) {
-			v = var+i;
-			if(v->node == node && v->name == n)
-				v->addr = 2;
-		}
-	}
-}
 
 static char* regname[] = {
 	".R0",
@@ -149,1059 +105,32 @@ static char* regname[] = {
 	".F31",
 };
 
-static Node* regnodes[NREGVAR];
-
-static void walkvardef(Node *n, Reg *r, int active);
-
-void
-regopt(Prog *firstp)
+char**
+regnames(int *n)
 {
-	Reg *r, *r1;
-	Prog *p;
-	Graph *g;
-	ProgInfo info;
-	int i, z, active;
-	uint64 vreg, usedreg;
-	Bits bit;
-
-	if(first) {
-		fmtinstall('Q', Qconv);
-		first = 0;
-	}
-
-	mergetemp(firstp);
+	*n = NREGVAR;
+	return regname;
+}
 
-	/*
-	 * control flow is more complicated in generated go code
-	 * than in generated c code.  define pseudo-variables for
-	 * registers, so we have complete register usage information.
-	 */
-	nvar = NREGVAR;
-	memset(var, 0, NREGVAR*sizeof var[0]);
-	for(i=0; i<NREGVAR; i++) {
-		if(regnodes[i] == N)
-			regnodes[i] = newname(lookup(regname[i]));
-		var[i].node = regnodes[i];
-	}
+uint64
+excludedregs(void)
+{
+	uint64 regbits;
 
 	// Exclude registers with fixed functions
 	regbits = (1<<0)|RtoB(REGSP)|RtoB(REGG)|RtoB(REGTLS);
 	// Also exclude floating point registers with fixed constants
 	regbits |= RtoB(REG_F27)|RtoB(REG_F28)|RtoB(REG_F29)|RtoB(REG_F30)|RtoB(REG_F31);
-	externs = zbits;
-	params = zbits;
-	consts = zbits;
-	addrs = zbits;
-	ivar = zbits;
-	ovar = zbits;
-
-	/*
-	 * pass 1
-	 * build aux data structure
-	 * allocate pcs
-	 * find use and set of variables
-	 */
-	g = flowstart(firstp, sizeof(Reg));
-	if(g == nil) {
-		for(i=0; i<nvar; i++)
-			var[i].node->opt = nil;
-		return;
-	}
-
-	firstr = (Reg*)g->start;
-
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		p = r->f.prog;
-		if(p->as == AVARDEF || p->as == AVARKILL)
-			continue;
-		proginfo(&info, p);
-
-		// Avoid making variables for direct-called functions.
-		if(p->as == ABL && p->to.name == NAME_EXTERN)
-			continue;
-
-		// from vs to doesn't matter for registers
-		r->use1.b[0] |= info.reguse | info.regindex;
-		r->set.b[0] |= info.regset;
-
-		// Compute used register for from
-		bit = mkvar(r, &p->from);
-		if(info.flags & LeftAddr)
-			setaddrs(bit);
-		if(info.flags & LeftRead)
-			for(z=0; z<BITS; z++)
-				r->use1.b[z] |= bit.b[z];
-
-		// Compute used register for reg
-		if(info.flags & RegRead)
-			r->use1.b[0] |= RtoB(p->reg);
-
-		// Currently we never generate three register forms.
-		// If we do, this will need to change.
-		if(p->from3.type != TYPE_NONE)
-			fatal("regopt not implemented for from3");
-
-		// Compute used register for to
-		bit = mkvar(r, &p->to);
-		if(info.flags & RightAddr)
-			setaddrs(bit);
-		if(info.flags & RightRead)
-			for(z=0; z<BITS; z++)
-				r->use2.b[z] |= bit.b[z];
-		if(info.flags & RightWrite)
-			for(z=0; z<BITS; z++)
-				r->set.b[z] |= bit.b[z];
-	}
-
-	for(i=0; i<nvar; i++) {
-		Var *v = var+i;
-		if(v->addr) {
-			bit = blsh(i);
-			for(z=0; z<BITS; z++)
-				addrs.b[z] |= bit.b[z];
-		}
-
-		if(debug['R'] && debug['v'])
-			print("bit=%2d addr=%d et=%-6E w=%-2d s=%N + %lld\n",
-				i, v->addr, v->etype, v->width, v->node, v->offset);
-	}
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass1", &firstr->f, 1);
-
-	/*
-	 * pass 2
-	 * find looping structure
-	 */
-	flowrpo(g);
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass2", &firstr->f, 1);
-
-	/*
-	 * pass 2.5
-	 * iterate propagating fat vardef covering forward
-	 * r->act records vars with a VARDEF since the last CALL.
-	 * (r->act will be reused in pass 5 for something else,
-	 * but we'll be done with it by then.)
-	 */
-	active = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		r->f.active = 0;
-		r->act = zbits;
-	}
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		p = r->f.prog;
-		if(p->as == AVARDEF && isfat(((Node*)(p->to.node))->type) && ((Node*)(p->to.node))->opt != nil) {
-			active++;
-			walkvardef(p->to.node, r, active);
-		}
-	}
-
-	/*
-	 * pass 3
-	 * iterate propagating usage
-	 * 	back until flow graph is complete
-	 */
-loop1:
-	change = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->f.active = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		if(r->f.prog->as == ARET)
-			prop(r, zbits, zbits);
-loop11:
-	/* pick up unreachable code */
-	i = 0;
-	for(r = firstr; r != R; r = r1) {
-		r1 = (Reg*)r->f.link;
-		if(r1 && r1->f.active && !r->f.active) {
-			prop(r, zbits, zbits);
-			i = 1;
-		}
-	}
-	if(i)
-		goto loop11;
-	if(change)
-		goto loop1;
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass3", &firstr->f, 1);
-
-	/*
-	 * pass 4
-	 * iterate propagating register/variable synchrony
-	 * 	forward until graph is complete
-	 */
-loop2:
-	change = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->f.active = 0;
-	synch(firstr, zbits);
-	if(change)
-		goto loop2;
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass4", &firstr->f, 1);
-
-	/*
-	 * pass 4.5
-	 * move register pseudo-variables into regu.
-	 */
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		r->regu = (r->refbehind.b[0] | r->set.b[0]) & REGBITS;
-
-		r->set.b[0] &= ~REGBITS;
-		r->use1.b[0] &= ~REGBITS;
-		r->use2.b[0] &= ~REGBITS;
-		r->refbehind.b[0] &= ~REGBITS;
-		r->refahead.b[0] &= ~REGBITS;
-		r->calbehind.b[0] &= ~REGBITS;
-		r->calahead.b[0] &= ~REGBITS;
-		r->regdiff.b[0] &= ~REGBITS;
-		r->act.b[0] &= ~REGBITS;
-	}
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass4.5", &firstr->f, 1);
-
-	/*
-	 * pass 5
-	 * isolate regions
-	 * calculate costs (paint1)
-	 */
-	r = firstr;
-	if(r) {
-		for(z=0; z<BITS; z++)
-			bit.b[z] = (r->refahead.b[z] | r->calahead.b[z]) &
-			  ~(externs.b[z] | params.b[z] | addrs.b[z] | consts.b[z]);
-		if(bany(&bit) && !r->f.refset) {
-			// should never happen - all variables are preset
-			if(debug['w'])
-				print("%L: used and not set: %Q\n", r->f.prog->lineno, bit);
-			r->f.refset = 1;
-		}
-	}
-	for(r = firstr; r != R; r = (Reg*)r->f.link)
-		r->act = zbits;
-	rgp = region;
-	nregion = 0;
-	for(r = firstr; r != R; r = (Reg*)r->f.link) {
-		for(z=0; z<BITS; z++)
-			bit.b[z] = r->set.b[z] &
-			  ~(r->refahead.b[z] | r->calahead.b[z] | addrs.b[z]);
-		if(bany(&bit) && !r->f.refset) {
-			if(debug['w'])
-				print("%L: set and not used: %Q\n", r->f.prog->lineno, bit);
-			r->f.refset = 1;
-			excise(&r->f);
-		}
-		for(z=0; z<BITS; z++)
-			bit.b[z] = LOAD(r) & ~(r->act.b[z] | addrs.b[z]);
-		while(bany(&bit)) {
-			i = bnum(bit);
-			rgp->enter = r;
-			rgp->varno = i;
-			change = 0;
-			paint1(r, i);
-			biclr(&bit, i);
-			if(change <= 0)
-				continue;
-			rgp->cost = change;
-			nregion++;
-			if(nregion >= NRGN) {
-				if(debug['R'] && debug['v'])
-					print("too many regions\n");
-				goto brk;
-			}
-			rgp++;
-		}
-	}
-brk:
-	qsort(region, nregion, sizeof(region[0]), rcmp);
-
-	if(debug['R'] && debug['v'])
-		dumpit("pass5", &firstr->f, 1);
-
-	/*
-	 * pass 6
-	 * determine used registers (paint2)
-	 * replace code (paint3)
-	 */
-	rgp = region;
-	if(debug['R'] && debug['v'])
-		print("\nregisterizing\n");
-	for(i=0; i<nregion; i++) {
-		if(debug['R'] && debug['v'])
-			print("region %d: cost %d varno %d enter %lld\n", i, rgp->cost, rgp->varno, rgp->enter->f.prog->pc);
-		bit = blsh(rgp->varno);
-		usedreg = paint2(rgp->enter, rgp->varno, 0);
-		vreg = allreg(usedreg, rgp);
-		if(rgp->regno != 0) {
-			if(debug['R'] && debug['v']) {
-				Var *v;
-
-				v = var + rgp->varno;
-				print("registerize %N+%lld (bit=%2d et=%2E) in %R usedreg=%llx vreg=%llx\n",
-						v->node, v->offset, rgp->varno, v->etype, rgp->regno, usedreg, vreg);
-			}
-			paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
-		}
-		rgp++;
-	}
-
-	/*
-	 * free aux structures. peep allocates new ones.
-	 */
-	for(i=0; i<nvar; i++)
-		var[i].node->opt = nil;
-	flowend(g);
-	firstr = R;
-
-	if(debug['R'] && debug['v']) {
-		// Rebuild flow graph, since we inserted instructions
-		g = flowstart(firstp, sizeof(Reg));
-		firstr = (Reg*)g->start;
-		dumpit("pass6", &firstr->f, 1);
-		flowend(g);
-		firstr = R;
-	}
-
-	/*
-	 * pass 7
-	 * peep-hole on basic block
-	 */
-	if(!debug['R'] || debug['P'])
-		peep(firstp);
-
-	/*
-	 * eliminate nops
-	 */
-	for(p=firstp; p!=P; p=p->link) {
-		while(p->link != P && p->link->as == ANOP)
-			p->link = p->link->link;
-		if(p->to.type == TYPE_BRANCH)
-			while(p->to.u.branch != P && p->to.u.branch->as == ANOP)
-				p->to.u.branch = p->to.u.branch->link;
-	}
-
-	if(debug['R']) {
-		if(ostats.ncvtreg ||
-		   ostats.nspill ||
-		   ostats.ndelmov ||
-		   ostats.nvar ||
-		   0)
-			print("\nstats\n");
-
-		if(ostats.ncvtreg)
-			print("	%4d cvtreg\n", ostats.ncvtreg);
-		if(ostats.nspill)
-			print("	%4d spill\n", ostats.nspill);
-		if(ostats.ndelmov)
-			print("	%4d delmov\n", ostats.ndelmov);
-		if(ostats.nvar)
-			print("	%4d var\n", ostats.nvar);
-
-		memset(&ostats, 0, sizeof(ostats));
-	}
-
-	return;
-}
-
-static void
-walkvardef(Node *n, Reg *r, int active)
-{
-	Reg *r1, *r2;
-	int bn;
-	Var *v;
-	
-	for(r1=r; r1!=R; r1=(Reg*)r1->f.s1) {
-		if(r1->f.active == active)
-			break;
-		r1->f.active = active;
-		if(r1->f.prog->as == AVARKILL && r1->f.prog->to.node == n)
-			break;
-		for(v=n->opt; v!=nil; v=v->nextinnode) {
-			bn = v - var;
-			biset(&r1->act, bn);
-		}
-		if(r1->f.prog->as == ABL)
-			break;
-	}
-
-	for(r2=r; r2!=r1; r2=(Reg*)r2->f.s1)
-		if(r2->f.s2 != nil)
-			walkvardef(n, (Reg*)r2->f.s2, active);
-}
-
-/*
- * add mov b,rn
- * just after r
- */
-void
-addmove(Reg *r, int bn, int rn, int f)
-{
-	Prog *p, *p1, *p2;
-	Adr *a;
-	Var *v;
-
-	p1 = mal(sizeof(*p1));
-	*p1 = zprog;
-	p = r->f.prog;
-	
-	// If there's a stack fixup coming (ADD $n,R1 after BL newproc or BL deferproc),
-	// delay the load until after the fixup.
-	p2 = p->link;
-	if(p2 && p2->as == AADD && p2->to.reg == REGSP && p2->to.type == TYPE_REG)
-		p = p2;
-
-	p1->link = p->link;
-	p->link = p1;
-	p1->lineno = p->lineno;
-
-	v = var + bn;
-
-	a = &p1->to;
-	a->name = v->name;
-	a->node = v->node;
-	a->sym = linksym(v->node->sym);
-	a->offset = v->offset;
-	a->etype = v->etype;
-	a->type = TYPE_MEM;
-	if(a->etype == TARRAY)
-		a->type = TYPE_ADDR;
-	else if(a->sym == nil)
-		a->type = TYPE_CONST;
-
-	if(v->addr)
-		fatal("addmove: shouldn't be doing this %A\n", a);
-
-	switch(v->etype) {
-	default:
-		print("What is this %E\n", v->etype);
-
-	case TINT8:
-		p1->as = AMOVB;
-		break;
-	case TBOOL:
-	case TUINT8:
-//print("movbu %E %d %S\n", v->etype, bn, v->sym);
-		p1->as = AMOVBZ;
-		break;
-	case TINT16:
-		p1->as = AMOVH;
-		break;
-	case TUINT16:
-		p1->as = AMOVHZ;
-		break;
-	case TINT32:
-		p1->as = AMOVW;
-		break;
-	case TUINT32:
-	case TPTR32:
-		p1->as = AMOVWZ;
-		break;
-	case TINT64:
-	case TUINT64:
-	case TPTR64:
-		p1->as = AMOVD;
-		break;
-	case TFLOAT32:
-		p1->as = AFMOVS;
-		break;
-	case TFLOAT64:
-		p1->as = AFMOVD;
-		break;
-	}
-
-	p1->from.type = TYPE_REG;
-	p1->from.reg = rn;
-	if(!f) {
-		p1->from = *a;
-		*a = zprog.from;
-		a->type = TYPE_REG;
-		a->reg = rn;
-		if(v->etype == TUINT8 || v->etype == TBOOL)
-			p1->as = AMOVBZ;
-		if(v->etype == TUINT16)
-			p1->as = AMOVHZ;
-	}
-	if(debug['R'])
-		print("%P\t.a%P\n", p, p1);
-	ostats.nspill++;
-}
-
-static int
-overlap(int64 o1, int w1, int64 o2, int w2)
-{
-	int64 t1, t2;
-
-	t1 = o1+w1;
-	t2 = o2+w2;
-
-	if(!(t1 > o2 && t2 > o1))
-		return 0;
-
-	return 1;
-}
-
-Bits
-mkvar(Reg *r, Adr *a)
-{
-	USED(r);
-	Var *v;
-	int i, t, n, et, z, flag;
-	int64 w;
-	int64 o;
-	Bits bit;
-	Node *node;
-
-	// mark registers used
-	t = a->type;
-	switch(t) {
-	default:
-		print("type %d %d %D\n", t, a->name, a);
-		goto none;
-
-	case TYPE_NONE:
-		goto none;
-
-	case TYPE_BRANCH:
-	case TYPE_CONST:
-	case TYPE_FCONST:
-	case TYPE_SCONST:
-	case TYPE_MEM:
-	case TYPE_ADDR:
-		break;
-
-	case TYPE_REG:
-		if(a->reg != 0) {
-			bit = zbits;
-			bit.b[0] = RtoB(a->reg);
-			return bit;
-		}
-		break;
-	}
-
-	switch(a->name) {
-	default:
-		goto none;
-
-	case NAME_EXTERN:
-	case NAME_STATIC:
-	case NAME_AUTO:
-	case NAME_PARAM:
-		n = a->name;
-		break;
-	}
-
-	node = a->node;
-	if(node == N || node->op != ONAME || node->orig == N)
-		goto none;
-	node = node->orig;
-	if(node->orig != node)
-		fatal("%D: bad node", a);
-	if(node->sym == S || node->sym->name[0] == '.')
-		goto none;
-	et = a->etype;
-	o = a->offset;
-	w = a->width;
-	if(w < 0)
-		fatal("bad width %lld for %D", w, a);
-
-	flag = 0;
-	for(i=0; i<nvar; i++) {
-		v = var+i;
-		if(v->node == node && v->name == n) {
-			if(v->offset == o)
-			if(v->etype == et)
-			if(v->width == w)
-				return blsh(i);
-
-			// if they overlap, disable both
-			if(overlap(v->offset, v->width, o, w)) {
-				v->addr = 1;
-				flag = 1;
-			}
-		}
-	}
-
-	switch(et) {
-	case 0:
-	case TFUNC:
-		goto none;
-	}
-
-	if(nvar >= NVAR) {
-		if(debug['w'] > 1 && node != N)
-			fatal("variable not optimized: %#N", node);
-		
-		// If we're not tracking a word in a variable, mark the rest as
-		// having its address taken, so that we keep the whole thing
-		// live at all calls. otherwise we might optimize away part of
-		// a variable but not all of it.
-		for(i=0; i<nvar; i++) {
-			v = var+i;
-			if(v->node == node)
-				v->addr = 1;
-		}
-		goto none;
-	}
-
-	i = nvar;
-	nvar++;
-	v = var+i;
-	v->offset = o;
-	v->name = n;
-	v->etype = et;
-	v->width = w;
-	v->addr = flag;		// funny punning
-	v->node = node;
-	
-	// node->opt is the head of a linked list
-	// of Vars within the given Node, so that
-	// we can start at a Var and find all the other
-	// Vars in the same Go variable.
-	v->nextinnode = node->opt;
-	node->opt = v;
-
-	bit = blsh(i);
-	if(n == NAME_EXTERN || n == NAME_STATIC)
-		for(z=0; z<BITS; z++)
-			externs.b[z] |= bit.b[z];
-	if(n == NAME_PARAM)
-		for(z=0; z<BITS; z++)
-			params.b[z] |= bit.b[z];
-
-	if(node->class == PPARAM)
-		for(z=0; z<BITS; z++)
-			ivar.b[z] |= bit.b[z];
-	if(node->class == PPARAMOUT)
-		for(z=0; z<BITS; z++)
-			ovar.b[z] |= bit.b[z];
-
-	// Treat values with their address taken as live at calls,
-	// because the garbage collector's liveness analysis in ../gc/plive.c does.
-	// These must be consistent or else we will elide stores and the garbage
-	// collector will see uninitialized data.
-	// The typical case where our own analysis is out of sync is when the
-	// node appears to have its address taken but that code doesn't actually
-	// get generated and therefore doesn't show up as an address being
-	// taken when we analyze the instruction stream.
-	// One instance of this case is when a closure uses the same name as
-	// an outer variable for one of its own variables declared with :=.
-	// The parser flags the outer variable as possibly shared, and therefore
-	// sets addrtaken, even though it ends up not being actually shared.
-	// If we were better about _ elision, _ = &x would suffice too.
-	// The broader := in a closure problem is mentioned in a comment in
-	// closure.c:/^typecheckclosure and dcl.c:/^oldname.
-	if(node->addrtaken)
-		v->addr = 1;
-
-	// Disable registerization for globals, because:
-	// (1) we might panic at any time and we want the recovery code
-	// to see the latest values (issue 1304).
-	// (2) we don't know what pointers might point at them and we want
-	// loads via those pointers to see updated values and vice versa (issue 7995).
-	//
-	// Disable registerization for results if using defer, because the deferred func
-	// might recover and return, causing the current values to be used.
-	if(node->class == PEXTERN || (hasdefer && node->class == PPARAMOUT))
-		v->addr = 1;
-
-	if(debug['R'])
-		print("bit=%2d et=%2E w=%lld+%lld %#N %D flag=%d\n", i, et, o, w, node, a, v->addr);
-	ostats.nvar++;
-
-	return bit;
-
-none:
-	return zbits;
-}
-
-void
-prop(Reg *r, Bits ref, Bits cal)
-{
-	Reg *r1, *r2;
-	int z, i, j;
-	Var *v, *v1;
-
-	for(r1 = r; r1 != R; r1 = (Reg*)r1->f.p1) {
-		for(z=0; z<BITS; z++) {
-			ref.b[z] |= r1->refahead.b[z];
-			if(ref.b[z] != r1->refahead.b[z]) {
-				r1->refahead.b[z] = ref.b[z];
-				change++;
-			}
-			cal.b[z] |= r1->calahead.b[z];
-			if(cal.b[z] != r1->calahead.b[z]) {
-				r1->calahead.b[z] = cal.b[z];
-				change++;
-			}
-		}
-		switch(r1->f.prog->as) {
-		case ABL:
-			if(noreturn(r1->f.prog))
-				break;
-
-			// Mark all input variables (ivar) as used, because that's what the
-			// liveness bitmaps say. The liveness bitmaps say that so that a
-			// panic will not show stale values in the parameter dump.
-			// Mark variables with a recent VARDEF (r1->act) as used,
-			// so that the optimizer flushes initializations to memory,
-			// so that if a garbage collection happens during this CALL,
-			// the collector will see initialized memory. Again this is to
-			// match what the liveness bitmaps say.
-			for(z=0; z<BITS; z++) {
-				cal.b[z] |= ref.b[z] | externs.b[z] | ivar.b[z] | r1->act.b[z];
-				ref.b[z] = 0;
-			}
-			
-			// cal.b is the current approximation of what's live across the call.
-			// Every bit in cal.b is a single stack word. For each such word,
-			// find all the other tracked stack words in the same Go variable
-			// (struct/slice/string/interface) and mark them live too.
-			// This is necessary because the liveness analysis for the garbage
-			// collector works at variable granularity, not at word granularity.
-			// It is fundamental for slice/string/interface: the garbage collector
-			// needs the whole value, not just some of the words, in order to
-			// interpret the other bits correctly. Specifically, slice needs a consistent
-			// ptr and cap, string needs a consistent ptr and len, and interface
-			// needs a consistent type word and data word.
-			for(z=0; z<BITS; z++) {
-				if(cal.b[z] == 0)
-					continue;
-				for(i=0; i<64; i++) {
-					if(z*64+i >= nvar || ((cal.b[z]>>i)&1) == 0)
-						continue;
-					v = var+z*64+i;
-					if(v->node->opt == nil) // v represents fixed register, not Go variable
-						continue;
-
-					// v->node->opt is the head of a linked list of Vars
-					// corresponding to tracked words from the Go variable v->node.
-					// Walk the list and set all the bits.
-					// For a large struct this could end up being quadratic:
-					// after the first setting, the outer loop (for z, i) would see a 1 bit
-					// for all of the remaining words in the struct, and for each such
-					// word would go through and turn on all the bits again.
-					// To avoid the quadratic behavior, we only turn on the bits if
-					// v is the head of the list or if the head's bit is not yet turned on.
-					// This will set the bits at most twice, keeping the overall loop linear.
-					v1 = v->node->opt;
-					j = v1 - var;
-					if(v == v1 || !btest(&cal, j)) {
-						for(; v1 != nil; v1 = v1->nextinnode) {
-							j = v1 - var;
-							biset(&cal, j);
-						}
-					}
-				}
-			}
-			break;
-
-		case ATEXT:
-			for(z=0; z<BITS; z++) {
-				cal.b[z] = 0;
-				ref.b[z] = 0;
-			}
-			break;
-
-		case ARET:
-			for(z=0; z<BITS; z++) {
-				cal.b[z] = externs.b[z] | ovar.b[z];
-				ref.b[z] = 0;
-			}
-			break;
-		}
-		for(z=0; z<BITS; z++) {
-			ref.b[z] = (ref.b[z] & ~r1->set.b[z]) |
-				r1->use1.b[z] | r1->use2.b[z];
-			cal.b[z] &= ~(r1->set.b[z] | r1->use1.b[z] | r1->use2.b[z]);
-			r1->refbehind.b[z] = ref.b[z];
-			r1->calbehind.b[z] = cal.b[z];
-		}
-		if(r1->f.active)
-			break;
-		r1->f.active = 1;
-	}
-	for(; r != r1; r = (Reg*)r->f.p1)
-		for(r2 = (Reg*)r->f.p2; r2 != R; r2 = (Reg*)r2->f.p2link)
-			prop(r2, r->refbehind, r->calbehind);
-}
-
-void
-synch(Reg *r, Bits dif)
-{
-	Reg *r1;
-	int z;
-
-	for(r1 = r; r1 != R; r1 = (Reg*)r1->f.s1) {
-		for(z=0; z<BITS; z++) {
-			dif.b[z] = (dif.b[z] &
-				~(~r1->refbehind.b[z] & r1->refahead.b[z])) |
-					r1->set.b[z] | r1->regdiff.b[z];
-			if(dif.b[z] != r1->regdiff.b[z]) {
-				r1->regdiff.b[z] = dif.b[z];
-				change++;
-			}
-		}
-		if(r1->f.active)
-			break;
-		r1->f.active = 1;
-		for(z=0; z<BITS; z++)
-			dif.b[z] &= ~(~r1->calbehind.b[z] & r1->calahead.b[z]);
-		if(r1->f.s2 != nil)
-			synch((Reg*)r1->f.s2, dif);
-	}
+	return regbits;
 }
 
 uint64
-allreg(uint64 b, Rgn *r)
+doregbits(int r)
 {
-	Var *v;
-	int i;
-
-	v = var + r->varno;
-	r->regno = 0;
-	switch(v->etype) {
-
-	default:
-		fatal("unknown etype %d/%E", bitno(b), v->etype);
-		break;
-
-	case TINT8:
-	case TUINT8:
-	case TINT16:
-	case TUINT16:
-	case TINT32:
-	case TUINT32:
-	case TINT64:
-	case TUINT64:
-	case TINT:
-	case TUINT:
-	case TUINTPTR:
-	case TBOOL:
-	case TPTR32:
-	case TPTR64:
-		i = BtoR(~b);
-		if(i && r->cost > 0) {
-			r->regno = i;
-			return RtoB(i);
-		}
-		break;
-
-	case TFLOAT32:
-	case TFLOAT64:
-		i = BtoF(~b);
-		if(i && r->cost > 0) {
-			r->regno = i;
-			return RtoB(i);
-		}
-		break;
-	}
+	USED(r);
 	return 0;
 }
 
-void
-paint1(Reg *r, int bn)
-{
-	Reg *r1;
-	int z;
-	uint64 bb;
-
-	z = bn/64;
-	bb = 1LL<<(bn%64);
-	if(r->act.b[z] & bb)
-		return;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(r1->act.b[z] & bb)
-			break;
-		r = r1;
-	}
-
-	if(LOAD(r) & ~(r->set.b[z]&~(r->use1.b[z]|r->use2.b[z])) & bb) {
-		change -= CLOAD * r->f.loop;
-	}
-	for(;;) {
-		r->act.b[z] |= bb;
-
-		if(r->f.prog->as != ANOP) { // don't give credit for NOPs
-			if(r->use1.b[z] & bb)
-				change += CREF * r->f.loop;
-			if((r->use2.b[z]|r->set.b[z]) & bb)
-				change += CREF * r->f.loop;
-		}
-
-		if(STORE(r) & r->regdiff.b[z] & bb) {
-			change -= CLOAD * r->f.loop;
-		}
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					paint1(r1, bn);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				paint1(r1, bn);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(r->act.b[z] & bb)
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-}
-
-uint64
-paint2(Reg *r, int bn, int depth)
-{
-	Reg *r1;
-	int z;
-	uint64 bb, vreg;
-
-	z = bn/64;
-	bb = 1LL << (bn%64);
-	vreg = regbits;
-	if(!(r->act.b[z] & bb))
-		return vreg;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(!(r1->act.b[z] & bb))
-			break;
-		r = r1;
-	}
-	for(;;) {
-		if(debug['R'] && debug['v'])
-			print("  paint2 %d %P\n", depth, r->f.prog);
-
-		r->act.b[z] &= ~bb;
-
-		vreg |= r->regu;
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					vreg |= paint2(r1, bn, depth+1);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				vreg |= paint2(r1, bn, depth+1);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(!(r->act.b[z] & bb))
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-	return vreg;
-}
-
-void
-paint3(Reg *r, int bn, uint64 rb, int rn)
-{
-	Reg *r1;
-	Prog *p;
-	int z;
-	uint64 bb;
-
-	z = bn/64;
-	bb = 1LL << (bn%64);
-	if(r->act.b[z] & bb)
-		return;
-	for(;;) {
-		if(!(r->refbehind.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.p1;
-		if(r1 == R)
-			break;
-		if(!(r1->refahead.b[z] & bb))
-			break;
-		if(r1->act.b[z] & bb)
-			break;
-		r = r1;
-	}
-
-	if(LOAD(r) & ~(r->set.b[z] & ~(r->use1.b[z]|r->use2.b[z])) & bb)
-		addmove(r, bn, rn, 0);
-	for(;;) {
-		r->act.b[z] |= bb;
-		p = r->f.prog;
-
-		if(r->use1.b[z] & bb) {
-			if(debug['R'] && debug['v'])
-				print("%P", p);
-			addreg(&p->from, rn);
-			if(debug['R'] && debug['v'])
-				print(" ===change== %P\n", p);
-		}
-		if((r->use2.b[z]|r->set.b[z]) & bb) {
-			if(debug['R'] && debug['v'])
-				print("%P", p);
-			addreg(&p->to, rn);
-			if(debug['R'] && debug['v'])
-				print(" ===change== %P\n", p);
-		}
-
-		if(STORE(r) & r->regdiff.b[z] & bb)
-			addmove(r, bn, rn, 1);
-		r->regu |= rb;
-
-		if(r->refbehind.b[z] & bb)
-			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
-				if(r1->refahead.b[z] & bb)
-					paint3(r1, bn, rb, rn);
-
-		if(!(r->refahead.b[z] & bb))
-			break;
-		r1 = (Reg*)r->f.s2;
-		if(r1 != R)
-			if(r1->refbehind.b[z] & bb)
-				paint3(r1, bn, rb, rn);
-		r = (Reg*)r->f.s1;
-		if(r == R)
-			break;
-		if(r->act.b[z] & bb)
-			break;
-		if(!(r->refbehind.b[z] & bb))
-			break;
-	}
-}
-
-void
-addreg(Adr *a, int rn)
-{
-	a->sym = nil;
-	a->node = nil;
-	a->name = NAME_NONE;
-	a->type = TYPE_REG;
-	a->reg = rn;
-
-	ostats.ncvtreg++;
-}
-
 /*
  * track register variables including external registers:
  *	bit	reg
@@ -1241,78 +170,3 @@ BtoF(uint64 b)
 		return 0;
 	return bitno(b) + REG_F0;
 }
-
-void
-dumpone(Flow *f, int isreg)
-{
-	int z;
-	Bits bit;
-	Reg *r;
-
-	print("%d:%P", f->loop, f->prog);
-	if(isreg) {	
-		r = (Reg*)f;
-		for(z=0; z<BITS; z++)
-			bit.b[z] =
-				r->set.b[z] |
-				r->use1.b[z] |
-				r->use2.b[z] |
-				r->refbehind.b[z] |
-				r->refahead.b[z] |
-				r->calbehind.b[z] |
-				r->calahead.b[z] |
-				r->regdiff.b[z] |
-				r->act.b[z] |
-					0;
-		if(bany(&bit)) {
-			print("\t");
-			if(bany(&r->set))
-				print(" s:%Q", r->set);
-			if(bany(&r->use1))
-				print(" u1:%Q", r->use1);
-			if(bany(&r->use2))
-				print(" u2:%Q", r->use2);
-			if(bany(&r->refbehind))
-				print(" rb:%Q ", r->refbehind);
-			if(bany(&r->refahead))
-				print(" ra:%Q ", r->refahead);
-			if(bany(&r->calbehind))
-				print(" cb:%Q ", r->calbehind);
-			if(bany(&r->calahead))
-				print(" ca:%Q ", r->calahead);
-			if(bany(&r->regdiff))
-				print(" d:%Q ", r->regdiff);
-			if(bany(&r->act))
-				print(" a:%Q ", r->act);
-		}
-	}
-	print("\n");
-}
-
-
-void
-dumpit(char *str, Flow *r0, int isreg)
-{
-	Flow *r, *r1;
-
-	print("\n%s\n", str);
-	for(r = r0; r != nil; r = r->link) {
-		dumpone(r, isreg);
-		r1 = r->p2;
-		if(r1 != nil) {
-			print("	pred:");
-			for(; r1 != nil; r1 = r1->p2link)
-				print(" %.4ud", (int)r1->prog->pc);
-			print("\n");
-		}
-		// Print successors if it's not just the next one
-		if(r->s1 != r->link || r->s2 != nil) {
-			print("	succ:");
-			if(r->s1 != nil)
-				print(" %.4ud", (int)r->s1->prog->pc);
-			if(r->s2 != nil)
-				print(" %.4ud", (int)r->s2->prog->pc);
-			print("\n");
-		}
-	}
-}
diff --git a/src/cmd/9l/9.out.h b/src/cmd/9l/9.out.h
index b99e7e6f4d..6e95698db9 100644
--- a/src/cmd/9l/9.out.h
+++ b/src/cmd/9l/9.out.h
@@ -204,7 +204,7 @@ enum
 	REGRT1		= REG_R3,	/* reserved for runtime, duffzero and duffcopy */
 	REGRT2		= REG_R4,	/* reserved for runtime, duffcopy */
 	REGMIN		= REG_R7,	/* register variables allocated from here to REGMAX */
-	REGENV		= REG_R11,	/* environment for closures */
+	REGCTXT		= REG_R11,	/* context for closures */
 	REGTLS		= REG_R13,	/* C ABI TLS base pointer */
 	REGMAX		= REG_R27,
 	REGEXT		= REG_R30,	/* external registers allocated from here down */
diff --git a/src/cmd/9l/asm.c b/src/cmd/9l/asm.c
index 391f9562cf..01e17aaa09 100644
--- a/src/cmd/9l/asm.c
+++ b/src/cmd/9l/asm.c
@@ -35,15 +35,6 @@
 #include	"../ld/elf.h"
 #include	"../ld/dwarf.h"
 
-
-// TODO(austin): ABI v1 uses /usr/lib/ld.so.1
-char linuxdynld[] = "/lib64/ld64.so.1";
-char freebsddynld[] = "XXX";
-char openbsddynld[] = "XXX";
-char netbsddynld[] = "XXX";
-char dragonflydynld[] = "XXX";
-char solarisdynld[] = "XXX";
-
 static int
 needlib(char *name)
 {
@@ -64,8 +55,6 @@ needlib(char *name)
 	return 0;
 }
 
-int	nelfsym = 1;
-
 static void	gencallstub(int abicase, LSym *stub, LSym *targ);
 static void	addpltsym(Link*, LSym*);
 static LSym*	ensureglinkresolver(void);
@@ -129,7 +118,7 @@ gentext(void)
 	// This assumes "case 1" from the ABI, where the caller needs
 	// us to save and restore the TOC pointer.
 	pprevtextp = &ctxt->textp;
-	for(s=*pprevtextp; s!=S; pprevtextp=&s->next, s=*pprevtextp) {
+	for(s=*pprevtextp; s!=nil; pprevtextp=&s->next, s=*pprevtextp) {
 		for(r=s->r; r<s->r+s->nr; r++) {
 			if(!(r->type == 256 + R_PPC64_REL24 &&
 			     r->sym->type == SDYNIMPORT))
@@ -797,14 +786,14 @@ asmb(void)
 	switch(HEADTYPE) {
 	default:
 	case Hplan9:	/* plan 9 */
-		LPUT(0x647);			/* magic */
-		LPUT(segtext.filelen);			/* sizes */
-		LPUT(segdata.filelen);
-		LPUT(segdata.len - segdata.filelen);
-		LPUT(symsize);			/* nsyms */
-		LPUT(entryvalue());		/* va of entry */
-		LPUT(0L);
-		LPUT(lcsize);
+		thearch.lput(0x647);			/* magic */
+		thearch.lput(segtext.filelen);			/* sizes */
+		thearch.lput(segdata.filelen);
+		thearch.lput(segdata.len - segdata.filelen);
+		thearch.lput(symsize);			/* nsyms */
+		thearch.lput(entryvalue());		/* va of entry */
+		thearch.lput(0L);
+		thearch.lput(lcsize);
 		break;
 	case Hlinux:
 	case Hfreebsd:
@@ -824,18 +813,3 @@ asmb(void)
 		print("total=%lld\n", segtext.filelen+segdata.len+symsize+lcsize);
 	}
 }
-
-vlong
-rnd(vlong v, int32 r)
-{
-	vlong c;
-
-	if(r <= 0)
-		return v;
-	v += r - 1;
-	c = v % r;
-	if(c < 0)
-		c += r;
-	v -= c;
-	return v;
-}
diff --git a/src/cmd/9l/l.h b/src/cmd/9l/l.h
index 9d8a4fae2c..a634cf80b4 100644
--- a/src/cmd/9l/l.h
+++ b/src/cmd/9l/l.h
@@ -44,55 +44,21 @@ enum
 	IntSize = 8,
 	RegSize = 8,
 	MaxAlign = 32,	// max data alignment
-	FuncAlign = 8
-};
-
-#define	P		((Prog*)0)
-#define	S		((LSym*)0)
-
-enum
-{
-	FPCHIP		= 1,
-	STRINGSZ	= 200,
-	MAXHIST		= 20,				/* limit of path elements for history symbols */
-	DATBLK		= 1024,
-	NHASH		= 10007,
-	NHUNK		= 100000,
-	MINSIZ		= 64,
-	NENT		= 100,
-	NSCHED		= 20,
+	FuncAlign = 8,
 	MINLC		= 4,
-
-	Roffset	= 22,		/* no. bits for offset in relocation address */
-	Rindex	= 10		/* no. bits for index in relocation address */
 };
 
-EXTERN	int32	autosize;
-EXTERN	LSym*	datap;
-EXTERN	int	debug[128];
-EXTERN	int32	lcsize;
-EXTERN	char	literal[32];
-EXTERN	int	nerrors;
-EXTERN	vlong	instoffset;
-EXTERN	char*	rpath;
-EXTERN	vlong	pc;
-EXTERN	int32	symsize;
-EXTERN	int32	staticgen;
-EXTERN	Prog*	lastp;
-EXTERN	vlong	textsize;
-
-void	asmb(void);
 void	adddynlib(char *lib);
 void	adddynrel(LSym *s, Reloc *r);
+void	adddynrela(LSym *rela, LSym *s, Reloc *r);
 void	adddynsym(Link *ctxt, LSym *s);
 int	archreloc(Reloc *r, LSym *s, vlong *val);
 vlong	archrelocvariant(Reloc *r, LSym *s, vlong t);
+void	asmb(void);
+int	elfreloc1(Reloc *r, vlong sectoff);
+void	elfsetupplt(void);
 void	listinit(void);
-vlong	rnd(vlong, int32);
-
-#define	LPUT(a)	(ctxt->arch->endian == BigEndian ? lputb(a):lputl(a))
-#define	WPUT(a)	(ctxt->arch->endian == BigEndian ? wputb(a):wputl(a))
-#define	VPUT(a)	(ctxt->arch->endian == BigEndian ? vputb(a):vputl(a))
+int	machoreloc1(Reloc *r, vlong sectoff);
 
 /* Used by ../ld/dwarf.c */
 enum
diff --git a/src/cmd/9l/obj.c b/src/cmd/9l/obj.c
index 77f665e5a2..17dd5caacd 100644
--- a/src/cmd/9l/obj.c
+++ b/src/cmd/9l/obj.c
@@ -36,8 +36,12 @@
 #include	"../ld/dwarf.h"
 #include	<ar.h>
 
-char *thestring = "ppc64";
-LinkArch *thelinkarch;
+void
+main(int argc, char **argv)
+{
+	linkarchinit();
+	ldmain(argc, argv);
+}
 
 void
 linkarchinit(void)
@@ -47,6 +51,45 @@ linkarchinit(void)
 		thelinkarch = &linkppc64le;
 	else
 		thelinkarch = &linkppc64;
+
+	thearch.thechar = thechar;
+	thearch.ptrsize = thelinkarch->ptrsize;
+	thearch.intsize = thelinkarch->ptrsize;
+	thearch.regsize = thelinkarch->regsize;
+	thearch.funcalign = FuncAlign;
+	thearch.maxalign = MaxAlign;
+	thearch.minlc = MINLC;
+	thearch.dwarfregsp = DWARFREGSP;
+
+	thearch.adddynlib = adddynlib;
+	thearch.adddynrel = adddynrel;
+	thearch.adddynsym = adddynsym;
+	thearch.archinit = archinit;
+	thearch.archreloc = archreloc;
+	thearch.archrelocvariant = archrelocvariant;
+	thearch.asmb = asmb;
+	thearch.elfreloc1 = elfreloc1;
+	thearch.elfsetupplt = elfsetupplt;
+	thearch.gentext = gentext;
+	thearch.listinit = listinit;
+	thearch.machoreloc1 = machoreloc1;
+	if(thelinkarch == &linkppc64le) {
+		thearch.lput = lputl;
+		thearch.wput = wputl;
+		thearch.vput = vputl;
+	} else {
+		thearch.lput = lputb;
+		thearch.wput = wputb;
+		thearch.vput = vputb;
+	}
+
+	// TODO(austin): ABI v1 uses /usr/lib/ld.so.1
+	thearch.linuxdynld = "/lib64/ld64.so.1";
+	thearch.freebsddynld = "XXX";
+	thearch.openbsddynld = "XXX";
+	thearch.netbsddynld = "XXX";
+	thearch.dragonflydynld = "XXX";
+	thearch.solarisdynld = "XXX";
 }
 
 void
diff --git a/src/cmd/api/goapi.go b/src/cmd/api/goapi.go
index 1519d96ccc..85988e3bb7 100644
--- a/src/cmd/api/goapi.go
+++ b/src/cmd/api/goapi.go
@@ -107,22 +107,12 @@ func setContexts() {
 	}
 }
 
-var (
-	internalPkg = regexp.MustCompile(`(^|/)internal($|/)`)
-	hashRx      = regexp.MustCompile(`^[0-9a-f]{7,40}$`)
-)
-
-func isDevelVersion(v string) bool {
-	if strings.Contains(v, "devel") {
-		return true
-	}
-	return hashRx.MatchString(v)
-}
+var internalPkg = regexp.MustCompile(`(^|/)internal($|/)`)
 
 func main() {
 	flag.Parse()
 
-	if v := runtime.Version(); !strings.Contains(v, "weekly") && !isDevelVersion(v) {
+	if !strings.Contains(runtime.Version(), "weekly") && !strings.Contains(runtime.Version(), "devel") {
 		if *nextFile != "" {
 			fmt.Printf("Go version is %q, ignoring -next %s\n", runtime.Version(), *nextFile)
 			*nextFile = ""
@@ -293,7 +283,7 @@ func compareAPI(w io.Writer, features, required, optional, exception []string) (
 				delete(optionalSet, newFeature)
 			} else {
 				fmt.Fprintf(w, "+%s\n", newFeature)
-				if !*allowNew || !isDevelVersion(runtime.Version()) {
+				if !*allowNew || !strings.Contains(runtime.Version(), "devel") {
 					ok = false // we're in lock-down mode for next release
 				}
 			}
diff --git a/src/cmd/cgo/out.go b/src/cmd/cgo/out.go
index a4bd5579d3..346ae94546 100644
--- a/src/cmd/cgo/out.go
+++ b/src/cmd/cgo/out.go
@@ -636,7 +636,7 @@ func (p *Package) writeExports(fgo2, fm io.Writer) {
 	fgcc := creat(*objDir + "_cgo_export.c")
 	fgcch := creat(*objDir + "_cgo_export.h")
 
-	fmt.Fprintf(fgcch, "/* Created by cgo - DO NOT EDIT. */\n")
+	fmt.Fprintf(fgcch, "/* Created by cgo - DO NOT EDIT. */\n/*  This file is arch-specific.  */\n")
 	fmt.Fprintf(fgcch, "%s\n", p.Preamble)
 	fmt.Fprintf(fgcch, "%s\n", p.gccExportHeaderProlog())
 
@@ -1310,6 +1310,10 @@ typedef double GoFloat64;
 typedef __complex float GoComplex64;
 typedef __complex double GoComplex128;
 
+// static assertion to make sure the file is being used on architecture
+// at least with matching size of GoInt.
+typedef char _check_for_GOINTBITS_bit_pointer_matching_GoInt[sizeof(void*)==GOINTBITS/8 ? 1:-1];
+
 typedef struct { char *p; GoInt n; } GoString;
 typedef void *GoMap;
 typedef void *GoChan;
diff --git a/src/cmd/dist/build.go b/src/cmd/dist/build.go
index 5cfc47ce84..81249d4645 100644
--- a/src/cmd/dist/build.go
+++ b/src/cmd/dist/build.go
@@ -32,7 +32,6 @@ var (
 	workdir          string
 	tooldir          string
 	gochar           string
-	goversion        string
 	oldgoos          string
 	oldgoarch        string
 	oldgochar        string
@@ -225,8 +224,6 @@ func xinit() {
 	os.Setenv("LANG", "C")
 	os.Setenv("LANGUAGE", "en_US.UTF8")
 
-	goversion = findgoversion()
-
 	workdir = xworkdir()
 	xatexit(rmworkdir)
 
@@ -428,6 +425,7 @@ func setup() {
 	}
 
 	// For release, make sure excluded things are excluded.
+	goversion := findgoversion()
 	if strings.HasPrefix(goversion, "release.") || (strings.HasPrefix(goversion, "go") && !strings.Contains(goversion, "beta")) {
 		for _, dir := range unreleased {
 			if p := pathf("%s/%s", goroot, dir); isdir(p) {
@@ -533,16 +531,16 @@ var deptab = []struct {
 		"$GOROOT/pkg/obj/${GOHOSTOS}_$GOHOSTARCH/libgc.a",
 	}},
 	{"cmd/5l", []string{
-		"../ld/*",
+		"$GOROOT/pkg/obj/${GOHOSTOS}_$GOHOSTARCH/libld.a",
 	}},
 	{"cmd/6l", []string{
-		"../ld/*",
+		"$GOROOT/pkg/obj/${GOHOSTOS}_$GOHOSTARCH/libld.a",
 	}},
 	{"cmd/8l", []string{
-		"../ld/*",
+		"$GOROOT/pkg/obj/${GOHOSTOS}_$GOHOSTARCH/libld.a",
 	}},
 	{"cmd/9l", []string{
-		"../ld/*",
+		"$GOROOT/pkg/obj/${GOHOSTOS}_$GOHOSTARCH/libld.a",
 	}},
 	{"cmd/go", []string{
 		"zdefaultcc.go",
@@ -619,7 +617,7 @@ func install(dir string) {
 		}
 		// disable word wrapping in error messages
 		gccargs = append(gccargs, "-fmessage-length=0")
-		if gohostos == "darwin" {
+		if gohostos == "darwin" && gohostarch != "arm" {
 			// golang.org/issue/5261
 			gccargs = append(gccargs, "-mmacosx-version-min=10.6")
 		}
@@ -634,7 +632,7 @@ func install(dir string) {
 
 	// Legacy C exceptions.
 	switch dir {
-	case "lib9", "libbio", "liblink", "cmd/gc":
+	case "lib9", "libbio", "liblink", "cmd/gc", "cmd/ld":
 		islib = true
 		isgo = false
 	case "cmd/5a", "cmd/5g", "cmd/5l",
@@ -914,7 +912,7 @@ func install(dir string) {
 					"-D", fmt.Sprintf("GOHOSTOS=%q", gohostos),
 					"-D", fmt.Sprintf("GOHOSTARCH=%q", gohostarch),
 					"-D", fmt.Sprintf("GOROOT=%q", goroot_final),
-					"-D", fmt.Sprintf("GOVERSION=%q", goversion),
+					"-D", fmt.Sprintf("GOVERSION=%q", findgoversion()),
 					"-D", fmt.Sprintf("GOARM=%q", goarm),
 					"-D", fmt.Sprintf("GO386=%q", go386),
 					"-D", fmt.Sprintf("GO_EXTLINK_ENABLED=%q", goextlinkenabled),
@@ -1117,6 +1115,7 @@ var buildorder = []string{
 	"liblink",
 
 	"cmd/gc",  // must be before g
+	"cmd/ld",  // must be before l
 	"cmd/%sl", // must be before a, g
 	"cmd/%sa",
 	"cmd/%sg",
@@ -1505,5 +1504,5 @@ func cmdbanner() {
 // Version prints the Go version.
 func cmdversion() {
 	xflagparse(0)
-	xprintf("%s\n", goversion)
+	xprintf("%s\n", findgoversion())
 }
diff --git a/src/cmd/dist/buildgc.go b/src/cmd/dist/buildgc.go
index bc1898e04e..b59aa5b5a4 100644
--- a/src/cmd/dist/buildgc.go
+++ b/src/cmd/dist/buildgc.go
@@ -87,7 +87,7 @@ func mkanames(dir, file string) {
 				line = line[:i]
 			}
 			line = line[2:]
-			fmt.Fprintf(&out, "\t\"%s\",\n", line)
+			fmt.Fprintf(&out, "\t\"%s\",\n", strings.TrimSpace(line))
 		}
 	}
 	fmt.Fprintf(&out, "};\n")
diff --git a/src/cmd/dist/buildruntime.go b/src/cmd/dist/buildruntime.go
index ff2a489b55..1f8276468e 100644
--- a/src/cmd/dist/buildruntime.go
+++ b/src/cmd/dist/buildruntime.go
@@ -28,7 +28,7 @@ func mkzversion(dir, file string) {
 			"const defaultGoroot = `%s`\n"+
 			"const theVersion = `%s`\n"+
 			"const goexperiment = `%s`\n"+
-			"var buildVersion = theVersion\n", goroot_final, goversion, os.Getenv("GOEXPERIMENT"))
+			"var buildVersion = theVersion\n", goroot_final, findgoversion(), os.Getenv("GOEXPERIMENT"))
 
 	writefile(out, file, 0)
 }
@@ -56,7 +56,7 @@ func mkzbootstrap(file string) {
 			"const defaultGOARCH = `%s`\n"+
 			"const version = `%s`\n"+
 			"const goexperiment = `%s`\n",
-		goroot_final, goarm, gohostos, gohostarch, goversion, os.Getenv("GOEXPERIMENT"))
+		goroot_final, goarm, gohostos, gohostarch, findgoversion(), os.Getenv("GOEXPERIMENT"))
 
 	writefile(out, file, 0)
 }
diff --git a/src/cmd/dist/util.go b/src/cmd/dist/util.go
index decbb0ffd2..df3e8154e2 100644
--- a/src/cmd/dist/util.go
+++ b/src/cmd/dist/util.go
@@ -414,7 +414,7 @@ func main() {
 
 	if gohostarch == "" {
 		// Default Unix system.
-		out := run("", CheckExit, "uname", "-m")
+		out := run("", CheckExit, "uname", "-m", "-v")
 		switch {
 		case strings.Contains(out, "x86_64"), strings.Contains(out, "amd64"):
 			gohostarch = "amd64"
@@ -426,6 +426,10 @@ func main() {
 			gohostarch = "ppc64le"
 		case strings.Contains(out, "ppc64"):
 			gohostarch = "ppc64"
+		case gohostos == "darwin":
+			if strings.Contains(out, "RELEASE_ARM_") {
+				gohostarch = "arm"
+			}
 		default:
 			fatal("unknown architecture: %s", out)
 		}
@@ -491,6 +495,12 @@ func xgetgoarm() string {
 		// NaCl guarantees VFPv3 and is always cross-compiled.
 		return "7"
 	}
+	if goos == "darwin" {
+		// Assume all darwin/arm devices are have VFPv3. This
+		// port is also mostly cross-compiled, so it makes little
+		// sense to auto-detect the setting.
+		return "7"
+	}
 	if gohostarch != "arm" || goos != gohostos {
 		// Conservative default for cross-compilation.
 		return "5"
@@ -499,24 +509,38 @@ func xgetgoarm() string {
 		// FreeBSD has broken VFP support.
 		return "5"
 	}
-	if xtryexecfunc(useVFPv3) {
+	if goos != "linux" {
+		// All other arm platforms that we support
+		// require ARMv7.
 		return "7"
 	}
-	if xtryexecfunc(useVFPv1) {
-		return "6"
-	}
-	return "5"
-}
+	cpuinfo := readfile("/proc/cpuinfo")
+	goarm := "5"
+	for _, line := range splitlines(cpuinfo) {
+		line := strings.SplitN(line, ":", 2)
+		if len(line) < 2 {
+			continue
+		}
+		if strings.TrimSpace(line[0]) != "Features" {
+			continue
+		}
+		features := splitfields(line[1])
+		sort.Strings(features) // so vfpv3 sorts after vfp
 
-func xtryexecfunc(f func()) bool {
-	// TODO(rsc): Implement.
-	// The C cmd/dist used this to test whether certain assembly
-	// sequences could be executed properly. It used signals and
-	// timers and sigsetjmp, which is basically not possible in Go.
-	// We probably have to invoke ourselves as a subprocess instead,
-	// to contain the fault/timeout.
-	return false
+		// Infer GOARM value from the vfp features available
+		// on this host. Values of GOARM detected are:
+		// 5: no vfp support was found
+		// 6: vfp (v1) support was detected, but no higher
+		// 7: vfpv3 support was detected.
+		// This matches the assertions in runtime.checkarm.
+		for _, f := range features {
+			switch f {
+			case "vfp":
+				goarm = "6"
+			case "vfpv3":
+				goarm = "7"
+			}
+		}
+	}
+	return goarm
 }
-
-func useVFPv1()
-func useVFPv3()
diff --git a/src/cmd/dist/vfp_arm.s b/src/cmd/dist/vfp_arm.s
deleted file mode 100644
index 39052dbb30..0000000000
--- a/src/cmd/dist/vfp_arm.s
+++ /dev/null
@@ -1,15 +0,0 @@
-// Copyright 2015 The Go Authors.  All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#include "textflag.h"
-
-// try to run "vmov.f64 d0, d0" instruction
-TEXT ·useVFPv1(SB),NOSPLIT,$0
-	WORD $0xeeb00b40	// vmov.f64 d0, d0
-	RET
-
-// try to run VFPv3-only "vmov.f64 d0, #112" instruction
-TEXT ·useVFPv3(SB),NOSPLIT,$0
-	WORD $0xeeb70b00	// vmov.f64 d0, #112
-	RET
diff --git a/src/cmd/dist/vfp_default.s b/src/cmd/dist/vfp_default.s
deleted file mode 100644
index c795b357f7..0000000000
--- a/src/cmd/dist/vfp_default.s
+++ /dev/null
@@ -1,14 +0,0 @@
-// Copyright 2015 The Go Authors.  All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-// +build !arm
-
-#include "textflag.h"
-
-TEXT ·useVFPv1(SB),NOSPLIT,$0
-	RET
-
-TEXT ·useVFPv3(SB),NOSPLIT,$0
-	RET
-
diff --git a/src/cmd/gc/align.c b/src/cmd/gc/align.c
index 26d72e87fc..59ff0abf73 100644
--- a/src/cmd/gc/align.c
+++ b/src/cmd/gc/align.c
@@ -35,7 +35,7 @@ offmod(Type *t)
 			fatal("offmod: not TFIELD: %lT", f);
 		f->width = o;
 		o += widthptr;
-		if(o >= arch.MAXWIDTH) {
+		if(o >= thearch.MAXWIDTH) {
 			yyerror("interface too large");
 			o = widthptr;
 		}
@@ -86,7 +86,7 @@ widstruct(Type *errtype, Type *t, vlong o, int flag)
 		if(w == 0)
 			lastzero = o;
 		o += w;
-		if(o >= arch.MAXWIDTH) {
+		if(o >= thearch.MAXWIDTH) {
 			yyerror("type %lT too large", errtype);
 			o = 8;  // small but nonzero
 		}
@@ -260,7 +260,7 @@ dowidth(Type *t)
 
 			dowidth(t->type);
 			if(t->type->width != 0) {
-				cap = (arch.MAXWIDTH-1) / t->type->width;
+				cap = (thearch.MAXWIDTH-1) / t->type->width;
 				if(t->bound > cap)
 					yyerror("type %lT larger than address space", t);
 			}
@@ -613,15 +613,15 @@ typeinit(void)
 	simtype[TFUNC] = tptr;
 	simtype[TUNSAFEPTR] = tptr;
 
-	/* pick up the backend arch.typedefs */
-	for(i=0; arch.typedefs[i].name; i++) {
-		s = lookup(arch.typedefs[i].name);
-		s1 = pkglookup(arch.typedefs[i].name, builtinpkg);
+	/* pick up the backend thearch.typedefs */
+	for(i=0; thearch.typedefs[i].name; i++) {
+		s = lookup(thearch.typedefs[i].name);
+		s1 = pkglookup(thearch.typedefs[i].name, builtinpkg);
 
-		etype = arch.typedefs[i].etype;
+		etype = thearch.typedefs[i].etype;
 		if(etype < 0 || etype >= nelem(types))
 			fatal("typeinit: %s bad etype", s->name);
-		sameas = arch.typedefs[i].sameas;
+		sameas = thearch.typedefs[i].sameas;
 		if(sameas < 0 || sameas >= nelem(types))
 			fatal("typeinit: %s bad sameas", s->name);
 		simtype[etype] = sameas;
diff --git a/src/cmd/gc/builtin.c b/src/cmd/gc/builtin.c
index f154ae70b1..1178f622e5 100644
--- a/src/cmd/gc/builtin.c
+++ b/src/cmd/gc/builtin.c
@@ -1,4 +1,7 @@
 // AUTO-GENERATED by mkbuiltin; DO NOT EDIT
+#include <u.h>
+#include <libc.h>
+#include "go.h"
 char *runtimeimport =
 	"package runtime\n"
 	"import runtime \"runtime\"\n"
@@ -36,10 +39,10 @@ char *runtimeimport =
 	"func @\"\".intstring (? *[4]byte, ? int64) (? string)\n"
 	"func @\"\".slicebytetostring (? *[32]byte, ? []byte) (? string)\n"
 	"func @\"\".slicebytetostringtmp (? []byte) (? string)\n"
-	"func @\"\".slicerunetostring (? []rune) (? string)\n"
-	"func @\"\".stringtoslicebyte (? string) (? []byte)\n"
+	"func @\"\".slicerunetostring (? *[32]byte, ? []rune) (? string)\n"
+	"func @\"\".stringtoslicebyte (? *[32]byte, ? string) (? []byte)\n"
 	"func @\"\".stringtoslicebytetmp (? string) (? []byte)\n"
-	"func @\"\".stringtoslicerune (? string) (? []rune)\n"
+	"func @\"\".stringtoslicerune (? *[32]rune, ? string) (? []rune)\n"
 	"func @\"\".stringiter (? string, ? int) (? int)\n"
 	"func @\"\".stringiter2 (? string, ? int) (@\"\".retk·1 int, @\"\".retv·2 rune)\n"
 	"func @\"\".slicecopy (@\"\".to·2 any, @\"\".fr·3 any, @\"\".wid·4 uintptr) (? int)\n"
@@ -65,7 +68,7 @@ char *runtimeimport =
 	"func @\"\".efaceeq (@\"\".i1·2 any, @\"\".i2·3 any) (@\"\".ret·1 bool)\n"
 	"func @\"\".ifacethash (@\"\".i1·2 any) (@\"\".ret·1 uint32)\n"
 	"func @\"\".efacethash (@\"\".i1·2 any) (@\"\".ret·1 uint32)\n"
-	"func @\"\".makemap (@\"\".mapType·2 *byte, @\"\".hint·3 int64) (@\"\".hmap·1 map[any]any)\n"
+	"func @\"\".makemap (@\"\".mapType·2 *byte, @\"\".hint·3 int64, @\"\".mapbuf·4 *any, @\"\".bucketbuf·5 *any) (@\"\".hmap·1 map[any]any)\n"
 	"func @\"\".mapaccess1 (@\"\".mapType·2 *byte, @\"\".hmap·3 map[any]any, @\"\".key·4 *any) (@\"\".val·1 *any)\n"
 	"func @\"\".mapaccess1_fast32 (@\"\".mapType·2 *byte, @\"\".hmap·3 map[any]any, @\"\".key·4 any) (@\"\".val·1 *any)\n"
 	"func @\"\".mapaccess1_fast64 (@\"\".mapType·2 *byte, @\"\".hmap·3 map[any]any, @\"\".key·4 any) (@\"\".val·1 *any)\n"
diff --git a/src/cmd/gc/closure.c b/src/cmd/gc/closure.c
index 5d25ffe4ad..0391ece379 100644
--- a/src/cmd/gc/closure.c
+++ b/src/cmd/gc/closure.c
@@ -105,6 +105,7 @@ typecheckclosure(Node *func, int top)
 	oldfn = curfn;
 	typecheck(&func->ntype, Etype);
 	func->type = func->ntype->type;
+	func->top = top;
 
 	// Type check the body now, but only if we're inside a function.
 	// At top level (in a variable initialization: curfn==nil) we're not
@@ -119,9 +120,6 @@ typecheckclosure(Node *func, int top)
 		curfn = oldfn;
 	}
 
-	// Remember closure context for capturevars.
-	func->etype = (top & (Ecall|Eproc)) == Ecall;
-
 	// Create top-level function 
 	xtop = list(xtop, makeclosure(func));
 }
@@ -172,22 +170,17 @@ makeclosure(Node *func)
 // It decides whether each variable captured by a closure should be captured
 // by value or by reference.
 // We use value capturing for values <= 128 bytes that are never reassigned
-// after declaration.
+// after capturing (effectively constant).
 void
 capturevars(Node *xfunc)
 {
-	Node *func, *v, *addr, *cv, *outer;
-	NodeList *l, *body;
-	char *p;
-	vlong offset;
-	int nvar, lno;
+	Node *func, *v, *outer;
+	NodeList *l;
+	int lno;
 
 	lno = lineno;
 	lineno = xfunc->lineno;
 
-	nvar = 0;
-	body = nil;
-	offset = widthptr;
 	func = xfunc->closure;
 	func->enter = nil;
 	for(l=func->cvars; l; l=l->next) {
@@ -202,7 +195,6 @@ capturevars(Node *xfunc)
 			v->op = OXXX;
 			continue;
 		}
-		nvar++;
 
 		// type check the & of closed variables outside the closure,
 		// so that the outer frame also grabs them and knows they escape.
@@ -213,53 +205,150 @@ capturevars(Node *xfunc)
 		if(outer->class != PPARAMOUT && !v->closure->addrtaken && !v->closure->assigned && v->type->width <= 128)
 			v->byval = 1;
 		else {
+			v->closure->addrtaken = 1;
 			outer = nod(OADDR, outer, N);
-			// For a closure that is called in place, but not
-			// inside a go statement, avoid moving variables to the heap.
-			outer->etype = func->etype;
 		}
-		if(debug['m'] > 1)
+		if(debug['m'] > 1) {
+			Sym *name;
+			char *how;
+			name = nil;
+			if(v->curfn && v->curfn->nname)
+				name = v->curfn->nname->sym;
+			how = "ref";
+			if(v->byval)
+				how = "value";
 			warnl(v->lineno, "%S capturing by %s: %S (addr=%d assign=%d width=%d)",
-				(v->curfn && v->curfn->nname) ? v->curfn->nname->sym : S, v->byval ? "value" : "ref",
+				name, how,
 				v->sym, v->closure->addrtaken, v->closure->assigned, (int32)v->type->width);
+		}
 		typecheck(&outer, Erv);
 		func->enter = list(func->enter, outer);
+	}
+
+	lineno = lno;
+}
+
+// transformclosure is called in a separate phase after escape analysis.
+// It transform closure bodies to properly reference captured variables.
+void
+transformclosure(Node *xfunc)
+{
+	Node *func, *cv, *addr, *v, *f;
+	NodeList *l, *body;
+	Type **param, *fld;
+	vlong offset;
+	int lno, nvar;
+
+	lno = lineno;
+	lineno = xfunc->lineno;
+	func = xfunc->closure;
 
-		// declare variables holding addresses taken from closure
-		// and initialize in entry prologue.
-		addr = nod(ONAME, N, N);
-		p = smprint("&%s", v->sym->name);
-		addr->sym = lookup(p);
-		free(p);
-		addr->ntype = nod(OIND, typenod(v->type), N);
-		addr->class = PAUTO;
-		addr->addable = 1;
-		addr->ullman = 1;
-		addr->used = 1;
-		addr->curfn = xfunc;
-		xfunc->dcl = list(xfunc->dcl, addr);
-		v->heapaddr = addr;
-		cv = nod(OCLOSUREVAR, N, N);
-		if(v->byval) {
+	if(func->top&Ecall) {
+		// If the closure is directly called, we transform it to a plain function call
+		// with variables passed as args. This avoids allocation of a closure object.
+		// Here we do only a part of the transformation. Walk of OCALLFUNC(OCLOSURE)
+		// will complete the transformation later.
+		// For illustration, the following closure:
+		//	func(a int) {
+		//		println(byval)
+		//		byref++
+		//	}(42)
+		// becomes:
+		//	func(a int, byval int, &byref *int) {
+		//		println(byval)
+		//		(*&byref)++
+		//	}(42, byval, &byref)
+
+		// f is ONAME of the actual function.
+		f = xfunc->nname;
+		// Get pointer to input arguments and rewind to the end.
+		// We are going to append captured variables to input args.
+		param = &getinargx(f->type)->type;
+		for(; *param; param = &(*param)->down) {
+		}
+		for(l=func->cvars; l; l=l->next) {
+			v = l->n;
+			if(v->op == OXXX)
+				continue;
+			fld = typ(TFIELD);
+			fld->funarg = 1;
+			if(v->byval) {
+				// If v is captured by value, we merely downgrade it to PPARAM.
+				v->class = PPARAM;
+				v->ullman = 1;
+				fld->nname = v;
+			} else {
+				// If v of type T is captured by reference,
+				// we introduce function param &v *T
+				// and v remains PPARAMREF with &v heapaddr
+				// (accesses will implicitly deref &v).
+				snprint(namebuf, sizeof namebuf, "&%s", v->sym->name);
+				addr = newname(lookup(namebuf));
+				addr->type = ptrto(v->type);
+				addr->class = PPARAM;
+				v->heapaddr = addr;
+				fld->nname = addr;
+			}
+			fld->type = fld->nname->type;
+			fld->sym = fld->nname->sym;
+			// Declare the new param and append it to input arguments.
+			xfunc->dcl = list(xfunc->dcl, fld->nname);
+			*param = fld;
+			param = &fld->down;
+		}
+		// Recalculate param offsets.
+		if(f->type->width > 0)
+			fatal("transformclosure: width is already calculated");
+		dowidth(f->type);
+		xfunc->type = f->type; // update type of ODCLFUNC
+	} else {
+		// The closure is not called, so it is going to stay as closure.
+		nvar = 0;
+		body = nil;
+		offset = widthptr;
+		for(l=func->cvars; l; l=l->next) {
+			v = l->n;
+			if(v->op == OXXX)
+				continue;
+			nvar++;
+			// cv refers to the field inside of closure OSTRUCTLIT.
+			cv = nod(OCLOSUREVAR, N, N);
 			cv->type = v->type;
-			offset = rnd(offset, v->type->align);
+			if(!v->byval)
+				cv->type = ptrto(v->type);
+			offset = rnd(offset, cv->type->align);
 			cv->xoffset = offset;
-			offset += v->type->width;
-			body = list(body, nod(OAS, addr, nod(OADDR, cv, N)));
-		} else {
-			v->closure->addrtaken = 1;
-			cv->type = ptrto(v->type);
-			offset = rnd(offset, widthptr);
-			cv->xoffset = offset;
-			offset += widthptr;
-			body = list(body, nod(OAS, addr, cv));
+			offset += cv->type->width;
+
+			if(v->byval && v->type->width <= 2*widthptr && thearch.thechar == '6') {
+				//  If it is a small variable captured by value, downgrade it to PAUTO.
+				// This optimization is currently enabled only for amd64, see:
+				// https://github.com/golang/go/issues/9865
+				v->class = PAUTO;
+				v->ullman = 1;
+				xfunc->dcl = list(xfunc->dcl, v);
+				body = list(body, nod(OAS, v, cv));
+			} else {
+				// Declare variable holding addresses taken from closure
+				// and initialize in entry prologue.
+				snprint(namebuf, sizeof namebuf, "&%s", v->sym->name);
+				addr = newname(lookup(namebuf));
+				addr->ntype = nod(OIND, typenod(v->type), N);
+				addr->class = PAUTO;
+				addr->used = 1;
+				addr->curfn = xfunc;
+				xfunc->dcl = list(xfunc->dcl, addr);
+				v->heapaddr = addr;
+				if(v->byval)
+					cv = nod(OADDR, cv, N);
+				body = list(body, nod(OAS, addr, cv));
+			}
 		}
+		typechecklist(body, Etop);
+		walkstmtlist(body);
+		xfunc->enter = body;
+		xfunc->needctxt = nvar > 0;
 	}
-	typechecklist(body, Etop);
-	walkstmtlist(body);
-	xfunc->enter = body;
-	xfunc->needctxt = nvar > 0;
-	func->etype = 0;
 
 	lineno = lno;
 }
diff --git a/src/cmd/gc/const.c b/src/cmd/gc/const.c
index e418b9c561..50accb93a3 100644
--- a/src/cmd/gc/const.c
+++ b/src/cmd/gc/const.c
@@ -696,14 +696,7 @@ evconst(Node *n)
 	// run op
 	switch(TUP(n->op, v.ctype)) {
 	default:
-	illegal:
-		if(!n->diag) {
-			yyerror("illegal constant expression: %T %O %T",
-				nl->type, n->op, nr->type);
-			n->diag = 1;
-		}
-		return;
-
+		goto illegal;
 	case TUP(OADD, CTINT):
 	case TUP(OADD, CTRUNE):
 		mpaddfixfix(v.u.xval, rv.u.xval, 0);
@@ -1036,6 +1029,14 @@ setfalse:
 	*n = *nodbool(0);
 	n->orig = norig;
 	return;
+
+illegal:
+	if(!n->diag) {
+		yyerror("illegal constant expression: %T %O %T",
+			nl->type, n->op, nr->type);
+		n->diag = 1;
+	}
+	return;
 }
 
 Node*
@@ -1214,28 +1215,31 @@ defaultlit(Node **np, Type *t)
 	case CTCPLX:
 		t1 = types[TCOMPLEX128];
 		goto num;
-	num:
-		if(t != T) {
-			if(isint[t->etype]) {
-				t1 = t;
-				n->val = toint(n->val);
-			}
-			else
-			if(isfloat[t->etype]) {
-				t1 = t;
-				n->val = toflt(n->val);
-			}
-			else
-			if(iscomplex[t->etype]) {
-				t1 = t;
-				n->val = tocplx(n->val);
-			}
+	}
+	lineno = lno;
+	return;
+
+num:
+	if(t != T) {
+		if(isint[t->etype]) {
+			t1 = t;
+			n->val = toint(n->val);
+		}
+		else
+		if(isfloat[t->etype]) {
+			t1 = t;
+			n->val = toflt(n->val);
+		}
+		else
+		if(iscomplex[t->etype]) {
+			t1 = t;
+			n->val = tocplx(n->val);
 		}
-		overflow(n->val, t1);
-		convlit(np, t1);
-		break;
 	}
+	overflow(n->val, t1);
+	convlit(np, t1);
 	lineno = lno;
+	return;
 }
 
 /*
diff --git a/src/cmd/gc/cplx.c b/src/cmd/gc/cplx.c
index d3fb952558..26c21df82a 100644
--- a/src/cmd/gc/cplx.c
+++ b/src/cmd/gc/cplx.c
@@ -78,8 +78,8 @@ complexmove(Node *f, Node *t)
 		subnode(&n1, &n2, f);
 		subnode(&n3, &n4, t);
 
-		arch.cgen(&n1, &n3);
-		arch.cgen(&n2, &n4);
+		thearch.cgen(&n1, &n3);
+		thearch.cgen(&n2, &n4);
 		break;
 	}
 }
@@ -151,9 +151,9 @@ complexgen(Node *n, Node *res)
 		if(res->addable) {
 			subnode(&n1, &n2, res);
 			tempname(&tmp, n1.type);
-			arch.cgen(n->left, &tmp);
-			arch.cgen(n->right, &n2);
-			arch.cgen(&tmp, &n1);
+			thearch.cgen(n->left, &tmp);
+			thearch.cgen(n->right, &n2);
+			thearch.cgen(&tmp, &n1);
 			return;
 		}
 		break;
@@ -168,10 +168,10 @@ complexgen(Node *n, Node *res)
 		}
 		subnode(&n1, &n2, nl);
 		if(n->op == OREAL) {
-			arch.cgen(&n1, res);
+			thearch.cgen(&n1, res);
 			return;
 		}
-		arch.cgen(&n2, res);
+		thearch.cgen(&n2, res);
 		return;
 	}
 
@@ -191,9 +191,9 @@ complexgen(Node *n, Node *res)
 	}
 
 	if(!res->addable) {
-		arch.igen(res, &n1, N);
-		arch.cgen(n, &n1);
-		arch.regfree(&n1);
+		thearch.igen(res, &n1, N);
+		thearch.cgen(n, &n1);
+		thearch.regfree(&n1);
 		return;
 	}
 	if(n->addable) {
@@ -214,9 +214,9 @@ complexgen(Node *n, Node *res)
 	case OCALLFUNC:
 	case OCALLMETH:
 	case OCALLINTER:
-		arch.igen(n, &n1, res);
+		thearch.igen(n, &n1, res);
 		complexmove(&n1, res);
-		arch.regfree(&n1);
+		thearch.regfree(&n1);
 		return;
 
 	case OCONV:
@@ -239,18 +239,18 @@ complexgen(Node *n, Node *res)
 	if(nr != N) {
 		if(nl->ullman > nr->ullman && !nl->addable) {
 			tempname(&tnl, nl->type);
-			arch.cgen(nl, &tnl);
+			thearch.cgen(nl, &tnl);
 			nl = &tnl;
 		}
 		if(!nr->addable) {
 			tempname(&tnr, nr->type);
-			arch.cgen(nr, &tnr);
+			thearch.cgen(nr, &tnr);
 			nr = &tnr;
 		}
 	}
 	if(!nl->addable) {
 		tempname(&tnl, nl->type);
-		arch.cgen(nl, &tnl);
+		thearch.cgen(nl, &tnl);
 		nl = &tnl;
 	}
 
@@ -289,18 +289,18 @@ complexbool(int op, Node *nl, Node *nr, int true, int likely, Prog *to)
 	if(nr != N) {
 		if(nl->ullman > nr->ullman && !nl->addable) {
 			tempname(&tnl, nl->type);
-			arch.cgen(nl, &tnl);
+			thearch.cgen(nl, &tnl);
 			nl = &tnl;
 		}
 		if(!nr->addable) {
 			tempname(&tnr, nr->type);
-			arch.cgen(nr, &tnr);
+			thearch.cgen(nr, &tnr);
 			nr = &tnr;
 		}
 	}
 	if(!nl->addable) {
 		tempname(&tnl, nl->type);
-		arch.cgen(nl, &tnl);
+		thearch.cgen(nl, &tnl);
 		nl = &tnl;
 	}
 
@@ -331,7 +331,7 @@ complexbool(int op, Node *nl, Node *nr, int true, int likely, Prog *to)
 	if(op == ONE)
 		true = !true;
 
-	arch.bgen(&na, true, likely, to);
+	thearch.bgen(&na, true, likely, to);
 }
 
 void
@@ -387,7 +387,7 @@ minus(Node *nl, Node *res)
 	ra.op = OMINUS;
 	ra.left = nl;
 	ra.type = nl->type;
-	arch.cgen(&ra, res);
+	thearch.cgen(&ra, res);
 }
 
 // build and execute tree
@@ -424,14 +424,14 @@ complexadd(int op, Node *nl, Node *nr, Node *res)
 	ra.left = &n1;
 	ra.right = &n3;
 	ra.type = n1.type;
-	arch.cgen(&ra, &n5);
+	thearch.cgen(&ra, &n5);
 
 	memset(&ra, 0, sizeof(ra));
 	ra.op = op;
 	ra.left = &n2;
 	ra.right = &n4;
 	ra.type = n2.type;
-	arch.cgen(&ra, &n6);
+	thearch.cgen(&ra, &n6);
 }
 
 // build and execute tree
@@ -467,7 +467,7 @@ complexmul(Node *nl, Node *nr, Node *res)
 	ra.left = &rm1;
 	ra.right = &rm2;
 	ra.type = rm1.type;
-	arch.cgen(&ra, &tmp);
+	thearch.cgen(&ra, &tmp);
 
 	// imag part
 	memset(&rm1, 0, sizeof(rm1));
@@ -487,8 +487,8 @@ complexmul(Node *nl, Node *nr, Node *res)
 	ra.left = &rm1;
 	ra.right = &rm2;
 	ra.type = rm1.type;
-	arch.cgen(&ra, &n6);
+	thearch.cgen(&ra, &n6);
 
 	// tmp ->real part
-	arch.cgen(&tmp, &n5);
+	thearch.cgen(&tmp, &n5);
 }
diff --git a/src/cmd/gc/dcl.c b/src/cmd/gc/dcl.c
index f47ca2b50e..9a6c0023f5 100644
--- a/src/cmd/gc/dcl.c
+++ b/src/cmd/gc/dcl.c
@@ -1435,7 +1435,7 @@ addmethod(Sym *sf, Type *t, int local, int nointerface)
 }
 
 void
-funccompile(Node *n, int isclosure)
+funccompile(Node *n)
 {
 	stksize = BADWIDTH;
 	maxarg = 0;
@@ -1449,20 +1449,6 @@ funccompile(Node *n, int isclosure)
 	// assign parameter offsets
 	checkwidth(n->type);
 	
-	// record offset to actual frame pointer.
-	// for closure, have to skip over leading pointers and PC slot.
-	// TODO(rsc): this is the old jit closure handling code.
-	// with the new closures, isclosure is always 0; delete this block.
-	nodfp->xoffset = 0;
-	if(isclosure) {
-		NodeList *l;
-		for(l=n->nname->ntype->list; l; l=l->next) {
-			nodfp->xoffset += widthptr;
-			if(l->n->left == N)	// found slot for PC
-				break;
-		}
-	}
-
 	if(curfn)
 		fatal("funccompile %S inside %S", n->nname->sym, curfn->nname->sym);
 
diff --git a/src/cmd/gc/esc.c b/src/cmd/gc/esc.c
index 4f77983926..5b09c0b7fb 100644
--- a/src/cmd/gc/esc.c
+++ b/src/cmd/gc/esc.c
@@ -693,12 +693,10 @@ esc(EscState *e, Node *n, Node *up)
 	case OMAKEMAP:
 	case OMAKESLICE:
 	case ONEW:
-		n->escloopdepth = e->loopdepth;
-		n->esc = EscNone;  // until proven otherwise
-		e->noesc = list(e->noesc, n);
-		break;
-
+	case OARRAYRUNESTR:
 	case OARRAYBYTESTR:
+	case OSTRARRAYRUNE:
+	case OSTRARRAYBYTE:
 	case ORUNESTR:
 		n->escloopdepth = e->loopdepth;
 		n->esc = EscNone;  // until proven otherwise
@@ -824,7 +822,10 @@ escassign(EscState *e, Node *dst, Node *src)
 	case OMAKECHAN:
 	case OMAKEMAP:
 	case OMAKESLICE:
+	case OARRAYRUNESTR:
 	case OARRAYBYTESTR:
+	case OSTRARRAYRUNE:
+	case OSTRARRAYBYTE:
 	case OADDSTR:
 	case ONEW:
 	case OCLOSURE:
@@ -1249,7 +1250,10 @@ escwalk(EscState *e, int level, Node *dst, Node *src)
 	case OMAKECHAN:
 	case OMAKEMAP:
 	case OMAKESLICE:
+	case OARRAYRUNESTR:
 	case OARRAYBYTESTR:
+	case OSTRARRAYRUNE:
+	case OSTRARRAYBYTE:
 	case OADDSTR:
 	case OMAPLIT:
 	case ONEW:
diff --git a/src/cmd/gc/export.c b/src/cmd/gc/export.c
index 8685aa0def..91f4957dc2 100644
--- a/src/cmd/gc/export.c
+++ b/src/cmd/gc/export.c
@@ -538,7 +538,7 @@ dumpasmhdr(void)
 	b = Bopen(asmhdr, OWRITE);
 	if(b == nil)
 		fatal("open %s: %r", asmhdr);
-	Bprint(b, "// generated by %cg -asmhdr from package %s\n\n", arch.thechar, localpkg->name);
+	Bprint(b, "// generated by %cg -asmhdr from package %s\n\n", thearch.thechar, localpkg->name);
 	for(l=asmlist; l; l=l->next) {
 		n = l->n;
 		if(isblanksym(n->sym))
diff --git a/src/cmd/gc/fmt.c b/src/cmd/gc/fmt.c
index f9eb0ba2b3..bd0a29f6e1 100644
--- a/src/cmd/gc/fmt.c
+++ b/src/cmd/gc/fmt.c
@@ -171,7 +171,7 @@ goopnames[] =
 };
 
 // Fmt "%O":  Node opcodes
-static int
+int
 Oconv(Fmt *fp)
 {
 	int o;
@@ -1558,7 +1558,7 @@ Sconv(Fmt *fp)
 // Flags: 'l' print definition, not name
 //	  'h' omit 'func' and receiver from function types, short type names
 //	  'u' package name, not prefix (FTypeId mode, sticky)
-static int
+int
 Tconv(Fmt *fp)
 {
 	Type *t;
diff --git a/src/cmd/gc/gen.c b/src/cmd/gc/gen.c
index 20c17bd46a..72bcc492a5 100644
--- a/src/cmd/gc/gen.c
+++ b/src/cmd/gc/gen.c
@@ -265,7 +265,7 @@ gen(Node *n)
 //dump("gen", n);
 
 	lno = setlineno(n);
-	wasregalloc = arch.anyregalloc();
+	wasregalloc = thearch.anyregalloc();
 
 	if(n == N)
 		goto ret;
@@ -304,8 +304,8 @@ gen(Node *n)
 
 		// if there are pending gotos, resolve them all to the current pc.
 		for(p1=lab->gotopc; p1; p1=p2) {
-			p2 = arch.unpatch(p1);
-			arch.patch(p1, pc);
+			p2 = unpatch(p1);
+			patch(p1, pc);
 		}
 		lab->gotopc = P;
 		if(lab->labelpc == P)
@@ -332,9 +332,9 @@ gen(Node *n)
 		// of the label in the OLABEL case above.)
 		lab = newlab(n);
 		if(lab->labelpc != P)
-			arch.gjmp(lab->labelpc);
+			gjmp(lab->labelpc);
 		else
-			lab->gotopc = arch.gjmp(lab->gotopc);
+			lab->gotopc = gjmp(lab->gotopc);
 		break;
 
 	case OBREAK:
@@ -349,14 +349,14 @@ gen(Node *n)
 				yyerror("invalid break label %S", n->left->sym);
 				break;
 			}
-			arch.gjmp(lab->breakpc);
+			gjmp(lab->breakpc);
 			break;
 		}
 		if(breakpc == P) {
 			yyerror("break is not in a loop");
 			break;
 		}
-		arch.gjmp(breakpc);
+		gjmp(breakpc);
 		break;
 
 	case OCONTINUE:
@@ -371,20 +371,20 @@ gen(Node *n)
 				yyerror("invalid continue label %S", n->left->sym);
 				break;
 			}
-			arch.gjmp(lab->continpc);
+			gjmp(lab->continpc);
 			break;
 		}
 		if(continpc == P) {
 			yyerror("continue is not in a loop");
 			break;
 		}
-		arch.gjmp(continpc);
+		gjmp(continpc);
 		break;
 
 	case OFOR:
 		sbreak = breakpc;
-		p1 = arch.gjmp(P);			//		goto test
-		breakpc = arch.gjmp(P);		// break:	goto done
+		p1 = gjmp(P);			//		goto test
+		breakpc = gjmp(P);		// break:	goto done
 		scontin = continpc;
 		continpc = pc;
 
@@ -394,11 +394,11 @@ gen(Node *n)
 			lab->continpc = continpc;
 		}
 		gen(n->nincr);				// contin:	incr
-		arch.patch(p1, pc);				// test:
-		arch.bgen(n->ntest, 0, -1, breakpc);		//		if(!test) goto break
+		patch(p1, pc);				// test:
+		thearch.bgen(n->ntest, 0, -1, breakpc);		//		if(!test) goto break
 		genlist(n->nbody);				//		body
-		arch.gjmp(continpc);
-		arch.patch(breakpc, pc);			// done:
+		gjmp(continpc);
+		patch(breakpc, pc);			// done:
 		continpc = scontin;
 		breakpc = sbreak;
 		if(lab) {
@@ -408,29 +408,29 @@ gen(Node *n)
 		break;
 
 	case OIF:
-		p1 = arch.gjmp(P);			//		goto test
-		p2 = arch.gjmp(P);			// p2:		goto else
-		arch.patch(p1, pc);				// test:
-		arch.bgen(n->ntest, 0, -n->likely, p2);		//		if(!test) goto p2
+		p1 = gjmp(P);			//		goto test
+		p2 = gjmp(P);			// p2:		goto else
+		patch(p1, pc);				// test:
+		thearch.bgen(n->ntest, 0, -n->likely, p2);		//		if(!test) goto p2
 		genlist(n->nbody);				//		then
-		p3 = arch.gjmp(P);			//		goto done
-		arch.patch(p2, pc);				// else:
+		p3 = gjmp(P);			//		goto done
+		patch(p2, pc);				// else:
 		genlist(n->nelse);				//		else
-		arch.patch(p3, pc);				// done:
+		patch(p3, pc);				// done:
 		break;
 
 	case OSWITCH:
 		sbreak = breakpc;
-		p1 = arch.gjmp(P);			//		goto test
-		breakpc = arch.gjmp(P);		// break:	goto done
+		p1 = gjmp(P);			//		goto test
+		breakpc = gjmp(P);		// break:	goto done
 
 		// define break label
 		if((lab = stmtlabel(n)) != L)
 			lab->breakpc = breakpc;
 
-		arch.patch(p1, pc);				// test:
+		patch(p1, pc);				// test:
 		genlist(n->nbody);				//		switch(test) body
-		arch.patch(breakpc, pc);			// done:
+		patch(breakpc, pc);			// done:
 		breakpc = sbreak;
 		if(lab != L)
 			lab->breakpc = P;
@@ -438,25 +438,21 @@ gen(Node *n)
 
 	case OSELECT:
 		sbreak = breakpc;
-		p1 = arch.gjmp(P);			//		goto test
-		breakpc = arch.gjmp(P);		// break:	goto done
+		p1 = gjmp(P);			//		goto test
+		breakpc = gjmp(P);		// break:	goto done
 
 		// define break label
 		if((lab = stmtlabel(n)) != L)
 			lab->breakpc = breakpc;
 
-		arch.patch(p1, pc);				// test:
+		patch(p1, pc);				// test:
 		genlist(n->nbody);				//		select() body
-		arch.patch(breakpc, pc);			// done:
+		patch(breakpc, pc);			// done:
 		breakpc = sbreak;
 		if(lab != L)
 			lab->breakpc = P;
 		break;
 
-	case OASOP:
-		arch.cgen_asop(n);
-		break;
-
 	case ODCL:
 		cgen_dcl(n->left);
 		break;
@@ -472,11 +468,11 @@ gen(Node *n)
 		break;
 
 	case OCALLINTER:
-		arch.cgen_callinter(n, N, 0);
+		thearch.cgen_callinter(n, N, 0);
 		break;
 
 	case OCALLFUNC:
-		arch.cgen_call(n, 0);
+		thearch.cgen_call(n, 0);
 		break;
 
 	case OPROC:
@@ -489,7 +485,7 @@ gen(Node *n)
 
 	case ORETURN:
 	case ORETJMP:
-		arch.cgen_ret(n);
+		thearch.cgen_ret(n);
 		break;
 	
 	case OCHECKNIL:
@@ -502,7 +498,7 @@ gen(Node *n)
 	}
 
 ret:
-	if(arch.anyregalloc() != wasregalloc) {
+	if(thearch.anyregalloc() != wasregalloc) {
 		dump("node", n);
 		fatal("registers left allocated");
 	}
@@ -536,7 +532,7 @@ cgen_callmeth(Node *n, int proc)
 
 	if(n2.left->op == ONAME)
 		n2.left->class = PFUNC;
-	arch.cgen_call(&n2, proc);
+	thearch.cgen_call(&n2, proc);
 }
 
 /*
@@ -554,11 +550,11 @@ cgen_proc(Node *n, int proc)
 		break;
 
 	case OCALLINTER:
-		arch.cgen_callinter(n->left, N, proc);
+		thearch.cgen_callinter(n->left, N, proc);
 		break;
 
 	case OCALLFUNC:
-		arch.cgen_call(n->left, proc);
+		thearch.cgen_call(n->left, proc);
 		break;
 	}
 
@@ -601,7 +597,7 @@ cgen_discard(Node *nr)
 	switch(nr->op) {
 	case ONAME:
 		if(!(nr->class & PHEAP) && nr->class != PEXTERN && nr->class != PFUNC && nr->class != PPARAMREF)
-			arch.gused(nr);
+			gused(nr);
 		break;
 
 	// unary
@@ -643,7 +639,7 @@ cgen_discard(Node *nr)
 	default:
 		tempname(&tmp, nr->type);
 		cgen_as(&tmp, nr);
-		arch.gused(&tmp);
+		gused(&tmp);
 	}
 }
 
@@ -705,7 +701,7 @@ clearslim(Node *n)
 	}
 
 	ullmancalc(&z);
-	arch.cgen(&z, n);
+	thearch.cgen(&z, n);
 }
 
 /*
@@ -739,10 +735,10 @@ cgen_as(Node *nl, Node *nr)
 		tl = nl->type;
 		if(tl == T)
 			return;
-		if(arch.isfat(tl)) {
+		if(isfat(tl)) {
 			if(nl->op == ONAME)
 				gvardef(nl);
-			arch.clearfat(nl);
+			thearch.clearfat(nl);
 			return;
 		}
 		clearslim(nl);
@@ -753,7 +749,7 @@ cgen_as(Node *nl, Node *nr)
 	if(tl == T)
 		return;
 
-	arch.cgen(nr, nl);
+	thearch.cgen(nr, nl);
 }
 
 /*
@@ -773,17 +769,17 @@ cgen_eface(Node *n, Node *res)
 	Node *tmp;
 
 	tmp = temp(types[tptr]);
-	arch.cgen(n->right, tmp);
+	thearch.cgen(n->right, tmp);
 
 	gvardef(res);
 
 	dst = *res;
 	dst.type = types[tptr];
 	dst.xoffset += widthptr;
-	arch.cgen(tmp, &dst);
+	thearch.cgen(tmp, &dst);
 
 	dst.xoffset -= widthptr;
-	arch.cgen(n->left, &dst);
+	thearch.cgen(n->left, &dst);
 }
 
 /*
@@ -824,7 +820,7 @@ cgen_slice(Node *n, Node *res)
 
 	if(isnil(n->left)) {
 		tempname(&src, n->left->type);
-		arch.cgen(n->left, &src);
+		thearch.cgen(n->left, &src);
 	} else
 		src = *n->left;
 	if(n->op == OSLICE || n->op == OSLICE3 || n->op == OSLICESTR)
@@ -833,11 +829,11 @@ cgen_slice(Node *n, Node *res)
 	if(n->op == OSLICEARR || n->op == OSLICE3ARR) {
 		if(!isptr[n->left->type->etype])
 			fatal("slicearr is supposed to work on pointer: %+N\n", n);
-		arch.cgen(&src, base);
+		thearch.cgen(&src, base);
 		cgen_checknil(base);
 	} else {
 		src.type = types[tptr];
-		arch.cgen(&src, base);
+		thearch.cgen(&src, base);
 	}
 	
 	// committed to the update
@@ -846,9 +842,9 @@ cgen_slice(Node *n, Node *res)
 	// compute len and cap.
 	// len = n-i, cap = m-i, and offs = i*width.
 	// computing offs last lets the multiply overwrite i.
-	arch.cgen(len, tmplen);
+	thearch.cgen(len, tmplen);
 	if(n->op != OSLICESTR)
-		arch.cgen(cap, tmpcap);
+		thearch.cgen(cap, tmpcap);
 
 	// if new cap != 0 { base += add }
 	// This avoids advancing base past the end of the underlying array/string,
@@ -857,40 +853,40 @@ cgen_slice(Node *n, Node *res)
 	// In essence we are replacing x[i:j:k] where i == j == k
 	// or x[i:j] where i == j == cap(x) with x[0:0:0].
 	if(offs != N) {
-		p1 = arch.gjmp(P);
-		p2 = arch.gjmp(P);
-		arch.patch(p1, pc);
+		p1 = gjmp(P);
+		p2 = gjmp(P);
+		patch(p1, pc);
 
 		nodconst(&con, tmpcap->type, 0);
 		cmp = nod(OEQ, tmpcap, &con);
 		typecheck(&cmp, Erv);
-		arch.bgen(cmp, 1, -1, p2);
+		thearch.bgen(cmp, 1, -1, p2);
 
 		add = nod(OADD, base, offs);
 		typecheck(&add, Erv);
-		arch.cgen(add, base);
+		thearch.cgen(add, base);
 
-		arch.patch(p2, pc);
+		patch(p2, pc);
 	}
 
 	// dst.array = src.array  [ + lo *width ]
 	dst = *res;
 	dst.xoffset += Array_array;
 	dst.type = types[tptr];
-	arch.cgen(base, &dst);
+	thearch.cgen(base, &dst);
 
 	// dst.len = hi [ - lo ]
 	dst = *res;
 	dst.xoffset += Array_nel;
 	dst.type = types[simtype[TUINT]];
-	arch.cgen(tmplen, &dst);
+	thearch.cgen(tmplen, &dst);
 
 	if(n->op != OSLICESTR) {
 		// dst.cap = cap [ - lo ]
 		dst = *res;
 		dst.xoffset += Array_cap;
 		dst.type = types[simtype[TUINT]];
-		arch.cgen(tmpcap, &dst);
+		thearch.cgen(tmpcap, &dst);
 	}
 }
 
diff --git a/src/cmd/gc/go.h b/src/cmd/gc/go.h
index 2aa7838c93..c5ef74586d 100644
--- a/src/cmd/gc/go.h
+++ b/src/cmd/gc/go.h
@@ -331,6 +331,7 @@ struct	Node
 	// ONAME closure param with PPARAMREF
 	Node*	outer;	// outer PPARAMREF in nested closure
 	Node*	closure;	// ONAME/PHEAP <-> ONAME/PPARAMREF
+	int	top;	// top context (Ecall, Eproc, etc)
 
 	// ONAME substitute while inlining
 	Node* inlvar;
@@ -598,7 +599,7 @@ enum
 	OCHECKNIL, // emit code to ensure pointer/interface not nil
 	OVARKILL, // variable is dead
 
-	// arch-specific registers
+	// thearch-specific registers
 	OREGISTER,	// a register, such as AX.
 	OINDREG,	// offset plus indirect of a register, such as 8(SP).
 
@@ -742,6 +743,7 @@ struct	Var
 	Node*	node;
 	Var*	nextinnode;
 	int	width;
+	int	id;
 	char	name;
 	char	etype;
 	char	addr;
@@ -1075,6 +1077,7 @@ Node*	closurebody(NodeList *body);
 void	closurehdr(Node *ntype);
 void	typecheckclosure(Node *func, int top);
 void	capturevars(Node *func);
+void	transformclosure(Node *func);
 Node*	walkclosure(Node *func, NodeList **init);
 void	typecheckpartialcall(Node*, Node*);
 Node*	walkpartialcall(Node*, NodeList**);
@@ -1128,7 +1131,7 @@ void	dumpdcl(char *st);
 Node*	embedded(Sym *s, Pkg *pkg);
 Node*	fakethis(void);
 void	funcbody(Node *n);
-void	funccompile(Node *n, int isclosure);
+void	funccompile(Node *n);
 void	funchdr(Node *n);
 Type*	functype(Node *this, NodeList *in, NodeList *out);
 void	ifacedcl(Node *n);
@@ -1321,7 +1324,9 @@ Sym*	typenamesym(Type *t);
 Sym*	tracksym(Type *t);
 Sym*	typesymprefix(char *prefix, Type *t);
 int	haspointers(Type *t);
+Type*	hmap(Type *t);
 Type*	hiter(Type* t);
+Type*	mapbucket(Type *t);
 
 /*
  *	select.c
@@ -1444,6 +1449,7 @@ void	warn(char *fmt, ...);
 void	warnl(int line, char *fmt, ...);
 void	yyerror(char *fmt, ...);
 void	yyerrorl(int line, char *fmt, ...);
+void	adderrorname(Node*);
 
 /*
  *	swt.c
@@ -1492,7 +1498,7 @@ Node*	outervalue(Node*);
 void	usefield(Node*);
 
 /*
- *	arch-specific ggen.c/gsubr.c/gobj.c/pgen.c/plive.c
+ *	thearch-specific ggen.c/gsubr.c/gobj.c/pgen.c/plive.c
  */
 #define	P	((Prog*)0)
 
@@ -1514,7 +1520,6 @@ void	gvarkill(Node*);
 void	movelarge(NodeList*);
 void	liveness(Node*, Prog*, Sym*, Sym*);
 void	twobitwalktype1(Type*, vlong*, Bvec*);
-void	nopout(Prog*);
 
 #pragma	varargck	type	"B"	Mpint*
 #pragma	varargck	type	"E"	int
@@ -1558,9 +1563,12 @@ struct Flow {
 	
 	int32	active;	// usable by client
 
+	int32	id;		// sequence number in flow graph
 	int32	rpo;		// reverse post ordering
 	uint16	loop;		// x5 for every loop
 	uchar	refset;		// diagnostic generated
+	
+	void*	data;	// for use by client
 };
 
 struct Graph
@@ -1659,61 +1667,83 @@ struct Arch
 	LinkArch *thelinkarch;
 	Typedef *typedefs;
 
+	int	REGSP;
+	int	REGCTXT;
 	vlong MAXWIDTH;
 
-	void (*afunclit)(Addr*, Node*);
 	int (*anyregalloc)(void);
 	void (*betypeinit)(void);
 	void (*bgen)(Node*, int, int, Prog*);
 	void (*cgen)(Node*, Node*);
-	void (*cgen_asop)(Node*);
 	void (*cgen_call)(Node*, int);
 	void (*cgen_callinter)(Node*, Node*, int);
 	void (*cgen_ret)(Node*);
 	void (*clearfat)(Node*);
-	void (*clearp)(Prog*);
 	void (*defframe)(Prog*);
-	int (*dgostringptr)(Sym*, int, char*);
-	int (*dgostrlitptr)(Sym*, int, Strlit*);
-	int (*dsname)(Sym*, int, char*, int);
-	int (*dsymptr)(Sym*, int, Sym*, int);
-	void (*dumpdata)(void);
-	void (*dumpit)(char*, Flow*, int);
 	void (*excise)(Flow*);
 	void (*expandchecks)(Prog*);
-	void (*fixautoused)(Prog*);
 	void (*gclean)(void);
-	void	(*gdata)(Node*, Node*, int);
-	void	(*gdatacomplex)(Node*, Mpcplx*);
-	void	(*gdatastring)(Node*, Strlit*);
-	void	(*ggloblnod)(Node*);
-	void	(*ggloblsym)(Sym*, int32, int8);
 	void (*ginit)(void);
 	Prog*	(*gins)(int, Node*, Node*);
 	void	(*ginscall)(Node*, int);
-	Prog*	(*gjmp)(Prog*);
-	void (*gtrack)(Sym*);
-	void	(*gused)(Node*);
 	void	(*igen)(Node*, Node*, Node*);
-	int (*isfat)(Type*);
 	void (*linkarchinit)(void);
-	void (*markautoused)(Prog*);
-	void (*naddr)(Node*, Addr*, int);
-	Plist* (*newplist)(void);
-	Node* (*nodarg)(Type*, int);
-	void (*patch)(Prog*, Prog*);
+	void (*peep)(Prog*);
 	void (*proginfo)(ProgInfo*, Prog*);
 	void (*regalloc)(Node*, Type*, Node*);
 	void (*regfree)(Node*);
-	void (*regopt)(Prog*);
 	int (*regtyp)(Addr*);
 	int (*sameaddr)(Addr*, Addr*);
 	int (*smallindir)(Addr*, Addr*);
 	int (*stackaddr)(Addr*);
-	Prog* (*unpatch)(Prog*);
+	uint64 (*excludedregs)(void);
+	uint64 (*RtoB)(int);
+	uint64 (*FtoB)(int);
+	int (*BtoR)(uint64);
+	int (*BtoF)(uint64);
+	int (*optoas)(int, Type*);
+	uint64 (*doregbits)(int);
+	char **(*regnames)(int*);
 };
 
-EXTERN Arch arch;
+void afunclit(Addr*, Node*);
+void clearp(Prog*);
+int dgostringptr(Sym*, int, char*);
+int dgostrlitptr(Sym*, int, Strlit*);
+int dsname(Sym*, int, char*, int);
+int dsymptr(Sym*, int, Sym*, int);
+void dumpdata(void);
+void fixautoused(Prog*);
+void	gdata(Node*, Node*, int);
+void	gdatacomplex(Node*, Mpcplx*);
+void	gdatastring(Node*, Strlit*);
+void	ggloblnod(Node*);
+void	ggloblsym(Sym*, int32, int8);
+Prog*	gjmp(Prog*);
+void gtrack(Sym*);
+void	gused(Node*);
+int isfat(Type*);
+void markautoused(Prog*);
+void naddr(Node*, Addr*, int);
+Plist* newplist(void);
+Node* nodarg(Type*, int);
+void patch(Prog*, Prog*);
+Prog* unpatch(Prog*);
+void datagostring(Strlit *sval, Addr *a);
+int ismem(Node*);
+int samereg(Node*, Node*);
+void	regopt(Prog*);
+int	Tconv(Fmt*);
+int	Oconv(Fmt*);
+Prog*	gbranch(int as, Type *t, int likely);
+void	nodindreg(Node *n, Type *t, int r);
+void	nodreg(Node *n, Type *t, int r);
+Prog*	prog(int as);
+void	datastring(char*, int, Addr*);
+
+EXTERN int32	pcloc;
+
+EXTERN Arch thearch;
 
 EXTERN Node *newproc;
 EXTERN Node *deferproc;
diff --git a/src/cmd/gc/gsubr.c b/src/cmd/gc/gsubr.c
new file mode 100644
index 0000000000..5175ae34f5
--- /dev/null
+++ b/src/cmd/gc/gsubr.c
@@ -0,0 +1,654 @@
+// Derived from Inferno utils/6c/txt.c
+// http://code.google.com/p/inferno-os/source/browse/utils/6c/txt.c
+//
+//	Copyright © 1994-1999 Lucent Technologies Inc.  All rights reserved.
+//	Portions Copyright © 1995-1997 C H Forsyth (forsyth@terzarima.net)
+//	Portions Copyright © 1997-1999 Vita Nuova Limited
+//	Portions Copyright © 2000-2007 Vita Nuova Holdings Limited (www.vitanuova.com)
+//	Portions Copyright © 2004,2006 Bruce Ellis
+//	Portions Copyright © 2005-2007 C H Forsyth (forsyth@terzarima.net)
+//	Revisions Copyright © 2000-2007 Lucent Technologies Inc. and others
+//	Portions Copyright © 2009 The Go Authors.  All rights reserved.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a copy
+// of this software and associated documentation files (the "Software"), to deal
+// in the Software without restriction, including without limitation the rights
+// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+// copies of the Software, and to permit persons to whom the Software is
+// furnished to do so, subject to the following conditions:
+//
+// The above copyright notice and this permission notice shall be included in
+// all copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+// THE SOFTWARE.
+
+#include <u.h>
+#include <libc.h>
+#include "go.h"
+#include "../../runtime/funcdata.h"
+#include "../ld/textflag.h"
+
+void
+ggloblnod(Node *nam)
+{
+	Prog *p;
+
+	p = thearch.gins(AGLOBL, nam, N);
+	p->lineno = nam->lineno;
+	p->from.sym->gotype = linksym(ngotype(nam));
+	p->to.sym = nil;
+	p->to.type = TYPE_CONST;
+	p->to.offset = nam->type->width;
+	if(nam->readonly)
+		p->from3.offset = RODATA;
+	if(nam->type != T && !haspointers(nam->type))
+		p->from3.offset |= NOPTR;
+}
+
+void
+gtrack(Sym *s)
+{
+	Prog *p;
+	
+	p = thearch.gins(AUSEFIELD, N, N);
+	p->from.type = TYPE_MEM;
+	p->from.name = NAME_EXTERN;
+	p->from.sym = linksym(s);
+}
+
+void
+ggloblsym(Sym *s, int32 width, int8 flags)
+{
+	Prog *p;
+
+	p = thearch.gins(AGLOBL, N, N);
+	p->from.type = TYPE_MEM;
+	p->from.name = NAME_EXTERN;
+	p->from.sym = linksym(s);
+	p->to.type = TYPE_CONST;
+	p->to.offset = width;
+	p->from3.offset = flags;
+}
+
+void
+clearp(Prog *p)
+{
+	nopout(p);
+	p->as = AEND;
+	p->pc = pcloc;
+	pcloc++;
+}
+
+static int ddumped;
+static Prog *dfirst;
+static Prog *dpc;
+
+/*
+ * generate and return proc with p->as = as,
+ * linked into program. pc is next instruction.
+ */
+Prog*
+prog(int as)
+{
+	Prog *p;
+
+	if(as == ADATA || as == AGLOBL) {
+		if(ddumped)
+			fatal("already dumped data");
+		if(dpc == nil) {
+			dpc = mal(sizeof(*dpc));
+			dfirst = dpc;
+		}
+		p = dpc;
+		dpc = mal(sizeof(*dpc));
+		p->link = dpc;
+	} else {
+		p = pc;
+		pc = mal(sizeof(*pc));
+		clearp(pc);
+		p->link = pc;
+	}
+
+	if(lineno == 0) {
+		if(debug['K'])
+			warn("prog: line 0");
+	}
+
+	p->as = as;
+	p->lineno = lineno;
+	return p;
+}
+
+void
+dumpdata(void)
+{
+	ddumped = 1;
+	if(dfirst == nil)
+		return;
+	newplist();
+	*pc = *dfirst;
+	pc = dpc;
+	clearp(pc);
+}
+
+/*
+ * generate a branch.
+ * t is ignored.
+ * likely values are for branch prediction:
+ *	-1 unlikely
+ *	0 no opinion
+ *	+1 likely
+ */
+Prog*
+gbranch(int as, Type *t, int likely)
+{
+	Prog *p;
+	
+	USED(t);
+
+	p = prog(as);
+	p->to.type = TYPE_BRANCH;
+	p->to.u.branch = P;
+	if(as != AJMP && likely != 0 && thearch.thechar != '9') {
+		p->from.type = TYPE_CONST;
+		p->from.offset = likely > 0;
+	}
+	return p;
+}
+
+/*
+ * patch previous branch to jump to to.
+ */
+void
+patch(Prog *p, Prog *to)
+{
+	if(p->to.type != TYPE_BRANCH)
+		fatal("patch: not a branch");
+	p->to.u.branch = to;
+	p->to.offset = to->pc;
+}
+
+Prog*
+unpatch(Prog *p)
+{
+	Prog *q;
+
+	if(p->to.type != TYPE_BRANCH)
+		fatal("unpatch: not a branch");
+	q = p->to.u.branch;
+	p->to.u.branch = P;
+	p->to.offset = 0;
+	return q;
+}
+
+/*
+ * start a new Prog list.
+ */
+Plist*
+newplist(void)
+{
+	Plist *pl;
+
+	pl = linknewplist(ctxt);
+
+	pc = mal(sizeof(*pc));
+	clearp(pc);
+	pl->firstpc = pc;
+
+	return pl;
+}
+
+void
+gused(Node *n)
+{
+	thearch.gins(ANOP, n, N);	// used
+}
+
+Prog*
+gjmp(Prog *to)
+{
+	Prog *p;
+
+	p = gbranch(AJMP, T, 0);
+	if(to != P)
+		patch(p, to);
+	return p;
+}
+
+int
+isfat(Type *t)
+{
+	if(t != T)
+	switch(t->etype) {
+	case TSTRUCT:
+	case TARRAY:
+	case TSTRING:
+	case TINTER:	// maybe remove later
+		return 1;
+	}
+	return 0;
+}
+
+/*
+ * naddr of func generates code for address of func.
+ * if using opcode that can take address implicitly,
+ * call afunclit to fix up the argument.
+ */
+void
+afunclit(Addr *a, Node *n)
+{
+	if(a->type == TYPE_ADDR && a->name == NAME_EXTERN) {
+		a->type = TYPE_MEM;
+		a->sym = linksym(n->sym);
+	}
+}
+
+/*
+ * initialize n to be register r of type t.
+ */
+void
+nodreg(Node *n, Type *t, int r)
+{
+	if(t == T)
+		fatal("nodreg: t nil");
+
+	memset(n, 0, sizeof(*n));
+	n->op = OREGISTER;
+	n->addable = 1;
+	ullmancalc(n);
+	n->val.u.reg = r;
+	n->type = t;
+}
+
+/*
+ * initialize n to be indirect of register r; n is type t.
+ */
+void
+nodindreg(Node *n, Type *t, int r)
+{
+	nodreg(n, t, r);
+	n->op = OINDREG;
+}
+
+/*
+ * Is this node a memory operand?
+ */
+int
+ismem(Node *n)
+{
+	switch(n->op) {
+	case OITAB:
+	case OSPTR:
+	case OLEN:
+	case OCAP:
+	case OINDREG:
+	case ONAME:
+	case OPARAM:
+	case OCLOSUREVAR:
+		return 1;
+	case OADDR:
+		return thearch.thechar == '6' || thearch.thechar == '9'; // because 6g uses PC-relative addressing; TODO(rsc): not sure why 9g too
+	}
+	return 0;
+}
+
+// Sweep the prog list to mark any used nodes.
+void
+markautoused(Prog* p)
+{
+	for (; p; p = p->link) {
+		if (p->as == ATYPE || p->as == AVARDEF || p->as == AVARKILL)
+			continue;
+
+		if (p->from.node)
+			((Node*)(p->from.node))->used = 1;
+
+		if (p->to.node)
+			((Node*)(p->to.node))->used = 1;
+	}
+}
+
+// Fixup instructions after allocauto (formerly compactframe) has moved all autos around.
+void
+fixautoused(Prog *p)
+{
+	Prog **lp;
+
+	for (lp=&p; (p=*lp) != P; ) {
+		if (p->as == ATYPE && p->from.node && p->from.name == NAME_AUTO && !((Node*)(p->from.node))->used) {
+			*lp = p->link;
+			continue;
+		}
+		if ((p->as == AVARDEF || p->as == AVARKILL) && p->to.node && !((Node*)(p->to.node))->used) {
+			// Cannot remove VARDEF instruction, because - unlike TYPE handled above -
+			// VARDEFs are interspersed with other code, and a jump might be using the
+			// VARDEF as a target. Replace with a no-op instead. A later pass will remove
+			// the no-ops.
+			nopout(p);
+			continue;
+		}
+		if (p->from.name == NAME_AUTO && p->from.node)
+			p->from.offset += ((Node*)(p->from.node))->stkdelta;
+
+		if (p->to.name == NAME_AUTO && p->to.node)
+			p->to.offset += ((Node*)(p->to.node))->stkdelta;
+
+		lp = &p->link;
+	}
+}
+
+int
+samereg(Node *a, Node *b)
+{
+	if(a == N || b == N)
+		return 0;
+	if(a->op != OREGISTER)
+		return 0;
+	if(b->op != OREGISTER)
+		return 0;
+	if(a->val.u.reg != b->val.u.reg)
+		return 0;
+	return 1;
+}
+
+Node*
+nodarg(Type *t, int fp)
+{
+	Node *n;
+	NodeList *l;
+	Type *first;
+	Iter savet;
+
+	// entire argument struct, not just one arg
+	if(t->etype == TSTRUCT && t->funarg) {
+		n = nod(ONAME, N, N);
+		n->sym = lookup(".args");
+		n->type = t;
+		first = structfirst(&savet, &t);
+		if(first == nil)
+			fatal("nodarg: bad struct");
+		if(first->width == BADWIDTH)
+			fatal("nodarg: offset not computed for %T", t);
+		n->xoffset = first->width;
+		n->addable = 1;
+		goto fp;
+	}
+
+	if(t->etype != TFIELD)
+		fatal("nodarg: not field %T", t);
+	
+	if(fp == 1) {
+		for(l=curfn->dcl; l; l=l->next) {
+			n = l->n;
+			if((n->class == PPARAM || n->class == PPARAMOUT) && !isblanksym(t->sym) && n->sym == t->sym)
+				return n;
+		}
+	}
+
+	n = nod(ONAME, N, N);
+	n->type = t->type;
+	n->sym = t->sym;
+	
+	if(t->width == BADWIDTH)
+		fatal("nodarg: offset not computed for %T", t);
+	n->xoffset = t->width;
+	n->addable = 1;
+	n->orig = t->nname;
+
+fp:
+	// Rewrite argument named _ to __,
+	// or else the assignment to _ will be
+	// discarded during code generation.
+	if(isblank(n))
+		n->sym = lookup("__");
+
+	switch(fp) {
+	case 0:		// output arg
+		n->op = OINDREG;
+		n->val.u.reg = thearch.REGSP;
+		if(thearch.thechar == '5')
+			n->xoffset += 4;
+		if(thearch.thechar == '9')
+			n->xoffset += 8;
+		break;
+
+	case 1:		// input arg
+		n->class = PPARAM;
+		break;
+
+	case 2:		// offset output arg
+fatal("shouldn't be used");
+		n->op = OINDREG;
+		n->val.u.reg = thearch.REGSP;
+		n->xoffset += types[tptr]->width;
+		break;
+	}
+	n->typecheck = 1;
+	return n;
+}
+
+/*
+ * generate code to compute n;
+ * make a refer to result.
+ */
+void
+naddr(Node *n, Addr *a, int canemitcode)
+{
+	Sym *s;
+
+	*a = zprog.from;
+	if(n == N)
+		return;
+
+	if(n->type != T && n->type->etype != TIDEAL) {
+		// TODO(rsc): This is undone by the selective clearing of width below,
+		// to match architectures that were not as aggressive in setting width
+		// during naddr. Those widths must be cleared to avoid triggering
+		// failures in gins when it detects real but heretofore latent (and one
+		// hopes innocuous) type mismatches.
+		// The type mismatches should be fixed and the clearing below removed.
+		dowidth(n->type);
+		a->width = n->type->width;
+	}
+
+	switch(n->op) {
+	default:
+		fatal("naddr: bad %O %D", n->op, a);
+		break;
+
+	case OREGISTER:
+		a->type = TYPE_REG;
+		a->reg = n->val.u.reg;
+		a->sym = nil;
+		if(thearch.thechar == '8') // TODO(rsc): Never clear a->width.
+			a->width = 0;
+		break;
+
+	case OINDREG:
+		a->type = TYPE_MEM;
+		a->reg = n->val.u.reg;
+		a->sym = linksym(n->sym);
+		a->offset = n->xoffset;
+		if(a->offset != (int32)a->offset)
+			yyerror("offset %lld too large for OINDREG", a->offset);
+		if(thearch.thechar == '8') // TODO(rsc): Never clear a->width.
+			a->width = 0;
+		break;
+
+	case OPARAM:
+		// n->left is PHEAP ONAME for stack parameter.
+		// compute address of actual parameter on stack.
+		a->etype = simtype[n->left->type->etype];
+		a->width = n->left->type->width;
+		a->offset = n->xoffset;
+		a->sym = linksym(n->left->sym);
+		a->type = TYPE_MEM;
+		a->name = NAME_PARAM;
+		a->node = n->left->orig;
+		break;
+	
+	case OCLOSUREVAR:
+		if(!curfn->needctxt)
+			fatal("closurevar without needctxt");
+		a->type = TYPE_MEM;
+		a->reg = thearch.REGCTXT;
+		a->sym = nil;
+		a->offset = n->xoffset;
+		break;
+	
+	case OCFUNC:
+		naddr(n->left, a, canemitcode);
+		a->sym = linksym(n->left->sym);
+		break;
+
+	case ONAME:
+		a->etype = 0;
+		if(n->type != T)
+			a->etype = simtype[n->type->etype];
+		a->offset = n->xoffset;
+		s = n->sym;
+		a->node = n->orig;
+		//if(a->node >= (Node*)&n)
+		//	fatal("stack node");
+		if(s == S)
+			s = lookup(".noname");
+		if(n->method) {
+			if(n->type != T)
+			if(n->type->sym != S)
+			if(n->type->sym->pkg != nil)
+				s = pkglookup(s->name, n->type->sym->pkg);
+		}
+
+		a->type = TYPE_MEM;
+		switch(n->class) {
+		default:
+			fatal("naddr: ONAME class %S %d\n", n->sym, n->class);
+		case PEXTERN:
+			a->name = NAME_EXTERN;
+			break;
+		case PAUTO:
+			a->name = NAME_AUTO;
+			break;
+		case PPARAM:
+		case PPARAMOUT:
+			a->name = NAME_PARAM;
+			break;
+		case PFUNC:
+			a->name = NAME_EXTERN;
+			a->type = TYPE_ADDR;
+			a->width = widthptr;
+			s = funcsym(s);			
+			break;
+		}
+		a->sym = linksym(s);
+		break;
+
+	case OLITERAL:
+		if(thearch.thechar == '8')
+			a->width = 0;
+		switch(n->val.ctype) {
+		default:
+			fatal("naddr: const %lT", n->type);
+			break;
+		case CTFLT:
+			a->type = TYPE_FCONST;
+			a->u.dval = mpgetflt(n->val.u.fval);
+			break;
+		case CTINT:
+		case CTRUNE:
+			a->sym = nil;
+			a->type = TYPE_CONST;
+			a->offset = mpgetfix(n->val.u.xval);
+			break;
+		case CTSTR:
+			datagostring(n->val.u.sval, a);
+			break;
+		case CTBOOL:
+			a->sym = nil;
+			a->type = TYPE_CONST;
+			a->offset = n->val.u.bval;
+			break;
+		case CTNIL:
+			a->sym = nil;
+			a->type = TYPE_CONST;
+			a->offset = 0;
+			break;
+		}
+		break;
+
+	case OADDR:
+		naddr(n->left, a, canemitcode);
+		a->etype = tptr;
+		if(thearch.thechar != '5' && thearch.thechar != '9') // TODO(rsc): Do this even for arm, ppc64.
+			a->width = widthptr;
+		if(a->type != TYPE_MEM)
+			fatal("naddr: OADDR %D (from %O)", a, n->left->op);
+		a->type = TYPE_ADDR;
+		break;
+	
+	case OITAB:
+		// itable of interface value
+		naddr(n->left, a, canemitcode);
+		if(a->type == TYPE_CONST && a->offset == 0)
+			break;  // itab(nil)
+		a->etype = tptr;
+		a->width = widthptr;
+		break;
+
+	case OSPTR:
+		// pointer in a string or slice
+		naddr(n->left, a, canemitcode);
+		if(a->type == TYPE_CONST && a->offset == 0)
+			break;	// ptr(nil)
+		a->etype = simtype[tptr];
+		a->offset += Array_array;
+		a->width = widthptr;
+		break;
+
+	case OLEN:
+		// len of string or slice
+		naddr(n->left, a, canemitcode);
+		if(a->type == TYPE_CONST && a->offset == 0)
+			break;	// len(nil)
+		a->etype = simtype[TUINT];
+		if(thearch.thechar == '9')
+			a->etype = simtype[TINT];
+		a->offset += Array_nel;
+		if(thearch.thechar != '5') // TODO(rsc): Do this even on arm.
+			a->width = widthint;
+		break;
+
+	case OCAP:
+		// cap of string or slice
+		naddr(n->left, a, canemitcode);
+		if(a->type == TYPE_CONST && a->offset == 0)
+			break;	// cap(nil)
+		a->etype = simtype[TUINT];
+		if(thearch.thechar == '9')
+			a->etype = simtype[TINT];
+		a->offset += Array_cap;
+		if(thearch.thechar != '5') // TODO(rsc): Do this even on arm.
+			a->width = widthint;
+		break;
+
+//	case OADD:
+//		if(n->right->op == OLITERAL) {
+//			v = n->right->vconst;
+//			naddr(n->left, a, canemitcode);
+//		} else
+//		if(n->left->op == OLITERAL) {
+//			v = n->left->vconst;
+//			naddr(n->right, a, canemitcode);
+//		} else
+//			goto bad;
+//		a->offset += v;
+//		break;
+
+	}
+}
diff --git a/src/cmd/gc/init.c b/src/cmd/gc/init.c
index 918d37180b..c769ec27f0 100644
--- a/src/cmd/gc/init.c
+++ b/src/cmd/gc/init.c
@@ -191,5 +191,5 @@ fninit(NodeList *n)
 	typecheck(&fn, Etop);
 	typechecklist(r, Etop);
 	curfn = nil;
-	funccompile(fn, 0);
+	funccompile(fn);
 }
diff --git a/src/cmd/gc/lex.c b/src/cmd/gc/lex.c
index 042099bd5e..a4b832aa0b 100644
--- a/src/cmd/gc/lex.c
+++ b/src/cmd/gc/lex.c
@@ -95,7 +95,7 @@ enum
 void
 usage(void)
 {
-	print("usage: %cg [options] file.go...\n", arch.thechar);
+	print("usage: %cg [options] file.go...\n", thearch.thechar);
 	flagprint(1);
 	exits("usage");
 }
@@ -140,7 +140,7 @@ doversion(void)
 	sep = "";
 	if(*p)
 		sep = " ";
-	print("%cg version %s%s%s\n", arch.thechar, getgoversion(), sep, p);
+	print("%cg version %s%s%s\n", thearch.thechar, getgoversion(), sep, p);
 	exits(0);
 }
 
@@ -161,15 +161,15 @@ gcmain(int argc, char *argv[])
 	// Tell the FPU to handle all exceptions.
 	setfcr(FPPDBL|FPRNR);
 #endif
-	// Allow GOARCH=arch.thestring or GOARCH=arch.thestringsuffix,
+	// Allow GOARCH=thearch.thestring or GOARCH=thearch.thestringsuffix,
 	// but not other values.	
 	p = getgoarch();
-	if(strncmp(p, arch.thestring, strlen(arch.thestring)) != 0)
-		sysfatal("cannot use %cg with GOARCH=%s", arch.thechar, p);
+	if(strncmp(p, thearch.thestring, strlen(thearch.thestring)) != 0)
+		sysfatal("cannot use %cg with GOARCH=%s", thearch.thechar, p);
 	goarch = p;
 
-	arch.linkarchinit();
-	ctxt = linknew(arch.thelinkarch);
+	thearch.linkarchinit();
+	ctxt = linknew(thearch.thelinkarch);
 	ctxt->diag = yyerror;
 	ctxt->bso = &bstdout;
 	Binit(&bstdout, 1, OWRITE);
@@ -179,6 +179,7 @@ gcmain(int argc, char *argv[])
 	
 	// pseudo-package, for scoping
 	builtinpkg = mkpkg(newstrlit("go.builtin"));
+	builtinpkg->prefix = "go.builtin"; // not go%2ebuiltin
 
 	// pseudo-package, accessed by import "unsafe"
 	unsafepkg = mkpkg(newstrlit("unsafe"));
@@ -267,7 +268,7 @@ gcmain(int argc, char *argv[])
 	flagcount("wb", "enable write barrier", &use_writebarrier);
 	flagcount("x", "debug lexer", &debug['x']);
 	flagcount("y", "debug declarations in canned imports (with -d)", &debug['y']);
-	if(arch.thechar == '6')
+	if(thearch.thechar == '6')
 		flagcount("largemodel", "generate code that assumes a large memory model", &flag_largemodel);
 
 	flagparse(&argc, &argv, usage);
@@ -308,7 +309,7 @@ gcmain(int argc, char *argv[])
 	if(debug['l'] <= 1)
 		debug['l'] = 1 - debug['l'];
 
-	if(arch.thechar == '8') {
+	if(thearch.thechar == '8') {
 		p = getgo386();
 		if(strcmp(p, "387") == 0)
 			use_sse = 0;
@@ -319,7 +320,7 @@ gcmain(int argc, char *argv[])
 	}
 
 	fmtinstallgo();
-	arch.betypeinit();
+	thearch.betypeinit();
 	if(widthptr == 0)
 		fatal("betypeinit failed");
 
@@ -404,8 +405,9 @@ gcmain(int argc, char *argv[])
 		}
 	}
 
-	// Phase 4: Decide how to capture variables
-	// and transform closure bodies accordingly.
+	// Phase 4: Decide how to capture closed variables.
+	// This needs to run before escape analysis,
+	// because variables captured by value do not escape.
 	for(l=xtop; l; l=l->next) {
 		if(l->n->op == ODCLFUNC && l->n->closure) {
 			curfn = l->n;
@@ -456,15 +458,26 @@ gcmain(int argc, char *argv[])
 	// Move large values off stack too.
 	movelarge(xtop);
 
-	// Phase 7: Compile top level functions.
+	// Phase 7: Transform closure bodies to properly reference captured variables.
+	// This needs to happen before walk, because closures must be transformed
+	// before walk reaches a call of a closure.
+	for(l=xtop; l; l=l->next) {
+		if(l->n->op == ODCLFUNC && l->n->closure) {
+			curfn = l->n;
+			transformclosure(l->n);
+		}
+	}
+	curfn = N;
+
+	// Phase 8: Compile top level functions.
 	for(l=xtop; l; l=l->next)
 		if(l->n->op == ODCLFUNC)
-			funccompile(l->n, 0);
+			funccompile(l->n);
 
 	if(nsavederrors+nerrors == 0)
 		fninit(xtop);
 
-	// Phase 8: Check external declarations.
+	// Phase 9: Check external declarations.
 	for(l=externdcl; l; l=l->next)
 		if(l->n->op == ONAME)
 			typecheck(&l->n, Erv);
@@ -588,7 +601,7 @@ findpkg(Strlit *name)
 		snprint(namebuf, sizeof(namebuf), "%Z.a", name);
 		if(access(namebuf, 0) >= 0)
 			return 1;
-		snprint(namebuf, sizeof(namebuf), "%Z.%c", name, arch.thechar);
+		snprint(namebuf, sizeof(namebuf), "%Z.%c", name, thearch.thechar);
 		if(access(namebuf, 0) >= 0)
 			return 1;
 		return 0;
@@ -610,7 +623,7 @@ findpkg(Strlit *name)
 		snprint(namebuf, sizeof(namebuf), "%s/%Z.a", p->dir, name);
 		if(access(namebuf, 0) >= 0)
 			return 1;
-		snprint(namebuf, sizeof(namebuf), "%s/%Z.%c", p->dir, name, arch.thechar);
+		snprint(namebuf, sizeof(namebuf), "%s/%Z.%c", p->dir, name, thearch.thechar);
 		if(access(namebuf, 0) >= 0)
 			return 1;
 	}
@@ -627,7 +640,7 @@ findpkg(Strlit *name)
 		snprint(namebuf, sizeof(namebuf), "%s/pkg/%s_%s%s%s/%Z.a", goroot, goos, goarch, suffixsep, suffix, name);
 		if(access(namebuf, 0) >= 0)
 			return 1;
-		snprint(namebuf, sizeof(namebuf), "%s/pkg/%s_%s%s%s/%Z.%c", goroot, goos, goarch, suffixsep, suffix, name, arch.thechar);
+		snprint(namebuf, sizeof(namebuf), "%s/pkg/%s_%s%s%s/%Z.%c", goroot, goos, goarch, suffixsep, suffix, name, thearch.thechar);
 		if(access(namebuf, 0) >= 0)
 			return 1;
 	}
@@ -647,7 +660,7 @@ importfile(Val *f, int line)
 	Biobuf *imp;
 	char *file, *p, *q, *tag;
 	int32 c;
-	int len;
+	int n;
 	Strlit *path;
 	char *cleanbuf, *prefix;
 
@@ -745,8 +758,8 @@ importfile(Val *f, int line)
 	}
 	file = strdup(namebuf);
 
-	len = strlen(namebuf);
-	if(len > 2 && namebuf[len-2] == '.' && namebuf[len-1] == 'a') {
+	n = strlen(namebuf);
+	if(n > 2 && namebuf[n-2] == '.' && namebuf[n-1] == 'a') {
 		if(!skiptopkgdef(imp)) {
 			yyerror("import %s: not a package file", file);
 			errorexit();
@@ -770,7 +783,7 @@ importfile(Val *f, int line)
 
 	// assume files move (get installed)
 	// so don't record the full path.
-	linehist(file + len - path->len - 2, -1, 1);	// acts as #pragma lib
+	linehist(file + n - path->len - 2, -1, 1);	// acts as #pragma lib
 
 	/*
 	 * position the input right
@@ -974,17 +987,7 @@ l0:
 			rune = c;
 			clen += runetochar(cp+clen, &rune);
 		}
-
-	strlit:
-		*(int32*)cp = clen-sizeof(int32);	// length
-		do {
-			cp[clen++] = 0;
-		} while(clen & MAXALIGN);
-		yylval.val.u.sval = (Strlit*)cp;
-		yylval.val.ctype = CTSTR;
-		DBG("lex: string literal\n");
-		strcpy(litbuf, "string literal");
-		return LLITERAL;
+		goto strlit;
 
 	case '\'':
 		/* '.' */
@@ -1279,7 +1282,7 @@ talph:
 		if(c >= Runeself) {
 			ungetc(c);
 			rune = getr();
-			// 0xb7 · is used for internal names
+			// 0xb7 · is used for internal names
 			if(!isalpharune(rune) && !isdigitrune(rune) && (importpkg == nil || rune != 0xb7))
 				yyerror("invalid identifier character U+%04x", rune);
 			cp += runetochar(cp, &rune);
@@ -1467,6 +1470,17 @@ caseout:
 	strcpy(litbuf, "literal ");
 	strcat(litbuf, lexbuf);
 	return LLITERAL;
+
+strlit:
+	*(int32*)cp = clen-sizeof(int32);	// length
+	do {
+		cp[clen++] = 0;
+	} while(clen & MAXALIGN);
+	yylval.val.u.sval = (Strlit*)cp;
+	yylval.val.ctype = CTSTR;
+	DBG("lex: string literal\n");
+	strcpy(litbuf, "string literal");
+	return LLITERAL;
 }
 
 static void pragcgo(char*);
@@ -2270,10 +2284,10 @@ lexfini(void)
 	}
 
 	// backend-specific builtin types (e.g. int).
-	for(i=0; arch.typedefs[i].name; i++) {
-		s = lookup(arch.typedefs[i].name);
+	for(i=0; thearch.typedefs[i].name; i++) {
+		s = lookup(thearch.typedefs[i].name);
 		if(s->def == N) {
-			s->def = typenod(types[arch.typedefs[i].etype]);
+			s->def = typenod(types[thearch.typedefs[i].etype]);
 			s->origpkg = builtinpkg;
 		}
 	}
@@ -2571,6 +2585,6 @@ mkpackage(char* pkgname)
 		p = strrchr(namebuf, '.');
 		if(p != nil)
 			*p = 0;
-		outfile = smprint("%s.%c", namebuf, arch.thechar);
+		outfile = smprint("%s.%c", namebuf, thearch.thechar);
 	}
 }
diff --git a/src/cmd/gc/mkbuiltin b/src/cmd/gc/mkbuiltin
index 1dab1c9a05..696aa82424 100755
--- a/src/cmd/gc/mkbuiltin
+++ b/src/cmd/gc/mkbuiltin
@@ -20,6 +20,10 @@ GC=${GOCHAR}g
 gcc -o mkbuiltin1 mkbuiltin1.c
 rm -f _builtin.c
 echo "// AUTO-GENERATED by mkbuiltin; DO NOT EDIT" >>_builtin.c
+echo "#include <u.h>" >>_builtin.c
+echo "#include <libc.h>" >>_builtin.c
+echo '#include "go.h"' >>_builtin.c
+
 for i in runtime unsafe
 do
 	go tool $GC -A $i.go
diff --git a/src/cmd/gc/mparith1.c b/src/cmd/gc/mparith1.c
index d33a81e09d..6a0eb2d6d9 100644
--- a/src/cmd/gc/mparith1.c
+++ b/src/cmd/gc/mparith1.c
@@ -235,9 +235,9 @@ mppow10flt(Mpflt *a, int p)
 static void
 mphextofix(Mpint *a, char *s, int n)
 {
-	char *hexdigitp, *end, c;
+	char c;
 	long d;
-	int bit;
+	int bit, hexdigitp, end;
 
 	while(*s == '0') {
 		s++;
@@ -250,9 +250,9 @@ mphextofix(Mpint *a, char *s, int n)
 		return;
 	}
 
-	end = s+n-1;
-	for(hexdigitp=end; hexdigitp>=s; hexdigitp--) {
-		c = *hexdigitp;
+	end = n-1;
+	for(hexdigitp=end; hexdigitp>=0; hexdigitp--) {
+		c = s[hexdigitp];
 		if(c >= '0' && c <= '9')
 			d = c-'0';
 		else if(c >= 'A' && c <= 'F')
@@ -334,18 +334,24 @@ mpatoflt(Mpflt *a, char *as)
 				break;
 			}
 		}
-		if(start == nil)
+		if(start == nil) {
+			yyerror("malformed hex constant: %s", as);
 			goto bad;
+		}
 
 		mphextofix(&a->val, start, s-start);
-		if(a->val.ovf)
+		if(a->val.ovf) {
+			yyerror("constant too large: %s", as);
 			goto bad;
+		}
 		a->exp = 0;
 		mpnorm(a);
 	}
 	for(;;) {
-		switch(c = *s++) {
+		c = *s++;
+		switch(c) {
 		default:
+			yyerror("malformed constant: %s (at %c)", as, c);
 			goto bad;
 
 		case '-':
@@ -357,8 +363,10 @@ mpatoflt(Mpflt *a, char *as)
 			continue;
 
 		case '.':
-			if(base == 16)
+			if(base == 16) {
+				yyerror("decimal point in hex constant: %s", as);
 				goto bad;
+			}
 			dp = 1;
 			continue;
 
@@ -414,8 +422,10 @@ mpatoflt(Mpflt *a, char *as)
 	}
 
 	if(eb) {
-		if(dp)
+		if(dp) {
+			yyerror("decimal point and binary point in constant: %s", as);
 			goto bad;
+		}
 		mpsetexp(a, a->exp+ex);
 		goto out;
 	}
@@ -444,7 +454,6 @@ out:
 	return;
 
 bad:
-	yyerror("constant too large: %s", as);
 	mpmovecflt(a, 0.0);
 }
 
@@ -483,6 +492,7 @@ mpatofix(Mpint *a, char *as)
 			c = *s++;
 			continue;
 		}
+		yyerror("malformed decimal constant: %s", as);
 		goto bad;
 	}
 	goto out;
@@ -498,6 +508,7 @@ oct:
 			c = *s++;
 			continue;
 		}
+		yyerror("malformed octal constant: %s", as);
 		goto bad;
 	}
 	goto out;
@@ -511,11 +522,14 @@ hex:
 			c = *s;
 			continue;
 		}
+		yyerror("malformed hex constant: %s", as);
 		goto bad;
 	}
 	mphextofix(a, s0, s-s0);
-	if(a->ovf)
+	if(a->ovf) {
+		yyerror("constant too large: %s", as);
 		goto bad;
+	}
 
 out:
 	if(f)
@@ -523,14 +537,14 @@ out:
 	return;
 
 bad:
-	yyerror("constant too large: %s", as);
 	mpmovecfix(a, 0);
 }
 
 int
 Bconv(Fmt *fp)
 {
-	char buf[500], *p;
+	char buf[500];
+	int p;
 	Mpint *xval, q, r, ten, sixteen;
 	int f, digit;
 
@@ -542,8 +556,8 @@ Bconv(Fmt *fp)
 		mpnegfix(&q);
 	}
 
-	p = &buf[sizeof(buf)];
-	*--p = 0;
+	p = sizeof(buf);
+	buf[--p] = 0;
 	if(fp->flags & FmtSharp) {
 		// Hexadecimal
 		mpmovecfix(&sixteen, 16);
@@ -551,27 +565,27 @@ Bconv(Fmt *fp)
 			mpdivmodfixfix(&q, &r, &q, &sixteen);
 			digit = mpgetfix(&r);
 			if(digit < 10)
-				*--p = digit + '0';
+				buf[--p] = digit + '0';
 			else
-				*--p = digit - 10 + 'A';
+				buf[--p] = digit - 10 + 'A';
 			if(mptestfix(&q) <= 0)
 				break;
 		}
-		*--p = 'x';
-		*--p = '0';
+		buf[--p] = 'x';
+		buf[--p] = '0';
 	} else {
 		// Decimal
 		mpmovecfix(&ten, 10);
 		for(;;) {
 			mpdivmodfixfix(&q, &r, &q, &ten);
-			*--p = mpgetfix(&r) + '0';
+			buf[--p] = mpgetfix(&r) + '0';
 			if(mptestfix(&q) <= 0)
 				break;
 		}
 	}
 	if(f)
-		*--p = '-';
-	return fmtstrcpy(fp, p);
+		buf[--p] = '-';
+	return fmtstrcpy(fp, &buf[p]);
 }
 
 int
diff --git a/src/cmd/gc/mparith2.c b/src/cmd/gc/mparith2.c
index fd9f591cea..37aafbb5f5 100644
--- a/src/cmd/gc/mparith2.c
+++ b/src/cmd/gc/mparith2.c
@@ -14,12 +14,10 @@ static int
 mplen(Mpint *a)
 {
 	int i, n;
-	long *a1;
 
 	n = -1;
-	a1 = &a->a[0];
 	for(i=0; i<Mpprec; i++) {
-		if(*a1++ != 0)
+		if(a->a[i] != 0)
 			n = i;
 	}
 	return n+1;
@@ -32,19 +30,18 @@ mplen(Mpint *a)
 static void
 mplsh(Mpint *a, int quiet)
 {
-	long *a1, x;
+	long x;
 	int i, c;
 
 	c = 0;
-	a1 = &a->a[0];
 	for(i=0; i<Mpprec; i++) {
-		x = (*a1 << 1) + c;
+		x = (a->a[i] << 1) + c;
 		c = 0;
 		if(x >= Mpbase) {
 			x -= Mpbase;
 			c = 1;
 		}
-		*a1++ = x;
+		a->a[i] = x;
 	}
 	a->ovf = c;
 	if(a->ovf && !quiet)
@@ -58,20 +55,17 @@ mplsh(Mpint *a, int quiet)
 static void
 mplshw(Mpint *a, int quiet)
 {
-	long *a1;
 	int i;
 
-	a1 = &a->a[Mpprec-1];
-	if(*a1) {
+	i = Mpprec-1;
+	if(a->a[i]) {
 		a->ovf = 1;
 		if(!quiet)
 			yyerror("constant shift overflow");
 	}
-	for(i=1; i<Mpprec; i++) {
-		a1[0] = a1[-1];
-		a1--;
-	}
-	a1[0] = 0;
+	for(; i > 0; i--)
+		a->a[i] = a->a[i-1];
+	a->a[i] = 0;
 }
 
 //
@@ -81,15 +75,14 @@ mplshw(Mpint *a, int quiet)
 static void
 mprsh(Mpint *a)
 {
-	long *a1, x, lo;
+	long x, lo;
 	int i, c;
 
 	c = 0;
 	lo = a->a[0] & 1;
-	a1 = &a->a[Mpprec];
-	for(i=0; i<Mpprec; i++) {
-		x = *--a1;
-		*a1 = (x + c) >> 1;
+	for(i=Mpprec-1; i>=0; i--) {
+		x = a->a[i];
+		a->a[i] = (x + c) >> 1;
 		c = 0;
 		if(x & 1)
 			c = Mpbase;
@@ -105,16 +98,14 @@ mprsh(Mpint *a)
 static void
 mprshw(Mpint *a)
 {
-	long *a1, lo;
+	long lo;
 	int i;
 
 	lo = a->a[0];
-	a1 = &a->a[0];
-	for(i=1; i<Mpprec; i++) {
-		a1[0] = a1[1];
-		a1++;
+	for(i=0; i<Mpprec-1; i++) {
+		a->a[i] = a->a[i+1];
 	}
-	a1[0] = 0;
+	a->a[i] = 0;
 	if(a->neg && lo != 0)
 		mpaddcfix(a, -1);
 }
@@ -125,7 +116,7 @@ mprshw(Mpint *a)
 static int
 mpcmp(Mpint *a, Mpint *b)
 {
-	long x, *a1, *b1;
+	long x;
 	int i;
 
 	if(a->ovf || b->ovf) {
@@ -134,11 +125,8 @@ mpcmp(Mpint *a, Mpint *b)
 		return 0;
 	}
 
-	a1 = &a->a[0] + Mpprec;
-	b1 = &b->a[0] + Mpprec;
-
-	for(i=0; i<Mpprec; i++) {
-		x = *--a1 - *--b1;
+	for(i=Mpprec-1; i>=0; i--) {
+		x = a->a[i] - b->a[i];
 		if(x > 0)
 			return +1;
 		if(x < 0)
@@ -154,19 +142,18 @@ mpcmp(Mpint *a, Mpint *b)
 static void
 mpneg(Mpint *a)
 {
-	long x, *a1;
+	long x;
 	int i, c;
 
-	a1 = &a->a[0];
 	c = 0;
 	for(i=0; i<Mpprec; i++) {
-		x = -*a1 -c;
+		x = -a->a[i] -c;
 		c = 0;
 		if(x < 0) {
 			x += Mpbase;
 			c = 1;
 		}
-		*a1++ = x;
+		a->a[i] = x;
 	}
 }
 
@@ -202,7 +189,7 @@ void
 mpaddfixfix(Mpint *a, Mpint *b, int quiet)
 {
 	int i, c;
-	long x, *a1, *b1;
+	long x;
 
 	if(a->ovf || b->ovf) {
 		if(nsavederrors+nerrors == 0)
@@ -212,20 +199,18 @@ mpaddfixfix(Mpint *a, Mpint *b, int quiet)
 	}
 
 	c = 0;
-	a1 = &a->a[0];
-	b1 = &b->a[0];
 	if(a->neg != b->neg)
 		goto sub;
 
 	// perform a+b
 	for(i=0; i<Mpprec; i++) {
-		x = *a1 + *b1++ + c;
+		x = a->a[i] + b->a[i] + c;
 		c = 0;
 		if(x >= Mpbase) {
 			x -= Mpbase;
 			c = 1;
 		}
-		*a1++ = x;
+		a->a[i] = x;
 	}
 	a->ovf = c;
 	if(a->ovf && !quiet)
@@ -242,26 +227,26 @@ sub:
 
 	case 1:
 		for(i=0; i<Mpprec; i++) {
-			x = *a1 - *b1++ - c;
+			x = a->a[i] - b->a[i] - c;
 			c = 0;
 			if(x < 0) {
 				x += Mpbase;
 				c = 1;
 			}
-			*a1++ = x;
+			a->a[i] = x;
 		}
 		break;
 
 	case -1:
 		a->neg ^= 1;
 		for(i=0; i<Mpprec; i++) {
-			x = *b1++ - *a1 - c;
+			x = b->a[i] - a->a[i] - c;
 			c = 0;
 			if(x < 0) {
 				x += Mpbase;
 				c = 1;
 			}
-			*a1++ = x;
+			a->a[i] = x;
 		}
 		break;
 	}
@@ -272,8 +257,9 @@ mpmulfixfix(Mpint *a, Mpint *b)
 {
 
 	int i, j, na, nb;
-	long *a1, x;
+	long x;
 	Mpint s, q;
+	Mpint *c;
 
 	if(a->ovf || b->ovf) {
 		if(nsavederrors+nerrors == 0)
@@ -288,17 +274,17 @@ mpmulfixfix(Mpint *a, Mpint *b)
 	nb = mplen(b);
 	if(na > nb) {
 		mpmovefixfix(&s, a);
-		a1 = &b->a[0];
+		c = b;
 		na = nb;
 	} else {
 		mpmovefixfix(&s, b);
-		a1 = &a->a[0];
+		c = a;
 	}
 	s.neg = 0;
 
 	mpmovecfix(&q, 0);
 	for(i=0; i<na; i++) {
-		x = *a1++;
+		x = c->a[i];
 		for(j=0; j<Mpscale; j++) {
 			if(x & 1) {
 				if(s.ovf) {
@@ -326,7 +312,7 @@ mpmulfract(Mpint *a, Mpint *b)
 {
 
 	int i, j;
-	long *a1, x;
+	long x;
 	Mpint s, q;
 
 	if(a->ovf || b->ovf) {
@@ -337,16 +323,16 @@ mpmulfract(Mpint *a, Mpint *b)
 	}
 
 	mpmovefixfix(&s, b);
-	a1 = &a->a[Mpprec];
 	s.neg = 0;
 	mpmovecfix(&q, 0);
 
-	x = *--a1;
+	i = Mpprec-1;
+	x = a->a[i];
 	if(x != 0)
 		yyerror("mpmulfract not normal");
 
-	for(i=0; i<Mpprec-1; i++) {
-		x = *--a1;
+	for(i--; i >= 0; i--) {
+		x = a->a[i];
 		if(x == 0) {
 			mprshw(&s);
 			continue;
@@ -369,7 +355,7 @@ void
 mporfixfix(Mpint *a, Mpint *b)
 {
 	int i;
-	long x, *a1, *b1;
+	long x;
 
 	x = 0;
 	if(a->ovf || b->ovf) {
@@ -386,11 +372,9 @@ mporfixfix(Mpint *a, Mpint *b)
 	if(b->neg)
 		mpneg(b);
 
-	a1 = &a->a[0];
-	b1 = &b->a[0];
 	for(i=0; i<Mpprec; i++) {
-		x = *a1 | *b1++;
-		*a1++ = x;
+		x = a->a[i] | b->a[i];
+		a->a[i] = x;
 	}
 
 	if(b->neg)
@@ -405,7 +389,7 @@ void
 mpandfixfix(Mpint *a, Mpint *b)
 {
 	int i;
-	long x, *a1, *b1;
+	long x;
 
 	x = 0;
 	if(a->ovf || b->ovf) {
@@ -422,11 +406,9 @@ mpandfixfix(Mpint *a, Mpint *b)
 	if(b->neg)
 		mpneg(b);
 
-	a1 = &a->a[0];
-	b1 = &b->a[0];
 	for(i=0; i<Mpprec; i++) {
-		x = *a1 & *b1++;
-		*a1++ = x;
+		x = a->a[i] & b->a[i];
+		a->a[i] = x;
 	}
 
 	if(b->neg)
@@ -441,7 +423,7 @@ void
 mpandnotfixfix(Mpint *a, Mpint *b)
 {
 	int i;
-	long x, *a1, *b1;
+	long x;
 
 	x = 0;
 	if(a->ovf || b->ovf) {
@@ -458,11 +440,9 @@ mpandnotfixfix(Mpint *a, Mpint *b)
 	if(b->neg)
 		mpneg(b);
 
-	a1 = &a->a[0];
-	b1 = &b->a[0];
 	for(i=0; i<Mpprec; i++) {
-		x = *a1 & ~*b1++;
-		*a1++ = x;
+		x = a->a[i] & ~b->a[i];
+		a->a[i] = x;
 	}
 
 	if(b->neg)
@@ -477,7 +457,7 @@ void
 mpxorfixfix(Mpint *a, Mpint *b)
 {
 	int i;
-	long x, *a1, *b1;
+	long x;
 
 	x = 0;
 	if(a->ovf || b->ovf) {
@@ -494,11 +474,9 @@ mpxorfixfix(Mpint *a, Mpint *b)
 	if(b->neg)
 		mpneg(b);
 
-	a1 = &a->a[0];
-	b1 = &b->a[0];
 	for(i=0; i<Mpprec; i++) {
-		x = *a1 ^ *b1++;
-		*a1++ = x;
+		x = a->a[i] ^ b->a[i];
+		a->a[i] = x;
 	}
 
 	if(b->neg)
@@ -585,7 +563,6 @@ void
 mpmovecfix(Mpint *a, vlong c)
 {
 	int i;
-	long *a1;
 	vlong x;
 
 	a->neg = 0;
@@ -597,9 +574,8 @@ mpmovecfix(Mpint *a, vlong c)
 		x = -(uvlong)x;
 	}
 
-	a1 = &a->a[0];
 	for(i=0; i<Mpprec; i++) {
-		*a1++ = x&Mpmask;
+		a->a[i] = x&Mpmask;
 		x >>= Mpscale;
 	}
 }
@@ -658,13 +634,11 @@ mpdivmodfixfix(Mpint *q, Mpint *r, Mpint *n, Mpint *d)
 static int
 mpiszero(Mpint *a)
 {
-	long *a1;
 	int i;
-	a1 = &a->a[0] + Mpprec;
-	for(i=0; i<Mpprec; i++) {
-		if(*--a1 != 0)
+
+	for(i=Mpprec-1; i>=0; i--)
+		if(a->a[i] != 0)
 			return 0;
-	}
 	return 1;
 }
 
@@ -673,16 +647,15 @@ mpdivfract(Mpint *a, Mpint *b)
 {
 	Mpint n, d;
 	int i, j, neg;
-	long *a1, x;
+	long x;
 
 	mpmovefixfix(&n, a);	// numerator
 	mpmovefixfix(&d, b);	// denominator
-	a1 = &a->a[Mpprec];	// quotient
 
 	neg = n.neg ^ d.neg;
 	n.neg = 0;
 	d.neg = 0;
-	for(i=0; i<Mpprec; i++) {
+	for(i=Mpprec-1; i >= 0; i--) {
 		x = 0;
 		for(j=0; j<Mpscale; j++) {
 			x <<= 1;
@@ -693,7 +666,7 @@ mpdivfract(Mpint *a, Mpint *b)
 			}
 			mprsh(&d);
 		}
-		*--a1 = x;
+		a->a[i] = x;
 	}
 	a->neg = neg;
 }
diff --git a/src/cmd/gc/obj.c b/src/cmd/gc/obj.c
index a4d470615a..3983f99d6d 100644
--- a/src/cmd/gc/obj.c
+++ b/src/cmd/gc/obj.c
@@ -95,9 +95,9 @@ dumpobj(void)
 	externdcl = tmp;
 
 	zero = pkglookup("zerovalue", runtimepkg);
-	arch.ggloblsym(zero, zerosize, DUPOK|RODATA);
+	ggloblsym(zero, zerosize, DUPOK|RODATA);
 
-	arch.dumpdata();
+	dumpdata();
 	writeobj(ctxt, bout);
 
 	if(writearchive) {
@@ -106,7 +106,7 @@ dumpobj(void)
 		if(size&1)
 			Bputc(bout, 0);
 		Bseek(bout, startobj - ArhdrSize, 0);
-		snprint(namebuf, sizeof namebuf, "_go_.%c", arch.thechar);
+		snprint(namebuf, sizeof namebuf, "_go_.%c", thearch.thechar);
 		formathdr(arhdr, namebuf, size);
 		Bwrite(bout, arhdr, ArhdrSize);
 	}
@@ -133,13 +133,13 @@ dumpglobls(void)
 			continue;
 		dowidth(n->type);
 
-		arch.ggloblnod(n);
+		ggloblnod(n);
 	}
 	
 	for(l=funcsyms; l; l=l->next) {
 		n = l->n;
-		arch.dsymptr(n->sym, 0, n->sym->def->shortname->sym, 0);
-		arch.ggloblsym(n->sym, widthptr, DUPOK|RODATA);
+		dsymptr(n->sym, 0, n->sym->def->shortname->sym, 0);
+		ggloblsym(n->sym, widthptr, DUPOK|RODATA);
 	}
 	
 	// Do not reprocess funcsyms on next dumpglobls call.
@@ -250,7 +250,7 @@ stringsym(char *s, int len)
 	off = 0;
 	
 	// string header
-	off = arch.dsymptr(sym, off, sym, widthptr+widthint);
+	off = dsymptr(sym, off, sym, widthptr+widthint);
 	off = duintxx(sym, off, len, widthint);
 	
 	// string data
@@ -258,11 +258,11 @@ stringsym(char *s, int len)
 		m = 8;
 		if(m > len-n)
 			m = len-n;
-		off = arch.dsname(sym, off, s+n, m);
+		off = dsname(sym, off, s+n, m);
 	}
 	off = duint8(sym, off, 0);  // terminating NUL for runtime
 	off = (off+widthptr-1)&~(widthptr-1);  // round to pointer alignment
-	arch.ggloblsym(sym, off, DUPOK|RODATA);
+	ggloblsym(sym, off, DUPOK|RODATA);
 
 	return sym;	
 }
@@ -283,14 +283,216 @@ slicebytes(Node *nam, char *s, int len)
 		m = 8;
 		if(m > len-n)
 			m = len-n;
-		off = arch.dsname(sym, off, s+n, m);
+		off = dsname(sym, off, s+n, m);
 	}
-	arch.ggloblsym(sym, off, NOPTR);
+	ggloblsym(sym, off, NOPTR);
 	
 	if(nam->op != ONAME)
 		fatal("slicebytes %N", nam);
 	off = nam->xoffset;
-	off = arch.dsymptr(nam->sym, off, sym, 0);
+	off = dsymptr(nam->sym, off, sym, 0);
 	off = duintxx(nam->sym, off, len, widthint);
 	duintxx(nam->sym, off, len, widthint);
 }
+
+int
+dsname(Sym *s, int off, char *t, int n)
+{
+	Prog *p;
+
+	p = thearch.gins(ADATA, N, N);
+	p->from.type = TYPE_MEM;
+	p->from.name = NAME_EXTERN;
+	p->from.offset = off;
+	p->from.sym = linksym(s);
+	p->from3.type = TYPE_CONST;
+	p->from3.offset = n;
+	
+	p->to.type = TYPE_SCONST;
+	memmove(p->to.u.sval, t, n);
+	return off + n;
+}
+
+/*
+ * make a refer to the data s, s+len
+ * emitting DATA if needed.
+ */
+void
+datastring(char *s, int len, Addr *a)
+{
+	Sym *sym;
+	
+	sym = stringsym(s, len);
+	a->type = TYPE_MEM;
+	a->name = NAME_EXTERN;
+	a->sym = linksym(sym);
+	a->node = sym->def;
+	a->offset = widthptr+widthint;  // skip header
+	a->etype = simtype[TINT];
+}
+
+/*
+ * make a refer to the string sval,
+ * emitting DATA if needed.
+ */
+void
+datagostring(Strlit *sval, Addr *a)
+{
+	Sym *sym;
+
+	sym = stringsym(sval->s, sval->len);
+	a->type = TYPE_MEM;
+	a->name = NAME_EXTERN;
+	a->sym = linksym(sym);
+	a->node = sym->def;
+	a->offset = 0;  // header
+	a->etype = TSTRING;
+}
+
+void
+gdata(Node *nam, Node *nr, int wid)
+{
+	Prog *p;
+
+	if(nr->op == OLITERAL) {
+		switch(nr->val.ctype) {
+		case CTCPLX:
+			gdatacomplex(nam, nr->val.u.cval);
+			return;
+		case CTSTR:
+			gdatastring(nam, nr->val.u.sval);
+			return;
+		}
+	}
+	p = thearch.gins(ADATA, nam, nr);
+	p->from3.type = TYPE_CONST;
+	p->from3.offset = wid;
+}
+
+void
+gdatacomplex(Node *nam, Mpcplx *cval)
+{
+	Prog *p;
+	int w;
+
+	w = cplxsubtype(nam->type->etype);
+	w = types[w]->width;
+
+	p = thearch.gins(ADATA, nam, N);
+	p->from3.type = TYPE_CONST;
+	p->from3.offset = w;
+	p->to.type = TYPE_FCONST;
+	p->to.u.dval = mpgetflt(&cval->real);
+
+	p = thearch.gins(ADATA, nam, N);
+	p->from3.type = TYPE_CONST;
+	p->from3.offset = w;
+	p->from.offset += w;
+	p->to.type = TYPE_FCONST;
+	p->to.u.dval = mpgetflt(&cval->imag);
+}
+
+void
+gdatastring(Node *nam, Strlit *sval)
+{
+	Prog *p;
+	Node nod1;
+
+	p = thearch.gins(ADATA, nam, N);
+	datastring(sval->s, sval->len, &p->to);
+	p->from3.type = TYPE_CONST;
+	p->from3.offset = types[tptr]->width;
+	p->to.type = TYPE_ADDR;
+//print("%P\n", p);
+
+	nodconst(&nod1, types[TINT], sval->len);
+	p = thearch.gins(ADATA, nam, &nod1);
+	p->from3.type = TYPE_CONST;
+	p->from3.offset = widthint;
+	p->from.offset += widthptr;
+}
+
+int
+dstringptr(Sym *s, int off, char *str)
+{
+	Prog *p;
+
+	off = rnd(off, widthptr);
+	p = thearch.gins(ADATA, N, N);
+	p->from.type = TYPE_MEM;
+	p->from.name = NAME_EXTERN;
+	p->from.sym = linksym(s);
+	p->from.offset = off;
+	p->from3.type = TYPE_CONST;
+	p->from3.offset = widthptr;
+
+	datastring(str, strlen(str)+1, &p->to);
+	p->to.type = TYPE_ADDR;
+	p->to.etype = simtype[TINT];
+	off += widthptr;
+
+	return off;
+}
+
+int
+dgostrlitptr(Sym *s, int off, Strlit *lit)
+{
+	Prog *p;
+
+	if(lit == nil)
+		return duintptr(s, off, 0);
+
+	off = rnd(off, widthptr);
+	p = thearch.gins(ADATA, N, N);
+	p->from.type = TYPE_MEM;
+	p->from.name = NAME_EXTERN;
+	p->from.sym = linksym(s);
+	p->from.offset = off;
+	p->from3.type = TYPE_CONST;
+	p->from3.offset = widthptr;
+	datagostring(lit, &p->to);
+	p->to.type = TYPE_ADDR;
+	p->to.etype = simtype[TINT];
+	off += widthptr;
+
+	return off;
+}
+
+int
+dgostringptr(Sym *s, int off, char *str)
+{
+	int n;
+	Strlit *lit;
+
+	if(str == nil)
+		return duintptr(s, off, 0);
+
+	n = strlen(str);
+	lit = mal(sizeof *lit + n);
+	strcpy(lit->s, str);
+	lit->len = n;
+	return dgostrlitptr(s, off, lit);
+}
+
+int
+dsymptr(Sym *s, int off, Sym *x, int xoff)
+{
+	Prog *p;
+
+	off = rnd(off, widthptr);
+
+	p = thearch.gins(ADATA, N, N);
+	p->from.type = TYPE_MEM;
+	p->from.name = NAME_EXTERN;
+	p->from.sym = linksym(s);
+	p->from.offset = off;
+	p->from3.type = TYPE_CONST;
+	p->from3.offset = widthptr;
+	p->to.type = TYPE_ADDR;
+	p->to.name = NAME_EXTERN;
+	p->to.sym = linksym(x);
+	p->to.offset = xoff;
+	off += widthptr;
+
+	return off;
+}
diff --git a/src/cmd/gc/pgen.c b/src/cmd/gc/pgen.c
index ec1628651e..3df78e7f9a 100644
--- a/src/cmd/gc/pgen.c
+++ b/src/cmd/gc/pgen.c
@@ -30,7 +30,7 @@ makefuncdatasym(char *namefmt, int64 funcdatakind)
 	pnod = newname(sym);
 	pnod->class = PEXTERN;
 	nodconst(&nod, types[TINT32], funcdatakind);
-	arch.gins(AFUNCDATA, &nod, pnod);
+	thearch.gins(AFUNCDATA, &nod, pnod);
 	return sym;
 }
 
@@ -39,7 +39,7 @@ makefuncdatasym(char *namefmt, int64 funcdatakind)
 // where a complete initialization (definition) of a variable begins.
 // Since the liveness analysis can see initialization of single-word
 // variables quite easy, gvardef is usually only called for multi-word
-// or 'fat' variables, those satisfying arch.isfat(n->type).
+// or 'fat' variables, those satisfying isfat(n->type).
 // However, gvardef is also called when a non-fat variable is initialized
 // via a block move; the only time this happens is when you have
 //	return f()
@@ -103,7 +103,7 @@ gvardefx(Node *n, int as)
 	case PAUTO:
 	case PPARAM:
 	case PPARAMOUT:
-		arch.gins(as, N, n);
+		thearch.gins(as, N, n);
 	}
 }
 
@@ -161,6 +161,7 @@ compile(Node *fn)
 	Iter save;
 	vlong oldstksize;
 	NodeList *l;
+	Node *nam;
 	Sym *gcargs;
 	Sym *gclocals;
 
@@ -223,13 +224,16 @@ compile(Node *fn)
 	continpc = P;
 	breakpc = P;
 
-	pl = arch.newplist();
+	pl = newplist();
 	pl->name = linksym(curfn->nname->sym);
 
 	setlineno(curfn);
 
 	nodconst(&nod1, types[TINT32], 0);
-	ptxt = arch.gins(ATEXT, isblank(curfn->nname) ? N : curfn->nname, &nod1);
+	nam = curfn->nname;
+	if(isblank(nam))
+		nam = N;
+	ptxt = thearch.gins(ATEXT, nam, &nod1);
 	if(fn->dupok)
 		ptxt->from3.offset |= DUPOK;
 	if(fn->wrapper)
@@ -247,15 +251,15 @@ compile(Node *fn)
 			ptxt->from3.offset |= WRAPPER;
 	}	
 	
-	arch.afunclit(&ptxt->from, curfn->nname);
+	afunclit(&ptxt->from, curfn->nname);
 
-	arch.ginit();
+	thearch.ginit();
 
 	gcargs = makefuncdatasym("gcargs·%d", FUNCDATA_ArgsPointerMaps);
 	gclocals = makefuncdatasym("gclocals·%d", FUNCDATA_LocalsPointerMaps);
 
 	for(t=curfn->paramfld; t; t=t->down)
-		arch.gtrack(tracksym(t->type));
+		gtrack(tracksym(t->type));
 
 	for(l=fn->dcl; l; l=l->next) {
 		n = l->n;
@@ -266,7 +270,7 @@ compile(Node *fn)
 		case PPARAM:
 		case PPARAMOUT:
 			nodconst(&nod1, types[TUINTPTR], l->n->type->width);
-			p = arch.gins(ATYPE, l->n, &nod1);
+			p = thearch.gins(ATYPE, l->n, &nod1);
 			p->from.gotype = linksym(ngotype(l->n));
 			break;
 		}
@@ -274,7 +278,7 @@ compile(Node *fn)
 
 	genlist(curfn->enter);
 	genlist(curfn->nbody);
-	arch.gclean();
+	thearch.gclean();
 	checklabels();
 	if(nerrors != 0)
 		goto ret;
@@ -282,18 +286,18 @@ compile(Node *fn)
 		lineno = curfn->endlineno;
 
 	if(curfn->type->outtuple != 0)
-		arch.ginscall(throwreturn, 0);
+		thearch.ginscall(throwreturn, 0);
 
-	arch.ginit();
+	thearch.ginit();
 	// TODO: Determine when the final cgen_ret can be omitted. Perhaps always?
-	arch.cgen_ret(nil);
+	thearch.cgen_ret(nil);
 	if(hasdefer) {
 		// deferreturn pretends to have one uintptr argument.
 		// Reserve space for it so stack scanner is happy.
 		if(maxarg < widthptr)
 			maxarg = widthptr;
 	}
-	arch.gclean();
+	thearch.gclean();
 	if(nerrors != 0)
 		goto ret;
 
@@ -302,10 +306,10 @@ compile(Node *fn)
 
 	fixjmp(ptxt);
 	if(!debug['N'] || debug['R'] || debug['P']) {
-		arch.regopt(ptxt);
+		regopt(ptxt);
 		nilopt(ptxt);
 	}
-	arch.expandchecks(ptxt);
+	thearch.expandchecks(ptxt);
 
 	oldstksize = stksize;
 	allocauto(ptxt);
@@ -325,9 +329,9 @@ compile(Node *fn)
 	gcsymdup(gcargs);
 	gcsymdup(gclocals);
 
-	arch.defframe(ptxt);
+	thearch.defframe(ptxt);
 
-	if(0)
+	if(debug['f'])
 		frame(0);
 
 	// Remove leftover instrumentation from the instruction stream.
@@ -369,7 +373,7 @@ emitptrargsmap(void)
 		for(j = 0; j < bv->n; j += 32)
 			off = duint32(sym, off, bv->b[j/32]);
 	}
-	arch.ggloblsym(sym, off, RODATA);
+	ggloblsym(sym, off, RODATA);
 	free(bv);
 }
 
@@ -438,7 +442,7 @@ allocauto(Prog* ptxt)
 		if (ll->n->class == PAUTO)
 			ll->n->used = 0;
 
-	arch.markautoused(ptxt);
+	markautoused(ptxt);
 
 	listsort(&curfn->dcl, cmpstackvar);
 
@@ -448,7 +452,7 @@ allocauto(Prog* ptxt)
 	if (n->class == PAUTO && n->op == ONAME && !n->used) {
 		// No locals used at all
 		curfn->dcl = nil;
-		arch.fixautoused(ptxt);
+		fixautoused(ptxt);
 		return;
 	}
 
@@ -469,13 +473,13 @@ allocauto(Prog* ptxt)
 
 		dowidth(n->type);
 		w = n->type->width;
-		if(w >= arch.MAXWIDTH || w < 0)
+		if(w >= thearch.MAXWIDTH || w < 0)
 			fatal("bad width");
 		stksize += w;
 		stksize = rnd(stksize, n->type->align);
 		if(haspointers(n->type))
 			stkptrsize = stksize;
-		if(arch.thechar == '5' || arch.thechar == '9')
+		if(thearch.thechar == '5' || thearch.thechar == '9')
 			stksize = rnd(stksize, widthptr);
 		if(stksize >= (1ULL<<31)) {
 			setlineno(curfn);
@@ -486,7 +490,7 @@ allocauto(Prog* ptxt)
 	stksize = rnd(stksize, widthreg);
 	stkptrsize = rnd(stkptrsize, widthreg);
 
-	arch.fixautoused(ptxt);
+	fixautoused(ptxt);
 
 	// The debug information needs accurate offsets on the symbols.
 	for(ll = curfn->dcl; ll != nil; ll=ll->next) {
@@ -532,12 +536,12 @@ cgen_checknil(Node *n)
 		dump("checknil", n);
 		fatal("bad checknil");
 	}
-	if(((arch.thechar == '5' || arch.thechar == '9') && n->op != OREGISTER) || !n->addable || n->op == OLITERAL) {
-		arch.regalloc(&reg, types[tptr], n);
-		arch.cgen(n, &reg);
-		arch.gins(ACHECKNIL, &reg, N);
-		arch.regfree(&reg);
+	if(((thearch.thechar == '5' || thearch.thechar == '9') && n->op != OREGISTER) || !n->addable || n->op == OLITERAL) {
+		thearch.regalloc(&reg, types[tptr], n);
+		thearch.cgen(n, &reg);
+		thearch.gins(ACHECKNIL, &reg, N);
+		thearch.regfree(&reg);
 		return;
 	}
-	arch.gins(ACHECKNIL, n, N);
+	thearch.gins(ACHECKNIL, n, N);
 }
diff --git a/src/cmd/gc/plive.c b/src/cmd/gc/plive.c
index 0461085b5b..c0d1e57932 100644
--- a/src/cmd/gc/plive.c
+++ b/src/cmd/gc/plive.c
@@ -677,7 +677,7 @@ progeffects(Prog *prog, Array *vars, Bvec *uevar, Bvec *varkill, Bvec *avarinit)
 	bvresetall(varkill);
 	bvresetall(avarinit);
 
-	arch.proginfo(&info, prog);
+	thearch.proginfo(&info, prog);
 	if(prog->as == ARET) {
 		// Return instructions implicitly read all the arguments.  For
 		// the sake of correctness, out arguments must be read.  For the
@@ -701,7 +701,7 @@ progeffects(Prog *prog, Array *vars, Bvec *uevar, Bvec *varkill, Bvec *avarinit)
 				// If we added it to uevar too, we'd not see any kill
 				// and decide that the varible was live entry, which it is not.
 				// So only use uevar in the non-addrtaken case.
-				// The p->to.type == arch.D_NONE limits the bvset to
+				// The p->to.type == thearch.D_NONE limits the bvset to
 				// non-tail-call return instructions; see note above
 				// the for loop for details.
 				if(!node->addrtaken && prog->to.type == TYPE_NONE)
@@ -744,7 +744,7 @@ progeffects(Prog *prog, Array *vars, Bvec *uevar, Bvec *varkill, Bvec *avarinit)
 					if(info.flags & (LeftRead | LeftAddr))
 						bvset(uevar, pos);
 					if(info.flags & LeftWrite)
-						if(from->node != nil && !arch.isfat(((Node*)(from->node))->type))
+						if(from->node != nil && !isfat(((Node*)(from->node))->type))
 							bvset(varkill, pos);
 				}
 			}
@@ -780,7 +780,7 @@ Next:
 					if((info.flags & RightRead) || (info.flags & (RightAddr|RightWrite)) == RightAddr)
 						bvset(uevar, pos);
 					if(info.flags & RightWrite)
-						if(to->node != nil && (!arch.isfat(((Node*)(to->node))->type) || prog->as == AVARDEF))
+						if(to->node != nil && (!isfat(((Node*)(to->node))->type) || prog->as == AVARDEF))
 							bvset(varkill, pos);
 				}
 			}
@@ -1218,7 +1218,7 @@ unlinkedprog(int as)
 	Prog *p;
 
 	p = mal(sizeof(*p));
-	arch.clearp(p);
+	clearp(p);
 	p->as = as;
 	return p;
 }
@@ -1235,8 +1235,8 @@ newpcdataprog(Prog *prog, int32 index)
 	nodconst(&to, types[TINT32], index);
 	pcdata = unlinkedprog(APCDATA);
 	pcdata->lineno = prog->lineno;
-	arch.naddr(&from, &pcdata->from, 0);
-	arch.naddr(&to, &pcdata->to, 0);
+	naddr(&from, &pcdata->from, 0);
+	naddr(&to, &pcdata->to, 0);
 	return pcdata;
 }
 
@@ -1909,15 +1909,15 @@ static void
 twobitwritesymbol(Array *arr, Sym *sym)
 {
 	Bvec *bv;
-	int off, i, j, len;
+	int off, i, j, n;
 	uint32 word;
 
-	len = arraylength(arr);
+	n = arraylength(arr);
 	off = 0;
 	off += 4; // number of bitmaps, to fill in later
 	bv = *(Bvec**)arrayget(arr, 0);
 	off = duint32(sym, off, bv->n); // number of bits in each bitmap
-	for(i = 0; i < len; i++) {
+	for(i = 0; i < n; i++) {
 		// bitmap words
 		bv = *(Bvec**)arrayget(arr, i);
 		if(bv == nil)
@@ -1932,7 +1932,7 @@ twobitwritesymbol(Array *arr, Sym *sym)
 		}
 	}
 	duint32(sym, 0, i); // number of bitmaps
-	arch.ggloblsym(sym, off, RODATA);
+	ggloblsym(sym, off, RODATA);
 }
 
 static void
diff --git a/src/cmd/gc/popt.c b/src/cmd/gc/popt.c
index f71702431a..b02d58e663 100644
--- a/src/cmd/gc/popt.c
+++ b/src/cmd/gc/popt.c
@@ -35,6 +35,7 @@
 #include	<u.h>
 #include	<libc.h>
 #include	"go.h"
+#include	"popt.h"
 
 // p is a call instruction. Does the call fail to return?
 int
@@ -202,7 +203,7 @@ fixjmp(Prog *firstp)
 // Control flow analysis. The Flow structures hold predecessor and successor
 // information as well as basic loop analysis.
 //
-//	graph = flowstart(firstp, sizeof(Flow));
+//	graph = flowstart(firstp, 0);
 //	... use flow graph ...
 //	flowend(graph); // free graph
 //
@@ -214,25 +215,26 @@ fixjmp(Prog *firstp)
 // f->p1 and this list:
 //
 //	for(f2 = f->p2; f2 != nil; f2 = f2->p2link)
-//	
-// Often the Flow struct is embedded as the first field inside a larger struct S.
-// In that case casts are needed to convert Flow* to S* in many places but the
-// idea is the same. Pass sizeof(S) instead of sizeof(Flow) to flowstart.
+//
+// The size argument to flowstart specifies an amount of zeroed memory
+// to allocate in every f->data field, for use by the client.
+// If size == 0, f->data will be nil.
 
 Graph*
 flowstart(Prog *firstp, int size)
 {
-	int nf;
+	int id, nf;
 	Flow *f, *f1, *start, *last;
 	Graph *graph;
 	Prog *p;
 	ProgInfo info;
+	char *data;
 
 	// Count and mark instructions to annotate.
 	nf = 0;
 	for(p = firstp; p != P; p = p->link) {
 		p->opt = nil; // should be already, but just in case
-		arch.proginfo(&info, p);
+		thearch.proginfo(&info, p);
 		if(info.flags & Skip)
 			continue;
 		p->opt = (void*)1;
@@ -248,12 +250,16 @@ flowstart(Prog *firstp, int size)
 	}
 
 	// Allocate annotations and assign to instructions.
-	graph = calloc(sizeof *graph + size*nf, 1);
+	graph = calloc(sizeof *graph + sizeof(Flow)*nf + size*nf, 1);
 	if(graph == nil)
 		fatal("out of memory");
 	start = (Flow*)(graph+1);
 	last = nil;
 	f = start;
+	data = (char*)(f+nf);
+	if(size == 0)
+		data = nil;
+	id = 0;
 	for(p = firstp; p != P; p = p->link) {
 		if(p->opt == nil)
 			continue;
@@ -262,14 +268,17 @@ flowstart(Prog *firstp, int size)
 		if(last)
 			last->link = f;
 		last = f;
-		
-		f = (Flow*)((uchar*)f + size);
+		f->data = data;
+		f->id = id;
+		f++;
+		id++;
+		data += size;
 	}
 
 	// Fill in pred/succ information.
 	for(f = start; f != nil; f = f->link) {
 		p = f->prog;
-		arch.proginfo(&info, p);
+		thearch.proginfo(&info, p);
 		if(!(info.flags & Break)) {
 			f1 = f->link;
 			f->s1 = f1;
@@ -383,10 +392,6 @@ loophead(int32 *idom, Flow *r)
 	return 0;
 }
 
-enum {
-	LOOP = 3,
-};
-
 static void
 loopmark(Flow **rpo2r, int32 head, Flow *r)
 {
@@ -498,13 +503,12 @@ uniqs(Flow *r)
 // ACM TOPLAS 1999.
 
 typedef struct TempVar TempVar;
-typedef struct TempFlow TempFlow;
 
 struct TempVar
 {
 	Node *node;
-	TempFlow *def; // definition of temp var
-	TempFlow *use; // use list, chained through TempFlow.uselink
+	Flow *def; // definition of temp var
+	Flow *use; // use list, chained through Flow.data
 	TempVar *freelink; // next free temp in Type.opt list
 	TempVar *merge; // merge var with this one
 	vlong start; // smallest Prog.pc in live range
@@ -513,12 +517,6 @@ struct TempVar
 	uchar removed; // removed from program
 };
 
-struct TempFlow
-{
-	Flow	f;
-	TempFlow *uselink;
-};
-
 static int
 startcmp(const void *va, const void *vb)
 {
@@ -531,6 +529,14 @@ startcmp(const void *va, const void *vb)
 		return -1;
 	if(a->start > b->start)
 		return +1;
+	// Order what's left by id or symbol name,
+	// just so that sort is forced into a specific ordering,
+	// so that the result of the sort does not depend on
+	// the sort implementation.
+	if(a->def != b->def)
+		return a->def->id - b->def->id;
+	if(a->node != b->node)
+		return strcmp(a->node->sym->name, b->node->sym->name);
 	return 0;
 }
 
@@ -541,15 +547,15 @@ canmerge(Node *n)
 	return n->class == PAUTO && strncmp(n->sym->name, "autotmp", 7) == 0;
 }
 
-static void mergewalk(TempVar*, TempFlow*, uint32);
-static void varkillwalk(TempVar*, TempFlow*, uint32);
+static void mergewalk(TempVar*, Flow*, uint32);
+static void varkillwalk(TempVar*, Flow*, uint32);
 
 void
 mergetemp(Prog *firstp)
 {
 	int i, j, nvar, ninuse, nfree, nkill;
 	TempVar *var, *v, *v1, **bystart, **inuse;
-	TempFlow *r;
+	Flow *f;
 	NodeList *l, **lp;
 	Node *n;
 	Prog *p, *p1;
@@ -558,9 +564,9 @@ mergetemp(Prog *firstp)
 	int32 gen;
 	Graph *g;
 
-	enum { Debug = 0 };
+	enum { debugmerge = 1 };
 
-	g = flowstart(firstp, sizeof(TempFlow));
+	g = flowstart(firstp, 0);
 	if(g == nil)
 		return;
 	
@@ -585,9 +591,9 @@ mergetemp(Prog *firstp)
 	// We assume that the earliest reference to a temporary is its definition.
 	// This is not true of variables in general but our temporaries are all
 	// single-use (that's why we have so many!).
-	for(r = (TempFlow*)g->start; r != nil; r = (TempFlow*)r->f.link) {
-		p = r->f.prog;
-		arch.proginfo(&info, p);
+	for(f = g->start; f != nil; f = f->link) {
+		p = f->prog;
+		thearch.proginfo(&info, p);
 
 		if(p->from.node != N && ((Node*)(p->from.node))->opt && p->to.node != N && ((Node*)(p->to.node))->opt)
 			fatal("double node %P", p);
@@ -598,32 +604,33 @@ mergetemp(Prog *firstp)
 			v = n->opt;
 		if(v != nil) {
 		   	if(v->def == nil)
-		   		v->def = r;
-			r->uselink = v->use;
-			v->use = r;
+		   		v->def = f;
+		   	f->data = v->use;
+			v->use = f;
 			if(n == p->from.node && (info.flags & LeftAddr))
 				v->addr = 1;
 		}
 	}
 	
-	if(Debug > 1)
-		arch.dumpit("before", g->start, 0);
+	if(debugmerge > 1 && debug['v'])
+		dumpit("before", g->start, 0);
 	
 	nkill = 0;
 
 	// Special case.
-	for(v = var; v < var+nvar; v++) {
+	for(i = 0; i < nvar; i++) {
+		v = &var[i];
 		if(v->addr)
 			continue;
 		// Used in only one instruction, which had better be a write.
-		if((r = v->use) != nil && r->uselink == nil) {
-			p = r->f.prog;
-			arch.proginfo(&info, p);
+		if((f = v->use) != nil && (Flow*)f->data == nil) {
+			p = f->prog;
+			thearch.proginfo(&info, p);
 			if(p->to.node == v->node && (info.flags & RightWrite) && !(info.flags & RightRead)) {
 				p->as = ANOP;
 				p->to = zprog.to;
 				v->removed = 1;
-				if(Debug)
+				if(debugmerge > 0 && debug['v'])
 					print("drop write-only %S\n", v->node->sym);
 			} else
 				fatal("temp used and not set: %P", p);
@@ -633,11 +640,11 @@ mergetemp(Prog *firstp)
 		
 		// Written in one instruction, read in the next, otherwise unused,
 		// no jumps to the next instruction. Happens mainly in 386 compiler.
-		if((r = v->use) != nil && r->f.link == &r->uselink->f && r->uselink->uselink == nil && uniqp(r->f.link) == &r->f) {
-			p = r->f.prog;
-			arch.proginfo(&info, p);
-			p1 = r->f.link->prog;
-			arch.proginfo(&info1, p1);
+		if((f = v->use) != nil && f->link == (Flow*)f->data && (Flow*)((Flow*)f->data)->data == nil && uniqp(f->link) == f) {
+			p = f->prog;
+			thearch.proginfo(&info, p);
+			p1 = f->link->prog;
+			thearch.proginfo(&info1, p1);
 			enum {
 				SizeAny = SizeB | SizeW | SizeL | SizeQ | SizeF | SizeD,
 			};
@@ -645,9 +652,9 @@ mergetemp(Prog *firstp)
 			   !((info.flags|info1.flags) & (LeftAddr|RightAddr)) &&
 			   (info.flags & SizeAny) == (info1.flags & SizeAny)) {
 				p1->from = p->from;
-				arch.excise(&r->f);
+				thearch.excise(f);
 				v->removed = 1;
-				if(Debug)
+				if(debugmerge > 0 && debug['v'])
 					print("drop immediate-use %S\n", v->node->sym);
 			}
 			nkill++;
@@ -657,16 +664,17 @@ mergetemp(Prog *firstp)
 
 	// Traverse live range of each variable to set start, end.
 	// Each flood uses a new value of gen so that we don't have
-	// to clear all the r->f.active words after each variable.
+	// to clear all the r->active words after each variable.
 	gen = 0;
-	for(v = var; v < var+nvar; v++) {
+	for(i = 0; i < nvar; i++) {
+		v = &var[i];
 		gen++;
-		for(r = v->use; r != nil; r = r->uselink)
-			mergewalk(v, r, gen);
+		for(f = v->use; f != nil; f = (Flow*)f->data)
+			mergewalk(v, f, gen);
 		if(v->addr) {
 			gen++;
-			for(r = v->use; r != nil; r = r->uselink)
-				varkillwalk(v, r, gen);
+			for(f = v->use; f != nil; f = (Flow*)f->data)
+				varkillwalk(v, f, gen);
 		}
 	}
 
@@ -687,6 +695,9 @@ mergetemp(Prog *firstp)
 	nfree = nvar;
 	for(i=0; i<nvar; i++) {
 		v = bystart[i];
+		if(debugmerge > 0 && debug['v'])
+			print("consider %#N: removed=%d\n", v->node, v->removed);
+			
 		if(v->removed)
 			continue;
 
@@ -696,10 +707,14 @@ mergetemp(Prog *firstp)
 			inuse[--nfree] = v1;
 		}
 
+		if(debugmerge > 0 && debug['v'])
+			print("consider %#N: removed=%d nfree=%d nvar=%d\n", v->node, v->removed, nfree, nvar);
 		// Find old temp to reuse if possible.
 		t = v->node->type;
 		for(j=nfree; j<nvar; j++) {
 			v1 = inuse[j];
+			if(debugmerge > 0 && debug['v'])
+				print("consider %#N: maybe %#N: type=%T,%T addrtaken=%d,%d\n", v->node, v1->node, t, v1->node->type, v->node->addrtaken, v1->node->addrtaken);
 			// Require the types to match but also require the addrtaken bits to match.
 			// If a variable's address is taken, that disables registerization for the individual
 			// words of the variable (for example, the base,len,cap of a slice).
@@ -725,9 +740,10 @@ mergetemp(Prog *firstp)
 		inuse[j] = v;
 	}
 
-	if(Debug) {
+	if(debugmerge > 0 && debug['v']) {
 		print("%S [%d - %d]\n", curfn->nname->sym, nvar, nkill);
-		for(v=var; v<var+nvar; v++) {
+		for(i = 0; i < nvar; i++) {
+			v = &var[i];
 			print("var %#N %T %lld-%lld", v->node, v->node->type, v->start, v->end);
 			if(v->addr)
 				print(" addr=1");
@@ -735,18 +751,18 @@ mergetemp(Prog *firstp)
 				print(" dead=1");
 			if(v->merge)
 				print(" merge %#N", v->merge->node);
-			if(v->start == v->end)
-				print(" %P", v->def->f.prog);
+			if(v->start == v->end && v->def != nil)
+				print(" %P", v->def->prog);
 			print("\n");
 		}
 	
-		if(Debug > 1)
-			arch.dumpit("after", g->start, 0);
+		if(debugmerge > 1 && debug['v'])
+			dumpit("after", g->start, 0);
 	}
 
 	// Update node references to use merged temporaries.
-	for(r = (TempFlow*)g->start; r != nil; r = (TempFlow*)r->f.link) {
-		p = r->f.prog;
+	for(f = g->start; f != nil; f = f->link) {
+		p = f->prog;
 		if((n = p->from.node) != N && (v = n->opt) != nil && v->merge != nil)
 			p->from.node = v->merge->node;
 		if((n = p->to.node) != N && (v = n->opt) != nil && v->merge != nil)
@@ -766,8 +782,9 @@ mergetemp(Prog *firstp)
 	}
 
 	// Clear aux structures.
-	for(v=var; v<var+nvar; v++)
-		v->node->opt = nil;
+	for(i = 0; i < nvar; i++)
+		var[i].node->opt = nil;
+
 	free(var);
 	free(bystart);
 	free(inuse);
@@ -775,40 +792,40 @@ mergetemp(Prog *firstp)
 }
 
 static void
-mergewalk(TempVar *v, TempFlow *r0, uint32 gen)
+mergewalk(TempVar *v, Flow *f0, uint32 gen)
 {
 	Prog *p;
-	TempFlow *r1, *r, *r2;
+	Flow *f1, *f, *f2;
 	
-	for(r1 = r0; r1 != nil; r1 = (TempFlow*)r1->f.p1) {
-		if(r1->f.active == gen)
+	for(f1 = f0; f1 != nil; f1 = f1->p1) {
+		if(f1->active == gen)
 			break;
-		r1->f.active = gen;
-		p = r1->f.prog;
+		f1->active = gen;
+		p = f1->prog;
 		if(v->end < p->pc)
 			v->end = p->pc;
-		if(r1 == v->def) {
+		if(f1 == v->def) {
 			v->start = p->pc;
 			break;
 		}
 	}
 	
-	for(r = r0; r != r1; r = (TempFlow*)r->f.p1)
-		for(r2 = (TempFlow*)r->f.p2; r2 != nil; r2 = (TempFlow*)r2->f.p2link)
-			mergewalk(v, r2, gen);
+	for(f = f0; f != f1; f = f->p1)
+		for(f2 = f->p2; f2 != nil; f2 = f2->p2link)
+			mergewalk(v, f2, gen);
 }
 
 static void
-varkillwalk(TempVar *v, TempFlow *r0, uint32 gen)
+varkillwalk(TempVar *v, Flow *f0, uint32 gen)
 {
 	Prog *p;
-	TempFlow *r1, *r;
+	Flow *f1, *f;
 	
-	for(r1 = r0; r1 != nil; r1 = (TempFlow*)r1->f.s1) {
-		if(r1->f.active == gen)
+	for(f1 = f0; f1 != nil; f1 = f1->s1) {
+		if(f1->active == gen)
 			break;
-		r1->f.active = gen;
-		p = r1->f.prog;
+		f1->active = gen;
+		p = f1->prog;
 		if(v->end < p->pc)
 			v->end = p->pc;
 		if(v->start > p->pc)
@@ -817,8 +834,8 @@ varkillwalk(TempVar *v, TempFlow *r0, uint32 gen)
 			break;
 	}
 	
-	for(r = r0; r != r1; r = (TempFlow*)r->f.s1)
-		varkillwalk(v, (TempFlow*)r->f.s2, gen);
+	for(f = f0; f != f1; f = f->s1)
+		varkillwalk(v, f->s2, gen);
 }
 
 // Eliminate redundant nil pointer checks.
@@ -836,62 +853,58 @@ varkillwalk(TempVar *v, TempFlow *r0, uint32 gen)
 // each load.
 	
 typedef struct NilVar NilVar;
-typedef struct NilFlow NilFlow;
 
-struct NilFlow {
-	Flow f;
-	int kill;
-};
+static void nilwalkback(Flow *rcheck);
+static void nilwalkfwd(Flow *rcheck);
 
-static void nilwalkback(NilFlow *rcheck);
-static void nilwalkfwd(NilFlow *rcheck);
+static int killed; // f->data is either nil or &killed
 
 void
 nilopt(Prog *firstp)
 {
-	NilFlow *r;
+	Flow *f;
 	Prog *p;
 	Graph *g;
 	int ncheck, nkill;
 
-	g = flowstart(firstp, sizeof(NilFlow));
+	g = flowstart(firstp, 0);
 	if(g == nil)
 		return;
 
 	if(debug_checknil > 1 /* || strcmp(curfn->nname->sym->name, "f1") == 0 */)
-		arch.dumpit("nilopt", g->start, 0);
+		dumpit("nilopt", g->start, 0);
 
 	ncheck = 0;
 	nkill = 0;
-	for(r = (NilFlow*)g->start; r != nil; r = (NilFlow*)r->f.link) {
-		p = r->f.prog;
-		if(p->as != ACHECKNIL || !arch.regtyp(&p->from))
+	for(f = g->start; f != nil; f = f->link) {
+		p = f->prog;
+		if(p->as != ACHECKNIL || !thearch.regtyp(&p->from))
 			continue;
 		ncheck++;
-		if(arch.stackaddr(&p->from)) {
+		if(thearch.stackaddr(&p->from)) {
 			if(debug_checknil && p->lineno > 1)
 				warnl(p->lineno, "removed nil check of SP address");
-			r->kill = 1;
+			f->data = &killed;
 			continue;
 		}
-		nilwalkfwd(r);
-		if(r->kill) {
+		nilwalkfwd(f);
+		if(f->data != nil) {
 			if(debug_checknil && p->lineno > 1)
 				warnl(p->lineno, "removed nil check before indirect");
 			continue;
 		}
-		nilwalkback(r);
-		if(r->kill) {
+		nilwalkback(f);
+		if(f->data != nil) {
 			if(debug_checknil && p->lineno > 1)
 				warnl(p->lineno, "removed repeated nil check");
 			continue;
 		}
 	}
 	
-	for(r = (NilFlow*)g->start; r != nil; r = (NilFlow*)r->f.link) {
-		if(r->kill) {
+	for(f = g->start; f != nil; f = f->link) {
+		if(f->data != nil) {
 			nkill++;
-			arch.excise(&r->f);
+			thearch.excise(f);
 		}
 	}
 
@@ -902,72 +915,72 @@ nilopt(Prog *firstp)
 }
 
 static void
-nilwalkback(NilFlow *rcheck)
+nilwalkback(Flow *fcheck)
 {
 	Prog *p;
 	ProgInfo info;
-	NilFlow *r;
+	Flow *f;
 	
-	for(r = rcheck; r != nil; r = (NilFlow*)uniqp(&r->f)) {
-		p = r->f.prog;
-		arch.proginfo(&info, p);
-		if((info.flags & RightWrite) && arch.sameaddr(&p->to, &rcheck->f.prog->from)) {
+	for(f = fcheck; f != nil; f = uniqp(f)) {
+		p = f->prog;
+		thearch.proginfo(&info, p);
+		if((info.flags & RightWrite) && thearch.sameaddr(&p->to, &fcheck->prog->from)) {
 			// Found initialization of value we're checking for nil.
 			// without first finding the check, so this one is unchecked.
 			return;
 		}
-		if(r != rcheck && p->as == ACHECKNIL && arch.sameaddr(&p->from, &rcheck->f.prog->from)) {
-			rcheck->kill = 1;
+		if(f != fcheck && p->as == ACHECKNIL && thearch.sameaddr(&p->from, &fcheck->prog->from)) {
+			fcheck->data = &killed;
 			return;
 		}
 	}
 
 	// Here is a more complex version that scans backward across branches.
-	// It assumes rcheck->kill = 1 has been set on entry, and its job is to find a reason
-	// to keep the check (setting rcheck->kill = 0).
+	// It assumes fcheck->kill = 1 has been set on entry, and its job is to find a reason
+	// to keep the check (setting fcheck->kill = 0).
 	// It doesn't handle copying of aggregates as well as I would like,
 	// nor variables with their address taken,
 	// and it's too subtle to turn on this late in Go 1.2. Perhaps for Go 1.3.
 	/*
-	for(r1 = r0; r1 != nil; r1 = (NilFlow*)r1->f.p1) {
-		if(r1->f.active == gen)
+	for(f1 = f0; f1 != nil; f1 = f1->p1) {
+		if(f1->active == gen)
 			break;
-		r1->f.active = gen;
-		p = r1->f.prog;
+		f1->active = gen;
+		p = f1->prog;
 		
 		// If same check, stop this loop but still check
 		// alternate predecessors up to this point.
-		if(r1 != rcheck && p->as == ACHECKNIL && arch.sameaddr(&p->from, &rcheck->f.prog->from))
+		if(f1 != fcheck && p->as == ACHECKNIL && thearch.sameaddr(&p->from, &fcheck->prog->from))
 			break;
 
-		arch.proginfo(&info, p);
-		if((info.flags & RightWrite) && arch.sameaddr(&p->to, &rcheck->f.prog->from)) {
+		thearch.proginfo(&info, p);
+		if((info.flags & RightWrite) && thearch.sameaddr(&p->to, &fcheck->prog->from)) {
 			// Found initialization of value we're checking for nil.
 			// without first finding the check, so this one is unchecked.
-			rcheck->kill = 0;
+			fcheck->kill = 0;
 			return;
 		}
 		
-		if(r1->f.p1 == nil && r1->f.p2 == nil) {
-			print("lost pred for %P\n", rcheck->f.prog);
-			for(r1=r0; r1!=nil; r1=(NilFlow*)r1->f.p1) {
-				arch.proginfo(&info, r1->f.prog);
-				print("\t%P %d %d %D %D\n", r1->f.prog, info.flags&RightWrite, arch.sameaddr(&r1->f.prog->to, &rcheck->f.prog->from), &r1->f.prog->to, &rcheck->f.prog->from);
+		if(f1->p1 == nil && f1->p2 == nil) {
+			print("lost pred for %P\n", fcheck->prog);
+			for(f1=f0; f1!=nil; f1=f1->p1) {
+				thearch.proginfo(&info, f1->prog);
+				print("\t%P %d %d %D %D\n", r1->prog, info.flags&RightWrite, thearch.sameaddr(&f1->prog->to, &fcheck->prog->from), &f1->prog->to, &fcheck->prog->from);
 			}
 			fatal("lost pred trail");
 		}
 	}
 
-	for(r = r0; r != r1; r = (NilFlow*)r->f.p1)
-		for(r2 = (NilFlow*)r->f.p2; r2 != nil; r2 = (NilFlow*)r2->f.p2link)
-			nilwalkback(rcheck, r2, gen);
+	for(f = f0; f != f1; f = f->p1)
+		for(f2 = f->p2; f2 != nil; f2 = f2->p2link)
+			nilwalkback(fcheck, f2, gen);
 	*/
 }
 
 static void
-nilwalkfwd(NilFlow *rcheck)
+nilwalkfwd(Flow *fcheck)
 {
-	NilFlow *r, *last;
+	Flow *f, *last;
 	Prog *p;
 	ProgInfo info;
 	
@@ -979,16 +992,16 @@ nilwalkfwd(NilFlow *rcheck)
 	//	_ = *x // should panic
 	//	for {} // no writes but infinite loop may be considered visible
 	last = nil;
-	for(r = (NilFlow*)uniqs(&rcheck->f); r != nil; r = (NilFlow*)uniqs(&r->f)) {
-		p = r->f.prog;
-		arch.proginfo(&info, p);
+	for(f = uniqs(fcheck); f != nil; f = uniqs(f)) {
+		p = f->prog;
+		thearch.proginfo(&info, p);
 		
-		if((info.flags & LeftRead) && arch.smallindir(&p->from, &rcheck->f.prog->from)) {
-			rcheck->kill = 1;
+		if((info.flags & LeftRead) && thearch.smallindir(&p->from, &fcheck->prog->from)) {
+			fcheck->data = &killed;
 			return;
 		}
-		if((info.flags & (RightRead|RightWrite)) && arch.smallindir(&p->to, &rcheck->f.prog->from)) {
-			rcheck->kill = 1;
+		if((info.flags & (RightRead|RightWrite)) && thearch.smallindir(&p->to, &fcheck->prog->from)) {
+			fcheck->data = &killed;
 			return;
 		}
 		
@@ -996,17 +1009,14 @@ nilwalkfwd(NilFlow *rcheck)
 		if(p->as == ACHECKNIL)
 			return;
 		// Stop if value is lost.
-		if((info.flags & RightWrite) && arch.sameaddr(&p->to, &rcheck->f.prog->from))
+		if((info.flags & RightWrite) && thearch.sameaddr(&p->to, &fcheck->prog->from))
 			return;
 		// Stop if memory write.
-		if((info.flags & RightWrite) && !arch.regtyp(&p->to))
+		if((info.flags & RightWrite) && !thearch.regtyp(&p->to))
 			return;
 		// Stop if we jump backward.
-		// This test is valid because all the NilFlow* are pointers into
-		// a single contiguous array. We will need to add an explicit
-		// numbering when the code is converted to Go.
-		if(last != nil && r <= last)
+		if(last != nil && f->id <= last->id)
 			return;
-		last = r;
+		last = f;
 	}
 }
diff --git a/src/cmd/6g/opt.h b/src/cmd/gc/popt.h
index 11befb6ad1..0a30e81f55 100644
--- a/src/cmd/6g/opt.h
+++ b/src/cmd/gc/popt.h
@@ -28,38 +28,28 @@
 // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 // THE SOFTWARE.
 
-
 #define	Z	N
 #define	Adr	Addr
 
-#define	D_HI	TYPE_NONE
-#define	D_LO	TYPE_NONE
-
 #define	BLOAD(r)	band(bnot(r->refbehind), r->refahead)
 #define	BSTORE(r)	band(bnot(r->calbehind), r->calahead)
 #define	LOAD(r)		(~r->refbehind.b[z] & r->refahead.b[z])
 #define	STORE(r)	(~r->calbehind.b[z] & r->calahead.b[z])
 
-#define	CLOAD	5
-#define	CREF	5
-#define	CINF	1000
-#define	LOOP	3
-
-typedef	struct	Reg	Reg;
-typedef	struct	Rgn	Rgn;
-
-/*c2go
-extern Node *Z;
 enum
 {
-	D_HI = TYPE_NONE,
-	D_LO = TYPE_NONE,
 	CLOAD = 5,
 	CREF = 5,
 	CINF = 1000,
 	LOOP = 3,
 };
 
+typedef	struct	Reg	Reg;
+typedef	struct	Rgn	Rgn;
+
+/*c2go
+extern Node *Z;
+
 uint32 BLOAD(Reg*);
 uint32 BSTORE(Reg*);
 uint64 LOAD(Reg*);
@@ -69,11 +59,8 @@ uint64 STORE(Reg*);
 // A Reg is a wrapper around a single Prog (one instruction) that holds
 // register optimization information while the optimizer runs.
 // r->prog is the instruction.
-// r->prog->opt points back to r.
 struct	Reg
 {
-	Flow	f;
-
 	Bits	set;  		// regopt variables written by this instruction.
 	Bits	use1; 		// regopt variables read by prog->from.
 	Bits	use2; 		// regopt variables read by prog->to.
@@ -91,7 +78,7 @@ struct	Reg
 	Bits	regdiff;
 	Bits	act;
 
-	int32	regu;		// register used bitmap
+	uint64	regu;		// register used bitmap
 };
 #define	R	((Reg*)0)
 /*c2go extern Reg *R; */
@@ -110,21 +97,18 @@ struct	Reg
 // cost.
 struct	Rgn
 {
-	Reg*	enter;
+	Flow*	enter;
 	short	cost;
 	short	varno;
 	short	regno;
 };
 
-EXTERN	int32	exregoffset;		// not set
-EXTERN	int32	exfregoffset;		// not set
 EXTERN	Reg	zreg;
 EXTERN	Rgn	region[NRGN];
 EXTERN	Rgn*	rgp;
 EXTERN	int	nregion;
 EXTERN	int	nvar;
-EXTERN	int32	regbits;
-EXTERN	int32	exregbits;
+EXTERN	uint64	regbits;
 EXTERN	Bits	externs;
 EXTERN	Bits	params;
 EXTERN	Bits	consts;
@@ -147,34 +131,19 @@ EXTERN	struct
 /*
  * reg.c
  */
-int	rcmp(const void*, const void*);
 void	regopt(Prog*);
-void	addmove(Reg*, int, int, int);
-Bits	mkvar(Reg*, Adr*);
-void	prop(Reg*, Bits, Bits);
-void	synch(Reg*, Bits);
-uint32	allreg(uint32, Rgn*);
-void	paint1(Reg*, int);
-uint32	paint2(Reg*, int, int);
-void	paint3(Reg*, int, uint32, int);
-void	addreg(Adr*, int);
 void	dumpone(Flow*, int);
 void	dumpit(char*, Flow*, int);
 
 /*
  * peep.c
- */
 void	peep(Prog*);
 void	excise(Flow*);
 int	copyu(Prog*, Adr*, Adr*);
-
-uint32	RtoB(int);
-uint32	FtoB(int);
-int	BtoR(uint32);
-int	BtoF(uint32);
+ */
 
 /*
  * prog.c
- */
 
 void proginfo(ProgInfo*, Prog*);
+ */
diff --git a/src/cmd/gc/racewalk.c b/src/cmd/gc/racewalk.c
index f3134dab23..3aa7e36386 100644
--- a/src/cmd/gc/racewalk.c
+++ b/src/cmd/gc/racewalk.c
@@ -24,7 +24,6 @@ static void racewalknode(Node **np, NodeList **init, int wr, int skip);
 static int callinstr(Node **n, NodeList **init, int wr, int skip);
 static Node* uintptraddr(Node *n);
 static void makeaddable(Node *n);
-static Node* basenod(Node *n);
 static void foreach(Node *n, void(*f)(Node*, void*), void *c);
 static void hascallspred(Node *n, void *c);
 static void appendinit(Node **np, NodeList *init);
@@ -155,12 +154,8 @@ racewalknode(Node **np, NodeList **init, int wr, int skip)
 	default:
 		fatal("racewalk: unknown node type %O", n->op);
 
-	case OASOP:
 	case OAS:
-	case OAS2:
-	case OAS2RECV:
 	case OAS2FUNC:
-	case OAS2MAPR:
 		racewalknode(&n->left, init, 1, 0);
 		racewalknode(&n->right, init, 0, 0);
 		goto ret;
@@ -210,6 +205,7 @@ racewalknode(Node **np, NodeList **init, int wr, int skip)
 	case OCALLFUNC:
 		// Instrument dst argument of runtime.writebarrier* calls
 		// as we do not instrument runtime code.
+		// typedslicecopy is instrumented in runtime.
 		if(n->left->sym != S && n->left->sym->pkg == runtimepkg && 
 		(strncmp(n->left->sym->name, "writebarrier", 12) == 0 || strcmp(n->left->sym->name, "typedmemmove") == 0)) {
 			// Find the dst argument.
@@ -350,7 +346,7 @@ racewalknode(Node **np, NodeList **init, int wr, int skip)
 		goto ret;
 
 	case OEFACE:
-		racewalknode(&n->left, init, 0, 0);
+		// n->left is Type* which is not interesting.
 		racewalknode(&n->right, init, 0, 0);
 		goto ret;
 
@@ -393,6 +389,10 @@ racewalknode(Node **np, NodeList **init, int wr, int skip)
 	case OARRAYLIT: // lowered to assignments
 	case OMAPLIT:
 	case OSTRUCTLIT:
+	case OAS2:
+	case OAS2RECV:
+	case OAS2MAPR:
+	case OASOP:
 		yyerror("racewalk: %O must be lowered by now", n->op);
 		goto ret;
 
@@ -489,7 +489,7 @@ callinstr(Node **np, NodeList **init, int wr, int skip)
 	if(isartificial(n))
 		return 0;
 
-	b = basenod(n);
+	b = outervalue(n);
 	// it skips e.g. stores to ... parameter array
 	if(isartificial(b))
 		return 0;
@@ -499,7 +499,7 @@ callinstr(Node **np, NodeList **init, int wr, int skip)
 	// that has got a pointer inside. Whether it points to
 	// the heap or not is impossible to know at compile time
 	if((class&PHEAP) || class == PPARAMREF || class == PEXTERN
-		|| b->op == OINDEX || b->op == ODOTPTR || b->op == OIND || b->op == OXDOT) {
+		|| b->op == OINDEX || b->op == ODOTPTR || b->op == OIND) {
 		hascalls = 0;
 		foreach(n, hascallspred, &hascalls);
 		if(hascalls) {
@@ -568,25 +568,6 @@ uintptraddr(Node *n)
 	return r;
 }
 
-// basenod returns the simplest child node of n pointing to the same
-// memory area.
-static Node*
-basenod(Node *n)
-{
-	for(;;) {
-		if(n->op == ODOT || n->op == OXDOT || n->op == OCONVNOP || n->op == OCONV || n->op == OPAREN) {
-			n = n->left;
-			continue;
-		}
-		if(n->op == OINDEX && isfixedarray(n->type)) {
-			n = n->left;
-			continue;
-		}
-		break;
-	}
-	return n;
-}
-
 static Node*
 detachexpr(Node *n, NodeList **init)
 {
diff --git a/src/cmd/gc/range.c b/src/cmd/gc/range.c
index ff05820b58..ff9de6c349 100644
--- a/src/cmd/gc/range.c
+++ b/src/cmd/gc/range.c
@@ -13,6 +13,7 @@
 void
 typecheckrange(Node *n)
 {
+	int toomany;
 	char *why;
 	Type *t, *t1, *t2;
 	Node *v1, *v2;
@@ -41,6 +42,7 @@ typecheckrange(Node *n)
 		t = t->type;
 	n->type = t;
 
+	toomany = 0;
 	switch(t->etype) {
 	default:
 		yyerror("cannot range over %lN", n->right);
@@ -64,7 +66,7 @@ typecheckrange(Node *n)
 		t1 = t->type;
 		t2 = nil;
 		if(count(n->list) == 2)
-			goto toomany;
+			toomany = 1;
 		break;
 
 	case TSTRING:
@@ -73,10 +75,8 @@ typecheckrange(Node *n)
 		break;
 	}
 
-	if(count(n->list) > 2) {
-	toomany:
+	if(count(n->list) > 2 || toomany)
 		yyerror("too many variables in range");
-	}
 
 	v1 = N;
 	if(n->list)
diff --git a/src/cmd/gc/reflect.c b/src/cmd/gc/reflect.c
index 61a63c0528..14a1f13e33 100644
--- a/src/cmd/gc/reflect.c
+++ b/src/cmd/gc/reflect.c
@@ -117,16 +117,29 @@ enum {
 };
 
 static Type*
+makefield(char *name, Type *t)
+{
+	Type *f;
+
+	f = typ(TFIELD);
+	f->type = t;
+	f->sym = mal(sizeof(Sym));
+	f->sym->name = name;
+	return f;
+}
+
+Type*
 mapbucket(Type *t)
 {
 	Type *keytype, *valtype;
-	Type *bucket;
-	Type *overflowfield, *keysfield, *valuesfield;
-	int32 offset;
+	Type *bucket, *arr;
+	Type *field[4];
+	int32 n;
 
 	if(t->bucket != T)
 		return t->bucket;
 
+	bucket = typ(TSTRUCT);
 	keytype = t->down;
 	valtype = t->type;
 	dowidth(keytype);
@@ -136,119 +149,74 @@ mapbucket(Type *t)
 	if(valtype->width > MAXVALSIZE)
 		valtype = ptrto(valtype);
 
-	bucket = typ(TSTRUCT);
-	bucket->noalg = 1;
-
 	// The first field is: uint8 topbits[BUCKETSIZE].
-	// We don't need to encode it as GC doesn't care about it.
-	offset = BUCKETSIZE * 1;
-
-	keysfield = typ(TFIELD);
-	keysfield->type = typ(TARRAY);
-	keysfield->type->type = keytype;
-	keysfield->type->bound = BUCKETSIZE;
-	keysfield->type->width = BUCKETSIZE * keytype->width;
-	keysfield->width = offset;
-	keysfield->sym = mal(sizeof(Sym));
-	keysfield->sym->name = "keys";
-	offset += BUCKETSIZE * keytype->width;
+	arr = typ(TARRAY);
+	arr->type = types[TUINT8];
+	arr->bound = BUCKETSIZE;
+	field[0] = makefield("topbits", arr);
+	arr = typ(TARRAY);
+	arr->type = keytype;
+	arr->bound = BUCKETSIZE;
+	field[1] = makefield("keys", arr);
+	arr = typ(TARRAY);
+	arr->type = valtype;
+	arr->bound = BUCKETSIZE;
+	field[2] = makefield("values", arr);
+	field[3] = makefield("overflow", ptrto(bucket));
 
-	valuesfield = typ(TFIELD);
-	valuesfield->type = typ(TARRAY);
-	valuesfield->type->type = valtype;
-	valuesfield->type->bound = BUCKETSIZE;
-	valuesfield->type->width = BUCKETSIZE * valtype->width;
-	valuesfield->width = offset;
-	valuesfield->sym = mal(sizeof(Sym));
-	valuesfield->sym->name = "values";
-	offset += BUCKETSIZE * valtype->width;
+	// link up fields
+	bucket->noalg = 1;
+	bucket->local = t->local;
+	bucket->type = field[0];
+	for(n = 0; n < nelem(field)-1; n++)
+		field[n]->down = field[n+1];
+	field[nelem(field)-1]->down = T;
+	dowidth(bucket);
 
-	overflowfield = typ(TFIELD);
-	overflowfield->type = ptrto(bucket);
-	overflowfield->width = offset;         // "width" is offset in structure
-	overflowfield->sym = mal(sizeof(Sym)); // not important but needs to be set to give this type a name
-	overflowfield->sym->name = "overflow";
-	offset += widthptr;
-	
 	// Pad to the native integer alignment.
-	// This is usually the same as widthptr; the exception (as usual) is nacl/amd64.
+	// This is usually the same as widthptr; the exception (as usual) is amd64p32.
 	if(widthreg > widthptr)
-		offset += widthreg - widthptr;
-
-	// link up fields
-	bucket->type = keysfield;
-	keysfield->down = valuesfield;
-	valuesfield->down = overflowfield;
-	overflowfield->down = T;
+		bucket->width += widthreg - widthptr;
 
 	// See comment on hmap.overflow in ../../runtime/hashmap.go.
 	if(!haspointers(t->type) && !haspointers(t->down))
 		bucket->haspointers = 1;  // no pointers
 
-	bucket->width = offset;
-	bucket->local = t->local;
 	t->bucket = bucket;
 	bucket->map = t;
 	return bucket;
 }
 
-// Builds a type respresenting a Hmap structure for
-// the given map type.  This type is not visible to users -
-// we include only enough information to generate a correct GC
-// program for it.
+// Builds a type representing a Hmap structure for the given map type.
 // Make sure this stays in sync with ../../runtime/hashmap.go!
-static Type*
+Type*
 hmap(Type *t)
 {
 	Type *h, *bucket;
-	Type *bucketsfield, *oldbucketsfield, *overflowfield;
-	int32 offset;
+	Type *field[8];
+	int32 n;
 
 	if(t->hmap != T)
 		return t->hmap;
 
 	bucket = mapbucket(t);
+	field[0] = makefield("count", types[TINT]);
+	field[1] = makefield("flags", types[TUINT8]);
+	field[2] = makefield("B", types[TUINT8]);
+	field[3] = makefield("hash0", types[TUINT32]);
+	field[4] = makefield("buckets", ptrto(bucket));
+	field[5] = makefield("oldbuckets", ptrto(bucket));
+	field[6] = makefield("nevacuate", types[TUINTPTR]);
+	field[7] = makefield("overflow", types[TUNSAFEPTR]);
+
 	h = typ(TSTRUCT);
 	h->noalg = 1;
-
-	offset = widthint; // count
-	offset += 1;       // flags
-	offset += 1;       // B
-	offset += 2;       // padding
-	offset += 4;       // hash0
-	offset = (offset + widthptr - 1) / widthptr * widthptr;
-	
-	bucketsfield = typ(TFIELD);
-	bucketsfield->type = ptrto(bucket);
-	bucketsfield->width = offset;
-	bucketsfield->sym = mal(sizeof(Sym));
-	bucketsfield->sym->name = "buckets";
-	offset += widthptr;
-
-	oldbucketsfield = typ(TFIELD);
-	oldbucketsfield->type = ptrto(bucket);
-	oldbucketsfield->width = offset;
-	oldbucketsfield->sym = mal(sizeof(Sym));
-	oldbucketsfield->sym->name = "oldbuckets";
-	offset += widthptr;
-
-	offset += widthptr; // nevacuate
-
-	overflowfield = typ(TFIELD);
-	overflowfield->type = types[TUNSAFEPTR];
-	overflowfield->width = offset;
-	overflowfield->sym = mal(sizeof(Sym));
-	overflowfield->sym->name = "overflow";
-	offset += widthptr;
-
-	// link up fields
-	h->type = bucketsfield;
-	bucketsfield->down = oldbucketsfield;
-	oldbucketsfield->down = overflowfield;
-	overflowfield->down = T;
-
-	h->width = offset;
 	h->local = t->local;
+	h->type = field[0];
+	for(n = 0; n < nelem(field)-1; n++)
+		field[n]->down = field[n+1];
+	field[nelem(field)-1]->down = T;
+	dowidth(h);
 	t->hmap = h;
 	h->map = t;
 	return h;
@@ -257,8 +225,8 @@ hmap(Type *t)
 Type*
 hiter(Type *t)
 {
-	int32 n, off;
-	Type *field[9];
+	int32 n;
+	Type *field[12];
 	Type *i;
 
 	if(t->hiter != T)
@@ -272,73 +240,37 @@ hiter(Type *t)
 	//    h *Hmap
 	//    buckets *Bucket
 	//    bptr *Bucket
-	//    overflow unsafe.Pointer
-	//    other [4]uintptr
+	//    overflow0 unsafe.Pointer
+	//    overflow1 unsafe.Pointer
+	//    startBucket uintptr
+	//    stuff uintptr
+	//    bucket uintptr
+	//    checkBucket uintptr
 	// }
 	// must match ../../runtime/hashmap.c:hash_iter.
-	field[0] = typ(TFIELD);
-	field[0]->type = ptrto(t->down);
-	field[0]->sym = mal(sizeof(Sym));
-	field[0]->sym->name = "key";
-	
-	field[1] = typ(TFIELD);
-	field[1]->type = ptrto(t->type);
-	field[1]->sym = mal(sizeof(Sym));
-	field[1]->sym->name = "val";
-	
-	field[2] = typ(TFIELD);
-	field[2]->type = ptrto(types[TUINT8]); // TODO: is there a Type type?
-	field[2]->sym = mal(sizeof(Sym));
-	field[2]->sym->name = "t";
-	
-	field[3] = typ(TFIELD);
-	field[3]->type = ptrto(hmap(t));
-	field[3]->sym = mal(sizeof(Sym));
-	field[3]->sym->name = "h";
-	
-	field[4] = typ(TFIELD);
-	field[4]->type = ptrto(mapbucket(t));
-	field[4]->sym = mal(sizeof(Sym));
-	field[4]->sym->name = "buckets";
-	
-	field[5] = typ(TFIELD);
-	field[5]->type = ptrto(mapbucket(t));
-	field[5]->sym = mal(sizeof(Sym));
-	field[5]->sym->name = "bptr";
-	
-	field[6] = typ(TFIELD);
-	field[6]->type = types[TUNSAFEPTR];
-	field[6]->sym = mal(sizeof(Sym));
-	field[6]->sym->name = "overflow0";
-
-	field[7] = typ(TFIELD);
-	field[7]->type = types[TUNSAFEPTR];
-	field[7]->sym = mal(sizeof(Sym));
-	field[7]->sym->name = "overflow1";
-
-	// all other non-pointer fields
-	field[8] = typ(TFIELD);
-	field[8]->type = typ(TARRAY);
-	field[8]->type->type = types[TUINTPTR];
-	field[8]->type->bound = 4;
-	field[8]->type->width = 4 * widthptr;
-	field[8]->sym = mal(sizeof(Sym));
-	field[8]->sym->name = "other";
+	field[0] = makefield("key", ptrto(t->down));
+	field[1] = makefield("val", ptrto(t->type));
+	field[2] = makefield("t", ptrto(types[TUINT8]));
+	field[3] = makefield("h", ptrto(hmap(t)));
+	field[4] = makefield("buckets", ptrto(mapbucket(t)));
+	field[5] = makefield("bptr", ptrto(mapbucket(t)));
+	field[6] = makefield("overflow0", types[TUNSAFEPTR]);
+	field[7] = makefield("overflow1", types[TUNSAFEPTR]);
+	field[8] = makefield("startBucket", types[TUINTPTR]);
+	field[9] = makefield("stuff", types[TUINTPTR]); // offset+wrapped+B+I
+	field[10] = makefield("bucket", types[TUINTPTR]);
+	field[11] = makefield("checkBucket", types[TUINTPTR]);
 	
 	// build iterator struct holding the above fields
 	i = typ(TSTRUCT);
 	i->noalg = 1;
 	i->type = field[0];
-	off = 0;
-	for(n = 0; n < nelem(field)-1; n++) {
+	for(n = 0; n < nelem(field)-1; n++)
 		field[n]->down = field[n+1];
-		field[n]->width = off;
-		off += field[n]->type->width;
-	}
 	field[nelem(field)-1]->down = T;
-	off += field[nelem(field)-1]->type->width;
-	if(off != 12 * widthptr)
-		yyerror("hash_iter size not correct %d %d", off, 11 * widthptr);
+	dowidth(i);
+	if(i->width != 12 * widthptr)
+		yyerror("hash_iter size not correct %d %d", i->width, 12 * widthptr);
 	t->hiter = i;
 	i->map = t;
 	return i;
@@ -547,15 +479,15 @@ dimportpath(Pkg *p)
 	n->xoffset = 0;
 	p->pathsym = n->sym;
 
-	arch.gdatastring(n, p->path);
-	arch.ggloblsym(n->sym, types[TSTRING]->width, DUPOK|RODATA);
+	gdatastring(n, p->path);
+	ggloblsym(n->sym, types[TSTRING]->width, DUPOK|RODATA);
 }
 
 static int
 dgopkgpath(Sym *s, int ot, Pkg *pkg)
 {
 	if(pkg == nil)
-		return arch.dgostringptr(s, ot, nil);
+		return dgostringptr(s, ot, nil);
 
 	// Emit reference to go.importpath.""., which 6l will
 	// rewrite using the correct import path.  Every package
@@ -565,11 +497,11 @@ dgopkgpath(Sym *s, int ot, Pkg *pkg)
 
 		if(ns == nil)
 			ns = pkglookup("importpath.\"\".", mkpkg(newstrlit("go")));
-		return arch.dsymptr(s, ot, ns, 0);
+		return dsymptr(s, ot, ns, 0);
 	}
 
 	dimportpath(pkg);
-	return arch.dsymptr(s, ot, pkg->pathsym, 0);
+	return dsymptr(s, ot, pkg->pathsym, 0);
 }
 
 /*
@@ -589,7 +521,7 @@ dextratype(Sym *sym, int off, Type *t, int ptroff)
 
 	// fill in *extraType pointer in header
 	off = rnd(off, widthptr);
-	arch.dsymptr(sym, ptroff, sym, off);
+	dsymptr(sym, ptroff, sym, off);
 
 	n = 0;
 	for(a=m; a; a=a->link) {
@@ -600,18 +532,18 @@ dextratype(Sym *sym, int off, Type *t, int ptroff)
 	ot = off;
 	s = sym;
 	if(t->sym) {
-		ot = arch.dgostringptr(s, ot, t->sym->name);
+		ot = dgostringptr(s, ot, t->sym->name);
 		if(t != types[t->etype] && t != errortype)
 			ot = dgopkgpath(s, ot, t->sym->pkg);
 		else
-			ot = arch.dgostringptr(s, ot, nil);
+			ot = dgostringptr(s, ot, nil);
 	} else {
-		ot = arch.dgostringptr(s, ot, nil);
-		ot = arch.dgostringptr(s, ot, nil);
+		ot = dgostringptr(s, ot, nil);
+		ot = dgostringptr(s, ot, nil);
 	}
 
 	// slice header
-	ot = arch.dsymptr(s, ot, s, ot + widthptr + 2*widthint);
+	ot = dsymptr(s, ot, s, ot + widthptr + 2*widthint);
 	ot = duintxx(s, ot, n, widthint);
 	ot = duintxx(s, ot, n, widthint);
 
@@ -619,16 +551,16 @@ dextratype(Sym *sym, int off, Type *t, int ptroff)
 	for(a=m; a; a=a->link) {
 		// method
 		// ../../runtime/type.go:/method
-		ot = arch.dgostringptr(s, ot, a->name);
+		ot = dgostringptr(s, ot, a->name);
 		ot = dgopkgpath(s, ot, a->pkg);
-		ot = arch.dsymptr(s, ot, dtypesym(a->mtype), 0);
-		ot = arch.dsymptr(s, ot, dtypesym(a->type), 0);
+		ot = dsymptr(s, ot, dtypesym(a->mtype), 0);
+		ot = dsymptr(s, ot, dtypesym(a->type), 0);
 		if(a->isym)
-			ot = arch.dsymptr(s, ot, a->isym, 0);
+			ot = dsymptr(s, ot, a->isym, 0);
 		else
 			ot = duintptr(s, ot, 0);
 		if(a->tsym)
-			ot = arch.dsymptr(s, ot, a->tsym, 0);
+			ot = dsymptr(s, ot, a->tsym, 0);
 		else
 			ot = duintptr(s, ot, 0);
 	}
@@ -816,43 +748,43 @@ dcommontype(Sym *s, int ot, Type *t)
 		i |= KindGCProg;
 	ot = duint8(s, ot, i);  // kind
 	if(algsym == S)
-		ot = arch.dsymptr(s, ot, algarray, alg*sizeofAlg);
+		ot = dsymptr(s, ot, algarray, alg*sizeofAlg);
 	else
-		ot = arch.dsymptr(s, ot, algsym, 0);
+		ot = dsymptr(s, ot, algsym, 0);
 	// gc
 	if(gcprog) {
 		gengcprog(t, &gcprog0, &gcprog1);
 		if(gcprog0 != S)
-			ot = arch.dsymptr(s, ot, gcprog0, 0);
+			ot = dsymptr(s, ot, gcprog0, 0);
 		else
 			ot = duintptr(s, ot, 0);
-		ot = arch.dsymptr(s, ot, gcprog1, 0);
+		ot = dsymptr(s, ot, gcprog1, 0);
 	} else {
 		gengcmask(t, gcmask);
 		x1 = 0;
 		for(i=0; i<8; i++)
 			x1 = x1<<8 | gcmask[i];
 		if(widthptr == 4) {
-			p = smprint("gcbits.%#016llux", x1);
+			p = smprint("gcbits.0x%016llux", x1);
 		} else {
 			x2 = 0;
 			for(i=0; i<8; i++)
 				x2 = x2<<8 | gcmask[i+8];
-			p = smprint("gcbits.%#016llux%016llux", x1, x2);
+			p = smprint("gcbits.0x%016llux%016llux", x1, x2);
 		}
 		sbits = pkglookup(p, runtimepkg);
 		if((sbits->flags & SymUniq) == 0) {
 			sbits->flags |= SymUniq;
 			for(i = 0; i < 2*widthptr; i++)
 				duint8(sbits, i, gcmask[i]);
-			arch.ggloblsym(sbits, 2*widthptr, DUPOK|RODATA);
+			ggloblsym(sbits, 2*widthptr, DUPOK|RODATA);
 		}
-		ot = arch.dsymptr(s, ot, sbits, 0);
+		ot = dsymptr(s, ot, sbits, 0);
 		ot = duintptr(s, ot, 0);
 	}
 	p = smprint("%-uT", t);
 	//print("dcommontype: %s\n", p);
-	ot = arch.dgostringptr(s, ot, p);	// string
+	ot = dgostringptr(s, ot, p);	// string
 	free(p);
 
 	// skip pointer to extraType,
@@ -861,8 +793,8 @@ dcommontype(Sym *s, int ot, Type *t)
 	// otherwise linker will assume 0.
 	ot += widthptr;
 
-	ot = arch.dsymptr(s, ot, sptr, 0);  // ptrto type
-	ot = arch.dsymptr(s, ot, zero, 0);  // ptr to zero value
+	ot = dsymptr(s, ot, sptr, 0);  // ptrto type
+	ot = dsymptr(s, ot, zero, 0);  // ptr to zero value
 	return ot;
 }
 
@@ -1092,15 +1024,15 @@ ok:
 			s2 = dtypesym(t2);
 			ot = dcommontype(s, ot, t);
 			xt = ot - 3*widthptr;
-			ot = arch.dsymptr(s, ot, s1, 0);
-			ot = arch.dsymptr(s, ot, s2, 0);
+			ot = dsymptr(s, ot, s1, 0);
+			ot = dsymptr(s, ot, s2, 0);
 			ot = duintptr(s, ot, t->bound);
 		} else {
 			// ../../runtime/type.go:/SliceType
 			s1 = dtypesym(t->type);
 			ot = dcommontype(s, ot, t);
 			xt = ot - 3*widthptr;
-			ot = arch.dsymptr(s, ot, s1, 0);
+			ot = dsymptr(s, ot, s1, 0);
 		}
 		break;
 
@@ -1109,7 +1041,7 @@ ok:
 		s1 = dtypesym(t->type);
 		ot = dcommontype(s, ot, t);
 		xt = ot - 3*widthptr;
-		ot = arch.dsymptr(s, ot, s1, 0);
+		ot = dsymptr(s, ot, s1, 0);
 		ot = duintptr(s, ot, t->chan);
 		break;
 
@@ -1130,21 +1062,21 @@ ok:
 
 		// two slice headers: in and out.
 		ot = rnd(ot, widthptr);
-		ot = arch.dsymptr(s, ot, s, ot+2*(widthptr+2*widthint));
+		ot = dsymptr(s, ot, s, ot+2*(widthptr+2*widthint));
 		n = t->thistuple + t->intuple;
 		ot = duintxx(s, ot, n, widthint);
 		ot = duintxx(s, ot, n, widthint);
-		ot = arch.dsymptr(s, ot, s, ot+1*(widthptr+2*widthint)+n*widthptr);
+		ot = dsymptr(s, ot, s, ot+1*(widthptr+2*widthint)+n*widthptr);
 		ot = duintxx(s, ot, t->outtuple, widthint);
 		ot = duintxx(s, ot, t->outtuple, widthint);
 
 		// slice data
 		for(t1=getthisx(t)->type; t1; t1=t1->down, n++)
-			ot = arch.dsymptr(s, ot, dtypesym(t1->type), 0);
+			ot = dsymptr(s, ot, dtypesym(t1->type), 0);
 		for(t1=getinargx(t)->type; t1; t1=t1->down, n++)
-			ot = arch.dsymptr(s, ot, dtypesym(t1->type), 0);
+			ot = dsymptr(s, ot, dtypesym(t1->type), 0);
 		for(t1=getoutargx(t)->type; t1; t1=t1->down, n++)
-			ot = arch.dsymptr(s, ot, dtypesym(t1->type), 0);
+			ot = dsymptr(s, ot, dtypesym(t1->type), 0);
 		break;
 
 	case TINTER:
@@ -1158,14 +1090,14 @@ ok:
 		// ../../runtime/type.go:/InterfaceType
 		ot = dcommontype(s, ot, t);
 		xt = ot - 3*widthptr;
-		ot = arch.dsymptr(s, ot, s, ot+widthptr+2*widthint);
+		ot = dsymptr(s, ot, s, ot+widthptr+2*widthint);
 		ot = duintxx(s, ot, n, widthint);
 		ot = duintxx(s, ot, n, widthint);
 		for(a=m; a; a=a->link) {
 			// ../../runtime/type.go:/imethod
-			ot = arch.dgostringptr(s, ot, a->name);
+			ot = dgostringptr(s, ot, a->name);
 			ot = dgopkgpath(s, ot, a->pkg);
-			ot = arch.dsymptr(s, ot, dtypesym(a->type), 0);
+			ot = dsymptr(s, ot, dtypesym(a->type), 0);
 		}
 		break;
 
@@ -1177,10 +1109,10 @@ ok:
 		s4 = dtypesym(hmap(t));
 		ot = dcommontype(s, ot, t);
 		xt = ot - 3*widthptr;
-		ot = arch.dsymptr(s, ot, s1, 0);
-		ot = arch.dsymptr(s, ot, s2, 0);
-		ot = arch.dsymptr(s, ot, s3, 0);
-		ot = arch.dsymptr(s, ot, s4, 0);
+		ot = dsymptr(s, ot, s1, 0);
+		ot = dsymptr(s, ot, s2, 0);
+		ot = dsymptr(s, ot, s3, 0);
+		ot = dsymptr(s, ot, s4, 0);
 		if(t->down->width > MAXKEYSIZE) {
 			ot = duint8(s, ot, widthptr);
 			ot = duint8(s, ot, 1); // indirect
@@ -1210,7 +1142,7 @@ ok:
 		s1 = dtypesym(t->type);
 		ot = dcommontype(s, ot, t);
 		xt = ot - 3*widthptr;
-		ot = arch.dsymptr(s, ot, s1, 0);
+		ot = dsymptr(s, ot, s1, 0);
 		break;
 
 	case TSTRUCT:
@@ -1223,32 +1155,32 @@ ok:
 		}
 		ot = dcommontype(s, ot, t);
 		xt = ot - 3*widthptr;
-		ot = arch.dsymptr(s, ot, s, ot+widthptr+2*widthint);
+		ot = dsymptr(s, ot, s, ot+widthptr+2*widthint);
 		ot = duintxx(s, ot, n, widthint);
 		ot = duintxx(s, ot, n, widthint);
 		for(t1=t->type; t1!=T; t1=t1->down) {
 			// ../../runtime/type.go:/structField
 			if(t1->sym && !t1->embedded) {
-				ot = arch.dgostringptr(s, ot, t1->sym->name);
+				ot = dgostringptr(s, ot, t1->sym->name);
 				if(exportname(t1->sym->name))
-					ot = arch.dgostringptr(s, ot, nil);
+					ot = dgostringptr(s, ot, nil);
 				else
 					ot = dgopkgpath(s, ot, t1->sym->pkg);
 			} else {
-				ot = arch.dgostringptr(s, ot, nil);
+				ot = dgostringptr(s, ot, nil);
 				if(t1->type->sym != S && t1->type->sym->pkg == builtinpkg)
 					ot = dgopkgpath(s, ot, localpkg);
 				else
-					ot = arch.dgostringptr(s, ot, nil);
+					ot = dgostringptr(s, ot, nil);
 			}
-			ot = arch.dsymptr(s, ot, dtypesym(t1->type), 0);
-			ot = arch.dgostrlitptr(s, ot, t1->note);
+			ot = dsymptr(s, ot, dtypesym(t1->type), 0);
+			ot = dgostrlitptr(s, ot, t1->note);
 			ot = duintptr(s, ot, t1->width);	// field offset
 		}
 		break;
 	}
 	ot = dextratype(s, ot, t, xt);
-	arch.ggloblsym(s, ot, dupok|RODATA);
+	ggloblsym(s, ot, dupok|RODATA);
 
 	// generate typelink.foo pointing at s = type.foo.
 	// The linker will leave a table of all the typelinks for
@@ -1261,8 +1193,8 @@ ok:
 		case TCHAN:
 		case TMAP:
 			slink = typelinksym(t);
-			arch.dsymptr(slink, 0, s, 0);
-			arch.ggloblsym(slink, widthptr, dupok|RODATA);
+			dsymptr(slink, 0, s, 0);
+			ggloblsym(slink, widthptr, dupok|RODATA);
 		}
 	}
 
@@ -1354,18 +1286,18 @@ dalgsym(Type *t)
 		hashfunc = pkglookup(p, typepkg);
 		free(p);
 		ot = 0;
-		ot = arch.dsymptr(hashfunc, ot, pkglookup("memhash_varlen", runtimepkg), 0);
+		ot = dsymptr(hashfunc, ot, pkglookup("memhash_varlen", runtimepkg), 0);
 		ot = duintxx(hashfunc, ot, t->width, widthptr); // size encoded in closure
-		arch.ggloblsym(hashfunc, ot, DUPOK|RODATA);
+		ggloblsym(hashfunc, ot, DUPOK|RODATA);
 
 		// make equality closure
 		p = smprint(".eqfunc%lld", t->width);
 		eqfunc = pkglookup(p, typepkg);
 		free(p);
 		ot = 0;
-		ot = arch.dsymptr(eqfunc, ot, pkglookup("memequal_varlen", runtimepkg), 0);
+		ot = dsymptr(eqfunc, ot, pkglookup("memequal_varlen", runtimepkg), 0);
 		ot = duintxx(eqfunc, ot, t->width, widthptr);
-		arch.ggloblsym(eqfunc, ot, DUPOK|RODATA);
+		ggloblsym(eqfunc, ot, DUPOK|RODATA);
 	} else {
 		// generate an alg table specific to this type
 		s = typesymprefix(".alg", t);
@@ -1378,16 +1310,16 @@ dalgsym(Type *t)
 		geneq(eq, t);
 
 		// make Go funcs (closures) for calling hash and equal from Go
-		arch.dsymptr(hashfunc, 0, hash, 0);
-		arch.ggloblsym(hashfunc, widthptr, DUPOK|RODATA);
-		arch.dsymptr(eqfunc, 0, eq, 0);
-		arch.ggloblsym(eqfunc, widthptr, DUPOK|RODATA);
+		dsymptr(hashfunc, 0, hash, 0);
+		ggloblsym(hashfunc, widthptr, DUPOK|RODATA);
+		dsymptr(eqfunc, 0, eq, 0);
+		ggloblsym(eqfunc, widthptr, DUPOK|RODATA);
 	}
 	// ../../runtime/alg.go:/typeAlg
 	ot = 0;
-	ot = arch.dsymptr(s, ot, hashfunc, 0);
-	ot = arch.dsymptr(s, ot, eqfunc, 0);
-	arch.ggloblsym(s, ot, DUPOK|RODATA);
+	ot = dsymptr(s, ot, hashfunc, 0);
+	ot = dsymptr(s, ot, eqfunc, 0);
+	ggloblsym(s, ot, DUPOK|RODATA);
 	return s;
 }
 
@@ -1417,7 +1349,7 @@ usegcprog(Type *t)
 
 // Generates sparse GC bitmask (4 bits per word).
 static void
-gengcmask(Type *t, uint8 gcmask[16])
+gengcmask(Type *t, uint8 *gcmask)
 {
 	Bvec *vec;
 	vlong xoffset, nptr, i, j;
@@ -1570,7 +1502,7 @@ gengcprog(Type *t, Sym **pgc0, Sym **pgc1)
 	// Don't generate it if it's too large, runtime will unroll directly into GC bitmap.
 	if(size <= MaxGCMask) {
 		gc0 = typesymprefix(".gc", t);
-		arch.ggloblsym(gc0, size, DUPOK|NOPTR);
+		ggloblsym(gc0, size, DUPOK|NOPTR);
 		*pgc0 = gc0;
 	}
 
@@ -1580,7 +1512,7 @@ gengcprog(Type *t, Sym **pgc0, Sym **pgc1)
 	xoffset = 0;
 	gengcprog1(&g, t, &xoffset);
 	ot = proggenfini(&g);
-	arch.ggloblsym(gc1, ot, DUPOK|RODATA);
+	ggloblsym(gc1, ot, DUPOK|RODATA);
 	*pgc1 = gc1;
 }
 
diff --git a/src/cmd/gc/reg.c b/src/cmd/gc/reg.c
new file mode 100644
index 0000000000..67409c2127
--- /dev/null
+++ b/src/cmd/gc/reg.c
@@ -0,0 +1,1233 @@
+// Derived from Inferno utils/6c/reg.c
+// http://code.google.com/p/inferno-os/source/browse/utils/6c/reg.c
+//
+//	Copyright © 1994-1999 Lucent Technologies Inc.  All rights reserved.
+//	Portions Copyright © 1995-1997 C H Forsyth (forsyth@terzarima.net)
+//	Portions Copyright © 1997-1999 Vita Nuova Limited
+//	Portions Copyright © 2000-2007 Vita Nuova Holdings Limited (www.vitanuova.com)
+//	Portions Copyright © 2004,2006 Bruce Ellis
+//	Portions Copyright © 2005-2007 C H Forsyth (forsyth@terzarima.net)
+//	Revisions Copyright © 2000-2007 Lucent Technologies Inc. and others
+//	Portions Copyright © 2009 The Go Authors.  All rights reserved.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a copy
+// of this software and associated documentation files (the "Software"), to deal
+// in the Software without restriction, including without limitation the rights
+// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+// copies of the Software, and to permit persons to whom the Software is
+// furnished to do so, subject to the following conditions:
+//
+// The above copyright notice and this permission notice shall be included in
+// all copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+// THE SOFTWARE.
+
+#include <u.h>
+#include <libc.h>
+#include "go.h"
+#include "popt.h"
+
+static	Flow*	firstf;
+static	int	first	= 1;
+
+static void	addmove(Flow*, int, int, int);
+static Bits	mkvar(Flow*, Adr*);
+static void	prop(Flow*, Bits, Bits);
+static void	synch(Flow*, Bits);
+static uint64	allreg(uint64, Rgn*);
+static void	paint1(Flow*, int);
+static uint64	paint2(Flow*, int, int);
+static void	paint3(Flow*, int, uint64, int);
+static void	addreg(Adr*, int);
+
+static int
+rcmp(const void *a1, const void *a2)
+{
+	Rgn *p1, *p2;
+
+	p1 = (Rgn*)a1;
+	p2 = (Rgn*)a2;
+	if(p1->cost != p2->cost)
+		return p2->cost - p1->cost;
+	if(p1->varno != p2->varno)
+		return p2->varno - p1->varno;
+	if(p1->enter != p2->enter)
+		return p2->enter->id - p1->enter->id;
+	return 0;
+}
+
+static void
+setaddrs(Bits bit)
+{
+	int i, n;
+	Var *v;
+	Node *node;
+
+	while(bany(&bit)) {
+		// convert each bit to a variable
+		i = bnum(bit);
+		node = var[i].node;
+		n = var[i].name;
+		biclr(&bit, i);
+
+		// disable all pieces of that variable
+		for(i=0; i<nvar; i++) {
+			v = var+i;
+			if(v->node == node && v->name == n)
+				v->addr = 2;
+		}
+	}
+}
+
+static Node* regnodes[64];
+
+static void walkvardef(Node *n, Flow *r, int active);
+
+void
+regopt(Prog *firstp)
+{
+	Flow *f, *f1;
+	Reg *r;
+	Prog *p;
+	Graph *g;
+	ProgInfo info;
+	int i, z, active;
+	uint64 vreg, usedreg;
+	uint64 mask;
+	int nreg;
+	char **regnames;
+	Bits bit;
+	Rgn *rgp;
+
+	if(first) {
+		fmtinstall('Q', Qconv);
+		first = 0;
+	}
+
+	mergetemp(firstp);
+
+	/*
+	 * control flow is more complicated in generated go code
+	 * than in generated c code.  define pseudo-variables for
+	 * registers, so we have complete register usage information.
+	 */
+	regnames = thearch.regnames(&nreg);
+	nvar = nreg;
+	memset(var, 0, nreg*sizeof var[0]);
+	for(i=0; i<nreg; i++) {
+		if(regnodes[i] == N)
+			regnodes[i] = newname(lookup(regnames[i]));
+		var[i].node = regnodes[i];
+	}
+
+	regbits = thearch.excludedregs();
+	externs = zbits;
+	params = zbits;
+	consts = zbits;
+	addrs = zbits;
+	ivar = zbits;
+	ovar = zbits;
+
+	/*
+	 * pass 1
+	 * build aux data structure
+	 * allocate pcs
+	 * find use and set of variables
+	 */
+	g = flowstart(firstp, sizeof(Reg));
+	if(g == nil) {
+		for(i=0; i<nvar; i++)
+			var[i].node->opt = nil;
+		return;
+	}
+
+	firstf = g->start;
+
+	for(f = firstf; f != nil; f = f->link) {
+		p = f->prog;
+		if(p->as == AVARDEF || p->as == AVARKILL)
+			continue;
+		thearch.proginfo(&info, p);
+
+		// Avoid making variables for direct-called functions.
+		if(p->as == ACALL && p->to.type == TYPE_MEM && p->to.name == NAME_EXTERN)
+			continue;
+
+		// from vs to doesn't matter for registers.
+		r = (Reg*)f->data;
+		r->use1.b[0] |= info.reguse | info.regindex;
+		r->set.b[0] |= info.regset;
+
+		bit = mkvar(f, &p->from);
+		if(bany(&bit)) {
+			if(info.flags & LeftAddr)
+				setaddrs(bit);
+			if(info.flags & LeftRead)
+				for(z=0; z<BITS; z++)
+					r->use1.b[z] |= bit.b[z];
+			if(info.flags & LeftWrite)
+				for(z=0; z<BITS; z++)
+					r->set.b[z] |= bit.b[z];
+		}
+
+		// Compute used register for reg
+		if(info.flags & RegRead)
+			r->use1.b[0] |= thearch.RtoB(p->reg);
+
+		// Currently we never generate three register forms.
+		// If we do, this will need to change.
+		if(p->from3.type != TYPE_NONE)
+			fatal("regopt not implemented for from3");
+
+		bit = mkvar(f, &p->to);
+		if(bany(&bit)) {	
+			if(info.flags & RightAddr)
+				setaddrs(bit);
+			if(info.flags & RightRead)
+				for(z=0; z<BITS; z++)
+					r->use2.b[z] |= bit.b[z];
+			if(info.flags & RightWrite)
+				for(z=0; z<BITS; z++)
+					r->set.b[z] |= bit.b[z];
+		}
+	}
+
+	for(i=0; i<nvar; i++) {
+		Var *v;
+		v = var+i;
+		if(v->addr) {
+			bit = blsh(i);
+			for(z=0; z<BITS; z++)
+				addrs.b[z] |= bit.b[z];
+		}
+
+		if(debug['R'] && debug['v'])
+			print("bit=%2d addr=%d et=%E w=%-2d s=%N + %lld\n",
+				i, v->addr, v->etype, v->width, v->node, v->offset);
+	}
+
+	if(debug['R'] && debug['v'])
+		dumpit("pass1", firstf, 1);
+
+	/*
+	 * pass 2
+	 * find looping structure
+	 */
+	flowrpo(g);
+
+	if(debug['R'] && debug['v'])
+		dumpit("pass2", firstf, 1);
+
+	/*
+	 * pass 2.5
+	 * iterate propagating fat vardef covering forward
+	 * r->act records vars with a VARDEF since the last CALL.
+	 * (r->act will be reused in pass 5 for something else,
+	 * but we'll be done with it by then.)
+	 */
+	active = 0;
+	for(f = firstf; f != nil; f = f->link) {
+		f->active = 0;
+		r = (Reg*)f->data;
+		r->act = zbits;
+	}
+	for(f = firstf; f != nil; f = f->link) {
+		p = f->prog;
+		if(p->as == AVARDEF && isfat(((Node*)(p->to.node))->type) && ((Node*)(p->to.node))->opt != nil) {
+			active++;
+			walkvardef(p->to.node, f, active);
+		}
+	}
+
+	/*
+	 * pass 3
+	 * iterate propagating usage
+	 * 	back until flow graph is complete
+	 */
+loop1:
+	change = 0;
+	for(f = firstf; f != nil; f = f->link)
+		f->active = 0;
+	for(f = firstf; f != nil; f = f->link)
+		if(f->prog->as == ARET)
+			prop(f, zbits, zbits);
+loop11:
+	/* pick up unreachable code */
+	i = 0;
+	for(f = firstf; f != nil; f = f1) {
+		f1 = f->link;
+		if(f1 && f1->active && !f->active) {
+			prop(f, zbits, zbits);
+			i = 1;
+		}
+	}
+	if(i)
+		goto loop11;
+	if(change)
+		goto loop1;
+
+	if(debug['R'] && debug['v'])
+		dumpit("pass3", firstf, 1);
+
+	/*
+	 * pass 4
+	 * iterate propagating register/variable synchrony
+	 * 	forward until graph is complete
+	 */
+loop2:
+	change = 0;
+	for(f = firstf; f != nil; f = f->link)
+		f->active = 0;
+	synch(firstf, zbits);
+	if(change)
+		goto loop2;
+
+	if(debug['R'] && debug['v'])
+		dumpit("pass4", firstf, 1);
+
+	/*
+	 * pass 4.5
+	 * move register pseudo-variables into regu.
+	 */
+	if(nreg == 64)
+		mask = ~0ULL; // can't rely on C to shift by 64
+	else
+		mask = (1ULL<<nreg) - 1;
+	for(f = firstf; f != nil; f = f->link) {
+		r = (Reg*)f->data;
+		r->regu = (r->refbehind.b[0] | r->set.b[0]) & mask;
+		r->set.b[0] &= ~mask;
+		r->use1.b[0] &= ~mask;
+		r->use2.b[0] &= ~mask;
+		r->refbehind.b[0] &= ~mask;
+		r->refahead.b[0] &= ~mask;
+		r->calbehind.b[0] &= ~mask;
+		r->calahead.b[0] &= ~mask;
+		r->regdiff.b[0] &= ~mask;
+		r->act.b[0] &= ~mask;
+	}
+
+	if(debug['R'] && debug['v'])
+		dumpit("pass4.5", firstf, 1);
+
+	/*
+	 * pass 5
+	 * isolate regions
+	 * calculate costs (paint1)
+	 */
+	f = firstf;
+	if(f) {
+		r = (Reg*)f->data;
+		for(z=0; z<BITS; z++)
+			bit.b[z] = (r->refahead.b[z] | r->calahead.b[z]) &
+			  ~(externs.b[z] | params.b[z] | addrs.b[z] | consts.b[z]);
+		if(bany(&bit) && !f->refset) {
+			// should never happen - all variables are preset
+			if(debug['w'])
+				print("%L: used and not set: %Q\n", f->prog->lineno, bit);
+			f->refset = 1;
+		}
+	}
+	for(f = firstf; f != nil; f = f->link)
+		((Reg*)f->data)->act = zbits;
+	nregion = 0;
+	for(f = firstf; f != nil; f = f->link) {
+		r = (Reg*)f->data;
+		for(z=0; z<BITS; z++)
+			bit.b[z] = r->set.b[z] &
+			  ~(r->refahead.b[z] | r->calahead.b[z] | addrs.b[z]);
+		if(bany(&bit) && !f->refset) {
+			if(debug['w'])
+				print("%L: set and not used: %Q\n", f->prog->lineno, bit);
+			f->refset = 1;
+			thearch.excise(f);
+		}
+		for(z=0; z<BITS; z++)
+			bit.b[z] = LOAD(r) & ~(r->act.b[z] | addrs.b[z]);
+		while(bany(&bit)) {
+			i = bnum(bit);
+			change = 0;
+			paint1(f, i);
+			biclr(&bit, i);
+			if(change <= 0)
+				continue;
+			if(nregion >= NRGN) {
+				if(debug['R'] && debug['v'])
+					print("too many regions\n");
+				goto brk;
+			}
+			rgp = &region[nregion];
+			rgp->enter = f;
+			rgp->varno = i;
+			rgp->cost = change;
+			nregion++;
+		}
+	}
+brk:
+	qsort(region, nregion, sizeof(region[0]), rcmp);
+
+	if(debug['R'] && debug['v'])
+		dumpit("pass5", firstf, 1);
+
+	/*
+	 * pass 6
+	 * determine used registers (paint2)
+	 * replace code (paint3)
+	 */
+	if(debug['R'] && debug['v'])
+		print("\nregisterizing\n");
+	for(i=0; i<nregion; i++) {
+		rgp = &region[i];
+		if(debug['R'] && debug['v'])
+			print("region %d: cost %d varno %d enter %lld\n", i, rgp->cost, rgp->varno, rgp->enter->prog->pc);
+		bit = blsh(rgp->varno);
+		usedreg = paint2(rgp->enter, rgp->varno, 0);
+		vreg = allreg(usedreg, rgp);
+		if(rgp->regno != 0) {
+			if(debug['R'] && debug['v']) {
+				Var *v;
+
+				v = var + rgp->varno;
+				print("registerize %N+%lld (bit=%2d et=%E) in %R usedreg=%#llx vreg=%#llx\n",
+						v->node, v->offset, rgp->varno, v->etype, rgp->regno, usedreg, vreg);
+			}
+			paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
+		}
+	}
+
+	/*
+	 * free aux structures. peep allocates new ones.
+	 */
+	for(i=0; i<nvar; i++)
+		var[i].node->opt = nil;
+	flowend(g);
+	firstf = nil;
+
+	if(debug['R'] && debug['v']) {
+		// Rebuild flow graph, since we inserted instructions
+		g = flowstart(firstp, 0);
+		firstf = g->start;
+		dumpit("pass6", firstf, 0);
+		flowend(g);
+		firstf = nil;
+	}
+
+	/*
+	 * pass 7
+	 * peep-hole on basic block
+	 */
+	if(!debug['R'] || debug['P'])
+		thearch.peep(firstp);
+
+	/*
+	 * eliminate nops
+	 */
+	for(p=firstp; p!=P; p=p->link) {
+		while(p->link != P && p->link->as == ANOP)
+			p->link = p->link->link;
+		if(p->to.type == TYPE_BRANCH)
+			while(p->to.u.branch != P && p->to.u.branch->as == ANOP)
+				p->to.u.branch = p->to.u.branch->link;
+	}
+
+	if(debug['R']) {
+		if(ostats.ncvtreg ||
+		   ostats.nspill ||
+		   ostats.nreload ||
+		   ostats.ndelmov ||
+		   ostats.nvar ||
+		   ostats.naddr ||
+		   0)
+			print("\nstats\n");
+
+		if(ostats.ncvtreg)
+			print("	%4d cvtreg\n", ostats.ncvtreg);
+		if(ostats.nspill)
+			print("	%4d spill\n", ostats.nspill);
+		if(ostats.nreload)
+			print("	%4d reload\n", ostats.nreload);
+		if(ostats.ndelmov)
+			print("	%4d delmov\n", ostats.ndelmov);
+		if(ostats.nvar)
+			print("	%4d var\n", ostats.nvar);
+		if(ostats.naddr)
+			print("	%4d addr\n", ostats.naddr);
+
+		memset(&ostats, 0, sizeof(ostats));
+	}
+}
+
+static void
+walkvardef(Node *n, Flow *f, int active)
+{
+	Flow *f1, *f2;
+	int bn;
+	Var *v;
+	
+	for(f1=f; f1!=nil; f1=f1->s1) {
+		if(f1->active == active)
+			break;
+		f1->active = active;
+		if(f1->prog->as == AVARKILL && f1->prog->to.node == n)
+			break;
+		for(v=n->opt; v!=nil; v=v->nextinnode) {
+			bn = v->id;
+			biset(&((Reg*)f1->data)->act, bn);
+		}
+		if(f1->prog->as == ACALL)
+			break;
+	}
+
+	for(f2=f; f2!=f1; f2=f2->s1)
+		if(f2->s2 != nil)
+			walkvardef(n, f2->s2, active);
+}
+
+/*
+ * add mov b,rn
+ * just after r
+ */
+static void
+addmove(Flow *r, int bn, int rn, int f)
+{
+	Prog *p, *p1;
+	Adr *a;
+	Var *v;
+
+	p1 = mal(sizeof(*p1));
+	clearp(p1);
+	p1->pc = 9999;
+
+	p = r->prog;
+	p1->link = p->link;
+	p->link = p1;
+	p1->lineno = p->lineno;
+
+	v = var + bn;
+
+	a = &p1->to;
+	a->offset = v->offset;
+	a->etype = v->etype;
+	a->type = TYPE_MEM;
+	a->name = v->name;
+	a->node = v->node;
+	a->sym = linksym(v->node->sym);
+	/* NOTE(rsc): 9g did
+	if(a->etype == TARRAY)
+		a->type = TYPE_ADDR;
+	else if(a->sym == nil)
+		a->type = TYPE_CONST;
+	*/
+
+	p1->as = thearch.optoas(OAS, types[(uchar)v->etype]);
+	// TODO(rsc): Remove special case here.
+	if((thearch.thechar == '9' || thearch.thechar == '5') && v->etype == TBOOL)
+		p1->as = thearch.optoas(OAS, types[TUINT8]);
+	p1->from.type = TYPE_REG;
+	p1->from.reg = rn;
+	p1->from.name = NAME_NONE;
+	if(!f) {
+		p1->from = *a;
+		*a = zprog.from;
+		a->type = TYPE_REG;
+		a->reg = rn;
+	}
+	if(debug['R'] && debug['v'])
+		print("%P ===add=== %P\n", p, p1);
+	ostats.nspill++;
+}
+
+static int
+overlap(int64 o1, int w1, int64 o2, int w2)
+{
+	int64 t1, t2;
+
+	t1 = o1+w1;
+	t2 = o2+w2;
+
+	if(!(t1 > o2 && t2 > o1))
+		return 0;
+
+	return 1;
+}
+
+static Bits
+mkvar(Flow *f, Adr *a)
+{
+	Var *v;
+	int i, n, et, z, flag;
+	int64 w;
+	uint64 regu;
+	int64 o;
+	Bits bit;
+	Node *node;
+	Reg *r;
+	
+
+	/*
+	 * mark registers used
+	 */
+	if(a->type == TYPE_NONE)
+		goto none;
+
+	r = (Reg*)f->data;
+	r->use1.b[0] |= thearch.doregbits(a->index); // TODO: Use RtoB
+
+	switch(a->type) {
+	default:
+		regu = thearch.doregbits(a->reg) | thearch.RtoB(a->reg); // TODO: Use RtoB
+		if(regu == 0)
+			goto none;
+		bit = zbits;
+		bit.b[0] = regu;
+		return bit;
+
+	case TYPE_ADDR:
+		// TODO(rsc): Remove special case here.
+		if(thearch.thechar == '9' || thearch.thechar == '5')
+			goto memcase;
+		a->type = TYPE_MEM;
+		bit = mkvar(f, a);
+		setaddrs(bit);
+		a->type = TYPE_ADDR;
+		ostats.naddr++;
+		goto none;
+
+	case TYPE_MEM:
+	memcase:
+		if(r != R) {
+			r->use1.b[0] |= thearch.RtoB(a->reg);
+			/* NOTE: 5g did
+				if(r->f.prog->scond & (C_PBIT|C_WBIT))
+					r->set.b[0] |= RtoB(a->reg);
+			*/
+		}
+		switch(a->name) {
+		default:
+			goto none;
+		case NAME_EXTERN:
+		case NAME_STATIC:
+		case NAME_PARAM:
+		case NAME_AUTO:
+			n = a->name;
+			break;
+		}
+	}
+
+	node = a->node;
+	if(node == N || node->op != ONAME || node->orig == N)
+		goto none;
+	node = node->orig;
+	if(node->orig != node)
+		fatal("%D: bad node", a);
+	if(node->sym == S || node->sym->name[0] == '.')
+		goto none;
+	et = a->etype;
+	o = a->offset;
+	w = a->width;
+	if(w < 0)
+		fatal("bad width %lld for %D", w, a);
+
+	flag = 0;
+	for(i=0; i<nvar; i++) {
+		v = var+i;
+		if(v->node == node && v->name == n) {
+			if(v->offset == o)
+			if(v->etype == et)
+			if(v->width == w) {
+				// TODO(rsc): Remove special case for arm here.
+				if(!flag || thearch.thechar != '5')
+					return blsh(i);
+			}
+
+			// if they overlap, disable both
+			if(overlap(v->offset, v->width, o, w)) {
+//				print("disable overlap %s %d %d %d %d, %E != %E\n", s->name, v->offset, v->width, o, w, v->etype, et);
+				v->addr = 1;
+				flag = 1;
+			}
+		}
+	}
+
+	switch(et) {
+	case 0:
+	case TFUNC:
+		goto none;
+	}
+
+	if(nvar >= NVAR) {
+		if(debug['w'] > 1 && node != N)
+			fatal("variable not optimized: %#N", node);
+		
+		// If we're not tracking a word in a variable, mark the rest as
+		// having its address taken, so that we keep the whole thing
+		// live at all calls. otherwise we might optimize away part of
+		// a variable but not all of it.
+		for(i=0; i<nvar; i++) {
+			v = var+i;
+			if(v->node == node)
+				v->addr = 1;
+		}
+		goto none;
+	}
+
+	i = nvar;
+	nvar++;
+	v = var+i;
+	v->id = i;
+	v->offset = o;
+	v->name = n;
+	v->etype = et;
+	v->width = w;
+	v->addr = flag;		// funny punning
+	v->node = node;
+	
+	// node->opt is the head of a linked list
+	// of Vars within the given Node, so that
+	// we can start at a Var and find all the other
+	// Vars in the same Go variable.
+	v->nextinnode = node->opt;
+	node->opt = v;
+
+	bit = blsh(i);
+	if(n == NAME_EXTERN || n == NAME_STATIC)
+		for(z=0; z<BITS; z++)
+			externs.b[z] |= bit.b[z];
+	if(n == NAME_PARAM)
+		for(z=0; z<BITS; z++)
+			params.b[z] |= bit.b[z];
+
+	if(node->class == PPARAM)
+		for(z=0; z<BITS; z++)
+			ivar.b[z] |= bit.b[z];
+	if(node->class == PPARAMOUT)
+		for(z=0; z<BITS; z++)
+			ovar.b[z] |= bit.b[z];
+
+	// Treat values with their address taken as live at calls,
+	// because the garbage collector's liveness analysis in ../gc/plive.c does.
+	// These must be consistent or else we will elide stores and the garbage
+	// collector will see uninitialized data.
+	// The typical case where our own analysis is out of sync is when the
+	// node appears to have its address taken but that code doesn't actually
+	// get generated and therefore doesn't show up as an address being
+	// taken when we analyze the instruction stream.
+	// One instance of this case is when a closure uses the same name as
+	// an outer variable for one of its own variables declared with :=.
+	// The parser flags the outer variable as possibly shared, and therefore
+	// sets addrtaken, even though it ends up not being actually shared.
+	// If we were better about _ elision, _ = &x would suffice too.
+	// The broader := in a closure problem is mentioned in a comment in
+	// closure.c:/^typecheckclosure and dcl.c:/^oldname.
+	if(node->addrtaken)
+		v->addr = 1;
+
+	// Disable registerization for globals, because:
+	// (1) we might panic at any time and we want the recovery code
+	// to see the latest values (issue 1304).
+	// (2) we don't know what pointers might point at them and we want
+	// loads via those pointers to see updated values and vice versa (issue 7995).
+	//
+	// Disable registerization for results if using defer, because the deferred func
+	// might recover and return, causing the current values to be used.
+	if(node->class == PEXTERN || (hasdefer && node->class == PPARAMOUT))
+		v->addr = 1;
+
+	if(debug['R'])
+		print("bit=%2d et=%E w=%lld+%lld %#N %D flag=%d\n", i, et, o, w, node, a, v->addr);
+	ostats.nvar++;
+
+	return bit;
+
+none:
+	return zbits;
+}
+
+static void
+prop(Flow *f, Bits ref, Bits cal)
+{
+	Flow *f1, *f2;
+	Reg *r, *r1;
+	int z, i;
+	Var *v, *v1;
+
+	for(f1 = f; f1 != nil; f1 = f1->p1) {
+		r1 = (Reg*)f1->data;
+		for(z=0; z<BITS; z++) {
+			ref.b[z] |= r1->refahead.b[z];
+			if(ref.b[z] != r1->refahead.b[z]) {
+				r1->refahead.b[z] = ref.b[z];
+				change++;
+			}
+			cal.b[z] |= r1->calahead.b[z];
+			if(cal.b[z] != r1->calahead.b[z]) {
+				r1->calahead.b[z] = cal.b[z];
+				change++;
+			}
+		}
+		switch(f1->prog->as) {
+		case ACALL:
+			if(noreturn(f1->prog))
+				break;
+
+			// Mark all input variables (ivar) as used, because that's what the
+			// liveness bitmaps say. The liveness bitmaps say that so that a
+			// panic will not show stale values in the parameter dump.
+			// Mark variables with a recent VARDEF (r1->act) as used,
+			// so that the optimizer flushes initializations to memory,
+			// so that if a garbage collection happens during this CALL,
+			// the collector will see initialized memory. Again this is to
+			// match what the liveness bitmaps say.
+			for(z=0; z<BITS; z++) {
+				cal.b[z] |= ref.b[z] | externs.b[z] | ivar.b[z] | r1->act.b[z];
+				ref.b[z] = 0;
+			}
+			
+			// cal.b is the current approximation of what's live across the call.
+			// Every bit in cal.b is a single stack word. For each such word,
+			// find all the other tracked stack words in the same Go variable
+			// (struct/slice/string/interface) and mark them live too.
+			// This is necessary because the liveness analysis for the garbage
+			// collector works at variable granularity, not at word granularity.
+			// It is fundamental for slice/string/interface: the garbage collector
+			// needs the whole value, not just some of the words, in order to
+			// interpret the other bits correctly. Specifically, slice needs a consistent
+			// ptr and cap, string needs a consistent ptr and len, and interface
+			// needs a consistent type word and data word.
+			for(z=0; z<BITS; z++) {
+				if(cal.b[z] == 0)
+					continue;
+				for(i=0; i<64; i++) {
+					if(z*64+i >= nvar || ((cal.b[z]>>i)&1) == 0)
+						continue;
+					v = var+z*64+i;
+					if(v->node->opt == nil) // v represents fixed register, not Go variable
+						continue;
+
+					// v->node->opt is the head of a linked list of Vars
+					// corresponding to tracked words from the Go variable v->node.
+					// Walk the list and set all the bits.
+					// For a large struct this could end up being quadratic:
+					// after the first setting, the outer loop (for z, i) would see a 1 bit
+					// for all of the remaining words in the struct, and for each such
+					// word would go through and turn on all the bits again.
+					// To avoid the quadratic behavior, we only turn on the bits if
+					// v is the head of the list or if the head's bit is not yet turned on.
+					// This will set the bits at most twice, keeping the overall loop linear.
+					v1 = v->node->opt;
+					if(v == v1 || !btest(&cal, v1->id)) {
+						for(; v1 != nil; v1 = v1->nextinnode) {
+							biset(&cal, v1->id);
+						}
+					}
+				}
+			}
+			break;
+
+		case ATEXT:
+			for(z=0; z<BITS; z++) {
+				cal.b[z] = 0;
+				ref.b[z] = 0;
+			}
+			break;
+
+		case ARET:
+			for(z=0; z<BITS; z++) {
+				cal.b[z] = externs.b[z] | ovar.b[z];
+				ref.b[z] = 0;
+			}
+			break;
+		}
+		for(z=0; z<BITS; z++) {
+			ref.b[z] = (ref.b[z] & ~r1->set.b[z]) |
+				r1->use1.b[z] | r1->use2.b[z];
+			cal.b[z] &= ~(r1->set.b[z] | r1->use1.b[z] | r1->use2.b[z]);
+			r1->refbehind.b[z] = ref.b[z];
+			r1->calbehind.b[z] = cal.b[z];
+		}
+		if(f1->active)
+			break;
+		f1->active = 1;
+	}
+
+	for(; f != f1; f = f->p1) {
+		r = (Reg*)f->data;
+		for(f2 = f->p2; f2 != nil; f2 = f2->p2link)
+			prop(f2, r->refbehind, r->calbehind);
+	}
+}
+
+static void
+synch(Flow *f, Bits dif)
+{
+	Flow *f1;
+	Reg *r1;
+	int z;
+
+	for(f1 = f; f1 != nil; f1 = f1->s1) {
+		r1 = (Reg*)f1->data;
+		for(z=0; z<BITS; z++) {
+			dif.b[z] = (dif.b[z] &
+				~(~r1->refbehind.b[z] & r1->refahead.b[z])) |
+					r1->set.b[z] | r1->regdiff.b[z];
+			if(dif.b[z] != r1->regdiff.b[z]) {
+				r1->regdiff.b[z] = dif.b[z];
+				change++;
+			}
+		}
+		if(f1->active)
+			break;
+		f1->active = 1;
+		for(z=0; z<BITS; z++)
+			dif.b[z] &= ~(~r1->calbehind.b[z] & r1->calahead.b[z]);
+		if(f1->s2 != nil)
+			synch(f1->s2, dif);
+	}
+}
+
+static uint64
+allreg(uint64 b, Rgn *r)
+{
+	Var *v;
+	int i;
+
+	v = var + r->varno;
+	r->regno = 0;
+	switch(v->etype) {
+
+	default:
+		fatal("unknown etype %d/%E", bitno(b), v->etype);
+		break;
+
+	case TINT8:
+	case TUINT8:
+	case TINT16:
+	case TUINT16:
+	case TINT32:
+	case TUINT32:
+	case TINT64:
+	case TUINT64:
+	case TINT:
+	case TUINT:
+	case TUINTPTR:
+	case TBOOL:
+	case TPTR32:
+	case TPTR64:
+		i = thearch.BtoR(~b);
+		if(i && r->cost > 0) {
+			r->regno = i;
+			return thearch.RtoB(i);
+		}
+		break;
+
+	case TFLOAT32:
+	case TFLOAT64:
+		i = thearch.BtoF(~b);
+		if(i && r->cost > 0) {
+			r->regno = i;
+			return thearch.FtoB(i);
+		}
+		break;
+	}
+	return 0;
+}
+
+static void
+paint1(Flow *f, int bn)
+{
+	Flow *f1;
+	Reg *r, *r1;
+	int z;
+	uint64 bb;
+
+	z = bn/64;
+	bb = 1LL<<(bn%64);
+	r = (Reg*)f->data;
+	if(r->act.b[z] & bb)
+		return;
+	for(;;) {
+		if(!(r->refbehind.b[z] & bb))
+			break;
+		f1 = f->p1;
+		if(f1 == nil)
+			break;
+		r1 = (Reg*)f1->data;
+		if(!(r1->refahead.b[z] & bb))
+			break;
+		if(r1->act.b[z] & bb)
+			break;
+		f = f1;
+		r = r1;
+	}
+
+	if(LOAD(r) & ~(r->set.b[z]&~(r->use1.b[z]|r->use2.b[z])) & bb) {
+		change -= CLOAD * f->loop;
+	}
+	for(;;) {
+		r->act.b[z] |= bb;
+
+		if(f->prog->as != ANOP) { // don't give credit for NOPs
+			if(r->use1.b[z] & bb)
+				change += CREF * f->loop;
+			if((r->use2.b[z]|r->set.b[z]) & bb)
+				change += CREF * f->loop;
+		}
+
+		if(STORE(r) & r->regdiff.b[z] & bb) {
+			change -= CLOAD * f->loop;
+		}
+
+		if(r->refbehind.b[z] & bb)
+			for(f1 = f->p2; f1 != nil; f1 = f1->p2link)
+				if(((Reg*)f1->data)->refahead.b[z] & bb)
+					paint1(f1, bn);
+
+		if(!(r->refahead.b[z] & bb))
+			break;
+		f1 = f->s2;
+		if(f1 != nil)
+			if(((Reg*)f1->data)->refbehind.b[z] & bb)
+				paint1(f1, bn);
+		f = f->s1;
+		if(f == nil)
+			break;
+		r = (Reg*)f->data;
+		if(r->act.b[z] & bb)
+			break;
+		if(!(r->refbehind.b[z] & bb))
+			break;
+	}
+}
+
+static uint64
+paint2(Flow *f, int bn, int depth)
+{
+	Flow *f1;
+	Reg *r, *r1;
+	int z;
+	uint64 bb, vreg;
+
+	z = bn/64;
+	bb = 1LL << (bn%64);
+	vreg = regbits;
+	r = (Reg*)f->data;
+	if(!(r->act.b[z] & bb))
+		return vreg;
+	for(;;) {
+		if(!(r->refbehind.b[z] & bb))
+			break;
+		f1 = f->p1;
+		if(f1 == nil)
+			break;
+		r1 = (Reg*)f1->data;
+		if(!(r1->refahead.b[z] & bb))
+			break;
+		if(!(r1->act.b[z] & bb))
+			break;
+		f = f1;
+		r = r1;
+	}
+	for(;;) {
+		if(debug['R'] && debug['v'])
+			print("  paint2 %d %P\n", depth, f->prog);
+
+		r->act.b[z] &= ~bb;
+
+		vreg |= r->regu;
+
+		if(r->refbehind.b[z] & bb)
+			for(f1 = f->p2; f1 != nil; f1 = f1->p2link)
+				if(((Reg*)f1->data)->refahead.b[z] & bb)
+					vreg |= paint2(f1, bn, depth+1);
+
+		if(!(r->refahead.b[z] & bb))
+			break;
+		f1 = f->s2;
+		if(f1 != nil)
+			if(((Reg*)f1->data)->refbehind.b[z] & bb)
+				vreg |= paint2(f1, bn, depth+1);
+		f = f->s1;
+		if(f == nil)
+			break;
+		r = (Reg*)f->data;
+		if(!(r->act.b[z] & bb))
+			break;
+		if(!(r->refbehind.b[z] & bb))
+			break;
+	}
+
+	return vreg;
+}
+
+static void
+paint3(Flow *f, int bn, uint64 rb, int rn)
+{
+	Flow *f1;
+	Reg *r, *r1;
+	Prog *p;
+	int z;
+	uint64 bb;
+
+	z = bn/64;
+	bb = 1LL << (bn%64);
+	r = (Reg*)f->data;
+	if(r->act.b[z] & bb)
+		return;
+	for(;;) {
+		if(!(r->refbehind.b[z] & bb))
+			break;
+		f1 = f->p1;
+		if(f1 == nil)
+			break;
+		r1 = (Reg*)f1->data;
+		if(!(r1->refahead.b[z] & bb))
+			break;
+		if(r1->act.b[z] & bb)
+			break;
+		f = f1;
+		r = r1;
+	}
+
+	if(LOAD(r) & ~(r->set.b[z] & ~(r->use1.b[z]|r->use2.b[z])) & bb)
+		addmove(f, bn, rn, 0);
+	for(;;) {
+		r->act.b[z] |= bb;
+		p = f->prog;
+
+		if(r->use1.b[z] & bb) {
+			if(debug['R'] && debug['v'])
+				print("%P", p);
+			addreg(&p->from, rn);
+			if(debug['R'] && debug['v'])
+				print(" ===change== %P\n", p);
+		}
+		if((r->use2.b[z]|r->set.b[z]) & bb) {
+			if(debug['R'] && debug['v'])
+				print("%P", p);
+			addreg(&p->to, rn);
+			if(debug['R'] && debug['v'])
+				print(" ===change== %P\n", p);
+		}
+
+		if(STORE(r) & r->regdiff.b[z] & bb)
+			addmove(f, bn, rn, 1);
+		r->regu |= rb;
+
+		if(r->refbehind.b[z] & bb)
+			for(f1 = f->p2; f1 != nil; f1 = f1->p2link)
+				if(((Reg*)f1->data)->refahead.b[z] & bb)
+					paint3(f1, bn, rb, rn);
+
+		if(!(r->refahead.b[z] & bb))
+			break;
+		f1 = f->s2;
+		if(f1 != nil)
+			if(((Reg*)f1->data)->refbehind.b[z] & bb)
+				paint3(f1, bn, rb, rn);
+		f = f->s1;
+		if(f == nil)
+			break;
+		r = (Reg*)f->data;
+		if(r->act.b[z] & bb)
+			break;
+		if(!(r->refbehind.b[z] & bb))
+			break;
+	}
+}
+
+static void
+addreg(Adr *a, int rn)
+{
+	a->sym = nil;
+	a->node = nil;
+	a->offset = 0;
+	a->type = TYPE_REG;
+	a->reg = rn;
+	a->name = 0;
+
+	ostats.ncvtreg++;
+}
+
+void
+dumpone(Flow *f, int isreg)
+{
+	int z;
+	Bits bit;
+	Reg *r;
+
+	print("%d:%P", f->loop, f->prog);
+	if(isreg) {	
+		r = (Reg*)f->data;
+		for(z=0; z<BITS; z++)
+			bit.b[z] =
+				r->set.b[z] |
+				r->use1.b[z] |
+				r->use2.b[z] |
+				r->refbehind.b[z] |
+				r->refahead.b[z] |
+				r->calbehind.b[z] |
+				r->calahead.b[z] |
+				r->regdiff.b[z] |
+				r->act.b[z] |
+					0;
+		if(bany(&bit)) {
+			print("\t");
+			if(bany(&r->set))
+				print(" s:%Q", r->set);
+			if(bany(&r->use1))
+				print(" u1:%Q", r->use1);
+			if(bany(&r->use2))
+				print(" u2:%Q", r->use2);
+			if(bany(&r->refbehind))
+				print(" rb:%Q ", r->refbehind);
+			if(bany(&r->refahead))
+				print(" ra:%Q ", r->refahead);
+			if(bany(&r->calbehind))
+				print(" cb:%Q ", r->calbehind);
+			if(bany(&r->calahead))
+				print(" ca:%Q ", r->calahead);
+			if(bany(&r->regdiff))
+				print(" d:%Q ", r->regdiff);
+			if(bany(&r->act))
+				print(" a:%Q ", r->act);
+		}
+	}
+	print("\n");
+}
+
+void
+dumpit(char *str, Flow *r0, int isreg)
+{
+	Flow *r, *r1;
+
+	print("\n%s\n", str);
+	for(r = r0; r != nil; r = r->link) {
+		dumpone(r, isreg);
+		r1 = r->p2;
+		if(r1 != nil) {
+			print("	pred:");
+			for(; r1 != nil; r1 = r1->p2link)
+				print(" %.4ud", (int)r1->prog->pc);
+			if(r->p1 != nil)
+				print(" (and %.4ud)", (int)r->p1->prog->pc);
+			else
+				print(" (only)");
+			print("\n");
+		}
+		// Print successors if it's not just the next one
+		if(r->s1 != r->link || r->s2 != nil) {
+			print("	succ:");
+			if(r->s1 != nil)
+				print(" %.4ud", (int)r->s1->prog->pc);
+			if(r->s2 != nil)
+				print(" %.4ud", (int)r->s2->prog->pc);
+			print("\n");
+		}
+	}
+}
diff --git a/src/cmd/gc/runtime.go b/src/cmd/gc/runtime.go
index 80550f856d..0a4c1b8cbb 100644
--- a/src/cmd/gc/runtime.go
+++ b/src/cmd/gc/runtime.go
@@ -50,10 +50,10 @@ func eqstring(string, string) bool
 func intstring(*[4]byte, int64) string
 func slicebytetostring(*[32]byte, []byte) string
 func slicebytetostringtmp([]byte) string
-func slicerunetostring([]rune) string
-func stringtoslicebyte(string) []byte
+func slicerunetostring(*[32]byte, []rune) string
+func stringtoslicebyte(*[32]byte, string) []byte
 func stringtoslicebytetmp(string) []byte
-func stringtoslicerune(string) []rune
+func stringtoslicerune(*[32]rune, string) []rune
 func stringiter(string, int) int
 func stringiter2(string, int) (retk int, retv rune)
 func slicecopy(to any, fr any, wid uintptr) int
@@ -86,7 +86,7 @@ func ifacethash(i1 any) (ret uint32)
 func efacethash(i1 any) (ret uint32)
 
 // *byte is really *runtime.Type
-func makemap(mapType *byte, hint int64) (hmap map[any]any)
+func makemap(mapType *byte, hint int64, mapbuf *any, bucketbuf *any) (hmap map[any]any)
 func mapaccess1(mapType *byte, hmap map[any]any, key *any) (val *any)
 func mapaccess1_fast32(mapType *byte, hmap map[any]any, key any) (val *any)
 func mapaccess1_fast64(mapType *byte, hmap map[any]any, key any) (val *any)
diff --git a/src/cmd/gc/select.c b/src/cmd/gc/select.c
index 5d3b71164a..537d0ca928 100644
--- a/src/cmd/gc/select.c
+++ b/src/cmd/gc/select.c
@@ -133,18 +133,16 @@ walkselect(Node *sel)
 				break;
 
 			case OSELRECV:
-				ch = n->right->left;
-			Selrecv1:
-				if(n->left == N)
-					n = n->right;
-				else
-					n->op = OAS;
-				break;
-			
 			case OSELRECV2:
 				ch = n->right->left;
-				if(n->ntest == N)
-					goto Selrecv1;
+				if(n->op == OSELRECV || n->ntest == N) {
+					if(n->left == N)
+						n = n->right;
+					else
+						n->op = OAS;
+					break;
+				}
+			
 				if(n->left == N) {
 					typecheck(&nblank, Erv | Easgn);
 					n->left = nblank;
diff --git a/src/cmd/gc/sinit.c b/src/cmd/gc/sinit.c
index 3d303eda09..1015950e41 100644
--- a/src/cmd/gc/sinit.c
+++ b/src/cmd/gc/sinit.c
@@ -302,13 +302,13 @@ staticcopy(Node *l, Node *r, NodeList **out)
 	case OLITERAL:
 		if(iszero(r))
 			return 1;
-		arch.gdata(l, r, l->type->width);
+		gdata(l, r, l->type->width);
 		return 1;
 
 	case OADDR:
 		switch(r->left->op) {
 		case ONAME:
-			arch.gdata(l, r, l->type->width);
+			gdata(l, r, l->type->width);
 			return 1;
 		}
 		break;
@@ -322,7 +322,7 @@ staticcopy(Node *l, Node *r, NodeList **out)
 		case OSTRUCTLIT:
 		case OMAPLIT:
 			// copy pointer
-			arch.gdata(l, nod(OADDR, r->nname, N), l->type->width);
+			gdata(l, nod(OADDR, r->nname, N), l->type->width);
 			return 1;
 		}
 		break;
@@ -333,11 +333,11 @@ staticcopy(Node *l, Node *r, NodeList **out)
 			a = r->nname;
 			n1 = *l;
 			n1.xoffset = l->xoffset + Array_array;
-			arch.gdata(&n1, nod(OADDR, a, N), widthptr);
+			gdata(&n1, nod(OADDR, a, N), widthptr);
 			n1.xoffset = l->xoffset + Array_nel;
-			arch.gdata(&n1, r->right, widthint);
+			gdata(&n1, r->right, widthint);
 			n1.xoffset = l->xoffset + Array_cap;
-			arch.gdata(&n1, r->right, widthint);
+			gdata(&n1, r->right, widthint);
 			return 1;
 		}
 		// fall through
@@ -349,7 +349,7 @@ staticcopy(Node *l, Node *r, NodeList **out)
 			n1.xoffset = l->xoffset + e->xoffset;
 			n1.type = e->expr->type;
 			if(e->expr->op == OLITERAL)
-				arch.gdata(&n1, e->expr, n1.type->width);
+				gdata(&n1, e->expr, n1.type->width);
 			else {
 				ll = nod(OXXX, N, N);
 				*ll = n1;
@@ -394,14 +394,14 @@ staticassign(Node *l, Node *r, NodeList **out)
 	case OLITERAL:
 		if(iszero(r))
 			return 1;
-		arch.gdata(l, r, l->type->width);
+		gdata(l, r, l->type->width);
 		return 1;
 
 	case OADDR:
 		if(stataddr(&nam, r->left)) {
 			n1 = *r;
 			n1.left = &nam;
-			arch.gdata(l, &n1, l->type->width);
+			gdata(l, &n1, l->type->width);
 			return 1;
 		}
 	
@@ -417,7 +417,7 @@ staticassign(Node *l, Node *r, NodeList **out)
 			// Init pointer.
 			a = staticname(r->left->type, 1);
 			r->nname = a;
-			arch.gdata(l, nod(OADDR, a, N), l->type->width);
+			gdata(l, nod(OADDR, a, N), l->type->width);
 			// Init underlying literal.
 			if(!staticassign(a, r->left, out))
 				*out = list(*out, nod(OAS, a, r->left));
@@ -444,11 +444,11 @@ staticassign(Node *l, Node *r, NodeList **out)
 			r->nname = a;
 			n1 = *l;
 			n1.xoffset = l->xoffset + Array_array;
-			arch.gdata(&n1, nod(OADDR, a, N), widthptr);
+			gdata(&n1, nod(OADDR, a, N), widthptr);
 			n1.xoffset = l->xoffset + Array_nel;
-			arch.gdata(&n1, r->right, widthint);
+			gdata(&n1, r->right, widthint);
 			n1.xoffset = l->xoffset + Array_cap;
-			arch.gdata(&n1, r->right, widthint);
+			gdata(&n1, r->right, widthint);
 			// Fall through to init underlying array.
 			l = a;
 		}
@@ -462,7 +462,7 @@ staticassign(Node *l, Node *r, NodeList **out)
 			n1.xoffset = l->xoffset + e->xoffset;
 			n1.type = e->expr->type;
 			if(e->expr->op == OLITERAL)
-				arch.gdata(&n1, e->expr, n1.type->width);
+				gdata(&n1, e->expr, n1.type->width);
 			else {
 				a = nod(OXXX, N, N);
 				*a = n1;
@@ -1223,7 +1223,7 @@ stataddr(Node *nam, Node *n)
 		if(l < 0)
 			break;
 		// Check for overflow.
-		if(n->type->width != 0 && arch.MAXWIDTH/n->type->width <= l)
+		if(n->type->width != 0 && thearch.MAXWIDTH/n->type->width <= l)
 			break;
  		nam->xoffset += l*n->type->width;
 		nam->type = n->type;
@@ -1303,16 +1303,16 @@ gen_as_init(Node *n)
 	case TPTR64:
 	case TFLOAT32:
 	case TFLOAT64:
-		arch.gdata(&nam, nr, nr->type->width);
+		gdata(&nam, nr, nr->type->width);
 		break;
 
 	case TCOMPLEX64:
 	case TCOMPLEX128:
-		arch.gdatacomplex(&nam, nr->val.u.cval);
+		gdatacomplex(&nam, nr->val.u.cval);
 		break;
 
 	case TSTRING:
-		arch.gdatastring(&nam, nr->val.u.sval);
+		gdatastring(&nam, nr->val.u.sval);
 		break;
 	}
 
@@ -1320,7 +1320,7 @@ yes:
 	return 1;
 
 slice:
-	arch.gused(N); // in case the data is the dest of a goto
+	gused(N); // in case the data is the dest of a goto
 	nl = nr;
 	if(nr == N || nr->op != OADDR)
 		goto no;
@@ -1333,14 +1333,14 @@ slice:
 		goto no;
 
 	nam.xoffset += Array_array;
-	arch.gdata(&nam, nl, types[tptr]->width);
+	gdata(&nam, nl, types[tptr]->width);
 
 	nam.xoffset += Array_nel-Array_array;
 	nodconst(&nod1, types[TINT], nr->type->bound);
-	arch.gdata(&nam, &nod1, widthint);
+	gdata(&nam, &nod1, widthint);
 
 	nam.xoffset += Array_cap-Array_nel;
-	arch.gdata(&nam, &nod1, widthint);
+	gdata(&nam, &nod1, widthint);
 
 	goto yes;
 
diff --git a/src/cmd/gc/subr.c b/src/cmd/gc/subr.c
index 3ed194ee8f..f739a72499 100644
--- a/src/cmd/gc/subr.c
+++ b/src/cmd/gc/subr.c
@@ -38,6 +38,19 @@ parserline(void)
 	return lineno;
 }
 
+void
+adderrorname(Node *n)
+{
+	char *old;
+	
+	if(n->op != ODOT)
+		return;
+	old = smprint("%L: undefined: %N\n", n->lineno, n->left);
+	if(nerr > 0 && err[nerr-1].lineno == n->lineno && strcmp(err[nerr-1].msg, old) == 0)
+		err[nerr-1].msg = smprint("%L: undefined: %N in %N\n", n->lineno, n->left, n);
+	free(old);
+}
+
 static void
 adderr(int line, char *fmt, va_list arg)
 {
@@ -1211,7 +1224,7 @@ assignop(Type *src, Type *dst, char **why)
 		*why = "";
 
 	// TODO(rsc,lvd): This behaves poorly in the presence of inlining.
-	// https://code.google.com/p/go/issues/detail?id=2795
+	// https://golang.org/issue/2795
 	if(safemode && importpkg == nil && src != T && src->etype == TUNSAFEPTR) {
 		yyerror("cannot use unsafe.Pointer");
 		errorexit();
@@ -1703,36 +1716,31 @@ ptrto(Type *t)
 void
 frame(int context)
 {
-	char *p;
 	NodeList *l;
 	Node *n;
-	int flag;
+	vlong w;
 
-	p = "stack";
-	l = nil;
-	if(curfn)
-		l = curfn->dcl;
 	if(context) {
-		p = "external";
+		print("--- external frame ---\n");
 		l = externdcl;
-	}
+	} else if(curfn) {
+		print("--- %S frame ---\n", curfn->nname->sym);
+		l = curfn->dcl;
+	} else
+		return;
 
-	flag = 1;
 	for(; l; l=l->next) {
 		n = l->n;
+		w = -1;
+		if(n->type)
+			w = n->type->width;
 		switch(n->op) {
 		case ONAME:
-			if(flag)
-				print("--- %s frame ---\n", p);
-			print("%O %S G%d %T\n", n->op, n->sym, n->vargen, n->type);
-			flag = 0;
+			print("%O %S G%d %T width=%lld\n", n->op, n->sym, n->vargen, n->type, w);
 			break;
 
 		case OTYPE:
-			if(flag)
-				print("--- %s frame ---\n", p);
-			print("%O %T\n", n->op, n->type);
-			flag = 0;
+			print("%O %T width=%lld\n", n->op, n->type, w);
 			break;
 		}
 	}
@@ -2121,10 +2129,10 @@ setmaxarg(Type *t, int32 extra)
 
 	dowidth(t);
 	w = t->argwid;
-	if(w >= arch.MAXWIDTH)
+	if(w >= thearch.MAXWIDTH)
 		fatal("bad argwid %T", t);
 	w += extra;
-	if(w >= arch.MAXWIDTH)
+	if(w >= thearch.MAXWIDTH)
 		fatal("bad argwid %d + %T", extra, t);
 	if(w > maxarg)
 		maxarg = w;
@@ -2628,7 +2636,7 @@ genwrapper(Type *rcvr, Type *method, Sym *newnam, int iface)
 	inl_nonlocal = 0;
 
 	curfn = nil;
-	funccompile(fn, 0);
+	funccompile(fn);
 }
 
 static Node*
@@ -2876,7 +2884,7 @@ genhash(Sym *sym, Type *t)
 	// an unexported field of type unsafe.Pointer.
 	old_safemode = safemode;
 	safemode = 0;
-	funccompile(fn, 0);
+	funccompile(fn);
 	safemode = old_safemode;
 }
 
@@ -3096,7 +3104,7 @@ geneq(Sym *sym, Type *t)
 	// an unexported field of type unsafe.Pointer.
 	old_safemode = safemode;
 	safemode = 0;
-	funccompile(fn, 0);
+	funccompile(fn);
 	safemode = old_safemode;
 }
 
diff --git a/src/cmd/gc/swt.c b/src/cmd/gc/swt.c
index ca5455d479..0dc0065ed9 100644
--- a/src/cmd/gc/swt.c
+++ b/src/cmd/gc/swt.c
@@ -454,31 +454,22 @@ exprbsw(Case *c0, int ncase, int arg)
 			n = c0->node;
 			lno = setlineno(n);
 
-			if(assignop(n->left->type, exprname->type, nil) == OCONVIFACE ||
-			   assignop(exprname->type, n->left->type, nil) == OCONVIFACE)
-				goto snorm;
-
-			switch(arg) {
-			case Strue:
+			if((arg != Strue && arg != Sfalse) ||
+			   assignop(n->left->type, exprname->type, nil) == OCONVIFACE ||
+			   assignop(exprname->type, n->left->type, nil) == OCONVIFACE) {
 				a = nod(OIF, N, N);
-				a->ntest = n->left;			// if val
+				a->ntest = nod(OEQ, exprname, n->left);	// if name == val
+				typecheck(&a->ntest, Erv);
 				a->nbody = list1(n->right);			// then goto l
-				break;
-
-			case Sfalse:
+			} else if(arg == Strue) {
 				a = nod(OIF, N, N);
-				a->ntest = nod(ONOT, n->left, N);	// if !val
-				typecheck(&a->ntest, Erv);
+				a->ntest = n->left;			// if val
 				a->nbody = list1(n->right);			// then goto l
-				break;
-
-			default:
-			snorm:
+			} else { // arg == Sfalse
 				a = nod(OIF, N, N);
-				a->ntest = nod(OEQ, exprname, n->left);	// if name == val
+				a->ntest = nod(ONOT, n->left, N);	// if !val
 				typecheck(&a->ntest, Erv);
 				a->nbody = list1(n->right);			// then goto l
-				break;
 			}
 
 			cas = list(cas, a);
@@ -503,7 +494,7 @@ exprbsw(Case *c0, int ncase, int arg)
 
 /*
  * normal (expression) switch.
- * rebulid case statements into if .. goto
+ * rebuild case statements into if .. goto
  */
 static void
 exprswitch(Node *sw)
@@ -533,12 +524,15 @@ exprswitch(Node *sw)
 	 */
 	exprname = N;
 	cas = nil;
-	if(arg != Strue && arg != Sfalse) {
+	if(arg == Strue || arg == Sfalse)
+		exprname = nodbool(arg == Strue);
+	else if(consttype(sw->ntest) >= 0)
+		// leave constants to enable dead code elimination (issue 9608)
+		exprname = sw->ntest;
+	else {
 		exprname = temp(sw->ntest->type);
 		cas = list1(nod(OAS, exprname, sw->ntest));
 		typechecklist(cas, Etop);
-	} else {
-		exprname = nodbool(arg == Strue);
 	}
 
 	c0 = mkcaselist(sw, arg);
diff --git a/src/cmd/gc/typecheck.c b/src/cmd/gc/typecheck.c
index 635d2c4170..649f1c5120 100644
--- a/src/cmd/gc/typecheck.c
+++ b/src/cmd/gc/typecheck.c
@@ -555,168 +555,7 @@ reswitch:
 		if(l->type == T || r->type == T)
 			goto error;
 		op = n->op;
-	arith:
-		if(op == OLSH || op == ORSH)
-			goto shift;
-		// ideal mixed with non-ideal
-		defaultlit2(&l, &r, 0);
-		n->left = l;
-		n->right = r;
-		if(l->type == T || r->type == T)
-			goto error;
-		t = l->type;
-		if(t->etype == TIDEAL)
-			t = r->type;
-		et = t->etype;
-		if(et == TIDEAL)
-			et = TINT;
-		if(iscmp[n->op] && t->etype != TIDEAL && !eqtype(l->type, r->type)) {
-			// comparison is okay as long as one side is
-			// assignable to the other.  convert so they have
-			// the same type.
-			//
-			// the only conversion that isn't a no-op is concrete == interface.
-			// in that case, check comparability of the concrete type.
-			if(r->type->etype != TBLANK && (aop = assignop(l->type, r->type, nil)) != 0) {
-				if(isinter(r->type) && !isinter(l->type) && algtype1(l->type, nil) == ANOEQ) {
-					yyerror("invalid operation: %N (operator %O not defined on %s)", n, op, typekind(l->type));
-					goto error;
-				}
-				l = nod(aop, l, N);
-				l->type = r->type;
-				l->typecheck = 1;
-				n->left = l;
-				t = l->type;
-				goto converted;
-			}
-			if(l->type->etype != TBLANK && (aop = assignop(r->type, l->type, nil)) != 0) {
-				if(isinter(l->type) && !isinter(r->type) && algtype1(r->type, nil) == ANOEQ) {
-					yyerror("invalid operation: %N (operator %O not defined on %s)", n, op, typekind(r->type));
-					goto error;
-				}
-				r = nod(aop, r, N);
-				r->type = l->type;
-				r->typecheck = 1;
-				n->right = r;
-				t = r->type;
-			}
-		converted:
-			et = t->etype;
-		}
-		if(t->etype != TIDEAL && !eqtype(l->type, r->type)) {
-			defaultlit2(&l, &r, 1);
-			if(n->op == OASOP && n->implicit) {
-				yyerror("invalid operation: %N (non-numeric type %T)", n, l->type);
-				goto error;
-			}
-			yyerror("invalid operation: %N (mismatched types %T and %T)", n, l->type, r->type);
-			goto error;
-		}
-		if(!okfor[op][et]) {
-			yyerror("invalid operation: %N (operator %O not defined on %s)", n, op, typekind(t));
-			goto error;
-		}
-		// okfor allows any array == array, map == map, func == func.
-		// restrict to slice/map/func == nil and nil == slice/map/func.
-		if(isfixedarray(l->type) && algtype1(l->type, nil) == ANOEQ) {
-			yyerror("invalid operation: %N (%T cannot be compared)", n, l->type);
-			goto error;
-		}
-		if(isslice(l->type) && !isnil(l) && !isnil(r)) {
-			yyerror("invalid operation: %N (slice can only be compared to nil)", n);
-			goto error;
-		}
-		if(l->type->etype == TMAP && !isnil(l) && !isnil(r)) {
-			yyerror("invalid operation: %N (map can only be compared to nil)", n);
-			goto error;
-		}
-		if(l->type->etype == TFUNC && !isnil(l) && !isnil(r)) {
-			yyerror("invalid operation: %N (func can only be compared to nil)", n);
-			goto error;
-		}
-		if(l->type->etype == TSTRUCT && algtype1(l->type, &badtype) == ANOEQ) {
-			yyerror("invalid operation: %N (struct containing %T cannot be compared)", n, badtype);
-			goto error;
-		}
-		
-		t = l->type;
-		if(iscmp[n->op]) {
-			evconst(n);
-			t = idealbool;
-			if(n->op != OLITERAL) {
-				defaultlit2(&l, &r, 1);
-				n->left = l;
-				n->right = r;
-			}
-		} else if(n->op == OANDAND || n->op == OOROR) {
-			if(l->type == r->type)
-				t = l->type;
-			else if(l->type == idealbool)
-				t = r->type;
-			else if(r->type == idealbool)
-				t = l->type;
-		// non-comparison operators on ideal bools should make them lose their ideal-ness
-		} else if(t == idealbool)
-			t = types[TBOOL];
-
-		if(et == TSTRING) {
-			if(iscmp[n->op]) {
-				n->etype = n->op;
-				n->op = OCMPSTR;
-			} else if(n->op == OADD) {
-				// create OADDSTR node with list of strings in x + y + z + (w + v) + ...
-				n->op = OADDSTR;
-				if(l->op == OADDSTR)
-					n->list = l->list;
-				else
-					n->list = list1(l);
-				if(r->op == OADDSTR)
-					n->list = concat(n->list, r->list);
-				else
-					n->list = list(n->list, r);
-				n->left = N;
-				n->right = N;
-			}
-		}
-		if(et == TINTER) {
-			if(l->op == OLITERAL && l->val.ctype == CTNIL) {
-				// swap for back end
-				n->left = r;
-				n->right = l;
-			} else if(r->op == OLITERAL && r->val.ctype == CTNIL) {
-				// leave alone for back end
-			} else {
-				n->etype = n->op;
-				n->op = OCMPIFACE;
-			}
-		}
-
-		if((op == ODIV || op == OMOD) && isconst(r, CTINT))
-		if(mpcmpfixc(r->val.u.xval, 0) == 0) {
-			yyerror("division by zero");
-			goto error;
-		} 
-
-		n->type = t;
-		goto ret;
-
-	shift:
-		defaultlit(&r, types[TUINT]);
-		n->right = r;
-		t = r->type;
-		if(!isint[t->etype] || issigned[t->etype]) {
-			yyerror("invalid operation: %N (shift count type %T, must be unsigned integer)", n, r->type);
-			goto error;
-		}
-		t = l->type;
-		if(t != T && t->etype != TIDEAL && !isint[t->etype]) {
-			yyerror("invalid operation: %N (shift of type %T)", n, t);
-			goto error;
-		}
-		// no defaultlit for left
-		// the outer context gives the type
-		n->type = l->type;
-		goto ret;
+		goto arith;
 
 	case OCOM:
 	case OMINUS:
@@ -776,12 +615,14 @@ reswitch:
 	case ODOT:
 		typecheck(&n->left, Erv|Etype);
 		defaultlit(&n->left, T);
-		if((t = n->left->type) == T)
-			goto error;
 		if(n->right->op != ONAME) {
 			yyerror("rhs of . must be a name");	// impossible
 			goto error;
 		}
+		if((t = n->left->type) == T) {
+			adderrorname(n);
+			goto error;
+		}
 		r = n->right;
 
 		if(n->left->op == OTYPE) {
@@ -1442,41 +1283,7 @@ reswitch:
 		goto ret;
 
 	case OCONV:
-	doconv:
-		ok |= Erv;
-		saveorignode(n);
-		typecheck(&n->left, Erv | (top & (Eindir | Eiota)));
-		convlit1(&n->left, n->type, 1);
-		if((t = n->left->type) == T || n->type == T)
-			goto error;
-		if((n->op = convertop(t, n->type, &why)) == 0) {
-			if(!n->diag && !n->type->broke) {
-				yyerror("cannot convert %lN to type %T%s", n->left, n->type, why);
-				n->diag = 1;
-			}
-			n->op = OCONV;
-		}
-		switch(n->op) {
-		case OCONVNOP:
-			if(n->left->op == OLITERAL && n->type != types[TBOOL]) {
-				r = nod(OXXX, N, N);
-				n->op = OCONV;
-				n->orig = r;
-				*r = *n;
-				n->op = OLITERAL;
-				n->val = n->left->val;
-			}
-			break;
-		case OSTRARRAYBYTE:
-			// do not use stringtoarraylit.
-			// generated code and compiler memory footprint is better without it.
-			break;
-		case OSTRARRAYRUNE:
-			if(n->left->op == OLITERAL)
-				stringtoarraylit(&n);
-			break;
-		}
-		goto ret;
+		goto doconv;
 
 	case OMAKE:
 		ok |= Erv;
@@ -1494,13 +1301,14 @@ reswitch:
 
 		switch(t->etype) {
 		default:
-		badmake:
 			yyerror("cannot make type %T", t);
 			goto error;
 
 		case TARRAY:
-			if(!isslice(t))
-				goto badmake;
+			if(!isslice(t)) {
+				yyerror("cannot make type %T", t);
+				goto error;
+			}
 			if(args == nil) {
 				yyerror("missing len argument to make(%T)", t);
 				goto error;
@@ -1793,6 +1601,216 @@ reswitch:
 			checkwidth(n->left->type);
 		goto ret;
 	}
+	goto ret;
+
+arith:
+	if(op == OLSH || op == ORSH)
+		goto shift;
+	// ideal mixed with non-ideal
+	defaultlit2(&l, &r, 0);
+	n->left = l;
+	n->right = r;
+	if(l->type == T || r->type == T)
+		goto error;
+	t = l->type;
+	if(t->etype == TIDEAL)
+		t = r->type;
+	et = t->etype;
+	if(et == TIDEAL)
+		et = TINT;
+	aop = 0;
+	if(iscmp[n->op] && t->etype != TIDEAL && !eqtype(l->type, r->type)) {
+		// comparison is okay as long as one side is
+		// assignable to the other.  convert so they have
+		// the same type.
+		//
+		// the only conversion that isn't a no-op is concrete == interface.
+		// in that case, check comparability of the concrete type.
+		// The conversion allocates, so only do it if the concrete type is huge.
+		if(r->type->etype != TBLANK && (aop = assignop(l->type, r->type, nil)) != 0) {
+			if(isinter(r->type) && !isinter(l->type) && algtype1(l->type, nil) == ANOEQ) {
+				yyerror("invalid operation: %N (operator %O not defined on %s)", n, op, typekind(l->type));
+				goto error;
+			}
+			dowidth(l->type);
+			if(isinter(r->type) == isinter(l->type) || l->type->width >= 1<<16) {
+				l = nod(aop, l, N);
+				l->type = r->type;
+				l->typecheck = 1;
+				n->left = l;
+			}
+			t = r->type;
+			goto converted;
+		}
+		if(l->type->etype != TBLANK && (aop = assignop(r->type, l->type, nil)) != 0) {
+			if(isinter(l->type) && !isinter(r->type) && algtype1(r->type, nil) == ANOEQ) {
+				yyerror("invalid operation: %N (operator %O not defined on %s)", n, op, typekind(r->type));
+				goto error;
+			}
+			dowidth(r->type);
+			if(isinter(r->type) == isinter(l->type) || r->type->width >= 1<<16) {
+				r = nod(aop, r, N);
+				r->type = l->type;
+				r->typecheck = 1;
+				n->right = r;
+			}
+			t = l->type;
+		}
+	converted:
+		et = t->etype;
+	}
+	if(t->etype != TIDEAL && !eqtype(l->type, r->type)) {
+		defaultlit2(&l, &r, 1);
+		if(n->op == OASOP && n->implicit) {
+			yyerror("invalid operation: %N (non-numeric type %T)", n, l->type);
+			goto error;
+		}
+		if(isinter(r->type) == isinter(l->type) || aop == 0) {
+			yyerror("invalid operation: %N (mismatched types %T and %T)", n, l->type, r->type);
+			goto error;
+		}
+	}
+	if(!okfor[op][et]) {
+		yyerror("invalid operation: %N (operator %O not defined on %s)", n, op, typekind(t));
+		goto error;
+	}
+	// okfor allows any array == array, map == map, func == func.
+	// restrict to slice/map/func == nil and nil == slice/map/func.
+	if(isfixedarray(l->type) && algtype1(l->type, nil) == ANOEQ) {
+		yyerror("invalid operation: %N (%T cannot be compared)", n, l->type);
+		goto error;
+	}
+	if(isslice(l->type) && !isnil(l) && !isnil(r)) {
+		yyerror("invalid operation: %N (slice can only be compared to nil)", n);
+		goto error;
+	}
+	if(l->type->etype == TMAP && !isnil(l) && !isnil(r)) {
+		yyerror("invalid operation: %N (map can only be compared to nil)", n);
+		goto error;
+	}
+	if(l->type->etype == TFUNC && !isnil(l) && !isnil(r)) {
+		yyerror("invalid operation: %N (func can only be compared to nil)", n);
+		goto error;
+	}
+	if(l->type->etype == TSTRUCT && algtype1(l->type, &badtype) == ANOEQ) {
+		yyerror("invalid operation: %N (struct containing %T cannot be compared)", n, badtype);
+		goto error;
+	}
+	
+	t = l->type;
+	if(iscmp[n->op]) {
+		evconst(n);
+		t = idealbool;
+		if(n->op != OLITERAL) {
+			defaultlit2(&l, &r, 1);
+			n->left = l;
+			n->right = r;
+		}
+	} else if(n->op == OANDAND || n->op == OOROR) {
+		if(l->type == r->type)
+			t = l->type;
+		else if(l->type == idealbool)
+			t = r->type;
+		else if(r->type == idealbool)
+			t = l->type;
+	// non-comparison operators on ideal bools should make them lose their ideal-ness
+	} else if(t == idealbool)
+		t = types[TBOOL];
+
+	if(et == TSTRING) {
+		if(iscmp[n->op]) {
+			n->etype = n->op;
+			n->op = OCMPSTR;
+		} else if(n->op == OADD) {
+			// create OADDSTR node with list of strings in x + y + z + (w + v) + ...
+			n->op = OADDSTR;
+			if(l->op == OADDSTR)
+				n->list = l->list;
+			else
+				n->list = list1(l);
+			if(r->op == OADDSTR)
+				n->list = concat(n->list, r->list);
+			else
+				n->list = list(n->list, r);
+			n->left = N;
+			n->right = N;
+		}
+	}
+	if(et == TINTER) {
+		if(l->op == OLITERAL && l->val.ctype == CTNIL) {
+			// swap for back end
+			n->left = r;
+			n->right = l;
+		} else if(r->op == OLITERAL && r->val.ctype == CTNIL) {
+			// leave alone for back end
+		} else if(isinter(r->type) == isinter(l->type)) {
+			n->etype = n->op;
+			n->op = OCMPIFACE;
+		}
+	}
+
+	if((op == ODIV || op == OMOD) && isconst(r, CTINT))
+	if(mpcmpfixc(r->val.u.xval, 0) == 0) {
+		yyerror("division by zero");
+		goto error;
+	} 
+
+	n->type = t;
+	goto ret;
+
+shift:
+	defaultlit(&r, types[TUINT]);
+	n->right = r;
+	t = r->type;
+	if(!isint[t->etype] || issigned[t->etype]) {
+		yyerror("invalid operation: %N (shift count type %T, must be unsigned integer)", n, r->type);
+		goto error;
+	}
+	t = l->type;
+	if(t != T && t->etype != TIDEAL && !isint[t->etype]) {
+		yyerror("invalid operation: %N (shift of type %T)", n, t);
+		goto error;
+	}
+	// no defaultlit for left
+	// the outer context gives the type
+	n->type = l->type;
+	goto ret;
+
+doconv:
+	ok |= Erv;
+	saveorignode(n);
+	typecheck(&n->left, Erv | (top & (Eindir | Eiota)));
+	convlit1(&n->left, n->type, 1);
+	if((t = n->left->type) == T || n->type == T)
+		goto error;
+	if((n->op = convertop(t, n->type, &why)) == 0) {
+		if(!n->diag && !n->type->broke) {
+			yyerror("cannot convert %lN to type %T%s", n->left, n->type, why);
+			n->diag = 1;
+		}
+		n->op = OCONV;
+	}
+	switch(n->op) {
+	case OCONVNOP:
+		if(n->left->op == OLITERAL && n->type != types[TBOOL]) {
+			r = nod(OXXX, N, N);
+			n->op = OCONV;
+			n->orig = r;
+			*r = *n;
+			n->op = OLITERAL;
+			n->val = n->left->val;
+		}
+		break;
+	case OSTRARRAYBYTE:
+		// do not use stringtoarraylit.
+		// generated code and compiler memory footprint is better without it.
+		break;
+	case OSTRARRAYRUNE:
+		if(n->left->op == OLITERAL)
+			stringtoarraylit(&n);
+		break;
+	}
+	goto ret;
 
 ret:
 	t = n->type;
@@ -1913,7 +1931,7 @@ checkdefergo(Node *n)
 	case OPRINTN:
 	case ORECOVER:
 		// ok
-		break;
+		return;
 	case OAPPEND:
 	case OCAP:
 	case OCOMPLEX:
@@ -1927,23 +1945,21 @@ checkdefergo(Node *n)
 	case OREAL:
 	case OLITERAL: // conversion or unsafe.Alignof, Offsetof, Sizeof
 		if(n->left->orig != N && n->left->orig->op == OCONV)
-			goto conv;
-		yyerror("%s discards result of %N", what, n->left);
-		break;
-	default:
-	conv:
-		// type is broken or missing, most likely a method call on a broken type
-		// we will warn about the broken type elsewhere. no need to emit a potentially confusing error
-		if(n->left->type == T || n->left->type->broke)
 			break;
+		yyerror("%s discards result of %N", what, n->left);
+		return;
+	}
 
-		if(!n->diag) {
-			// The syntax made sure it was a call, so this must be
-			// a conversion.
-			n->diag = 1;
-			yyerror("%s requires function call, not conversion", what);
-		}
-		break;
+	// type is broken or missing, most likely a method call on a broken type
+	// we will warn about the broken type elsewhere. no need to emit a potentially confusing error
+	if(n->left->type == T || n->left->type->broke)
+		return;
+
+	if(!n->diag) {
+		// The syntax made sure it was a call, so this must be
+		// a conversion.
+		n->diag = 1;
+		yyerror("%s requires function call, not conversion", what);
 	}
 }
 
@@ -3036,15 +3052,20 @@ typecheckas2(Node *n)
 			goto out;
 		switch(r->op) {
 		case OINDEXMAP:
-			n->op = OAS2MAPR;
-			goto common;
 		case ORECV:
-			n->op = OAS2RECV;
-			goto common;
 		case ODOTTYPE:
-			n->op = OAS2DOTTYPE;
-			r->op = ODOTTYPE2;
-		common:
+			switch(r->op) {
+			case OINDEXMAP:
+				n->op = OAS2MAPR;
+				break;
+			case ORECV:
+				n->op = OAS2RECV;
+				break;
+			case ODOTTYPE:
+				n->op = OAS2DOTTYPE;
+				r->op = ODOTTYPE2;
+				break;
+			}
 			if(l->type != T)
 				checkassignto(r->type, l);
 			if(l->defn == n)
@@ -3293,6 +3314,8 @@ typecheckdef(Node *n)
 			n->diag = 1;
 			if(n->lineno != 0)
 				lineno = n->lineno;
+			// Note: adderrorname looks for this string and
+			// adds context about the outer expression
 			yyerror("undefined: %S", n->sym);
 		}
 		return n;
diff --git a/src/cmd/gc/walk.c b/src/cmd/gc/walk.c
index efb283a1b8..0e69f88b66 100644
--- a/src/cmd/gc/walk.c
+++ b/src/cmd/gc/walk.c
@@ -628,6 +628,26 @@ walkexpr(Node **np, NodeList **init)
 		goto ret;
 
 	case OCALLFUNC:
+		if(n->left->op == OCLOSURE) {
+			// Transform direct call of a closure to call of a normal function.
+			// transformclosure already did all preparation work.
+
+			// Append captured variables to argument list.
+			n->list = concat(n->list, n->left->enter);
+			n->left->enter = NULL;
+			// Replace OCLOSURE with ONAME/PFUNC.
+			n->left = n->left->closure->nname;
+			// Update type of OCALLFUNC node.
+			// Output arguments had not changed, but their offsets could.
+			if(n->left->type->outtuple == 1) {
+				t = getoutargx(n->left->type)->type;
+				if(t->etype == TFIELD)
+					t = t->type;
+				n->type = t;
+			} else
+				n->type = getoutargx(n->left->type);
+		}
+
 		t = n->left->type;
 		if(n->list && n->list->n->op == OAS)
 			goto ret;
@@ -927,7 +947,7 @@ walkexpr(Node **np, NodeList **init)
 				l->class = PEXTERN;
 				l->xoffset = 0;
 				sym->def = l;
-				arch.ggloblsym(sym, widthptr, DUPOK|NOPTR);
+				ggloblsym(sym, widthptr, DUPOK|NOPTR);
 			}
 			l = nod(OADDR, sym->def, N);
 			l->addable = 1;
@@ -995,7 +1015,7 @@ walkexpr(Node **np, NodeList **init)
 
 	case OCONV:
 	case OCONVNOP:
-		if(arch.thechar == '5') {
+		if(thearch.thechar == '5') {
 			if(isfloat[n->left->type->etype]) {
 				if(n->type->etype == TINT64) {
 					n = mkcall("float64toint64", n->type, init, conv(n->left, types[TFLOAT64]));
@@ -1273,6 +1293,7 @@ walkexpr(Node **np, NodeList **init)
 				conv(n->right, types[TSTRING]));
 
 			// quick check of len before full compare for == or !=
+			// eqstring assumes that the lengths are equal
 			if(n->etype == OEQ) {
 				// len(left) == len(right) && eqstring(left, right)
 				r = nod(OANDAND, nod(OEQ, nod(OLEN, n->left, N), nod(OLEN, n->right, N)), r);
@@ -1329,12 +1350,32 @@ walkexpr(Node **np, NodeList **init)
 		t = n->type;
 
 		fn = syslook("makemap", 1);
-		argtype(fn, t->down);	// any-1
-		argtype(fn, t->type);	// any-2
 
-		n = mkcall1(fn, n->type, init,
-			typename(n->type),
-			conv(n->left, types[TINT64]));
+		a = nodnil(); // hmap buffer
+		r = nodnil(); // bucket buffer
+		if(n->esc == EscNone) {
+			// Allocate hmap buffer on stack.
+			var = temp(hmap(t));
+			a = nod(OAS, var, N); // zero temp
+			typecheck(&a, Etop);
+			*init = list(*init, a);
+			a = nod(OADDR, var, N);
+
+			// Allocate one bucket on stack.
+			// Maximum key/value size is 128 bytes, larger objects
+			// are stored with an indirection. So max bucket size is 2048+eps.
+			var = temp(mapbucket(t));
+			r = nod(OAS, var, N); // zero temp
+			typecheck(&r, Etop);
+			*init = list(*init, r);
+			r = nod(OADDR, var, N);
+		}
+
+		argtype(fn, hmap(t));	// hmap buffer
+		argtype(fn, mapbucket(t));	// bucket buffer
+		argtype(fn, t->down);	// key type
+		argtype(fn, t->type);	// value type
+		n = mkcall1(fn, n->type, init, typename(n->type), conv(n->left, types[TINT64]), a, r);
 		goto ret;
 
 	case OMAKESLICE:
@@ -1397,13 +1438,25 @@ walkexpr(Node **np, NodeList **init)
 		goto ret;
 
 	case OARRAYRUNESTR:
-		// slicerunetostring([]rune) string;
-		n = mkcall("slicerunetostring", n->type, init, n->left);
+		// slicerunetostring(*[32]byte, []rune) string;
+		a = nodnil();
+		if(n->esc == EscNone) {
+			// Create temporary buffer for string on stack.
+			t = aindex(nodintconst(tmpstringbufsize), types[TUINT8]);
+			a = nod(OADDR, temp(t), N);
+		}
+		n = mkcall("slicerunetostring", n->type, init, a, n->left);
 		goto ret;
 
 	case OSTRARRAYBYTE:
-		// stringtoslicebyte(string) []byte;
-		n = mkcall("stringtoslicebyte", n->type, init, conv(n->left, types[TSTRING]));
+		// stringtoslicebyte(*32[byte], string) []byte;
+		a = nodnil();
+		if(n->esc == EscNone) {
+			// Create temporary buffer for slice on stack.
+			t = aindex(nodintconst(tmpstringbufsize), types[TUINT8]);
+			a = nod(OADDR, temp(t), N);
+		}
+		n = mkcall("stringtoslicebyte", n->type, init, a, conv(n->left, types[TSTRING]));
 		goto ret;
 
 	case OSTRARRAYBYTETMP:
@@ -1412,8 +1465,14 @@ walkexpr(Node **np, NodeList **init)
 		goto ret;
 
 	case OSTRARRAYRUNE:
-		// stringtoslicerune(string) []rune
-		n = mkcall("stringtoslicerune", n->type, init, n->left);
+		// stringtoslicerune(*[32]rune, string) []rune
+		a = nodnil();
+		if(n->esc == EscNone) {
+			// Create temporary buffer for slice on stack.
+			t = aindex(nodintconst(tmpstringbufsize), types[TINT32]);
+			a = nod(OADDR, temp(t), N);
+		}
+		n = mkcall("stringtoslicerune", n->type, init, a, n->left);
 		goto ret;
 
 	case OCMPIFACE:
@@ -1601,7 +1660,7 @@ ascompatet(int op, NodeList *nl, Type **nr, int fp, NodeList **init)
 			l = tmp;
 		}
 
-		a = nod(OAS, l, arch.nodarg(r, fp));
+		a = nod(OAS, l, nodarg(r, fp));
 		a = convas(a, init);
 		ullmancalc(a);
 		if(a->ullman >= UINF) {
@@ -1654,7 +1713,7 @@ mkdotargslice(NodeList *lr0, NodeList *nn, Type *l, int fp, NodeList **init, Nod
 		walkexpr(&n, init);
 	}
 
-	a = nod(OAS, arch.nodarg(l, fp), n);
+	a = nod(OAS, nodarg(l, fp), n);
 	nn = list(nn, convas(a, init));
 	return nn;
 }
@@ -1734,7 +1793,7 @@ ascompatte(int op, Node *call, int isddd, Type **nl, NodeList *lr, int fp, NodeL
 	if(r != N && lr->next == nil && r->type->etype == TSTRUCT && r->type->funarg) {
 		// optimization - can do block copy
 		if(eqtypenoname(r->type, *nl)) {
-			a = arch.nodarg(*nl, fp);
+			a = nodarg(*nl, fp);
 			r = nod(OCONVNOP, r, N);
 			r->type = a->type;
 			nn = list1(convas(nod(OAS, a, r), init));
@@ -1771,7 +1830,7 @@ loop:
 		// argument to a ddd parameter then it is
 		// passed thru unencapsulated
 		if(r != N && lr->next == nil && isddd && eqtype(l->type, r->type)) {
-			a = nod(OAS, arch.nodarg(l, fp), r);
+			a = nod(OAS, nodarg(l, fp), r);
 			a = convas(a, init);
 			nn = list(nn, a);
 			goto ret;
@@ -1796,7 +1855,7 @@ loop:
 		goto ret;
 	}
 
-	a = nod(OAS, arch.nodarg(l, fp), r);
+	a = nod(OAS, nodarg(l, fp), r);
 	a = convas(a, init);
 	nn = list(nn, a);
 
@@ -1944,8 +2003,7 @@ isstack(Node *n)
 {
 	Node *defn;
 
-	while(n->op == ODOT || n->op == OPAREN || n->op == OCONVNOP || n->op == OINDEX && isfixedarray(n->left->type))
-		n = n->left;
+	n = outervalue(n);
 
 	// If n is *autotmp and autotmp = &foo, replace n with foo.
 	// We introduce such temps when initializing struct literals.
@@ -1976,8 +2034,7 @@ isstack(Node *n)
 static int
 isglobal(Node *n)
 {
-	while(n->op == ODOT || n->op == OPAREN || n->op == OCONVNOP || n->op == OINDEX && isfixedarray(n->left->type))
-		n = n->left;
+	n = outervalue(n);
 
 	switch(n->op) {
 	case ONAME:
@@ -2336,7 +2393,9 @@ Node*
 outervalue(Node *n)
 {	
 	for(;;) {
-		if(n->op == ODOT || n->op == OPAREN) {
+		if(n->op == OXDOT)
+			fatal("OXDOT in walk");
+		if(n->op == ODOT || n->op == OPAREN || n->op == OCONVNOP) {
 			n = n->left;
 			continue;
 		}
@@ -2559,7 +2618,7 @@ paramstoheap(Type **argin, int out)
 			// Defer might stop a panic and show the
 			// return values as they exist at the time of panic.
 			// Make sure to zero them on entry to the function.
-			nn = list(nn, nod(OAS, arch.nodarg(t, 1), N));
+			nn = list(nn, nod(OAS, nodarg(t, 1), N));
 		}
 		if(v == N || !(v->class & PHEAP))
 			continue;
@@ -3300,10 +3359,51 @@ static void
 walkcompare(Node **np, NodeList **init)
 {
 	Node *n, *l, *r, *call, *a, *li, *ri, *expr, *cmpl, *cmpr;
+	Node *x, *ok;
 	int andor, i, needsize;
 	Type *t, *t1;
 	
 	n = *np;
+
+	// Given interface value l and concrete value r, rewrite
+	//   l == r
+	// to
+	//   x, ok := l.(type(r)); ok && x == r
+	// Handle != similarly.
+	// This avoids the allocation that would be required
+	// to convert r to l for comparison.
+	l = N;
+	r = N;
+	if(isinter(n->left->type) && !isinter(n->right->type)) {
+		l = n->left;
+		r = n->right;
+	} else if(!isinter(n->left->type) && isinter(n->right->type)) {
+		l = n->right;
+		r = n->left;
+	}
+	if(l != N) {
+		x = temp(r->type);
+		ok = temp(types[TBOOL]);
+
+		// l.(type(r))
+		a = nod(ODOTTYPE, l, N);
+		a->type = r->type;
+
+		// x, ok := l.(type(r))
+		expr = nod(OAS2, N, N);
+		expr->list = list1(x);
+		expr->list = list(expr->list, ok);
+		expr->rlist = list1(a);
+		typecheck(&expr, Etop);
+		walkexpr(&expr, init);
+
+		if(n->op == OEQ)
+			r = nod(OANDAND, ok, nod(OEQ, x, r));
+		else
+			r = nod(OOROR, nod(ONOT, ok, N), nod(ONE, x, r));
+		*init = list(*init, expr);
+		goto ret;
+	}
 	
 	// Must be comparison of array or struct.
 	// Otherwise back end handles it.
@@ -3447,7 +3547,7 @@ walkrotate(Node **np)
 	Node *l, *r;
 	Node *n;
 
-	if(arch.thechar == '9')
+	if(thearch.thechar == '9')
 		return;
 	
 	n = *np;
@@ -3575,7 +3675,7 @@ walkdiv(Node **np, NodeList **init)
 	Magic m;
 
 	// TODO(minux)
-	if(arch.thechar == '9')
+	if(thearch.thechar == '9')
 		return;
 
 	n = *np;
diff --git a/src/cmd/go/doc.go b/src/cmd/go/doc.go
index 398f83d113..7c92389767 100644
--- a/src/cmd/go/doc.go
+++ b/src/cmd/go/doc.go
@@ -14,32 +14,32 @@ Usage:
 
 The commands are:
 
-    build       compile packages and dependencies
-    clean       remove object files
-    env         print Go environment information
-    fix         run go tool fix on packages
-    fmt         run gofmt on package sources
-    generate    generate Go files by processing source
-    get         download and install packages and dependencies
-    install     compile and install packages and dependencies
-    list        list packages
-    run         compile and run Go program
-    test        test packages
-    tool        run specified go tool
-    version     print Go version
-    vet         run go tool vet on packages
+	build       compile packages and dependencies
+	clean       remove object files
+	env         print Go environment information
+	fix         run go tool fix on packages
+	fmt         run gofmt on package sources
+	generate    generate Go files by processing source
+	get         download and install packages and dependencies
+	install     compile and install packages and dependencies
+	list        list packages
+	run         compile and run Go program
+	test        test packages
+	tool        run specified go tool
+	version     print Go version
+	vet         run go tool vet on packages
 
 Use "go help [command]" for more information about a command.
 
 Additional help topics:
 
-    c           calling between Go and C
-    filetype    file types
-    gopath      GOPATH environment variable
-    importpath  import path syntax
-    packages    description of package lists
-    testflag    description of testing flags
-    testfunc    description of testing functions
+	c           calling between Go and C
+	filetype    file types
+	gopath      GOPATH environment variable
+	importpath  import path syntax
+	packages    description of package lists
+	testflag    description of testing flags
+	testfunc    description of testing functions
 
 Use "go help [topic]" for more information about that topic.
 
@@ -310,6 +310,7 @@ The generator is run in the package's source directory.
 Go generate accepts one specific flag:
 
 	-run=""
+		TODO: This flag is unimplemented.
 		if non-empty, specifies a regular expression to
 		select directives whose command matches the expression.
 
diff --git a/src/cmd/go/generate.go b/src/cmd/go/generate.go
index 8c1e3ee230..3c0af8760b 100644
--- a/src/cmd/go/generate.go
+++ b/src/cmd/go/generate.go
@@ -106,6 +106,7 @@ The generator is run in the package's source directory.
 Go generate accepts one specific flag:
 
 	-run=""
+		TODO: This flag is unimplemented.
 		if non-empty, specifies a regular expression to
 		select directives whose command matches the expression.
 
diff --git a/src/cmd/go/main.go b/src/cmd/go/main.go
index 1482a39582..d5e86395a2 100644
--- a/src/cmd/go/main.go
+++ b/src/cmd/go/main.go
@@ -180,13 +180,13 @@ Usage:
 
 The commands are:
 {{range .}}{{if .Runnable}}
-    {{.Name | printf "%-11s"}} {{.Short}}{{end}}{{end}}
+	{{.Name | printf "%-11s"}} {{.Short}}{{end}}{{end}}
 
 Use "go help [command]" for more information about a command.
 
 Additional help topics:
 {{range .}}{{if not .Runnable}}
-    {{.Name | printf "%-11s"}} {{.Short}}{{end}}{{end}}
+	{{.Name | printf "%-11s"}} {{.Short}}{{end}}{{end}}
 
 Use "go help [topic]" for more information about that topic.
 
diff --git a/src/cmd/go/mkdoc.sh b/src/cmd/go/mkdoc.sh
index e15e8809ce..507a8aeddc 100755
--- a/src/cmd/go/mkdoc.sh
+++ b/src/cmd/go/mkdoc.sh
@@ -1,9 +1,12 @@
-#!/bin/sh
+#!/bin/bash
 # Copyright 2012 The Go Authors.  All rights reserved.
 # Use of this source code is governed by a BSD-style
 # license that can be found in the LICENSE file.
 
-go install # So the next line will produce updated documentation.
-go help documentation | sed 's; \*/; * /;' >doc.go
+set -e
+
+go build -o go.latest
+./go.latest help documentation | sed 's; \*/; * /;' >doc.go
 gofmt -w doc.go
+rm go.latest
 
diff --git a/src/cmd/go/test.go b/src/cmd/go/test.go
index 60f6b16c9a..a8110f385d 100644
--- a/src/cmd/go/test.go
+++ b/src/cmd/go/test.go
@@ -346,11 +346,11 @@ func runTest(cmd *Command, args []string) {
 	// been given on the command line (implicit current directory)
 	// or when benchmarking.
 	// Also stream if we're showing output anyway with a
-	// single package under test.  In that case, streaming the
-	// output produces the same result as not streaming,
-	// just more immediately.
+	// single package under test or if parallelism is set to 1.
+	// In these cases, streaming the output produces the same result
+	// as not streaming, just more immediately.
 	testStreamOutput = len(pkgArgs) == 0 || testBench ||
-		(len(pkgs) <= 1 && testShowPass)
+		(testShowPass && (len(pkgs) == 1 || buildP == 1))
 
 	var b builder
 	b.init()
diff --git a/src/cmd/gofmt/doc.go b/src/cmd/gofmt/doc.go
index 3fc0439548..9d0cd32862 100644
--- a/src/cmd/gofmt/doc.go
+++ b/src/cmd/gofmt/doc.go
@@ -87,6 +87,13 @@ When invoked with -s gofmt will make the following source transformations where
 		for x, _ = range v {...}
 	will be simplified to:
 		for x = range v {...}
+
+	A range of the form:
+		for _ = range v {...}
+	will be simplified to:
+		for range v {...}
+
+This may result in changes that are incompatible with earlier versions of Go.
 */
 package main
 
diff --git a/src/cmd/ld/Makefile b/src/cmd/ld/Makefile
new file mode 100644
index 0000000000..3f528d7517
--- /dev/null
+++ b/src/cmd/ld/Makefile
@@ -0,0 +1,5 @@
+# Copyright 2012 The Go Authors. All rights reserved.
+# Use of this source code is governed by a BSD-style
+# license that can be found in the LICENSE file.
+
+include ../../Make.dist
diff --git a/src/cmd/ld/data.c b/src/cmd/ld/data.c
index 0f287c202f..580d36f7e2 100644
--- a/src/cmd/ld/data.c
+++ b/src/cmd/ld/data.c
@@ -30,11 +30,14 @@
 
 // Data layout and relocation.
 
-#include	"l.h"
-#include	"../ld/lib.h"
-#include	"../ld/elf.h"
-#include	"../ld/macho.h"
-#include	"../ld/pe.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
+#include	"lib.h"
+#include	"elf.h"
+#include	"macho.h"
+#include	"pe.h"
 #include	"../../runtime/mgc0.h"
 
 void	dynreloc(void);
@@ -150,7 +153,7 @@ relocsym(LSym *s)
 			diag("%s: invalid relocation %d+%d not in [%d,%d)", s->name, off, siz, 0, s->np);
 			continue;
 		}
-		if(r->sym != S && ((r->sym->type & (SMASK | SHIDDEN)) == 0 || (r->sym->type & SMASK) == SXREF)) {
+		if(r->sym != nil && ((r->sym->type & (SMASK | SHIDDEN)) == 0 || (r->sym->type & SMASK) == SXREF)) {
 			diag("%s: not defined", r->sym->name);
 			continue;
 		}
@@ -160,9 +163,9 @@ relocsym(LSym *s)
 			continue;
 
 		// Solaris needs the ability to reference dynimport symbols.
-		if(HEADTYPE != Hsolaris && r->sym != S && r->sym->type == SDYNIMPORT)
+		if(HEADTYPE != Hsolaris && r->sym != nil && r->sym->type == SDYNIMPORT)
 			diag("unhandled relocation for %s (type %d rtype %d)", r->sym->name, r->sym->type, r->type);
-		if(r->sym != S && r->sym->type != STLSBSS && !r->sym->reachable)
+		if(r->sym != nil && r->sym->type != STLSBSS && !r->sym->reachable)
 			diag("unreachable sym in relocation: %s %s", s->name, r->sym->name);
 
 		// Android emulates runtime.tlsg as a regular variable.
@@ -172,11 +175,11 @@ relocsym(LSym *s)
 		switch(r->type) {
 		default:
 			o = 0;
-			if(archreloc(r, s, &o) < 0)
+			if(thearch.archreloc(r, s, &o) < 0)
 				diag("unknown reloc %d", r->type);
 			break;
 		case R_TLS:
-			if(linkmode == LinkInternal && iself && thechar == '5') {
+			if(linkmode == LinkInternal && iself && thearch.thechar == '5') {
 				// On ELF ARM, the thread pointer is 8 bytes before
 				// the start of the thread-local data block, so add 8
 				// to the actual TLS offset (r->sym->value).
@@ -189,7 +192,7 @@ relocsym(LSym *s)
 			}
 			r->done = 0;
 			o = 0;
-			if(thechar != '6')
+			if(thearch.thechar != '6')
 				o = r->add;
 			break;
 		case R_TLS_LE:
@@ -199,7 +202,7 @@ relocsym(LSym *s)
 				r->xsym = ctxt->tlsg;
 				r->xadd = r->add;
 				o = 0;
-				if(thechar != '6')
+				if(thearch.thechar != '6')
 					o = r->add;
 				break;
 			}
@@ -213,7 +216,7 @@ relocsym(LSym *s)
 				r->xsym = ctxt->tlsg;
 				r->xadd = r->add;
 				o = 0;
-				if(thechar != '6')
+				if(thearch.thechar != '6')
 					o = r->add;
 				break;
 			}
@@ -241,7 +244,7 @@ relocsym(LSym *s)
 
 				o = r->xadd;
 				if(iself) {
-					if(thechar == '6')
+					if(thearch.thechar == '6')
 						o = 0;
 				} else if(HEADTYPE == Hdarwin) {
 					if(rs->type != SHOSTOBJ)
@@ -258,7 +261,7 @@ relocsym(LSym *s)
 			// fail at runtime. See http://golang.org/issue/7980.
 			// Instead of special casing only amd64, we treat this as an error on all
 			// 64-bit architectures so as to be future-proof.
-			if((int32)o < 0 && PtrSize > 4 && siz == 4) {
+			if((int32)o < 0 && thearch.ptrsize > 4 && siz == 4) {
 				diag("non-pc-relative relocation address is too big: %#llux", o);
 				errorexit();
 			}
@@ -283,7 +286,7 @@ relocsym(LSym *s)
 
 				o = r->xadd;
 				if(iself) {
-					if(thechar == '6')
+					if(thearch.thechar == '6')
 						o = 0;
 				} else if(HEADTYPE == Hdarwin) {
 					if(r->type == R_CALL) {
@@ -314,7 +317,7 @@ relocsym(LSym *s)
 			break;
 		}
 		if(r->variant != RV_NONE)
-			o = archrelocvariant(r, s, o);
+			o = thearch.archrelocvariant(r, s, o);
 //print("relocate %s %#llux (%#llux+%#llux, size %d) => %s %#llux +%#llx [%llx]\n", s->name, (uvlong)(s->value+off), (uvlong)s->value, (uvlong)r->off, r->siz, r->sym ? r->sym->name : "<nil>", (uvlong)symaddr(r->sym), (vlong)r->add, (vlong)o);
 		switch(siz) {
 		default:
@@ -363,9 +366,9 @@ reloc(void)
 		Bprint(&bso, "%5.2f reloc\n", cputime());
 	Bflush(&bso);
 
-	for(s=ctxt->textp; s!=S; s=s->next)
+	for(s=ctxt->textp; s!=nil; s=s->next)
 		relocsym(s);
-	for(s=datap; s!=S; s=s->next)
+	for(s=datap; s!=nil; s=s->next)
 		relocsym(s);
 }
 
@@ -392,7 +395,7 @@ dynrelocsym(LSym *s)
 				r->add = targ->plt;
 
 				// jmp *addr
-				if(thechar == '8') {
+				if(thearch.thechar == '8') {
 					adduint8(ctxt, rel, 0xff);
 					adduint8(ctxt, rel, 0x25);
 					addaddr(ctxt, rel, targ);
@@ -414,10 +417,10 @@ dynrelocsym(LSym *s)
 	}
 
 	for(r=s->r; r<s->r+s->nr; r++) {
-		if(r->sym != S && r->sym->type == SDYNIMPORT || r->type >= 256) {
-			if(r->sym != S && !r->sym->reachable)
+		if(r->sym != nil && r->sym->type == SDYNIMPORT || r->type >= 256) {
+			if(r->sym != nil && !r->sym->reachable)
 				diag("internal inconsistency: dynamic symbol %s is not reachable.", r->sym->name);
-			adddynrel(s, r);
+			thearch.adddynrel(s, r);
 		}
 	}
 }
@@ -435,9 +438,9 @@ dynreloc(void)
 		Bprint(&bso, "%5.2f reloc\n", cputime());
 	Bflush(&bso);
 
-	for(s=ctxt->textp; s!=S; s=s->next)
+	for(s=ctxt->textp; s!=nil; s=s->next)
 		dynrelocsym(s);
-	for(s=datap; s!=S; s=s->next)
+	for(s=datap; s!=nil; s=s->next)
 		dynrelocsym(s);
 	if(iself)
 		elfdynhash();
@@ -650,7 +653,7 @@ addstrdata(char *name, char *value)
 	s->dupok = 1;
 	reachable = s->reachable;
 	addaddr(ctxt, s, sp);
-	adduintxx(ctxt, s, strlen(value), PtrSize);
+	adduintxx(ctxt, s, strlen(value), thearch.ptrsize);
 
 	// addstring, addaddr, etc., mark the symbols as reachable.
 	// In this case that is not necessarily true, so stick to what
@@ -701,7 +704,7 @@ symalign(LSym *s)
 	if(s->align != 0)
 		return s->align;
 
-	align = MaxAlign;
+	align = thearch.maxalign;
 	while(align > s->size && align > 1)
 		align >>= 1;
 	if(align < s->align)
@@ -723,7 +726,7 @@ maxalign(LSym *s, int type)
 	int32 align, max;
 	
 	max = 0;
-	for(; s != S && s->type <= type; s = s->next) {
+	for(; s != nil && s->type <= type; s = s->next) {
 		align = symalign(s);
 		if(max < align)
 			max = align;
@@ -789,7 +792,7 @@ proggenskip(ProgGen *g, vlong off, vlong v)
 	vlong i;
 
 	for(i = off; i < off+v; i++) {
-		if((i%PtrSize) == 0)
+		if((i%thearch.ptrsize) == 0)
 			proggendata(g, BitsScalar);
 	}
 }
@@ -802,7 +805,7 @@ proggenarray(ProgGen *g, vlong len)
 
 	proggendataflush(g);
 	proggenemit(g, insArray);
-	for(i = 0; i < PtrSize; i++, len >>= 8)
+	for(i = 0; i < thearch.ptrsize; i++, len >>= 8)
 		proggenemit(g, len);
 }
 
@@ -843,39 +846,39 @@ proggenaddsym(ProgGen *g, LSym *s)
 	// and not SDATA, but sometimes that doesn't happen.
 	// Leave debugging the SDATA issue for the Go rewrite.
 
-	if(s->gotype == nil && s->size >= PtrSize && s->name[0] != '.') {
+	if(s->gotype == nil && s->size >= thearch.ptrsize && s->name[0] != '.') {
 		// conservative scan
 		diag("missing Go type information for global symbol: %s size %d", s->name, (int)s->size);
-		if((s->size%PtrSize) || (g->pos%PtrSize))
+		if((s->size%thearch.ptrsize) || (g->pos%thearch.ptrsize))
 			diag("proggenaddsym: unaligned conservative symbol %s: size=%lld pos=%lld",
 				s->name, s->size, g->pos);
-		size = (s->size+PtrSize-1)/PtrSize*PtrSize;
-		if(size < 32*PtrSize) {
+		size = (s->size+thearch.ptrsize-1)/thearch.ptrsize*thearch.ptrsize;
+		if(size < 32*thearch.ptrsize) {
 			// Emit small symbols as data.
-			for(i = 0; i < size/PtrSize; i++)
+			for(i = 0; i < size/thearch.ptrsize; i++)
 				proggendata(g, BitsPointer);
 		} else {
 			// Emit large symbols as array.
-			proggenarray(g, size/PtrSize);
+			proggenarray(g, size/thearch.ptrsize);
 			proggendata(g, BitsPointer);
 			proggenarrayend(g);
 		}
 		g->pos = s->value + size;
-	} else if(s->gotype == nil || decodetype_noptr(s->gotype) || s->size < PtrSize || s->name[0] == '.') {
+	} else if(s->gotype == nil || decodetype_noptr(s->gotype) || s->size < thearch.ptrsize || s->name[0] == '.') {
 		// no scan
-		if(s->size < 32*PtrSize) {
+		if(s->size < 32*thearch.ptrsize) {
 			// Emit small symbols as data.
 			// This case also handles unaligned and tiny symbols, so tread carefully.
 			for(i = s->value; i < s->value+s->size; i++) {
-				if((i%PtrSize) == 0)
+				if((i%thearch.ptrsize) == 0)
 					proggendata(g, BitsScalar);
 			}
 		} else {
 			// Emit large symbols as array.
-			if((s->size%PtrSize) || (g->pos%PtrSize))
+			if((s->size%thearch.ptrsize) || (g->pos%thearch.ptrsize))
 				diag("proggenaddsym: unaligned noscan symbol %s: size=%lld pos=%lld",
 					s->name, s->size, g->pos);
-			proggenarray(g, s->size/PtrSize);
+			proggenarray(g, s->size/thearch.ptrsize);
 			proggendata(g, BitsScalar);
 			proggenarrayend(g);
 		}
@@ -885,7 +888,7 @@ proggenaddsym(ProgGen *g, LSym *s)
 		proggendataflush(g);
 		gcprog = decodetype_gcprog(s->gotype);
 		size = decodetype_size(s->gotype);
-		if((size%PtrSize) || (g->pos%PtrSize))
+		if((size%thearch.ptrsize) || (g->pos%thearch.ptrsize))
 			diag("proggenaddsym: unaligned gcprog symbol %s: size=%lld pos=%lld",
 				s->name, s->size, g->pos);
 		for(i = 0; i < gcprog->np-1; i++)
@@ -895,11 +898,11 @@ proggenaddsym(ProgGen *g, LSym *s)
 		// gc mask, it's small so emit as data
 		mask = decodetype_gcmask(s->gotype);
 		size = decodetype_size(s->gotype);
-		if((size%PtrSize) || (g->pos%PtrSize))
+		if((size%thearch.ptrsize) || (g->pos%thearch.ptrsize))
 			diag("proggenaddsym: unaligned gcmask symbol %s: size=%lld pos=%lld",
 				s->name, s->size, g->pos);
-		for(i = 0; i < size; i += PtrSize)
-			proggendata(g, (mask[i/PtrSize/2]>>((i/PtrSize%2)*4+2))&BitsMask);
+		for(i = 0; i < size; i += thearch.ptrsize)
+			proggendata(g, (mask[i/thearch.ptrsize/2]>>((i/thearch.ptrsize%2)*4+2))&BitsMask);
 		g->pos = s->value + size;
 	}
 }
@@ -935,7 +938,7 @@ dodata(void)
 	last = nil;
 	datap = nil;
 
-	for(s=ctxt->allsym; s!=S; s=s->allsym) {
+	for(s=ctxt->allsym; s!=nil; s=s->allsym) {
 		if(!s->reachable || s->special)
 			continue;
 		if(STEXT < s->type && s->type < SXREF) {
@@ -1134,7 +1137,7 @@ dodata(void)
 	
 	if(iself && linkmode == LinkExternal && s != nil && s->type == STLSBSS && HEADTYPE != Hopenbsd) {
 		sect = addsection(&segdata, ".tbss", 06);
-		sect->align = PtrSize;
+		sect->align = thearch.ptrsize;
 		sect->vaddr = 0;
 		datsize = 0;
 		for(; s != nil && s->type == STLSBSS; s = s->next) {
@@ -1311,9 +1314,9 @@ textaddress(void)
 		else
 			va = rnd(va, funcalign);
 		sym->value = 0;
-		for(sub = sym; sub != S; sub = sub->sub)
+		for(sub = sym; sub != nil; sub = sub->sub)
 			sub->value += va;
-		if(sym->size == 0 && sym->sub != S)
+		if(sym->size == 0 && sym->sub != nil)
 			ctxt->cursym = sym;
 		if(sym->size < MINFUNC)
 			va += MINFUNC; // spacing required for findfunctab
diff --git a/src/cmd/ld/decodesym.c b/src/cmd/ld/decodesym.c
index eedb246789..c194a1f767 100644
--- a/src/cmd/ld/decodesym.c
+++ b/src/cmd/ld/decodesym.c
@@ -2,7 +2,10 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-#include	"l.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
 #include	"lib.h"
 #include	"../../runtime/typekind.h"
 
@@ -70,42 +73,42 @@ decode_inuxi(uchar* p, int sz)
 static int
 commonsize(void)
 {
-	return 8*PtrSize + 8;
+	return 8*thearch.ptrsize + 8;
 }
 
 // Type.commonType.kind
 uint8
 decodetype_kind(LSym *s)
 {
-	return s->p[1*PtrSize + 7] & KindMask;	//  0x13 / 0x1f
+	return s->p[1*thearch.ptrsize + 7] & KindMask;	//  0x13 / 0x1f
 }
 
 // Type.commonType.kind
 uint8
 decodetype_noptr(LSym *s)
 {
-	return s->p[1*PtrSize + 7] & KindNoPointers;	//  0x13 / 0x1f
+	return s->p[1*thearch.ptrsize + 7] & KindNoPointers;	//  0x13 / 0x1f
 }
 
 // Type.commonType.kind
 uint8
 decodetype_usegcprog(LSym *s)
 {
-	return s->p[1*PtrSize + 7] & KindGCProg;	//  0x13 / 0x1f
+	return s->p[1*thearch.ptrsize + 7] & KindGCProg;	//  0x13 / 0x1f
 }
 
 // Type.commonType.size
 vlong
 decodetype_size(LSym *s)
 {
-	return decode_inuxi(s->p, PtrSize);	 // 0x8 / 0x10
+	return decode_inuxi(s->p, thearch.ptrsize);	 // 0x8 / 0x10
 }
 
 // Type.commonType.gc
 LSym*
 decodetype_gcprog(LSym *s)
 {
-	return decode_reloc_sym(s, 1*PtrSize + 8 + 2*PtrSize);
+	return decode_reloc_sym(s, 1*thearch.ptrsize + 8 + 2*thearch.ptrsize);
 }
 
 uint8*
@@ -113,7 +116,7 @@ decodetype_gcmask(LSym *s)
 {
 	LSym *mask;
 	
-	mask = decode_reloc_sym(s, 1*PtrSize + 8 + 1*PtrSize);
+	mask = decode_reloc_sym(s, 1*thearch.ptrsize + 8 + 1*thearch.ptrsize);
 	return mask->p;
 }
 
@@ -127,7 +130,7 @@ decodetype_arrayelem(LSym *s)
 vlong
 decodetype_arraylen(LSym *s)
 {
-	return decode_inuxi(s->p + commonsize()+2*PtrSize, PtrSize);
+	return decode_inuxi(s->p + commonsize()+2*thearch.ptrsize, thearch.ptrsize);
 }
 
 // Type.PtrType.elem
@@ -147,7 +150,7 @@ decodetype_mapkey(LSym *s)
 LSym*
 decodetype_mapvalue(LSym *s)
 {
-	return decode_reloc_sym(s, commonsize()+PtrSize);	// 0x20 / 0x38
+	return decode_reloc_sym(s, commonsize()+thearch.ptrsize);	// 0x20 / 0x38
 }
 
 // Type.ChanType.elem
@@ -168,13 +171,13 @@ decodetype_funcdotdotdot(LSym *s)
 int
 decodetype_funcincount(LSym *s)
 {
-	return decode_inuxi(s->p + commonsize()+2*PtrSize, IntSize);
+	return decode_inuxi(s->p + commonsize()+2*thearch.ptrsize, thearch.intsize);
 }
 
 int
 decodetype_funcoutcount(LSym *s)
 {
-	return decode_inuxi(s->p + commonsize()+3*PtrSize + 2*IntSize, IntSize);
+	return decode_inuxi(s->p + commonsize()+3*thearch.ptrsize + 2*thearch.intsize, thearch.intsize);
 }
 
 LSym*
@@ -182,10 +185,10 @@ decodetype_funcintype(LSym *s, int i)
 {
 	Reloc *r;
 
-	r = decode_reloc(s, commonsize() + PtrSize);
+	r = decode_reloc(s, commonsize() + thearch.ptrsize);
 	if (r == nil)
 		return nil;
-	return decode_reloc_sym(r->sym, r->add + i * PtrSize);
+	return decode_reloc_sym(r->sym, r->add + i * thearch.ptrsize);
 }
 
 LSym*
@@ -193,23 +196,23 @@ decodetype_funcouttype(LSym *s, int i)
 {
 	Reloc *r;
 
-	r = decode_reloc(s, commonsize() + 2*PtrSize + 2*IntSize);
+	r = decode_reloc(s, commonsize() + 2*thearch.ptrsize + 2*thearch.intsize);
 	if (r == nil)
 		return nil;
-	return decode_reloc_sym(r->sym, r->add + i * PtrSize);
+	return decode_reloc_sym(r->sym, r->add + i * thearch.ptrsize);
 }
 
 // Type.StructType.fields.Slice::len
 int
 decodetype_structfieldcount(LSym *s)
 {
-	return decode_inuxi(s->p + commonsize() + PtrSize, IntSize);
+	return decode_inuxi(s->p + commonsize() + thearch.ptrsize, thearch.intsize);
 }
 
 static int
 structfieldsize(void)
 {
-	return 5*PtrSize;
+	return 5*thearch.ptrsize;
 }
 
 // Type.StructType.fields[]-> name, typ and offset.
@@ -219,7 +222,7 @@ decodetype_structfieldname(LSym *s, int i)
 	Reloc *r;
 
 	// go.string."foo"  0x28 / 0x40
-	s = decode_reloc_sym(s, commonsize() + PtrSize + 2*IntSize + i*structfieldsize());
+	s = decode_reloc_sym(s, commonsize() + thearch.ptrsize + 2*thearch.intsize + i*structfieldsize());
 	if (s == nil)			// embedded structs have a nil name.
 		return nil;
 	r = decode_reloc(s, 0);		// s has a pointer to the string data at offset 0
@@ -231,18 +234,18 @@ decodetype_structfieldname(LSym *s, int i)
 LSym*
 decodetype_structfieldtype(LSym *s, int i)
 {
-	return decode_reloc_sym(s, commonsize() + PtrSize + 2*IntSize + i*structfieldsize() + 2*PtrSize);
+	return decode_reloc_sym(s, commonsize() + thearch.ptrsize + 2*thearch.intsize + i*structfieldsize() + 2*thearch.ptrsize);
 }
 
 vlong
 decodetype_structfieldoffs(LSym *s, int i)
 {
-	return decode_inuxi(s->p + commonsize() + PtrSize + 2*IntSize + i*structfieldsize() + 4*PtrSize, IntSize);
+	return decode_inuxi(s->p + commonsize() + thearch.ptrsize + 2*thearch.intsize + i*structfieldsize() + 4*thearch.ptrsize, thearch.intsize);
 }
 
 // InterfaceTYpe.methods.len
 vlong
 decodetype_ifacemethodcount(LSym *s)
 {
-	return decode_inuxi(s->p + commonsize() + PtrSize, IntSize);
+	return decode_inuxi(s->p + commonsize() + thearch.ptrsize, thearch.intsize);
 }
diff --git a/src/cmd/ld/dwarf.c b/src/cmd/ld/dwarf.c
index 061171ea0b..53f0194b31 100644
--- a/src/cmd/ld/dwarf.c
+++ b/src/cmd/ld/dwarf.c
@@ -12,13 +12,16 @@
 //   - file:line info for variables
 //   - make strings a typedef so prettyprinters can see the underlying string type
 //
-#include	"l.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
 #include	"lib.h"
-#include	"../ld/dwarf.h"
-#include	"../ld/dwarf_defs.h"
-#include	"../ld/elf.h"
-#include	"../ld/macho.h"
-#include	"../ld/pe.h"
+#include	"dwarf.h"
+#include	"dwarf_defs.h"
+#include	"elf.h"
+#include	"macho.h"
+#include	"pe.h"
 #include	"../../runtime/typekind.h"
 
 /*
@@ -75,12 +78,12 @@ static char  gdbscript[1024];
 static void
 addrput(vlong addr)
 {
-	switch(PtrSize) {
+	switch(thearch.ptrsize) {
 	case 4:
-		LPUT(addr);
+		thearch.lput(addr);
 		break;
 	case 8:
-		VPUT(addr);
+		thearch.vput(addr);
 		break;
 	}
 }
@@ -629,14 +632,14 @@ adddwarfrel(LSym* sec, LSym* sym, vlong offsetbase, int siz, vlong addend)
 	r->type = R_ADDR;
 	r->add = addend;
 	r->xadd = addend;
-	if(iself && thechar == '6')
+	if(iself && thearch.thechar == '6')
 		addend = 0;
 	switch(siz) {
 	case 4:
-		LPUT(addend);
+		thearch.lput(addend);
 		break;
 	case 8:
-		VPUT(addend);
+		thearch.vput(addend);
 		break;
 	default:
 		diag("bad size in adddwarfrel");
@@ -663,7 +666,7 @@ putattr(int abbrev, int form, int cls, vlong value, char *data)
 	case DW_FORM_addr:	// address
 		if(linkmode == LinkExternal) {
 			value -= ((LSym*)data)->value;
-			adddwarfrel(infosec, (LSym*)data, infoo, PtrSize, value);
+			adddwarfrel(infosec, (LSym*)data, infoo, thearch.ptrsize, value);
 			break;
 		}
 		addrput(value);
@@ -671,11 +674,11 @@ putattr(int abbrev, int form, int cls, vlong value, char *data)
 
 	case DW_FORM_block1:	// block
 		if(cls == DW_CLS_ADDRESS) {
-			cput(1+PtrSize);
+			cput(1+thearch.ptrsize);
 			cput(DW_OP_addr);
 			if(linkmode == LinkExternal) {
 				value -= ((LSym*)data)->value;
-				adddwarfrel(infosec, (LSym*)data, infoo, PtrSize, value);
+				adddwarfrel(infosec, (LSym*)data, infoo, thearch.ptrsize, value);
 				break;
 			}
 			addrput(value);
@@ -689,14 +692,14 @@ putattr(int abbrev, int form, int cls, vlong value, char *data)
 
 	case DW_FORM_block2:	// block
 		value &= 0xffff;
-		WPUT(value);
+		thearch.wput(value);
 		while(value--)
 			cput(*data++);
 		break;
 
 	case DW_FORM_block4:	// block
 		value &= 0xffffffff;
-		LPUT(value);
+		thearch.lput(value);
 		while(value--)
 			cput(*data++);
 		break;
@@ -712,7 +715,7 @@ putattr(int abbrev, int form, int cls, vlong value, char *data)
 		break;
 
 	case DW_FORM_data2:	// constant
-		WPUT(value);
+		thearch.wput(value);
 		break;
 
 	case DW_FORM_data4:	// constant, {line,loclist,mac,rangelist}ptr
@@ -720,11 +723,11 @@ putattr(int abbrev, int form, int cls, vlong value, char *data)
 			adddwarfrel(infosec, linesym, infoo, 4, value);
 			break;
 		}
-		LPUT(value);
+		thearch.lput(value);
 		break;
 
 	case DW_FORM_data8:	// constant, {line,loclist,mac,rangelist}ptr
-		VPUT(value);
+		thearch.vput(value);
 		break;
 
 	case DW_FORM_sdata:	// constant
@@ -750,16 +753,16 @@ putattr(int abbrev, int form, int cls, vlong value, char *data)
 		// (> 4 GB of debug info aka "64-bit") unit, which we don't implement.
 		if (data == nil) {
 			diag("dwarf: null reference in %d", abbrev);
-			if(PtrSize == 8)
-				VPUT(0); // invalid dwarf, gdb will complain.
+			if(thearch.ptrsize == 8)
+				thearch.vput(0); // invalid dwarf, gdb will complain.
 			else
-				LPUT(0); // invalid dwarf, gdb will complain.
+				thearch.lput(0); // invalid dwarf, gdb will complain.
 		} else {
 			off = ((DWDie*)data)->offs;
 			if (off == 0)
 				fwdcount++;
 			if(linkmode == LinkExternal) {
-				adddwarfrel(infosec, infosym, infoo, PtrSize, off);
+				adddwarfrel(infosec, infosym, infoo, thearch.ptrsize, off);
 				break;
 			}
 			addrput(off);
@@ -1236,17 +1239,17 @@ synthesizemaptypes(DWDie *die)
 
 		// compute size info like hashmap.c does.
 		a = getattr(keytype, DW_AT_byte_size);
-		keysize = a ? a->value : PtrSize;  // We don't store size with Pointers
+		keysize = a ? a->value : thearch.ptrsize;  // We don't store size with Pointers
 		a = getattr(valtype, DW_AT_byte_size);
-		valsize = a ? a->value : PtrSize;
+		valsize = a ? a->value : thearch.ptrsize;
 		indirect_key = 0;
 		indirect_val = 0;
 		if(keysize > MaxKeySize) {
-			keysize = PtrSize;
+			keysize = thearch.ptrsize;
 			indirect_key = 1;
 		}
 		if(valsize > MaxValSize) {
-			valsize = PtrSize;
+			valsize = thearch.ptrsize;
 			indirect_val = 1;
 		}
 
@@ -1288,12 +1291,12 @@ synthesizemaptypes(DWDie *die)
 		fld = newdie(dwhb, DW_ABRV_STRUCTFIELD, "overflow");
 		newrefattr(fld, DW_AT_type, defptrto(dwhb));
 		newmemberoffsetattr(fld, BucketSize + BucketSize * (keysize + valsize));
-		if(RegSize > PtrSize) {
+		if(thearch.regsize > thearch.ptrsize) {
 			fld = newdie(dwhb, DW_ABRV_STRUCTFIELD, "pad");
 			newrefattr(fld, DW_AT_type, find_or_diag(&dwtypes, "uintptr"));
-			newmemberoffsetattr(fld, BucketSize + BucketSize * (keysize + valsize) + PtrSize);
+			newmemberoffsetattr(fld, BucketSize + BucketSize * (keysize + valsize) + thearch.ptrsize);
 		}
-		newattr(dwhb, DW_AT_byte_size, DW_CLS_CONSTANT, BucketSize + BucketSize * keysize + BucketSize * valsize + RegSize, 0);
+		newattr(dwhb, DW_AT_byte_size, DW_CLS_CONSTANT, BucketSize + BucketSize * keysize + BucketSize * valsize + thearch.regsize, 0);
 
 		// Construct hash<K,V>
 		dwh = newdie(&dwtypes, DW_ABRV_STRUCTTYPE,
@@ -1332,7 +1335,7 @@ synthesizechantypes(DWDie *die)
 			continue;
 		elemtype = (DWDie*) getattr(die, DW_AT_go_elem)->data;
 		a = getattr(elemtype, DW_AT_byte_size);
-		elemsize = a ? a->value : PtrSize;
+		elemsize = a ? a->value : thearch.ptrsize;
 
 		// sudog<T>
 		dws = newdie(&dwtypes, DW_ABRV_STRUCTTYPE,
@@ -1414,7 +1417,7 @@ movetomodule(DWDie *parent)
 	die->link = parent->child;
 }
 
-// If the pcln table contains runtime/string.goc, use that to set gdbscript path.
+// If the pcln table contains runtime/runtime.go, use that to set gdbscript path.
 static void
 finddebugruntimepath(LSym *s)
 {
@@ -1427,7 +1430,7 @@ finddebugruntimepath(LSym *s)
 
 	for(i=0; i<s->pcln->nfile; i++) {
 		f = s->pcln->file[i];
-		if((p = strstr(f->name, "runtime/string.goc")) != nil) {
+		if((p = strstr(f->name, "runtime/runtime.go")) != nil) {
 			*p = '\0';
 			snprint(gdbscript, sizeof gdbscript, "%sruntime/runtime-gdb.py", f->name);
 			*p = 'r';
@@ -1519,9 +1522,9 @@ flushunit(DWDie *dwinfo, vlong pc, LSym *pcsym, vlong unitstart, int32 header_le
 
 		here = cpos();
 		cseek(unitstart);
-		LPUT(here - unitstart - sizeof(int32));	 // unit_length
-		WPUT(2);  // dwarf version
-		LPUT(header_length); // header length starting here
+		thearch.lput(here - unitstart - sizeof(int32));	 // unit_length
+		thearch.wput(2);  // dwarf version
+		thearch.lput(header_length); // header length starting here
 		cseek(here);
 	}
 }
@@ -1540,14 +1543,14 @@ writelines(void)
 	Pciter pcfile, pcline;
 	LSym **files, *f;
 
-	if(linesec == S)
+	if(linesec == nil)
 		linesec = linklookup(ctxt, ".dwarfline", 0);
 	linesec->nr = 0;
 
 	unitstart = -1;
 	headerend = -1;
 	epc = 0;
-	epcs = S;
+	epcs = nil;
 	lineo = cpos();
 	dwinfo = nil;
 	
@@ -1565,9 +1568,9 @@ writelines(void)
 
 	// Write .debug_line Line Number Program Header (sec 6.2.4)
 	// Fields marked with (*) must be changed for 64-bit dwarf
-	LPUT(0);   // unit_length (*), will be filled in by flushunit.
-	WPUT(2);   // dwarf version (appendix F)
-	LPUT(0);   // header_length (*), filled in by flushunit.
+	thearch.lput(0);   // unit_length (*), will be filled in by flushunit.
+	thearch.wput(2);   // dwarf version (appendix F)
+	thearch.lput(0);   // header_length (*), filled in by flushunit.
 	// cpos == unitstart + 4 + 2 + 4
 	cput(1);   // minimum_instruction_length
 	cput(1);   // default_is_stmt
@@ -1598,14 +1601,14 @@ writelines(void)
 	headerend = cpos();
 
 	cput(0);  // start extended opcode
-	uleb128put(1 + PtrSize);
+	uleb128put(1 + thearch.ptrsize);
 	cput(DW_LNE_set_address);
 
 	pc = s->value;
 	line = 1;
 	file = 1;
 	if(linkmode == LinkExternal)
-		adddwarfrel(linesec, s, lineo, PtrSize, 0);
+		adddwarfrel(linesec, s, lineo, thearch.ptrsize, 0);
 	else
 		addrput(pc);
 
@@ -1661,7 +1664,7 @@ writelines(void)
 			switch (a->name) {
 			case A_AUTO:
 				dt = DW_ABRV_AUTO;
-				offs = a->aoffset - PtrSize;
+				offs = a->aoffset - thearch.ptrsize;
 				break;
 			case A_PARAM:
 				dt = DW_ABRV_PARAM;
@@ -1710,7 +1713,7 @@ writelines(void)
 enum
 {
 	CIERESERVE = 16,
-	DATAALIGNMENTFACTOR = -4,	// TODO -PtrSize?
+	DATAALIGNMENTFACTOR = -4,	// TODO -thearch.ptrsize?
 	FAKERETURNCOLUMN = 16		// TODO gdb6 doesn't like > 15?
 };
 
@@ -1727,10 +1730,10 @@ putpccfadelta(vlong deltapc, vlong cfa)
 		cput(deltapc);
 	} else if (deltapc < 0x10000) {
 		cput(DW_CFA_advance_loc2);
-		WPUT(deltapc);
+		thearch.wput(deltapc);
 	} else {
 		cput(DW_CFA_advance_loc4);
-		LPUT(deltapc);
+		thearch.lput(deltapc);
 	}
 }
 
@@ -1742,14 +1745,14 @@ writeframes(void)
 	Pciter pcsp;
 	uint32 nextpc;
 
-	if(framesec == S)
+	if(framesec == nil)
 		framesec = linklookup(ctxt, ".dwarfframe", 0);
 	framesec->nr = 0;
 	frameo = cpos();
 
 	// Emit the CIE, Section 6.4.1
-	LPUT(CIERESERVE);	// initial length, must be multiple of PtrSize
-	LPUT(0xffffffff);	// cid.
+	thearch.lput(CIERESERVE);	// initial length, must be multiple of thearch.ptrsize
+	thearch.lput(0xffffffff);	// cid.
 	cput(3);		// dwarf version (appendix F)
 	cput(0);		// augmentation ""
 	uleb128put(1);		// code_alignment_factor
@@ -1757,11 +1760,11 @@ writeframes(void)
 	uleb128put(FAKERETURNCOLUMN);	// return_address_register
 
 	cput(DW_CFA_def_cfa);
-	uleb128put(DWARFREGSP);	// register SP (**ABI-dependent, defined in l.h)
-	uleb128put(PtrSize);	// offset
+	uleb128put(thearch.dwarfregsp);	// register SP (**ABI-dependent, defined in l.h)
+	uleb128put(thearch.ptrsize);	// offset
 
 	cput(DW_CFA_offset + FAKERETURNCOLUMN);	 // return address
-	uleb128put(-PtrSize / DATAALIGNMENTFACTOR);  // at cfa - x*4
+	uleb128put(-thearch.ptrsize / DATAALIGNMENTFACTOR);  // at cfa - x*4
 
 	// 4 is to exclude the length field.
 	pad = CIERESERVE + frameo + 4 - cpos();
@@ -1778,8 +1781,8 @@ writeframes(void)
 
 		fdeo = cpos();
 		// Emit a FDE, Section 6.4.1, starting wit a placeholder.
-		LPUT(0);	// length, must be multiple of PtrSize
-		LPUT(0);	// Pointer to the CIE above, at offset 0
+		thearch.lput(0);	// length, must be multiple of thearch.ptrsize
+		thearch.lput(0);	// Pointer to the CIE above, at offset 0
 		addrput(0);	// initial location
 		addrput(0);	// address range
 
@@ -1792,23 +1795,23 @@ writeframes(void)
 				if(nextpc < pcsp.pc)
 					continue;
 			}
-			putpccfadelta(nextpc - pcsp.pc, PtrSize + pcsp.value);
+			putpccfadelta(nextpc - pcsp.pc, thearch.ptrsize + pcsp.value);
 		}
 
 		fdesize = cpos() - fdeo - 4;	// exclude the length field.
-		pad = rnd(fdesize, PtrSize) - fdesize;
+		pad = rnd(fdesize, thearch.ptrsize) - fdesize;
 		strnput("", pad);
 		fdesize += pad;
 
 		// Emit the FDE header for real, Section 6.4.1.
 		cseek(fdeo);
-		LPUT(fdesize);
+		thearch.lput(fdesize);
 		if(linkmode == LinkExternal) {
 			adddwarfrel(framesec, framesym, frameo, 4, 0);
-			adddwarfrel(framesec, s, frameo, PtrSize, 0);
+			adddwarfrel(framesec, s, frameo, thearch.ptrsize, 0);
 		}
 		else {
-			LPUT(0);
+			thearch.lput(0);
 			addrput(s->value);
 		}
 		addrput(s->size);
@@ -1834,11 +1837,11 @@ writeinfo(void)
 	vlong unitstart, here;
 
 	fwdcount = 0;
-	if (infosec == S)
+	if (infosec == nil)
 		infosec = linklookup(ctxt, ".dwarfinfo", 0);
 	infosec->nr = 0;
 
-	if(arangessec == S)
+	if(arangessec == nil)
 		arangessec = linklookup(ctxt, ".dwarfaranges", 0);
 	arangessec->nr = 0;
 
@@ -1848,22 +1851,22 @@ writeinfo(void)
 		// Write .debug_info Compilation Unit Header (sec 7.5.1)
 		// Fields marked with (*) must be changed for 64-bit dwarf
 		// This must match COMPUNITHEADERSIZE above.
-		LPUT(0);	// unit_length (*), will be filled in later.
-		WPUT(2);	// dwarf version (appendix F)
+		thearch.lput(0);	// unit_length (*), will be filled in later.
+		thearch.wput(2);	// dwarf version (appendix F)
 
 		// debug_abbrev_offset (*)
 		if(linkmode == LinkExternal)
 			adddwarfrel(infosec, abbrevsym, infoo, 4, 0);
 		else
-			LPUT(0);
+			thearch.lput(0);
 
-		cput(PtrSize);	// address_size
+		cput(thearch.ptrsize);	// address_size
 
 		putdie(compunit);
 
 		here = cpos();
 		cseek(unitstart);
-		LPUT(here - unitstart - 4);	// exclude the length field.
+		thearch.lput(here - unitstart - 4);	// exclude the length field.
 		cseek(here);
 	}
 	cflush();
@@ -1910,22 +1913,22 @@ writepub(int (*ispub)(DWDie*))
 			unitend = infoo + infosize;
 
 		// Write .debug_pubnames/types	Header (sec 6.1.1)
-		LPUT(0);			// unit_length (*), will be filled in later.
-		WPUT(2);			// dwarf version (appendix F)
-		LPUT(unitstart);		// debug_info_offset (of the Comp unit Header)
-		LPUT(unitend - unitstart);	// debug_info_length
+		thearch.lput(0);			// unit_length (*), will be filled in later.
+		thearch.wput(2);			// dwarf version (appendix F)
+		thearch.lput(unitstart);		// debug_info_offset (of the Comp unit Header)
+		thearch.lput(unitend - unitstart);	// debug_info_length
 
 		for (die = compunit->child; die != nil; die = die->link) {
 			if (!ispub(die)) continue;
-			LPUT(die->offs - unitstart);
+			thearch.lput(die->offs - unitstart);
 			dwa = getattr(die, DW_AT_name);
 			strnput(dwa->data, dwa->value + 1);
 		}
-		LPUT(0);
+		thearch.lput(0);
 
 		here = cpos();
 		cseek(sectionstart);
-		LPUT(here - sectionstart - 4);	// exclude the length field.
+		thearch.lput(here - sectionstart - 4);	// exclude the length field.
 		cseek(here);
 
 	}
@@ -1947,7 +1950,7 @@ writearanges(void)
 	vlong value;
 
 	sectionstart = cpos();
-	headersize = rnd(4+2+4+1+1, PtrSize);  // don't count unit_length field itself
+	headersize = rnd(4+2+4+1+1, thearch.ptrsize);  // don't count unit_length field itself
 
 	for (compunit = dwroot.child; compunit != nil; compunit = compunit->link) {
 		b = getattr(compunit,  DW_AT_low_pc);
@@ -1958,21 +1961,21 @@ writearanges(void)
 			continue;
 
 		// Write .debug_aranges	 Header + entry	 (sec 6.1.2)
-		LPUT(headersize + 4*PtrSize - 4);	// unit_length (*)
-		WPUT(2);	// dwarf version (appendix F)
+		thearch.lput(headersize + 4*thearch.ptrsize - 4);	// unit_length (*)
+		thearch.wput(2);	// dwarf version (appendix F)
 
 		value = compunit->offs - COMPUNITHEADERSIZE;	// debug_info_offset
 		if(linkmode == LinkExternal)
 			adddwarfrel(arangessec, infosym, sectionstart, 4, value);
 		else
-			LPUT(value);
+			thearch.lput(value);
 
-		cput(PtrSize);	// address_size
+		cput(thearch.ptrsize);	// address_size
 		cput(0);	// segment_size
-		strnput("", headersize - (4+2+4+1+1));	// align to PtrSize
+		strnput("", headersize - (4+2+4+1+1));	// align to thearch.ptrsize
 
 		if(linkmode == LinkExternal)
-			adddwarfrel(arangessec, (LSym*)b->data, sectionstart, PtrSize, b->value-((LSym*)b->data)->value);
+			adddwarfrel(arangessec, (LSym*)b->data, sectionstart, thearch.ptrsize, b->value-((LSym*)b->data)->value);
 		else
 			addrput(b->value);
 
@@ -2016,9 +2019,9 @@ writedwarfreloc(LSym* s)
 	start = cpos();
 	for(r = s->r; r < s->r+s->nr; r++) {
 		if(iself)
-			i = elfreloc1(r, r->off);
+			i = thearch.elfreloc1(r, r->off);
 		else if(HEADTYPE == Hdarwin)
-			i = machoreloc1(r, r->off);
+			i = thearch.machoreloc1(r, r->off);
 		else
 			i = -1;
 		if(i < 0)
@@ -2062,7 +2065,7 @@ dwarfemitdebugsections(void)
 
 	die = newdie(&dwtypes, DW_ABRV_BASETYPE, "uintptr");  // needed for array size
 	newattr(die, DW_AT_encoding,  DW_CLS_CONSTANT, DW_ATE_unsigned, 0);
-	newattr(die, DW_AT_byte_size, DW_CLS_CONSTANT, PtrSize, 0);
+	newattr(die, DW_AT_byte_size, DW_CLS_CONSTANT, thearch.ptrsize, 0);
 	newattr(die, DW_AT_go_kind, DW_CLS_CONSTANT, KindUintptr, 0);
 
 	// Needed by the prettyprinter code for interface inspection.
@@ -2196,7 +2199,7 @@ dwarfaddshstrings(LSym *shstrtab)
 	elfstrdbg[ElfStrDebugStr]      = addstring(shstrtab, ".debug_str");
 	elfstrdbg[ElfStrGDBScripts]    = addstring(shstrtab, ".debug_gdb_scripts");
 	if(linkmode == LinkExternal) {
-		if(thechar == '6' || thechar == '9') {
+		if(thearch.thechar == '6' || thearch.thechar == '9') {
 			elfstrdbg[ElfStrRelDebugInfo] = addstring(shstrtab, ".rela.debug_info");
 			elfstrdbg[ElfStrRelDebugAranges] = addstring(shstrtab, ".rela.debug_aranges");
 			elfstrdbg[ElfStrRelDebugLine] = addstring(shstrtab, ".rela.debug_line");
@@ -2251,17 +2254,17 @@ dwarfaddelfrelocheader(int elfstr, ElfShdr *shdata, vlong off, vlong size)
 	ElfShdr *sh;
 
 	sh = newElfShdr(elfstrdbg[elfstr]);
-	if(thechar == '6' || thechar == '9') {
+	if(thearch.thechar == '6' || thearch.thechar == '9') {
 		sh->type = SHT_RELA;
 	} else {
 		sh->type = SHT_REL;
 	}
-	sh->entsize = PtrSize*(2+(sh->type==SHT_RELA));
+	sh->entsize = thearch.ptrsize*(2+(sh->type==SHT_RELA));
 	sh->link = elfshname(".symtab")->shnum;
 	sh->info = shdata->shnum;
 	sh->off = off;
 	sh->size = size;
-	sh->addralign = PtrSize;
+	sh->addralign = thearch.ptrsize;
 	
 }
 
diff --git a/src/cmd/ld/elf.c b/src/cmd/ld/elf.c
index 89a0a5e87f..12ced98107 100644
--- a/src/cmd/ld/elf.c
+++ b/src/cmd/ld/elf.c
@@ -2,9 +2,12 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-#include	"l.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
 #include	"lib.h"
-#include	"../ld/elf.h"
+#include	"elf.h"
 
 /*
  * We use the 64-bit data structures on both 32- and 64-bit machines
@@ -15,6 +18,8 @@
 
 int	iself;
 
+int	nelfsym = 1;
+
 static	int	elf64;
 static	ElfEhdr	hdr;
 static	ElfPhdr	*phdr[NSECT];
@@ -42,7 +47,7 @@ elfinit(void)
 {
 	iself = 1;
 
-	switch(thechar) {
+	switch(thearch.thechar) {
 	// 64-bit architectures
 	case '9':
 		if(ctxt->arch->endian == BigEndian)
@@ -77,14 +82,14 @@ elfinit(void)
 void
 elf64phdr(ElfPhdr *e)
 {
-	LPUT(e->type);
-	LPUT(e->flags);
-	VPUT(e->off);
-	VPUT(e->vaddr);
-	VPUT(e->paddr);
-	VPUT(e->filesz);
-	VPUT(e->memsz);
-	VPUT(e->align);
+	thearch.lput(e->type);
+	thearch.lput(e->flags);
+	thearch.vput(e->off);
+	thearch.vput(e->vaddr);
+	thearch.vput(e->paddr);
+	thearch.vput(e->filesz);
+	thearch.vput(e->memsz);
+	thearch.vput(e->align);
 }
 
 void
@@ -103,44 +108,44 @@ elf32phdr(ElfPhdr *e)
 		e->filesz += frag;
 		e->memsz += frag;
 	}
-	LPUT(e->type);
-	LPUT(e->off);
-	LPUT(e->vaddr);
-	LPUT(e->paddr);
-	LPUT(e->filesz);
-	LPUT(e->memsz);
-	LPUT(e->flags);
-	LPUT(e->align);
+	thearch.lput(e->type);
+	thearch.lput(e->off);
+	thearch.lput(e->vaddr);
+	thearch.lput(e->paddr);
+	thearch.lput(e->filesz);
+	thearch.lput(e->memsz);
+	thearch.lput(e->flags);
+	thearch.lput(e->align);
 }
 
 void
 elf64shdr(ElfShdr *e)
 {
-	LPUT(e->name);
-	LPUT(e->type);
-	VPUT(e->flags);
-	VPUT(e->addr);
-	VPUT(e->off);
-	VPUT(e->size);
-	LPUT(e->link);
-	LPUT(e->info);
-	VPUT(e->addralign);
-	VPUT(e->entsize);
+	thearch.lput(e->name);
+	thearch.lput(e->type);
+	thearch.vput(e->flags);
+	thearch.vput(e->addr);
+	thearch.vput(e->off);
+	thearch.vput(e->size);
+	thearch.lput(e->link);
+	thearch.lput(e->info);
+	thearch.vput(e->addralign);
+	thearch.vput(e->entsize);
 }
 
 void
 elf32shdr(ElfShdr *e)
 {
-	LPUT(e->name);
-	LPUT(e->type);
-	LPUT(e->flags);
-	LPUT(e->addr);
-	LPUT(e->off);
-	LPUT(e->size);
-	LPUT(e->link);
-	LPUT(e->info);
-	LPUT(e->addralign);
-	LPUT(e->entsize);
+	thearch.lput(e->name);
+	thearch.lput(e->type);
+	thearch.lput(e->flags);
+	thearch.lput(e->addr);
+	thearch.lput(e->off);
+	thearch.lput(e->size);
+	thearch.lput(e->link);
+	thearch.lput(e->info);
+	thearch.lput(e->addralign);
+	thearch.lput(e->entsize);
 }
 
 uint32
@@ -231,19 +236,19 @@ elf64writehdr(void)
 
 	for (i = 0; i < EI_NIDENT; i++)
 		cput(hdr.ident[i]);
-	WPUT(hdr.type);
-	WPUT(hdr.machine);
-	LPUT(hdr.version);
-	VPUT(hdr.entry);
-	VPUT(hdr.phoff);
-	VPUT(hdr.shoff);
-	LPUT(hdr.flags);
-	WPUT(hdr.ehsize);
-	WPUT(hdr.phentsize);
-	WPUT(hdr.phnum);
-	WPUT(hdr.shentsize);
-	WPUT(hdr.shnum);
-	WPUT(hdr.shstrndx);
+	thearch.wput(hdr.type);
+	thearch.wput(hdr.machine);
+	thearch.lput(hdr.version);
+	thearch.vput(hdr.entry);
+	thearch.vput(hdr.phoff);
+	thearch.vput(hdr.shoff);
+	thearch.lput(hdr.flags);
+	thearch.wput(hdr.ehsize);
+	thearch.wput(hdr.phentsize);
+	thearch.wput(hdr.phnum);
+	thearch.wput(hdr.shentsize);
+	thearch.wput(hdr.shnum);
+	thearch.wput(hdr.shstrndx);
 	return ELF64HDRSIZE;
 }
 
@@ -254,19 +259,19 @@ elf32writehdr(void)
 
 	for (i = 0; i < EI_NIDENT; i++)
 		cput(hdr.ident[i]);
-	WPUT(hdr.type);
-	WPUT(hdr.machine);
-	LPUT(hdr.version);
-	LPUT(hdr.entry);
-	LPUT(hdr.phoff);
-	LPUT(hdr.shoff);
-	LPUT(hdr.flags);
-	WPUT(hdr.ehsize);
-	WPUT(hdr.phentsize);
-	WPUT(hdr.phnum);
-	WPUT(hdr.shentsize);
-	WPUT(hdr.shnum);
-	WPUT(hdr.shstrndx);
+	thearch.wput(hdr.type);
+	thearch.wput(hdr.machine);
+	thearch.lput(hdr.version);
+	thearch.lput(hdr.entry);
+	thearch.lput(hdr.phoff);
+	thearch.lput(hdr.shoff);
+	thearch.lput(hdr.flags);
+	thearch.wput(hdr.ehsize);
+	thearch.wput(hdr.phentsize);
+	thearch.wput(hdr.phnum);
+	thearch.wput(hdr.shentsize);
+	thearch.wput(hdr.shnum);
+	thearch.wput(hdr.shstrndx);
 	return ELF32HDRSIZE;
 }
 
@@ -381,9 +386,9 @@ elfwritenotehdr(char *str, uint32 namesz, uint32 descsz, uint32 tag)
 
 	// Write Elf_Note header.
 	cseek(sh->off);
-	LPUT(namesz);
-	LPUT(descsz);
-	LPUT(tag);
+	thearch.lput(namesz);
+	thearch.lput(descsz);
+	thearch.lput(tag);
 
 	return sh;
 }
@@ -416,7 +421,7 @@ elfwritenetbsdsig(void)
 
 	// Followed by NetBSD string and version.
 	cwrite(ELF_NOTE_NETBSD_NAME, ELF_NOTE_NETBSD_NAMESZ + 1);
-	LPUT(ELF_NOTE_NETBSD_VERSION);
+	thearch.lput(ELF_NOTE_NETBSD_VERSION);
 
 	return sh->size;
 }
@@ -449,7 +454,7 @@ elfwriteopenbsdsig(void)
 
 	// Followed by OpenBSD string and version.
 	cwrite(ELF_NOTE_OPENBSD_NAME, ELF_NOTE_OPENBSD_NAMESZ);
-	LPUT(ELF_NOTE_OPENBSD_VERSION);
+	thearch.lput(ELF_NOTE_OPENBSD_VERSION);
 
 	return sh->size;
 }
@@ -525,7 +530,6 @@ elfwritebuildinfo(void)
 	return sh->size;
 }
 
-extern int nelfsym;
 int elfverneed;
 
 typedef struct Elfaux Elfaux;
@@ -610,7 +614,7 @@ elfdynhash(void)
 	memset(need, 0, nsym * sizeof need[0]);
 	memset(chain, 0, nsym * sizeof chain[0]);
 	memset(buckets, 0, nbucket * sizeof buckets[0]);
-	for(sy=ctxt->allsym; sy!=S; sy=sy->allsym) {
+	for(sy=ctxt->allsym; sy!=nil; sy=sy->allsym) {
 		if (sy->dynid <= 0)
 			continue;
 
@@ -690,7 +694,7 @@ elfdynhash(void)
 		elfwritedynentsym(s, DT_VERSYM, linklookup(ctxt, ".gnu.version", 0));
 	}
 
-	if(thechar == '6' || thechar == '9') {
+	if(thearch.thechar == '6' || thearch.thechar == '9') {
 		sy = linklookup(ctxt, ".rela.plt", 0);
 		if(sy->size > 0) {
 			elfwritedynent(s, DT_PLTREL, DT_RELA);
@@ -816,7 +820,7 @@ elfshreloc(Section *sect)
 	if(strcmp(sect->name, ".shstrtab") == 0 || strcmp(sect->name, ".tbss") == 0)
 		return nil;
 
-	if(thechar == '6' || thechar == '9') {
+	if(thearch.thechar == '6' || thearch.thechar == '9') {
 		prefix = ".rela";
 		typ = SHT_RELA;
 	} else {
@@ -827,12 +831,12 @@ elfshreloc(Section *sect)
 	snprint(buf, sizeof buf, "%s%s", prefix, sect->name);
 	sh = elfshname(buf);
 	sh->type = typ;
-	sh->entsize = RegSize*(2+(typ==SHT_RELA));
+	sh->entsize = thearch.regsize*(2+(typ==SHT_RELA));
 	sh->link = elfshname(".symtab")->shnum;
 	sh->info = sect->elfsect->shnum;
 	sh->off = sect->reloff;
 	sh->size = sect->rellen;
-	sh->addralign = RegSize;
+	sh->addralign = thearch.regsize;
 	return sh;
 }
 
@@ -875,7 +879,7 @@ elfrelocsect(Section *sect, LSym *first)
 			}
 			if(r->xsym->elfsym == 0)
 				diag("reloc %d to non-elf symbol %s (outer=%s) %d", r->type, r->sym->name, r->xsym->name, r->sym->type);
-			if(elfreloc1(r, sym->value+r->off - sect->vaddr) < 0)
+			if(thearch.elfreloc1(r, sym->value+r->off - sect->vaddr) < 0)
 				diag("unsupported obj reloc %d/%d to %s", r->type, r->siz, r->sym->name);
 		}
 	}
@@ -943,7 +947,7 @@ doelf(void)
 		debug['s'] = 0;
 		debug['d'] = 1;
 
-		if(thechar == '6' || thechar == '9') {
+		if(thearch.thechar == '6' || thearch.thechar == '9') {
 			addstring(shstrtab, ".rela.text");
 			addstring(shstrtab, ".rela.rodata");
 			addstring(shstrtab, ".rela.typelink");
@@ -966,7 +970,7 @@ doelf(void)
 
 	if(flag_shared) {
 		addstring(shstrtab, ".init_array");
-		if(thechar == '6' || thechar == '9')
+		if(thearch.thechar == '6' || thearch.thechar == '9')
 			addstring(shstrtab, ".rela.init_array");
 		else
 			addstring(shstrtab, ".rel.init_array");
@@ -983,13 +987,13 @@ doelf(void)
 		addstring(shstrtab, ".interp");
 		addstring(shstrtab, ".hash");
 		addstring(shstrtab, ".got");
-		if(thechar == '9')
+		if(thearch.thechar == '9')
 			addstring(shstrtab, ".glink");
 		addstring(shstrtab, ".got.plt");
 		addstring(shstrtab, ".dynamic");
 		addstring(shstrtab, ".dynsym");
 		addstring(shstrtab, ".dynstr");
-		if(thechar == '6' || thechar == '9') {
+		if(thearch.thechar == '6' || thearch.thechar == '9') {
 			addstring(shstrtab, ".rela");
 			addstring(shstrtab, ".rela.plt");
 		} else {
@@ -1004,7 +1008,7 @@ doelf(void)
 		s = linklookup(ctxt, ".dynsym", 0);
 		s->type = SELFROSECT;
 		s->reachable = 1;
-		if(thechar == '6' || thechar == '9')
+		if(thearch.thechar == '6' || thearch.thechar == '9')
 			s->size += ELF64SYMSIZE;
 		else
 			s->size += ELF32SYMSIZE;
@@ -1018,7 +1022,7 @@ doelf(void)
 		dynstr = s;
 
 		/* relocation table */
-		if(thechar == '6' || thechar == '9')
+		if(thearch.thechar == '6' || thearch.thechar == '9')
 			s = linklookup(ctxt, ".rela", 0);
 		else
 			s = linklookup(ctxt, ".rel", 0);
@@ -1031,7 +1035,7 @@ doelf(void)
 		s->type = SELFGOT; // writable
 
 		/* ppc64 glink resolver */
-		if(thechar == '9') {
+		if(thearch.thechar == '9') {
 			s = linklookup(ctxt, ".glink", 0);
 			s->reachable = 1;
 			s->type = SELFRXSECT;
@@ -1048,16 +1052,16 @@ doelf(void)
 
 		s = linklookup(ctxt, ".plt", 0);
 		s->reachable = 1;
-		if(thechar == '9')
+		if(thearch.thechar == '9')
 			// In the ppc64 ABI, .plt is a data section
 			// written by the dynamic linker.
 			s->type = SELFSECT;
 		else
 			s->type = SELFRXSECT;
 		
-		elfsetupplt();
+		thearch.elfsetupplt();
 		
-		if(thechar == '6' || thechar == '9')
+		if(thearch.thechar == '6' || thearch.thechar == '9')
 			s = linklookup(ctxt, ".rela.plt", 0);
 		else
 			s = linklookup(ctxt, ".rel.plt", 0);
@@ -1082,13 +1086,13 @@ doelf(void)
 		 */
 		elfwritedynentsym(s, DT_HASH, linklookup(ctxt, ".hash", 0));
 		elfwritedynentsym(s, DT_SYMTAB, linklookup(ctxt, ".dynsym", 0));
-		if(thechar == '6' || thechar == '9')
+		if(thearch.thechar == '6' || thearch.thechar == '9')
 			elfwritedynent(s, DT_SYMENT, ELF64SYMSIZE);
 		else
 			elfwritedynent(s, DT_SYMENT, ELF32SYMSIZE);
 		elfwritedynentsym(s, DT_STRTAB, linklookup(ctxt, ".dynstr", 0));
 		elfwritedynentsymsize(s, DT_STRSZ, linklookup(ctxt, ".dynstr", 0));
-		if(thechar == '6' || thechar == '9') {
+		if(thearch.thechar == '6' || thearch.thechar == '9') {
 			elfwritedynentsym(s, DT_RELA, linklookup(ctxt, ".rela", 0));
 			elfwritedynentsymsize(s, DT_RELASZ, linklookup(ctxt, ".rela", 0));
 			elfwritedynent(s, DT_RELAENT, ELF64RELASIZE);
@@ -1100,12 +1104,12 @@ doelf(void)
 		if(rpath)
 			elfwritedynent(s, DT_RUNPATH, addstring(dynstr, rpath));
 
-		if(thechar == '9')
+		if(thearch.thechar == '9')
 			elfwritedynentsym(s, DT_PLTGOT, linklookup(ctxt, ".plt", 0));
 		else
 			elfwritedynentsym(s, DT_PLTGOT, linklookup(ctxt, ".got.plt", 0));
 
-		if(thechar == '9')
+		if(thearch.thechar == '9')
 			elfwritedynent(s, DT_PPC64_OPT, 0);
 
 		// Solaris dynamic linker can't handle an empty .rela.plt if
@@ -1167,7 +1171,7 @@ asmbelf(vlong symo)
 	Section *sect;
 
 	eh = getElfEhdr();
-	switch(thechar) {
+	switch(thearch.thechar) {
 	default:
 		diag("unknown architecture in asmbelf");
 		errorexit();
@@ -1228,22 +1232,22 @@ asmbelf(vlong symo)
 		if(interpreter == nil) {
 			switch(HEADTYPE) {
 			case Hlinux:
-				interpreter = linuxdynld;
+				interpreter = thearch.linuxdynld;
 				break;
 			case Hfreebsd:
-				interpreter = freebsddynld;
+				interpreter = thearch.freebsddynld;
 				break;
 			case Hnetbsd:
-				interpreter = netbsddynld;
+				interpreter = thearch.netbsddynld;
 				break;
 			case Hopenbsd:
-				interpreter = openbsddynld;
+				interpreter = thearch.openbsddynld;
 				break;
 			case Hdragonfly:
-				interpreter = dragonflydynld;
+				interpreter = thearch.dragonflydynld;
 				break;
 			case Hsolaris:
-				interpreter = solarisdynld;
+				interpreter = thearch.solarisdynld;
 				break;
 			}
 		}
@@ -1304,7 +1308,7 @@ asmbelf(vlong symo)
 			sh->entsize = ELF64SYMSIZE;
 		else
 			sh->entsize = ELF32SYMSIZE;
-		sh->addralign = RegSize;
+		sh->addralign = thearch.regsize;
 		sh->link = elfshname(".dynstr")->shnum;
 		// sh->info = index of first non-local symbol (number of local symbols)
 		shsym(sh, linklookup(ctxt, ".dynsym", 0));
@@ -1327,7 +1331,7 @@ asmbelf(vlong symo)
 			sh = elfshname(".gnu.version_r");
 			sh->type = SHT_GNU_VERNEED;
 			sh->flags = SHF_ALLOC;
-			sh->addralign = RegSize;
+			sh->addralign = thearch.regsize;
 			sh->info = elfverneed;
 			sh->link = elfshname(".dynstr")->shnum;
 			shsym(sh, linklookup(ctxt, ".gnu.version_r", 0));
@@ -1340,7 +1344,7 @@ asmbelf(vlong symo)
 			sh->type = SHT_RELA;
 			sh->flags = SHF_ALLOC;
 			sh->entsize = ELF64RELASIZE;
-			sh->addralign = RegSize;
+			sh->addralign = thearch.regsize;
 			sh->link = elfshname(".dynsym")->shnum;
 			sh->info = elfshname(".plt")->shnum;
 			shsym(sh, linklookup(ctxt, ".rela.plt", 0));
@@ -1402,15 +1406,15 @@ asmbelf(vlong symo)
 			sh = elfshname(".got");
 			sh->type = SHT_PROGBITS;
 			sh->flags = SHF_ALLOC+SHF_WRITE;
-			sh->entsize = RegSize;
-			sh->addralign = RegSize;
+			sh->entsize = thearch.regsize;
+			sh->addralign = thearch.regsize;
 			shsym(sh, linklookup(ctxt, ".got", 0));
 
 			sh = elfshname(".got.plt");
 			sh->type = SHT_PROGBITS;
 			sh->flags = SHF_ALLOC+SHF_WRITE;
-			sh->entsize = RegSize;
-			sh->addralign = RegSize;
+			sh->entsize = thearch.regsize;
+			sh->addralign = thearch.regsize;
 			shsym(sh, linklookup(ctxt, ".got.plt", 0));
 		}
 		
@@ -1418,7 +1422,7 @@ asmbelf(vlong symo)
 		sh->type = SHT_HASH;
 		sh->flags = SHF_ALLOC;
 		sh->entsize = 4;
-		sh->addralign = RegSize;
+		sh->addralign = thearch.regsize;
 		sh->link = elfshname(".dynsym")->shnum;
 		shsym(sh, linklookup(ctxt, ".hash", 0));
 
@@ -1426,8 +1430,8 @@ asmbelf(vlong symo)
 		sh = elfshname(".dynamic");
 		sh->type = SHT_DYNAMIC;
 		sh->flags = SHF_ALLOC+SHF_WRITE;
-		sh->entsize = 2*RegSize;
-		sh->addralign = RegSize;
+		sh->entsize = 2*thearch.regsize;
+		sh->addralign = thearch.regsize;
 		sh->link = elfshname(".dynstr")->shnum;
 		shsym(sh, linklookup(ctxt, ".dynamic", 0));
 		ph = newElfPhdr();
@@ -1446,7 +1450,7 @@ asmbelf(vlong symo)
 			ph->type = PT_TLS;
 			ph->flags = PF_R;
 			ph->memsz = -ctxt->tlsoffset;
-			ph->align = RegSize;
+			ph->align = thearch.regsize;
 		}
 	}
 
@@ -1454,12 +1458,12 @@ asmbelf(vlong symo)
 		ph = newElfPhdr();
 		ph->type = PT_GNU_STACK;
 		ph->flags = PF_W+PF_R;
-		ph->align = RegSize;
+		ph->align = thearch.regsize;
 		
 		ph = newElfPhdr();
 		ph->type = PT_PAX_FLAGS;
 		ph->flags = 0x2a00; // mprotect, randexec, emutramp disabled
-		ph->align = RegSize;
+		ph->align = thearch.regsize;
 	}
 
 elfobj:
@@ -1501,7 +1505,7 @@ elfobj:
 	if(linkmode == LinkInternal && !debug['d'] && HEADTYPE != Hopenbsd) {
 		sh = elfshname(".tbss");
 		sh->type = SHT_NOBITS;
-		sh->addralign = RegSize;
+		sh->addralign = thearch.regsize;
 		sh->size = -ctxt->tlsoffset;
 		sh->flags = SHF_ALLOC | SHF_TLS | SHF_WRITE;
 	}
@@ -1511,8 +1515,8 @@ elfobj:
 		sh->type = SHT_SYMTAB;
 		sh->off = symo;
 		sh->size = symsize;
-		sh->addralign = RegSize;
-		sh->entsize = 8+2*RegSize;
+		sh->addralign = thearch.regsize;
+		sh->entsize = 8+2*thearch.regsize;
 		sh->link = elfshname(".strtab")->shnum;
 		sh->info = elfglobalsymndx;
 
diff --git a/src/cmd/ld/elf.h b/src/cmd/ld/elf.h
index 16c052fcb5..85589bd70f 100644
--- a/src/cmd/ld/elf.h
+++ b/src/cmd/ld/elf.h
@@ -1021,12 +1021,6 @@ void	dwarfaddelfsectionsyms(void);
 void	dwarfaddelfheaders(void);
 void	asmbelf(vlong symo);
 void	asmbelfsetup(void);
-extern char linuxdynld[];
-extern char freebsddynld[];
-extern char netbsddynld[];
-extern char openbsddynld[];
-extern char dragonflydynld[];
-extern char solarisdynld[];
 int	elfreloc1(Reloc*, vlong sectoff);
 void	putelfsectionsyms(void);
 
diff --git a/src/cmd/ld/go.c b/src/cmd/ld/go.c
index 9c296b740f..8b1123d3de 100644
--- a/src/cmd/ld/go.c
+++ b/src/cmd/ld/go.c
@@ -4,8 +4,11 @@
 
 // go-specific code shared across loaders (5l, 6l, 8l).
 
-#include	"l.h"
-#include	"../ld/lib.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
+#include	"lib.h"
 
 // accumulate all type information from .6 files.
 // check for inconsistencies.
@@ -401,7 +404,7 @@ loadcgo(char *file, char *pkg, char *p, int n)
 				// allow #pragma dynimport _ _ "foo.so"
 				// to force a link of foo.so.
 				havedynamic = 1;
-				adddynlib(lib);
+				thearch.adddynlib(lib);
 				continue;
 			}
 
@@ -518,7 +521,7 @@ static LSym *emarkq;
 static void
 mark1(LSym *s, LSym *parent)
 {
-	if(s == S || s->reachable)
+	if(s == nil || s->reachable)
 		return;
 	if(strncmp(s->name, "go.weak.", 8) == 0)
 		return;
@@ -544,7 +547,7 @@ markflood(void)
 	LSym *s;
 	int i;
 	
-	for(s=markq; s!=S; s=s->queue) {
+	for(s=markq; s!=nil; s=s->queue) {
 		if(s->type == STEXT) {
 			if(debug['v'] > 1)
 				Bprint(&bso, "marktext %s\n", s->name);
@@ -608,7 +611,7 @@ deadcode(void)
 	markflood();
 	
 	// keep each beginning with 'typelink.' if the symbol it points at is being kept.
-	for(s = ctxt->allsym; s != S; s = s->allsym) {
+	for(s = ctxt->allsym; s != nil; s = s->allsym) {
 		if(strncmp(s->name, "go.typelink.", 12) == 0)
 			s->reachable = s->nr==1 && s->r[0].sym->reachable;
 	}
@@ -630,7 +633,7 @@ deadcode(void)
 	else
 		last->next = nil;
 	
-	for(s = ctxt->allsym; s != S; s = s->allsym)
+	for(s = ctxt->allsym; s != nil; s = s->allsym)
 		if(strncmp(s->name, "go.weak.", 8) == 0) {
 			s->special = 1;  // do not lay out in data segment
 			s->reachable = 1;
@@ -639,7 +642,7 @@ deadcode(void)
 	
 	// record field tracking references
 	fmtstrinit(&fmt);
-	for(s = ctxt->allsym; s != S; s = s->allsym) {
+	for(s = ctxt->allsym; s != nil; s = s->allsym) {
 		if(strncmp(s->name, "go.track.", 9) == 0) {
 			s->special = 1;  // do not lay out in data segment
 			s->hide = 1;
@@ -668,7 +671,7 @@ doweak(void)
 
 	// resolve weak references only if
 	// target symbol will be in binary anyway.
-	for(s = ctxt->allsym; s != S; s = s->allsym) {
+	for(s = ctxt->allsym; s != nil; s = s->allsym) {
 		if(strncmp(s->name, "go.weak.", 8) == 0) {
 			t = linkrlookup(ctxt, s->name+8, s->version);
 			if(t && t->type != 0 && t->reachable) {
@@ -693,7 +696,7 @@ addexport(void)
 		return;
 
 	for(i=0; i<ndynexp; i++)
-		adddynsym(ctxt, dynexp[i]);
+		thearch.adddynsym(ctxt, dynexp[i]);
 }
 
 /* %Z from gc, for quoting import paths */
diff --git a/src/cmd/ld/ldelf.c b/src/cmd/ld/ldelf.c
index 29d32d283d..7beff337f2 100644
--- a/src/cmd/ld/ldelf.c
+++ b/src/cmd/ld/ldelf.c
@@ -25,9 +25,12 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 THE SOFTWARE.
 */
 
-#include	"l.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
 #include	"lib.h"
-#include	"../ld/elf.h"
+#include	"elf.h"
 
 enum
 {
@@ -414,7 +417,7 @@ ldelf(Biobuf *f, char *pkg, int64 len, char *pn)
 		return;
 	}
 
-	switch(thechar) {
+	switch(thearch.thechar) {
 	default:
 		diag("%s: elf %s unimplemented", pn, thestring);
 		return;
@@ -590,7 +593,7 @@ ldelf(Biobuf *f, char *pkg, int64 len, char *pn)
 		if(sym.shndx >= obj->nsect || sym.shndx == 0)
 			continue;
 		// even when we pass needSym == 1 to readsym, it might still return nil to skip some unwanted symbols
-		if(sym.sym == S)
+		if(sym.sym == nil)
 			continue;
 		sect = obj->sect+sym.shndx;
 		if(sect->sym == nil) {
@@ -600,7 +603,7 @@ ldelf(Biobuf *f, char *pkg, int64 len, char *pn)
 			continue;
 		}
 		s = sym.sym;
-		if(s->outer != S) {
+		if(s->outer != nil) {
 			if(s->dupok)
 				continue;
 			diag("%s: duplicate symbol reference: %s in both %s and %s", pn, s->name, s->outer->name, sect->sym->name);
@@ -632,7 +635,7 @@ ldelf(Biobuf *f, char *pkg, int64 len, char *pn)
 	// This keeps textp in increasing address order.
 	for(i=0; i<obj->nsect; i++) {
 		s = obj->sect[i].sym;
-		if(s == S)
+		if(s == nil)
 			continue;
 		if(s->sub)
 			s->sub = listsort(s->sub, valuecmp, offsetof(LSym, sub));
@@ -645,7 +648,7 @@ ldelf(Biobuf *f, char *pkg, int64 len, char *pn)
 			else
 				ctxt->textp = s;
 			ctxt->etextp = s;
-			for(s = s->sub; s != S; s = s->sub) {
+			for(s = s->sub; s != nil; s = s->sub) {
 				if(s->onlist)
 					sysfatal("symbol %s listed multiple times", s->name);
 				s->onlist = 1;
@@ -700,7 +703,7 @@ ldelf(Biobuf *f, char *pkg, int64 len, char *pn)
 				continue;
 			}
 			if((info >> 32) == 0) { // absolute relocation, don't bother reading the null symbol
-				rp->sym = S;
+				rp->sym = nil;
 			} else {
 				if(readsym(obj, info>>32, &sym, 0) < 0)
 					goto bad;
@@ -844,7 +847,7 @@ readsym(ElfObj *obj, int i, ElfSym *sym, int needSym)
 			}
 			break;
 		case ElfSymBindLocal:
-			if(thechar == '5' && (strncmp(sym->name, "$a", 2) == 0 || strncmp(sym->name, "$d", 2) == 0)) {
+			if(thearch.thechar == '5' && (strncmp(sym->name, "$a", 2) == 0 || strncmp(sym->name, "$d", 2) == 0)) {
 				// binutils for arm generate these mapping
 				// symbols, ignore these
 				break;
@@ -905,7 +908,7 @@ rbyoff(const void *va, const void *vb)
 static int
 reltype(char *pn, int elftype, uchar *siz)
 {
-	switch(R(thechar, elftype)) {
+	switch(R(thearch.thechar, elftype)) {
 	default:
 		diag("%s: unknown relocation type %d; compiled without -fpic?", pn, elftype);
 	case R('9', R_PPC64_TOC16):
diff --git a/src/cmd/ld/ldmacho.c b/src/cmd/ld/ldmacho.c
index 71cfa63dec..4cbfa8d1ce 100644
--- a/src/cmd/ld/ldmacho.c
+++ b/src/cmd/ld/ldmacho.c
@@ -25,7 +25,10 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 THE SOFTWARE.
 */
 
-#include	"l.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
 #include	"lib.h"
 
 enum {
@@ -477,7 +480,7 @@ ldmacho(Biobuf *f, char *pkg, int64 len, char *pn)
 	m->len = len;
 	m->name = pn;
 	
-	switch(thechar) {
+	switch(thearch.thechar) {
 	default:
 		diag("%s: mach-o %s unimplemented", pn, thestring);
 		return;
@@ -627,7 +630,7 @@ ldmacho(Biobuf *f, char *pkg, int64 len, char *pn)
 			werrstr("reference to invalid section %s/%s", sect->segname, sect->name);
 			continue;
 		}
-		if(s->outer != S) {
+		if(s->outer != nil) {
 			if(s->dupok)
 				continue;
 			diag("%s: duplicate symbol reference: %s in both %s and %s", pn, s->name, s->outer->name, sect->sym->name);
@@ -652,13 +655,13 @@ ldmacho(Biobuf *f, char *pkg, int64 len, char *pn)
 	// This keeps textp in increasing address order.
 	for(i=0; i<c->seg.nsect; i++) {
 		sect = &c->seg.sect[i];
-		if((s = sect->sym) == S)
+		if((s = sect->sym) == nil)
 			continue;
 		if(s->sub) {
 			s->sub = listsort(s->sub, valuecmp, offsetof(LSym, sub));
 			
 			// assign sizes, now that we know symbols in sorted order.
-			for(s1 = s->sub; s1 != S; s1 = s1->sub) {
+			for(s1 = s->sub; s1 != nil; s1 = s1->sub) {
 				if(s1->sub)
 					s1->size = s1->sub->value - s1->value;
 				else
@@ -674,7 +677,7 @@ ldmacho(Biobuf *f, char *pkg, int64 len, char *pn)
 			else
 				ctxt->textp = s;
 			ctxt->etextp = s;
-			for(s1 = s->sub; s1 != S; s1 = s1->sub) {
+			for(s1 = s->sub; s1 != nil; s1 = s1->sub) {
 				if(s1->onlist)
 					sysfatal("symbol %s listed multiple times", s1->name);
 				s1->onlist = 1;
@@ -687,7 +690,7 @@ ldmacho(Biobuf *f, char *pkg, int64 len, char *pn)
 	// load relocations
 	for(i=0; i<c->seg.nsect; i++) {
 		sect = &c->seg.sect[i];
-		if((s = sect->sym) == S)
+		if((s = sect->sym) == nil)
 			continue;
 		macholoadrel(m, sect);
 		if(sect->rel == nil)
@@ -700,7 +703,7 @@ ldmacho(Biobuf *f, char *pkg, int64 len, char *pn)
 				int k;
 				MachoSect *ks;
 
-				if(thechar != '8') {
+				if(thearch.thechar != '8') {
 					// mach-o only uses scattered relocation on 32-bit platforms
 					diag("unexpected scattered relocation");
 					continue;
@@ -750,7 +753,7 @@ ldmacho(Biobuf *f, char *pkg, int64 len, char *pn)
 				werrstr("unsupported scattered relocation: invalid address %#ux", rel->addr);
 				goto bad;
 			foundk:
-				if(ks->sym != S) {
+				if(ks->sym != nil) {
 					rp->sym = ks->sym;
 					rp->add += rel->value - ks->addr;
 				} else if(strcmp(ks->segname, "__IMPORT") == 0 && strcmp(ks->name, "__pointers") == 0) {
@@ -788,7 +791,7 @@ ldmacho(Biobuf *f, char *pkg, int64 len, char *pn)
 			rp->off = rel->addr;
 
 			// Handle X86_64_RELOC_SIGNED referencing a section (rel->extrn == 0).
-			if (thechar == '6' && rel->extrn == 0 && rel->type == 1) {
+			if (thearch.thechar == '6' && rel->extrn == 0 && rel->type == 1) {
 				// Calculate the addend as the offset into the section.
 				//
 				// The rip-relative offset stored in the object file is encoded
@@ -812,7 +815,7 @@ ldmacho(Biobuf *f, char *pkg, int64 len, char *pn)
 			// For i386 Mach-O PC-relative, the addend is written such that
 			// it *is* the PC being subtracted.  Use that to make
 			// it match our version of PC-relative.
-			if(rel->pcrel && thechar == '8')
+			if(rel->pcrel && thearch.thechar == '8')
 				rp->add += rp->off+rp->siz;
 			if(!rel->extrn) {
 				if(rel->symnum < 1 || rel->symnum > c->seg.nsect) {
@@ -828,7 +831,7 @@ ldmacho(Biobuf *f, char *pkg, int64 len, char *pn)
 				// include that information in the addend.
 				// We only care about the delta from the 
 				// section base.
-				if(thechar == '8')
+				if(thearch.thechar == '8')
 					rp->add -= c->seg.sect[rel->symnum-1].addr;
 			} else {
 				if(rel->symnum >= symtab->nsym) {
diff --git a/src/cmd/ld/ldpe.c b/src/cmd/ld/ldpe.c
index 4f5e51f2f1..da11a3ea18 100644
--- a/src/cmd/ld/ldpe.c
+++ b/src/cmd/ld/ldpe.c
@@ -2,9 +2,12 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-#include	"l.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
 #include	"lib.h"
-#include	"../ld/pe.h"
+#include	"pe.h"
 
 #define IMAGE_SCN_MEM_DISCARDABLE 0x2000000
 
@@ -363,7 +366,7 @@ ldpe(Biobuf *f, char *pkg, int64 len, char *pn)
 		if(sect == nil) 
 			return;
 
-		if(s->outer != S) {
+		if(s->outer != nil) {
 			if(s->dupok)
 				continue;
 			diag("%s: duplicate symbol reference: %s in both %s and %s", pn, s->name, s->outer->name, sect->sym->name);
@@ -386,7 +389,7 @@ ldpe(Biobuf *f, char *pkg, int64 len, char *pn)
 	// This keeps textp in increasing address order.
 	for(i=0; i<obj->nsect; i++) {
 		s = obj->sect[i].sym;
-		if(s == S)
+		if(s == nil)
 			continue;
 		if(s->sub)
 			s->sub = listsort(s->sub, valuecmp, offsetof(LSym, sub));
@@ -399,7 +402,7 @@ ldpe(Biobuf *f, char *pkg, int64 len, char *pn)
 			else
 				ctxt->textp = s;
 			ctxt->etextp = s;
-			for(s = s->sub; s != S; s = s->sub) {
+			for(s = s->sub; s != nil; s = s->sub) {
 				if(s->onlist)
 					sysfatal("symbol %s listed multiple times", s->name);
 				s->onlist = 1;
@@ -458,7 +461,7 @@ readsym(PeObj *obj, int i, PeSym **y)
 		name = sym->name;
 		if(strncmp(name, "__imp_", 6) == 0)
 			name = &name[6]; // __imp_Name => Name
-		if(thechar == '8' && name[0] == '_')
+		if(thearch.thechar == '8' && name[0] == '_')
 			name = &name[1]; // _Name => Name
 	}
 	// remove last @XXX
diff --git a/src/cmd/ld/lib.c b/src/cmd/ld/lib.c
index fa08bc5f03..72c903b4d2 100644
--- a/src/cmd/ld/lib.c
+++ b/src/cmd/ld/lib.c
@@ -29,10 +29,13 @@
 // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 // THE SOFTWARE.
 
-#include	"l.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
 #include	"lib.h"
-#include	"../ld/elf.h"
-#include	"../ld/dwarf.h"
+#include	"elf.h"
+#include	"dwarf.h"
 #include	"../../runtime/stack.h"
 #include	"../../runtime/funcdata.h"
 
@@ -105,7 +108,7 @@ libinit(void)
 {
 	char *suffix, *suffixsep;
 
-	funcalign = FuncAlign;
+	funcalign = thearch.funcalign;
 	fmtinstall('i', iconv);
 	fmtinstall('Y', Yconv);
 	fmtinstall('Z', Zconv);
@@ -191,7 +194,7 @@ loadlib(void)
 	}
 
 	loadinternal("runtime");
-	if(thechar == '5')
+	if(thearch.thechar == '5')
 		loadinternal("math");
 	if(flag_race)
 		loadinternal("runtime/race");
@@ -212,6 +215,14 @@ loadlib(void)
 		// Force external linking for android.
 		if(strcmp(goos, "android") == 0)
 			linkmode = LinkExternal;
+
+		// cgo on Darwin must use external linking
+		// we can always use external linking, but then there will be circular
+		// dependency problems when compiling natively (external linking requires
+		// runtime/cgo, runtime/cgo requires cmd/cgo, but cmd/cgo needs to be
+		// compiled using external linking.)
+		if(thearch.thechar == '5' && HEADTYPE == Hdarwin && iscgo)
+			linkmode = LinkExternal;
 	}
 
 	if(linkmode == LinkExternal && !iscgo) {
@@ -243,7 +254,7 @@ loadlib(void)
 	if(linkmode == LinkInternal) {
 		// Drop all the cgo_import_static declarations.
 		// Turns out we won't be needing them.
-		for(s = ctxt->allsym; s != S; s = s->allsym)
+		for(s = ctxt->allsym; s != nil; s = s->allsym)
 			if(s->type == SHOSTOBJ) {
 				// If a symbol was marked both
 				// cgo_import_static and cgo_import_dynamic,
@@ -257,8 +268,14 @@ loadlib(void)
 	}
 	
 	tlsg = linklookup(ctxt, "runtime.tlsg", 0);
-	tlsg->type = STLSBSS;
-	tlsg->size = PtrSize;
+	// For most ports, runtime.tlsg is a placeholder symbol for TLS
+	// relocation. However, the Android and Darwin ports need it to
+	// be a real variable. Instead of hard-coding which platforms
+	// need it to be a real variable, we set the type to STLSBSS only
+	// when the runtime has not declared its type already.
+	if(tlsg->type == 0)
+		tlsg->type = STLSBSS;
+	tlsg->size = thearch.ptrsize;
 	tlsg->hide = 1;
 	tlsg->reachable = 1;
 	ctxt->tlsg = tlsg;
@@ -426,7 +443,7 @@ dowrite(int fd, char *p, int n)
 	while(n > 0) {
 		m = write(fd, p, n);
 		if(m <= 0) {
-			ctxt->cursym = S;
+			ctxt->cursym = nil;
 			diag("write error: %r");
 			errorexit();
 		}
@@ -515,7 +532,7 @@ hostobjs(void)
 		h = &hostobj[i];
 		f = Bopen(h->file, OREAD);
 		if(f == nil) {
-			ctxt->cursym = S;
+			ctxt->cursym = nil;
 			diag("cannot reopen %s: %r", h->pn);
 			errorexit();
 		}
@@ -589,7 +606,7 @@ hostlink(void)
 	if(extld == nil)
 		extld = "gcc";
 	argv[argc++] = extld;
-	switch(thechar){
+	switch(thearch.thechar){
 	case '8':
 		argv[argc++] = "-m32";
 		break;
@@ -637,7 +654,7 @@ hostlink(void)
 		h = &hostobj[i];
 		f = Bopen(h->file, OREAD);
 		if(f == nil) {
-			ctxt->cursym = S;
+			ctxt->cursym = nil;
 			diag("cannot reopen %s: %r", h->pn);
 			errorexit();
 		}
@@ -646,7 +663,7 @@ hostlink(void)
 		argv[argc++] = p;
 		w = create(p, 1, 0775);
 		if(w < 0) {
-			ctxt->cursym = S;
+			ctxt->cursym = nil;
 			diag("cannot create %s: %r", p);
 			errorexit();
 		}
@@ -658,7 +675,7 @@ hostlink(void)
 			len -= n;
 		}
 		if(close(w) < 0) {
-			ctxt->cursym = S;
+			ctxt->cursym = nil;
 			diag("cannot write %s: %r", p);
 			errorexit();
 		}
@@ -705,7 +722,7 @@ hostlink(void)
 	}
 
 	if(runcmd(argv) < 0) {
-		ctxt->cursym = S;
+		ctxt->cursym = nil;
 		diag("%s: running %s failed: %r", argv0, argv[0]);
 		errorexit();
 	}
@@ -760,7 +777,7 @@ ldobj(Biobuf *f, char *pkg, int64 len, char *pn, char *file, int whence)
 	line[n] = '\0';
 	if(strncmp(line, "go object ", 10) != 0) {
 		if(strlen(pn) > 3 && strcmp(pn+strlen(pn)-3, ".go") == 0) {
-			print("%cl: input %s is not .%c file (use %cg to compile .go files)\n", thechar, pn, thechar, thechar);
+			print("%cl: input %s is not .%c file (use %cg to compile .go files)\n", thearch.thechar, pn, thearch.thechar, thearch.thechar);
 			errorexit();
 		}
 		if(strcmp(line, thestring) == 0) {
@@ -846,7 +863,7 @@ mywhatsys(void)
 	goarch = getgoarch();
 
 	if(strncmp(goarch, thestring, strlen(thestring)) != 0)
-		sysfatal("cannot use %cc with GOARCH=%s", thechar, goarch);
+		sysfatal("cannot use %cc with GOARCH=%s", thearch.thechar, goarch);
 }
 
 int
@@ -974,7 +991,7 @@ addsection(Segment *seg, char *name, int rwx)
 	sect->rwx = rwx;
 	sect->name = name;
 	sect->seg = seg;
-	sect->align = PtrSize; // everything is at least pointer-aligned
+	sect->align = thearch.ptrsize; // everything is at least pointer-aligned
 	*l = sect;
 	return sect;
 }
@@ -1032,20 +1049,21 @@ static void stkbroke(Chain*, int);
 static LSym *morestack;
 static LSym *newstack;
 
-enum
-{
-	HasLinkRegister = (thechar == '5' || thechar == '9'),
-};
-
 // TODO: Record enough information in new object files to
 // allow stack checks here.
 
 static int
+haslinkregister(void)
+{
+	return thearch.thechar == '5' || thearch.thechar == '9';
+}
+
+static int
 callsize(void)
 {
-	if(HasLinkRegister)
+	if(haslinkregister())
 		return 0;
-	return RegSize;
+	return thearch.regsize;
 }
 
 void
@@ -1166,8 +1184,8 @@ stkcheck(Chain *up, int depth)
 				// to StackLimit beyond the frame size.
 				if(strncmp(r->sym->name, "runtime.morestack", 17) == 0) {
 					limit = StackLimit + s->locals;
-					if(HasLinkRegister)
-						limit += RegSize;
+					if(haslinkregister())
+						limit += thearch.regsize;
 				}
 				break;
 
@@ -1215,8 +1233,8 @@ stkprint(Chain *ch, int limit)
 		else
 			print("\t%d\tguaranteed after split check in %s\n", ch->limit, name);
 	} else {
-		stkprint(ch->up, ch->limit + (!HasLinkRegister)*RegSize);
-		if(!HasLinkRegister)
+		stkprint(ch->up, ch->limit + callsize());
+		if(!haslinkregister())
 			print("\t%d\ton entry to %s\n", ch->limit, name);
 	}
 	if(ch->limit != limit)
@@ -1232,7 +1250,7 @@ Yconv(Fmt *fp)
 	char *str;
 
 	s = va_arg(fp->args, LSym*);
-	if (s == S) {
+	if (s == nil) {
 		fmtprint(fp, "<nil>");
 	} else {
 		fmtstrinit(&fmt);
@@ -1317,7 +1335,7 @@ cwrite(void *buf, int n)
 void
 usage(void)
 {
-	fprint(2, "usage: %cl [options] main.%c\n", thechar, thechar);
+	fprint(2, "usage: %cl [options] main.%c\n", thearch.thechar, thearch.thechar);
 	flagprint(2);
 	exits("usage");
 }
@@ -1346,7 +1364,7 @@ setinterp(char *s)
 void
 doversion(void)
 {
-	print("%cl version %s\n", thechar, getgoversion());
+	print("%cl version %s\n", thearch.thechar, getgoversion());
 	errorexit();
 }
 
@@ -1366,7 +1384,7 @@ genasmsym(void (*put)(LSym*, char*, int, vlong, vlong, int, LSym*))
 	if(s->type == STEXT)
 		put(s, s->name, 'T', s->value, s->size, s->version, 0);
 
-	for(s=ctxt->allsym; s!=S; s=s->allsym) {
+	for(s=ctxt->allsym; s!=nil; s=s->allsym) {
 		if(s->hide || (s->name[0] == '.' && s->version == 0 && strcmp(s->name, ".rathole") != 0))
 			continue;
 		switch(s->type&SMASK) {
@@ -1406,7 +1424,7 @@ genasmsym(void (*put)(LSym*, char*, int, vlong, vlong, int, LSym*))
 		put(s, s->name, 'T', s->value, s->size, s->version, s->gotype);
 
 		// NOTE(ality): acid can't produce a stack trace without .frame symbols
-		put(nil, ".frame", 'm', s->locals+PtrSize, 0, 0, 0);
+		put(nil, ".frame", 'm', s->locals+thearch.ptrsize, 0, 0, 0);
 
 		for(a=s->autom; a; a=a->link) {
 			// Emit a or p according to actual offset, even if label is wrong.
@@ -1418,7 +1436,7 @@ genasmsym(void (*put)(LSym*, char*, int, vlong, vlong, int, LSym*))
 			if(a->name == A_PARAM)
 				off = a->aoffset;
 			else
-				off = a->aoffset - PtrSize;
+				off = a->aoffset - thearch.ptrsize;
 			
 			// FP
 			if(off >= 0) {
@@ -1427,8 +1445,8 @@ genasmsym(void (*put)(LSym*, char*, int, vlong, vlong, int, LSym*))
 			}
 			
 			// SP
-			if(off <= -PtrSize) {
-				put(nil, a->asym->name, 'a', -(off+PtrSize), 0, 0, a->gotype);
+			if(off <= -thearch.ptrsize) {
+				put(nil, a->asym->name, 'a', -(off+thearch.ptrsize), 0, 0, a->gotype);
 				continue;
 			}
 			
@@ -1549,7 +1567,7 @@ diag(char *fmt, ...)
 
 	tn = "";
 	sep = "";
-	if(ctxt->cursym != S) {
+	if(ctxt->cursym != nil) {
 		tn = ctxt->cursym->name;
 		sep = ": ";
 	}
@@ -1617,3 +1635,18 @@ checkgo(void)
 		}
 	}
 }
+
+vlong
+rnd(vlong v, vlong r)
+{
+	vlong c;
+
+	if(r <= 0)
+		return v;
+	v += r - 1;
+	c = v % r;
+	if(c < 0)
+		c += r;
+	v -= c;
+	return v;
+}
diff --git a/src/cmd/ld/lib.h b/src/cmd/ld/lib.h
index fd84c8bccb..f6a89535da 100644
--- a/src/cmd/ld/lib.h
+++ b/src/cmd/ld/lib.h
@@ -28,6 +28,58 @@
 // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 // THE SOFTWARE.
 
+#ifndef	EXTERN
+#define	EXTERN	extern
+#endif
+
+typedef struct Arch Arch;
+struct Arch {
+	int	thechar;
+	int	ptrsize;
+	int	intsize;
+	int	regsize;
+	int	funcalign;
+	int	maxalign;
+	int	minlc;
+	int	dwarfregsp;
+	
+	char *linuxdynld;
+	char *freebsddynld;
+	char *netbsddynld;
+	char *openbsddynld;
+	char *dragonflydynld;
+	char *solarisdynld;
+	
+	void	(*adddynlib)(char*);
+	void	(*adddynrel)(LSym*, Reloc*);
+	void	(*adddynsym)(Link*, LSym*);
+	void	(*archinit)(void);
+	int	(*archreloc)(Reloc*, LSym*, vlong*);
+	vlong	(*archrelocvariant)(Reloc*, LSym*, vlong);
+	void	(*asmb)(void);
+	int	(*elfreloc1)(Reloc*, vlong);
+	void	(*elfsetupplt)(void);
+	void	(*gentext)(void);
+	void	(*listinit)(void);
+	int	(*machoreloc1)(Reloc*, vlong);
+	
+	void	(*lput)(uint32);
+	void	(*wput)(uint16);
+	void	(*vput)(uint64);
+};
+
+vlong rnd(vlong, vlong);
+
+EXTERN	Arch	thearch;
+EXTERN	LSym*	datap;
+EXTERN	int	debug[128];
+EXTERN	char	literal[32];
+EXTERN	int32	lcsize;
+EXTERN	char*	rpath;
+EXTERN	int32	spsize;
+EXTERN	LSym*	symlist;
+EXTERN	int32	symsize;
+
 // Terrible but standard terminology.
 // A segment describes a block of file to load into memory.
 // A section further describes the pieces of that block for
@@ -71,8 +123,8 @@ struct Section
 extern	char	symname[];
 
 EXTERN	char*	INITENTRY;
-extern	char*	thestring;
-extern	LinkArch*	thelinkarch;
+EXTERN	char*	thestring;
+EXTERN	LinkArch*	thelinkarch;
 EXTERN	char*	outfile;
 EXTERN	int	ndynexp;
 EXTERN	LSym**	dynexp;
@@ -82,6 +134,7 @@ EXTERN	int	havedynamic;
 EXTERN	int	funcalign;
 EXTERN	int	iscgo;
 EXTERN	int	elfglobalsymndx;
+extern	int	nelfsym;
 EXTERN	char*	flag_installsuffix;
 EXTERN	int	flag_race;
 EXTERN	int flag_shared;
@@ -250,8 +303,8 @@ void	libinit(void);
 LSym*	listsort(LSym *l, int (*cmp)(LSym*, LSym*), int off);
 void	loadinternal(char *name);
 void	loadlib(void);
-void	lputb(int32 l);
-void	lputl(int32 l);
+void	lputb(uint32 l);
+void	lputl(uint32 l);
 void*	mal(uint32 n);
 void	mark(LSym *s);
 void	mywhatsys(void);
@@ -289,4 +342,8 @@ void	zerosig(char *sp);
 void	archinit(void);
 void	diag(char *fmt, ...);
 
+void	ldmain(int, char**);
+
 #pragma	varargck	argpos	diag	1
+
+#define	SYMDEF	"__.GOSYMDEF"
\ No newline at end of file
diff --git a/src/cmd/ld/macho.c b/src/cmd/ld/macho.c
index ffb20b3a55..b9b341031b 100644
--- a/src/cmd/ld/macho.c
+++ b/src/cmd/ld/macho.c
@@ -5,10 +5,13 @@
 // Mach-O file writing
 // http://developer.apple.com/mac/library/DOCUMENTATION/DeveloperTools/Conceptual/MachORuntime/Reference/reference.html
 
-#include "l.h"
-#include "../ld/dwarf.h"
-#include "../ld/lib.h"
-#include "../ld/macho.h"
+#include <u.h>
+#include <libc.h>
+#include <bio.h>
+#include <link.h>
+#include "dwarf.h"
+#include "lib.h"
+#include "macho.h"
 
 static	int	macho64;
 static	MachoHdr	hdr;
@@ -41,7 +44,7 @@ static	void	machodysymtab(void);
 void
 machoinit(void)
 {
-	switch(thechar) {
+	switch(thearch.thechar) {
 	// 64-bit architectures
 	case '6':
 	case '9':
@@ -146,85 +149,85 @@ machowrite(void)
 	}
 
 	if(macho64)
-		LPUT(0xfeedfacf);
+		thearch.lput(0xfeedfacf);
 	else
-		LPUT(0xfeedface);
-	LPUT(hdr.cpu);
-	LPUT(hdr.subcpu);
+		thearch.lput(0xfeedface);
+	thearch.lput(hdr.cpu);
+	thearch.lput(hdr.subcpu);
 	if(linkmode == LinkExternal)
-		LPUT(1);	/* file type - mach object */
+		thearch.lput(1);	/* file type - mach object */
 	else
-		LPUT(2);	/* file type - mach executable */
-	LPUT(nload+nseg+ndebug);
-	LPUT(loadsize);
-	LPUT(1);	/* flags - no undefines */
+		thearch.lput(2);	/* file type - mach executable */
+	thearch.lput(nload+nseg+ndebug);
+	thearch.lput(loadsize);
+	thearch.lput(1);	/* flags - no undefines */
 	if(macho64)
-		LPUT(0);	/* reserved */
+		thearch.lput(0);	/* reserved */
 
 	for(i=0; i<nseg; i++) {
 		s = &seg[i];
 		if(macho64) {
-			LPUT(25);	/* segment 64 */
-			LPUT(72+80*s->nsect);
+			thearch.lput(25);	/* segment 64 */
+			thearch.lput(72+80*s->nsect);
 			strnput(s->name, 16);
-			VPUT(s->vaddr);
-			VPUT(s->vsize);
-			VPUT(s->fileoffset);
-			VPUT(s->filesize);
-			LPUT(s->prot1);
-			LPUT(s->prot2);
-			LPUT(s->nsect);
-			LPUT(s->flag);
+			thearch.vput(s->vaddr);
+			thearch.vput(s->vsize);
+			thearch.vput(s->fileoffset);
+			thearch.vput(s->filesize);
+			thearch.lput(s->prot1);
+			thearch.lput(s->prot2);
+			thearch.lput(s->nsect);
+			thearch.lput(s->flag);
 		} else {
-			LPUT(1);	/* segment 32 */
-			LPUT(56+68*s->nsect);
+			thearch.lput(1);	/* segment 32 */
+			thearch.lput(56+68*s->nsect);
 			strnput(s->name, 16);
-			LPUT(s->vaddr);
-			LPUT(s->vsize);
-			LPUT(s->fileoffset);
-			LPUT(s->filesize);
-			LPUT(s->prot1);
-			LPUT(s->prot2);
-			LPUT(s->nsect);
-			LPUT(s->flag);
+			thearch.lput(s->vaddr);
+			thearch.lput(s->vsize);
+			thearch.lput(s->fileoffset);
+			thearch.lput(s->filesize);
+			thearch.lput(s->prot1);
+			thearch.lput(s->prot2);
+			thearch.lput(s->nsect);
+			thearch.lput(s->flag);
 		}
 		for(j=0; j<s->nsect; j++) {
 			t = &s->sect[j];
 			if(macho64) {
 				strnput(t->name, 16);
 				strnput(t->segname, 16);
-				VPUT(t->addr);
-				VPUT(t->size);
-				LPUT(t->off);
-				LPUT(t->align);
-				LPUT(t->reloc);
-				LPUT(t->nreloc);
-				LPUT(t->flag);
-				LPUT(t->res1);	/* reserved */
-				LPUT(t->res2);	/* reserved */
-				LPUT(0);	/* reserved */
+				thearch.vput(t->addr);
+				thearch.vput(t->size);
+				thearch.lput(t->off);
+				thearch.lput(t->align);
+				thearch.lput(t->reloc);
+				thearch.lput(t->nreloc);
+				thearch.lput(t->flag);
+				thearch.lput(t->res1);	/* reserved */
+				thearch.lput(t->res2);	/* reserved */
+				thearch.lput(0);	/* reserved */
 			} else {
 				strnput(t->name, 16);
 				strnput(t->segname, 16);
-				LPUT(t->addr);
-				LPUT(t->size);
-				LPUT(t->off);
-				LPUT(t->align);
-				LPUT(t->reloc);
-				LPUT(t->nreloc);
-				LPUT(t->flag);
-				LPUT(t->res1);	/* reserved */
-				LPUT(t->res2);	/* reserved */
+				thearch.lput(t->addr);
+				thearch.lput(t->size);
+				thearch.lput(t->off);
+				thearch.lput(t->align);
+				thearch.lput(t->reloc);
+				thearch.lput(t->nreloc);
+				thearch.lput(t->flag);
+				thearch.lput(t->res1);	/* reserved */
+				thearch.lput(t->res2);	/* reserved */
 			}
 		}
 	}
 
 	for(i=0; i<nload; i++) {
 		l = &load[i];
-		LPUT(l->type);
-		LPUT(4*(l->ndata+2));
+		thearch.lput(l->type);
+		thearch.lput(4*(l->ndata+2));
 		for(j=0; j<l->ndata; j++)
-			LPUT(l->data[j]);
+			thearch.lput(l->data[j]);
 	}
 
 	return cpos() - o1;
@@ -353,10 +356,14 @@ asmbmacho(void)
 	/* apple MACH */
 	va = INITTEXT - HEADR;
 	mh = getMachoHdr();
-	switch(thechar){
+	switch(thearch.thechar){
 	default:
 		diag("unknown mach architecture");
 		errorexit();
+	case '5':
+		mh->cpu = MACHO_CPU_ARM;
+		mh->subcpu = MACHO_SUBCPU_ARMV7;
+		break;
 	case '6':
 		mh->cpu = MACHO_CPU_AMD64;
 		mh->subcpu = MACHO_SUBCPU_X86;
@@ -412,10 +419,16 @@ asmbmacho(void)
 		machoshbits(ms, sect, "__DATA");
 
 	if(linkmode != LinkExternal) {
-		switch(thechar) {
+		switch(thearch.thechar) {
 		default:
 			diag("unknown macho architecture");
 			errorexit();
+		case '5':
+			ml = newMachoLoad(5, 17+2);	/* unix thread */
+			ml->data[0] = 1;	/* thread type */
+			ml->data[1] = 17;	/* word count */
+			ml->data[2+15] = entryvalue();	/* start pc */
+			break;
 		case '6':
 			ml = newMachoLoad(5, 42+2);	/* unix thread */
 			ml->data[0] = 4;	/* thread type */
@@ -605,7 +618,7 @@ machosymtab(void)
 			adduint8(ctxt, symtab, 0x01); // type N_EXT, external symbol
 			adduint8(ctxt, symtab, 0); // no section
 			adduint16(ctxt, symtab, 0); // desc
-			adduintxx(ctxt, symtab, 0, PtrSize); // no value
+			adduintxx(ctxt, symtab, 0, thearch.ptrsize); // no value
 		} else {
 			if(s->cgoexport)
 				adduint8(ctxt, symtab, 0x0f);
@@ -620,7 +633,7 @@ machosymtab(void)
 			} else
 				adduint8(ctxt, symtab, o->sect->extnum);
 			adduint16(ctxt, symtab, 0); // desc
-			adduintxx(ctxt, symtab, symaddr(s), PtrSize);
+			adduintxx(ctxt, symtab, symaddr(s), thearch.ptrsize);
 		}
 	}
 }
@@ -746,7 +759,7 @@ machorelocsect(Section *sect, LSym *first)
 		for(r = sym->r; r < sym->r+sym->nr; r++) {
 			if(r->done)
 				continue;
-			if(machoreloc1(r, sym->value+r->off - sect->vaddr) < 0)
+			if(thearch.machoreloc1(r, sym->value+r->off - sect->vaddr) < 0)
 				diag("unsupported obj reloc %d/%d to %s", r->type, r->siz, r->sym->name);
 		}
 	}
diff --git a/src/cmd/ld/macho.h b/src/cmd/ld/macho.h
index d759f4b0f8..d21109b38f 100644
--- a/src/cmd/ld/macho.h
+++ b/src/cmd/ld/macho.h
@@ -66,6 +66,9 @@ enum {
 	MACHO_CPU_AMD64 = (1<<24)|7,
 	MACHO_CPU_386 = 7,
 	MACHO_SUBCPU_X86 = 3,
+	MACHO_CPU_ARM = 12,
+	MACHO_SUBCPU_ARM = 0,
+	MACHO_SUBCPU_ARMV7 = 9,
 
 	MACHO32SYMSIZE = 12,
 	MACHO64SYMSIZE = 16,
@@ -80,6 +83,9 @@ enum {
 	MACHO_X86_64_RELOC_SIGNED_2 = 7,
 	MACHO_X86_64_RELOC_SIGNED_4 = 8,
 	
+	MACHO_ARM_RELOC_VANILLA = 0,
+	MACHO_ARM_RELOC_BR24 = 5,
+	
 	MACHO_GENERIC_RELOC_VANILLA = 0,
 	
 	MACHO_FAKE_GOTPCREL = 100,
diff --git a/src/cmd/ld/pcln.c b/src/cmd/ld/pcln.c
index f889b2c3ea..adb7661301 100644
--- a/src/cmd/ld/pcln.c
+++ b/src/cmd/ld/pcln.c
@@ -2,7 +2,10 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-#include	"l.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
 #include	"lib.h"
 #include	"../../runtime/funcdata.h"
 
@@ -104,6 +107,15 @@ renumberfiles(Link *ctxt, LSym **files, int nfiles, Pcdata *d)
 	*d = out;	
 }
 
+static int
+container(LSym *s)
+{
+	// We want to generate func table entries only for the "lowest level" symbols,
+	// not containers of subsymbols.
+	if(s != nil && s->sub != nil)
+		return 1;
+	return 0;
+}
 
 // pclntab initializes the pclntab symbol with
 // runtime function and file name information.
@@ -111,7 +123,7 @@ void
 pclntab(void)
 {
 	int32 i, nfunc, start, funcstart;
-	LSym *ftab, *s;
+	LSym *ftab, *s, *last;
 	int32 off, end, frameptrsize;
 	int64 funcdata_bytes;
 	Pcln *pcln;
@@ -125,35 +137,41 @@ pclntab(void)
 
 	// See golang.org/s/go12symtab for the format. Briefly:
 	//	8-byte header
-	//	nfunc [PtrSize bytes]
-	//	function table, alternating PC and offset to func struct [each entry PtrSize bytes]
-	//	end PC [PtrSize bytes]
+	//	nfunc [thearch.ptrsize bytes]
+	//	function table, alternating PC and offset to func struct [each entry thearch.ptrsize bytes]
+	//	end PC [thearch.ptrsize bytes]
 	//	offset to file table [4 bytes]
 	nfunc = 0;
-	for(ctxt->cursym = ctxt->textp; ctxt->cursym != nil; ctxt->cursym = ctxt->cursym->next)
-		nfunc++;
-	symgrow(ctxt, ftab, 8+PtrSize+nfunc*2*PtrSize+PtrSize+4);
+	for(ctxt->cursym = ctxt->textp; ctxt->cursym != nil; ctxt->cursym = ctxt->cursym->next) {
+		if(!container(ctxt->cursym))
+			nfunc++;
+	}
+	symgrow(ctxt, ftab, 8+thearch.ptrsize+nfunc*2*thearch.ptrsize+thearch.ptrsize+4);
 	setuint32(ctxt, ftab, 0, 0xfffffffb);
-	setuint8(ctxt, ftab, 6, MINLC);
-	setuint8(ctxt, ftab, 7, PtrSize);
-	setuintxx(ctxt, ftab, 8, nfunc, PtrSize);
+	setuint8(ctxt, ftab, 6, thearch.minlc);
+	setuint8(ctxt, ftab, 7, thearch.ptrsize);
+	setuintxx(ctxt, ftab, 8, nfunc, thearch.ptrsize);
 
 	nfunc = 0;
-	for(ctxt->cursym = ctxt->textp; ctxt->cursym != nil; ctxt->cursym = ctxt->cursym->next, nfunc++) {
+	last = nil;
+	for(ctxt->cursym = ctxt->textp; ctxt->cursym != nil; ctxt->cursym = ctxt->cursym->next) {
+		last = ctxt->cursym;
+		if(container(ctxt->cursym))
+			continue;
 		pcln = ctxt->cursym->pcln;
 		if(pcln == nil)
 			pcln = &zpcln;
 	
 		funcstart = ftab->np;
-		funcstart += -ftab->np & (PtrSize-1);
+		funcstart += -ftab->np & (thearch.ptrsize-1);
 
-		setaddr(ctxt, ftab, 8+PtrSize+nfunc*2*PtrSize, ctxt->cursym);
-		setuintxx(ctxt, ftab, 8+PtrSize+nfunc*2*PtrSize+PtrSize, funcstart, PtrSize);
+		setaddr(ctxt, ftab, 8+thearch.ptrsize+nfunc*2*thearch.ptrsize, ctxt->cursym);
+		setuintxx(ctxt, ftab, 8+thearch.ptrsize+nfunc*2*thearch.ptrsize+thearch.ptrsize, funcstart, thearch.ptrsize);
 
 		// fixed size of struct, checked below
 		off = funcstart;
-		end = funcstart + PtrSize + 3*4 + 5*4 + pcln->npcdata*4 + pcln->nfuncdata*PtrSize;
-		if(pcln->nfuncdata > 0 && (end&(PtrSize-1)))
+		end = funcstart + thearch.ptrsize + 3*4 + 5*4 + pcln->npcdata*4 + pcln->nfuncdata*thearch.ptrsize;
+		if(pcln->nfuncdata > 0 && (end&(thearch.ptrsize-1)))
 			end += 4;
 		symgrow(ctxt, ftab, end);
 
@@ -173,7 +191,7 @@ pclntab(void)
 		// when a called function doesn't have argument information.
 		// We need to make sure everything has argument information
 		// and then remove this.
-		frameptrsize = PtrSize;
+		frameptrsize = thearch.ptrsize;
 		if(ctxt->cursym->leaf)
 			frameptrsize = 0;
 		off = setuint32(ctxt, ftab, off, ctxt->cursym->locals + frameptrsize);
@@ -203,38 +221,38 @@ pclntab(void)
 		// funcdata, must be pointer-aligned and we're only int32-aligned.
 		// Missing funcdata will be 0 (nil pointer).
 		if(pcln->nfuncdata > 0) {
-			if(off&(PtrSize-1))
+			if(off&(thearch.ptrsize-1))
 				off += 4;
 			for(i=0; i<pcln->nfuncdata; i++) {
 				if(pcln->funcdata[i] == nil)
-					setuintxx(ctxt, ftab, off+PtrSize*i, pcln->funcdataoff[i], PtrSize);
+					setuintxx(ctxt, ftab, off+thearch.ptrsize*i, pcln->funcdataoff[i], thearch.ptrsize);
 				else {
 					// TODO: Dedup.
 					funcdata_bytes += pcln->funcdata[i]->size;
-					setaddrplus(ctxt, ftab, off+PtrSize*i, pcln->funcdata[i], pcln->funcdataoff[i]);
+					setaddrplus(ctxt, ftab, off+thearch.ptrsize*i, pcln->funcdata[i], pcln->funcdataoff[i]);
 				}
 			}
-			off += pcln->nfuncdata*PtrSize;
+			off += pcln->nfuncdata*thearch.ptrsize;
 		}
 
 		if(off != end) {
-			diag("bad math in functab: funcstart=%d off=%d but end=%d (npcdata=%d nfuncdata=%d ptrsize=%d)", funcstart, off, end, pcln->npcdata, pcln->nfuncdata, PtrSize);
+			diag("bad math in functab: funcstart=%d off=%d but end=%d (npcdata=%d nfuncdata=%d ptrsize=%d)", funcstart, off, end, pcln->npcdata, pcln->nfuncdata, thearch.ptrsize);
 			errorexit();
 		}
 	
-		// Final entry of table is just end pc.
-		if(ctxt->cursym->next == nil)
-			setaddrplus(ctxt, ftab, 8+PtrSize+(nfunc+1)*2*PtrSize, ctxt->cursym, ctxt->cursym->size);
+		nfunc++;
 	}
+	// Final entry of table is just end pc.
+	setaddrplus(ctxt, ftab, 8+thearch.ptrsize+nfunc*2*thearch.ptrsize, last, last->size);
 	
 	// Start file table.
 	start = ftab->np;
-	start += -ftab->np & (PtrSize-1);
-	setuint32(ctxt, ftab, 8+PtrSize+nfunc*2*PtrSize+PtrSize, start);
+	start += -ftab->np & (thearch.ptrsize-1);
+	setuint32(ctxt, ftab, 8+thearch.ptrsize+nfunc*2*thearch.ptrsize+thearch.ptrsize, start);
 
 	symgrow(ctxt, ftab, start+(ctxt->nhistfile+1)*4);
 	setuint32(ctxt, ftab, start, ctxt->nhistfile);
-	for(s = ctxt->filesyms; s != S; s = s->next)
+	for(s = ctxt->filesyms; s != nil; s = s->next)
 		setuint32(ctxt, ftab, start + s->value*4, ftabaddstring(ftab, s->name));
 
 	ftab->size = ftab->np;
@@ -246,6 +264,8 @@ pclntab(void)
 enum {
 	BUCKETSIZE = 256*MINFUNC,
 	SUBBUCKETS = 16,
+	SUBBUCKETSIZE = BUCKETSIZE/SUBBUCKETS,
+	NOIDX = 0x7fffffff
 };
 
 // findfunctab generates a lookup table to quickly find the containing
@@ -253,9 +273,10 @@ enum {
 void
 findfunctab(void)
 {
-	LSym *t, *s;
-	int32 idx, bidx, i, j, nbuckets;
-	vlong min, max;
+	LSym *t, *s, *e;
+	int32 idx, i, j, nbuckets, n, base;
+	vlong min, max, p, q;
+	int32 *indexes;
 
 	t = linklookup(ctxt, "runtime.findfunctab", 0);
 	t->type = SRODATA;
@@ -267,33 +288,60 @@ findfunctab(void)
 	for(s = ctxt->textp; s != nil; s = s->next)
 		max = s->value + s->size;
 
+	// for each subbucket, compute the minimum of all symbol indexes
+	// that map to that subbucket.
+	n = (max-min+SUBBUCKETSIZE-1)/SUBBUCKETSIZE;
+	indexes = (int32*)malloc(n*4);
+	if(indexes == nil) {
+		diag("out of memory");
+		errorexit();
+	}
+	for(i = 0; i < n; i++)
+		indexes[i] = NOIDX;
+	idx = 0;
+	for(s = ctxt->textp; s != nil; s = s->next) {
+		if(container(s))
+			continue;
+		p = s->value;
+		e = s->next;
+		while(container(e))
+			e = e->next;
+		if(e != nil)
+			q = e->value;
+		else
+			q = max;
+
+		//print("%d: [%lld %lld] %s\n", idx, p, q, s->name);
+		for(; p < q; p += SUBBUCKETSIZE) {
+			i = (p - min) / SUBBUCKETSIZE;
+			if(indexes[i] > idx)
+				indexes[i] = idx;
+		}
+		i = (q - 1 - min) / SUBBUCKETSIZE;
+		if(indexes[i] > idx)
+			indexes[i] = idx;
+		idx++;
+	}
+
 	// allocate table
 	nbuckets = (max-min+BUCKETSIZE-1)/BUCKETSIZE;
-	symgrow(ctxt, t, nbuckets * (4+SUBBUCKETS));
+	symgrow(ctxt, t, 4*nbuckets + n);
 
 	// fill in table
-	s = ctxt->textp;
-	idx = 0;
 	for(i = 0; i < nbuckets; i++) {
-		// Find first function which overlaps this bucket.
-		// Only do leaf symbols; skip symbols which are just containers (sub != nil but outer == nil).
-		while(s != nil && (s->value+s->size <= min + i * BUCKETSIZE || s->sub != nil && s->outer == nil)) {
-			s = s->next;
-			idx++;
-		}
-		// record this function in bucket header
-		setuint32(ctxt, t, i*(4+SUBBUCKETS), idx);
-		bidx = idx;
-
-		// compute SUBBUCKETS deltas
-		for(j = 0; j < SUBBUCKETS; j++) {
-			while(s != nil && (s->value+s->size <= min + i * BUCKETSIZE + j * (BUCKETSIZE/SUBBUCKETS) || s->sub != nil && s->outer == nil)) {
-				s = s->next;
-				idx++;
+		base = indexes[i*SUBBUCKETS];
+		if(base == NOIDX)
+			diag("hole in findfunctab");
+		setuint32(ctxt, t, i*(4+SUBBUCKETS), base);
+		for(j = 0; j < SUBBUCKETS && i*SUBBUCKETS+j < n; j++) {
+			idx = indexes[i*SUBBUCKETS+j];
+			if(idx == NOIDX)
+				diag("hole in findfunctab");
+			if(idx - base >= 256) {
+				diag("too many functions in a findfunc bucket! %d/%d %d %d", i, nbuckets, j, idx-base);
 			}
-			if(idx - bidx >= 256)
-				diag("too many functions in a findfunc bucket! %d %s", idx-bidx, s->name);
-			setuint8(ctxt, t, i*(4+SUBBUCKETS)+4+j, idx-bidx);
+			setuint8(ctxt, t, i*(4+SUBBUCKETS)+4+j, idx-base);
 		}
 	}
+	free(indexes);
 }
diff --git a/src/cmd/ld/pe.c b/src/cmd/ld/pe.c
index c26cd5264a..cb87778abd 100644
--- a/src/cmd/ld/pe.c
+++ b/src/cmd/ld/pe.c
@@ -5,10 +5,13 @@
 // PE (Portable Executable) file writing
 // http://www.microsoft.com/whdc/system/platform/firmware/PECOFF.mspx
 
-#include "l.h"
-#include "../ld/lib.h"
-#include "../ld/pe.h"
-#include "../ld/dwarf.h"
+#include <u.h>
+#include <libc.h>
+#include <bio.h>
+#include <link.h>
+#include "lib.h"
+#include "pe.h"
+#include "dwarf.h"
 
 // DOS stub that prints out
 // "This program cannot be run in DOS mode."
@@ -139,7 +142,7 @@ peinit(void)
 {
 	int32 l;
 
-	switch(thechar) {
+	switch(thearch.thechar) {
 	// 64-bit architectures
 	case '6':
 		pe64 = 1;
@@ -204,7 +207,7 @@ initdynimport(void)
 
 	dr = nil;
 	m = nil;
-	for(s = ctxt->allsym; s != S; s = s->allsym) {
+	for(s = ctxt->allsym; s != nil; s = s->allsym) {
 		if(!s->reachable || s->type != SDYNIMPORT)
 			continue;
 		for(d = dr; d != nil; d = d->next) {
@@ -234,9 +237,9 @@ initdynimport(void)
 			m->s->sub = dynamic->sub;
 			dynamic->sub = m->s;
 			m->s->value = dynamic->size;
-			dynamic->size += PtrSize;
+			dynamic->size += thearch.ptrsize;
 		}
-		dynamic->size += PtrSize;
+		dynamic->size += thearch.ptrsize;
 	}
 		
 	return dr;
@@ -344,7 +347,7 @@ initdynexport(void)
 	LSym *s;
 	
 	nexport = 0;
-	for(s = ctxt->allsym; s != S; s = s->allsym) {
+	for(s = ctxt->allsym; s != nil; s = s->allsym) {
 		if(!s->reachable || !(s->cgoexport & CgoExportDynamic))
 			continue;
 		if(nexport+1 > sizeof(dexport)/sizeof(dexport[0])) {
@@ -611,7 +614,7 @@ asmbpe(void)
 {
 	IMAGE_SECTION_HEADER *t, *d;
 
-	switch(thechar) {
+	switch(thearch.thechar) {
 	default:
 		diag("unknown PE architecture");
 		errorexit();
@@ -646,7 +649,9 @@ asmbpe(void)
 	addpersrc();
 
 	fh.NumberOfSections = nsect;
-	fh.TimeDateStamp = time(0);
+	// Being able to produce identical output for identical input is
+	// much more beneficial than having build timestamp in the header.
+	fh.TimeDateStamp = 0;
 	fh.Characteristics = IMAGE_FILE_RELOCS_STRIPPED|
 		IMAGE_FILE_EXECUTABLE_IMAGE|IMAGE_FILE_DEBUG_STRIPPED;
 	if (pe64) {
diff --git a/src/cmd/ld/pobj.c b/src/cmd/ld/pobj.c
index 8ecd18b817..0a48eb3a76 100644
--- a/src/cmd/ld/pobj.c
+++ b/src/cmd/ld/pobj.c
@@ -31,31 +31,33 @@
 // Reading object files.
 
 #define	EXTERN
-#include	"l.h"
-#include	"../ld/lib.h"
-#include	"../ld/elf.h"
-#include	"../ld/macho.h"
-#include	"../ld/dwarf.h"
-#include	"../ld/pe.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
+#include	"lib.h"
+#include	"elf.h"
+#include	"macho.h"
+#include	"dwarf.h"
+#include	"pe.h"
 #include	<ar.h>
 
 char	*noname		= "<none>";
 char*	paramspace	= "FP";
 
 void
-main(int argc, char *argv[])
+ldmain(int argc, char **argv)
 {
 	int i;
 
-	linkarchinit();
 	ctxt = linknew(thelinkarch);
-	ctxt->thechar = thechar;
+	ctxt->thechar = thearch.thechar;
 	ctxt->thestring = thestring;
 	ctxt->diag = diag;
 	ctxt->bso = &bso;
 
 	Binit(&bso, 1, OWRITE);
-	listinit();
+	thearch.listinit();
 	memset(debug, 0, sizeof(debug));
 	nerrors = 0;
 	outfile = nil;
@@ -73,30 +75,30 @@ main(int argc, char *argv[])
 		if(strcmp(argv[i], "-crash_for_testing") == 0)
 			*(volatile int*)0 = 0;
 	
-	if(thechar == '5' && ctxt->goarm == 5)
+	if(thearch.thechar == '5' && ctxt->goarm == 5)
 		debug['F'] = 1;
 
 	flagcount("1", "use alternate profiling code", &debug['1']);
-	if(thechar == '6')
+	if(thearch.thechar == '6')
 		flagcount("8", "assume 64-bit addresses", &debug['8']);
 	flagfn1("B", "info: define ELF NT_GNU_BUILD_ID note", addbuildinfo);
 	flagcount("C", "check Go calls to C code", &debug['C']);
 	flagint64("D", "addr: data address", &INITDAT);
 	flagstr("E", "sym: entry symbol", &INITENTRY);
-	if(thechar == '5')
+	if(thearch.thechar == '5')
 		flagcount("G", "debug pseudo-ops", &debug['G']);
 	flagfn1("I", "interp: set ELF interp", setinterp);
 	flagfn1("L", "dir: add dir to library path", Lflag);
 	flagfn1("H", "head: header type", setheadtype);
 	flagcount("K", "add stack underflow checks", &debug['K']);
-	if(thechar == '5')
+	if(thearch.thechar == '5')
 		flagcount("M", "disable software div/mod", &debug['M']);
 	flagcount("O", "print pc-line tables", &debug['O']);
 	flagcount("Q", "debug byte-register code gen", &debug['Q']);
-	if(thechar == '5')
+	if(thearch.thechar == '5')
 		flagcount("P", "debug code generation", &debug['P']);
 	flagint32("R", "rnd: address rounding", &INITRND);
-	flagcount("S", "check type signatures", &debug['S']);
+	flagcount("nil", "check type signatures", &debug['S']);
 	flagint64("T", "addr: text address", &INITTEXT);
 	flagfn0("V", "print version and exit", doversion);
 	flagcount("W", "disassemble input", &debug['W']);
@@ -117,7 +119,7 @@ main(int argc, char *argv[])
 	flagstr("r", "dir1:dir2:...: set ELF dynamic linker search path", &rpath);
 	flagcount("race", "enable race detector", &flag_race);
 	flagcount("s", "disable symbol table", &debug['s']);
-	if(thechar == '5' || thechar == '6')
+	if(thearch.thechar == '5' || thearch.thechar == '6')
 		flagcount("shared", "generate shared object (implies -linkmode external)", &flag_shared);
 	flagstr("tmpdir", "dir: leave temporary files in this directory", &tmpdir);
 	flagcount("u", "reject unsafe packages", &debug['u']);
@@ -139,9 +141,9 @@ main(int argc, char *argv[])
 
 	if(outfile == nil) {
 		if(HEADTYPE == Hwindows)
-			outfile = smprint("%c.out.exe", thechar);
+			outfile = smprint("%c.out.exe", thearch.thechar);
 		else
-			outfile = smprint("%c.out", thechar);
+			outfile = smprint("%c.out", thearch.thechar);
 	}
 	libinit(); // creates outfile
 
@@ -151,7 +153,7 @@ main(int argc, char *argv[])
 	if(headstring == nil)
 		headstring = headstr(HEADTYPE);
 
-	archinit();
+	thearch.archinit();
 	ctxt->debugfloat = debug['F'];
 
 	if(debug['v'])
@@ -165,7 +167,7 @@ main(int argc, char *argv[])
 	addlibpath(ctxt, "command line", "command line", argv[0], "main");
 	loadlib();
 	
-	if(thechar == '5') {
+	if(thearch.thechar == '5') {
 		// mark some functions that are only referenced after linker code editing
 		if(debug['F'])
 			mark(linkrlookup(ctxt, "_sfloat", 0));
@@ -184,7 +186,7 @@ main(int argc, char *argv[])
 	if(HEADTYPE == Hwindows)
 		dope();
 	addexport();
-	gentext();		// trampolines, call stubs, etc.
+	thearch.gentext();		// trampolines, call stubs, etc.
 	textaddress();
 	pclntab();
 	findfunctab();
@@ -193,7 +195,7 @@ main(int argc, char *argv[])
 	address();
 	doweak();
 	reloc();
-	asmb();
+	thearch.asmb();
 	undef();
 	hostlink();
 	if(debug['v']) {
diff --git a/src/cmd/ld/symtab.c b/src/cmd/ld/symtab.c
index c3a72c3cf6..55ef608c1d 100644
--- a/src/cmd/ld/symtab.c
+++ b/src/cmd/ld/symtab.c
@@ -30,9 +30,12 @@
 
 // Symbol table.
 
-#include	"l.h"
-#include	"../ld/lib.h"
-#include	"../ld/elf.h"
+#include	<u.h>
+#include	<libc.h>
+#include	<bio.h>
+#include	<link.h>
+#include	"lib.h"
+#include	"elf.h"
 
 static int maxelfstr;
 
@@ -76,24 +79,24 @@ putelfstr(char *s)
 static void
 putelfsyment(int off, vlong addr, vlong size, int info, int shndx, int other)
 {
-	switch(thechar) {
+	switch(thearch.thechar) {
 	case '6':
 	case '9':
-		LPUT(off);
+		thearch.lput(off);
 		cput(info);
 		cput(other);
-		WPUT(shndx);
-		VPUT(addr);
-		VPUT(size);
+		thearch.wput(shndx);
+		thearch.vput(addr);
+		thearch.vput(size);
 		symsize += ELF64SYMSIZE;
 		break;
 	default:
-		LPUT(off);
-		LPUT(addr);
-		LPUT(size);
+		thearch.lput(off);
+		thearch.lput(addr);
+		thearch.lput(size);
 		cput(info);
 		cput(other);
-		WPUT(shndx);
+		thearch.wput(shndx);
 		symsize += ELF32SYMSIZE;
 		break;
 	}
@@ -172,7 +175,7 @@ putelfsymshndx(vlong sympos, int shndx)
 	vlong here;
 
 	here = cpos();
-	switch(thechar) {
+	switch(thearch.thechar) {
 	case '6':
 		cseek(sympos+6);
 		break;
@@ -180,7 +183,7 @@ putelfsymshndx(vlong sympos, int shndx)
 		cseek(sympos+14);
 		break;
 	}
-	WPUT(shndx);
+	thearch.wput(shndx);
 	cseek(here);
 }
 
@@ -218,7 +221,7 @@ asmelfsym(void)
 	elfglobalsymndx = numelfsym;
 	genasmsym(putelfsym);
 	
-	for(s=ctxt->allsym; s!=S; s=s->allsym) {
+	for(s=ctxt->allsym; s!=nil; s=s->allsym) {
 		if(s->type != SHOSTOBJ && !(s->type == SDYNIMPORT && s->reachable))
 			continue;
 		if(s->type == SDYNIMPORT)
@@ -253,7 +256,7 @@ putplan9sym(LSym *x, char *s, int t, vlong addr, vlong size, int ver, LSym *go)
 	case 'Z':
 	case 'm':
 		l = 4;
-		if(HEADTYPE == Hplan9 && thechar == '6' && !debug['8']) {
+		if(HEADTYPE == Hplan9 && thearch.thechar == '6' && !debug['8']) {
 			lputb(addr>>32);
 			l = 8;
 		}
@@ -307,7 +310,7 @@ wputb(ushort w)
 }
 
 void
-lputb(int32 l)
+lputb(uint32 l)
 {
 	cput(l>>24);
 	cput(l>>16);
@@ -316,7 +319,7 @@ lputb(int32 l)
 }
 
 void
-lputl(int32 l)
+lputl(uint32 l)
 {
 	cput(l);
 	cput(l>>8);
@@ -408,7 +411,7 @@ symtab(void)
 	// within a type they sort by size, so the .* symbols
 	// just defined above will be first.
 	// hide the specific symbols.
-	for(s = ctxt->allsym; s != S; s = s->allsym) {
+	for(s = ctxt->allsym; s != nil; s = s->allsym) {
 		if(!s->reachable || s->special || s->type != SRODATA)
 			continue;
 		if(strncmp(s->name, "type.", 5) == 0) {
diff --git a/src/cmd/pack/pack_test.go b/src/cmd/pack/pack_test.go
index fbc6a3c698..0c58d628b4 100644
--- a/src/cmd/pack/pack_test.go
+++ b/src/cmd/pack/pack_test.go
@@ -257,15 +257,8 @@ func TestLargeDefs(t *testing.T) {
 		}
 	}
 
-	n := 10000
-	if testing.Short() {
-		// Issue 9656: 10,000 is too aggressive for several
-		// builders, with ~120 MB of disk consumed. 1,000 is
-		// still enough to exercise the old bug.
-		n = 1000
-	}
 	printf("package large\n\ntype T struct {\n")
-	for i := 0; i < n; i++ {
+	for i := 0; i < 1000; i++ {
 		printf("f%d int `tag:\"", i)
 		for j := 0; j < 100; j++ {
 			printf("t%d=%d,", j, j)
diff --git a/src/cmd/pprof/internal/report/report.go b/src/cmd/pprof/internal/report/report.go
index e5977fd034..586f41d4d6 100644
--- a/src/cmd/pprof/internal/report/report.go
+++ b/src/cmd/pprof/internal/report/report.go
@@ -1531,7 +1531,7 @@ func memoryLabel(value int64, fromUnit, toUnit string) (v float64, u string, ok
 		output, toUnit = float64(value)/1024, "kB"
 	case "mb", "mbyte", "megabyte":
 		output, toUnit = float64(value)/(1024*1024), "MB"
-	case "gb", "gbyte", "giggabyte":
+	case "gb", "gbyte", "gigabyte":
 		output, toUnit = float64(value)/(1024*1024*1024), "GB"
 	}
 	return output, toUnit, true
diff --git a/src/cmd/yacc/yacc.go b/src/cmd/yacc/yacc.go
index 0a69c36afc..24692dc166 100644
--- a/src/cmd/yacc/yacc.go
+++ b/src/cmd/yacc/yacc.go
@@ -603,6 +603,7 @@ outer:
 			}
 			levprd[nprod] |= ACTFLAG
 			fmt.Fprintf(fcode, "\n\tcase %v:", nprod)
+			fmt.Fprintf(fcode, "\n\t\t%sDollar = %sS[%spt-%v:%spt+1]", prefix, prefix, prefix, mem-1, prefix)
 			cpyact(curprod, mem)
 
 			// action within rule...
@@ -1345,7 +1346,7 @@ loop:
 				ungetrune(finput, c)
 				continue loop
 			}
-			fmt.Fprintf(fcode, "%sS[%spt-%v]", prefix, prefix, max-j-1)
+			fmt.Fprintf(fcode, "%sDollar[%v]", prefix, j)
 
 			// put out the proper tag
 			if ntypes != 0 {
@@ -3264,6 +3265,7 @@ func $$Parse($$lex $$Lexer) int {
 	var $$n int
 	var $$lval $$SymType
 	var $$VAL $$SymType
+	var $$Dollar []$$SymType
 	$$S := make([]$$SymType, $$MaxDepth)
 
 	Nerrs := 0   /* number of errors */
diff --git a/src/compress/bzip2/bzip2_test.go b/src/compress/bzip2/bzip2_test.go
index fb79d089eb..77c50dfe94 100644
--- a/src/compress/bzip2/bzip2_test.go
+++ b/src/compress/bzip2/bzip2_test.go
@@ -200,7 +200,7 @@ func BenchmarkDecodeDigits(b *testing.B) { benchmarkDecode(b, digits) }
 func BenchmarkDecodeTwain(b *testing.B)  { benchmarkDecode(b, twain) }
 
 func TestBufferOverrun(t *testing.T) {
-	// Tests https://code.google.com/p/go/issues/detail?id=5747.
+	// Tests https://golang.org/issue/5747.
 	buffer := bytes.NewReader([]byte(bufferOverrunBase64))
 	decoder := base64.NewDecoder(base64.StdEncoding, buffer)
 	decompressor := NewReader(decoder)
@@ -209,7 +209,7 @@ func TestBufferOverrun(t *testing.T) {
 }
 
 func TestOutOfRangeSelector(t *testing.T) {
-	// Tests https://code.google.com/p/go/issues/detail?id=8363.
+	// Tests https://golang.org/issue/8363.
 	buffer := bytes.NewReader(outOfRangeSelector)
 	decompressor := NewReader(buffer)
 	// This shouldn't panic.
diff --git a/src/compress/flate/deflate_test.go b/src/compress/flate/deflate_test.go
index 730234c385..53bfd26438 100644
--- a/src/compress/flate/deflate_test.go
+++ b/src/compress/flate/deflate_test.go
@@ -407,7 +407,7 @@ func TestWriterDict(t *testing.T) {
 	}
 }
 
-// See http://code.google.com/p/go/issues/detail?id=2508
+// See http://golang.org/issue/2508
 func TestRegression2508(t *testing.T) {
 	if testing.Short() {
 		t.Logf("test disabled with -short")
diff --git a/src/crypto/tls/handshake_server.go b/src/crypto/tls/handshake_server.go
index a46133439d..c87da50df9 100644
--- a/src/crypto/tls/handshake_server.go
+++ b/src/crypto/tls/handshake_server.go
@@ -172,7 +172,7 @@ Curves:
 		// Although sending an empty NPN extension is reasonable, Firefox has
 		// had a bug around this. Best to send nothing at all if
 		// config.NextProtos is empty. See
-		// https://code.google.com/p/go/issues/detail?id=5445.
+		// https://golang.org/issue/5445.
 		if hs.clientHello.nextProtoNeg && len(config.NextProtos) > 0 {
 			hs.hello.nextProtoNeg = true
 			hs.hello.nextProtos = config.NextProtos
diff --git a/src/crypto/x509/verify.go b/src/crypto/x509/verify.go
index ec1981423d..0181f140fa 100644
--- a/src/crypto/x509/verify.go
+++ b/src/crypto/x509/verify.go
@@ -323,6 +323,8 @@ nextIntermediate:
 }
 
 func matchHostnames(pattern, host string) bool {
+	host = strings.TrimSuffix(host, ".")
+
 	if len(pattern) == 0 || len(host) == 0 {
 		return false
 	}
diff --git a/src/crypto/x509/x509_test.go b/src/crypto/x509/x509_test.go
index bd7cbed8a2..45d49ce3e3 100644
--- a/src/crypto/x509/x509_test.go
+++ b/src/crypto/x509/x509_test.go
@@ -161,11 +161,16 @@ var matchHostnamesTests = []matchHostnamesTest{
 	{"", "b.b.c", false},
 	{"a.b.c", "", false},
 	{"example.com", "example.com", true},
+	{"example.com", "example.com.", true},
 	{"example.com", "www.example.com", false},
 	{"*.example.com", "www.example.com", true},
+	{"*.example.com", "www.example.com.", true},
 	{"*.example.com", "xyz.www.example.com", false},
 	{"*.*.example.com", "xyz.www.example.com", true},
 	{"*.www.*.com", "xyz.www.example.com", true},
+	{"", ".", false},
+	{".", "", false},
+	{".", ".", false},
 }
 
 func TestMatchHostnames(t *testing.T) {
diff --git a/src/encoding/xml/marshal.go b/src/encoding/xml/marshal.go
index 8c6342013d..9d6045c916 100644
--- a/src/encoding/xml/marshal.go
+++ b/src/encoding/xml/marshal.go
@@ -179,17 +179,24 @@ var (
 )
 
 // EncodeToken writes the given XML token to the stream.
-// It returns an error if StartElement and EndElement tokens are not properly matched.
+// It returns an error if StartElement and EndElement tokens are not
+// properly matched.
 //
-// EncodeToken does not call Flush, because usually it is part of a larger operation
-// such as Encode or EncodeElement (or a custom Marshaler's MarshalXML invoked
-// during those), and those will call Flush when finished.
-// Callers that create an Encoder and then invoke EncodeToken directly, without
-// using Encode or EncodeElement, need to call Flush when finished to ensure
-// that the XML is written to the underlying writer.
+// EncodeToken does not call Flush, because usually it is part of a
+// larger operation such as Encode or EncodeElement (or a custom
+// Marshaler's MarshalXML invoked during those), and those will call
+// Flush when finished. Callers that create an Encoder and then invoke
+// EncodeToken directly, without using Encode or EncodeElement, need to
+// call Flush when finished to ensure that the XML is written to the
+// underlying writer.
 //
-// EncodeToken allows writing a ProcInst with Target set to "xml" only as the first token
-// in the stream.
+// EncodeToken allows writing a ProcInst with Target set to "xml" only
+// as the first token in the stream.
+//
+// When encoding a StartElement holding an XML namespace prefix
+// declaration for a prefix that is not already declared, contained
+// elements (including the StartElement itself) will use the declared
+// prefix when encoding names with matching namespace URIs.
 func (enc *Encoder) EncodeToken(t Token) error {
 	p := &enc.p
 	switch t := t.(type) {
@@ -256,19 +263,27 @@ type printer struct {
 	depth      int
 	indentedIn bool
 	putNewline bool
+	defaultNS  string
 	attrNS     map[string]string // map prefix -> name space
 	attrPrefix map[string]string // map name space -> prefix
-	prefixes   []string
+	prefixes   []printerPrefix
 	tags       []Name
 }
 
-// createAttrPrefix finds the name space prefix attribute to use for the given name space,
-// defining a new prefix if necessary. It returns the prefix.
-func (p *printer) createAttrPrefix(url string) string {
-	if prefix := p.attrPrefix[url]; prefix != "" {
-		return prefix
-	}
+// printerPrefix holds a namespace undo record.
+// When an element is popped, the prefix record
+// is set back to the recorded URL. The empty
+// prefix records the URL for the default name space.
+//
+// The start of an element is recorded with an element
+// that has mark=true.
+type printerPrefix struct {
+	prefix string
+	url    string
+	mark   bool
+}
 
+func (p *printer) prefixForNS(url string, isAttr bool) string {
 	// The "http://www.w3.org/XML/1998/namespace" name space is predefined as "xml"
 	// and must be referred to that way.
 	// (The "http://www.w3.org/2000/xmlns/" name space is also predefined as "xmlns",
@@ -276,12 +291,97 @@ func (p *printer) createAttrPrefix(url string) string {
 	if url == xmlURL {
 		return "xml"
 	}
+	if !isAttr && url == p.defaultNS {
+		// We can use the default name space.
+		return ""
+	}
+	return p.attrPrefix[url]
+}
 
-	// Need to define a new name space.
-	if p.attrPrefix == nil {
-		p.attrPrefix = make(map[string]string)
-		p.attrNS = make(map[string]string)
+// defineNS pushes any namespace definition found in the given attribute.
+// If ignoreNonEmptyDefault is true, an xmlns="nonempty"
+// attribute will be ignored.
+func (p *printer) defineNS(attr Attr, ignoreNonEmptyDefault bool) error {
+	var prefix string
+	if attr.Name.Local == "xmlns" {
+		if attr.Name.Space != "" && attr.Name.Space != "xml" && attr.Name.Space != xmlURL {
+			return fmt.Errorf("xml: cannot redefine xmlns attribute prefix")
+		}
+	} else if attr.Name.Space == "xmlns" && attr.Name.Local != "" {
+		prefix = attr.Name.Local
+		if attr.Value == "" {
+			// Technically, an empty XML namespace is allowed for an attribute.
+			// From http://www.w3.org/TR/xml-names11/#scoping-defaulting:
+			//
+			// 	The attribute value in a namespace declaration for a prefix may be
+			//	empty. This has the effect, within the scope of the declaration, of removing
+			//	any association of the prefix with a namespace name.
+			//
+			// However our namespace prefixes here are used only as hints. There's
+			// no need to respect the removal of a namespace prefix, so we ignore it.
+			return nil
+		}
+	} else {
+		// Ignore: it's not a namespace definition
+		return nil
+	}
+	if prefix == "" {
+		if attr.Value == p.defaultNS {
+			// No need for redefinition.
+			return nil
+		}
+		if attr.Value != "" && ignoreNonEmptyDefault {
+			// We have an xmlns="..." value but
+			// it can't define a name space in this context,
+			// probably because the element has an empty
+			// name space. In this case, we just ignore
+			// the name space declaration.
+			return nil
+		}
+	} else if _, ok := p.attrPrefix[attr.Value]; ok {
+		// There's already a prefix for the given name space,
+		// so use that. This prevents us from
+		// having two prefixes for the same name space
+		// so attrNS and attrPrefix can remain bijective.
+		return nil
 	}
+	p.pushPrefix(prefix, attr.Value)
+	return nil
+}
+
+// createNSPrefix creates a name space prefix attribute
+// to use for the given name space, defining a new prefix
+// if necessary.
+// If isAttr is true, the prefix is to be created for an attribute
+// prefix, which means that the default name space cannot
+// be used.
+func (p *printer) createNSPrefix(url string, isAttr bool) {
+	if _, ok := p.attrPrefix[url]; ok {
+		// We already have a prefix for the given URL.
+		return
+	}
+	switch {
+	case !isAttr && url == p.defaultNS:
+		// We can use the default name space.
+		return
+	case url == "":
+		// The only way we can encode names in the empty
+		// name space is by using the default name space,
+		// so we must use that.
+		if p.defaultNS != "" {
+			// The default namespace is non-empty, so we
+			// need to set it to empty.
+			p.pushPrefix("", "")
+		}
+		return
+	case url == xmlURL:
+		return
+	}
+	// TODO If the URL is an existing prefix, we could
+	// use it as is. That would enable the
+	// marshaling of elements that had been unmarshaled
+	// and with a name space prefix that was not found.
+	// although technically it would be incorrect.
 
 	// Pick a name. We try to use the final element of the path
 	// but fall back to _.
@@ -306,39 +406,98 @@ func (p *printer) createAttrPrefix(url string) string {
 		}
 	}
 
-	p.attrPrefix[url] = prefix
-	p.attrNS[prefix] = url
-
-	p.WriteString(`xmlns:`)
-	p.WriteString(prefix)
-	p.WriteString(`="`)
-	EscapeText(p, []byte(url))
-	p.WriteString(`" `)
+	p.pushPrefix(prefix, url)
+}
 
-	p.prefixes = append(p.prefixes, prefix)
+// writeNamespaces writes xmlns attributes for all the
+// namespace prefixes that have been defined in
+// the current element.
+func (p *printer) writeNamespaces() {
+	for i := len(p.prefixes) - 1; i >= 0; i-- {
+		prefix := p.prefixes[i]
+		if prefix.mark {
+			return
+		}
+		p.WriteString(" ")
+		if prefix.prefix == "" {
+			// Default name space.
+			p.WriteString(`xmlns="`)
+		} else {
+			p.WriteString("xmlns:")
+			p.WriteString(prefix.prefix)
+			p.WriteString(`="`)
+		}
+		EscapeText(p, []byte(p.nsForPrefix(prefix.prefix)))
+		p.WriteString(`"`)
+	}
+}
 
-	return prefix
+// pushPrefix pushes a new prefix on the prefix stack
+// without checking to see if it is already defined.
+func (p *printer) pushPrefix(prefix, url string) {
+	p.prefixes = append(p.prefixes, printerPrefix{
+		prefix: prefix,
+		url:    p.nsForPrefix(prefix),
+	})
+	p.setAttrPrefix(prefix, url)
 }
 
-// deleteAttrPrefix removes an attribute name space prefix.
-func (p *printer) deleteAttrPrefix(prefix string) {
-	delete(p.attrPrefix, p.attrNS[prefix])
-	delete(p.attrNS, prefix)
+// nsForPrefix returns the name space for the given
+// prefix. Note that this is not valid for the
+// empty attribute prefix, which always has an empty
+// name space.
+func (p *printer) nsForPrefix(prefix string) string {
+	if prefix == "" {
+		return p.defaultNS
+	}
+	return p.attrNS[prefix]
 }
 
+// markPrefix marks the start of an element on the prefix
+// stack.
 func (p *printer) markPrefix() {
-	p.prefixes = append(p.prefixes, "")
+	p.prefixes = append(p.prefixes, printerPrefix{
+		mark: true,
+	})
 }
 
+// popPrefix pops all defined prefixes for the current
+// element.
 func (p *printer) popPrefix() {
 	for len(p.prefixes) > 0 {
 		prefix := p.prefixes[len(p.prefixes)-1]
 		p.prefixes = p.prefixes[:len(p.prefixes)-1]
-		if prefix == "" {
+		if prefix.mark {
 			break
 		}
-		p.deleteAttrPrefix(prefix)
+		p.setAttrPrefix(prefix.prefix, prefix.url)
+	}
+}
+
+// setAttrPrefix sets an attribute name space prefix.
+// If url is empty, the attribute is removed.
+// If prefix is empty, the default name space is set.
+func (p *printer) setAttrPrefix(prefix, url string) {
+	if prefix == "" {
+		p.defaultNS = url
+		return
+	}
+	if url == "" {
+		delete(p.attrPrefix, p.attrNS[prefix])
+		delete(p.attrNS, prefix)
+		return
+	}
+	if p.attrPrefix == nil {
+		// Need to define a new name space.
+		p.attrPrefix = make(map[string]string)
+		p.attrNS = make(map[string]string)
 	}
+	// Remove any old prefix value. This is OK because we maintain a
+	// strict one-to-one mapping between prefix and URL (see
+	// defineNS)
+	delete(p.attrPrefix, p.attrNS[prefix])
+	p.attrPrefix[url] = prefix
+	p.attrNS[prefix] = url
 }
 
 var (
@@ -376,23 +535,23 @@ func (p *printer) marshalValue(val reflect.Value, finfo *fieldInfo, startTemplat
 
 	// Check for marshaler.
 	if val.CanInterface() && typ.Implements(marshalerType) {
-		return p.marshalInterface(val.Interface().(Marshaler), defaultStart(typ, finfo, startTemplate))
+		return p.marshalInterface(val.Interface().(Marshaler), p.defaultStart(typ, finfo, startTemplate))
 	}
 	if val.CanAddr() {
 		pv := val.Addr()
 		if pv.CanInterface() && pv.Type().Implements(marshalerType) {
-			return p.marshalInterface(pv.Interface().(Marshaler), defaultStart(pv.Type(), finfo, startTemplate))
+			return p.marshalInterface(pv.Interface().(Marshaler), p.defaultStart(pv.Type(), finfo, startTemplate))
 		}
 	}
 
 	// Check for text marshaler.
 	if val.CanInterface() && typ.Implements(textMarshalerType) {
-		return p.marshalTextInterface(val.Interface().(encoding.TextMarshaler), defaultStart(typ, finfo, startTemplate))
+		return p.marshalTextInterface(val.Interface().(encoding.TextMarshaler), p.defaultStart(typ, finfo, startTemplate))
 	}
 	if val.CanAddr() {
 		pv := val.Addr()
 		if pv.CanInterface() && pv.Type().Implements(textMarshalerType) {
-			return p.marshalTextInterface(pv.Interface().(encoding.TextMarshaler), defaultStart(pv.Type(), finfo, startTemplate))
+			return p.marshalTextInterface(pv.Interface().(encoding.TextMarshaler), p.defaultStart(pv.Type(), finfo, startTemplate))
 		}
 	}
 
@@ -419,6 +578,10 @@ func (p *printer) marshalValue(val reflect.Value, finfo *fieldInfo, startTemplat
 	// 3. type name
 	var start StartElement
 
+	// Historic behaviour: elements use the default name space
+	// they are contained in by default.
+	start.Name.Space = p.defaultNS
+
 	if startTemplate != nil {
 		start.Name = startTemplate.Name
 		start.Attr = append(start.Attr, startTemplate.Attr...)
@@ -431,7 +594,10 @@ func (p *printer) marshalValue(val reflect.Value, finfo *fieldInfo, startTemplat
 		}
 	}
 	if start.Name.Local == "" && finfo != nil {
-		start.Name.Space, start.Name.Local = finfo.xmlns, finfo.name
+		start.Name.Local = finfo.name
+		if finfo.xmlns != "" {
+			start.Name.Space = finfo.xmlns
+		}
 	}
 	if start.Name.Local == "" {
 		name := typ.Name()
@@ -440,6 +606,9 @@ func (p *printer) marshalValue(val reflect.Value, finfo *fieldInfo, startTemplat
 		}
 		start.Name.Local = name
 	}
+	// Historic behaviour: an element that's in a namespace sets
+	// the default namespace for all elements contained within it.
+	start.setDefaultNamespace()
 
 	// Attributes
 	for i := range tinfo.fields {
@@ -552,7 +721,7 @@ func (p *printer) marshalValue(val reflect.Value, finfo *fieldInfo, startTemplat
 
 // defaultStart returns the default start element to use,
 // given the reflect type, field info, and start template.
-func defaultStart(typ reflect.Type, finfo *fieldInfo, startTemplate *StartElement) StartElement {
+func (p *printer) defaultStart(typ reflect.Type, finfo *fieldInfo, startTemplate *StartElement) StartElement {
 	var start StartElement
 	// Precedence for the XML element name is as above,
 	// except that we do not look inside structs for the first field.
@@ -569,6 +738,12 @@ func defaultStart(typ reflect.Type, finfo *fieldInfo, startTemplate *StartElemen
 		// since it has the Marshaler methods.
 		start.Name.Local = typ.Elem().Name()
 	}
+	// Historic behaviour: elements use the name space of
+	// the element they are contained in by default.
+	if start.Name.Space == "" {
+		start.Name.Space = p.defaultNS
+	}
+	start.setDefaultNamespace()
 	return start
 }
 
@@ -613,29 +788,44 @@ func (p *printer) writeStart(start *StartElement) error {
 
 	p.tags = append(p.tags, start.Name)
 	p.markPrefix()
+	// Define any name spaces explicitly declared in the attributes.
+	// We do this as a separate pass so that explicitly declared prefixes
+	// will take precedence over implicitly declared prefixes
+	// regardless of the order of the attributes.
+	ignoreNonEmptyDefault := start.Name.Space == ""
+	for _, attr := range start.Attr {
+		if err := p.defineNS(attr, ignoreNonEmptyDefault); err != nil {
+			return err
+		}
+	}
+	// Define any new name spaces implied by the attributes.
+	for _, attr := range start.Attr {
+		name := attr.Name
+		// From http://www.w3.org/TR/xml-names11/#defaulting
+		// "Default namespace declarations do not apply directly
+		// to attribute names; the interpretation of unprefixed
+		// attributes is determined by the element on which they
+		// appear."
+		// This means we don't need to create a new namespace
+		// when an attribute name space is empty.
+		if name.Space != "" && !name.isNamespace() {
+			p.createNSPrefix(name.Space, true)
+		}
+	}
+	p.createNSPrefix(start.Name.Space, false)
 
 	p.writeIndent(1)
 	p.WriteByte('<')
-	p.WriteString(start.Name.Local)
-
-	if start.Name.Space != "" {
-		p.WriteString(` xmlns="`)
-		p.EscapeString(start.Name.Space)
-		p.WriteByte('"')
-	}
-
-	// Attributes
+	p.writeName(start.Name, false)
+	p.writeNamespaces()
 	for _, attr := range start.Attr {
 		name := attr.Name
-		if name.Local == "" {
+		if name.Local == "" || name.isNamespace() {
+			// Namespaces have already been written by writeNamespaces above.
 			continue
 		}
 		p.WriteByte(' ')
-		if name.Space != "" {
-			p.WriteString(p.createAttrPrefix(name.Space))
-			p.WriteByte(':')
-		}
-		p.WriteString(name.Local)
+		p.writeName(name, true)
 		p.WriteString(`="`)
 		p.EscapeString(attr.Value)
 		p.WriteByte('"')
@@ -644,6 +834,16 @@ func (p *printer) writeStart(start *StartElement) error {
 	return nil
 }
 
+// writeName writes the given name. It assumes
+// that p.createNSPrefix(name) has already been called.
+func (p *printer) writeName(name Name, isAttr bool) {
+	if prefix := p.prefixForNS(name.Space, isAttr); prefix != "" {
+		p.WriteString(prefix)
+		p.WriteByte(':')
+	}
+	p.WriteString(name.Local)
+}
+
 func (p *printer) writeEnd(name Name) error {
 	if name.Local == "" {
 		return fmt.Errorf("xml: end tag with no name")
@@ -662,7 +862,7 @@ func (p *printer) writeEnd(name Name) error {
 	p.writeIndent(-1)
 	p.WriteByte('<')
 	p.WriteByte('/')
-	p.WriteString(name.Local)
+	p.writeName(name, false)
 	p.WriteByte('>')
 	p.popPrefix()
 	return nil
diff --git a/src/encoding/xml/marshal_test.go b/src/encoding/xml/marshal_test.go
index cdd52ff97f..cc6994338d 100644
--- a/src/encoding/xml/marshal_test.go
+++ b/src/encoding/xml/marshal_test.go
@@ -1194,41 +1194,363 @@ func TestStructPointerMarshal(t *testing.T) {
 }
 
 var encodeTokenTests = []struct {
-	tok  Token
+	desc string
+	toks []Token
 	want string
-	ok   bool
-}{
-	{StartElement{Name{"space", "local"}, nil}, "<local xmlns=\"space\">", true},
-	{StartElement{Name{"space", ""}, nil}, "", false},
-	{EndElement{Name{"space", ""}}, "", false},
-	{CharData("foo"), "foo", true},
-	{Comment("foo"), "<!--foo-->", true},
-	{Comment("foo-->"), "", false},
-	{ProcInst{"Target", []byte("Instruction")}, "<?Target Instruction?>", true},
-	{ProcInst{"", []byte("Instruction")}, "", false},
-	{ProcInst{"Target", []byte("Instruction?>")}, "", false},
-	{Directive("foo"), "<!foo>", true},
-	{Directive("foo>"), "", false},
-}
+	err  string
+}{{
+	desc: "start element with name space",
+	toks: []Token{
+		StartElement{Name{"space", "local"}, nil},
+	},
+	want: `<space:local xmlns:space="space">`,
+}, {
+	desc: "start element with no name",
+	toks: []Token{
+		StartElement{Name{"space", ""}, nil},
+	},
+	err: "xml: start tag with no name",
+}, {
+	desc: "end element with no name",
+	toks: []Token{
+		EndElement{Name{"space", ""}},
+	},
+	err: "xml: end tag with no name",
+}, {
+	desc: "char data",
+	toks: []Token{
+		CharData("foo"),
+	},
+	want: `foo`,
+}, {
+	desc: "char data with escaped chars",
+	toks: []Token{
+		CharData(" \t\n"),
+	},
+	want: ` &#x9;&#xA;`,
+}, {
+	desc: "comment",
+	toks: []Token{
+		Comment("foo"),
+	},
+	want: `<!--foo-->`,
+}, {
+	desc: "comment with invalid content",
+	toks: []Token{
+		Comment("foo-->"),
+	},
+	err: "xml: EncodeToken of Comment containing --> marker",
+}, {
+	desc: "proc instruction",
+	toks: []Token{
+		ProcInst{"Target", []byte("Instruction")},
+	},
+	want: `<?Target Instruction?>`,
+}, {
+	desc: "proc instruction with empty target",
+	toks: []Token{
+		ProcInst{"", []byte("Instruction")},
+	},
+	err: "xml: EncodeToken of ProcInst with invalid Target",
+}, {
+	desc: "proc instruction with bad content",
+	toks: []Token{
+		ProcInst{"", []byte("Instruction?>")},
+	},
+	err: "xml: EncodeToken of ProcInst with invalid Target",
+}, {
+	desc: "directive",
+	toks: []Token{
+		Directive("foo"),
+	},
+	want: `<!foo>`,
+}, {
+	desc: "directive instruction with bad name",
+	toks: []Token{
+		Directive("foo>"),
+	},
+	err: "xml: EncodeToken of Directive containing > marker",
+}, {
+	desc: "end tag without start tag",
+	toks: []Token{
+		EndElement{Name{"foo", "bar"}},
+	},
+	err: "xml: end tag </bar> without start tag",
+}, {
+	desc: "mismatching end tag local name",
+	toks: []Token{
+		StartElement{Name{"", "foo"}, nil},
+		EndElement{Name{"", "bar"}},
+	},
+	err:  "xml: end tag </bar> does not match start tag <foo>",
+	want: `<foo>`,
+}, {
+	desc: "mismatching end tag namespace",
+	toks: []Token{
+		StartElement{Name{"space", "foo"}, nil},
+		EndElement{Name{"another", "foo"}},
+	},
+	err:  "xml: end tag </foo> in namespace another does not match start tag <foo> in namespace space",
+	want: `<space:foo xmlns:space="space">`,
+}, {
+	desc: "start element with explicit namespace",
+	toks: []Token{
+		StartElement{Name{"space", "local"}, []Attr{
+			{Name{"xmlns", "x"}, "space"},
+			{Name{"space", "foo"}, "value"},
+		}},
+	},
+	want: `<x:local xmlns:x="space" x:foo="value">`,
+}, {
+	desc: "start element with explicit namespace and colliding prefix",
+	toks: []Token{
+		StartElement{Name{"space", "local"}, []Attr{
+			{Name{"xmlns", "x"}, "space"},
+			{Name{"space", "foo"}, "value"},
+			{Name{"x", "bar"}, "other"},
+		}},
+	},
+	want: `<x:local xmlns:x_1="x" xmlns:x="space" x:foo="value" x_1:bar="other">`,
+}, {
+	desc: "start element using previously defined namespace",
+	toks: []Token{
+		StartElement{Name{"", "local"}, []Attr{
+			{Name{"xmlns", "x"}, "space"},
+		}},
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"space", "x"}, "y"},
+		}},
+	},
+	want: `<local xmlns:x="space"><x:foo x:x="y">`,
+}, {
+	desc: "nested name space with same prefix",
+	toks: []Token{
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"xmlns", "x"}, "space1"},
+		}},
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"xmlns", "x"}, "space2"},
+		}},
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"space1", "a"}, "space1 value"},
+			{Name{"space2", "b"}, "space2 value"},
+		}},
+		EndElement{Name{"", "foo"}},
+		EndElement{Name{"", "foo"}},
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"space1", "a"}, "space1 value"},
+			{Name{"space2", "b"}, "space2 value"},
+		}},
+	},
+	want: `<foo xmlns:x="space1"><foo xmlns:x="space2"><foo xmlns:space1="space1" space1:a="space1 value" x:b="space2 value"></foo></foo><foo xmlns:space2="space2" x:a="space1 value" space2:b="space2 value">`,
+}, {
+	desc: "start element defining several prefixes for the same name space",
+	toks: []Token{
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"xmlns", "a"}, "space"},
+			{Name{"xmlns", "b"}, "space"},
+			{Name{"space", "x"}, "value"},
+		}},
+	},
+	want: `<a:foo xmlns:a="space" a:x="value">`,
+}, {
+	desc: "nested element redefines name space",
+	toks: []Token{
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"xmlns", "x"}, "space"},
+		}},
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"xmlns", "y"}, "space"},
+			{Name{"space", "a"}, "value"},
+		}},
+	},
+	want: `<foo xmlns:x="space"><x:foo x:a="value">`,
+}, {
+	desc: "nested element creates alias for default name space",
+	toks: []Token{
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"", "xmlns"}, "space"},
+		}},
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"xmlns", "y"}, "space"},
+			{Name{"space", "a"}, "value"},
+		}},
+	},
+	want: `<foo xmlns="space"><foo xmlns:y="space" y:a="value">`,
+}, {
+	desc: "nested element defines default name space with existing prefix",
+	toks: []Token{
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"xmlns", "x"}, "space"},
+		}},
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"", "xmlns"}, "space"},
+			{Name{"space", "a"}, "value"},
+		}},
+	},
+	want: `<foo xmlns:x="space"><foo xmlns="space" x:a="value">`,
+}, {
+	desc: "nested element uses empty attribute name space when default ns defined",
+	toks: []Token{
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"", "xmlns"}, "space"},
+		}},
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"", "attr"}, "value"},
+		}},
+	},
+	want: `<foo xmlns="space"><foo attr="value">`,
+}, {
+	desc: "redefine xmlns",
+	toks: []Token{
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"foo", "xmlns"}, "space"},
+		}},
+	},
+	err: `xml: cannot redefine xmlns attribute prefix`,
+}, {
+	desc: "xmlns with explicit name space #1",
+	toks: []Token{
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"xml", "xmlns"}, "space"},
+		}},
+	},
+	want: `<foo xmlns="space">`,
+}, {
+	desc: "xmlns with explicit name space #2",
+	toks: []Token{
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{xmlURL, "xmlns"}, "space"},
+		}},
+	},
+	want: `<foo xmlns="space">`,
+}, {
+	desc: "empty name space declaration is ignored",
+	toks: []Token{
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"xmlns", "foo"}, ""},
+		}},
+	},
+	want: `<foo>`,
+}, {
+	desc: "attribute with no name is ignored",
+	toks: []Token{
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"", ""}, "value"},
+		}},
+	},
+	want: `<foo>`,
+}, {
+	desc: "namespace URL with non-valid name",
+	toks: []Token{
+		StartElement{Name{"/34", "foo"}, []Attr{
+			{Name{"/34", "x"}, "value"},
+		}},
+	},
+	want: `<_:foo xmlns:_="/34" _:x="value">`,
+}, {
+	desc: "nested element resets default namespace to empty",
+	toks: []Token{
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"", "xmlns"}, "space"},
+		}},
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"", "xmlns"}, ""},
+			{Name{"", "x"}, "value"},
+			{Name{"space", "x"}, "value"},
+		}},
+	},
+	want: `<foo xmlns="space"><foo xmlns:space="space" xmlns="" x="value" space:x="value">`,
+}, {
+	desc: "nested element requires empty default name space",
+	toks: []Token{
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"", "xmlns"}, "space"},
+		}},
+		StartElement{Name{"", "foo"}, nil},
+	},
+	want: `<foo xmlns="space"><foo xmlns="">`,
+}, {
+	desc: "attribute uses name space from xmlns",
+	toks: []Token{
+		StartElement{Name{"some/space", "foo"}, []Attr{
+			{Name{"", "attr"}, "value"},
+			{Name{"some/space", "other"}, "other value"},
+		}},
+	},
+	want: `<space:foo xmlns:space="some/space" attr="value" space:other="other value">`,
+}, {
+	desc: "default name space should not be used by attributes",
+	toks: []Token{
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"", "xmlns"}, "space"},
+			{Name{"xmlns", "bar"}, "space"},
+			{Name{"space", "baz"}, "foo"},
+		}},
+		StartElement{Name{"space", "baz"}, nil},
+		EndElement{Name{"space", "baz"}},
+		EndElement{Name{"space", "foo"}},
+	},
+	want: `<foo xmlns:bar="space" xmlns="space" bar:baz="foo"><baz></baz></foo>`,
+}, {
+	desc: "default name space not used by attributes, not explicitly defined",
+	toks: []Token{
+		StartElement{Name{"space", "foo"}, []Attr{
+			{Name{"", "xmlns"}, "space"},
+			{Name{"space", "baz"}, "foo"},
+		}},
+		StartElement{Name{"space", "baz"}, nil},
+		EndElement{Name{"space", "baz"}},
+		EndElement{Name{"space", "foo"}},
+	},
+	want: `<foo xmlns:space="space" xmlns="space" space:baz="foo"><baz></baz></foo>`,
+}, {
+	desc: "impossible xmlns declaration",
+	toks: []Token{
+		StartElement{Name{"", "foo"}, []Attr{
+			{Name{"", "xmlns"}, "space"},
+		}},
+		StartElement{Name{"space", "bar"}, []Attr{
+			{Name{"space", "attr"}, "value"},
+		}},
+	},
+	want: `<foo><space:bar xmlns:space="space" space:attr="value">`,
+}}
 
 func TestEncodeToken(t *testing.T) {
-	for _, tt := range encodeTokenTests {
+loop:
+	for i, tt := range encodeTokenTests {
 		var buf bytes.Buffer
 		enc := NewEncoder(&buf)
-		err := enc.EncodeToken(tt.tok)
+		var err error
+		for j, tok := range tt.toks {
+			err = enc.EncodeToken(tok)
+			if err != nil && j < len(tt.toks)-1 {
+				t.Errorf("#%d %s token #%d: %v", i, tt.desc, j, err)
+				continue loop
+			}
+		}
+		errorf := func(f string, a ...interface{}) {
+			t.Errorf("#%d %s token #%d:%s", i, tt.desc, len(tt.toks)-1, fmt.Sprintf(f, a...))
+		}
 		switch {
-		case !tt.ok && err == nil:
-			t.Errorf("enc.EncodeToken(%#v): expected error; got none", tt.tok)
-		case tt.ok && err != nil:
-			t.Fatalf("enc.EncodeToken: %v", err)
-		case !tt.ok && err != nil:
-			// expected error, got one
+		case tt.err != "" && err == nil:
+			errorf(" expected error; got none")
+			continue
+		case tt.err == "" && err != nil:
+			errorf(" got error: %v", err)
+			continue
+		case tt.err != "" && err != nil && tt.err != err.Error():
+			errorf(" error mismatch; got %v, want %v", err, tt.err)
+			continue
 		}
 		if err := enc.Flush(); err != nil {
-			t.Fatalf("enc.EncodeToken: %v", err)
+			errorf(" %v", err)
+			continue
 		}
 		if got := buf.String(); got != tt.want {
-			t.Errorf("enc.EncodeToken = %s; want: %s", got, tt.want)
+			errorf("\ngot  %v\nwant %v", got, tt.want)
+			continue
 		}
 	}
 }
diff --git a/src/encoding/xml/read_test.go b/src/encoding/xml/read_test.go
index 01f55d0dd0..02f1e10c33 100644
--- a/src/encoding/xml/read_test.go
+++ b/src/encoding/xml/read_test.go
@@ -5,6 +5,8 @@
 package xml
 
 import (
+	"bytes"
+	"fmt"
 	"io"
 	"reflect"
 	"strings"
@@ -484,6 +486,34 @@ func TestUnmarshalNS(t *testing.T) {
 	}
 }
 
+func TestRoundTrip(t *testing.T) {
+	// From issue 7535
+	const s = `<ex:element xmlns:ex="http://example.com/schema"></ex:element>`
+	in := bytes.NewBufferString(s)
+	for i := 0; i < 10; i++ {
+		out := &bytes.Buffer{}
+		d := NewDecoder(in)
+		e := NewEncoder(out)
+
+		for {
+			t, err := d.Token()
+			if err == io.EOF {
+				break
+			}
+			if err != nil {
+				fmt.Println("failed:", err)
+				return
+			}
+			e.EncodeToken(t)
+		}
+		e.Flush()
+		in = out
+	}
+	if got := in.String(); got != s {
+		t.Errorf("have: %q\nwant: %q\n", got, s)
+	}
+}
+
 func TestMarshalNS(t *testing.T) {
 	dst := Tables{"hello", "world"}
 	data, err := Marshal(&dst)
@@ -607,7 +637,7 @@ func TestMarshalNSAttr(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Marshal: %v", err)
 	}
-	want := `<TableAttrs><TAttr xmlns:html4="http://www.w3.org/TR/html4/" html4:table="hello" xmlns:furniture="http://www.w3schools.com/furniture" furniture:table="world" xml:lang="en_US" xmlns:_xml="http://golang.org/xml/" _xml:other="other1" xmlns:_xmlfoo="http://golang.org/xmlfoo/" _xmlfoo:other="other2" xmlns:json="http://golang.org/json/" json:other="other3" xmlns:json_1="http://golang.org/2/json/" json_1:other="other4"></TAttr></TableAttrs>`
+	want := `<TableAttrs><TAttr xmlns:json_1="http://golang.org/2/json/" xmlns:json="http://golang.org/json/" xmlns:_xmlfoo="http://golang.org/xmlfoo/" xmlns:_xml="http://golang.org/xml/" xmlns:furniture="http://www.w3schools.com/furniture" xmlns:html4="http://www.w3.org/TR/html4/" html4:table="hello" furniture:table="world" xml:lang="en_US" _xml:other="other1" _xmlfoo:other="other2" json:other="other3" json_1:other="other4"></TAttr></TableAttrs>`
 	str := string(data)
 	if str != want {
 		t.Errorf("Marshal:\nhave: %#q\nwant: %#q\n", str, want)
@@ -694,7 +724,7 @@ type Pod struct {
 	Pea interface{} `xml:"Pea"`
 }
 
-// https://code.google.com/p/go/issues/detail?id=6836
+// https://golang.org/issue/6836
 func TestUnmarshalIntoInterface(t *testing.T) {
 	pod := new(Pod)
 	pod.Pea = new(Pea)
diff --git a/src/encoding/xml/xml.go b/src/encoding/xml/xml.go
index 5690b20256..0c64cd730d 100644
--- a/src/encoding/xml/xml.go
+++ b/src/encoding/xml/xml.go
@@ -35,15 +35,24 @@ func (e *SyntaxError) Error() string {
 	return "XML syntax error on line " + strconv.Itoa(e.Line) + ": " + e.Msg
 }
 
-// A Name represents an XML name (Local) annotated
-// with a name space identifier (Space).
-// In tokens returned by Decoder.Token, the Space identifier
-// is given as a canonical URL, not the short prefix used
-// in the document being parsed.
+// A Name represents an XML name (Local) annotated with a name space
+// identifier (Space). In tokens returned by Decoder.Token, the Space
+// identifier is given as a canonical URL, not the short prefix used in
+// the document being parsed.
+//
+// As a special case, XML namespace declarations will use the literal
+// string "xmlns" for the Space field instead of the fully resolved URL.
+// See Encoder.EncodeToken for more information on namespace encoding
+// behaviour.
 type Name struct {
 	Space, Local string
 }
 
+// isNamespace reports whether the name is a namespace-defining name.
+func (name Name) isNamespace() bool {
+	return name.Local == "xmlns" || name.Space == "xmlns"
+}
+
 // An Attr represents an attribute in an XML element (Name=Value).
 type Attr struct {
 	Name  Name
@@ -72,6 +81,24 @@ func (e StartElement) End() EndElement {
 	return EndElement{e.Name}
 }
 
+// setDefaultNamespace sets the namespace of the element
+// as the default for all elements contained within it.
+func (e *StartElement) setDefaultNamespace() {
+	if e.Name.Space == "" {
+		// If there's no namespace on the element, don't
+		// set the default. Strictly speaking this might be wrong, as
+		// we can't tell if the element had no namespace set
+		// or was just using the default namespace.
+		return
+	}
+	e.Attr = append(e.Attr, Attr{
+		Name: Name{
+			Local: "xmlns",
+		},
+		Value: e.Name.Space,
+	})
+}
+
 // An EndElement represents an XML end element.
 type EndElement struct {
 	Name Name
@@ -723,7 +750,7 @@ func (d *Decoder) rawToken() (Token, error) {
 		return nil, d.err
 	}
 
-	attr = make([]Attr, 0, 4)
+	attr = []Attr{}
 	for {
 		d.space()
 		if b, ok = d.mustgetc(); !ok {
@@ -747,7 +774,11 @@ func (d *Decoder) rawToken() (Token, error) {
 
 		n := len(attr)
 		if n >= cap(attr) {
-			nattr := make([]Attr, n, 2*cap(attr))
+			nCap := 2 * cap(attr)
+			if nCap == 0 {
+				nCap = 4
+			}
+			nattr := make([]Attr, n, nCap)
 			copy(nattr, attr)
 			attr = nattr
 		}
diff --git a/src/expvar/expvar_test.go b/src/expvar/expvar_test.go
index 765e3b757e..11e6497b96 100644
--- a/src/expvar/expvar_test.go
+++ b/src/expvar/expvar_test.go
@@ -7,8 +7,11 @@ package expvar
 import (
 	"bytes"
 	"encoding/json"
+	"net"
 	"net/http/httptest"
+	"runtime"
 	"strconv"
+	"sync"
 	"testing"
 )
 
@@ -47,6 +50,26 @@ func TestInt(t *testing.T) {
 	}
 }
 
+func BenchmarkIntAdd(b *testing.B) {
+	var v Int
+
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			v.Add(1)
+		}
+	})
+}
+
+func BenchmarkIntSet(b *testing.B) {
+	var v Int
+
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			v.Set(1)
+		}
+	})
+}
+
 func TestFloat(t *testing.T) {
 	RemoveAll()
 	reqs := NewFloat("requests-float")
@@ -73,6 +96,26 @@ func TestFloat(t *testing.T) {
 	}
 }
 
+func BenchmarkFloatAdd(b *testing.B) {
+	var f Float
+
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			f.Add(1.0)
+		}
+	})
+}
+
+func BenchmarkFloatSet(b *testing.B) {
+	var f Float
+
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			f.Set(1.0)
+		}
+	})
+}
+
 func TestString(t *testing.T) {
 	RemoveAll()
 	name := NewString("my-name")
@@ -90,6 +133,16 @@ func TestString(t *testing.T) {
 	}
 }
 
+func BenchmarkStringSet(b *testing.B) {
+	var s String
+
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			s.Set("red")
+		}
+	})
+}
+
 func TestMapCounter(t *testing.T) {
 	RemoveAll()
 	colors := NewMap("bike-shed-colors")
@@ -130,6 +183,38 @@ func TestMapCounter(t *testing.T) {
 	}
 }
 
+func BenchmarkMapSet(b *testing.B) {
+	m := new(Map).Init()
+
+	v := new(Int)
+
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			m.Set("red", v)
+		}
+	})
+}
+
+func BenchmarkMapAddSame(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		m := new(Map).Init()
+		m.Add("red", 1)
+		m.Add("red", 1)
+		m.Add("red", 1)
+		m.Add("red", 1)
+	}
+}
+
+func BenchmarkMapAddDifferent(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		m := new(Map).Init()
+		m.Add("red", 1)
+		m.Add("blue", 1)
+		m.Add("green", 1)
+		m.Add("yellow", 1)
+	}
+}
+
 func TestFunc(t *testing.T) {
 	RemoveAll()
 	var x interface{} = []string{"a", "b"}
@@ -165,3 +250,135 @@ func TestHandler(t *testing.T) {
 		t.Errorf("HTTP handler wrote:\n%s\nWant:\n%s", got, want)
 	}
 }
+
+func BenchmarkRealworldExpvarUsage(b *testing.B) {
+	var (
+		bytesSent Int
+		bytesRead Int
+	)
+
+	// The benchmark creates GOMAXPROCS client/server pairs.
+	// Each pair creates 4 goroutines: client reader/writer and server reader/writer.
+	// The benchmark stresses concurrent reading and writing to the same connection.
+	// Such pattern is used in net/http and net/rpc.
+
+	b.StopTimer()
+
+	P := runtime.GOMAXPROCS(0)
+	N := b.N / P
+	W := 1000
+
+	// Setup P client/server connections.
+	clients := make([]net.Conn, P)
+	servers := make([]net.Conn, P)
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		b.Fatalf("Listen failed: %v", err)
+	}
+	defer ln.Close()
+	done := make(chan bool)
+	go func() {
+		for p := 0; p < P; p++ {
+			s, err := ln.Accept()
+			if err != nil {
+				b.Errorf("Accept failed: %v", err)
+				return
+			}
+			servers[p] = s
+		}
+		done <- true
+	}()
+	for p := 0; p < P; p++ {
+		c, err := net.Dial("tcp", ln.Addr().String())
+		if err != nil {
+			b.Fatalf("Dial failed: %v", err)
+		}
+		clients[p] = c
+	}
+	<-done
+
+	b.StartTimer()
+
+	var wg sync.WaitGroup
+	wg.Add(4 * P)
+	for p := 0; p < P; p++ {
+		// Client writer.
+		go func(c net.Conn) {
+			defer wg.Done()
+			var buf [1]byte
+			for i := 0; i < N; i++ {
+				v := byte(i)
+				for w := 0; w < W; w++ {
+					v *= v
+				}
+				buf[0] = v
+				n, err := c.Write(buf[:])
+				if err != nil {
+					b.Errorf("Write failed: %v", err)
+					return
+				}
+
+				bytesSent.Add(int64(n))
+			}
+		}(clients[p])
+
+		// Pipe between server reader and server writer.
+		pipe := make(chan byte, 128)
+
+		// Server reader.
+		go func(s net.Conn) {
+			defer wg.Done()
+			var buf [1]byte
+			for i := 0; i < N; i++ {
+				n, err := s.Read(buf[:])
+
+				if err != nil {
+					b.Errorf("Read failed: %v", err)
+					return
+				}
+
+				bytesRead.Add(int64(n))
+				pipe <- buf[0]
+			}
+		}(servers[p])
+
+		// Server writer.
+		go func(s net.Conn) {
+			defer wg.Done()
+			var buf [1]byte
+			for i := 0; i < N; i++ {
+				v := <-pipe
+				for w := 0; w < W; w++ {
+					v *= v
+				}
+				buf[0] = v
+				n, err := s.Write(buf[:])
+				if err != nil {
+					b.Errorf("Write failed: %v", err)
+					return
+				}
+
+				bytesSent.Add(int64(n))
+			}
+			s.Close()
+		}(servers[p])
+
+		// Client reader.
+		go func(c net.Conn) {
+			defer wg.Done()
+			var buf [1]byte
+			for i := 0; i < N; i++ {
+				n, err := c.Read(buf[:])
+
+				if err != nil {
+					b.Errorf("Read failed: %v", err)
+					return
+				}
+
+				bytesRead.Add(int64(n))
+			}
+			c.Close()
+		}(clients[p])
+	}
+	wg.Wait()
+}
diff --git a/src/go/build/deps_test.go b/src/go/build/deps_test.go
index b3c1105156..98201a5d96 100644
--- a/src/go/build/deps_test.go
+++ b/src/go/build/deps_test.go
@@ -240,7 +240,7 @@ var pkgDeps = map[string][]string{
 	// Basic networking.
 	// Because net must be used by any package that wants to
 	// do networking portably, it must have a small dependency set: just L1+basic os.
-	"net": {"L1", "CGO", "os", "syscall", "time"},
+	"net": {"L1", "CGO", "os", "syscall", "time", "internal/syscall/windows"},
 
 	// NET enables use of basic network-related packages.
 	"NET": {
diff --git a/src/html/template/clone_test.go b/src/html/template/clone_test.go
index e11bff2c5d..5de3bc0eef 100644
--- a/src/html/template/clone_test.go
+++ b/src/html/template/clone_test.go
@@ -166,7 +166,7 @@ func TestCloneThenParse(t *testing.T) {
 	}
 }
 
-// https://code.google.com/p/go/issues/detail?id=5980
+// https://golang.org/issue/5980
 func TestFuncMapWorksAfterClone(t *testing.T) {
 	funcs := FuncMap{"customFunc": func() (string, error) {
 		return "", errors.New("issue5980")
diff --git a/src/html/template/content_test.go b/src/html/template/content_test.go
index 5f3ffe2d32..e698328693 100644
--- a/src/html/template/content_test.go
+++ b/src/html/template/content_test.go
@@ -260,7 +260,7 @@ func TestStringer(t *testing.T) {
 	}
 }
 
-// https://code.google.com/p/go/issues/detail?id=5982
+// https://golang.org/issue/5982
 func TestEscapingNilNonemptyInterfaces(t *testing.T) {
 	tmpl := Must(New("x").Parse("{{.E}}"))
 
diff --git a/src/image/image.go b/src/image/image.go
index e0ecd92baa..951cc8ae05 100644
--- a/src/image/image.go
+++ b/src/image/image.go
@@ -46,9 +46,9 @@ type Image interface {
 }
 
 // PalettedImage is an image whose colors may come from a limited palette.
-// If m is a PalettedImage and m.ColorModel() returns a PalettedColorModel p,
+// If m is a PalettedImage and m.ColorModel() returns a color.Palette p,
 // then m.At(x, y) should be equivalent to p[m.ColorIndexAt(x, y)]. If m's
-// color model is not a PalettedColorModel, then ColorIndexAt's behavior is
+// color model is not a color.Palette, then ColorIndexAt's behavior is
 // undefined.
 type PalettedImage interface {
 	// ColorIndexAt returns the palette index of the pixel at (x, y).
diff --git a/src/internal/syscall/windows/syscall_windows.go b/src/internal/syscall/windows/syscall_windows.go
new file mode 100644
index 0000000000..2541a83440
--- /dev/null
+++ b/src/internal/syscall/windows/syscall_windows.go
@@ -0,0 +1,99 @@
+// Copyright 2014 The Go Authors.  All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package windows
+
+import (
+	"syscall"
+)
+
+//go:generate go run ../../../syscall/mksyscall_windows.go -output zsyscall_windows.go syscall_windows.go
+
+const GAA_FLAG_INCLUDE_PREFIX = 0x00000010
+
+const IF_TYPE_SOFTWARE_LOOPBACK = 24
+
+type SocketAddress struct {
+	Sockaddr       *syscall.RawSockaddrAny
+	SockaddrLength int32
+}
+
+type IpAdapterUnicastAddress struct {
+	Length             uint32
+	Flags              uint32
+	Next               *IpAdapterUnicastAddress
+	Address            SocketAddress
+	PrefixOrigin       int32
+	SuffixOrigin       int32
+	DadState           int32
+	ValidLifetime      uint32
+	PreferredLifetime  uint32
+	LeaseLifetime      uint32
+	OnLinkPrefixLength uint8
+}
+
+type IpAdapterAnycastAddress struct {
+	Length  uint32
+	Flags   uint32
+	Next    *IpAdapterAnycastAddress
+	Address SocketAddress
+}
+
+type IpAdapterMulticastAddress struct {
+	Length  uint32
+	Flags   uint32
+	Next    *IpAdapterMulticastAddress
+	Address SocketAddress
+}
+
+type IpAdapterDnsServerAdapter struct {
+	Length   uint32
+	Reserved uint32
+	Next     *IpAdapterDnsServerAdapter
+	Address  SocketAddress
+}
+
+type IpAdapterPrefix struct {
+	Length       uint32
+	Flags        uint32
+	Next         *IpAdapterPrefix
+	Address      SocketAddress
+	PrefixLength uint32
+}
+
+type IpAdapterAddresses struct {
+	Length                uint32
+	IfIndex               uint32
+	Next                  *IpAdapterAddresses
+	AdapterName           *byte
+	FirstUnicastAddress   *IpAdapterUnicastAddress
+	FirstAnycastAddress   *IpAdapterAnycastAddress
+	FirstMulticastAddress *IpAdapterMulticastAddress
+	FirstDnsServerAddress *IpAdapterDnsServerAdapter
+	DnsSuffix             *uint16
+	Description           *uint16
+	FriendlyName          *uint16
+	PhysicalAddress       [syscall.MAX_ADAPTER_ADDRESS_LENGTH]byte
+	PhysicalAddressLength uint32
+	Flags                 uint32
+	Mtu                   uint32
+	IfType                uint32
+	OperStatus            uint32
+	Ipv6IfIndex           uint32
+	ZoneIndices           [16]uint32
+	FirstPrefix           *IpAdapterPrefix
+	/* more fields might be present here. */
+}
+
+const (
+	IfOperStatusUp             = 1
+	IfOperStatusDown           = 2
+	IfOperStatusTesting        = 3
+	IfOperStatusUnknown        = 4
+	IfOperStatusDormant        = 5
+	IfOperStatusNotPresent     = 6
+	IfOperStatusLowerLayerDown = 7
+)
+
+//sys GetAdaptersAddresses(family uint32, flags uint32, reserved uintptr, adapterAddresses *IpAdapterAddresses, sizeOfPointer *uint32) (errcode error) = iphlpapi.GetAdaptersAddresses
diff --git a/src/internal/syscall/windows/zsyscall_windows.go b/src/internal/syscall/windows/zsyscall_windows.go
new file mode 100644
index 0000000000..90e2034641
--- /dev/null
+++ b/src/internal/syscall/windows/zsyscall_windows.go
@@ -0,0 +1,20 @@
+// MACHINE GENERATED BY 'go generate' COMMAND; DO NOT EDIT
+
+package windows
+
+import "unsafe"
+import "syscall"
+
+var (
+	modiphlpapi = syscall.NewLazyDLL("iphlpapi.dll")
+
+	procGetAdaptersAddresses = modiphlpapi.NewProc("GetAdaptersAddresses")
+)
+
+func GetAdaptersAddresses(family uint32, flags uint32, reserved uintptr, adapterAddresses *IpAdapterAddresses, sizeOfPointer *uint32) (errcode error) {
+	r0, _, _ := syscall.Syscall6(procGetAdaptersAddresses.Addr(), 5, uintptr(family), uintptr(flags), uintptr(reserved), uintptr(unsafe.Pointer(adapterAddresses)), uintptr(unsafe.Pointer(sizeOfPointer)), 0)
+	if r0 != 0 {
+		errcode = syscall.Errno(r0)
+	}
+	return
+}
diff --git a/src/lib9/math_darwin_arm.c b/src/lib9/math_darwin_arm.c
new file mode 100644
index 0000000000..ff5b0ce07a
--- /dev/null
+++ b/src/lib9/math_darwin_arm.c
@@ -0,0 +1,102 @@
+// Copyright 2014 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Some Darwin/ARM libc versions fail to provide a standard compliant version
+// of frexp and ldexp that could handle denormal floating point numbers.
+// The frexp and ldexp implementations are translated from their Go version.
+
+#include <stdint.h>
+
+// Assume double and uint64_t are using the same endian.
+union dint64 {
+	double d;
+	uint64_t u;
+};
+
+static const uint64_t mask = 0x7FF, bias = 1023;
+static const int shift = 64 - 11 - 1;
+static const uint64_t uvnan = 0x7FF8000000000001ULL, uvinf = 0x7FF0000000000000ULL,
+	  uvneginf = 0xFFF0000000000000ULL;
+static const double smallestnormal = 2.2250738585072014e-308; // 2**-1022
+
+static inline uint64_t float64bits(double x) {
+	union dint64 u;
+	u.d = x;
+	return u.u;
+}
+static inline double float64frombits(uint64_t x) {
+	union dint64 u;
+	u.u = x;
+	return u.d;
+}
+static inline int isinf(double x) {
+	return float64bits(x) == uvinf || float64bits(x) == uvneginf;
+}
+static inline int isnan(double x) {
+	return x != x;
+}
+extern double fabs(double);
+static double normalize(double x, int *exp) {
+	if (fabs(x) < smallestnormal) {
+		*exp = -52;
+		return x * (double)(1LL<<52);
+	}
+	*exp = 0;
+	return x;
+}
+
+double ldexp(double frac, int exp) {
+	// special cases
+	if (frac == 0.0) return frac;
+	if (isinf(frac) || isnan(frac)) return frac;
+
+	int e;
+	frac = normalize(frac, &e);
+	exp += e;
+	uint64_t x = float64bits(frac);
+	exp += (int)((x>>shift)&mask) - bias;
+	if (exp < -1074) { // underflow
+		if (frac < 0.0) return float64frombits(1ULL<<63); // -0.0
+		return 0.0;
+	}
+	if (exp > 1023) { // overflow
+		if (frac < 0.0) return float64frombits(uvneginf);
+		return float64frombits(uvinf);
+	}
+	double m = 1;
+	if (exp < -1022) { // denormal
+		exp += 52;
+		m = 1.0 / (double)(1ULL<<52);
+	}
+	x &= ~(mask << shift);
+	x |= (uint64_t)(exp+bias) << shift;
+	return m * float64frombits(x);
+}
+
+double frexp(double f, int *exp) {
+	*exp = 0;
+	// special cases
+	if (f == 0.0) return f;
+	if (isinf(f) || isnan(f)) return f;
+
+	f = normalize(f, exp);
+	uint64_t x = float64bits(f);
+	*exp += (int)((x>>shift)&mask) - bias + 1;
+	x &= ~(mask << shift);
+	x |= (-1 + bias) << shift;
+	return float64frombits(x);
+}
+
+// On Darwin/ARM, the kernel insists on running VFP in runfast mode, and it
+// cannot deal with denormal floating point numbers in that mode, so we have
+// to disable the runfast mode if the client uses ldexp/frexp (i.e. 5g).
+void disable_vfp_runfast(void) __attribute__((constructor));
+void disable_vfp_runfast(void) {
+    __asm__ volatile (
+		      "fmrx r0, fpscr\n"
+		      "bic r0, r0, $0x03000000\n"
+		      "fmxr fpscr, r0\n"
+		      : : : "r0"
+		     );
+}
diff --git a/src/liblink/asm5.c b/src/liblink/asm5.c
index 8d597750b7..398b7841c2 100644
--- a/src/liblink/asm5.c
+++ b/src/liblink/asm5.c
@@ -96,6 +96,7 @@ static Optab	optab[] =
 	{ ABL,		C_NONE,	C_NONE,	C_SBRA,		 5, 4, 0 },
 	{ ABX,		C_NONE,	C_NONE,	C_SBRA,		 74, 20, 0 },
 	{ ABEQ,		C_NONE,	C_NONE,	C_SBRA,		 5, 4, 0 },
+	{ ABEQ,		C_RCON,	C_NONE,	C_SBRA,		 5, 4, 0 }, // prediction hinted form, hint ignored
 
 	{ AB,		C_NONE,	C_NONE,	C_ROREG,	 6, 4, 0,	LPOOL },
 	{ ABL,		C_NONE,	C_NONE,	C_ROREG,	 7, 4, 0 },
@@ -353,14 +354,6 @@ static	uchar	xcmp[C_GOK+1][C_GOK+1];
 
 static LSym *deferreturn;
 
-static void
-nocache(Prog *p)
-{
-	p->optab = 0;
-	p->from.class = 0;
-	p->to.class = 0;
-}
-
 /* size of a case statement including jump table */
 static int32
 casesz(Link *ctxt, Prog *p)
@@ -1638,8 +1631,11 @@ if(0 /*debug['G']*/) print("%ux: %s: arm %d\n", (uint32)(p->pc), p->from.sym->na
 			// runtime.tlsg is special.
 			// Its "address" is the offset from the TLS thread pointer
 			// to the thread-local g and m pointers.
-			// Emit a TLS relocation instead of a standard one.
-			if(rel->sym == ctxt->tlsg) {
+			// Emit a TLS relocation instead of a standard one if its
+			// type is not explicitly set by runtime. This assumes that
+			// all references to runtime.tlsg should be accompanied with
+			// its type declaration if necessary.
+			if(rel->sym == ctxt->tlsg && ctxt->tlsg->type == 0) {
 				rel->type = R_TLS;
 				if(ctxt->flag_shared)
 					rel->add += ctxt->pc - p->pcrel->pc - 8 - rel->siz;
diff --git a/src/liblink/go.c b/src/liblink/go.c
index 754a7cc15e..08ce82b30c 100644
--- a/src/liblink/go.c
+++ b/src/liblink/go.c
@@ -133,3 +133,24 @@ double2ieee(uint64 *ieee, float64 f)
 {
 	memmove(ieee, &f, 8);
 }
+
+void
+nopout(Prog *p)
+{
+	p->as = ANOP;
+	p->scond = zprog.scond;
+	p->from = zprog.from;
+	p->from3 = zprog.from3;
+	p->reg = zprog.reg;
+	p->to = zprog.to;
+}
+
+void
+nocache(Prog *p)
+{
+	p->optab = 0;
+	p->from.class = 0;
+	p->from3.class = 0;
+	p->to.class = 0;
+}
+
diff --git a/src/liblink/list5.c b/src/liblink/list5.c
index bca10f6899..3ea411d2b8 100644
--- a/src/liblink/list5.c
+++ b/src/liblink/list5.c
@@ -168,7 +168,7 @@ Dconv(Fmt *fp)
 		if(a->u.argsize == ArgsSizeUnknown)
 			sprint(str, "$%lld", a->offset);
 		else
-			sprint(str, "$%lld-%lld", a->offset, a->u.argsize);
+			sprint(str, "$%lld-%d", a->offset, a->u.argsize);
 		break;
 
 	case TYPE_SHIFT:
diff --git a/src/liblink/list6.c b/src/liblink/list6.c
index d9e0b45c7a..b324ec89d6 100644
--- a/src/liblink/list6.c
+++ b/src/liblink/list6.c
@@ -206,7 +206,7 @@ Dconv(Fmt *fp)
 		if(a->u.argsize == ArgsSizeUnknown)
 			sprint(str, "$%lld", a->offset);
 		else
-			sprint(str, "$%lld-%lld", a->offset, a->u.argsize);
+			sprint(str, "$%lld-%d", a->offset, a->u.argsize);
 		break;
 
 	case TYPE_FCONST:
diff --git a/src/liblink/list8.c b/src/liblink/list8.c
index dbba103280..07c1d7cd67 100644
--- a/src/liblink/list8.c
+++ b/src/liblink/list8.c
@@ -193,7 +193,7 @@ Dconv(Fmt *fp)
 		if(a->u.argsize == ArgsSizeUnknown)
 			sprint(str, "$%lld", a->offset);
 		else
-			sprint(str, "$%lld-%lld", a->offset, a->u.argsize);
+			sprint(str, "$%lld-%d", a->offset, a->u.argsize);
 		break;
 
 	case TYPE_FCONST:
diff --git a/src/liblink/list9.c b/src/liblink/list9.c
index a45e73e61e..ac37bb566e 100644
--- a/src/liblink/list9.c
+++ b/src/liblink/list9.c
@@ -172,7 +172,7 @@ Dconv(Fmt *fp)
 		if(a->u.argsize == ArgsSizeUnknown)
 			sprint(str, "$%lld", a->offset);
 		else
-			sprint(str, "$%lld-%lld", a->offset, a->u.argsize);
+			sprint(str, "$%lld-%d", a->offset, a->u.argsize);
 		break;
 
 	case TYPE_MEM:
diff --git a/src/liblink/obj5.c b/src/liblink/obj5.c
index 805d7a59c2..abddbc3c4f 100644
--- a/src/liblink/obj5.c
+++ b/src/liblink/obj5.c
@@ -185,14 +185,6 @@ linkcase(Prog *casep)
 }
 
 static void
-nocache5(Prog *p)
-{
-	p->optab = 0;
-	p->from.class = 0;
-	p->to.class = 0;
-}
-
-static void
 preprocess(Link *ctxt, LSym *cursym)
 {
 	Prog *p, *pl, *p1, *p2, *q, *q1, *q2;
@@ -478,7 +470,7 @@ preprocess(Link *ctxt, LSym *cursym)
 			break;
 
 		case ARET:
-			nocache5(p);
+			nocache(p);
 			if(cursym->text->mark & LEAF) {
 				if(!autosize) {
 					p->as = AB;
diff --git a/src/liblink/obj6.c b/src/liblink/obj6.c
index cddc723bf2..546c89d53c 100644
--- a/src/liblink/obj6.c
+++ b/src/liblink/obj6.c
@@ -35,18 +35,6 @@
 #include "../cmd/6l/6.out.h"
 #include "../runtime/stack.h"
 
-static void
-nopout(Prog *p)
-{
-	p->as = ANOP;
-	p->from.type = TYPE_NONE;
-	p->from.reg = 0;
-	p->from.name = 0;
-	p->to.type = TYPE_NONE;
-	p->to.reg = 0;
-	p->to.name = 0;
-}
-
 static void nacladdr(Link*, Prog*, Addr*);
 
 static int
diff --git a/src/liblink/sym.c b/src/liblink/sym.c
index 079f600aa9..cae7e4aafe 100644
--- a/src/liblink/sym.c
+++ b/src/liblink/sym.c
@@ -148,15 +148,15 @@ linknew(LinkArch *arch)
 		switch(ctxt->arch->thechar) {
 		default:
 			sysfatal("unknown thread-local storage offset for nacl/%s", ctxt->arch->name);
+		case '5':
+			ctxt->tlsoffset = 0;
+			break;
 		case '6':
 			ctxt->tlsoffset = 0;
 			break;
 		case '8':
 			ctxt->tlsoffset = -8;
 			break;
-		case '5':
-			ctxt->tlsoffset = 0;
-			break;
 		}
 		break;
 
@@ -174,6 +174,9 @@ linknew(LinkArch *arch)
 		case '8':
 			ctxt->tlsoffset = 0x468;
 			break;
+		case '5':
+			ctxt->tlsoffset = 0; // dummy value, not needed
+			break;
 		}
 		break;
 	}
diff --git a/src/math/all_test.go b/src/math/all_test.go
index 763efb2e64..0848c506e4 100644
--- a/src/math/all_test.go
+++ b/src/math/all_test.go
@@ -2606,7 +2606,7 @@ func TestLargeTan(t *testing.T) {
 
 // Check that math constants are accepted by compiler
 // and have right value (assumes strconv.ParseFloat works).
-// http://code.google.com/p/go/issues/detail?id=201
+// http://golang.org/issue/201
 
 type floatTest struct {
 	val  interface{}
diff --git a/src/math/big/float.go b/src/math/big/float.go
index ea42a9166e..06b3ae2f40 100644
--- a/src/math/big/float.go
+++ b/src/math/big/float.go
@@ -9,7 +9,7 @@
 // rounding mode of the result operand determines the rounding
 // mode of an operation. This is a from-scratch implementation.
 
-// CAUTION: WORK IN PROGRESS - ANY ASPECT OF THIS IMPLEMENTATION MAY CHANGE!
+// CAUTION: WORK IN PROGRESS - USE AT YOUR OWN RISK.
 
 package big
 
@@ -18,47 +18,38 @@ import (
 	"math"
 )
 
-// TODO(gri): Determine if there's a more natural way to set the precision.
-// Should there be a special meaning for prec 0? Such as "full precision"?
-// (would be possible for all ops except quotient).
-
 const debugFloat = true // enable for debugging
 
-// Internal representation: A floating-point value x != 0 consists
-// of a sign (x.neg), mantissa (x.mant), and exponent (x.exp) such
-// that
-//
-//   x = sign * 0.mantissa * 2**exponent
+// A Float represents a multi-precision floating point number of the form
 //
-// and the mantissa is interpreted as a value between 0.5 and 1:
+//   sign × mantissa × 2**exponent
 //
-//  0.5 <= mantissa < 1.0
-//
-// The mantissa bits are stored in the shortest nat slice long enough
-// to hold x.prec mantissa bits. The mantissa is normalized such that
-// the msb of x.mant == 1. Thus, if the precision is not a multiple of
-// the Word size _W, x.mant[0] contains trailing zero bits. The number
-// 0 is represented by an empty mantissa and a zero exponent.
-
-// A Float represents a multi-precision floating point number
-// of the form
+// with 0.5 <= mantissa < 1.0, and MinExp <= exponent <= MaxExp (with the
+// exception of 0 and Inf which have a 0 mantissa and special exponents).
 //
-//   sign * mantissa * 2**exponent
+// Each Float value also has a precision, rounding mode, and accuracy.
 //
-// Each value also has a precision, rounding mode, and accuracy value:
-// The precision is the number of mantissa bits used to represent a
-// value, and the result of operations is rounded to that many bits
-// according to the value's rounding mode (unless specified othewise).
-// The accuracy value indicates the rounding error with respect to the
-// exact (not rounded) value.
+// The precision is the number of mantissa bits used to represent the
+// value. The rounding mode specifies how a result should be rounded
+// to fit into the mantissa bits, and accuracy describes the rounding
+// error with respect to the exact result.
 //
-// The zero value for a Float represents the number 0.
+// All operations, including setters, that specify a *Float for the result,
+// usually via the receiver, round their result to the result's precision
+// and according to its rounding mode, unless specified otherwise. If the
+// result precision is 0 (see below), it is set to the precision of the
+// argument with the largest precision value before any rounding takes
+// place.
+// TODO(gri) should the rounding mode also be copied in this case?
 //
-// By setting the desired precision to 24 (or 53) and using ToNearestEven
-// rounding, Float arithmetic operations emulate the corresponding float32
-// or float64 IEEE-754 operations (except for denormalized numbers and NaNs).
+// By setting the desired precision to 24 or 53 and using ToNearestEven
+// rounding, Float operations produce the same results as the corresponding
+// float32 or float64 IEEE-754 arithmetic for normalized operands (no NaNs
+// or denormalized numbers). Additionally, positive and negative zeros and
+// infinities are fully supported.
 //
-// CAUTION: THIS IS WORK IN PROGRESS - USE AT YOUR OWN RISK.
+// The zero (uninitialized) value for a Float is ready to use and
+// represents the number +0.0 of 0 bit precision.
 //
 type Float struct {
 	mode RoundingMode
@@ -69,31 +60,42 @@ type Float struct {
 	prec uint // TODO(gri) make this a 32bit field
 }
 
+// Internal representation details: The mantissa bits x.mant of a Float x
+// are stored in the shortest nat slice long enough to hold x.prec bits.
+// Unless x is a zero or an infinity, x.mant is normalized such that the
+// msb of x.mant == 1. Thus, if the precision is not a multiple of the
+// the Word size _W, x.mant[0] contains trailing zero bits. Zero and Inf
+// values have an empty mantissa and a 0 or infExp exponent, respectively.
+
 // NewFloat returns a new Float with value x rounded
 // to prec bits according to the given rounding mode.
+// If prec == 0, the result has value 0.0 independent
+// of the value of x.
+// BUG(gri) For prec == 0 and x == Inf, the result
+// should be Inf as well.
+// TODO(gri) rethink this signature.
 func NewFloat(x float64, prec uint, mode RoundingMode) *Float {
-	// TODO(gri) should make this more efficient
-	z := new(Float).SetFloat64(x)
-	return z.Round(z, prec, mode)
+	var z Float
+	if prec > 0 {
+		// TODO(gri) should make this more efficient
+		z.SetFloat64(x)
+		return z.Round(&z, prec, mode)
+	}
+	z.mode = mode // TODO(gri) don't do this twice for prec > 0
+	return &z
 }
 
-// infExp is the exponent value for infinity.
-const infExp = 1<<31 - 1
+const (
+	MaxExp = math.MaxInt32 // largest supported exponent magnitude
+	infExp = -MaxExp - 1   // exponent for Inf values
+)
 
-// NewInf returns a new Float with value positive infinity (sign >= 0),
-// or negative infinity (sign < 0).
+// NewInf returns a new infinite Float value with value +Inf (sign >= 0),
+// or -Inf (sign < 0).
 func NewInf(sign int) *Float {
 	return &Float{neg: sign < 0, exp: infExp}
 }
 
-func (z *Float) setExp(e int64) {
-	e32 := int32(e)
-	if int64(e32) != e {
-		panic("exponent overflow") // TODO(gri) handle this gracefully
-	}
-	z.exp = e32
-}
-
 // Accuracy describes the rounding error produced by the most recent
 // operation that generated a Float value, relative to the exact value:
 //
@@ -155,7 +157,7 @@ func (mode RoundingMode) String() string {
 }
 
 // Precision returns the mantissa precision of x in bits.
-// The precision may be 0 if x == 0. // TODO(gri) Determine a better approach.
+// The precision may be 0 for |x| == 0 or |x| == Inf.
 func (x *Float) Precision() uint {
 	return uint(x.prec)
 }
@@ -170,11 +172,111 @@ func (x *Float) Mode() RoundingMode {
 	return x.mode
 }
 
+// Sign returns:
+//
+//	-1 if x <  0
+//	 0 if x == 0 or x == -0
+//	+1 if x >  0
+//
+func (x *Float) Sign() int {
+	s := 0
+	if len(x.mant) != 0 || x.exp == infExp {
+		s = 1 // non-zero x
+	}
+	if x.neg {
+		s = -s
+	}
+	return s
+}
+
+// MantExp breaks x into its mantissa and exponent components.
+// It returns mant and exp satisfying x == mant × 2**exp, with
+// the absolute value of mant satisfying 0.5 <= |mant| < 1.0.
+// mant has the same precision and rounding mode as x.
+//
+// Special cases are:
+//
+//	(  ±0).MantExp() =   ±0, 0
+//	(±Inf).MantExp() = ±Inf, 0
+//
+// MantExp does not modify x; the result mant is a new Float.
+func (x *Float) MantExp() (mant *Float, exp int) {
+	mant = new(Float).Copy(x)
+	if x.exp != infExp {
+		mant.exp = 0
+		exp = int(x.exp)
+	}
+	return
+}
+
+// SetMantExp is the inverse of MantExp. It sets z to mant × 2**exp and
+// and returns z. The result z has the same precision and rounding mode
+// as mant.
+//
+// Special cases are:
+//
+//	z.SetMantExp(  ±0, exp) =   ±0
+//	z.SetMantExp(±Inf, exp) = ±Inf
+//
+// The result is ±Inf if the magnitude of exp is > MaxExp.
+func (z *Float) SetMantExp(mant *Float, exp int) *Float {
+	z.Copy(mant)
+	if len(z.mant) == 0 || z.exp == infExp {
+		return z
+	}
+	z.setExp(int64(exp))
+	return z
+}
+
+// IsInt reports whether x is an integer.
+// ±Inf are not considered integers.
+func (x *Float) IsInt() bool {
+	if debugFloat {
+		x.validate()
+	}
+	// pick off easy cases
+	if x.exp <= 0 {
+		// |x| < 1 || |x| == Inf
+		return len(x.mant) == 0 && x.exp != infExp
+	}
+	// x.exp > 0
+	return x.prec <= uint(x.exp) || x.minPrec() <= uint(x.exp) // not enough bits for fractional mantissa
+}
+
+// IsInf reports whether x is an infinity, according to sign.
+// If sign > 0, IsInf reports whether x is positive infinity.
+// If sign < 0, IsInf reports whether x is negative infinity.
+// If sign == 0, IsInf reports whether x is either infinity.
+func (x *Float) IsInf(sign int) bool {
+	return x.exp == infExp && (sign == 0 || x.neg == (sign < 0))
+}
+
+// setExp sets the exponent for z.
+// If the exponent's magnitude is too large, z becomes ±Inf.
+func (z *Float) setExp(e int64) {
+	if -MaxExp <= e && e <= MaxExp {
+		if len(z.mant) == 0 {
+			e = 0
+		}
+		z.exp = int32(e)
+		return
+	}
+	// Inf
+	z.mant = z.mant[:0]
+	z.exp = infExp
+}
+
 // debugging support
 func (x *Float) validate() {
-	// assumes x != 0
 	const msb = 1 << (_W - 1)
 	m := len(x.mant)
+	if m == 0 {
+		// 0.0 or Inf
+		if x.exp != 0 && x.exp != infExp {
+			panic(fmt.Sprintf("empty matissa with invalid exponent %d", x.exp))
+		}
+		return
+	}
 	if x.mant[m-1]&msb == 0 {
 		panic(fmt.Sprintf("msb not set in last word %#x of %s", x.mant[m-1], x.Format('p', 0)))
 	}
@@ -185,21 +287,24 @@ func (x *Float) validate() {
 
 // round rounds z according to z.mode to z.prec bits and sets z.acc accordingly.
 // sbit must be 0 or 1 and summarizes any "sticky bit" information one might
-// have before calling round. z's mantissa must be normalized, with the msb set.
+// have before calling round. z's mantissa must be normalized (with the msb set)
+// or empty.
 func (z *Float) round(sbit uint) {
 	z.acc = Exact
 
-	// handle zero
+	// handle zero and Inf
 	m := uint(len(z.mant)) // mantissa length in words for current precision
 	if m == 0 {
-		z.exp = 0
+		if z.exp != infExp {
+			z.exp = 0
+		}
 		return
 	}
+	// z.prec > 0
 
 	if debugFloat {
 		z.validate()
 	}
-	// z.prec > 0
 
 	bits := m * _W // available mantissa bits
 	if bits == z.prec {
@@ -342,8 +447,9 @@ func (z *Float) round(sbit uint) {
 }
 
 // Round sets z to the value of x rounded according to mode to prec bits and returns z.
+// TODO(gri) rethink this signature.
 func (z *Float) Round(x *Float, prec uint, mode RoundingMode) *Float {
-	z.Set(x)
+	z.Copy(x)
 	z.prec = prec
 	z.mode = mode
 	z.round(0)
@@ -369,24 +475,33 @@ func nlz64(x uint64) uint {
 	panic("unreachable")
 }
 
-// SetUint64 sets z to x and returns z.
-// Precision is set to 64 bits.
+// SetUint64 sets z to the (possibly rounded) value of x and returns z.
+// If z's precision is 0, it is changed to 64 (and rounding will have
+// no effect).
 func (z *Float) SetUint64(x uint64) *Float {
+	if z.prec == 0 {
+		z.prec = 64
+	}
+	z.acc = Exact
 	z.neg = false
-	z.prec = 64
 	if x == 0 {
 		z.mant = z.mant[:0]
 		z.exp = 0
 		return z
 	}
+	// x != 0
 	s := nlz64(x)
 	z.mant = z.mant.setUint64(x << s)
-	z.exp = int32(64 - s)
+	z.exp = int32(64 - s) // always fits
+	if z.prec < 64 {
+		z.round(0)
+	}
 	return z
 }
 
-// SetInt64 sets z to x and returns z.
-// Precision is set to 64 bits.
+// SetInt64 sets z to the (possibly rounded) value of x and returns z.
+// If z's precision is 0, it is changed to 64 (and rounding will have
+// no effect).
 func (z *Float) SetInt64(x int64) *Float {
 	u := x
 	if u < 0 {
@@ -397,27 +512,40 @@ func (z *Float) SetInt64(x int64) *Float {
 	return z
 }
 
-// SetFloat64 sets z to x and returns z.
-// Precision is set to 53 bits.
-// TODO(gri) test denormals, +/-Inf, disallow NaN.
+// SetFloat64 sets z to the (possibly rounded) value of x and returns z.
+// If z's precision is 0, it is changed to 53 (and rounding will have
+// no effect).
+// If x is denormalized or NaN, the result is unspecified.
+// TODO(gri) should return nil in those cases
 func (z *Float) SetFloat64(x float64) *Float {
-	z.prec = 53
-	z.neg = math.Signbit(x) // handle -0 correctly (-0 == 0)
+	if z.prec == 0 {
+		z.prec = 53
+	}
+	z.acc = Exact
+	z.neg = math.Signbit(x) // handle -0 correctly
+	if math.IsInf(x, 0) {
+		z.mant = z.mant[:0]
+		z.exp = infExp
+		return z
+	}
 	if x == 0 {
 		z.mant = z.mant[:0]
 		z.exp = 0
 		return z
 	}
+	// x != 0
 	fmant, exp := math.Frexp(x) // get normalized mantissa
 	z.mant = z.mant.setUint64(1<<63 | math.Float64bits(fmant)<<11)
-	z.exp = int32(exp)
+	z.exp = int32(exp) // always fits
+	if z.prec < 53 {
+		z.round(0)
+	}
 	return z
 }
 
 // fnorm normalizes mantissa m by shifting it to the left
-// such that the msb of the most-significant word (msw)
-// is 1. It returns the shift amount.
-// It assumes that m is not the zero nat.
+// such that the msb of the most-significant word (msw) is 1.
+// It returns the shift amount. It assumes that len(m) != 0.
 func fnorm(m nat) uint {
 	if debugFloat && (len(m) == 0 || m[len(m)-1] == 0) {
 		panic("msw of mantissa is 0")
@@ -432,38 +560,80 @@ func fnorm(m nat) uint {
 	return s
 }
 
-// SetInt sets z to x and returns z.
-// Precision is set to the number of bits required to represent x accurately.
-// TODO(gri) what about precision for x == 0?
+// SetInt sets z to the (possibly rounded) value of x and returns z.
+// If z's precision is 0, it is changed to the larger of x.BitLen()
+// or 64 (and rounding will have no effect).
 func (z *Float) SetInt(x *Int) *Float {
+	// TODO(gri) can be more efficient if z.prec > 0
+	// but small compared to the size of x, or if there
+	// are many trailing 0's.
+	bits := uint(x.BitLen())
+	if z.prec == 0 {
+		z.prec = umax(bits, 64)
+	}
+	z.acc = Exact
+	z.neg = x.neg
 	if len(x.abs) == 0 {
-		z.neg = false
 		z.mant = z.mant[:0]
 		z.exp = 0
-		// z.prec = ?
 		return z
 	}
 	// x != 0
-	z.neg = x.neg
 	z.mant = z.mant.set(x.abs)
-	e := uint(len(z.mant))*_W - fnorm(z.mant)
-	z.exp = int32(e)
-	z.prec = e
+	fnorm(z.mant)
+	z.setExp(int64(bits))
+	if z.prec < bits {
+		z.round(0)
+	}
 	return z
 }
 
-// SetRat sets z to x rounded to the precision of z and returns z.
-func (z *Float) SetRat(x *Rat, prec uint) *Float {
-	panic("unimplemented")
+// SetRat sets z to the (possibly rounded) value of x and returns z.
+// If z's precision is 0, it is changed to the largest of a.BitLen(),
+// b.BitLen(), or 64; with x = a/b.
+func (z *Float) SetRat(x *Rat) *Float {
+	// TODO(gri) can be more efficient if x is an integer
+	var a, b Float
+	a.SetInt(x.Num())
+	b.SetInt(x.Denom())
+	if z.prec == 0 {
+		z.prec = umax(a.prec, b.prec)
+	}
+	return z.Quo(&a, &b)
 }
 
-// Set sets z to x, with the same precision as x, and returns z.
+// Set sets z to the (possibly rounded) value of x and returns z.
+// If z's precision is 0, it is changed to the precision of x
+// before setting z (and rounding will have no effect).
+// Rounding is performed according to z's precision and rounding
+// mode; and z's accuracy reports the result error relative to the
+// exact (not rounded) result.
 func (z *Float) Set(x *Float) *Float {
 	if z != x {
+		if z.prec == 0 {
+			z.prec = x.prec
+		}
+		z.acc = Exact
+		z.neg = x.neg
+		z.exp = x.exp
+		z.mant = z.mant.set(x.mant)
+		if z.prec < x.prec {
+			z.round(0)
+		}
+	}
+	return z
+}
+
+// Copy sets z to x, with the same precision and rounding mode as x,
+// and returns z.
+func (z *Float) Copy(x *Float) *Float {
+	if z != x {
+		z.acc = Exact
 		z.neg = x.neg
 		z.exp = x.exp
 		z.mant = z.mant.set(x.mant)
 		z.prec = x.prec
+		z.mode = x.mode
 	}
 	return z
 }
@@ -484,29 +654,119 @@ func high64(x nat) uint64 {
 	return v
 }
 
-// TODO(gri) FIX THIS (rounding mode, errors, accuracy, etc.)
-func (x *Float) Uint64() uint64 {
-	m := high64(x.mant)
-	s := x.exp
-	if s >= 0 {
-		return m >> (64 - uint(s))
+// minPrec returns the minimum precision required to represent
+// x without loss of accuracy.
+// TODO(gri) this might be useful to export, perhaps under a better name
+func (x *Float) minPrec() uint {
+	return uint(len(x.mant))*_W - x.mant.trailingZeroBits()
+}
+
+// Uint64 returns the unsigned integer resulting from truncating x
+// towards zero. If 0 <= x <= math.MaxUint64, the result is Exact
+// if x is an integer and Below otherwise.
+// The result is (0, Above) for x < 0, and (math.MaxUint64, Below)
+// for x > math.MaxUint64.
+func (x *Float) Uint64() (uint64, Accuracy) {
+	if debugFloat {
+		x.validate()
 	}
-	return 0 // imprecise
+	switch x.ord() {
+	case -2, -1:
+		// x < 0
+		return 0, Above
+	case 0:
+		// x == 0 || x == -0
+		return 0, Exact
+	case 1:
+		// 0 < x < +Inf
+		if x.exp <= 0 {
+			// 0 < x < 1
+			return 0, Below
+		}
+		// 1 <= x < +Inf
+		if x.exp <= 64 {
+			// u = trunc(x) fits into a uint64
+			u := high64(x.mant) >> (64 - uint32(x.exp))
+			if x.minPrec() <= 64 {
+				return u, Exact
+			}
+			return u, Below // x truncated
+		}
+		fallthrough // x too large
+	case 2:
+		// x == +Inf
+		return math.MaxUint64, Below
+	}
+	panic("unreachable")
 }
 
-// TODO(gri) FIX THIS (rounding mode, errors, etc.)
-func (x *Float) Int64() int64 {
-	v := int64(x.Uint64())
-	if x.neg {
-		return -v
+// Int64 returns the integer resulting from truncating x towards zero.
+// If math.MinInt64 <= x <= math.MaxInt64, the result is Exact if x is
+// an integer, and Above (x < 0) or Below (x > 0) otherwise.
+// The result is (math.MinInt64, Above) for x < math.MinInt64, and
+// (math.MaxInt64, Below) for x > math.MaxInt64.
+func (x *Float) Int64() (int64, Accuracy) {
+	if debugFloat {
+		x.validate()
 	}
-	return v
+
+	switch x.ord() {
+	case -2:
+		// x == -Inf
+		return math.MinInt64, Above
+	case 0:
+		// x == 0 || x == -0
+		return 0, Exact
+	case -1, 1:
+		// 0 < |x| < +Inf
+		acc := Below
+		if x.neg {
+			acc = Above
+		}
+		if x.exp <= 0 {
+			// 0 < |x| < 1
+			return 0, acc
+		}
+		// 1 <= |x| < +Inf
+		if x.exp <= 63 {
+			// i = trunc(x) fits into an int64 (excluding math.MinInt64)
+			i := int64(high64(x.mant) >> (64 - uint32(x.exp)))
+			if x.neg {
+				i = -i
+			}
+			if x.minPrec() <= 63 {
+				return i, Exact
+			}
+			return i, acc // x truncated
+		}
+		if x.neg {
+			// check for special case x == math.MinInt64 (i.e., x == -(0.5 << 64))
+			if x.exp == 64 && x.minPrec() == 1 {
+				acc = Exact
+			}
+			return math.MinInt64, acc
+		}
+		fallthrough
+	case 2:
+		// x == +Inf
+		return math.MaxInt64, Below
+	}
+	panic("unreachable")
 }
 
 // Float64 returns the closest float64 value of x
 // by rounding to nearest with 53 bits precision.
 // TODO(gri) implement/document error scenarios.
 func (x *Float) Float64() (float64, Accuracy) {
+	// x == ±Inf
+	if x.exp == infExp {
+		var sign int
+		if x.neg {
+			sign = -1
+		}
+		return math.Inf(sign), Exact
+	}
+	// x == 0
 	if len(x.mant) == 0 {
 		return 0, Exact
 	}
@@ -521,39 +781,66 @@ func (x *Float) Float64() (float64, Accuracy) {
 	return math.Float64frombits(s | e<<52 | m), r.acc
 }
 
-func (x *Float) Int() *Int {
-	if len(x.mant) == 0 {
-		return new(Int)
+// Int returns the result of truncating x towards zero; or nil
+// if x is an infinity. The result is Exact if x.IsInt();
+// otherwise it is Below for x > 0, and Above for x < 0.
+func (x *Float) Int() (res *Int, acc Accuracy) {
+	if debugFloat {
+		x.validate()
 	}
-	panic("unimplemented")
-}
-
-func (x *Float) Rat() *Rat {
-	panic("unimplemented")
-}
-
-func (x *Float) IsInt() bool {
-	if len(x.mant) == 0 {
-		return true
+	// accuracy for inexact results
+	acc = Below // truncation
+	if x.neg {
+		acc = Above
 	}
+	// pick off easy cases
 	if x.exp <= 0 {
-		return false
+		// |x| < 1 || |x| == Inf
+		if x.exp == infExp {
+			return nil, acc // ±Inf
+		}
+		if len(x.mant) == 0 {
+			acc = Exact // ±0
+		}
+		return new(Int), acc // ±0.xxx
 	}
-	if uint(x.exp) >= x.prec {
-		return true
+	// x.exp > 0
+	// x.mant[len(x.mant)-1] != 0
+	// determine minimum required precision for x
+	allBits := uint(len(x.mant)) * _W
+	exp := uint(x.exp)
+	if x.minPrec() <= exp {
+		acc = Exact
 	}
+	// shift mantissa as needed
+	res = &Int{neg: x.neg}
+	// TODO(gri) should have a shift that takes positive and negative shift counts
+	switch {
+	case exp > allBits:
+		res.abs = res.abs.shl(x.mant, exp-allBits)
+	default:
+		res.abs = res.abs.set(x.mant)
+	case exp < allBits:
+		res.abs = res.abs.shr(x.mant, allBits-exp)
+	}
+	return
+}
+
+// BUG(gri) Rat is not yet implemented
+func (x *Float) Rat() *Rat {
 	panic("unimplemented")
 }
 
-// Abs sets z to |x| (the absolute value of x) and returns z.
-// TODO(gri) should Abs (and Neg) below ignore z's precision and rounding mode?
+// Abs sets z to the (possibly rounded) value |x| (the absolute value of x)
+// and returns z.
 func (z *Float) Abs(x *Float) *Float {
 	z.Set(x)
 	z.neg = false
 	return z
 }
 
-// Neg sets z to x with its sign negated, and returns z.
+// Neg sets z to the (possibly rounded) value of x with its sign negated,
+// and returns z.
 func (z *Float) Neg(x *Float) *Float {
 	z.Set(x)
 	z.neg = !z.neg
@@ -561,7 +848,7 @@ func (z *Float) Neg(x *Float) *Float {
 }
 
 // z = x + y, ignoring signs of x and y.
-// x and y must not be 0.
+// x and y must not be 0 or an Inf.
 func (z *Float) uadd(x, y *Float) {
 	// Note: This implementation requires 2 shifts most of the
 	// time. It is also inefficient if exponents or precisions
@@ -603,7 +890,7 @@ func (z *Float) uadd(x, y *Float) {
 }
 
 // z = x - y for x >= y, ignoring signs of x and y.
-// x and y must not be zero.
+// x and y must not be 0 or an Inf.
 func (z *Float) usub(x, y *Float) {
 	// This code is symmetric to uadd.
 	// We have not factored the common code out because
@@ -643,7 +930,7 @@ func (z *Float) usub(x, y *Float) {
 }
 
 // z = x * y, ignoring signs of x and y.
-// x and y must not be zero.
+// x and y must not be 0 or an Inf.
 func (z *Float) umul(x, y *Float) {
 	if debugFloat && (len(x.mant) == 0 || len(y.mant) == 0) {
 		panic("umul called with 0 argument")
@@ -664,7 +951,7 @@ func (z *Float) umul(x, y *Float) {
 }
 
 // z = x / y, ignoring signs of x and y.
-// x and y must not be zero.
+// x and y must not be 0 or an Inf.
 func (z *Float) uquo(x, y *Float) {
 	if debugFloat && (len(x.mant) == 0 || len(y.mant) == 0) {
 		panic("uquo called with 0 argument")
@@ -708,7 +995,7 @@ func (z *Float) uquo(x, y *Float) {
 }
 
 // ucmp returns -1, 0, or 1, depending on whether x < y, x == y, or x > y,
-// while ignoring the signs of x and y. x and y must not be zero.
+// while ignoring the signs of x and y. x and y must not be 0 or an Inf.
 func (x *Float) ucmp(y *Float) int {
 	if debugFloat && (len(x.mant) == 0 || len(y.mant) == 0) {
 		panic("ucmp called with 0 argument")
@@ -765,16 +1052,24 @@ func (x *Float) ucmp(y *Float) int {
 // sign as x even when x is zero.
 
 // Add sets z to the rounded sum x+y and returns z.
+// If z's precision is 0, it is changed to the larger
+// of x's or y's precision before the operation.
 // Rounding is performed according to z's precision
 // and rounding mode; and z's accuracy reports the
 // result error relative to the exact (not rounded)
 // result.
 func (z *Float) Add(x, y *Float) *Float {
+	if z.prec == 0 {
+		z.prec = umax(x.prec, y.prec)
+	}
+
 	// TODO(gri) what about -0?
 	if len(y.mant) == 0 {
+		// TODO(gri) handle Inf
 		return z.Round(x, z.prec, z.mode)
 	}
 	if len(x.mant) == 0 {
+		// TODO(gri) handle Inf
 		return z.Round(y, z.prec, z.mode)
 	}
 
@@ -799,13 +1094,15 @@ func (z *Float) Add(x, y *Float) *Float {
 }
 
 // Sub sets z to the rounded difference x-y and returns z.
-// Rounding is performed according to z's precision
-// and rounding mode; and z's accuracy reports the
-// result error relative to the exact (not rounded)
-// result.
+// Precision, rounding, and accuracy reporting are as for Add.
 func (z *Float) Sub(x, y *Float) *Float {
+	if z.prec == 0 {
+		z.prec = umax(x.prec, y.prec)
+	}
+
 	// TODO(gri) what about -0?
 	if len(y.mant) == 0 {
+		// TODO(gri) handle Inf
 		return z.Round(x, z.prec, z.mode)
 	}
 	if len(x.mant) == 0 {
@@ -836,11 +1133,14 @@ func (z *Float) Sub(x, y *Float) *Float {
 }
 
 // Mul sets z to the rounded product x*y and returns z.
-// Rounding is performed according to z's precision
-// and rounding mode; and z's accuracy reports the
-// result error relative to the exact (not rounded)
-// result.
+// Precision, rounding, and accuracy reporting are as for Add.
 func (z *Float) Mul(x, y *Float) *Float {
+	if z.prec == 0 {
+		z.prec = umax(x.prec, y.prec)
+	}
+
+	// TODO(gri) handle Inf
+
 	// TODO(gri) what about -0?
 	if len(x.mant) == 0 || len(y.mant) == 0 {
 		z.neg = false
@@ -857,47 +1157,61 @@ func (z *Float) Mul(x, y *Float) *Float {
 }
 
 // Quo sets z to the rounded quotient x/y and returns z.
-// If y == 0, a division-by-zero run-time panic occurs. TODO(gri) this should become Inf
-// Rounding is performed according to z's precision
-// and rounding mode; and z's accuracy reports the
-// result error relative to the exact (not rounded)
-// result.
+// Precision, rounding, and accuracy reporting are as for Add.
 func (z *Float) Quo(x, y *Float) *Float {
-	// TODO(gri) what about -0?
+	if z.prec == 0 {
+		z.prec = umax(x.prec, y.prec)
+	}
+
+	// TODO(gri) handle Inf
+
+	// TODO(gri) check that this is correct
+	z.neg = x.neg != y.neg
+
+	if len(y.mant) == 0 {
+		z.setExp(infExp)
+		return z
+	}
+
 	if len(x.mant) == 0 {
-		z.neg = false
 		z.mant = z.mant[:0]
 		z.exp = 0
 		z.acc = Exact
 		return z
 	}
-	if len(y.mant) == 0 {
-		panic("division-by-zero") // TODO(gri) handle this better
-	}
 
 	// x, y != 0
 	z.uquo(x, y)
-	z.neg = x.neg != y.neg
 	return z
 }
 
 // Lsh sets z to the rounded x * (1<<s) and returns z.
+// If z's precision is 0, it is changed to x's precision.
 // Rounding is performed according to z's precision
 // and rounding mode; and z's accuracy reports the
 // result error relative to the exact (not rounded)
 // result.
 func (z *Float) Lsh(x *Float, s uint, mode RoundingMode) *Float {
+	if z.prec == 0 {
+		z.prec = x.prec
+	}
+
+	// TODO(gri) handle Inf
+
 	z.Round(x, z.prec, mode)
 	z.setExp(int64(z.exp) + int64(s))
 	return z
 }
 
 // Rsh sets z to the rounded x / (1<<s) and returns z.
-// Rounding is performed according to z's precision
-// and rounding mode; and z's accuracy reports the
-// result error relative to the exact (not rounded)
-// result.
+// Precision, rounding, and accuracy reporting are as for Lsh.
 func (z *Float) Rsh(x *Float, s uint, mode RoundingMode) *Float {
+	if z.prec == 0 {
+		z.prec = x.prec
+	}
+
+	// TODO(gri) handle Inf
+
 	z.Round(x, z.prec, mode)
 	z.setExp(int64(z.exp) - int64(s))
 	return z
@@ -909,49 +1223,60 @@ func (z *Float) Rsh(x *Float, s uint, mode RoundingMode) *Float {
 //    0 if x == y (incl. -0 == 0)
 //   +1 if x >  y
 //
+// Infinities with matching sign are equal.
 func (x *Float) Cmp(y *Float) int {
-	// special cases
-	switch {
-	case len(x.mant) == 0:
-		// 0 cmp y == -sign(y)
-		return -y.Sign()
-	case len(y.mant) == 0:
-		// x cmp 0 == sign(x)
-		return x.Sign()
+	if debugFloat {
+		x.validate()
+		y.validate()
 	}
-	// x != 0 && y != 0
 
-	// x cmp y == x cmp y
-	// x cmp (-y) == 1
-	// (-x) cmp y == -1
-	// (-x) cmp (-y) == -(x cmp y)
+	mx := x.ord()
+	my := y.ord()
 	switch {
-	case x.neg == y.neg:
-		r := x.ucmp(y)
-		if x.neg {
-			r = -r
-		}
-		return r
-	case x.neg:
+	case mx < my:
 		return -1
-	default:
-		return 1
+	case mx > my:
+		return +1
 	}
+	// mx == my
+
+	// only if |mx| == 1 we have to compare the mantissae
+	switch mx {
+	case -1:
+		return -x.ucmp(y)
+	case +1:
+		return +x.ucmp(y)
+	}
+
 	return 0
 }
 
-// Sign returns:
+func umax(x, y uint) uint {
+	if x > y {
+		return x
+	}
+	return y
+}
+
+// ord classifies x and returns:
 //
-//	-1 if x <  0
-//	 0 if x == 0 (incl. x == -0)
-//	+1 if x >  0
+//	-2 if -Inf == x
+//	-1 if -Inf < x < 0
+//	 0 if x == 0 (signed or unsigned)
+//	+1 if 0 < x < +Inf
+//	+2 if x == +Inf
 //
-func (x *Float) Sign() int {
+// TODO(gri) export (and remove IsInf)?
+func (x *Float) ord() int {
+	m := 1 // common case
 	if len(x.mant) == 0 {
-		return 0
+		m = 0
+		if x.exp == infExp {
+			m = 2
+		}
 	}
 	if x.neg {
-		return -1
+		m = -m
 	}
-	return 1
+	return m
 }
diff --git a/src/math/big/float_test.go b/src/math/big/float_test.go
index ec67a6d606..dd059ba4a5 100644
--- a/src/math/big/float_test.go
+++ b/src/math/big/float_test.go
@@ -6,11 +6,210 @@ package big
 
 import (
 	"fmt"
+	"math"
 	"sort"
 	"strconv"
+	"strings"
 	"testing"
 )
 
+func (x *Float) uint64() uint64 {
+	u, acc := x.Uint64()
+	if acc != Exact {
+		panic(fmt.Sprintf("%s is not a uint64", x.Format('g', 10)))
+	}
+	return u
+}
+
+func (x *Float) int64() int64 {
+	i, acc := x.Int64()
+	if acc != Exact {
+		panic(fmt.Sprintf("%s is not an int64", x.Format('g', 10)))
+	}
+	return i
+}
+
+func TestFloatZeroValue(t *testing.T) {
+	// zero (uninitialized) value is a ready-to-use 0.0
+	var x Float
+	if s := x.Format('f', 1); s != "0.0" {
+		t.Errorf("zero value = %s; want 0.0", s)
+	}
+
+	// zero value has precision 0
+	if prec := x.Precision(); prec != 0 {
+		t.Errorf("prec = %d; want 0", prec)
+	}
+
+	// zero value can be used in any and all positions of binary operations
+	make := func(x int) *Float {
+		if x == 0 {
+			return new(Float) // 0 translates into the zero value
+		}
+		return NewFloat(float64(x), 10, 0)
+	}
+	for _, test := range []struct {
+		z, x, y, want int
+		opname        rune
+		op            func(z, x, y *Float) *Float
+	}{
+		{0, 0, 0, 0, '+', (*Float).Add},
+		{0, 1, 2, 3, '+', (*Float).Add},
+		{1, 2, 0, 2, '+', (*Float).Add},
+		{2, 0, 1, 1, '+', (*Float).Add},
+
+		{0, 0, 0, 0, '-', (*Float).Sub},
+		{0, 1, 2, -1, '-', (*Float).Sub},
+		{1, 2, 0, 2, '-', (*Float).Sub},
+		{2, 0, 1, -1, '-', (*Float).Sub},
+
+		{0, 0, 0, 0, '*', (*Float).Mul},
+		{0, 1, 2, 2, '*', (*Float).Mul},
+		{1, 2, 0, 0, '*', (*Float).Mul},
+		{2, 0, 1, 0, '*', (*Float).Mul},
+
+		{0, 0, 0, 0, '/', (*Float).Quo}, // = +Inf
+		{0, 2, 1, 2, '/', (*Float).Quo},
+		{1, 2, 0, 0, '/', (*Float).Quo}, // = +Inf
+		{2, 0, 1, 0, '/', (*Float).Quo},
+	} {
+		z := make(test.z)
+		test.op(z, make(test.x), make(test.y))
+		got := 0
+		if !z.IsInf(0) {
+			got = int(z.int64())
+		}
+		if got != test.want {
+			t.Errorf("%d %c %d = %d; want %d", test.x, test.opname, test.y, got, test.want)
+		}
+	}
+
+	// TODO(gri) test how precision is set for zero value results
+}
+
+func makeFloat(s string) *Float {
+	if s == "Inf" || s == "+Inf" {
+		return NewInf(+1)
+	}
+	if s == "-Inf" {
+		return NewInf(-1)
+	}
+	var x Float
+	x.prec = 1000 // TODO(gri) find a better way to do this
+	if _, ok := x.SetString(s); !ok {
+		panic(fmt.Sprintf("%q is not a valid float", s))
+	}
+	return &x
+}
+
+func TestFloatSign(t *testing.T) {
+	for _, test := range []struct {
+		x string
+		s int
+	}{
+		{"-Inf", -1},
+		{"-1", -1},
+		{"-0", 0},
+		{"+0", 0},
+		{"+1", +1},
+		{"+Inf", +1},
+	} {
+		x := makeFloat(test.x)
+		s := x.Sign()
+		if s != test.s {
+			t.Errorf("%s.Sign() = %d; want %d", test.x, s, test.s)
+		}
+	}
+}
+
+// feq(x, y) is like x.Cmp(y) == 0 but it also considers the sign of 0 (0 != -0).
+func feq(x, y *Float) bool {
+	return x.Cmp(y) == 0 && x.neg == y.neg
+}
+
+func TestFloatMantExp(t *testing.T) {
+	for _, test := range []struct {
+		x    string
+		frac string
+		exp  int
+	}{
+		{"0", "0", 0},
+		{"+0", "0", 0},
+		{"-0", "-0", 0},
+		{"Inf", "+Inf", 0},
+		{"+Inf", "+Inf", 0},
+		{"-Inf", "-Inf", 0},
+		{"1.5", "0.75", 1},
+		{"1.024e3", "0.5", 11},
+		{"-0.125", "-0.5", -2},
+	} {
+		x := makeFloat(test.x)
+		frac := makeFloat(test.frac)
+		f, e := x.MantExp()
+		if !feq(f, frac) || e != test.exp {
+			t.Errorf("%s.MantExp() = %s, %d; want %s, %d", test.x, f.Format('g', 10), e, test.frac, test.exp)
+		}
+	}
+}
+
+func TestFloatSetMantExp(t *testing.T) {
+	for _, test := range []struct {
+		frac string
+		exp  int
+		z    string
+	}{
+		{"0", 0, "0"},
+		{"+0", 0, "0"},
+		{"-0", 0, "-0"},
+		{"Inf", 1234, "+Inf"},
+		{"+Inf", -1234, "+Inf"},
+		{"-Inf", -1234, "-Inf"},
+		{"0", -MaxExp - 1, "0"},
+		{"1", -MaxExp - 1, "+Inf"},  // exponent magnitude too large
+		{"-1", -MaxExp - 1, "-Inf"}, // exponent magnitude too large
+		{"0.75", 1, "1.5"},
+		{"0.5", 11, "1024"},
+		{"-0.5", -2, "-0.125"},
+	} {
+		frac := makeFloat(test.frac)
+		want := makeFloat(test.z)
+		var z Float
+		z.SetMantExp(frac, test.exp)
+		if !feq(&z, want) {
+			t.Errorf("SetMantExp(%s, %d) = %s; want %s", test.frac, test.exp, z.Format('g', 10), test.z)
+		}
+	}
+}
+
+func TestFloatIsInt(t *testing.T) {
+	for _, test := range []string{
+		"0 int",
+		"-0 int",
+		"1 int",
+		"-1 int",
+		"0.5",
+		"1.23",
+		"1.23e1",
+		"1.23e2 int",
+		"0.000000001e+8",
+		"0.000000001e+9 int",
+		"1.2345e200 int",
+		"Inf",
+		"+Inf",
+		"-Inf",
+	} {
+		s := strings.TrimSuffix(test, " int")
+		want := s != test
+		if got := makeFloat(s).IsInt(); got != want {
+			t.Errorf("%s.IsInt() == %t", s, got)
+		}
+	}
+}
+
+func TestFloatIsInf(t *testing.T) {
+	// TODO(gri) implement this
+}
+
 func fromBinary(s string) int64 {
 	x, err := strconv.ParseInt(s, 2, 64)
 	if err != nil {
@@ -66,7 +265,7 @@ func testFloatRound(t *testing.T, x, r int64, prec uint, mode RoundingMode) {
 	f.Round(f, prec, mode)
 
 	// check result
-	r1 := f.Int64()
+	r1 := f.int64()
 	p1 := f.Precision()
 	a1 := f.Accuracy()
 	if r1 != r || p1 != prec || a1 != a {
@@ -203,9 +402,21 @@ func TestFloatSetUint64(t *testing.T) {
 		1 << 32,
 		1<<64 - 1,
 	} {
-		f := new(Float).SetUint64(want)
-		if got := f.Uint64(); got != want {
-			t.Errorf("got %d (%s); want %d", got, f.Format('p', 0), want)
+		var f Float
+		f.SetUint64(want)
+		if got := f.uint64(); got != want {
+			t.Errorf("got %#x (%s); want %#x", got, f.Format('p', 0), want)
+		}
+	}
+
+	// test basic rounding behavior (exhaustive rounding testing is done elsewhere)
+	const x uint64 = 0x8765432187654321 // 64 bits needed
+	for prec := uint(1); prec <= 64; prec++ {
+		f := NewFloat(0, prec, ToZero).SetUint64(x)
+		got := f.uint64()
+		want := x &^ (1<<(64-prec) - 1) // cut off (round to zero) low 64-prec bits
+		if got != want {
+			t.Errorf("got %#x (%s); want %#x", got, f.Format('p', 0), want)
 		}
 	}
 }
@@ -225,12 +436,24 @@ func TestFloatSetInt64(t *testing.T) {
 			if i&1 != 0 {
 				want = -want
 			}
-			f := new(Float).SetInt64(want)
-			if got := f.Int64(); got != want {
-				t.Errorf("got %d (%s); want %d", got, f.Format('p', 0), want)
+			var f Float
+			f.SetInt64(want)
+			if got := f.int64(); got != want {
+				t.Errorf("got %#x (%s); want %#x", got, f.Format('p', 0), want)
 			}
 		}
 	}
+
+	// test basic rounding behavior (exhaustive rounding testing is done elsewhere)
+	const x int64 = 0x7654321076543210 // 63 bits needed
+	for prec := uint(1); prec <= 63; prec++ {
+		f := NewFloat(0, prec, ToZero).SetInt64(x)
+		got := f.int64()
+		want := x &^ (1<<(63-prec) - 1) // cut off (round to zero) low 63-prec bits
+		if got != want {
+			t.Errorf("got %#x (%s); want %#x", got, f.Format('p', 0), want)
+		}
+	}
 }
 
 func TestFloatSetFloat64(t *testing.T) {
@@ -244,21 +467,265 @@ func TestFloatSetFloat64(t *testing.T) {
 		3.14159265e10,
 		2.718281828e-123,
 		1.0 / 3,
+		math.Inf(-1),
+		math.Inf(0),
+		-math.Inf(1),
 	} {
 		for i := range [2]int{} {
 			if i&1 != 0 {
 				want = -want
 			}
-			f := new(Float).SetFloat64(want)
+			var f Float
+			f.SetFloat64(want)
 			if got, _ := f.Float64(); got != want {
 				t.Errorf("got %g (%s); want %g", got, f.Format('p', 0), want)
 			}
 		}
 	}
+
+	// test basic rounding behavior (exhaustive rounding testing is done elsewhere)
+	const x uint64 = 0x8765432143218 // 53 bits needed
+	for prec := uint(1); prec <= 52; prec++ {
+		f := NewFloat(0, prec, ToZero).SetFloat64(float64(x))
+		got, _ := f.Float64()
+		want := float64(x &^ (1<<(52-prec) - 1)) // cut off (round to zero) low 53-prec bits
+		if got != want {
+			t.Errorf("got %g (%s); want %g", got, f.Format('p', 0), want)
+		}
+	}
 }
 
 func TestFloatSetInt(t *testing.T) {
-	// TODO(gri) implement
+	for _, want := range []string{
+		"0",
+		"1",
+		"-1",
+		"1234567890",
+		"123456789012345678901234567890",
+		"123456789012345678901234567890123456789012345678901234567890123456789012345678901234567890",
+	} {
+		var x Int
+		_, ok := x.SetString(want, 0)
+		if !ok {
+			t.Errorf("invalid integer %s", want)
+			continue
+		}
+		n := x.BitLen()
+
+		var f Float
+		f.SetInt(&x)
+
+		// check precision
+		if n < 64 {
+			n = 64
+		}
+		if prec := f.Precision(); prec != uint(n) {
+			t.Errorf("got prec = %d; want %d", prec, n)
+		}
+
+		// check value
+		got := f.Format('g', 100)
+		if got != want {
+			t.Errorf("got %s (%s); want %s", got, f.Format('p', 0), want)
+		}
+	}
+
+	// TODO(gri) test basic rounding behavior
+}
+
+func TestFloatSetRat(t *testing.T) {
+	for _, want := range []string{
+		"0",
+		"1",
+		"-1",
+		"1234567890",
+		"123456789012345678901234567890",
+		"123456789012345678901234567890123456789012345678901234567890123456789012345678901234567890",
+		"1.2",
+		"3.14159265",
+		// TODO(gri) expand
+	} {
+		var x Rat
+		_, ok := x.SetString(want)
+		if !ok {
+			t.Errorf("invalid fraction %s", want)
+			continue
+		}
+		n := max(x.Num().BitLen(), x.Denom().BitLen())
+
+		var f1 Float
+		var f2 = NewFloat(0, 1000, 0) // set a high precision - TODO(gri) find a cleaner way
+		f1.SetRat(&x)
+		f2.SetRat(&x)
+
+		// check precision when set automatically
+		if n < 64 {
+			n = 64
+		}
+		if prec := f1.Precision(); prec != uint(n) {
+			t.Errorf("got prec = %d; want %d", prec, n)
+		}
+
+		got := f2.Format('g', 100)
+		if got != want {
+			t.Errorf("got %s (%s); want %s", got, f2.Format('p', 0), want)
+		}
+	}
+}
+
+func TestFloatUint64(t *testing.T) {
+	for _, test := range []struct {
+		x   string
+		out uint64
+		acc Accuracy
+	}{
+		{"-Inf", 0, Above},
+		{"-1", 0, Above},
+		{"-1e-1000", 0, Above},
+		{"-0", 0, Exact},
+		{"0", 0, Exact},
+		{"1e-1000", 0, Below},
+		{"1", 1, Exact},
+		{"1.000000000000000000001", 1, Below},
+		{"12345.0", 12345, Exact},
+		{"12345.000000000000000000001", 12345, Below},
+		{"18446744073709551615", 18446744073709551615, Exact},
+		{"18446744073709551615.000000000000000000001", math.MaxUint64, Below},
+		{"18446744073709551616", math.MaxUint64, Below},
+		{"1e10000", math.MaxUint64, Below},
+		{"+Inf", math.MaxUint64, Below},
+	} {
+		x := makeFloat(test.x)
+		out, acc := x.Uint64()
+		if out != test.out || acc != test.acc {
+			t.Errorf("%s: got %d (%s); want %d (%s)", test.x, out, acc, test.out, test.acc)
+		}
+	}
+}
+
+func TestFloatInt64(t *testing.T) {
+	for _, test := range []struct {
+		x   string
+		out int64
+		acc Accuracy
+	}{
+		{"-Inf", math.MinInt64, Above},
+		{"-1e10000", math.MinInt64, Above},
+		{"-9223372036854775809", math.MinInt64, Above},
+		{"-9223372036854775808.000000000000000000001", math.MinInt64, Above},
+		{"-9223372036854775808", -9223372036854775808, Exact},
+		{"-9223372036854775807.000000000000000000001", -9223372036854775807, Above},
+		{"-9223372036854775807", -9223372036854775807, Exact},
+		{"-12345.000000000000000000001", -12345, Above},
+		{"-12345.0", -12345, Exact},
+		{"-1.000000000000000000001", -1, Above},
+		{"-1", -1, Exact},
+		{"-1e-1000", 0, Above},
+		{"0", 0, Exact},
+		{"1e-1000", 0, Below},
+		{"1", 1, Exact},
+		{"1.000000000000000000001", 1, Below},
+		{"12345.0", 12345, Exact},
+		{"12345.000000000000000000001", 12345, Below},
+		{"9223372036854775807", 9223372036854775807, Exact},
+		{"9223372036854775807.000000000000000000001", math.MaxInt64, Below},
+		{"9223372036854775808", math.MaxInt64, Below},
+		{"1e10000", math.MaxInt64, Below},
+		{"+Inf", math.MaxInt64, Below},
+	} {
+		x := makeFloat(test.x)
+		out, acc := x.Int64()
+		if out != test.out || acc != test.acc {
+			t.Errorf("%s: got %d (%s); want %d (%s)", test.x, out, acc, test.out, test.acc)
+		}
+	}
+}
+
+func TestFloatInt(t *testing.T) {
+	for _, test := range []struct {
+		x   string
+		out string
+		acc Accuracy
+	}{
+		{"0", "0", Exact},
+		{"+0", "0", Exact},
+		{"-0", "0", Exact},
+		{"Inf", "nil", Below},
+		{"+Inf", "nil", Below},
+		{"-Inf", "nil", Above},
+		{"1", "1", Exact},
+		{"-1", "-1", Exact},
+		{"1.23", "1", Below},
+		{"-1.23", "-1", Above},
+		{"123e-2", "1", Below},
+		{"123e-3", "0", Below},
+		{"123e-4", "0", Below},
+		{"1e-1000", "0", Below},
+		{"-1e-1000", "0", Above},
+		{"1e+10", "10000000000", Exact},
+		{"1e+100", "10000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000", Exact},
+	} {
+		x := makeFloat(test.x)
+		out, acc := x.Int()
+		got := "nil"
+		if out != nil {
+			got = out.String()
+		}
+		if got != test.out || acc != test.acc {
+			t.Errorf("%s: got %s (%s); want %s (%s)", test.x, got, acc, test.out, test.acc)
+		}
+	}
+}
+
+func TestFloatRat(t *testing.T) {
+	// TODO(gri) implement this
+}
+
+func TestFloatAbs(t *testing.T) {
+	for _, test := range []string{
+		"0",
+		"1",
+		"1234",
+		"1.23e-2",
+		"1e-1000",
+		"1e1000",
+		"Inf",
+	} {
+		p := makeFloat(test)
+		a := new(Float).Abs(p)
+		if !feq(a, p) {
+			t.Errorf("%s: got %s; want %s", test, a.Format('g', 10), test)
+		}
+
+		n := makeFloat("-" + test)
+		a.Abs(n)
+		if !feq(a, p) {
+			t.Errorf("-%s: got %s; want %s", test, a.Format('g', 10), test)
+		}
+	}
+}
+
+func TestFloatNeg(t *testing.T) {
+	for _, test := range []string{
+		"0",
+		"1",
+		"1234",
+		"1.23e-2",
+		"1e-1000",
+		"1e1000",
+		"Inf",
+	} {
+		p1 := makeFloat(test)
+		n1 := makeFloat("-" + test)
+		n2 := new(Float).Neg(p1)
+		p2 := new(Float).Neg(n2)
+		if !feq(n2, n1) {
+			t.Errorf("%s: got %s; want %s", test, n2.Format('g', 10), n1.Format('g', 10))
+		}
+		if !feq(p2, p1) {
+			t.Errorf("%s: got %s; want %s", test, p2.Format('g', 10), p1.Format('g', 10))
+		}
+	}
 }
 
 // Selected precisions with which to run various tests.
@@ -386,6 +853,7 @@ func TestFloatAdd64(t *testing.T) {
 }
 
 func TestFloatMul(t *testing.T) {
+	// TODO(gri) implement this
 }
 
 // TestFloatMul64 tests that Float.Mul/Quo of numbers with
@@ -521,6 +989,51 @@ func TestFloatQuo(t *testing.T) {
 	}
 }
 
+// TestFloatQuoSmoke tests all divisions x/y for values x, y in the range [-n, +n];
+// it serves as a smoke test for basic correctness of division.
+func TestFloatQuoSmoke(t *testing.T) {
+	n := 1000
+	if testing.Short() {
+		n = 10
+	}
+
+	const dprec = 3         // max. precision variation
+	const prec = 10 + dprec // enough bits to hold n precisely
+	for x := -n; x <= n; x++ {
+		for y := -n; y < n; y++ {
+			if y == 0 {
+				continue
+			}
+
+			a := float64(x)
+			b := float64(y)
+			c := a / b
+
+			// vary operand precision (only ok as long as a, b can be represented correctly)
+			for ad := -dprec; ad <= dprec; ad++ {
+				for bd := -dprec; bd <= dprec; bd++ {
+					A := NewFloat(a, uint(prec+ad), 0)
+					B := NewFloat(b, uint(prec+bd), 0)
+					C := NewFloat(0, 53, 0).Quo(A, B) // C has float64 mantissa width
+
+					cc, acc := C.Float64()
+					if cc != c {
+						t.Errorf("%g/%g = %s; want %.5g\n", a, b, C.Format('g', 5), c)
+						continue
+					}
+					if acc != Exact {
+						t.Errorf("%g/%g got %s result; want exact result", a, b, acc)
+					}
+				}
+			}
+		}
+	}
+}
+
+func TestFloatCmp(t *testing.T) {
+	// TODO(gri) implement this
+}
+
 // normBits returns the normalized bits for x: It
 // removes multiple equal entries by treating them
 // as an addition (e.g., []int{5, 5} => []int{6}),
diff --git a/src/math/big/floatconv.go b/src/math/big/floatconv.go
index 06c1f14471..71920cd51c 100644
--- a/src/math/big/floatconv.go
+++ b/src/math/big/floatconv.go
@@ -57,6 +57,7 @@ func (z *Float) SetString(s string) (*Float, bool) {
 // with base 0 or 10 corresponds to the value 1.2 * 2**3.
 //
 // BUG(gri) This signature conflicts with Scan(s fmt.ScanState, ch rune) error.
+// TODO(gri) What should the default precision be?
 func (z *Float) Scan(r io.ByteScanner, base int) (f *Float, b int, err error) {
 	// sign
 	z.neg, err = scanSign(r)
@@ -162,8 +163,8 @@ func ParseFloat(s string, base int, prec uint, mode RoundingMode) (f *Float, b i
 // Format converts the floating-point number x to a string according
 // to the given format and precision prec. The format is one of:
 //
-//	'e'	-d.dddde±dd, decimal exponent
-//	'E'	-d.ddddE±dd, decimal exponent
+//	'e'	-d.dddde±dd, decimal exponent, at least two (possibly 0) exponent digits
+//	'E'	-d.ddddE±dd, decimal exponent, at least two (possibly 0) exponent digits
 //	'f'	-ddddd.dddd, no exponent
 //	'g'	like 'e' for large exponents, like 'f' otherwise
 //	'G'	like 'E' for large exponents, like 'f' otherwise
@@ -182,7 +183,7 @@ func ParseFloat(s string, base int, prec uint, mode RoundingMode) (f *Float, b i
 // number of digits necessary such that ParseFloat will return f exactly.
 // The prec value is ignored for the 'b' or 'p' format.
 //
-// BUG(gri) Currently, Format only accepts the 'b' and 'p' format.
+// BUG(gri) Currently, Format does not accept negative precisions.
 func (x *Float) Format(format byte, prec int) string {
 	const extra = 10 // TODO(gri) determine a good/better value here
 	return string(x.Append(make([]byte, 0, prec+extra), format, prec))
@@ -191,13 +192,27 @@ func (x *Float) Format(format byte, prec int) string {
 // Append appends the string form of the floating-point number x,
 // as generated by x.Format, to buf and returns the extended buffer.
 func (x *Float) Append(buf []byte, format byte, prec int) []byte {
+	// TODO(gri) factor out handling of sign?
+
+	// Inf
+	if x.IsInf(0) {
+		var ch byte = '+'
+		if x.neg {
+			ch = '-'
+		}
+		buf = append(buf, ch)
+		return append(buf, "Inf"...)
+	}
+
+	// easy formats
 	switch format {
 	case 'b':
 		return x.bstring(buf)
 	case 'p':
 		return x.pstring(buf)
 	}
-	return append(buf, fmt.Sprintf(`%%!c`, format)...)
+
+	return x.bigFtoa(buf, format, prec)
 }
 
 // BUG(gri): Currently, String uses the 'p' (rather than 'g') format.
@@ -211,7 +226,6 @@ func (x *Float) String() string {
 // The mantissa is normalized such that is uses x.Precision() bits in binary
 // representation.
 func (x *Float) bstring(buf []byte) []byte {
-	// TODO(gri) handle Inf
 	if x.neg {
 		buf = append(buf, '-')
 	}
@@ -239,7 +253,6 @@ func (x *Float) bstring(buf []byte) []byte {
 // ad returns the extended buffer.
 // The mantissa is normalized such that 0.5 <= 0.mantissa < 1.0.
 func (x *Float) pstring(buf []byte) []byte {
-	// TODO(gri) handle Inf
 	if x.neg {
 		buf = append(buf, '-')
 	}
diff --git a/src/math/big/floatconv_test.go b/src/math/big/floatconv_test.go
index 0e8bfb39ab..11e5df448a 100644
--- a/src/math/big/floatconv_test.go
+++ b/src/math/big/floatconv_test.go
@@ -5,6 +5,8 @@
 package big
 
 import (
+	"io"
+	"math"
 	"strconv"
 	"testing"
 )
@@ -71,37 +73,157 @@ func TestFloatSetFloat64String(t *testing.T) {
 	}
 }
 
-func TestFloatFormat(t *testing.T) {
+const (
+	below1e23 = 99999999999999974834176
+	above1e23 = 100000000000000008388608
+)
+
+func TestFloat64Format(t *testing.T) {
 	for _, test := range []struct {
-		x      string
+		x      float64
 		format byte
 		prec   int
 		want   string
 	}{
-		{"0", 'b', 0, "0"},
-		{"-0", 'b', 0, "-0"},
-		{"1.0", 'b', 0, "4503599627370496p-52"},
-		{"-1.0", 'b', 0, "-4503599627370496p-52"},
-		{"4503599627370496", 'b', 0, "4503599627370496p+0"},
+		{0, 'f', 0, "0"},
+		{math.Copysign(0, -1), 'f', 0, "-0"},
+		{1, 'f', 0, "1"},
+		{-1, 'f', 0, "-1"},
 
-		{"0", 'p', 0, "0"},
-		{"-0", 'p', 0, "-0"},
-		{"1024.0", 'p', 0, "0x.8p11"},
-		{"-1024.0", 'p', 0, "-0x.8p11"},
-	} {
-		f64, err := strconv.ParseFloat(test.x, 64)
-		if err != nil {
-			t.Error(err)
-			continue
-		}
+		{1.459, 'e', 0, "1e+00"},
+		{2.459, 'e', 1, "2.5e+00"},
+		{3.459, 'e', 2, "3.46e+00"},
+		{4.459, 'e', 3, "4.459e+00"},
+		{5.459, 'e', 4, "5.4590e+00"},
+
+		{1.459, 'f', 0, "1"},
+		{2.459, 'f', 1, "2.5"},
+		{3.459, 'f', 2, "3.46"},
+		{4.459, 'f', 3, "4.459"},
+		{5.459, 'f', 4, "5.4590"},
+
+		{0, 'b', 0, "0"},
+		{math.Copysign(0, -1), 'b', 0, "-0"},
+		{1.0, 'b', 0, "4503599627370496p-52"},
+		{-1.0, 'b', 0, "-4503599627370496p-52"},
+		{4503599627370496, 'b', 0, "4503599627370496p+0"},
+
+		{0, 'p', 0, "0"},
+		{math.Copysign(0, -1), 'p', 0, "-0"},
+		{1024.0, 'p', 0, "0x.8p11"},
+		{-1024.0, 'p', 0, "-0x.8p11"},
+
+		// all test cases below from strconv/ftoa_test.go
+		{1, 'e', 5, "1.00000e+00"},
+		{1, 'f', 5, "1.00000"},
+		{1, 'g', 5, "1"},
+		// {1, 'g', -1, "1"},
+		// {20, 'g', -1, "20"},
+		// {1234567.8, 'g', -1, "1.2345678e+06"},
+		// {200000, 'g', -1, "200000"},
+		// {2000000, 'g', -1, "2e+06"},
+
+		// g conversion and zero suppression
+		{400, 'g', 2, "4e+02"},
+		{40, 'g', 2, "40"},
+		{4, 'g', 2, "4"},
+		{.4, 'g', 2, "0.4"},
+		{.04, 'g', 2, "0.04"},
+		{.004, 'g', 2, "0.004"},
+		{.0004, 'g', 2, "0.0004"},
+		{.00004, 'g', 2, "4e-05"},
+		{.000004, 'g', 2, "4e-06"},
+
+		{0, 'e', 5, "0.00000e+00"},
+		{0, 'f', 5, "0.00000"},
+		{0, 'g', 5, "0"},
+		// {0, 'g', -1, "0"},
+
+		{-1, 'e', 5, "-1.00000e+00"},
+		{-1, 'f', 5, "-1.00000"},
+		{-1, 'g', 5, "-1"},
+		// {-1, 'g', -1, "-1"},
 
-		f := new(Float).SetFloat64(f64)
+		{12, 'e', 5, "1.20000e+01"},
+		{12, 'f', 5, "12.00000"},
+		{12, 'g', 5, "12"},
+		// {12, 'g', -1, "12"},
+
+		{123456700, 'e', 5, "1.23457e+08"},
+		{123456700, 'f', 5, "123456700.00000"},
+		{123456700, 'g', 5, "1.2346e+08"},
+		// {123456700, 'g', -1, "1.234567e+08"},
+
+		{1.2345e6, 'e', 5, "1.23450e+06"},
+		{1.2345e6, 'f', 5, "1234500.00000"},
+		{1.2345e6, 'g', 5, "1.2345e+06"},
+
+		{1e23, 'e', 17, "9.99999999999999916e+22"},
+		{1e23, 'f', 17, "99999999999999991611392.00000000000000000"},
+		{1e23, 'g', 17, "9.9999999999999992e+22"},
+
+		// {1e23, 'e', -1, "1e+23"},
+		// {1e23, 'f', -1, "100000000000000000000000"},
+		// {1e23, 'g', -1, "1e+23"},
+
+		{below1e23, 'e', 17, "9.99999999999999748e+22"},
+		{below1e23, 'f', 17, "99999999999999974834176.00000000000000000"},
+		{below1e23, 'g', 17, "9.9999999999999975e+22"},
+
+		// {below1e23, 'e', -1, "9.999999999999997e+22"},
+		// {below1e23, 'f', -1, "99999999999999970000000"},
+		// {below1e23, 'g', -1, "9.999999999999997e+22"},
+
+		{above1e23, 'e', 17, "1.00000000000000008e+23"},
+		{above1e23, 'f', 17, "100000000000000008388608.00000000000000000"},
+		// {above1e23, 'g', 17, "1.0000000000000001e+23"},
+
+		// {above1e23, 'e', -1, "1.0000000000000001e+23"},
+		// {above1e23, 'f', -1, "100000000000000010000000"},
+		// {above1e23, 'g', -1, "1.0000000000000001e+23"},
+
+		// {fdiv(5e-304, 1e20), 'g', -1, "5e-324"},
+		// {fdiv(-5e-304, 1e20), 'g', -1, "-5e-324"},
+
+		// {32, 'g', -1, "32"},
+		// {32, 'g', 0, "3e+01"},
+
+		// {100, 'x', -1, "%x"},
+
+		// {math.NaN(), 'g', -1, "NaN"},
+		// {-math.NaN(), 'g', -1, "NaN"},
+		{math.Inf(0), 'g', -1, "+Inf"},
+		{math.Inf(-1), 'g', -1, "-Inf"},
+		{-math.Inf(0), 'g', -1, "-Inf"},
+
+		{-1, 'b', -1, "-4503599627370496p-52"},
+
+		// fixed bugs
+		{0.9, 'f', 1, "0.9"},
+		{0.09, 'f', 1, "0.1"},
+		{0.0999, 'f', 1, "0.1"},
+		{0.05, 'f', 1, "0.1"},
+		{0.05, 'f', 0, "0"},
+		{0.5, 'f', 1, "0.5"},
+		{0.5, 'f', 0, "0"},
+		{1.5, 'f', 0, "2"},
+
+		// http://www.exploringbinary.com/java-hangs-when-converting-2-2250738585072012e-308/
+		// {2.2250738585072012e-308, 'g', -1, "2.2250738585072014e-308"},
+		// http://www.exploringbinary.com/php-hangs-on-numeric-value-2-2250738585072011e-308/
+		// {2.2250738585072011e-308, 'g', -1, "2.225073858507201e-308"},
+
+		// Issue 2625.
+		{383260575764816448, 'f', 0, "383260575764816448"},
+		// {383260575764816448, 'g', -1, "3.8326057576481645e+17"},
+	} {
+		f := new(Float).SetFloat64(test.x)
 		got := f.Format(test.format, test.prec)
 		if got != test.want {
 			t.Errorf("%v: got %s; want %s", test, got, test.want)
 		}
 
-		if test.format == 'b' && f64 == 0 {
+		if test.format == 'b' && test.x == 0 {
 			continue // 'b' format in strconv.Float requires knowledge of bias for 0.0
 		}
 		if test.format == 'p' {
@@ -109,9 +231,97 @@ func TestFloatFormat(t *testing.T) {
 		}
 
 		// verify that Float format matches strconv format
-		want := strconv.FormatFloat(f64, test.format, test.prec, 64)
+		want := strconv.FormatFloat(test.x, test.format, test.prec, 64)
 		if got != want {
-			t.Errorf("%v: got %s; want %s", test, got, want)
+			t.Errorf("%v: got %s; want %s (strconv)", test, got, want)
+		}
+	}
+}
+
+func TestFloatFormat(t *testing.T) {
+	for _, test := range []struct {
+		x      string
+		format byte
+		prec   int
+		want   string
+	}{
+		{"0", 'f', 0, "0"},
+		{"-0", 'f', 0, "-0"},
+		{"1", 'f', 0, "1"},
+		{"-1", 'f', 0, "-1"},
+
+		{"1.459", 'e', 0, "1e+00"},
+		{"2.459", 'e', 1, "2.5e+00"},
+		{"3.459", 'e', 2, "3.46e+00"},
+		{"4.459", 'e', 3, "4.459e+00"},
+		{"5.459", 'e', 4, "5.4590e+00"},
+
+		{"1.459", 'E', 0, "1E+00"},
+		{"2.459", 'E', 1, "2.5E+00"},
+		{"3.459", 'E', 2, "3.46E+00"},
+		{"4.459", 'E', 3, "4.459E+00"},
+		{"5.459", 'E', 4, "5.4590E+00"},
+
+		{"1.459", 'f', 0, "1"},
+		{"2.459", 'f', 1, "2.5"},
+		{"3.459", 'f', 2, "3.46"},
+		{"4.459", 'f', 3, "4.459"},
+		{"5.459", 'f', 4, "5.4590"},
+
+		{"1.459", 'g', 0, "1"},
+		{"2.459", 'g', 1, "2"},
+		{"3.459", 'g', 2, "3.5"},
+		{"4.459", 'g', 3, "4.46"},
+		{"5.459", 'g', 4, "5.459"},
+
+		{"1459", 'g', 0, "1e+03"},
+		{"2459", 'g', 1, "2e+03"},
+		{"3459", 'g', 2, "3.5e+03"},
+		{"4459", 'g', 3, "4.46e+03"},
+		{"5459", 'g', 4, "5459"},
+
+		{"1459", 'G', 0, "1E+03"},
+		{"2459", 'G', 1, "2E+03"},
+		{"3459", 'G', 2, "3.5E+03"},
+		{"4459", 'G', 3, "4.46E+03"},
+		{"5459", 'G', 4, "5459"},
+
+		{"3", 'e', 40, "3.0000000000000000000000000000000000000000e+00"},
+		{"3", 'f', 40, "3.0000000000000000000000000000000000000000"},
+		{"3", 'g', 40, "3"},
+
+		{"3e40", 'e', 40, "3.0000000000000000000000000000000000000000e+40"},
+		{"3e40", 'f', 4, "30000000000000000000000000000000000000000.0000"},
+		{"3e40", 'g', 40, "3e+40"},
+
+		// TODO(gri) need tests for actual large Floats
+
+		// These depend on the selected mantissa length to match strconv.FormatFloat.
+		// Disabled for now.
+		// {"0", 'b', 0, "0"},
+		// {"-0", 'b', 0, "-0"},
+		// {"1.0", 'b', 0, "4503599627370496p-52"},
+		// {"-1.0", 'b', 0, "-4503599627370496p-52"},
+		// {"4503599627370496", 'b', 0, "4503599627370496p+0"},
+
+		{"0", 'p', 0, "0"},
+		{"-0", 'p', 0, "-0"},
+		{"1024.0", 'p', 0, "0x.8p11"},
+		{"-1024.0", 'p', 0, "-0x.8p11"},
+
+		// unsupported format
+		{"3.14", 'x', 0, "%x"},
+	} {
+		f, _, err := ParseFloat(test.x, 0, 1000, ToNearestEven)
+		// TODO(gri) should we return io.EOF at the end?
+		if err != nil && err != io.EOF {
+			t.Errorf("%v: %s", test, err)
+			continue
+		}
+
+		got := f.Format(test.format, test.prec)
+		if got != test.want {
+			t.Errorf("%v: got %s; want %s", test, got, test.want)
 		}
 	}
 }
diff --git a/src/math/big/ftoa.go b/src/math/big/ftoa.go
new file mode 100644
index 0000000000..148081589d
--- /dev/null
+++ b/src/math/big/ftoa.go
@@ -0,0 +1,184 @@
+// Copyright 2015 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// This file implements the 'e', 'f', 'g' floating-point formats.
+// It is closely following the corresponding implementation in
+// strconv/ftoa.go, but modified and simplified for big.Float.
+
+// Algorithm:
+//   1) convert Float to multiprecision decimal
+//   2) round to desired precision
+//   3) read digits out and format
+
+package big
+
+import "strconv"
+
+// TODO(gri) Consider moving sign into decimal - could make the signatures below cleaner.
+
+// bigFtoa formats a float for the %e, %E, %f, %g, and %G formats.
+func (f *Float) bigFtoa(buf []byte, fmt byte, prec int) []byte {
+	// TODO(gri) handle Inf.
+
+	// 1) convert Float to multiprecision decimal
+	var d decimal
+	d.init(f.mant, int(f.exp)-f.mant.bitLen())
+
+	// 2) round to desired precision
+	shortest := false
+	if prec < 0 {
+		shortest = true
+		panic("unimplemented")
+		// TODO(gri) complete this
+		// roundShortest(&d, f.mant, int(f.exp))
+		// Precision for shortest representation mode.
+		switch fmt {
+		case 'e', 'E':
+			prec = len(d.mant) - 1
+		case 'f':
+			prec = max(len(d.mant)-d.exp, 0)
+		case 'g', 'G':
+			prec = len(d.mant)
+		}
+	} else {
+		// round appropriately
+		switch fmt {
+		case 'e', 'E':
+			// one digit before and number of digits after decimal point
+			d.round(1 + prec)
+		case 'f':
+			// number of digits before and after decimal point
+			d.round(d.exp + prec)
+		case 'g', 'G':
+			if prec == 0 {
+				prec = 1
+			}
+			d.round(prec)
+		}
+	}
+
+	// 3) read digits out and format
+	switch fmt {
+	case 'e', 'E':
+		return fmtE(buf, fmt, prec, f.neg, d)
+	case 'f':
+		return fmtF(buf, prec, f.neg, d)
+	case 'g', 'G':
+		// trim trailing fractional zeros in %e format
+		eprec := prec
+		if eprec > len(d.mant) && len(d.mant) >= d.exp {
+			eprec = len(d.mant)
+		}
+		// %e is used if the exponent from the conversion
+		// is less than -4 or greater than or equal to the precision.
+		// If precision was the shortest possible, use eprec = 6 for
+		// this decision.
+		if shortest {
+			eprec = 6
+		}
+		exp := d.exp - 1
+		if exp < -4 || exp >= eprec {
+			if prec > len(d.mant) {
+				prec = len(d.mant)
+			}
+			return fmtE(buf, fmt+'e'-'g', prec-1, f.neg, d)
+		}
+		if prec > d.exp {
+			prec = len(d.mant)
+		}
+		return fmtF(buf, max(prec-d.exp, 0), f.neg, d)
+	}
+
+	// unknown format
+	return append(buf, '%', fmt)
+}
+
+// %e: -d.ddddde±dd
+func fmtE(buf []byte, fmt byte, prec int, neg bool, d decimal) []byte {
+	// sign
+	if neg {
+		buf = append(buf, '-')
+	}
+
+	// first digit
+	ch := byte('0')
+	if len(d.mant) > 0 {
+		ch = d.mant[0]
+	}
+	buf = append(buf, ch)
+
+	// .moredigits
+	if prec > 0 {
+		buf = append(buf, '.')
+		i := 1
+		m := min(len(d.mant), prec+1)
+		if i < m {
+			buf = append(buf, d.mant[i:m]...)
+			i = m
+		}
+		for ; i <= prec; i++ {
+			buf = append(buf, '0')
+		}
+	}
+
+	// e±
+	buf = append(buf, fmt)
+	var exp int64
+	if len(d.mant) > 0 {
+		exp = int64(d.exp) - 1 // -1 because first digit was printed before '.'
+	}
+	if exp < 0 {
+		ch = '-'
+		exp = -exp
+	} else {
+		ch = '+'
+	}
+	buf = append(buf, ch)
+
+	// dd...d
+	if exp < 10 {
+		buf = append(buf, '0') // at least 2 exponent digits
+	}
+	return strconv.AppendInt(buf, exp, 10)
+}
+
+// %f: -ddddddd.ddddd
+func fmtF(buf []byte, prec int, neg bool, d decimal) []byte {
+	// sign
+	if neg {
+		buf = append(buf, '-')
+	}
+
+	// integer, padded with zeros as needed
+	if d.exp > 0 {
+		m := min(len(d.mant), d.exp)
+		buf = append(buf, d.mant[:m]...)
+		for ; m < d.exp; m++ {
+			buf = append(buf, '0')
+		}
+	} else {
+		buf = append(buf, '0')
+	}
+
+	// fraction
+	if prec > 0 {
+		buf = append(buf, '.')
+		for i := 0; i < prec; i++ {
+			ch := byte('0')
+			if j := d.exp + i; 0 <= j && j < len(d.mant) {
+				ch = d.mant[j]
+			}
+			buf = append(buf, ch)
+		}
+	}
+
+	return buf
+}
+
+func min(x, y int) int {
+	if x < y {
+		return x
+	}
+	return y
+}
diff --git a/src/math/big/int_test.go b/src/math/big/int_test.go
index 1418dca335..cb8c76bed1 100644
--- a/src/math/big/int_test.go
+++ b/src/math/big/int_test.go
@@ -655,7 +655,7 @@ var primes = []string{
 	"10953742525620032441",
 	"17908251027575790097",
 
-	// http://code.google.com/p/go/issues/detail?id=638
+	// http://golang.org/issue/638
 	"18699199384836356663",
 
 	"98920366548084643601728869055592650835572950932266967461790948584315647051443",
diff --git a/src/nacltest.bash b/src/nacltest.bash
index 6220d39f13..534f1ef5af 100755
--- a/src/nacltest.bash
+++ b/src/nacltest.bash
@@ -62,7 +62,7 @@ fi
 # Run host build to get toolchain for running zip generator.
 unset GOOS GOARCH
 if [ ! -f make.bash ]; then
-	echo 'nacl.bash must be run from $GOROOT/src' 1>&2
+	echo 'nacltest.bash must be run from $GOROOT/src' 1>&2
 	exit 1
 fi
 GOOS=$GOHOSTOS GOARCH=$GOHOSTARCH ./make.bash
diff --git a/src/net/dial.go b/src/net/dial.go
index e6f0436cdd..0424ed250f 100644
--- a/src/net/dial.go
+++ b/src/net/dial.go
@@ -159,13 +159,15 @@ func (d *Dialer) Dial(network, address string) (Conn, error) {
 	if err != nil {
 		return nil, &OpError{Op: "dial", Net: network, Addr: nil, Err: err}
 	}
-	dialer := func(deadline time.Time) (Conn, error) {
-		return dialSingle(network, address, d.LocalAddr, ra.toAddr(), deadline)
-	}
+	var dialer func(deadline time.Time) (Conn, error)
 	if ras, ok := ra.(addrList); ok && d.DualStack && network == "tcp" {
 		dialer = func(deadline time.Time) (Conn, error) {
 			return dialMulti(network, address, d.LocalAddr, ras, deadline)
 		}
+	} else {
+		dialer = func(deadline time.Time) (Conn, error) {
+			return dialSingle(network, address, d.LocalAddr, ra.toAddr(), deadline)
+		}
 	}
 	c, err := dial(network, ra.toAddr(), dialer, d.deadline())
 	if d.KeepAlive > 0 && err == nil {
diff --git a/src/net/hosts_test.go b/src/net/hosts_test.go
index 2fe358e079..5bb663b4c7 100644
--- a/src/net/hosts_test.go
+++ b/src/net/hosts_test.go
@@ -53,7 +53,7 @@ func TestLookupStaticHost(t *testing.T) {
 	hostsPath = p
 }
 
-// https://code.google.com/p/go/issues/detail?id=6646
+// https://golang.org/issue/6646
 func TestSingleLineHostsFile(t *testing.T) {
 	p := hostsPath
 	hostsPath = "testdata/hosts_singleline"
diff --git a/src/net/http/proxy_test.go b/src/net/http/proxy_test.go
index b6aed3792b..823d1447ee 100644
--- a/src/net/http/proxy_test.go
+++ b/src/net/http/proxy_test.go
@@ -18,7 +18,7 @@ var UseProxyTests = []struct {
 	match bool
 }{
 	// Never proxy localhost:
-	{"localhost:80", false},
+	{"localhost", false},
 	{"127.0.0.1", false},
 	{"127.0.0.2", false},
 	{"[::1]", false},
diff --git a/src/net/http/request.go b/src/net/http/request.go
index 63d7d44aa0..f7a7f19b9b 100644
--- a/src/net/http/request.go
+++ b/src/net/http/request.go
@@ -673,7 +673,7 @@ func ReadRequest(b *bufio.Reader) (req *Request, err error) {
 // MaxBytesReader is similar to io.LimitReader but is intended for
 // limiting the size of incoming request bodies. In contrast to
 // io.LimitReader, MaxBytesReader's result is a ReadCloser, returns a
-// non-EOF error for a Read beyond the limit, and Closes the
+// non-EOF error for a Read beyond the limit, and closes the
 // underlying reader when its Close method is called.
 //
 // MaxBytesReader prevents clients from accidentally or maliciously
diff --git a/src/net/http/serve_test.go b/src/net/http/serve_test.go
index 85d5705137..c21b57b57e 100644
--- a/src/net/http/serve_test.go
+++ b/src/net/http/serve_test.go
@@ -416,7 +416,7 @@ func TestServeMuxHandlerRedirects(t *testing.T) {
 	}
 }
 
-// Tests for http://code.google.com/p/go/issues/detail?id=900
+// Tests for http://golang.org/issue/900
 func TestMuxRedirectLeadingSlashes(t *testing.T) {
 	paths := []string{"//foo.txt", "///foo.txt", "/../../foo.txt"}
 	for _, path := range paths {
@@ -2124,7 +2124,7 @@ func TestDoubleHijack(t *testing.T) {
 	<-conn.closec
 }
 
-// http://code.google.com/p/go/issues/detail?id=5955
+// http://golang.org/issue/5955
 // Note that this does not test the "request too large"
 // exit path from the http server. This is intentional;
 // not sending Connection: close is just a minor wire
diff --git a/src/net/interface_test.go b/src/net/interface_test.go
index fbf15de4d3..15c0cd7be4 100644
--- a/src/net/interface_test.go
+++ b/src/net/interface_test.go
@@ -38,8 +38,7 @@ func ipv6LinkLocalUnicastAddr(ifi *Interface) string {
 		return ""
 	}
 	for _, ifa := range ifat {
-		switch ifa := ifa.(type) {
-		case *IPNet:
+		if ifa, ok := ifa.(*IPNet); ok {
 			if ifa.IP.To4() == nil && ifa.IP.IsLinkLocalUnicast() {
 				return ifa.IP.String()
 			}
@@ -49,10 +48,6 @@ func ipv6LinkLocalUnicastAddr(ifi *Interface) string {
 }
 
 func TestInterfaces(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("temporarily disabled until golang.org/issue/5395 is fixed")
-	}
-
 	ift, err := Interfaces()
 	if err != nil {
 		t.Fatal(err)
@@ -110,10 +105,6 @@ func TestInterfaces(t *testing.T) {
 }
 
 func TestInterfaceAddrs(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("temporarily disabled until golang.org/issue/5395 is fixed")
-	}
-
 	ift, err := Interfaces()
 	if err != nil {
 		t.Fatal(err)
diff --git a/src/net/interface_windows.go b/src/net/interface_windows.go
index 0759dc255d..438dc874d6 100644
--- a/src/net/interface_windows.go
+++ b/src/net/interface_windows.go
@@ -5,123 +5,139 @@
 package net
 
 import (
+	"internal/syscall/windows"
 	"os"
 	"syscall"
 	"unsafe"
 )
 
-func bytePtrToString(p *uint8) string {
-	a := (*[10000]uint8)(unsafe.Pointer(p))
-	i := 0
-	for a[i] != 0 {
-		i++
-	}
-	return string(a[:i])
-}
+func getAdapters() (*windows.IpAdapterAddresses, error) {
+	block := uint32(unsafe.Sizeof(windows.IpAdapterAddresses{}))
 
-func getAdapterList() (*syscall.IpAdapterInfo, error) {
-	b := make([]byte, 1000)
-	l := uint32(len(b))
-	a := (*syscall.IpAdapterInfo)(unsafe.Pointer(&b[0]))
-	// TODO(mikio): GetAdaptersInfo returns IP_ADAPTER_INFO that
-	// contains IPv4 address list only. We should use another API
-	// for fetching IPv6 stuff from the kernel.
-	err := syscall.GetAdaptersInfo(a, &l)
-	if err == syscall.ERROR_BUFFER_OVERFLOW {
-		b = make([]byte, l)
-		a = (*syscall.IpAdapterInfo)(unsafe.Pointer(&b[0]))
-		err = syscall.GetAdaptersInfo(a, &l)
-	}
-	if err != nil {
-		return nil, os.NewSyscallError("GetAdaptersInfo", err)
+	// pre-allocate a 15KB working buffer pointed to by the AdapterAddresses
+	// parameter.
+	// https://msdn.microsoft.com/en-us/library/windows/desktop/aa365915(v=vs.85).aspx
+	size := uint32(15000)
+
+	var addrs []windows.IpAdapterAddresses
+	for {
+		addrs = make([]windows.IpAdapterAddresses, size/block+1)
+		err := windows.GetAdaptersAddresses(syscall.AF_UNSPEC, windows.GAA_FLAG_INCLUDE_PREFIX, 0, &addrs[0], &size)
+		if err == nil {
+			break
+		}
+		if err.(syscall.Errno) != syscall.ERROR_BUFFER_OVERFLOW {
+			return nil, os.NewSyscallError("GetAdaptersAddresses", err)
+		}
 	}
-	return a, nil
+	return &addrs[0], nil
 }
 
-func getInterfaceList() ([]syscall.InterfaceInfo, error) {
+func getInterfaceInfos() ([]syscall.InterfaceInfo, error) {
 	s, err := sysSocket(syscall.AF_INET, syscall.SOCK_DGRAM, syscall.IPPROTO_UDP)
 	if err != nil {
 		return nil, os.NewSyscallError("Socket", err)
 	}
 	defer syscall.Closesocket(s)
 
-	ii := [20]syscall.InterfaceInfo{}
+	iia := [20]syscall.InterfaceInfo{}
 	ret := uint32(0)
-	size := uint32(unsafe.Sizeof(ii))
-	err = syscall.WSAIoctl(s, syscall.SIO_GET_INTERFACE_LIST, nil, 0, (*byte)(unsafe.Pointer(&ii[0])), size, &ret, nil, 0)
+	size := uint32(unsafe.Sizeof(iia))
+	err = syscall.WSAIoctl(s, syscall.SIO_GET_INTERFACE_LIST, nil, 0, (*byte)(unsafe.Pointer(&iia[0])), size, &ret, nil, 0)
 	if err != nil {
 		return nil, os.NewSyscallError("WSAIoctl", err)
 	}
-	c := ret / uint32(unsafe.Sizeof(ii[0]))
-	return ii[:c-1], nil
+	iilen := ret / uint32(unsafe.Sizeof(iia[0]))
+	return iia[:iilen-1], nil
+}
+
+func bytesEqualIP(a []byte, b []int8) bool {
+	for i := 0; i < len(a); i++ {
+		if a[i] != byte(b[i]) {
+			return false
+		}
+	}
+	return true
+}
+
+func findInterfaceInfo(iis []syscall.InterfaceInfo, paddr *windows.IpAdapterAddresses) *syscall.InterfaceInfo {
+	for _, ii := range iis {
+		iaddr := (*syscall.RawSockaddr)(unsafe.Pointer(&ii.Address))
+		puni := paddr.FirstUnicastAddress
+		for ; puni != nil; puni = puni.Next {
+			if iaddr.Family == puni.Address.Sockaddr.Addr.Family {
+				switch iaddr.Family {
+				case syscall.AF_INET:
+					a := (*syscall.RawSockaddrInet4)(unsafe.Pointer(&ii.Address)).Addr
+					if bytesEqualIP(a[:], puni.Address.Sockaddr.Addr.Data[2:]) {
+						return &ii
+					}
+				case syscall.AF_INET6:
+					a := (*syscall.RawSockaddrInet6)(unsafe.Pointer(&ii.Address)).Addr
+					if bytesEqualIP(a[:], puni.Address.Sockaddr.Addr.Data[2:]) {
+						return &ii
+					}
+				default:
+					continue
+				}
+			}
+		}
+	}
+	return nil
 }
 
 // If the ifindex is zero, interfaceTable returns mappings of all
 // network interfaces.  Otherwise it returns a mapping of a specific
 // interface.
 func interfaceTable(ifindex int) ([]Interface, error) {
-	ai, err := getAdapterList()
+	paddr, err := getAdapters()
 	if err != nil {
 		return nil, err
 	}
 
-	ii, err := getInterfaceList()
+	iis, err := getInterfaceInfos()
 	if err != nil {
 		return nil, err
 	}
 
 	var ift []Interface
-	for ; ai != nil; ai = ai.Next {
-		index := ai.Index
+	for ; paddr != nil; paddr = paddr.Next {
+		index := paddr.IfIndex
+		if paddr.Ipv6IfIndex != 0 {
+			index = paddr.Ipv6IfIndex
+		}
 		if ifindex == 0 || ifindex == int(index) {
+			ii := findInterfaceInfo(iis, paddr)
+			if ii == nil {
+				continue
+			}
 			var flags Flags
-
-			row := syscall.MibIfRow{Index: index}
-			e := syscall.GetIfEntry(&row)
-			if e != nil {
-				return nil, os.NewSyscallError("GetIfEntry", e)
+			if paddr.Flags&windows.IfOperStatusUp != 0 {
+				flags |= FlagUp
 			}
-
-			for _, ii := range ii {
-				ip := (*syscall.RawSockaddrInet4)(unsafe.Pointer(&ii.Address)).Addr
-				ipv4 := IPv4(ip[0], ip[1], ip[2], ip[3])
-				ipl := &ai.IpAddressList
-				for ipl != nil {
-					ips := bytePtrToString(&ipl.IpAddress.String[0])
-					if ipv4.Equal(parseIPv4(ips)) {
-						break
-					}
-					ipl = ipl.Next
-				}
-				if ipl == nil {
-					continue
-				}
-				if ii.Flags&syscall.IFF_UP != 0 {
-					flags |= FlagUp
-				}
-				if ii.Flags&syscall.IFF_LOOPBACK != 0 {
-					flags |= FlagLoopback
-				}
-				if ii.Flags&syscall.IFF_BROADCAST != 0 {
-					flags |= FlagBroadcast
-				}
-				if ii.Flags&syscall.IFF_POINTTOPOINT != 0 {
-					flags |= FlagPointToPoint
-				}
-				if ii.Flags&syscall.IFF_MULTICAST != 0 {
-					flags |= FlagMulticast
-				}
+			if paddr.IfType&windows.IF_TYPE_SOFTWARE_LOOPBACK != 0 {
+				flags |= FlagLoopback
+			}
+			if ii.Flags&syscall.IFF_BROADCAST != 0 {
+				flags |= FlagBroadcast
+			}
+			if ii.Flags&syscall.IFF_POINTTOPOINT != 0 {
+				flags |= FlagPointToPoint
+			}
+			if ii.Flags&syscall.IFF_MULTICAST != 0 {
+				flags |= FlagMulticast
 			}
-
-			name := bytePtrToString(&ai.AdapterName[0])
-
 			ifi := Interface{
 				Index:        int(index),
-				MTU:          int(row.Mtu),
-				Name:         name,
-				HardwareAddr: HardwareAddr(row.PhysAddr[:row.PhysAddrLen]),
-				Flags:        flags}
+				MTU:          int(paddr.Mtu),
+				Name:         syscall.UTF16ToString((*(*[10000]uint16)(unsafe.Pointer(paddr.FriendlyName)))[:]),
+				HardwareAddr: HardwareAddr(paddr.PhysicalAddress[:]),
+				Flags:        flags,
+			}
 			ift = append(ift, ifi)
+			if ifindex == int(ifi.Index) {
+				break
+			}
 		}
 	}
 	return ift, nil
@@ -131,28 +147,86 @@ func interfaceTable(ifindex int) ([]Interface, error) {
 // network interfaces.  Otherwise it returns addresses for a specific
 // interface.
 func interfaceAddrTable(ifi *Interface) ([]Addr, error) {
-	ai, err := getAdapterList()
+	paddr, err := getAdapters()
 	if err != nil {
 		return nil, err
 	}
 
 	var ifat []Addr
-	for ; ai != nil; ai = ai.Next {
-		index := ai.Index
+	for ; paddr != nil; paddr = paddr.Next {
+		index := paddr.IfIndex
+		if paddr.Ipv6IfIndex != 0 {
+			index = paddr.Ipv6IfIndex
+		}
 		if ifi == nil || ifi.Index == int(index) {
-			ipl := &ai.IpAddressList
-			for ; ipl != nil; ipl = ipl.Next {
-				ifa := IPAddr{IP: parseIPv4(bytePtrToString(&ipl.IpAddress.String[0]))}
-				ifat = append(ifat, ifa.toAddr())
+			puni := paddr.FirstUnicastAddress
+			for ; puni != nil; puni = puni.Next {
+				if sa, err := puni.Address.Sockaddr.Sockaddr(); err == nil {
+					switch sav := sa.(type) {
+					case *syscall.SockaddrInet4:
+						ifa := &IPNet{IP: make(IP, IPv4len), Mask: CIDRMask(int(puni.Address.SockaddrLength), 8*IPv4len)}
+						copy(ifa.IP, sav.Addr[:])
+						ifat = append(ifat, ifa)
+					case *syscall.SockaddrInet6:
+						ifa := &IPNet{IP: make(IP, IPv6len), Mask: CIDRMask(int(puni.Address.SockaddrLength), 8*IPv6len)}
+						copy(ifa.IP, sav.Addr[:])
+						ifat = append(ifat, ifa)
+					}
+				}
+			}
+			pany := paddr.FirstAnycastAddress
+			for ; pany != nil; pany = pany.Next {
+				if sa, err := pany.Address.Sockaddr.Sockaddr(); err == nil {
+					switch sav := sa.(type) {
+					case *syscall.SockaddrInet4:
+						ifa := &IPNet{IP: make(IP, IPv4len), Mask: CIDRMask(int(pany.Address.SockaddrLength), 8*IPv4len)}
+						copy(ifa.IP, sav.Addr[:])
+						ifat = append(ifat, ifa)
+					case *syscall.SockaddrInet6:
+						ifa := &IPNet{IP: make(IP, IPv6len), Mask: CIDRMask(int(pany.Address.SockaddrLength), 8*IPv6len)}
+						copy(ifa.IP, sav.Addr[:])
+						ifat = append(ifat, ifa)
+					}
+				}
 			}
 		}
 	}
+
 	return ifat, nil
 }
 
 // interfaceMulticastAddrTable returns addresses for a specific
 // interface.
 func interfaceMulticastAddrTable(ifi *Interface) ([]Addr, error) {
-	// TODO(mikio): Implement this like other platforms.
-	return nil, nil
+	paddr, err := getAdapters()
+	if err != nil {
+		return nil, err
+	}
+
+	var ifat []Addr
+	for ; paddr != nil; paddr = paddr.Next {
+		index := paddr.IfIndex
+		if paddr.Ipv6IfIndex != 0 {
+			index = paddr.Ipv6IfIndex
+		}
+		if ifi == nil || ifi.Index == int(index) {
+			pmul := paddr.FirstMulticastAddress
+			for ; pmul != nil; pmul = pmul.Next {
+				if sa, err := pmul.Address.Sockaddr.Sockaddr(); err == nil {
+					switch sav := sa.(type) {
+					case *syscall.SockaddrInet4:
+						ifa := &IPAddr{IP: make(IP, IPv4len)}
+						copy(ifa.IP, sav.Addr[:])
+						ifat = append(ifat, ifa.toAddr())
+					case *syscall.SockaddrInet6:
+						ifa := &IPAddr{IP: make(IP, IPv6len)}
+						copy(ifa.IP, sav.Addr[:])
+						ifat = append(ifat, ifa.toAddr())
+					}
+				}
+			}
+		}
+	}
+
+	return ifat, nil
 }
diff --git a/src/net/net.go b/src/net/net.go
index c850d2b1fd..339c972906 100644
--- a/src/net/net.go
+++ b/src/net/net.go
@@ -135,6 +135,8 @@ func (c *conn) Close() error {
 }
 
 // LocalAddr returns the local network address.
+// The Addr returned is shared by all invocations of LocalAddr, so
+// do not modify it.
 func (c *conn) LocalAddr() Addr {
 	if !c.ok() {
 		return nil
@@ -143,6 +145,8 @@ func (c *conn) LocalAddr() Addr {
 }
 
 // RemoteAddr returns the remote network address.
+// The Addr returned is shared by all invocations of RemoteAddr, so
+// do not modify it.
 func (c *conn) RemoteAddr() Addr {
 	if !c.ok() {
 		return nil
diff --git a/src/net/smtp/example_test.go b/src/net/smtp/example_test.go
index d551e365a9..16419f4276 100644
--- a/src/net/smtp/example_test.go
+++ b/src/net/smtp/example_test.go
@@ -46,14 +46,36 @@ func Example() {
 	}
 }
 
+// variables to make ExamplePlainAuth compile, without adding
+// unnecessary noise there.
+var (
+	from       = "gopher@example.net"
+	msg        = []byte("dummy message")
+	recipients = []string{"foo@example.com"}
+)
+
 func ExamplePlainAuth() {
+	// hostname is used by PlainAuth to validate the TLS certificate.
+	hostname := "mail.example.com"
+	auth := smtp.PlainAuth("", "user@example.com", "password", hostname)
+
+	err := smtp.SendMail(hostname+":25", auth, from, recipients, msg)
+	if err != nil {
+		log.Fatal(err)
+	}
+}
+
+func ExampleSendMail() {
 	// Set up authentication information.
 	auth := smtp.PlainAuth("", "user@example.com", "password", "mail.example.com")
 
 	// Connect to the server, authenticate, set the sender and recipient,
 	// and send the email all in one step.
 	to := []string{"recipient@example.net"}
-	msg := []byte("This is the email body.")
+	msg := []byte("To: recipient@example.net\r\n" +
+		"Subject: discount Gophers!\r\n" +
+		"\r\n" +
+		"This is the email body.\r\n")
 	err := smtp.SendMail("mail.example.com:25", auth, "sender@example.org", to, msg)
 	if err != nil {
 		log.Fatal(err)
diff --git a/src/net/smtp/smtp.go b/src/net/smtp/smtp.go
index c9b3c07aa8..81f3c0bd62 100644
--- a/src/net/smtp/smtp.go
+++ b/src/net/smtp/smtp.go
@@ -264,9 +264,9 @@ func (d *dataCloser) Close() error {
 }
 
 // Data issues a DATA command to the server and returns a writer that
-// can be used to write the data. The caller should close the writer
-// before calling any more methods on c.
-// A call to Data must be preceded by one or more calls to Rcpt.
+// can be used to write the mail headers and body. The caller should
+// close the writer before calling any more methods on c.  A call to
+// Data must be preceded by one or more calls to Rcpt.
 func (c *Client) Data() (io.WriteCloser, error) {
 	_, _, err := c.cmd(354, "DATA")
 	if err != nil {
@@ -281,6 +281,21 @@ var testHookStartTLS func(*tls.Config) // nil, except for tests
 // possible, authenticates with the optional mechanism a if possible,
 // and then sends an email from address from, to addresses to, with
 // message msg.
+//
+// The addresses in the to parameter are the SMTP RCPT addresses.
+//
+// The msg parameter should be an RFC 822-style email with headers
+// first, a blank line, and then the message body. The lines of msg
+// should be CRLF terminated.  The msg headers should usually include
+// fields such as "From", "To", "Subject", and "Cc".  Sending "Bcc"
+// messages is accomplished by including an email address in the to
+// parameter but not including it in the msg headers.
+//
+// The SendMail function and the the net/smtp package are low-level
+// mechanisms and provide no support for DKIM signing, MIME
+// attachments (see the mime/multipart package), or other mail
+// functionality. Higher-level packages exist outside of the standard
+// library.
 func SendMail(addr string, a Auth, from string, to []string, msg []byte) error {
 	c, err := Dial(addr)
 	if err != nil {
diff --git a/src/net/tcpsock_plan9.go b/src/net/tcpsock_plan9.go
index 52019d7b4e..ae2194277d 100644
--- a/src/net/tcpsock_plan9.go
+++ b/src/net/tcpsock_plan9.go
@@ -157,6 +157,8 @@ func (l *TCPListener) Close() error {
 }
 
 // Addr returns the listener's network address, a *TCPAddr.
+// The Addr returned is shared by all invocations of Addr, so
+// do not modify it.
 func (l *TCPListener) Addr() Addr { return l.fd.laddr }
 
 // SetDeadline sets the deadline associated with the listener.
diff --git a/src/net/tcpsock_posix.go b/src/net/tcpsock_posix.go
index dd78aefa77..aaff0acaa6 100644
--- a/src/net/tcpsock_posix.go
+++ b/src/net/tcpsock_posix.go
@@ -258,6 +258,8 @@ func (l *TCPListener) Close() error {
 }
 
 // Addr returns the listener's network address, a *TCPAddr.
+// The Addr returned is shared by all invocations of Addr, so
+// do not modify it.
 func (l *TCPListener) Addr() Addr { return l.fd.laddr }
 
 // SetDeadline sets the deadline associated with the listener.
diff --git a/src/net/udp_test.go b/src/net/udp_test.go
index 125bbca6c4..2be2c319a7 100644
--- a/src/net/udp_test.go
+++ b/src/net/udp_test.go
@@ -81,26 +81,26 @@ func TestWriteToUDP(t *testing.T) {
 		t.Skipf("skipping test on %q", runtime.GOOS)
 	}
 
-	l, err := ListenPacket("udp", "127.0.0.1:0")
+	c, err := ListenPacket("udp", "127.0.0.1:0")
 	if err != nil {
-		t.Fatalf("Listen failed: %v", err)
+		t.Fatal(err)
 	}
-	defer l.Close()
+	defer c.Close()
 
-	testWriteToConn(t, l.LocalAddr().String())
-	testWriteToPacketConn(t, l.LocalAddr().String())
+	testWriteToConn(t, c.LocalAddr().String())
+	testWriteToPacketConn(t, c.LocalAddr().String())
 }
 
 func testWriteToConn(t *testing.T, raddr string) {
 	c, err := Dial("udp", raddr)
 	if err != nil {
-		t.Fatalf("Dial failed: %v", err)
+		t.Fatal(err)
 	}
 	defer c.Close()
 
 	ra, err := ResolveUDPAddr("udp", raddr)
 	if err != nil {
-		t.Fatalf("ResolveUDPAddr failed: %v", err)
+		t.Fatal(err)
 	}
 
 	_, err = c.(*UDPConn).WriteToUDP([]byte("Connection-oriented mode socket"), ra)
@@ -121,36 +121,70 @@ func testWriteToConn(t *testing.T, raddr string) {
 
 	_, err = c.Write([]byte("Connection-oriented mode socket"))
 	if err != nil {
-		t.Fatalf("Write failed: %v", err)
+		t.Fatal(err)
+	}
+
+	_, _, err = c.(*UDPConn).WriteMsgUDP([]byte("Connection-oriented mode socket"), nil, ra)
+	if err == nil {
+		t.Fatal("WriteMsgUDP should fail")
+	}
+	if err != nil && err.(*OpError).Err != ErrWriteToConnected {
+		t.Fatalf("WriteMsgUDP should fail as ErrWriteToConnected: %v", err)
+	}
+	_, _, err = c.(*UDPConn).WriteMsgUDP([]byte("Connection-oriented mode socket"), nil, nil)
+	switch runtime.GOOS {
+	case "nacl", "windows": // see golang.org/issue/9252
+		t.Skipf("not implemented yet on %s", runtime.GOOS)
+	default:
+		if err != nil {
+			t.Fatal(err)
+		}
 	}
 }
 
 func testWriteToPacketConn(t *testing.T, raddr string) {
 	c, err := ListenPacket("udp", "127.0.0.1:0")
 	if err != nil {
-		t.Fatalf("ListenPacket failed: %v", err)
+		t.Fatal(err)
 	}
 	defer c.Close()
 
 	ra, err := ResolveUDPAddr("udp", raddr)
 	if err != nil {
-		t.Fatalf("ResolveUDPAddr failed: %v", err)
+		t.Fatal(err)
 	}
 
 	_, err = c.(*UDPConn).WriteToUDP([]byte("Connection-less mode socket"), ra)
 	if err != nil {
-		t.Fatalf("WriteToUDP failed: %v", err)
+		t.Fatal(err)
 	}
 
 	_, err = c.WriteTo([]byte("Connection-less mode socket"), ra)
 	if err != nil {
-		t.Fatalf("WriteTo failed: %v", err)
+		t.Fatal(err)
 	}
 
 	_, err = c.(*UDPConn).Write([]byte("Connection-less mode socket"))
 	if err == nil {
 		t.Fatal("Write should fail")
 	}
+
+	_, _, err = c.(*UDPConn).WriteMsgUDP([]byte("Connection-less mode socket"), nil, nil)
+	if err == nil {
+		t.Fatal("WriteMsgUDP should fail")
+	}
+	if err != nil && err.(*OpError).Err != errMissingAddress {
+		t.Fatalf("WriteMsgUDP should fail as errMissingAddress: %v", err)
+	}
+	_, _, err = c.(*UDPConn).WriteMsgUDP([]byte("Connection-less mode socket"), nil, ra)
+	switch runtime.GOOS {
+	case "nacl", "windows": // see golang.org/issue/9252
+		t.Skipf("not implemented yet on %s", runtime.GOOS)
+	default:
+		if err != nil {
+			t.Fatal(err)
+		}
+	}
 }
 
 var udpConnLocalNameTests = []struct {
diff --git a/src/net/udpsock_plan9.go b/src/net/udpsock_plan9.go
index 510ac5e4aa..269272632a 100644
--- a/src/net/udpsock_plan9.go
+++ b/src/net/udpsock_plan9.go
@@ -101,9 +101,11 @@ func (c *UDPConn) WriteTo(b []byte, addr Addr) (int, error) {
 	return c.WriteToUDP(b, a)
 }
 
-// WriteMsgUDP writes a packet to addr via c, copying the payload from
-// b and the associated out-of-band data from oob.  It returns the
-// number of payload and out-of-band bytes written.
+// WriteMsgUDP writes a packet to addr via c if c isn't connected, or
+// to c's remote destination address if c is connected (in which case
+// addr must be nil).  The payload is copied from b and the associated
+// out-of-band data is copied from oob.  It returns the number of
+// payload and out-of-band bytes written.
 func (c *UDPConn) WriteMsgUDP(b, oob []byte, addr *UDPAddr) (n, oobn int, err error) {
 	return 0, 0, syscall.EPLAN9
 }
diff --git a/src/net/udpsock_posix.go b/src/net/udpsock_posix.go
index a0533366a4..0770b7c5ce 100644
--- a/src/net/udpsock_posix.go
+++ b/src/net/udpsock_posix.go
@@ -139,17 +139,19 @@ func (c *UDPConn) WriteTo(b []byte, addr Addr) (int, error) {
 	return c.WriteToUDP(b, a)
 }
 
-// WriteMsgUDP writes a packet to addr via c, copying the payload from
-// b and the associated out-of-band data from oob.  It returns the
-// number of payload and out-of-band bytes written.
+// WriteMsgUDP writes a packet to addr via c if c isn't connected, or
+// to c's remote destination address if c is connected (in which case
+// addr must be nil).  The payload is copied from b and the associated
+// out-of-band data is copied from oob.  It returns the number of
+// payload and out-of-band bytes written.
 func (c *UDPConn) WriteMsgUDP(b, oob []byte, addr *UDPAddr) (n, oobn int, err error) {
 	if !c.ok() {
 		return 0, 0, syscall.EINVAL
 	}
-	if c.fd.isConnected {
+	if c.fd.isConnected && addr != nil {
 		return 0, 0, &OpError{"write", c.fd.net, addr, ErrWriteToConnected}
 	}
-	if addr == nil {
+	if !c.fd.isConnected && addr == nil {
 		return 0, 0, &OpError{Op: "write", Net: c.fd.net, Addr: nil, Err: errMissingAddress}
 	}
 	sa, err := addr.sockaddr(c.fd.family)
diff --git a/src/net/unixsock_plan9.go b/src/net/unixsock_plan9.go
index c60c1d83bb..64a511d648 100644
--- a/src/net/unixsock_plan9.go
+++ b/src/net/unixsock_plan9.go
@@ -115,6 +115,8 @@ func (l *UnixListener) Close() error {
 }
 
 // Addr returns the listener's network address.
+// The Addr returned is shared by all invocations of Addr, so
+// do not modify it.
 func (l *UnixListener) Addr() Addr { return nil }
 
 // SetDeadline sets the deadline associated with the listener.
diff --git a/src/net/unixsock_posix.go b/src/net/unixsock_posix.go
index 3c2e78bdca..d7127d9c69 100644
--- a/src/net/unixsock_posix.go
+++ b/src/net/unixsock_posix.go
@@ -321,6 +321,8 @@ func (l *UnixListener) Close() error {
 }
 
 // Addr returns the listener's network address.
+// The Addr returned is shared by all invocations of Addr, so
+// do not modify it.
 func (l *UnixListener) Addr() Addr { return l.fd.laddr }
 
 // SetDeadline sets the deadline associated with the listener.
diff --git a/src/os/os_test.go b/src/os/os_test.go
index a30a2b0313..5285b76024 100644
--- a/src/os/os_test.go
+++ b/src/os/os_test.go
@@ -21,7 +21,6 @@ import (
 	"sync"
 	"syscall"
 	"testing"
-	"text/template"
 	"time"
 )
 
@@ -491,6 +490,30 @@ func TestReaddirStatFailures(t *testing.T) {
 	}
 }
 
+// Readdir on a regular file should fail.
+func TestReaddirOfFile(t *testing.T) {
+	f, err := ioutil.TempFile("", "_Go_ReaddirOfFile")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer Remove(f.Name())
+	f.Write([]byte("foo"))
+	f.Close()
+	reg, err := Open(f.Name())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer reg.Close()
+
+	names, err := reg.Readdirnames(-1)
+	if err == nil {
+		t.Error("Readdirnames succeeded; want non-nil error")
+	}
+	if len(names) > 0 {
+		t.Errorf("unexpected dir names in regular file: %q", names)
+	}
+}
+
 func TestHardLink(t *testing.T) {
 	// Hardlinks are not supported under windows or Plan 9.
 	if runtime.GOOS == "plan9" {
@@ -922,7 +945,7 @@ func TestSeek(t *testing.T) {
 		if off != tt.out || err != nil {
 			if e, ok := err.(*PathError); ok && e.Err == syscall.EINVAL && tt.out > 1<<32 {
 				// Reiserfs rejects the big seeks.
-				// http://code.google.com/p/go/issues/detail?id=91
+				// http://golang.org/issue/91
 				break
 			}
 			t.Errorf("#%d: Seek(%v, %v) = %v, %v want %v, nil", i, tt.in, tt.whence, off, err, tt.out)
@@ -1304,39 +1327,9 @@ func testKillProcess(t *testing.T, processKiller func(p *Process)) {
 		t.Skipf("skipping on %s", runtime.GOOS)
 	}
 
-	dir, err := ioutil.TempDir("", "go-build")
-	if err != nil {
-		t.Fatalf("Failed to create temp directory: %v", err)
-	}
-	defer RemoveAll(dir)
-
-	src := filepath.Join(dir, "main.go")
-	f, err := Create(src)
-	if err != nil {
-		t.Fatalf("Failed to create %v: %v", src, err)
-	}
-	st := template.Must(template.New("source").Parse(`
-package main
-import "time"
-func main() {
-	time.Sleep(time.Second)
-}
-`))
-	err = st.Execute(f, nil)
-	if err != nil {
-		f.Close()
-		t.Fatalf("Failed to execute template: %v", err)
-	}
-	f.Close()
-
-	exe := filepath.Join(dir, "main.exe")
-	output, err := osexec.Command("go", "build", "-o", exe, src).CombinedOutput()
-	if err != nil {
-		t.Fatalf("Failed to build exe %v: %v %v", exe, err, string(output))
-	}
-
-	cmd := osexec.Command(exe)
-	err = cmd.Start()
+	// Re-exec the test binary itself to emulate "sleep 1".
+	cmd := osexec.Command(Args[0], "-test.run", "TestSleep")
+	err := cmd.Start()
 	if err != nil {
 		t.Fatalf("Failed to start test process: %v", err)
 	}
@@ -1350,6 +1343,15 @@ func main() {
 	}
 }
 
+// TestSleep emulates "sleep 1". It is a helper for testKillProcess, so we
+// don't have to rely on an external "sleep" command being available.
+func TestSleep(t *testing.T) {
+	if testing.Short() {
+		t.Skip("Skipping in short mode")
+	}
+	time.Sleep(time.Second)
+}
+
 func TestKillStartProcess(t *testing.T) {
 	testKillProcess(t, func(p *Process) {
 		err := p.Kill()
diff --git a/src/os/os_windows_test.go b/src/os/os_windows_test.go
index fd96713eac..3ea0fc7f4f 100644
--- a/src/os/os_windows_test.go
+++ b/src/os/os_windows_test.go
@@ -3,11 +3,15 @@ package os_test
 import (
 	"io/ioutil"
 	"os"
+	osexec "os/exec"
 	"path/filepath"
+	"strings"
 	"syscall"
 	"testing"
 )
 
+var supportJunctionLinks = true
+
 func init() {
 	tmpdir, err := ioutil.TempDir("", "symtest")
 	if err != nil {
@@ -16,14 +20,18 @@ func init() {
 	defer os.RemoveAll(tmpdir)
 
 	err = os.Symlink("target", filepath.Join(tmpdir, "symlink"))
-	if err == nil {
-		return
+	if err != nil {
+		err = err.(*os.LinkError).Err
+		switch err {
+		case syscall.EWINDOWS, syscall.ERROR_PRIVILEGE_NOT_HELD:
+			supportsSymlinks = false
+		}
 	}
+	defer os.Remove("target")
 
-	err = err.(*os.LinkError).Err
-	switch err {
-	case syscall.EWINDOWS, syscall.ERROR_PRIVILEGE_NOT_HELD:
-		supportsSymlinks = false
+	b, _ := osexec.Command("cmd", "/c", "mklink", "/?").Output()
+	if !strings.Contains(string(b), " /J ") {
+		supportJunctionLinks = false
 	}
 }
 
@@ -79,3 +87,33 @@ func TestSameWindowsFile(t *testing.T) {
 		t.Errorf("files should be same")
 	}
 }
+
+func TestStatJunctionLink(t *testing.T) {
+	if !supportJunctionLinks {
+		t.Skip("skipping because junction links are not supported")
+	}
+
+	dir, err := ioutil.TempDir("", "go-build")
+	if err != nil {
+		t.Fatalf("failed to create temp directory: %v", err)
+	}
+	defer os.RemoveAll(dir)
+
+	link := filepath.Join(filepath.Dir(dir), filepath.Base(dir)+"-link")
+
+	output, err := osexec.Command("cmd", "/c", "mklink", "/J", link, dir).CombinedOutput()
+	if err != nil {
+		t.Fatalf("failed to run mklink %v %v: %v %q", link, dir, err, output)
+	}
+	defer os.Remove(link)
+
+	fi, err := os.Stat(link)
+	if err != nil {
+		t.Fatalf("failed to stat link %v: %v", link, err)
+	}
+	expected := filepath.Base(dir)
+	got := fi.Name()
+	if !fi.IsDir() || expected != got {
+		t.Fatalf("link should point to %v but points to %v instead", expected, got)
+	}
+}
diff --git a/src/os/signal/signal.go b/src/os/signal/signal.go
index 3004275495..81906d6f40 100644
--- a/src/os/signal/signal.go
+++ b/src/os/signal/signal.go
@@ -5,8 +5,6 @@
 // Package signal implements access to incoming signals.
 package signal
 
-// BUG(rsc): This package is not yet implemented on Plan 9.
-
 import (
 	"os"
 	"sync"
diff --git a/src/os/signal/signal_plan9.go b/src/os/signal/signal_plan9.go
new file mode 100644
index 0000000000..45355da48a
--- /dev/null
+++ b/src/os/signal/signal_plan9.go
@@ -0,0 +1,55 @@
+// Copyright 2012 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package signal
+
+import (
+	"os"
+	"syscall"
+)
+
+var sigtab = make(map[os.Signal]int)
+
+// In sig.s; jumps to runtime.
+func signal_disable(uint32)
+func signal_enable(uint32)
+func signal_recv() string
+
+func init() {
+	signal_enable(0) // first call - initialize
+	go loop()
+}
+
+func loop() {
+	for {
+		process(syscall.Note(signal_recv()))
+	}
+}
+
+const numSig = 256
+
+func signum(sig os.Signal) int {
+	switch sig := sig.(type) {
+	case syscall.Note:
+		n, ok := sigtab[sig]
+		if !ok {
+			n = len(sigtab) + 1
+			if n > numSig {
+				return -1
+			}
+			sigtab[sig] = n
+		}
+		return n
+	default:
+		return -1
+	}
+}
+
+func enableSignal(sig int) {
+	signal_enable(uint32(sig))
+}
+
+func disableSignal(sig int) {
+	signal_disable(uint32(sig))
+}
diff --git a/src/os/signal/signal_plan9_test.go b/src/os/signal/signal_plan9_test.go
new file mode 100644
index 0000000000..10bfdc3ff1
--- /dev/null
+++ b/src/os/signal/signal_plan9_test.go
@@ -0,0 +1,181 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package signal
+
+import (
+	"os"
+	"runtime"
+	"syscall"
+	"testing"
+	"time"
+)
+
+func waitSig(t *testing.T, c <-chan os.Signal, sig os.Signal) {
+	select {
+	case s := <-c:
+		if s != sig {
+			t.Fatalf("signal was %v, want %v", s, sig)
+		}
+	case <-time.After(1 * time.Second):
+		t.Fatalf("timeout waiting for %v", sig)
+	}
+}
+
+// Test that basic signal handling works.
+func TestSignal(t *testing.T) {
+	// Ask for hangup
+	c := make(chan os.Signal, 1)
+	Notify(c, syscall.Note("hangup"))
+	defer Stop(c)
+
+	// Send this process a hangup
+	t.Logf("hangup...")
+	postNote(syscall.Getpid(), "hangup")
+	waitSig(t, c, syscall.Note("hangup"))
+
+	// Ask for everything we can get.
+	c1 := make(chan os.Signal, 1)
+	Notify(c1)
+
+	// Send this process an alarm
+	t.Logf("alarm...")
+	postNote(syscall.Getpid(), "alarm")
+	waitSig(t, c1, syscall.Note("alarm"))
+
+	// Send two more hangups, to make sure that
+	// they get delivered on c1 and that not reading
+	// from c does not block everything.
+	t.Logf("hangup...")
+	postNote(syscall.Getpid(), "hangup")
+	waitSig(t, c1, syscall.Note("hangup"))
+	t.Logf("hangup...")
+	postNote(syscall.Getpid(), "hangup")
+	waitSig(t, c1, syscall.Note("hangup"))
+
+	// The first SIGHUP should be waiting for us on c.
+	waitSig(t, c, syscall.Note("hangup"))
+}
+
+func TestStress(t *testing.T) {
+	dur := 3 * time.Second
+	if testing.Short() {
+		dur = 100 * time.Millisecond
+	}
+	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(4))
+	done := make(chan bool)
+	finished := make(chan bool)
+	go func() {
+		sig := make(chan os.Signal, 1)
+		Notify(sig, syscall.Note("alarm"))
+		defer Stop(sig)
+	Loop:
+		for {
+			select {
+			case <-sig:
+			case <-done:
+				break Loop
+			}
+		}
+		finished <- true
+	}()
+	go func() {
+	Loop:
+		for {
+			select {
+			case <-done:
+				break Loop
+			default:
+				postNote(syscall.Getpid(), "alarm")
+				runtime.Gosched()
+			}
+		}
+		finished <- true
+	}()
+	time.Sleep(dur)
+	close(done)
+	<-finished
+	<-finished
+	// When run with 'go test -cpu=1,2,4' alarm from this test can slip
+	// into subsequent TestSignal() causing failure.
+	// Sleep for a while to reduce the possibility of the failure.
+	time.Sleep(10 * time.Millisecond)
+}
+
+// Test that Stop cancels the channel's registrations.
+func TestStop(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping in short mode")
+	}
+	sigs := []string{
+		"alarm",
+		"hangup",
+	}
+
+	for _, sig := range sigs {
+		// Send the signal.
+		// If it's alarm, we should not see it.
+		// If it's hangup, maybe we'll die. Let the flag tell us what to do.
+		if sig != "hangup" {
+			postNote(syscall.Getpid(), sig)
+		}
+		time.Sleep(100 * time.Millisecond)
+
+		// Ask for signal
+		c := make(chan os.Signal, 1)
+		Notify(c, syscall.Note(sig))
+		defer Stop(c)
+
+		// Send this process that signal
+		postNote(syscall.Getpid(), sig)
+		waitSig(t, c, syscall.Note(sig))
+
+		Stop(c)
+		select {
+		case s := <-c:
+			t.Fatalf("unexpected signal %v", s)
+		case <-time.After(100 * time.Millisecond):
+			// nothing to read - good
+		}
+
+		// Send the signal.
+		// If it's alarm, we should not see it.
+		// If it's hangup, maybe we'll die. Let the flag tell us what to do.
+		if sig != "hangup" {
+			postNote(syscall.Getpid(), sig)
+		}
+
+		select {
+		case s := <-c:
+			t.Fatalf("unexpected signal %v", s)
+		case <-time.After(100 * time.Millisecond):
+			// nothing to read - good
+		}
+	}
+}
+
+func itoa(val int) string {
+	if val < 0 {
+		return "-" + itoa(-val)
+	}
+	var buf [32]byte // big enough for int64
+	i := len(buf) - 1
+	for val >= 10 {
+		buf[i] = byte(val%10 + '0')
+		i--
+		val /= 10
+	}
+	buf[i] = byte(val + '0')
+	return string(buf[i:])
+}
+
+func postNote(pid int, note string) error {
+	f, err := os.OpenFile("/proc/"+itoa(pid)+"/note", os.O_WRONLY, 0)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+	_, err = f.Write([]byte(note))
+	return err
+}
diff --git a/src/os/signal/signal_stub.go b/src/os/signal/signal_stub.go
deleted file mode 100644
index d0a6935ff2..0000000000
--- a/src/os/signal/signal_stub.go
+++ /dev/null
@@ -1,17 +0,0 @@
-// Copyright 2012 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-// +build plan9
-
-package signal
-
-import "os"
-
-const numSig = 0
-
-func signum(sig os.Signal) int { return -1 }
-
-func disableSignal(int) {}
-
-func enableSignal(int) {}
diff --git a/src/path/filepath/path_test.go b/src/path/filepath/path_test.go
index c4f74b97ff..293d316124 100644
--- a/src/path/filepath/path_test.go
+++ b/src/path/filepath/path_test.go
@@ -1008,7 +1008,7 @@ func TestDriveLetterInEvalSymlinks(t *testing.T) {
 	}
 }
 
-func TestBug3486(t *testing.T) { // http://code.google.com/p/go/issues/detail?id=3486
+func TestBug3486(t *testing.T) { // http://golang.org/issue/3486
 	root, err := filepath.EvalSymlinks(runtime.GOROOT() + "/test")
 	if err != nil {
 		t.Fatal(err)
diff --git a/src/race.bat b/src/race.bat
index 8858c57b06..027c475640 100644
--- a/src/race.bat
+++ b/src/race.bat
@@ -18,7 +18,7 @@ goto end
 set GOROOT=%CD%\..
 call make.bat --dist-tool >NUL
 if errorlevel 1 goto fail
-.\cmd\dist\dist env -wp >env.bat
+.\cmd\dist\dist env -w -p >env.bat
 if errorlevel 1 goto fail
 call env.bat
 del env.bat
diff --git a/src/reflect/all_test.go b/src/reflect/all_test.go
index 7d40f9a8b6..2be32f31b4 100644
--- a/src/reflect/all_test.go
+++ b/src/reflect/all_test.go
@@ -2735,6 +2735,8 @@ var tagGetTests = []struct {
 	{`protobuf:"PB(1,2)"`, `rotobuf`, ``},
 	{`protobuf:"PB(1,2)" json:"name"`, `json`, `name`},
 	{`protobuf:"PB(1,2)" json:"name"`, `protobuf`, `PB(1,2)`},
+	{`k0:"values contain spaces" k1:"and\ttabs"`, "k0", "values contain spaces"},
+	{`k0:"values contain spaces" k1:"and\ttabs"`, "k1", "and\ttabs"},
 }
 
 func TestTagGet(t *testing.T) {
diff --git a/src/reflect/type.go b/src/reflect/type.go
index ae7d165a68..0a8c40808a 100644
--- a/src/reflect/type.go
+++ b/src/reflect/type.go
@@ -762,8 +762,11 @@ type StructTag string
 // If the tag does not have the conventional format, the value
 // returned by Get is unspecified.
 func (tag StructTag) Get(key string) string {
+	// When modifying this code, also update the validateStructTag code
+	// in golang.org/x/tools/cmd/vet/structtag.go.
+
 	for tag != "" {
-		// skip leading space
+		// Skip leading space.
 		i := 0
 		for i < len(tag) && tag[i] == ' ' {
 			i++
@@ -773,19 +776,21 @@ func (tag StructTag) Get(key string) string {
 			break
 		}
 
-		// scan to colon.
-		// a space or a quote is a syntax error
+		// Scan to colon. A space, a quote or a control character is a syntax error.
+		// Strictly speaking, control chars include the range [0x7f, 0x9f], not just
+		// [0x00, 0x1f], but in practice, we ignore the multi-byte control characters
+		// as it is simpler to inspect the tag's bytes than the tag's runes.
 		i = 0
-		for i < len(tag) && tag[i] != ' ' && tag[i] != ':' && tag[i] != '"' {
+		for i < len(tag) && tag[i] > ' ' && tag[i] != ':' && tag[i] != '"' && tag[i] != 0x7f {
 			i++
 		}
-		if i+1 >= len(tag) || tag[i] != ':' || tag[i+1] != '"' {
+		if i == 0 || i+1 >= len(tag) || tag[i] != ':' || tag[i+1] != '"' {
 			break
 		}
 		name := string(tag[:i])
 		tag = tag[i+1:]
 
-		// scan quoted string to find value
+		// Scan quoted string to find value.
 		i = 1
 		for i < len(tag) && tag[i] != '"' {
 			if tag[i] == '\\' {
@@ -800,7 +805,10 @@ func (tag StructTag) Get(key string) string {
 		tag = tag[i+1:]
 
 		if key == name {
-			value, _ := strconv.Unquote(qvalue)
+			value, err := strconv.Unquote(qvalue)
+			if err != nil {
+				break
+			}
 			return value
 		}
 	}
diff --git a/src/reflect/value.go b/src/reflect/value.go
index 4060206eac..081c4d9d7b 100644
--- a/src/reflect/value.go
+++ b/src/reflect/value.go
@@ -2439,12 +2439,27 @@ func chansend(t *rtype, ch unsafe.Pointer, val unsafe.Pointer, nb bool) bool
 
 func makechan(typ *rtype, size uint64) (ch unsafe.Pointer)
 func makemap(t *rtype) (m unsafe.Pointer)
+
+//go:noescape
 func mapaccess(t *rtype, m unsafe.Pointer, key unsafe.Pointer) (val unsafe.Pointer)
+
 func mapassign(t *rtype, m unsafe.Pointer, key, val unsafe.Pointer)
+
+//go:noescape
 func mapdelete(t *rtype, m unsafe.Pointer, key unsafe.Pointer)
+
+// m escapes into the return value, but the caller of mapiterinit
+// doesn't let the return value escape.
+//go:noescape
 func mapiterinit(t *rtype, m unsafe.Pointer) unsafe.Pointer
+
+//go:noescape
 func mapiterkey(it unsafe.Pointer) (key unsafe.Pointer)
+
+//go:noescape
 func mapiternext(it unsafe.Pointer)
+
+//go:noescape
 func maplen(m unsafe.Pointer) int
 
 // call calls fn with a copy of the n argument bytes pointed at by arg.
diff --git a/src/runtime/asm_386.s b/src/runtime/asm_386.s
index 49bba32ebe..58a0d502bd 100644
--- a/src/runtime/asm_386.s
+++ b/src/runtime/asm_386.s
@@ -1298,26 +1298,22 @@ eq:
 	RET
 
 // eqstring tests whether two strings are equal.
+// The compiler guarantees that strings passed
+// to eqstring have equal length.
 // See runtime_test.go:eqstring_generic for
 // equivalent Go code.
 TEXT runtime·eqstring(SB),NOSPLIT,$0-17
-	MOVL	s1len+4(FP), AX
-	MOVL	s2len+12(FP), BX
-	CMPL	AX, BX
-	JNE	different
 	MOVL	s1str+0(FP), SI
 	MOVL	s2str+8(FP), DI
 	CMPL	SI, DI
 	JEQ	same
+	MOVL	s1len+4(FP), BX
 	CALL	runtime·memeqbody(SB)
 	MOVB	AX, v+16(FP)
 	RET
 same:
 	MOVB	$1, v+16(FP)
 	RET
-different:
-	MOVB	$0, v+16(FP)
-	RET
 
 TEXT bytes·Equal(SB),NOSPLIT,$0-25
 	MOVL	a_len+4(FP), BX
diff --git a/src/runtime/asm_amd64.s b/src/runtime/asm_amd64.s
index f09e5ae250..f6c1c5f6e6 100644
--- a/src/runtime/asm_amd64.s
+++ b/src/runtime/asm_amd64.s
@@ -1262,26 +1262,22 @@ eq:
 	RET
 
 // eqstring tests whether two strings are equal.
+// The compiler guarantees that strings passed
+// to eqstring have equal length.
 // See runtime_test.go:eqstring_generic for
 // equivalent Go code.
 TEXT runtime·eqstring(SB),NOSPLIT,$0-33
-	MOVQ	s1len+8(FP), AX
-	MOVQ	s2len+24(FP), BX
-	CMPQ	AX, BX
-	JNE	noteq
 	MOVQ	s1str+0(FP), SI
 	MOVQ	s2str+16(FP), DI
 	CMPQ	SI, DI
 	JEQ	eq
+	MOVQ	s1len+8(FP), BX
 	CALL	runtime·memeqbody(SB)
 	MOVB	AX, v+32(FP)
 	RET
 eq:
 	MOVB	$1, v+32(FP)
 	RET
-noteq:
-	MOVB	$0, v+32(FP)
-	RET
 
 // a in SI
 // b in DI
diff --git a/src/runtime/asm_amd64p32.s b/src/runtime/asm_amd64p32.s
index 77355bb998..f2324285a5 100644
--- a/src/runtime/asm_amd64p32.s
+++ b/src/runtime/asm_amd64p32.s
@@ -704,26 +704,22 @@ eq:
 	RET
 
 // eqstring tests whether two strings are equal.
+// The compiler guarantees that strings passed
+// to eqstring have equal length.
 // See runtime_test.go:eqstring_generic for
 // equivalent Go code.
 TEXT runtime·eqstring(SB),NOSPLIT,$0-17
-	MOVL	s1len+4(FP), AX
-	MOVL	s2len+12(FP), BX
-	CMPL	AX, BX
-	JNE	different
 	MOVL	s1str+0(FP), SI
 	MOVL	s2str+8(FP), DI
 	CMPL	SI, DI
 	JEQ	same
+	MOVL	s1len+4(FP), BX
 	CALL	runtime·memeqbody(SB)
 	MOVB	AX, v+16(FP)
 	RET
 same:
 	MOVB	$1, v+16(FP)
 	RET
-different:
-	MOVB	$0, v+16(FP)
-	RET
 
 // a in SI
 // b in DI
diff --git a/src/runtime/asm_arm.s b/src/runtime/asm_arm.s
index f3ce1a8f17..2efeaaa531 100644
--- a/src/runtime/asm_arm.s
+++ b/src/runtime/asm_arm.s
@@ -39,20 +39,8 @@ TEXT runtime·rt0_go(SB),NOSPLIT,$-4
 
 	BL	runtime·emptyfunc(SB)	// fault if stack check is wrong
 
-#ifndef GOOS_nacl
-	// if there is an _cgo_init, call it.
-	MOVW	_cgo_init(SB), R4
-	CMP	$0, R4
-	B.EQ	nocgo
-	MRC     15, 0, R0, C13, C0, 3 	// load TLS base pointer
-	MOVW 	R0, R3 			// arg 3: TLS base pointer
-	MOVW 	$runtime·tlsg(SB), R2 	// arg 2: tlsg
-	MOVW	$setg_gcc<>(SB), R1 	// arg 1: setg
-	MOVW	g, R0 			// arg 0: G
-	BL	(R4) // will clobber R0-R3
-#endif
+	BL	runtime·_initcgo(SB)	// will clobber R0-R3
 
-nocgo:
 	// update stackguard after _cgo_init
 	MOVW	(g_stack+stack_lo)(g), R0
 	ADD	$const__StackGuard, R0
@@ -806,21 +794,18 @@ eq:
 	RET
 
 // eqstring tests whether two strings are equal.
+// The compiler guarantees that strings passed
+// to eqstring have equal length.
 // See runtime_test.go:eqstring_generic for
 // equivalent Go code.
 TEXT runtime·eqstring(SB),NOSPLIT,$-4-17
-	MOVW	s1len+4(FP), R0
-	MOVW	s2len+12(FP), R1
-	MOVW	$0, R7
-	CMP	R0, R1
-	MOVB.NE R7, v+16(FP)
-	RET.NE
 	MOVW	s1str+0(FP), R2
 	MOVW	s2str+8(FP), R3
 	MOVW	$1, R8
 	MOVB	R8, v+16(FP)
 	CMP	R2, R3
 	RET.EQ
+	MOVW	s1len+4(FP), R0
 	ADD	R2, R0, R6
 loop:
 	CMP	R2, R6
@@ -829,14 +814,10 @@ loop:
 	MOVBU.P	1(R3), R5
 	CMP	R4, R5
 	BEQ	loop
-	MOVB	R7, v+16(FP)
+	MOVW	$0, R8
+	MOVB	R8, v+16(FP)
 	RET
 
-// void setg_gcc(G*); set g called from gcc.
-TEXT setg_gcc<>(SB),NOSPLIT,$0
-	MOVW	R0, g
-	B		runtime·save_g(SB)
-
 // TODO: share code with memeq?
 TEXT bytes·Equal(SB),NOSPLIT,$0
 	MOVW	a_len+4(FP), R1
diff --git a/src/runtime/asm_ppc64x.s b/src/runtime/asm_ppc64x.s
index 660c7cdfe5..694dfcc24e 100644
--- a/src/runtime/asm_ppc64x.s
+++ b/src/runtime/asm_ppc64x.s
@@ -969,30 +969,30 @@ eq:
 	RETURN
 
 // eqstring tests whether two strings are equal.
+// The compiler guarantees that strings passed
+// to eqstring have equal length.
 // See runtime_test.go:eqstring_generic for
 // equivalent Go code.
 TEXT runtime·eqstring(SB),NOSPLIT,$0-33
-	MOVD	s1len+8(FP), R4
-	MOVD	s2len+24(FP), R5
-	CMP	R4, R5
-	BNE	noteq
-
 	MOVD	s1str+0(FP), R3
 	MOVD	s2str+16(FP), R4
+	MOVD	$1, R5
+	MOVB	R5, ret+32(FP)
+	CMP	R3, R4
+	BNE	2(PC)
+	RETURN
+	MOVD	s1len+8(FP), R5
 	SUB	$1, R3
 	SUB	$1, R4
 	ADD	R3, R5, R8
 loop:
 	CMP	R3, R8
-	BNE	4(PC)
-	MOVD	$1, R3
-	MOVB	R3, ret+32(FP)
+	BNE	2(PC)
 	RETURN
 	MOVBZU	1(R3), R6
 	MOVBZU	1(R4), R7
 	CMP	R6, R7
 	BEQ	loop
-noteq:
 	MOVB	R0, ret+32(FP)
 	RETURN
 
diff --git a/src/runtime/cgo/cgo.go b/src/runtime/cgo/cgo.go
index 8528692f7b..9a41399cd6 100644
--- a/src/runtime/cgo/cgo.go
+++ b/src/runtime/cgo/cgo.go
@@ -11,7 +11,7 @@ package cgo
 
 /*
 
-#cgo darwin LDFLAGS: -lpthread
+#cgo darwin,!arm LDFLAGS: -lpthread
 #cgo dragonfly LDFLAGS: -lpthread
 #cgo freebsd LDFLAGS: -lpthread
 #cgo android LDFLAGS: -llog
diff --git a/src/runtime/cgo/gcc_arm.S b/src/runtime/cgo/gcc_arm.S
index d5833bfad0..980ab579e4 100644
--- a/src/runtime/cgo/gcc_arm.S
+++ b/src/runtime/cgo/gcc_arm.S
@@ -11,6 +11,10 @@
 #define EXT(s) s
 #endif
 
+// Apple's ld64 wants 4-byte alignment for ARM code sections.
+// .align in both Apple as and GNU as treat n as aligning to 2**n bytes.
+.align	2
+
 /*
  * void crosscall_arm1(void (*fn)(void), void (*setg_gcc)(void *g), void *g)
  *
diff --git a/src/runtime/cgo/gcc_darwin_arm.c b/src/runtime/cgo/gcc_darwin_arm.c
new file mode 100644
index 0000000000..d56c55777d
--- /dev/null
+++ b/src/runtime/cgo/gcc_darwin_arm.c
@@ -0,0 +1,99 @@
+// Copyright 2014 The Go Authors.  All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#include <string.h> /* for strerror */
+#include <pthread.h>
+#include <signal.h>
+#include <limits.h>
+#include "libcgo.h"
+
+#define magic (0xe696c4f4U)
+
+// inittls allocates a thread-local storage slot for g.
+//
+// It finds the first available slot using pthread_key_create and uses
+// it as the offset value for runtime.tlsg.
+static void
+inittls(void **tlsg, void **tlsbase)
+{
+	pthread_key_t k;
+	int i, err;
+
+	err = pthread_key_create(&k, nil);
+	if(err != 0) {
+		fprintf(stderr, "runtime/cgo: pthread_key_create failed: %d\n", err);
+		abort();
+	}
+	//fprintf(stderr, "runtime/cgo: k = %d, tlsbase = %p\n", (int)k, tlsbase); // debug
+	pthread_setspecific(k, (void*)magic);
+	// The first key should be at 258.
+	for (i=0; i<PTHREAD_KEYS_MAX; i++) {
+		if (*(tlsbase+i) == (void*)magic) {
+			*tlsg = (void*)(i*sizeof(void *));
+			pthread_setspecific(k, 0);
+			return;
+		}
+	}
+	fprintf(stderr, "runtime/cgo: could not find pthread key.\n");
+	abort();
+}
+
+static void *threadentry(void*);
+void (*setg_gcc)(void*);
+
+void
+_cgo_sys_thread_start(ThreadStart *ts)
+{
+	pthread_attr_t attr;
+	sigset_t ign, oset;
+	pthread_t p;
+	size_t size;
+	int err;
+
+	sigfillset(&ign);
+	pthread_sigmask(SIG_SETMASK, &ign, &oset);
+
+	pthread_attr_init(&attr);
+	size = 0;
+	pthread_attr_getstacksize(&attr, &size);
+	// Leave stacklo=0 and set stackhi=size; mstack will do the rest.
+	ts->g->stackhi = size;
+	err = pthread_create(&p, &attr, threadentry, ts);
+
+	pthread_sigmask(SIG_SETMASK, &oset, nil);
+
+	if (err != 0) {
+		fprintf(stderr, "runtime/cgo: pthread_create failed: %s\n", strerror(err));
+		abort();
+	}
+}
+
+extern void crosscall_arm1(void (*fn)(void), void (*setg_gcc)(void*), void *g);
+static void*
+threadentry(void *v)
+{
+	ThreadStart ts;
+
+	ts = *(ThreadStart*)v;
+	free(v);
+
+	crosscall_arm1(ts.fn, setg_gcc, (void*)ts.g);
+	return nil;
+}
+
+void
+x_cgo_init(G *g, void (*setg)(void*), void **tlsg, void **tlsbase)
+{
+	pthread_attr_t attr;
+	size_t size;
+
+	setg_gcc = setg;
+	pthread_attr_init(&attr);
+	pthread_attr_getstacksize(&attr, &size);
+	g->stacklo = (uintptr)&attr - size + 4096;
+	pthread_attr_destroy(&attr);
+
+	// yes, tlsbase from mrc might not be correctly aligned.
+	inittls(tlsg, (void**)((uintptr)tlsbase & ~3));
+}
diff --git a/src/runtime/compiler.go b/src/runtime/compiler.go
index 562a460226..f6edc95959 100644
--- a/src/runtime/compiler.go
+++ b/src/runtime/compiler.go
@@ -7,7 +7,7 @@ package runtime
 // Compiler is the name of the compiler toolchain that built the
 // running binary.  Known toolchains are:
 //
-//	gc      The 5g/6g/8g compiler suite at code.google.com/p/go.
+//	gc      The 5g/6g/8g compiler suite at go.googlesource.com/go.
 //	gccgo   The gccgo front end, part of the GCC compiler suite.
 //
 const Compiler = "gc"
diff --git a/src/runtime/crash_test.go b/src/runtime/crash_test.go
index 43cea9008a..715b2da232 100644
--- a/src/runtime/crash_test.go
+++ b/src/runtime/crash_test.go
@@ -10,6 +10,7 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"regexp"
 	"runtime"
 	"strings"
 	"sync"
@@ -17,17 +18,20 @@ import (
 	"text/template"
 )
 
-// testEnv excludes GODEBUG from the environment
-// to prevent its output from breaking tests that
-// are trying to parse other command output.
 func testEnv(cmd *exec.Cmd) *exec.Cmd {
 	if cmd.Env != nil {
 		panic("environment already set")
 	}
 	for _, env := range os.Environ() {
+		// Exclude GODEBUG from the environment to prevent its output
+		// from breaking tests that are trying to parse other command output.
 		if strings.HasPrefix(env, "GODEBUG=") {
 			continue
 		}
+		// Exclude GOTRACEBACK for the same reason.
+		if strings.HasPrefix(env, "GOTRACEBACK=") {
+			continue
+		}
 		cmd.Env = append(cmd.Env, env)
 	}
 	return cmd
@@ -217,6 +221,14 @@ func TestMainGoroutineId(t *testing.T) {
 	}
 }
 
+func TestNoHelperGoroutines(t *testing.T) {
+	output := executeTest(t, noHelperGoroutinesSource, nil)
+	matches := regexp.MustCompile(`goroutine [0-9]+ \[`).FindAllStringSubmatch(output, -1)
+	if len(matches) != 1 || matches[0][0] != "goroutine 1 [" {
+		t.Fatalf("want to see only goroutine 1, see:\n%s", output)
+	}
+}
+
 func TestBreakpoint(t *testing.T) {
 	output := executeTest(t, breakpointSource, nil)
 	want := "runtime.Breakpoint()"
@@ -431,6 +443,22 @@ func main() {
 }
 `
 
+const noHelperGoroutinesSource = `
+package main
+import (
+	"runtime"
+	"time"
+)
+func init() {
+	i := 0
+	runtime.SetFinalizer(&i, func(p *int) {})
+	time.AfterFunc(time.Hour, func() {})
+	panic("oops")
+}
+func main() {
+}
+`
+
 const breakpointSource = `
 package main
 import "runtime"
diff --git a/src/runtime/defs_darwin_arm.go b/src/runtime/defs_darwin_arm.go
new file mode 100644
index 0000000000..92bab509fb
--- /dev/null
+++ b/src/runtime/defs_darwin_arm.go
@@ -0,0 +1,245 @@
+// Note: cgo can't handle some Darwin/ARM structures, so this file can't
+// be auto generated by cgo yet.
+// Created based on output of `cgo -cdefs defs_darwin.go` and Darwin/ARM
+// specific header (mainly mcontext and ucontext related stuff)
+
+package runtime
+
+import "unsafe"
+
+const (
+	_EINTR  = 0x4
+	_EFAULT = 0xe
+
+	_PROT_NONE  = 0x0
+	_PROT_READ  = 0x1
+	_PROT_WRITE = 0x2
+	_PROT_EXEC  = 0x4
+
+	_MAP_ANON    = 0x1000
+	_MAP_PRIVATE = 0x2
+	_MAP_FIXED   = 0x10
+
+	_MADV_DONTNEED = 0x4
+	_MADV_FREE     = 0x5
+
+	_MACH_MSG_TYPE_MOVE_RECEIVE   = 0x10
+	_MACH_MSG_TYPE_MOVE_SEND      = 0x11
+	_MACH_MSG_TYPE_MOVE_SEND_ONCE = 0x12
+	_MACH_MSG_TYPE_COPY_SEND      = 0x13
+	_MACH_MSG_TYPE_MAKE_SEND      = 0x14
+	_MACH_MSG_TYPE_MAKE_SEND_ONCE = 0x15
+	_MACH_MSG_TYPE_COPY_RECEIVE   = 0x16
+
+	_MACH_MSG_PORT_DESCRIPTOR         = 0x0
+	_MACH_MSG_OOL_DESCRIPTOR          = 0x1
+	_MACH_MSG_OOL_PORTS_DESCRIPTOR    = 0x2
+	_MACH_MSG_OOL_VOLATILE_DESCRIPTOR = 0x3
+
+	_MACH_MSGH_BITS_COMPLEX = 0x80000000
+
+	_MACH_SEND_MSG  = 0x1
+	_MACH_RCV_MSG   = 0x2
+	_MACH_RCV_LARGE = 0x4
+
+	_MACH_SEND_TIMEOUT   = 0x10
+	_MACH_SEND_INTERRUPT = 0x40
+	_MACH_SEND_ALWAYS    = 0x10000
+	_MACH_SEND_TRAILER   = 0x20000
+	_MACH_RCV_TIMEOUT    = 0x100
+	_MACH_RCV_NOTIFY     = 0x200
+	_MACH_RCV_INTERRUPT  = 0x400
+	_MACH_RCV_OVERWRITE  = 0x1000
+
+	_NDR_PROTOCOL_2_0      = 0x0
+	_NDR_INT_BIG_ENDIAN    = 0x0
+	_NDR_INT_LITTLE_ENDIAN = 0x1
+	_NDR_FLOAT_IEEE        = 0x0
+	_NDR_CHAR_ASCII        = 0x0
+
+	_SA_SIGINFO   = 0x40
+	_SA_RESTART   = 0x2
+	_SA_ONSTACK   = 0x1
+	_SA_USERTRAMP = 0x100
+	_SA_64REGSET  = 0x200
+
+	_SIGHUP    = 0x1
+	_SIGINT    = 0x2
+	_SIGQUIT   = 0x3
+	_SIGILL    = 0x4
+	_SIGTRAP   = 0x5
+	_SIGABRT   = 0x6
+	_SIGEMT    = 0x7
+	_SIGFPE    = 0x8
+	_SIGKILL   = 0x9
+	_SIGBUS    = 0xa
+	_SIGSEGV   = 0xb
+	_SIGSYS    = 0xc
+	_SIGPIPE   = 0xd
+	_SIGALRM   = 0xe
+	_SIGTERM   = 0xf
+	_SIGURG    = 0x10
+	_SIGSTOP   = 0x11
+	_SIGTSTP   = 0x12
+	_SIGCONT   = 0x13
+	_SIGCHLD   = 0x14
+	_SIGTTIN   = 0x15
+	_SIGTTOU   = 0x16
+	_SIGIO     = 0x17
+	_SIGXCPU   = 0x18
+	_SIGXFSZ   = 0x19
+	_SIGVTALRM = 0x1a
+	_SIGPROF   = 0x1b
+	_SIGWINCH  = 0x1c
+	_SIGINFO   = 0x1d
+	_SIGUSR1   = 0x1e
+	_SIGUSR2   = 0x1f
+
+	_FPE_INTDIV = 0x7
+	_FPE_INTOVF = 0x8
+	_FPE_FLTDIV = 0x1
+	_FPE_FLTOVF = 0x2
+	_FPE_FLTUND = 0x3
+	_FPE_FLTRES = 0x4
+	_FPE_FLTINV = 0x5
+	_FPE_FLTSUB = 0x6
+
+	_BUS_ADRALN = 0x1
+	_BUS_ADRERR = 0x2
+	_BUS_OBJERR = 0x3
+
+	_SEGV_MAPERR = 0x1
+	_SEGV_ACCERR = 0x2
+
+	_ITIMER_REAL    = 0x0
+	_ITIMER_VIRTUAL = 0x1
+	_ITIMER_PROF    = 0x2
+
+	_EV_ADD       = 0x1
+	_EV_DELETE    = 0x2
+	_EV_CLEAR     = 0x20
+	_EV_RECEIPT   = 0x40
+	_EV_ERROR     = 0x4000
+	_EVFILT_READ  = -0x1
+	_EVFILT_WRITE = -0x2
+)
+
+type machbody struct {
+	msgh_descriptor_count uint32
+}
+
+type machheader struct {
+	msgh_bits        uint32
+	msgh_size        uint32
+	msgh_remote_port uint32
+	msgh_local_port  uint32
+	msgh_reserved    uint32
+	msgh_id          int32
+}
+
+type machndr struct {
+	mig_vers     uint8
+	if_vers      uint8
+	reserved1    uint8
+	mig_encoding uint8
+	int_rep      uint8
+	char_rep     uint8
+	float_rep    uint8
+	reserved2    uint8
+}
+
+type machport struct {
+	name        uint32
+	pad1        uint32
+	pad2        uint16
+	disposition uint8
+	_type       uint8
+}
+
+type stackt struct {
+	ss_sp    *byte
+	ss_size  uintptr
+	ss_flags int32
+}
+
+type sigactiont struct {
+	__sigaction_u [4]byte
+	sa_tramp      unsafe.Pointer
+	sa_mask       uint32
+	sa_flags      int32
+}
+
+type siginfo struct {
+	si_signo  int32
+	si_errno  int32
+	si_code   int32
+	si_pid    int32
+	si_uid    uint32
+	si_status int32
+	si_addr   *byte
+	si_value  [4]byte
+	si_band   int32
+	__pad     [7]uint32
+}
+
+type timeval struct {
+	tv_sec  int32
+	tv_usec int32
+}
+
+func (tv *timeval) set_usec(x int32) {
+	tv.tv_usec = x
+}
+
+type itimerval struct {
+	it_interval timeval
+	it_value    timeval
+}
+
+type timespec struct {
+	tv_sec  int32
+	tv_nsec int32
+}
+
+type floatstate32 struct {
+	r     [32]uint32
+	fpscr uint32
+}
+
+type regs32 struct {
+	r    [13]uint32 // r0 to r12
+	sp   uint32     // r13
+	lr   uint32     // r14
+	pc   uint32     // r15
+	cpsr uint32
+}
+
+type exceptionstate32 struct {
+	trapno     uint32 // NOTE: on 386, the trapno field is split into trapno and cpu
+	err        uint32
+	faultvaddr uint32
+}
+
+type mcontext32 struct {
+	es exceptionstate32
+	ss regs32
+	fs floatstate32
+}
+
+type ucontext struct {
+	uc_onstack  int32
+	uc_sigmask  uint32
+	uc_stack    stackt
+	uc_link     *ucontext
+	uc_mcsize   uint32
+	uc_mcontext *mcontext32
+}
+
+type keventt struct {
+	ident  uint32
+	filter int16
+	flags  uint16
+	fflags uint32
+	data   int32
+	udata  *byte
+}
diff --git a/src/runtime/defs_windows_386.go b/src/runtime/defs_windows_386.go
index abec2d839f..c860f74a3f 100644
--- a/src/runtime/defs_windows_386.go
+++ b/src/runtime/defs_windows_386.go
@@ -101,6 +101,12 @@ type context struct {
 	extendedregisters [512]uint8
 }
 
+func (c *context) ip() uintptr { return uintptr(c.eip) }
+func (c *context) sp() uintptr { return uintptr(c.esp) }
+
+func (c *context) setip(x uintptr) { c.eip = uint32(x) }
+func (c *context) setsp(x uintptr) { c.esp = uint32(x) }
+
 type overlapped struct {
 	internal     uint32
 	internalhigh uint32
diff --git a/src/runtime/defs_windows_amd64.go b/src/runtime/defs_windows_amd64.go
index 81b13597b7..d1e55ec426 100644
--- a/src/runtime/defs_windows_amd64.go
+++ b/src/runtime/defs_windows_amd64.go
@@ -116,6 +116,12 @@ type context struct {
 	lastexceptionfromrip uint64
 }
 
+func (c *context) ip() uintptr { return uintptr(c.rip) }
+func (c *context) sp() uintptr { return uintptr(c.rsp) }
+
+func (c *context) setip(x uintptr) { c.rip = uint64(x) }
+func (c *context) setsp(x uintptr) { c.rsp = uint64(x) }
+
 type overlapped struct {
 	internal     uint64
 	internalhigh uint64
diff --git a/src/runtime/gcwork.go b/src/runtime/gcwork.go
new file mode 100644
index 0000000000..cf5a97957f
--- /dev/null
+++ b/src/runtime/gcwork.go
@@ -0,0 +1,338 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package runtime
+
+import "unsafe"
+
+const (
+	_Debugwbufs  = true    // if true check wbufs consistency
+	_WorkbufSize = 1 * 256 // in bytes - if small wbufs are passed to GC in a timely fashion.
+)
+
+type workbufhdr struct {
+	node  lfnode // must be first
+	nobj  uintptr
+	id    uintptr
+	inuse bool       // This workbuf is in use by some gorotuine and is not on the work.empty/partial/full queues.
+	log   [4]uintptr // line numbers forming a history of ownership changes to workbuf
+}
+
+type workbuf struct {
+	workbufhdr
+	// account for the above fields
+	obj [(_WorkbufSize - unsafe.Sizeof(workbufhdr{})) / ptrSize]uintptr
+}
+
+// workbuf factory routines. These funcs are used to manage the
+// workbufs. They cache workbuf in the m struct field currentwbuf.
+// If the GC asks for some work these are the only routines that
+// make partially full wbufs available to the GC.
+// Each of the gets and puts also take an distinct integer that is used
+// to record a brief history of changes to ownership of the workbuf.
+// The convention is to use a unique line number but any encoding
+// is permissible. For example if you want to pass in 2 bits of information
+// you could simple add lineno1*100000+lineno2.
+
+// logget records the past few values of entry to aid in debugging.
+// logget checks the buffer b is not currently in use.
+func (b *workbuf) logget(entry uintptr) {
+	if !_Debugwbufs {
+		return
+	}
+	if b.inuse {
+		println("runtime: logget fails log entry=", entry,
+			"b.log[0]=", b.log[0], "b.log[1]=", b.log[1],
+			"b.log[2]=", b.log[2], "b.log[3]=", b.log[3])
+		throw("logget: get not legal")
+	}
+	b.inuse = true
+	copy(b.log[1:], b.log[:])
+	b.log[0] = entry
+}
+
+// logput records the past few values of entry to aid in debugging.
+// logput checks the buffer b is currently in use.
+func (b *workbuf) logput(entry uintptr) {
+	if !_Debugwbufs {
+		return
+	}
+	if !b.inuse {
+		println("runtime:logput fails log entry=", entry,
+			"b.log[0]=", b.log[0], "b.log[1]=", b.log[1],
+			"b.log[2]=", b.log[2], "b.log[3]=", b.log[3])
+		throw("logput: put not legal")
+	}
+	b.inuse = false
+	copy(b.log[1:], b.log[:])
+	b.log[0] = entry
+}
+
+func (b *workbuf) checknonempty() {
+	if b.nobj == 0 {
+		println("runtime: nonempty check fails",
+			"b.log[0]=", b.log[0], "b.log[1]=", b.log[1],
+			"b.log[2]=", b.log[2], "b.log[3]=", b.log[3])
+		throw("workbuf is empty")
+	}
+}
+
+func (b *workbuf) checkempty() {
+	if b.nobj != 0 {
+		println("runtime: empty check fails",
+			"b.log[0]=", b.log[0], "b.log[1]=", b.log[1],
+			"b.log[2]=", b.log[2], "b.log[3]=", b.log[3])
+		throw("workbuf is not empty")
+	}
+}
+
+// checknocurrentwbuf checks that the m's currentwbuf field is empty
+func checknocurrentwbuf() {
+	if getg().m.currentwbuf != 0 {
+		throw("unexpected currentwbuf")
+	}
+}
+
+// getempty pops an empty work buffer off the work.empty list,
+// allocating new buffers if none are available.
+// entry is used to record a brief history of ownership.
+//go:nowritebarrier
+func getempty(entry uintptr) *workbuf {
+	var b *workbuf
+	if work.empty != 0 {
+		b = (*workbuf)(lfstackpop(&work.empty))
+		if b != nil {
+			b.checkempty()
+		}
+	}
+	if b == nil {
+		b = (*workbuf)(persistentalloc(unsafe.Sizeof(*b), _CacheLineSize, &memstats.gc_sys))
+	}
+	b.logget(entry)
+	return b
+}
+
+// putempty puts a workbuf onto the work.empty list.
+// Upon entry this go routine owns b. The lfstackpush relinquishes ownership.
+//go:nowritebarrier
+func putempty(b *workbuf, entry uintptr) {
+	b.checkempty()
+	b.logput(entry)
+	lfstackpush(&work.empty, &b.node)
+}
+
+// putfull puts the workbuf on the work.full list for the GC.
+// putfull accepts partially full buffers so the GC can avoid competing
+// with the mutators for ownership of partially full buffers.
+//go:nowritebarrier
+func putfull(b *workbuf, entry uintptr) {
+	b.checknonempty()
+	b.logput(entry)
+	lfstackpush(&work.full, &b.node)
+}
+
+// getpartialorempty tries to return a partially empty
+// and if none are available returns an empty one.
+// entry is used to provide a brief histoy of ownership
+// using entry + xxx00000 to
+// indicating that two line numbers in the call chain.
+//go:nowritebarrier
+func getpartialorempty(entry uintptr) *workbuf {
+	var b *workbuf
+	// If this m has a buf in currentwbuf then as an optimization
+	// simply return that buffer. If it turns out currentwbuf
+	// is full, put it on the work.full queue and get another
+	// workbuf off the partial or empty queue.
+	if getg().m.currentwbuf != 0 {
+		b = (*workbuf)(unsafe.Pointer(xchguintptr(&getg().m.currentwbuf, 0)))
+		if b != nil {
+			if b.nobj <= uintptr(len(b.obj)) {
+				return b
+			}
+			putfull(b, entry+80100000)
+		}
+	}
+	b = (*workbuf)(lfstackpop(&work.partial))
+	if b != nil {
+		b.logget(entry)
+		return b
+	}
+	// Let getempty do the logget check but
+	// use the entry to encode that it passed
+	// through this routine.
+	b = getempty(entry + 80700000)
+	return b
+}
+
+// putpartial puts empty buffers on the work.empty queue,
+// full buffers on the work.full queue and
+// others on the work.partial queue.
+// entry is used to provide a brief histoy of ownership
+// using entry + xxx00000 to
+// indicating that two call chain line numbers.
+//go:nowritebarrier
+func putpartial(b *workbuf, entry uintptr) {
+	if b.nobj == 0 {
+		putempty(b, entry+81500000)
+	} else if b.nobj < uintptr(len(b.obj)) {
+		b.logput(entry)
+		lfstackpush(&work.partial, &b.node)
+	} else if b.nobj == uintptr(len(b.obj)) {
+		b.logput(entry)
+		lfstackpush(&work.full, &b.node)
+	} else {
+		throw("putpartial: bad Workbuf b.nobj")
+	}
+}
+
+// trygetfull tries to get a full or partially empty workbuffer.
+// If one is not immediately available return nil
+//go:nowritebarrier
+func trygetfull(entry uintptr) *workbuf {
+	b := (*workbuf)(lfstackpop(&work.full))
+	if b == nil {
+		b = (*workbuf)(lfstackpop(&work.partial))
+	}
+	if b != nil {
+		b.logget(entry)
+		b.checknonempty()
+		return b
+	}
+	// full and partial are both empty so see if there
+	// is an work available on currentwbuf.
+	// This is an optimization to shift
+	// processing from the STW marktermination phase into
+	// the concurrent mark phase.
+	if getg().m.currentwbuf != 0 {
+		b = (*workbuf)(unsafe.Pointer(xchguintptr(&getg().m.currentwbuf, 0)))
+		if b != nil {
+			if b.nobj != 0 {
+				return b
+			}
+			putempty(b, 839)
+			b = nil
+		}
+	}
+	return b
+}
+
+// Get a full work buffer off the work.full or a partially
+// filled one off the work.partial list. If nothing is available
+// wait until all the other gc helpers have finished and then
+// return nil.
+// getfull acts as a barrier for work.nproc helpers. As long as one
+// gchelper is actively marking objects it
+// may create a workbuffer that the other helpers can work on.
+// The for loop either exits when a work buffer is found
+// or when _all_ of the work.nproc GC helpers are in the loop
+// looking for work and thus not capable of creating new work.
+// This is in fact the termination condition for the STW mark
+// phase.
+//go:nowritebarrier
+func getfull(entry uintptr) *workbuf {
+	b := (*workbuf)(lfstackpop(&work.full))
+	if b != nil {
+		b.logget(entry)
+		b.checknonempty()
+		return b
+	}
+	b = (*workbuf)(lfstackpop(&work.partial))
+	if b != nil {
+		b.logget(entry)
+		return b
+	}
+	// Make sure that currentwbuf is also not a source for pointers to be
+	// processed. This is an optimization that shifts processing
+	// from the mark termination STW phase to the concurrent mark phase.
+	if getg().m.currentwbuf != 0 {
+		b = (*workbuf)(unsafe.Pointer(xchguintptr(&getg().m.currentwbuf, 0)))
+		if b != nil {
+			if b.nobj != 0 {
+				return b
+			}
+			putempty(b, 877)
+			b = nil
+		}
+	}
+
+	xadd(&work.nwait, +1)
+	for i := 0; ; i++ {
+		if work.full != 0 {
+			xadd(&work.nwait, -1)
+			b = (*workbuf)(lfstackpop(&work.full))
+			if b == nil {
+				b = (*workbuf)(lfstackpop(&work.partial))
+			}
+			if b != nil {
+				b.logget(entry)
+				b.checknonempty()
+				return b
+			}
+			xadd(&work.nwait, +1)
+		}
+		if work.nwait == work.nproc {
+			return nil
+		}
+		_g_ := getg()
+		if i < 10 {
+			_g_.m.gcstats.nprocyield++
+			procyield(20)
+		} else if i < 20 {
+			_g_.m.gcstats.nosyield++
+			osyield()
+		} else {
+			_g_.m.gcstats.nsleep++
+			usleep(100)
+		}
+	}
+}
+
+//go:nowritebarrier
+func handoff(b *workbuf) *workbuf {
+	// Make new buffer with half of b's pointers.
+	b1 := getempty(915)
+	n := b.nobj / 2
+	b.nobj -= n
+	b1.nobj = n
+	memmove(unsafe.Pointer(&b1.obj[0]), unsafe.Pointer(&b.obj[b.nobj]), n*unsafe.Sizeof(b1.obj[0]))
+	_g_ := getg()
+	_g_.m.gcstats.nhandoff++
+	_g_.m.gcstats.nhandoffcnt += uint64(n)
+
+	// Put b on full list - let first half of b get stolen.
+	putfull(b, 942)
+	return b1
+}
+
+// 1 when you are harvesting so that the write buffer code shade can
+// detect calls during a presumable STW write barrier.
+var harvestingwbufs uint32
+
+// harvestwbufs moves non-empty workbufs to work.full from  m.currentwuf
+// Must be in a STW phase.
+// xchguintptr is used since there are write barrier calls from the GC helper
+// routines even during a STW phase.
+// TODO: chase down write barrier calls in STW phase and understand and eliminate
+// them.
+//go:nowritebarrier
+func harvestwbufs() {
+	// announce to write buffer that you are harvesting the currentwbufs
+	atomicstore(&harvestingwbufs, 1)
+
+	for mp := allm; mp != nil; mp = mp.alllink {
+		wbuf := (*workbuf)(unsafe.Pointer(xchguintptr(&mp.currentwbuf, 0)))
+		// TODO: beat write barriers out of the mark termination and eliminate xchg
+		//		tempwbuf := (*workbuf)(unsafe.Pointer(tempm.currentwbuf))
+		//		tempm.currentwbuf = 0
+		if wbuf != nil {
+			if wbuf.nobj == 0 {
+				putempty(wbuf, 945)
+			} else {
+				putfull(wbuf, 947) //use full instead of partial so GC doesn't compete to get wbuf
+			}
+		}
+	}
+
+	atomicstore(&harvestingwbufs, 0)
+}
diff --git a/src/runtime/hashmap.go b/src/runtime/hashmap.go
index 058d1c76c4..c7c1198259 100644
--- a/src/runtime/hashmap.go
+++ b/src/runtime/hashmap.go
@@ -182,8 +182,14 @@ func (h *hmap) createOverflow() {
 	}
 }
 
-func makemap(t *maptype, hint int64) *hmap {
+// makemap implements a Go map creation make(map[k]v, hint)
+// If the compiler has determined that the map or the first bucket
+// can be created on the stack, h and/or bucket may be non-nil.
+// If h != nil, the map can be created directly in h.
+// If bucket != nil, bucket can be used as the first bucket.
+func makemap(t *maptype, hint int64, h *hmap, bucket unsafe.Pointer) *hmap {
 	if sz := unsafe.Sizeof(hmap{}); sz > 48 || sz != uintptr(t.hmap.size) {
+		println("runtime: sizeof(hmap) =", sz, ", t.hmap.size =", t.hmap.size)
 		throw("bad hmap size")
 	}
 
@@ -238,7 +244,7 @@ func makemap(t *maptype, hint int64) *hmap {
 	// allocate initial hash table
 	// if B == 0, the buckets field is allocated lazily later (in mapassign)
 	// If hint is large zeroing this memory could take a while.
-	var buckets unsafe.Pointer
+	buckets := bucket
 	if B != 0 {
 		if checkgc {
 			memstats.next_gc = memstats.heap_alloc
@@ -250,7 +256,9 @@ func makemap(t *maptype, hint int64) *hmap {
 	if checkgc {
 		memstats.next_gc = memstats.heap_alloc
 	}
-	h := (*hmap)(newobject(t.hmap))
+	if h == nil {
+		h = (*hmap)(newobject(t.hmap))
+	}
 	h.count = 0
 	h.B = B
 	h.flags = 0
@@ -956,7 +964,7 @@ func ismapkey(t *_type) bool {
 
 //go:linkname reflect_makemap reflect.makemap
 func reflect_makemap(t *maptype) *hmap {
-	return makemap(t, 0)
+	return makemap(t, 0, nil, nil)
 }
 
 //go:linkname reflect_mapaccess reflect.mapaccess
diff --git a/src/runtime/heapdump.go b/src/runtime/heapdump.go
index 6cbb5f3775..7b4a846195 100644
--- a/src/runtime/heapdump.go
+++ b/src/runtime/heapdump.go
@@ -344,7 +344,7 @@ func dumpgoroutine(gp *g) {
 	dumpint(uint64(gp.goid))
 	dumpint(uint64(gp.gopc))
 	dumpint(uint64(readgstatus(gp)))
-	dumpbool(gp.issystem)
+	dumpbool(isSystemGoroutine(gp))
 	dumpbool(false) // isbackground
 	dumpint(uint64(gp.waitsince))
 	dumpstr(gp.waitreason)
diff --git a/src/runtime/iface_test.go b/src/runtime/iface_test.go
index bca0ea0ee7..bfeb94b8aa 100644
--- a/src/runtime/iface_test.go
+++ b/src/runtime/iface_test.go
@@ -5,6 +5,7 @@
 package runtime_test
 
 import (
+	"runtime"
 	"testing"
 )
 
@@ -38,6 +39,47 @@ var (
 	tl TL
 )
 
+// Issue 9370
+func TestCmpIfaceConcreteAlloc(t *testing.T) {
+	if runtime.Compiler != "gc" {
+		t.Skip("skipping on non-gc compiler")
+	}
+
+	n := testing.AllocsPerRun(1, func() {
+		_ = e == ts
+		_ = i1 == ts
+		_ = e == 1
+	})
+
+	if n > 0 {
+		t.Fatalf("iface cmp allocs=%v; want 0", n)
+	}
+}
+
+func BenchmarkEqEfaceConcrete(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		_ = e == ts
+	}
+}
+
+func BenchmarkEqIfaceConcrete(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		_ = i1 == ts
+	}
+}
+
+func BenchmarkNeEfaceConcrete(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		_ = e != ts
+	}
+}
+
+func BenchmarkNeIfaceConcrete(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		_ = i1 != ts
+	}
+}
+
 func BenchmarkConvT2ESmall(b *testing.B) {
 	for i := 0; i < b.N; i++ {
 		e = ts
diff --git a/src/runtime/malloc.go b/src/runtime/malloc.go
index b8b1f4ed36..06ba124473 100644
--- a/src/runtime/malloc.go
+++ b/src/runtime/malloc.go
@@ -312,6 +312,7 @@ func profilealloc(mp *m, x unsafe.Pointer, size uintptr) {
 
 // For now this must be bracketed with a stoptheworld and a starttheworld to ensure
 // all go routines see the new barrier.
+//go:nowritebarrier
 func gcinstallmarkwb() {
 	gcphase = _GCmark
 }
@@ -389,6 +390,7 @@ func gcwork(force int32) {
 		gctimer.cycle.installmarkwb = nanotime()
 		systemstack(stoptheworld)
 		systemstack(gcinstallmarkwb)
+		systemstack(harvestwbufs)
 		systemstack(starttheworld)
 		gctimer.cycle.mark = nanotime()
 		systemstack(gcmark_m)
diff --git a/src/runtime/malloc1.go b/src/runtime/malloc1.go
index 4d0754ba9d..18d998b554 100644
--- a/src/runtime/malloc1.go
+++ b/src/runtime/malloc1.go
@@ -99,7 +99,7 @@ func mallocinit() {
 	var reserved bool
 
 	// limit = runtime.memlimit();
-	// See https://code.google.com/p/go/issues/detail?id=5049
+	// See https://golang.org/issue/5049
 	// TODO(rsc): Fix after 1.1.
 	limit = 0
 
diff --git a/src/runtime/map_test.go b/src/runtime/map_test.go
index 92da2d8209..55f1f82625 100644
--- a/src/runtime/map_test.go
+++ b/src/runtime/map_test.go
@@ -535,3 +535,13 @@ func benchmarkMapPop(b *testing.B, n int) {
 func BenchmarkMapPop100(b *testing.B)   { benchmarkMapPop(b, 100) }
 func BenchmarkMapPop1000(b *testing.B)  { benchmarkMapPop(b, 1000) }
 func BenchmarkMapPop10000(b *testing.B) { benchmarkMapPop(b, 10000) }
+
+func TestNonEscapingMap(t *testing.T) {
+	n := testing.AllocsPerRun(1000, func() {
+		m := make(map[int]int)
+		m[0] = 0
+	})
+	if n != 0 {
+		t.Fatalf("want 0 allocs, got %v", n)
+	}
+}
diff --git a/src/runtime/mapspeed_test.go b/src/runtime/mapspeed_test.go
index 119eb3f39c..b036d2a3ab 100644
--- a/src/runtime/mapspeed_test.go
+++ b/src/runtime/mapspeed_test.go
@@ -234,6 +234,15 @@ func BenchmarkNewEmptyMap(b *testing.B) {
 	}
 }
 
+func BenchmarkNewSmallMap(b *testing.B) {
+	b.ReportAllocs()
+	for i := 0; i < b.N; i++ {
+		m := make(map[int]int)
+		m[0] = 0
+		m[1] = 1
+	}
+}
+
 func BenchmarkMapIter(b *testing.B) {
 	m := make(map[int]bool)
 	for i := 0; i < 8; i++ {
diff --git a/src/runtime/mbarrier.go b/src/runtime/mbarrier.go
index 33d67c4976..f6e9269858 100644
--- a/src/runtime/mbarrier.go
+++ b/src/runtime/mbarrier.go
@@ -328,6 +328,13 @@ func typedslicecopy(typ *_type, dst, src slice) int {
 	dstp := unsafe.Pointer(dst.array)
 	srcp := unsafe.Pointer(src.array)
 
+	if raceenabled {
+		callerpc := getcallerpc(unsafe.Pointer(&typ))
+		pc := funcPC(slicecopy)
+		racewriterangepc(dstp, uintptr(n)*typ.size, callerpc, pc)
+		racereadrangepc(srcp, uintptr(n)*typ.size, callerpc, pc)
+	}
+
 	if !needwb() {
 		memmove(dstp, srcp, uintptr(n)*typ.size)
 		return int(n)
diff --git a/src/runtime/mem_linux.go b/src/runtime/mem_linux.go
index aa99b762bf..920fbcf6d6 100644
--- a/src/runtime/mem_linux.go
+++ b/src/runtime/mem_linux.go
@@ -54,7 +54,6 @@ func sysAlloc(n uintptr, stat *uint64) unsafe.Pointer {
 	if uintptr(p) < 4096 {
 		if uintptr(p) == _EACCES {
 			print("runtime: mmap: access denied\n")
-			print("if you're running SELinux, enable execmem for this process.\n")
 			exit(2)
 		}
 		if uintptr(p) == _EAGAIN {
diff --git a/src/runtime/mem_plan9.go b/src/runtime/mem_plan9.go
index 477a52700e..6ceed25d87 100644
--- a/src/runtime/mem_plan9.go
+++ b/src/runtime/mem_plan9.go
@@ -48,6 +48,7 @@ func sysFree(v unsafe.Pointer, n uintptr, stat *uint64) {
 	n = memRound(n)
 	if bloc == uintptr(v)+n {
 		bloc -= n
+		memclr(unsafe.Pointer(bloc), n)
 	}
 	unlock(&memlock)
 }
diff --git a/src/runtime/mem_windows.go b/src/runtime/mem_windows.go
index d72d49b975..a800ccae1d 100644
--- a/src/runtime/mem_windows.go
+++ b/src/runtime/mem_windows.go
@@ -18,16 +18,6 @@ const (
 	_PAGE_NOACCESS  = 0x0001
 )
 
-//go:cgo_import_dynamic runtime._VirtualAlloc VirtualAlloc "kernel32.dll"
-//go:cgo_import_dynamic runtime._VirtualFree VirtualFree "kernel32.dll"
-//go:cgo_import_dynamic runtime._VirtualProtect VirtualProtect "kernel32.dll"
-
-var (
-	_VirtualAlloc,
-	_VirtualFree,
-	_VirtualProtect stdFunction
-)
-
 //go:nosplit
 func sysAlloc(n uintptr, stat *uint64) unsafe.Pointer {
 	xadd64(stat, int64(n))
diff --git a/src/runtime/mfinal.go b/src/runtime/mfinal.go
index 28afa0dfab..525aa0955a 100644
--- a/src/runtime/mfinal.go
+++ b/src/runtime/mfinal.go
@@ -102,7 +102,10 @@ func wakefing() *g {
 	return res
 }
 
-var fingCreate uint32
+var (
+	fingCreate  uint32
+	fingRunning bool
+)
 
 func createfing() {
 	// start the finalizer goroutine exactly once
@@ -126,9 +129,7 @@ func runfinq() {
 			gp := getg()
 			fing = gp
 			fingwait = true
-			gp.issystem = true
 			goparkunlock(&finlock, "finalizer wait", traceEvGoBlock)
-			gp.issystem = false
 			continue
 		}
 		unlock(&finlock)
@@ -169,7 +170,9 @@ func runfinq() {
 				default:
 					throw("bad kind in runfinq")
 				}
+				fingRunning = true
 				reflectcall(nil, unsafe.Pointer(f.fn), frame, uint32(framesz), uint32(framesz))
+				fingRunning = false
 
 				// drop finalizer queue references to finalized object
 				f.fn = nil
diff --git a/src/runtime/mgc.go b/src/runtime/mgc.go
index 75b1e52916..1a3e70fcdd 100644
--- a/src/runtime/mgc.go
+++ b/src/runtime/mgc.go
@@ -122,8 +122,6 @@ const (
 	_DebugGC         = 0
 	_DebugGCPtrs     = false // if true, print trace of every pointer load during GC
 	_ConcurrentSweep = true
-
-	_WorkbufSize     = 4 * 256
 	_FinBlockSize    = 4 * 1024
 	_RootData        = 0
 	_RootBss         = 1
@@ -154,12 +152,6 @@ var gcpercent int32
 //
 var worldsema uint32 = 1
 
-type workbuf struct {
-	node lfnode // must be first
-	nobj uintptr
-	obj  [(_WorkbufSize - unsafe.Sizeof(lfnode{}) - ptrSize) / ptrSize]uintptr
-}
-
 var data, edata, bss, ebss, gcdata, gcbss struct{}
 
 var gcdatamask bitvector
@@ -347,7 +339,8 @@ func greyobject(obj, base, off uintptr, hbits heapBits, wbuf *workbuf) *workbuf
 
 	// If workbuf is full, obtain an empty one.
 	if wbuf.nobj >= uintptr(len(wbuf.obj)) {
-		wbuf = getempty(wbuf)
+		putfull(wbuf, 358)
+		wbuf = getempty(359)
 	}
 
 	wbuf.obj[wbuf.nobj] = obj
@@ -426,13 +419,10 @@ func scanobject(b, n uintptr, ptrmask *uint8, wbuf *workbuf) *workbuf {
 	return wbuf
 }
 
-// scanblock starts by scanning b as scanobject would.
-// If the gcphase is GCscan, that's all scanblock does.
-// Otherwise it traverses some fraction of the pointers it found in b, recursively.
-// As a special case, scanblock(nil, 0, nil) means to scan previously queued work,
-// stopping only when no work is left in the system.
+// scanblock scans b as scanobject would.
+// If the gcphase is GCscan, scanblock performs additional checks.
 //go:nowritebarrier
-func scanblock(b0, n0 uintptr, ptrmask *uint8) {
+func scanblock(b0, n0 uintptr, ptrmask *uint8, wbuf *workbuf) *workbuf {
 	// Use local copies of original parameters, so that a stack trace
 	// due to one of the throws below shows the original block
 	// base and extent.
@@ -443,48 +433,40 @@ func scanblock(b0, n0 uintptr, ptrmask *uint8) {
 	// 1. nil - obtain pointer mask from GC bitmap.
 	// 2. pointer to a compact mask (for stacks and data).
 
-	wbuf := getpartialorempty()
-	if b != 0 {
-		wbuf = scanobject(b, n, ptrmask, wbuf)
-		if gcphase == _GCscan {
-			if inheap(b) && ptrmask == nil {
-				// b is in heap, we are in GCscan so there should be a ptrmask.
-				throw("scanblock: In GCscan phase and inheap is true.")
-			}
-			// GCscan only goes one level deep since mark wb not turned on.
-			putpartial(wbuf)
-			return
+	if wbuf == nil {
+		wbuf = getpartialorempty(460) // no wbuf passed in.
+	}
+	wbuf = scanobject(b, n, ptrmask, wbuf)
+	if gcphase == _GCscan {
+		if inheap(b) && ptrmask == nil {
+			// b is in heap, we are in GCscan so there should be a ptrmask.
+			throw("scanblock: In GCscan phase and inheap is true.")
 		}
 	}
-
-	drainallwbufs := b == 0
-	drainworkbuf(wbuf, drainallwbufs)
+	return wbuf
 }
 
-// Scan objects in wbuf until wbuf is empty.
-// If drainallwbufs is true find all other available workbufs and repeat the process.
+// gcDrain scans objects in work buffers (starting with wbuf), blackening grey
+// objects until all work buffers have been drained.
 //go:nowritebarrier
-func drainworkbuf(wbuf *workbuf, drainallwbufs bool) {
+func gcDrain(wbuf *workbuf) {
+	if wbuf == nil {
+		wbuf = getpartialorempty(472)
+	}
+	checknocurrentwbuf()
 	if gcphase != _GCmark && gcphase != _GCmarktermination {
-		println("gcphase", gcphase)
-		throw("scanblock phase")
+		throw("scanblock phase incorrect")
 	}
 
 	for {
 		if wbuf.nobj == 0 {
-			if !drainallwbufs {
-				putempty(wbuf)
-				return
-			}
+			putempty(wbuf, 496)
 			// Refill workbuf from global queue.
-			wbuf = getfull(wbuf)
+			wbuf = getfull(504)
 			if wbuf == nil { // nil means out of work barrier reached
-				return
-			}
-
-			if wbuf.nobj <= 0 {
-				throw("runtime:scanblock getfull returns empty buffer")
+				break
 			}
+			wbuf.checknonempty()
 		}
 
 		// If another proc wants a pointer, give it some.
@@ -493,52 +475,61 @@ func drainworkbuf(wbuf *workbuf, drainallwbufs bool) {
 		}
 
 		// This might be a good place to add prefetch code...
-		// if(wbuf->nobj > 4) {
-		//         PREFETCH(wbuf->obj[wbuf->nobj - 3];
+		// if(wbuf.nobj > 4) {
+		//         PREFETCH(wbuf->obj[wbuf.nobj - 3];
 		//  }
 		wbuf.nobj--
 		b := wbuf.obj[wbuf.nobj]
+		// If the current wbuf is filled by the scan a new wbuf might be
+		// returned that could possibly hold only a single object. This
+		// could result in each iteration draining only a single object
+		// out of the wbuf passed in + a single object placed
+		// into an empty wbuf in scanobject so there could be
+		// a performance hit as we keep fetching fresh wbufs.
 		wbuf = scanobject(b, 0, nil, wbuf)
 	}
+	checknocurrentwbuf()
 }
 
-// Scan count objects starting with those in wbuf.
+// gcDrainN scans n objects starting with those in wbuf, blackening
+// grey objects.
 //go:nowritebarrier
-func drainobjects(wbuf *workbuf, count uintptr) {
-	for i := uintptr(0); i < count; i++ {
+func gcDrainN(wbuf *workbuf, n uintptr) *workbuf {
+	checknocurrentwbuf()
+	for i := uintptr(0); i < n; i++ {
 		if wbuf.nobj == 0 {
-			putempty(wbuf)
-			wbuf = trygetfull()
+			putempty(wbuf, 544)
+			wbuf = trygetfull(545)
 			if wbuf == nil {
-				return
+				return nil
 			}
 		}
 
 		// This might be a good place to add prefetch code...
-		// if(wbuf->nobj > 4) {
-		//         PREFETCH(wbuf->obj[wbuf->nobj - 3];
+		// if(wbuf.nobj > 4) {
+		//         PREFETCH(wbuf->obj[wbuf.nobj - 3];
 		//  }
 		wbuf.nobj--
 		b := wbuf.obj[wbuf.nobj]
 		wbuf = scanobject(b, 0, nil, wbuf)
 	}
-	putpartial(wbuf)
-	return
+	return wbuf
 }
 
 //go:nowritebarrier
 func markroot(desc *parfor, i uint32) {
 	// Note: if you add a case here, please also update heapdump.c:dumproots.
+	wbuf := (*workbuf)(unsafe.Pointer(xchguintptr(&getg().m.currentwbuf, 0)))
 	switch i {
 	case _RootData:
-		scanblock(uintptr(unsafe.Pointer(&data)), uintptr(unsafe.Pointer(&edata))-uintptr(unsafe.Pointer(&data)), gcdatamask.bytedata)
+		wbuf = scanblock(uintptr(unsafe.Pointer(&data)), uintptr(unsafe.Pointer(&edata))-uintptr(unsafe.Pointer(&data)), gcdatamask.bytedata, wbuf)
 
 	case _RootBss:
-		scanblock(uintptr(unsafe.Pointer(&bss)), uintptr(unsafe.Pointer(&ebss))-uintptr(unsafe.Pointer(&bss)), gcbssmask.bytedata)
+		wbuf = scanblock(uintptr(unsafe.Pointer(&bss)), uintptr(unsafe.Pointer(&ebss))-uintptr(unsafe.Pointer(&bss)), gcbssmask.bytedata, wbuf)
 
 	case _RootFinalizers:
 		for fb := allfin; fb != nil; fb = fb.alllink {
-			scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), uintptr(fb.cnt)*unsafe.Sizeof(fb.fin[0]), &finptrmask[0])
+			wbuf = scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), uintptr(fb.cnt)*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], wbuf)
 		}
 
 	case _RootSpans:
@@ -564,9 +555,9 @@ func markroot(desc *parfor, i uint32) {
 				// A finalizer can be set for an inner byte of an object, find object beginning.
 				p := uintptr(s.start<<_PageShift) + uintptr(spf.special.offset)/s.elemsize*s.elemsize
 				if gcphase != _GCscan {
-					scanblock(p, s.elemsize, nil) // scanned during mark phase
+					wbuf = scanblock(p, s.elemsize, nil, wbuf) // scanned during mark phase
 				}
-				scanblock(uintptr(unsafe.Pointer(&spf.fn)), ptrSize, &oneptr[0])
+				wbuf = scanblock(uintptr(unsafe.Pointer(&spf.fn)), ptrSize, &oneptr[0], wbuf)
 			}
 		}
 
@@ -626,157 +617,14 @@ func markroot(desc *parfor, i uint32) {
 			restartg(gp)
 		}
 	}
-}
-
-// Get an empty work buffer off the work.empty list,
-// allocating new buffers as needed.
-//go:nowritebarrier
-func getempty(b *workbuf) *workbuf {
-	if b != nil {
-		putfull(b)
-		b = nil
-	}
-	if work.empty != 0 {
-		b = (*workbuf)(lfstackpop(&work.empty))
-	}
-	if b != nil && b.nobj != 0 {
-		_g_ := getg()
-		print("m", _g_.m.id, ": getempty: popped b=", b, " with non-zero b.nobj=", b.nobj, "\n")
-		throw("getempty: workbuffer not empty, b->nobj not 0")
-	}
-	if b == nil {
-		b = (*workbuf)(persistentalloc(unsafe.Sizeof(*b), _CacheLineSize, &memstats.gc_sys))
-		b.nobj = 0
-	}
-	return b
-}
-
-//go:nowritebarrier
-func putempty(b *workbuf) {
-	if b.nobj != 0 {
-		throw("putempty: b->nobj not 0")
-	}
-	lfstackpush(&work.empty, &b.node)
-}
-
-//go:nowritebarrier
-func putfull(b *workbuf) {
-	if b.nobj <= 0 {
-		throw("putfull: b->nobj <= 0")
-	}
-	lfstackpush(&work.full, &b.node)
-}
-
-// Get an partially empty work buffer
-// if none are available get an empty one.
-//go:nowritebarrier
-func getpartialorempty() *workbuf {
-	b := (*workbuf)(lfstackpop(&work.partial))
-	if b == nil {
-		b = getempty(nil)
-	}
-	return b
-}
-
-//go:nowritebarrier
-func putpartial(b *workbuf) {
-	if b.nobj == 0 {
-		lfstackpush(&work.empty, &b.node)
-	} else if b.nobj < uintptr(len(b.obj)) {
-		lfstackpush(&work.partial, &b.node)
-	} else if b.nobj == uintptr(len(b.obj)) {
-		lfstackpush(&work.full, &b.node)
-	} else {
-		print("b=", b, " b.nobj=", b.nobj, " len(b.obj)=", len(b.obj), "\n")
-		throw("putpartial: bad Workbuf b.nobj")
-	}
-}
-
-// trygetfull tries to get a full or partially empty workbuffer.
-// if one is not immediately available return nil
-//go:nowritebarrier
-func trygetfull() *workbuf {
-	wbuf := (*workbuf)(lfstackpop(&work.full))
 	if wbuf == nil {
-		wbuf = (*workbuf)(lfstackpop(&work.partial))
-	}
-	return wbuf
-}
-
-// Get a full work buffer off the work.full or a partially
-// filled one off the work.partial list. If nothing is available
-// wait until all the other gc helpers have finished and then
-// return nil.
-// getfull acts as a barrier for work.nproc helpers. As long as one
-// gchelper is actively marking objects it
-// may create a workbuffer that the other helpers can work on.
-// The for loop either exits when a work buffer is found
-// or when _all_ of the work.nproc GC helpers are in the loop
-// looking for work and thus not capable of creating new work.
-// This is in fact the termination condition for the STW mark
-// phase.
-//go:nowritebarrier
-func getfull(b *workbuf) *workbuf {
-	if b != nil {
-		putempty(b)
-	}
-
-	b = (*workbuf)(lfstackpop(&work.full))
-	if b == nil {
-		b = (*workbuf)(lfstackpop(&work.partial))
-	}
-	if b != nil {
-		return b
-	}
-
-	xadd(&work.nwait, +1)
-	for i := 0; ; i++ {
-		if work.full != 0 {
-			xadd(&work.nwait, -1)
-			b = (*workbuf)(lfstackpop(&work.full))
-			if b == nil {
-				b = (*workbuf)(lfstackpop(&work.partial))
-			}
-			if b != nil {
-				return b
-			}
-			xadd(&work.nwait, +1)
-		}
-		if work.nwait == work.nproc {
-			return nil
-		}
-		_g_ := getg()
-		if i < 10 {
-			_g_.m.gcstats.nprocyield++
-			procyield(20)
-		} else if i < 20 {
-			_g_.m.gcstats.nosyield++
-			osyield()
-		} else {
-			_g_.m.gcstats.nsleep++
-			usleep(100)
-		}
+		return
+	} else {
+		putpartial(wbuf, 670)
 	}
 }
 
 //go:nowritebarrier
-func handoff(b *workbuf) *workbuf {
-	// Make new buffer with half of b's pointers.
-	b1 := getempty(nil)
-	n := b.nobj / 2
-	b.nobj -= n
-	b1.nobj = n
-	memmove(unsafe.Pointer(&b1.obj[0]), unsafe.Pointer(&b.obj[b.nobj]), n*unsafe.Sizeof(b1.obj[0]))
-	_g_ := getg()
-	_g_.m.gcstats.nhandoff++
-	_g_.m.gcstats.nhandoffcnt += uint64(n)
-
-	// Put b on full list - let first half of b get stolen.
-	lfstackpush(&work.full, &b.node)
-	return b1
-}
-
-//go:nowritebarrier
 func stackmapdata(stkmap *stackmap, n int32) bitvector {
 	if n < 0 || n >= stkmap.n {
 		throw("stackmapdata: index out of range")
@@ -786,13 +634,13 @@ func stackmapdata(stkmap *stackmap, n int32) bitvector {
 
 // Scan a stack frame: local variables and function arguments/results.
 //go:nowritebarrier
-func scanframe(frame *stkframe, unused unsafe.Pointer) bool {
+func scanframeworker(frame *stkframe, unused unsafe.Pointer, wbuf *workbuf) *workbuf {
 
 	f := frame.fn
 	targetpc := frame.continpc
 	if targetpc == 0 {
 		// Frame is dead.
-		return true
+		return wbuf
 	}
 	if _DebugGC > 1 {
 		print("scanframe ", funcname(f), "\n")
@@ -831,7 +679,7 @@ func scanframe(frame *stkframe, unused unsafe.Pointer) bool {
 		}
 		bv := stackmapdata(stkmap, pcdata)
 		size = (uintptr(bv.n) / typeBitsWidth) * ptrSize
-		scanblock(frame.varp-size, size, bv.bytedata)
+		wbuf = scanblock(frame.varp-size, size, bv.bytedata, wbuf)
 	}
 
 	// Scan arguments.
@@ -852,9 +700,9 @@ func scanframe(frame *stkframe, unused unsafe.Pointer) bool {
 			}
 			bv = stackmapdata(stkmap, pcdata)
 		}
-		scanblock(frame.argp, uintptr(bv.n)/typeBitsWidth*ptrSize, bv.bytedata)
+		wbuf = scanblock(frame.argp, uintptr(bv.n)/typeBitsWidth*ptrSize, bv.bytedata, wbuf)
 	}
-	return true
+	return wbuf
 }
 
 //go:nowritebarrier
@@ -889,8 +737,19 @@ func scanstack(gp *g) {
 		throw("can't scan gchelper stack")
 	}
 
+	wbuf := (*workbuf)(unsafe.Pointer(xchguintptr(&getg().m.currentwbuf, 0)))
+	scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
+		// Pick up wbuf as free variable so gentraceback and friends can
+		// keep the same signature.
+		wbuf = scanframeworker(frame, unused, wbuf)
+		return true
+	}
 	gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
 	tracebackdefers(gp, scanframe, nil)
+	wbuf = (*workbuf)(unsafe.Pointer(xchguintptr(&getg().m.currentwbuf, uintptr(unsafe.Pointer(wbuf)))))
+	if wbuf != nil {
+		throw("wbuf not nil after stack scans")
+	}
 	gp.gcscanvalid = true
 }
 
@@ -898,17 +757,34 @@ func scanstack(gp *g) {
 // The object is not nil and known to be in the heap.
 //go:nowritebarrier
 func shade(b uintptr) {
+	var wbuf *workbuf
+
 	if !inheap(b) {
 		throw("shade: passed an address not in the heap")
 	}
-
-	wbuf := getpartialorempty()
-
 	if obj, hbits := heapBitsForObject(b); obj != 0 {
-		wbuf = greyobject(obj, 0, 0, hbits, wbuf)
-	}
+		// TODO: this would be a great place to put a check to see
+		// if we are harvesting and if we are then we should
+		// figure out why there is a call to shade when the
+		// harvester thinks we are in a STW.
+		// if atomicload(&harvestingwbufs) == uint32(1) {
+		//	// Throw here to discover write barriers
+		//	// being executed during a STW.
+		// }
 
-	putpartial(wbuf)
+		wbuf = getpartialorempty(1181)
+		wbuf := greyobject(obj, 0, 0, hbits, wbuf)
+		checknocurrentwbuf()
+		// This is part of the write barrier so put the wbuf back.
+		if gcphase == _GCmarktermination {
+			putpartial(wbuf, 1191) // Put on full???
+		} else {
+			wbuf = (*workbuf)(unsafe.Pointer(xchguintptr(&getg().m.currentwbuf, uintptr(unsafe.Pointer(wbuf)))))
+			if wbuf != nil {
+				throw("m.currentwbuf lost in shade")
+			}
+		}
+	}
 }
 
 // gchelpwork does a small bounded amount of gc work. The purpose is to
@@ -931,10 +807,24 @@ func gchelpwork() {
 		// scanstack(gp)
 	case _GCmark:
 		// Get a full work buffer and empty it.
-		var wbuf *workbuf
-		wbuf = trygetfull()
+		m := getg().m
+		// drain your own currentwbuf first in the hopes that it will
+		// be more cache friendly.
+		wbuf := (*workbuf)(unsafe.Pointer(xchguintptr(&m.currentwbuf, 0)))
+		//		wbuf := (*workbuf)(unsafe.Pointer(m.currentwbuf))
+		//		m.currentwbuf = 0
+		if wbuf == nil {
+			wbuf = trygetfull(1228)
+		}
 		if wbuf != nil {
-			drainobjects(wbuf, uintptr(len(wbuf.obj))) // drain upto one buffer's worth of objects
+			wbuf = gcDrainN(wbuf, uintptr(len(wbuf.obj))) // drain upto one buffer's worth of objects
+			if wbuf != nil {
+				if wbuf.nobj != 0 {
+					putfull(wbuf, 1175)
+				} else {
+					putempty(wbuf, 1177)
+				}
+			}
 		}
 	case _GCmarktermination:
 		// We should never be here since the world is stopped.
@@ -1249,7 +1139,7 @@ func gchelper() {
 	// parallel mark for over GC roots
 	parfordo(work.markfor)
 	if gcphase != _GCscan {
-		scanblock(0, 0, nil) // blocks in getfull
+		gcDrain(nil) // blocks in getfull
 	}
 
 	if trace.enabled {
@@ -1380,7 +1270,7 @@ var heapminimum = uint64(4 << 20)
 
 func gcinit() {
 	if unsafe.Sizeof(workbuf{}) != _WorkbufSize {
-		throw("runtime: size of Workbuf is suboptimal")
+		throw("size of Workbuf is suboptimal")
 	}
 
 	work.markfor = parforalloc(_MaxGcproc)
@@ -1508,16 +1398,12 @@ func gcscan_m() {
 }
 
 // Mark all objects that are known about.
+// This is the concurrent mark phase.
 //go:nowritebarrier
 func gcmark_m() {
-	scanblock(0, 0, nil)
-}
-
-// For now this must be bracketed with a stoptheworld and a starttheworld to ensure
-// all go routines see the new barrier.
-//go:nowritebarrier
-func gcinstallmarkwb_m() {
-	gcphase = _GCmark
+	gcDrain(nil)
+	// TODO add another harvestwbuf and reset work.nwait=0, work.ndone=0, and work.nproc=1
+	// and repeat the above gcDrain.
 }
 
 // For now this must be bracketed with a stoptheworld and a starttheworld to ensure
@@ -1527,12 +1413,14 @@ func gcinstalloffwb_m() {
 	gcphase = _GCoff
 }
 
+// STW is in effect at this point.
 //TODO go:nowritebarrier
 func gc(start_time int64, eagersweep bool) {
 	if _DebugGCPtrs {
 		print("GC start\n")
 	}
 
+	gcphase = _GCmarktermination
 	if debug.allocfreetrace > 0 {
 		tracegc()
 	}
@@ -1571,12 +1459,10 @@ func gc(start_time int64, eagersweep bool) {
 	mheap_.gcspans = mheap_.allspans
 	work.spans = h_allspans
 	unlock(&mheap_.lock)
-	oldphase := gcphase
 
 	work.nwait = 0
 	work.ndone = 0
 	work.nproc = uint32(gcprocs())
-	gcphase = _GCmarktermination
 
 	// World is stopped so allglen will not change.
 	for i := uintptr(0); i < allglen; i++ {
@@ -1599,9 +1485,10 @@ func gc(start_time int64, eagersweep bool) {
 		t2 = nanotime()
 	}
 
+	harvestwbufs() // move local workbufs onto global queues where the GC can find them
 	gchelperstart()
 	parfordo(work.markfor)
-	scanblock(0, 0, nil)
+	gcDrain(nil)
 
 	if work.full != 0 {
 		throw("work.full != 0")
@@ -1610,7 +1497,7 @@ func gc(start_time int64, eagersweep bool) {
 		throw("work.partial != 0")
 	}
 
-	gcphase = oldphase
+	gcphase = _GCoff
 	var t3 int64
 	if debug.gctrace > 0 {
 		t3 = nanotime()
@@ -1755,7 +1642,7 @@ func readGCStats_m(pauses *[]uint64) {
 	p := *pauses
 	// Calling code in runtime/debug should make the slice large enough.
 	if cap(p) < len(memstats.pause_ns)+3 {
-		throw("runtime: short slice passed to readGCStats")
+		throw("short slice passed to readGCStats")
 	}
 
 	// Pass back: pauses, pause ends, last gc (absolute time), number of gc, total pause ns.
diff --git a/src/runtime/mgc0.go b/src/runtime/mgc0.go
index bbd786d519..f54d93377d 100644
--- a/src/runtime/mgc0.go
+++ b/src/runtime/mgc0.go
@@ -62,7 +62,6 @@ func clearpools() {
 // bggc holds the state of the backgroundgc.
 func backgroundgc() {
 	bggc.g = getg()
-	bggc.g.issystem = true
 	for {
 		gcwork(0)
 		lock(&bggc.lock)
@@ -73,7 +72,6 @@ func backgroundgc() {
 
 func bgsweep() {
 	sweep.g = getg()
-	getg().issystem = true
 	for {
 		for gosweepone() != ^uintptr(0) {
 			sweep.nbgsweep++
diff --git a/src/runtime/mheap.go b/src/runtime/mheap.go
index 11bc809ec1..13f1b363d0 100644
--- a/src/runtime/mheap.go
+++ b/src/runtime/mheap.go
@@ -318,8 +318,7 @@ HaveSpan:
 		t.needzero = s.needzero
 		s.state = _MSpanStack // prevent coalescing with s
 		t.state = _MSpanStack
-		mHeap_FreeSpanLocked(h, t, false, false)
-		t.unusedsince = s.unusedsince // preserve age (TODO: wrong: t is possibly merged and/or deallocated at this point)
+		mHeap_FreeSpanLocked(h, t, false, false, s.unusedsince)
 		s.state = _MSpanFree
 	}
 	s.unusedsince = 0
@@ -395,7 +394,7 @@ func mHeap_Grow(h *mheap, npage uintptr) bool {
 	h_spans[p+s.npages-1] = s
 	atomicstore(&s.sweepgen, h.sweepgen)
 	s.state = _MSpanInUse
-	mHeap_FreeSpanLocked(h, s, false, true)
+	mHeap_FreeSpanLocked(h, s, false, true, 0)
 	return true
 }
 
@@ -442,7 +441,7 @@ func mHeap_Free(h *mheap, s *mspan, acct int32) {
 			memstats.heap_alloc -= uint64(s.npages << _PageShift)
 			memstats.heap_objects--
 		}
-		mHeap_FreeSpanLocked(h, s, true, true)
+		mHeap_FreeSpanLocked(h, s, true, true, 0)
 		if trace.enabled {
 			traceHeapAlloc()
 		}
@@ -458,11 +457,11 @@ func mHeap_FreeStack(h *mheap, s *mspan) {
 	s.needzero = 1
 	lock(&h.lock)
 	memstats.stacks_inuse -= uint64(s.npages << _PageShift)
-	mHeap_FreeSpanLocked(h, s, true, true)
+	mHeap_FreeSpanLocked(h, s, true, true, 0)
 	unlock(&h.lock)
 }
 
-func mHeap_FreeSpanLocked(h *mheap, s *mspan, acctinuse, acctidle bool) {
+func mHeap_FreeSpanLocked(h *mheap, s *mspan, acctinuse, acctidle bool, unusedsince int64) {
 	switch s.state {
 	case _MSpanStack:
 		if s.ref != 0 {
@@ -488,7 +487,10 @@ func mHeap_FreeSpanLocked(h *mheap, s *mspan, acctinuse, acctidle bool) {
 
 	// Stamp newly unused spans. The scavenger will use that
 	// info to potentially give back some pages to the OS.
-	s.unusedsince = nanotime()
+	s.unusedsince = unusedsince
+	if unusedsince == 0 {
+		s.unusedsince = nanotime()
+	}
 	s.npreleased = 0
 
 	// Coalesce with earlier, later spans.
diff --git a/src/runtime/netpoll_windows.go b/src/runtime/netpoll_windows.go
index 8e0750d607..0861e20f0c 100644
--- a/src/runtime/netpoll_windows.go
+++ b/src/runtime/netpoll_windows.go
@@ -10,16 +10,6 @@ import (
 
 const _DWORD_MAX = 0xffffffff
 
-//go:cgo_import_dynamic runtime._CreateIoCompletionPort CreateIoCompletionPort "kernel32.dll"
-//go:cgo_import_dynamic runtime._GetQueuedCompletionStatus GetQueuedCompletionStatus "kernel32.dll"
-//go:cgo_import_dynamic runtime._WSAGetOverlappedResult WSAGetOverlappedResult "ws2_32.dll"
-
-var (
-	_CreateIoCompletionPort,
-	_GetQueuedCompletionStatus,
-	_WSAGetOverlappedResult stdFunction
-)
-
 const _INVALID_HANDLE_VALUE = ^uintptr(0)
 
 // net_op must be the same as beginning of net.operation. Keep these in sync.
diff --git a/src/runtime/os1_windows.go b/src/runtime/os1_windows.go
index 8655c083b2..2de6b09343 100644
--- a/src/runtime/os1_windows.go
+++ b/src/runtime/os1_windows.go
@@ -11,6 +11,7 @@ import (
 //go:cgo_import_dynamic runtime._AddVectoredExceptionHandler AddVectoredExceptionHandler "kernel32.dll"
 //go:cgo_import_dynamic runtime._CloseHandle CloseHandle "kernel32.dll"
 //go:cgo_import_dynamic runtime._CreateEventA CreateEventA "kernel32.dll"
+//go:cgo_import_dynamic runtime._CreateIoCompletionPort CreateIoCompletionPort "kernel32.dll"
 //go:cgo_import_dynamic runtime._CreateThread CreateThread "kernel32.dll"
 //go:cgo_import_dynamic runtime._CreateWaitableTimerA CreateWaitableTimerA "kernel32.dll"
 //go:cgo_import_dynamic runtime._CryptAcquireContextW CryptAcquireContextW "advapi32.dll"
@@ -21,6 +22,7 @@ import (
 //go:cgo_import_dynamic runtime._FreeEnvironmentStringsW FreeEnvironmentStringsW "kernel32.dll"
 //go:cgo_import_dynamic runtime._GetEnvironmentStringsW GetEnvironmentStringsW "kernel32.dll"
 //go:cgo_import_dynamic runtime._GetProcAddress GetProcAddress "kernel32.dll"
+//go:cgo_import_dynamic runtime._GetQueuedCompletionStatus GetQueuedCompletionStatus "kernel32.dll"
 //go:cgo_import_dynamic runtime._GetStdHandle GetStdHandle "kernel32.dll"
 //go:cgo_import_dynamic runtime._GetSystemInfo GetSystemInfo "kernel32.dll"
 //go:cgo_import_dynamic runtime._GetThreadContext GetThreadContext "kernel32.dll"
@@ -37,14 +39,22 @@ import (
 //go:cgo_import_dynamic runtime._SetWaitableTimer SetWaitableTimer "kernel32.dll"
 //go:cgo_import_dynamic runtime._Sleep Sleep "kernel32.dll"
 //go:cgo_import_dynamic runtime._SuspendThread SuspendThread "kernel32.dll"
+//go:cgo_import_dynamic runtime._VirtualAlloc VirtualAlloc "kernel32.dll"
+//go:cgo_import_dynamic runtime._VirtualFree VirtualFree "kernel32.dll"
+//go:cgo_import_dynamic runtime._VirtualProtect VirtualProtect "kernel32.dll"
+//go:cgo_import_dynamic runtime._WSAGetOverlappedResult WSAGetOverlappedResult "ws2_32.dll"
 //go:cgo_import_dynamic runtime._WaitForSingleObject WaitForSingleObject "kernel32.dll"
 //go:cgo_import_dynamic runtime._WriteFile WriteFile "kernel32.dll"
 //go:cgo_import_dynamic runtime._timeBeginPeriod timeBeginPeriod "winmm.dll"
 
 var (
+	// Following syscalls are available on every Windows PC.
+	// All these variables are set by the Windows executable
+	// loader before the Go program starts.
 	_AddVectoredExceptionHandler,
 	_CloseHandle,
 	_CreateEventA,
+	_CreateIoCompletionPort,
 	_CreateThread,
 	_CreateWaitableTimerA,
 	_CryptAcquireContextW,
@@ -55,6 +65,7 @@ var (
 	_FreeEnvironmentStringsW,
 	_GetEnvironmentStringsW,
 	_GetProcAddress,
+	_GetQueuedCompletionStatus,
 	_GetStdHandle,
 	_GetSystemInfo,
 	_GetThreadContext,
@@ -71,12 +82,36 @@ var (
 	_SetWaitableTimer,
 	_Sleep,
 	_SuspendThread,
+	_VirtualAlloc,
+	_VirtualFree,
+	_VirtualProtect,
+	_WSAGetOverlappedResult,
 	_WaitForSingleObject,
 	_WriteFile,
 	_timeBeginPeriod stdFunction
+
+	// Following syscalls are only available on some Windows PCs.
+	// We will load syscalls, if available, before using them.
+	_AddVectoredContinueHandler,
+	_GetQueuedCompletionStatusEx stdFunction
 )
 
-var _GetQueuedCompletionStatusEx stdFunction
+func loadOptionalSyscalls() {
+	var buf [50]byte // large enough for longest string
+	strtoptr := func(s string) uintptr {
+		buf[copy(buf[:], s)] = 0 // nil-terminated for OS
+		return uintptr(noescape(unsafe.Pointer(&buf[0])))
+	}
+	l := stdcall1(_LoadLibraryA, strtoptr("kernel32.dll"))
+	findfunc := func(name string) stdFunction {
+		f := stdcall2(_GetProcAddress, l, strtoptr(name))
+		return stdFunction(unsafe.Pointer(f))
+	}
+	if l != 0 {
+		_AddVectoredContinueHandler = findfunc("AddVectoredContinueHandler")
+		_GetQueuedCompletionStatusEx = findfunc("GetQueuedCompletionStatusEx")
+	}
+}
 
 // in sys_windows_386.s and sys_windows_amd64.s
 func externalthreadhandler()
@@ -117,34 +152,24 @@ func disableWER() {
 	stdcall1(_SetErrorMode, uintptr(errormode)|SEM_FAILCRITICALERRORS|SEM_NOGPFAULTERRORBOX|SEM_NOOPENFILEERRORBOX)
 }
 
-var (
-	kernel32Name                    = []byte("kernel32.dll\x00")
-	addVectoredContinueHandlerName  = []byte("AddVectoredContinueHandler\x00")
-	getQueuedCompletionStatusExName = []byte("GetQueuedCompletionStatusEx\x00")
-)
-
 func osinit() {
 	setBadSignalMsg()
 
-	kernel32 := stdcall1(_LoadLibraryA, uintptr(unsafe.Pointer(&kernel32Name[0])))
+	loadOptionalSyscalls()
 
 	disableWER()
 
 	externalthreadhandlerp = funcPC(externalthreadhandler)
 
 	stdcall2(_AddVectoredExceptionHandler, 1, funcPC(exceptiontramp))
-	addVectoredContinueHandler := uintptr(0)
-	if kernel32 != 0 {
-		addVectoredContinueHandler = stdcall2(_GetProcAddress, kernel32, uintptr(unsafe.Pointer(&addVectoredContinueHandlerName[0])))
-	}
-	if addVectoredContinueHandler == 0 || unsafe.Sizeof(&kernel32) == 4 {
+	if _AddVectoredContinueHandler == nil || unsafe.Sizeof(&_AddVectoredContinueHandler) == 4 {
 		// use SetUnhandledExceptionFilter for windows-386 or
 		// if VectoredContinueHandler is unavailable.
 		// note: SetUnhandledExceptionFilter handler won't be called, if debugging.
 		stdcall1(_SetUnhandledExceptionFilter, funcPC(lastcontinuetramp))
 	} else {
-		stdcall2(stdFunction(unsafe.Pointer(addVectoredContinueHandler)), 1, funcPC(firstcontinuetramp))
-		stdcall2(stdFunction(unsafe.Pointer(addVectoredContinueHandler)), 0, funcPC(lastcontinuetramp))
+		stdcall2(_AddVectoredContinueHandler, 1, funcPC(firstcontinuetramp))
+		stdcall2(_AddVectoredContinueHandler, 0, funcPC(lastcontinuetramp))
 	}
 
 	stdcall2(_SetConsoleCtrlHandler, funcPC(ctrlhandler), 1)
@@ -158,10 +183,6 @@ func osinit() {
 	// equivalent threads that all do a mix of GUI, IO, computations, etc.
 	// In such context dynamic priority boosting does nothing but harm, so we turn it off.
 	stdcall2(_SetProcessPriorityBoost, currentProcess, 1)
-
-	if kernel32 != 0 {
-		_GetQueuedCompletionStatusEx = stdFunction(unsafe.Pointer(stdcall2(_GetProcAddress, kernel32, uintptr(unsafe.Pointer(&getQueuedCompletionStatusExName[0])))))
-	}
 }
 
 //go:nosplit
@@ -506,7 +527,7 @@ func profilem(mp *m) {
 	r = (*context)(unsafe.Pointer((uintptr(unsafe.Pointer(&rbuf[15]))) &^ 15))
 	r.contextflags = _CONTEXT_CONTROL
 	stdcall2(_GetThreadContext, mp.thread, uintptr(unsafe.Pointer(r)))
-	dosigprof(r, gp, mp)
+	sigprof((*byte)(unsafe.Pointer(r.ip())), (*byte)(unsafe.Pointer(r.sp())), nil, gp, mp)
 }
 
 func profileloop1() {
diff --git a/src/runtime/os1_windows_386.go b/src/runtime/os1_windows_386.go
index 7b4fdfe94a..b7eae204d1 100644
--- a/src/runtime/os1_windows_386.go
+++ b/src/runtime/os1_windows_386.go
@@ -118,7 +118,3 @@ func sigenable(sig uint32) {
 
 func sigdisable(sig uint32) {
 }
-
-func dosigprof(r *context, gp *g, mp *m) {
-	sigprof((*byte)(unsafe.Pointer(uintptr(r.eip))), (*byte)(unsafe.Pointer(uintptr(r.esp))), nil, gp, mp)
-}
diff --git a/src/runtime/os1_windows_amd64.go b/src/runtime/os1_windows_amd64.go
index c211f6fd91..4163fcf23d 100644
--- a/src/runtime/os1_windows_amd64.go
+++ b/src/runtime/os1_windows_amd64.go
@@ -137,7 +137,3 @@ func sigenable(sig uint32) {
 
 func sigdisable(sig uint32) {
 }
-
-func dosigprof(r *context, gp *g, mp *m) {
-	sigprof((*byte)(unsafe.Pointer(uintptr(r.rip))), (*byte)(unsafe.Pointer(uintptr(r.rsp))), nil, gp, mp)
-}
diff --git a/src/runtime/os3_plan9.go b/src/runtime/os3_plan9.go
index 58ca0e7b2f..facaab2546 100644
--- a/src/runtime/os3_plan9.go
+++ b/src/runtime/os3_plan9.go
@@ -72,9 +72,9 @@ func sighandler(_ureg *ureg, note *byte, gp *g) int {
 		return _NCONT
 	}
 	if flags&_SigNotify != 0 {
-		// TODO(ality): See if os/signal wants it.
-		//if(sigsend(...))
-		//	return _NCONT;
+		if sendNote(note) {
+			return _NCONT
+		}
 	}
 	if flags&_SigKill != 0 {
 		goto Exit
diff --git a/src/runtime/os_darwin_arm.go b/src/runtime/os_darwin_arm.go
new file mode 100644
index 0000000000..d3336c012a
--- /dev/null
+++ b/src/runtime/os_darwin_arm.go
@@ -0,0 +1,17 @@
+// Copyright 2014 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package runtime
+
+func checkgoarm() {
+	return // TODO(minux)
+}
+
+//go:nosplit
+func cputicks() int64 {
+	// Currently cputicks() is used in blocking profiler and to seed runtime·fastrand1().
+	// runtime·nanotime() is a poor approximation of CPU ticks that is enough for the profiler.
+	// TODO: need more entropy to better seed fastrand1.
+	return nanotime()
+}
diff --git a/src/runtime/os_plan9.go b/src/runtime/os_plan9.go
index 679bf34519..870404b948 100644
--- a/src/runtime/os_plan9.go
+++ b/src/runtime/os_plan9.go
@@ -6,6 +6,11 @@ package runtime
 
 import "unsafe"
 
+func close(fd int32) int32
+
+//go:noescape
+func open(name *byte, mode, perm int32) int32
+
 //go:noescape
 func pread(fd int32, buf unsafe.Pointer, nbytes int32, offset int64) int32
 
diff --git a/src/runtime/os_windows_386.go b/src/runtime/os_windows_386.go
deleted file mode 100644
index 86a1906c0c..0000000000
--- a/src/runtime/os_windows_386.go
+++ /dev/null
@@ -1,11 +0,0 @@
-// Copyright 2014 The Go Authors.  All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-package runtime
-
-// contextPC returns the EIP (program counter) register from the context.
-func contextPC(r *context) uintptr { return uintptr(r.eip) }
-
-// contextSP returns the ESP (stack pointer) register from the context.
-func contextSP(r *context) uintptr { return uintptr(r.esp) }
diff --git a/src/runtime/os_windows_amd64.go b/src/runtime/os_windows_amd64.go
deleted file mode 100644
index 3f4d4d07cb..0000000000
--- a/src/runtime/os_windows_amd64.go
+++ /dev/null
@@ -1,11 +0,0 @@
-// Copyright 2014 The Go Authors.  All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-package runtime
-
-// contextPC returns the RIP (program counter) register from the context.
-func contextPC(r *context) uintptr { return uintptr(r.rip) }
-
-// contextSP returns the RSP (stack pointer) register from the context.
-func contextSP(r *context) uintptr { return uintptr(r.rsp) }
diff --git a/src/runtime/proc.go b/src/runtime/proc.go
index c8f6de1ac8..027416a9ec 100644
--- a/src/runtime/proc.go
+++ b/src/runtime/proc.go
@@ -32,7 +32,9 @@ func main() {
 		maxstacksize = 250000000
 	}
 
-	systemstack(newsysmon)
+	systemstack(func() {
+		newm(sysmon, nil)
+	})
 
 	// Lock the main goroutine onto this, the main OS thread,
 	// during initialization.  Most programs won't care, but a few
@@ -110,7 +112,6 @@ func init() {
 
 func forcegchelper() {
 	forcegc.g = getg()
-	forcegc.g.issystem = true
 	for {
 		lock(&forcegc.lock)
 		if forcegc.idle != 0 {
diff --git a/src/runtime/proc1.go b/src/runtime/proc1.go
index 70addbffad..471ffc83a3 100644
--- a/src/runtime/proc1.go
+++ b/src/runtime/proc1.go
@@ -81,10 +81,6 @@ func schedinit() {
 	}
 }
 
-func newsysmon() {
-	_newm(sysmon, nil)
-}
-
 func dumpgstatus(gp *g) {
 	_g_ := getg()
 	print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
@@ -638,7 +634,7 @@ func starttheworld() {
 			notewakeup(&mp.park)
 		} else {
 			// Start M to run P.  Do not start another M below.
-			_newm(nil, p)
+			newm(nil, p)
 			add = false
 		}
 	}
@@ -658,7 +654,7 @@ func starttheworld() {
 		// coordinate.  This lazy approach works out in practice:
 		// we don't mind if the first couple gc rounds don't have quite
 		// the maximum number of procs.
-		_newm(mhelpgc, nil)
+		newm(mhelpgc, nil)
 	}
 	_g_.m.locks--
 	if _g_.m.locks == 0 && _g_.preempt { // restore the preemption request in case we've cleared it in newstack
@@ -960,7 +956,7 @@ func unlockextra(mp *m) {
 }
 
 // Create a new m.  It will start off with a call to fn, or else the scheduler.
-func _newm(fn func(), _p_ *p) {
+func newm(fn func(), _p_ *p) {
 	mp := allocm(_p_)
 	mp.nextp = _p_
 	mp.mstartfn = *(*unsafe.Pointer)(unsafe.Pointer(&fn))
@@ -1037,7 +1033,7 @@ func startm(_p_ *p, spinning bool) {
 		if spinning {
 			fn = mspinning
 		}
-		_newm(fn, _p_)
+		newm(fn, _p_)
 		return
 	}
 	if mp.spinning {
@@ -2636,7 +2632,7 @@ func checkdead() {
 	lock(&allglock)
 	for i := 0; i < len(allgs); i++ {
 		gp := allgs[i]
-		if gp.issystem {
+		if isSystemGoroutine(gp) {
 			continue
 		}
 		s := readgstatus(gp)
@@ -2667,7 +2663,7 @@ func checkdead() {
 		}
 		mp := mget()
 		if mp == nil {
-			_newm(nil, _p_)
+			newm(nil, _p_)
 		} else {
 			mp.nextp = _p_
 			notewakeup(&mp.park)
diff --git a/src/runtime/proc_test.go b/src/runtime/proc_test.go
index 3b78b01ca3..88cd48486a 100644
--- a/src/runtime/proc_test.go
+++ b/src/runtime/proc_test.go
@@ -435,6 +435,18 @@ func BenchmarkCreateGoroutinesCapture(b *testing.B) {
 	}
 }
 
+func BenchmarkClosureCall(b *testing.B) {
+	sum := 0
+	off1 := 1
+	for i := 0; i < b.N; i++ {
+		off2 := 2
+		func() {
+			sum += i + off1 + off2
+		}()
+	}
+	_ = sum
+}
+
 type Matrix [][]float64
 
 func BenchmarkMatmult(b *testing.B) {
diff --git a/src/runtime/race/race_test.go b/src/runtime/race/race_test.go
index 7e0ee866a6..f4caff0ed4 100644
--- a/src/runtime/race/race_test.go
+++ b/src/runtime/race/race_test.go
@@ -152,7 +152,7 @@ func runTests() ([]byte, error) {
 		}
 		cmd.Env = append(cmd.Env, env)
 	}
-	cmd.Env = append(cmd.Env, `GORACE="suppress_equal_stacks=0 suppress_equal_addresses=0 exitcode=0"`)
+	cmd.Env = append(cmd.Env, `GORACE=suppress_equal_stacks=0 suppress_equal_addresses=0 exitcode=0`)
 	return cmd.CombinedOutput()
 }
 
diff --git a/src/runtime/race/testdata/slice_test.go b/src/runtime/race/testdata/slice_test.go
index 5702d1ac85..32ae878970 100644
--- a/src/runtime/race/testdata/slice_test.go
+++ b/src/runtime/race/testdata/slice_test.go
@@ -144,6 +144,54 @@ func TestNoRaceSliceCopyRead(t *testing.T) {
 	<-ch
 }
 
+func TestRacePointerSliceCopyRead(t *testing.T) {
+	ch := make(chan bool, 1)
+	a := make([]*int, 10)
+	b := make([]*int, 10)
+	go func() {
+		_ = a[5]
+		ch <- true
+	}()
+	copy(a, b)
+	<-ch
+}
+
+func TestNoRacePointerSliceWriteCopy(t *testing.T) {
+	ch := make(chan bool, 1)
+	a := make([]*int, 10)
+	b := make([]*int, 10)
+	go func() {
+		a[5] = new(int)
+		ch <- true
+	}()
+	copy(a[:5], b[:5])
+	<-ch
+}
+
+func TestRacePointerSliceCopyWrite2(t *testing.T) {
+	ch := make(chan bool, 1)
+	a := make([]*int, 10)
+	b := make([]*int, 10)
+	go func() {
+		b[5] = new(int)
+		ch <- true
+	}()
+	copy(a, b)
+	<-ch
+}
+
+func TestNoRacePointerSliceCopyRead(t *testing.T) {
+	ch := make(chan bool, 1)
+	a := make([]*int, 10)
+	b := make([]*int, 10)
+	go func() {
+		_ = b[5]
+		ch <- true
+	}()
+	copy(a, b)
+	<-ch
+}
+
 func TestNoRaceSliceWriteSlice2(t *testing.T) {
 	ch := make(chan bool, 1)
 	a := make([]float64, 10)
@@ -395,6 +443,53 @@ func TestRaceSliceAppendString(t *testing.T) {
 	<-c
 }
 
+func TestRacePointerSliceAppend(t *testing.T) {
+	c := make(chan bool, 1)
+	s := make([]*int, 10, 20)
+	go func() {
+		_ = append(s, new(int))
+		c <- true
+	}()
+	_ = append(s, new(int))
+	<-c
+}
+
+func TestRacePointerSliceAppendWrite(t *testing.T) {
+	c := make(chan bool, 1)
+	s := make([]*int, 10)
+	go func() {
+		_ = append(s, new(int))
+		c <- true
+	}()
+	s[0] = new(int)
+	<-c
+}
+
+func TestRacePointerSliceAppendSlice(t *testing.T) {
+	c := make(chan bool, 1)
+	s := make([]*int, 10)
+	go func() {
+		s2 := make([]*int, 10)
+		_ = append(s, s2...)
+		c <- true
+	}()
+	s[0] = new(int)
+	<-c
+}
+
+func TestRacePointerSliceAppendSlice2(t *testing.T) {
+	c := make(chan bool, 1)
+	s := make([]*int, 10)
+	s2foobar := make([]*int, 10)
+	go func() {
+		_ = append(s, s2foobar...)
+		c <- true
+	}()
+	println("WRITE:", &s2foobar[5])
+	s2foobar[5] = nil
+	<-c
+}
+
 func TestNoRaceSliceIndexAccess(t *testing.T) {
 	c := make(chan bool, 1)
 	s := make([]int, 10)
diff --git a/src/runtime/race_amd64.s b/src/runtime/race_amd64.s
index 972cbe3f8a..267cd6cec4 100644
--- a/src/runtime/race_amd64.s
+++ b/src/runtime/race_amd64.s
@@ -383,8 +383,9 @@ TEXT	runtime·racesymbolizethunk(SB), NOSPLIT, $56-8
 	MOVQ	g_m(R13), R13
 	MOVQ	m_g0(R13), R14
 	MOVQ	R14, g(R12)	// g = m->g0
-	MOVQ	RARG0, 0(SP)	// func arg
+	PUSHQ	RARG0	// func arg
 	CALL	runtime·racesymbolize(SB)
+	POPQ	R12
 	// All registers are smashed after Go code, reload.
 	get_tls(R12)
 	MOVQ	g(R12), R13
diff --git a/src/runtime/rt0_darwin_arm.s b/src/runtime/rt0_darwin_arm.s
new file mode 100644
index 0000000000..4d31e3a78a
--- /dev/null
+++ b/src/runtime/rt0_darwin_arm.s
@@ -0,0 +1,18 @@
+// Copyright 2014 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#include "textflag.h"
+
+TEXT _rt0_arm_darwin(SB),7,$-4
+	// prepare arguments for main (_rt0_go)
+	MOVW	(R13), R0	// argc
+	MOVW	$4(R13), R1		// argv
+	MOVW	$main(SB), R4
+	B		(R4)
+
+TEXT main(SB),NOSPLIT,$-8
+	// save argc and argv onto stack
+	MOVM.DB.W [R0-R1], (R13)
+	MOVW	$runtime·rt0_go(SB), R4
+	B		(R4)
diff --git a/src/runtime/runtime-gdb_test.go b/src/runtime/runtime-gdb_test.go
index 4a74dd372a..4911dc000d 100644
--- a/src/runtime/runtime-gdb_test.go
+++ b/src/runtime/runtime-gdb_test.go
@@ -1,21 +1,23 @@
 package runtime_test
 
 import (
+	"fmt"
 	"io/ioutil"
 	"os"
 	"os/exec"
 	"path/filepath"
+	"runtime"
 	"testing"
 )
 
 func checkGdbPython(t *testing.T) {
-	cmd := exec.Command("gdb", "-nx", "-q", "--batch", "-ex", "python import sys; print('golang gdb python support')")
+	cmd := exec.Command("gdb", "-nx", "-q", "--batch", "-iex", "python import sys; print('go gdb python support')")
 	out, err := cmd.CombinedOutput()
 
 	if err != nil {
-		t.Skipf("skipping due to issue running gdb%v", err)
+		t.Skipf("skipping due to issue running gdb: %v", err)
 	}
-	if string(out) != "golang gdb python support\n" {
+	if string(out) != "go gdb python support\n" {
 		t.Skipf("skipping due to lack of python gdb support: %s", out)
 	}
 }
@@ -29,7 +31,6 @@ func main() {
 `
 
 func TestGdbLoadRuntimeSupport(t *testing.T) {
-
 	checkGdbPython(t)
 
 	dir, err := ioutil.TempDir("", "go-build")
@@ -51,7 +52,8 @@ func TestGdbLoadRuntimeSupport(t *testing.T) {
 		t.Fatalf("building source %v\n%s", err, out)
 	}
 
-	got, _ := exec.Command("gdb", "-nx", "-q", "--batch", "-ex", "source runtime-gdb.py",
+	got, _ := exec.Command("gdb", "-nx", "-q", "--batch", "-iex",
+		fmt.Sprintf("add-auto-load-safe-path %s/src/runtime", runtime.GOROOT()),
 		filepath.Join(dir, "a.exe")).CombinedOutput()
 	if string(got) != "Loading Go Runtime support.\n" {
 		t.Fatalf("%s", got)
diff --git a/src/runtime/runtime2.go b/src/runtime/runtime2.go
index e38d11a59d..ca3e7d564e 100644
--- a/src/runtime/runtime2.go
+++ b/src/runtime/runtime2.go
@@ -199,7 +199,6 @@ type g struct {
 	waitsince    int64  // approx time when the g become blocked
 	waitreason   string // if status==gwaiting
 	schedlink    *g
-	issystem     bool // do not output in stack dump, ignore in deadlock detector
 	preempt      bool // preemption signal, duplicates stackguard0 = stackpreempt
 	paniconfault bool // panic (instead of crash) on unexpected fault address
 	preemptscan  bool // preempted g does scan for gc
@@ -275,6 +274,7 @@ type m struct {
 	waitsemacount uint32
 	waitsemalock  uint32
 	gcstats       gcstats
+	currentwbuf   uintptr // use locks or atomic operations such as xchguinptr to access.
 	needextram    bool
 	traceback     uint8
 	waitunlockf   unsafe.Pointer // todo go func(*g, unsafe.pointer) bool
diff --git a/src/runtime/signal_darwin_arm.go b/src/runtime/signal_darwin_arm.go
new file mode 100644
index 0000000000..1441a655ef
--- /dev/null
+++ b/src/runtime/signal_darwin_arm.go
@@ -0,0 +1,44 @@
+// Copyright 2014 The Go Authors.  All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package runtime
+
+import "unsafe"
+
+type sigctxt struct {
+	info *siginfo
+	ctxt unsafe.Pointer
+}
+
+func (c *sigctxt) regs() *regs32   { return &(*ucontext)(c.ctxt).uc_mcontext.ss }
+func (c *sigctxt) r0() uint32      { return c.regs().r[0] }
+func (c *sigctxt) r1() uint32      { return c.regs().r[1] }
+func (c *sigctxt) r2() uint32      { return c.regs().r[2] }
+func (c *sigctxt) r3() uint32      { return c.regs().r[3] }
+func (c *sigctxt) r4() uint32      { return c.regs().r[4] }
+func (c *sigctxt) r5() uint32      { return c.regs().r[5] }
+func (c *sigctxt) r6() uint32      { return c.regs().r[6] }
+func (c *sigctxt) r7() uint32      { return c.regs().r[7] }
+func (c *sigctxt) r8() uint32      { return c.regs().r[8] }
+func (c *sigctxt) r9() uint32      { return c.regs().r[9] }
+func (c *sigctxt) r10() uint32     { return c.regs().r[10] }
+func (c *sigctxt) fp() uint32      { return c.regs().r[11] }
+func (c *sigctxt) ip() uint32      { return c.regs().r[12] }
+func (c *sigctxt) sp() uint32      { return c.regs().sp }
+func (c *sigctxt) lr() uint32      { return c.regs().lr }
+func (c *sigctxt) pc() uint32      { return c.regs().pc }
+func (c *sigctxt) cpsr() uint32    { return c.regs().cpsr }
+func (c *sigctxt) fault() uint32   { return uint32(uintptr(unsafe.Pointer(c.info.si_addr))) }
+func (c *sigctxt) sigcode() uint32 { return uint32(c.info.si_code) }
+func (c *sigctxt) trap() uint32    { return 0 }
+func (c *sigctxt) error() uint32   { return 0 }
+func (c *sigctxt) oldmask() uint32 { return 0 }
+
+func (c *sigctxt) set_pc(x uint32)  { c.regs().pc = x }
+func (c *sigctxt) set_sp(x uint32)  { c.regs().sp = x }
+func (c *sigctxt) set_lr(x uint32)  { c.regs().lr = x }
+func (c *sigctxt) set_r10(x uint32) { c.regs().r[10] = x }
+
+func (c *sigctxt) set_sigcode(x uint32) { c.info.si_code = int32(x) }
+func (c *sigctxt) set_sigaddr(x uint32) { c.info.si_addr = (*byte)(unsafe.Pointer(uintptr(x))) }
diff --git a/src/runtime/signals_windows.h b/src/runtime/signals_windows.h
deleted file mode 100644
index 6943714b0f..0000000000
--- a/src/runtime/signals_windows.h
+++ /dev/null
@@ -1,3 +0,0 @@
-// Copyright 2011 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
diff --git a/src/runtime/sigqueue.go b/src/runtime/sigqueue.go
index fbe3425fa6..5cfc926f67 100644
--- a/src/runtime/sigqueue.go
+++ b/src/runtime/sigqueue.go
@@ -24,6 +24,8 @@
 // unnecessary rechecks of sig.mask, but it cannot lead to missed signals
 // nor deadlocks.
 
+// +build !plan9
+
 package runtime
 
 import "unsafe"
diff --git a/src/runtime/sigqueue_plan9.go b/src/runtime/sigqueue_plan9.go
new file mode 100644
index 0000000000..b029a300a8
--- /dev/null
+++ b/src/runtime/sigqueue_plan9.go
@@ -0,0 +1,115 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// This file implements runtime support for signal handling.
+
+package runtime
+
+const qsize = 64
+
+var sig struct {
+	q     noteQueue
+	inuse bool
+
+	lock     mutex
+	note     note
+	sleeping bool
+}
+
+type noteQueue struct {
+	lock mutex
+	data [qsize]*byte
+	ri   int
+	wi   int
+	full bool
+}
+
+func (q *noteQueue) push(item *byte) bool {
+	lock(&q.lock)
+	if q.full {
+		unlock(&q.lock)
+		return false
+	}
+	q.data[q.wi] = item
+	q.wi++
+	if q.wi == qsize {
+		q.wi = 0
+	}
+	if q.wi == q.ri {
+		q.full = true
+	}
+	unlock(&q.lock)
+	return true
+}
+
+func (q *noteQueue) pop() *byte {
+	lock(&q.lock)
+	q.full = false
+	if q.ri == q.wi {
+		unlock(&q.lock)
+		return nil
+	}
+	item := q.data[q.ri]
+	q.ri++
+	if q.ri == qsize {
+		q.ri = 0
+	}
+	unlock(&q.lock)
+	return item
+}
+
+// Called from sighandler to send a signal back out of the signal handling thread.
+// Reports whether the signal was sent. If not, the caller typically crashes the program.
+func sendNote(s *byte) bool {
+	if !sig.inuse {
+		return false
+	}
+
+	// Add signal to outgoing queue.
+	if !sig.q.push(s) {
+		return false
+	}
+
+	lock(&sig.lock)
+	if sig.sleeping {
+		sig.sleeping = false
+		notewakeup(&sig.note)
+	}
+	unlock(&sig.lock)
+
+	return true
+}
+
+// Called to receive the next queued signal.
+// Must only be called from a single goroutine at a time.
+func signal_recv() string {
+	for {
+		note := sig.q.pop()
+		if note != nil {
+			return gostring(note)
+		}
+
+		lock(&sig.lock)
+		sig.sleeping = true
+		noteclear(&sig.note)
+		unlock(&sig.lock)
+		notetsleepg(&sig.note, -1)
+	}
+}
+
+// Must only be called from a single goroutine at a time.
+func signal_enable(s uint32) {
+	if !sig.inuse {
+		// The first call to signal_enable is for us
+		// to use for initialization.  It does not pass
+		// signal information in m.
+		sig.inuse = true // enable reception of signals; cannot disable
+		noteclear(&sig.note)
+		return
+	}
+}
+
+// Must only be called from a single goroutine at a time.
+func signal_disable(s uint32) {
+}
diff --git a/src/runtime/stack.h b/src/runtime/stack.h
index b790e70103..88c7e02f40 100644
--- a/src/runtime/stack.h
+++ b/src/runtime/stack.h
@@ -7,14 +7,22 @@
 
 enum {
 #ifdef GOOS_windows
-#define StackSystem (512*sizeof(uintptr))
-#else
+#define STACKSYSTEM (512 * sizeof(uintptr))
+#endif // GOOS_windows
 #ifdef GOOS_plan9
-#define StackSystem (512)
-#else
-	StackSystem = 0,
-#endif	// Plan 9
-#endif	// Windows
+#define STACKSYSTEM	512
+#endif // GOOS_plan9
+#ifdef GOOS_darwin
+#ifdef GOARCH_arm
+#define STACKSYSTEM 1024
+#endif // GOARCH_arm
+#endif // GOOS_darwin
+
+#ifndef STACKSYSTEM
+#define STACKSYSTEM 0
+#endif
+
+	StackSystem = STACKSYSTEM,
 
 	StackBig = 4096,
 	StackGuard = 640 + StackSystem,
diff --git a/src/runtime/stack2.go b/src/runtime/stack2.go
index 8a78b1ad96..07a7d38f0c 100644
--- a/src/runtime/stack2.go
+++ b/src/runtime/stack2.go
@@ -57,9 +57,9 @@ functions to make sure that this limit cannot be violated.
 const (
 	// StackSystem is a number of additional bytes to add
 	// to each stack below the usual guard area for OS-specific
-	// purposes like signal handling. Used on Windows and on
-	// Plan 9 because they do not use a separate stack.
-	_StackSystem = goos_windows*512*ptrSize + goos_plan9*512
+	// purposes like signal handling. Used on Windows, Plan 9,
+	// and Darwin/ARM because they do not use a separate stack.
+	_StackSystem = goos_windows*512*ptrSize + goos_plan9*512 + goos_darwin*goarch_arm*1024
 
 	// The minimum size of stack used by Go code
 	_StackMin = 2048
diff --git a/src/runtime/string.go b/src/runtime/string.go
index 46c3502f77..0ba309cf02 100644
--- a/src/runtime/string.go
+++ b/src/runtime/string.go
@@ -129,8 +129,13 @@ func slicebytetostringtmp(b []byte) string {
 	return *(*string)(unsafe.Pointer(&b))
 }
 
-func stringtoslicebyte(s string) []byte {
-	b := rawbyteslice(len(s))
+func stringtoslicebyte(buf *tmpBuf, s string) []byte {
+	var b []byte
+	if buf != nil && len(s) <= len(buf) {
+		b = buf[:len(s)]
+	} else {
+		b = rawbyteslice(len(s))
+	}
 	copy(b, s)
 	return b
 }
@@ -147,7 +152,7 @@ func stringtoslicebytetmp(s string) []byte {
 	return *(*[]byte)(unsafe.Pointer(&ret))
 }
 
-func stringtoslicerune(s string) []rune {
+func stringtoslicerune(buf *[tmpStringBufSize]rune, s string) []rune {
 	// two passes.
 	// unlike slicerunetostring, no race because strings are immutable.
 	n := 0
@@ -157,7 +162,12 @@ func stringtoslicerune(s string) []rune {
 		s = s[k:]
 		n++
 	}
-	a := rawruneslice(n)
+	var a []rune
+	if buf != nil && n <= len(buf) {
+		a = buf[:n]
+	} else {
+		a = rawruneslice(n)
+	}
 	n = 0
 	for len(t) > 0 {
 		r, k := charntorune(t)
@@ -168,7 +178,7 @@ func stringtoslicerune(s string) []rune {
 	return a
 }
 
-func slicerunetostring(a []rune) string {
+func slicerunetostring(buf *tmpBuf, a []rune) string {
 	if raceenabled && len(a) > 0 {
 		racereadrangepc(unsafe.Pointer(&a[0]),
 			uintptr(len(a))*unsafe.Sizeof(a[0]),
@@ -180,7 +190,7 @@ func slicerunetostring(a []rune) string {
 	for _, r := range a {
 		size1 += runetochar(dum[:], r)
 	}
-	s, b := rawstring(size1 + 3)
+	s, b := rawstringtmp(buf, size1+3)
 	size2 := 0
 	for _, r := range a {
 		// check for race
@@ -309,11 +319,6 @@ func gobytes(p *byte, n int) []byte {
 	return x
 }
 
-func gostringsize(n int) string {
-	s, _ := rawstring(n)
-	return s
-}
-
 func gostring(p *byte) string {
 	l := findnull(p)
 	if l == 0 {
diff --git a/src/runtime/stubs3.go b/src/runtime/stubs3.go
deleted file mode 100644
index ffaa28775d..0000000000
--- a/src/runtime/stubs3.go
+++ /dev/null
@@ -1,12 +0,0 @@
-// Copyright 2014 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-// +build plan9
-
-package runtime
-
-func close(fd int32) int32
-
-//go:noescape
-func open(name *byte, mode, perm int32) int32
diff --git a/src/runtime/symtab.go b/src/runtime/symtab.go
index db20ab11e1..3e46d428f7 100644
--- a/src/runtime/symtab.go
+++ b/src/runtime/symtab.go
@@ -44,8 +44,8 @@ type functab struct {
 	funcoff uintptr
 }
 
-const minfunc = 16 // minimum function size
-const pcbucketsize = 256*minfunc // size of bucket in the pc->func lookup table
+const minfunc = 16                 // minimum function size
+const pcbucketsize = 256 * minfunc // size of bucket in the pc->func lookup table
 
 // findfunctab is an array of these structures.
 // Each bucket represents 4096 bytes of the text segment.
@@ -56,7 +56,7 @@ const pcbucketsize = 256*minfunc // size of bucket in the pc->func lookup table
 // index to find the target function.
 // This table uses 20 bytes for every 4096 bytes of code, or ~0.5% overhead.
 type findfuncbucket struct {
-	idx uint32
+	idx        uint32
 	subbuckets [16]byte
 }
 
@@ -154,9 +154,9 @@ func findfunc(pc uintptr) *_func {
 
 	x := pc - minpc
 	b := x / pcbucketsize
-	i := x % pcbucketsize / (pcbucketsize/nsub)
+	i := x % pcbucketsize / (pcbucketsize / nsub)
 
-	ffb := (*findfuncbucket)(add(unsafe.Pointer(&findfunctab), b * unsafe.Sizeof(findfuncbucket{})))
+	ffb := (*findfuncbucket)(add(unsafe.Pointer(&findfunctab), b*unsafe.Sizeof(findfuncbucket{})))
 	idx := ffb.idx + uint32(ffb.subbuckets[i])
 	if pc < ftab[idx].entry {
 		throw("findfunc: bad findfunctab entry")
diff --git a/src/runtime/sys_darwin_arm.s b/src/runtime/sys_darwin_arm.s
new file mode 100644
index 0000000000..e1b2b664b6
--- /dev/null
+++ b/src/runtime/sys_darwin_arm.s
@@ -0,0 +1,481 @@
+// Copyright 2014 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// System calls and other sys.stuff for ARM, Darwin
+// See http://fxr.watson.org/fxr/source/bsd/kern/syscalls.c?v=xnu-1228
+// or /usr/include/sys/syscall.h (on a Mac) for system call numbers.
+
+#include "go_asm.h"
+#include "go_tls.h"
+#include "textflag.h"
+
+// Copied from /usr/include/sys/syscall.h
+#define	SYS_exit           1
+#define	SYS_read           3
+#define	SYS_write          4
+#define	SYS_open           5
+#define	SYS_close          6
+#define	SYS_mmap           197
+#define	SYS_munmap         73
+#define	SYS_madvise        75
+#define	SYS_mincore        78
+#define	SYS_gettimeofday   116
+#define	SYS_kill           37
+#define	SYS_getpid         20
+#define	SYS___pthread_kill 328
+#define	SYS_setitimer      83
+#define	SYS___sysctl       202
+#define	SYS_sigprocmask    48
+#define	SYS_sigaction      46
+#define	SYS_sigreturn      184
+#define	SYS_select         93
+#define	SYS_bsdthread_register 366
+#define	SYS_bsdthread_create 360
+#define	SYS_bsdthread_terminate 361
+#define	SYS_kqueue         362
+#define	SYS_kevent         363
+#define	SYS_fcntl          92
+
+TEXT notok<>(SB),NOSPLIT,$0
+	MOVW	$0, R8
+	MOVW	R8, (R8)
+	B		0(PC)
+
+TEXT runtime·open(SB),NOSPLIT,$0
+	MOVW	name+0(FP), R0
+	MOVW	mode+4(FP), R1
+	MOVW	perm+8(FP), R2
+	MOVW	$SYS_open, R12
+	SWI	$0x80
+	MOVW	R0, ret+12(FP)
+	RET
+
+TEXT runtime·close(SB),NOSPLIT,$0
+	MOVW	fd+0(FP), R0
+	MOVW	$SYS_close, R12
+	SWI	$0x80
+	MOVW	R0, ret+4(FP)
+	RET
+
+TEXT runtime·write(SB),NOSPLIT,$0
+	MOVW	fd+0(FP), R0
+	MOVW	p+4(FP), R1
+	MOVW	n+8(FP), R2
+	MOVW	$SYS_write, R12
+	SWI	$0x80
+	MOVW	R0, ret+12(FP)
+	RET
+
+TEXT runtime·read(SB),NOSPLIT,$0
+	MOVW	fd+0(FP), R0
+	MOVW	p+4(FP), R1
+	MOVW	n+8(FP), R2
+	MOVW	$SYS_read, R12
+	SWI	$0x80
+	MOVW	R0, ret+12(FP)
+	RET
+
+TEXT runtime·exit(SB),NOSPLIT,$-4
+	MOVW	0(FP), R0
+	MOVW	$SYS_exit, R12
+	SWI	$0x80
+	MOVW	$1234, R0
+	MOVW	$1002, R1
+	MOVW	R0, (R1)	// fail hard
+
+// Exit this OS thread (like pthread_exit, which eventually
+// calls __bsdthread_terminate).
+TEXT runtime·exit1(SB),NOSPLIT,$0
+	MOVW	$SYS_bsdthread_terminate, R12
+	SWI	$0x80
+	MOVW	$1234, R0
+	MOVW	$1003, R1
+	MOVW	R0, (R1)	// fail hard
+
+TEXT runtime·raise(SB),NOSPLIT,$24
+	MOVW	$SYS_getpid, R12
+	SWI	$0x80
+	// arg 1 pid already in R0 from getpid
+	MOVW	sig+0(FP), R1	// arg 2 - signal
+	MOVW	$1, R2	// arg 3 - posix
+	MOVW	$SYS_kill, R12
+	SWI $0x80
+	RET
+
+TEXT runtime·mmap(SB),NOSPLIT,$0
+	MOVW	addr+0(FP), R0
+	MOVW	n+4(FP), R1
+	MOVW	prot+8(FP), R2
+	MOVW	flags+12(FP), R3
+	MOVW	fd+16(FP), R4
+	MOVW	off+20(FP), R5
+	MOVW	$0, R6 // off_t is uint64_t
+	MOVW	$SYS_mmap, R12
+	SWI	$0x80
+	MOVW	R0, ret+24(FP)
+	RET
+
+TEXT runtime·munmap(SB),NOSPLIT,$0
+	MOVW	0(FP), R0
+	MOVW	4(FP), R1
+	MOVW	$SYS_munmap, R12
+	SWI	$0x80
+	BL.CS	notok<>(SB)
+	RET
+
+TEXT runtime·madvise(SB),NOSPLIT,$0
+	MOVW	0(FP), R0
+	MOVW	4(FP), R1
+	MOVW	8(FP), R2
+	MOVW	$SYS_madvise, R12
+	SWI	$0x80
+	BL.CS	notok<>(SB)
+	RET
+
+TEXT runtime·setitimer(SB),NOSPLIT,$0
+	MOVW	0(FP), R0
+	MOVW	4(FP), R1
+	MOVW	8(FP), R2
+	MOVW	$SYS_setitimer, R12
+	SWI	$0x80
+	RET
+
+TEXT runtime·mincore(SB),NOSPLIT,$0
+	MOVW	0(FP), R0
+	MOVW	4(FP), R1
+	MOVW	8(FP), R2
+	MOVW	$SYS_mincore, R12
+	SWI	$0x80
+	RET
+
+TEXT time·now(SB), 7, $32
+	MOVW	$8(R13), R0  // timeval
+	MOVW	$0, R1  // zone
+	MOVW	$SYS_gettimeofday, R12
+	SWI	$0x80 // Note: R0 is tv_sec, R1 is tv_usec
+
+	MOVW    R1, R2  // usec
+
+	MOVW	R0, 0(FP)
+	MOVW	$0, R1
+	MOVW	R1, 4(FP)
+	MOVW	$1000, R3
+	MUL	R3, R2
+	MOVW	R2, 8(FP)
+	RET
+
+TEXT runtime·nanotime(SB),NOSPLIT,$32
+	MOVW	$8(R13), R0  // timeval
+	MOVW	$0, R1  // zone
+	MOVW	$SYS_gettimeofday, R12
+	SWI	$0x80 // Note: R0 is tv_sec, R1 is tv_usec
+
+	MOVW    R1, R2
+	MOVW	$1000000000, R3
+	MULLU	R0, R3, (R1, R0)
+	MOVW	$1000, R3
+	MOVW	$0, R4
+	MUL	R3, R2
+	ADD.S	R2, R0
+	ADC	R4, R1
+
+	MOVW	R0, 0(FP)
+	MOVW	R1, 4(FP)
+	RET
+
+// Sigtramp's job is to call the actual signal handler.
+// It is called with the following arguments on the stack:
+//	 LR  	"return address" - ignored
+//	 R0  	actual handler
+//	 R1  	siginfo style - ignored
+//	 R2   	signal number
+//	 R3   	siginfo
+//	 -4(FP)	context, beware that 0(FP) is the saved LR
+TEXT runtime·sigtramp(SB),NOSPLIT,$0
+	// this might be called in external code context,
+	// where g is not set.
+	// first save R0, because runtime·load_g will clobber it
+	MOVM.DB.W [R0], (R13)
+	MOVB	runtime·iscgo(SB), R0
+	CMP 	$0, R0
+	BL.NE	runtime·load_g(SB)
+
+	CMP 	$0, g
+	BNE 	cont
+	// fake function call stack frame for badsignal
+	// we only need to pass R2 (signal number), but
+	// badsignal will expect R2 at 4(R13), so we also
+	// push R1 onto stack. turns out we do need R1
+	// to do sigreturn.
+	MOVM.DB.W [R1,R2], (R13)
+	MOVW  	$runtime·badsignal(SB), R11
+	BL	(R11)
+	MOVM.IA.W [R1], (R13) // saved infostype
+	ADD		$(4+4), R13 // +4: also need to remove the pushed R0.
+	MOVW    -4(FP), R0 // load ucontext
+	B	ret
+
+cont:
+	// Restore R0
+	MOVM.IA.W (R13), [R0]
+
+	// NOTE: some Darwin/ARM kernels always use the main stack to run the
+	// signal handler. We need to switch to gsignal ourselves.
+	MOVW	g_m(g), R11
+	MOVW	m_gsignal(R11), R5
+	MOVW	(g_stack+stack_hi)(R5), R6
+	SUB		$28, R6
+
+	// copy arguments for call to sighandler
+	MOVW	R2, 4(R6) // signal num
+	MOVW	R3, 8(R6) // signal info
+	MOVW	g, 16(R6) // old_g
+	MOVW    -4(FP), R4
+	MOVW	R4, 12(R6) // context
+
+	// Backup ucontext and infostyle
+	MOVW    R4, 20(R6)
+	MOVW    R1, 24(R6)
+
+	// switch stack and g
+	MOVW	R6, R13 // sigtramp can not re-entrant, so no need to back up R13.
+	MOVW	R5, g
+
+	BL	(R0)
+
+	// call sigreturn
+	MOVW	20(R13), R0	// saved ucontext
+	MOVW	24(R13), R1	// saved infostyle
+ret:
+	MOVW	$SYS_sigreturn, R12 // sigreturn(ucontext, infostyle)
+	SWI	$0x80
+
+	// if sigreturn fails, we can do nothing but exit
+	B	runtime·exit(SB)
+
+TEXT runtime·sigprocmask(SB),NOSPLIT,$0
+	MOVW	0(FP), R0
+	MOVW	4(FP), R1
+	MOVW	8(FP), R2
+	MOVW	$SYS_sigprocmask, R12
+	SWI	$0x80
+	BL.CS	notok<>(SB)
+	RET
+
+TEXT runtime·sigaction(SB),NOSPLIT,$0
+	MOVW	0(FP), R0
+	MOVW	4(FP), R1
+	MOVW	8(FP), R2
+	MOVW	$SYS_sigaction, R12
+	SWI	$0x80
+	RET
+
+TEXT runtime·usleep(SB),NOSPLIT,$12
+	MOVW	usec+0(FP), R0
+	MOVW	R0, R1
+	MOVW	$1000000, R2
+	DIV     R2, R0
+	MOD     R2, R1
+	MOVW	R0, -12(SP)
+	MOVW	R1, -8(SP)
+
+	// select(0, 0, 0, 0, &tv)
+	MOVW	$0, R0
+	MOVW	$0, R1
+	MOVW	$0, R2
+	MOVW	$0, R3
+	MOVW	$-12(SP), R4
+	MOVW	$SYS_select, R12
+	SWI	$0x80
+	RET
+
+TEXT runtime·cas(SB),NOSPLIT,$0
+	B	runtime·armcas(SB)
+
+TEXT runtime·casp1(SB),NOSPLIT,$0
+	B	runtime·cas(SB)
+
+TEXT runtime·sysctl(SB),NOSPLIT,$0
+	MOVW	0(FP), R0
+	MOVW	4(FP), R1
+	MOVW	8(FP), R2
+	MOVW	12(FP), R3
+	MOVW	16(FP), R4
+	MOVW	20(FP), R5
+	MOVW	$SYS___sysctl, R12 // syscall entry
+	SWI	$0x80
+	BCC     sysctl_ret
+	RSB     $0, R0, R0
+	MOVW	R0, ret+24(FP)
+	RET
+sysctl_ret:
+	MOVW	$0, R0
+	MOVW	R0, ret+24(FP)
+	RET
+
+// Thread related functions
+// void bsdthread_create(void *stk, M *m, G *g, void (*fn)(void))
+TEXT runtime·bsdthread_create(SB),NOSPLIT,$0
+	// Set up arguments to bsdthread_create system call.
+	// The ones in quotes pass through to the thread callback
+	// uninterpreted, so we can put whatever we want there.
+	MOVW    fn+12(FP), R0   // "func"
+	MOVW    mm+4(FP), R1 // "arg"
+	MOVW    stk+0(FP), R2 // stack
+	MOVW    gg+8(FP), R3 // "pthread"
+	MOVW	$0x01000000, R4	// flags = PTHREAD_START_CUSTOM
+	MOVW	$0, R5	// paranoia
+	MOVW	$SYS_bsdthread_create, R12
+	SWI	$0x80
+	BCC		create_ret
+	RSB 	$0, R0, R0
+	MOVW	R0, ret+16(FP)
+	RET
+create_ret:
+	MOVW	$0, R0
+	MOVW	R0, ret+16(FP)
+	RET
+
+// The thread that bsdthread_create creates starts executing here,
+// because we registered this function using bsdthread_register
+// at startup.
+//	R0 = "pthread"
+//	R1 = mach thread port
+//	R2 = "func" (= fn)
+//	R3 = "arg" (= m)
+//	R4 = stack
+//	R5 = flags (= 0)
+// XXX: how to deal with R4/SP? ref: Libc-594.9.1/arm/pthreads/thread_start.s
+TEXT runtime·bsdthread_start(SB),NOSPLIT,$0
+	MOVW    R1, m_procid(R3) // thread port is m->procid
+	MOVW	m_g0(R3), g
+	MOVW	R3, g_m(g)
+	// ARM don't have runtime·stackcheck(SB)
+	// disable runfast mode of vfp
+	EOR     R12, R12
+	WORD    $0xeee1ca10 // fmxr	fpscr, ip
+	BL      (R2) // fn
+	BL      runtime·exit1(SB)
+	RET
+
+// int32 bsdthread_register(void)
+// registers callbacks for threadstart (see bsdthread_create above
+// and wqthread and pthsize (not used).  returns 0 on success.
+TEXT runtime·bsdthread_register(SB),NOSPLIT,$0
+	MOVW	$runtime·bsdthread_start(SB), R0	// threadstart
+	MOVW	$0, R1	// wqthread, not used by us
+	MOVW	$0, R2	// pthsize, not used by us
+	MOVW	$0, R3 	// dummy_value [sic]
+	MOVW	$0, R4	// targetconc_ptr
+	MOVW	$0, R5	// dispatchqueue_offset
+	MOVW	$SYS_bsdthread_register, R12	// bsdthread_register
+	SWI	$0x80
+	MOVW	R0, ret+0(FP)
+	RET
+
+// uint32 mach_msg_trap(void*, uint32, uint32, uint32, uint32, uint32, uint32)
+TEXT runtime·mach_msg_trap(SB),NOSPLIT,$0
+	MOVW    0(FP), R0
+	MOVW    4(FP), R1
+	MOVW    8(FP), R2
+	MOVW    12(FP), R3
+	MOVW    16(FP), R4
+	MOVW    20(FP), R5
+	MOVW    24(FP), R6
+	MVN     $30, R12
+	SWI	$0x80
+	MOVW	R0, 28(FP)
+	RET
+
+TEXT runtime·mach_task_self(SB),NOSPLIT,$0
+	MVN     $27, R12 // task_self_trap
+	SWI	$0x80
+	MOVW	R0, 0(FP)
+	RET
+
+TEXT runtime·mach_thread_self(SB),NOSPLIT,$0
+	MVN 	$26, R12 // thread_self_trap
+	SWI	$0x80
+	MOVW	R0, 0(FP)
+	RET
+
+TEXT runtime·mach_reply_port(SB),NOSPLIT,$0
+	MVN 	$25, R12	// mach_reply_port
+	SWI	$0x80
+	MOVW	R0, 0(FP)
+	RET
+
+// Mach provides trap versions of the semaphore ops,
+// instead of requiring the use of RPC.
+
+// uint32 mach_semaphore_wait(uint32)
+TEXT runtime·mach_semaphore_wait(SB),NOSPLIT,$0
+	MOVW	0(FP), R0
+	MVN 	$35, R12	// semaphore_wait_trap
+	SWI	$0x80
+	MOVW	R0, ret+4(FP)
+	RET
+
+// uint32 mach_semaphore_timedwait(uint32, uint32, uint32)
+TEXT runtime·mach_semaphore_timedwait(SB),NOSPLIT,$0
+	MOVW	0(FP), R0
+	MOVW	4(FP), R1
+	MOVW	8(FP), R2
+	MVN 	$37, R12	// semaphore_timedwait_trap
+	SWI	$0x80
+	MOVW	R0, ret+12(FP)
+	RET
+
+// uint32 mach_semaphore_signal(uint32)
+TEXT runtime·mach_semaphore_signal(SB),NOSPLIT,$0
+	MOVW    0(FP), R0
+	MVN 	$32, R12	// semaphore_signal_trap
+	SWI	$0x80
+	MOVW	R0, ret+4(FP)
+	RET
+
+// uint32 mach_semaphore_signal_all(uint32)
+TEXT runtime·mach_semaphore_signal_all(SB),NOSPLIT,$0
+	MOVW	0(FP), R0
+	MVN 	$33, R12	// semaphore_signal_all_trap
+	SWI	$0x80
+	MOVW	R0, ret+4(FP)
+	RET
+
+// int32 runtime·kqueue(void)
+TEXT runtime·kqueue(SB),NOSPLIT,$0
+	MOVW	$SYS_kqueue, R12
+	SWI	$0x80
+	RSB.CS	$0, R0, R0
+	MOVW	R0, ret+0(FP)
+	RET
+
+// int32 runtime·kevent(int kq, Kevent *changelist, int nchanges, Kevent *eventlist, int events, Timespec *timeout)
+TEXT runtime·kevent(SB),NOSPLIT,$0
+	MOVW	$SYS_kevent, R12
+	MOVW	kq+0(FP), R0
+	MOVW	changelist+4(FP), R1
+	MOVW	nchanges+8(FP), R2
+	MOVW	eventlist+12(FP), R3
+	MOVW	nevents+16(FP), R4
+	MOVW	timeout+20(FP), R5
+	SWI	$0x80
+	RSB.CS	$0, R0, R0
+	MOVW	R0, ret+24(FP)
+	RET
+
+// int32 runtime·closeonexec(int32 fd)
+TEXT runtime·closeonexec(SB),NOSPLIT,$0
+	MOVW	$SYS_fcntl, R12
+	MOVW	0(FP), R0
+	MOVW	$2, R1	// F_SETFD
+	MOVW	$1, R2	// FD_CLOEXEC
+	SWI	$0x80
+	RET
+
+// sigaltstack on some darwin/arm version is buggy and will always
+// run the signal handler on the main stack, so our sigtramp has
+// to do the stack switch ourselves.
+TEXT runtime·sigaltstack(SB),NOSPLIT,$0
+	RET
diff --git a/src/runtime/time.go b/src/runtime/time.go
index 50895ca8ec..6a2cc2136c 100644
--- a/src/runtime/time.go
+++ b/src/runtime/time.go
@@ -153,7 +153,6 @@ func deltimer(t *timer) bool {
 // If addtimer inserts a new earlier event, addtimer1 wakes timerproc early.
 func timerproc() {
 	timers.gp = getg()
-	timers.gp.issystem = true
 	for {
 		lock(&timers.lock)
 		timers.sleeping = false
diff --git a/src/runtime/tls_arm.s b/src/runtime/tls_arm.s
index 7c5c0e215e..d130d42cf2 100644
--- a/src/runtime/tls_arm.s
+++ b/src/runtime/tls_arm.s
@@ -15,8 +15,14 @@
 // Note: both functions will clobber R0 and R11 and
 // can be called from 5c ABI code.
 
-// On android, runtime.tlsg is a normal variable.
+// On android and darwin, runtime.tlsg is a normal variable.
 // TLS offset is computed in x_cgo_inittls.
+#ifdef GOOS_android
+#define TLSG_IS_VARIABLE
+#endif
+#ifdef GOOS_darwin
+#define TLSG_IS_VARIABLE
+#endif
 
 // save_g saves the g register into pthread-provided
 // thread-local memory, so that we can call externally compiled
@@ -34,10 +40,11 @@ TEXT runtime·save_g(SB),NOSPLIT,$-4
 	// a call to runtime.read_tls_fallback which jumps to __kuser_get_tls.
 	// The replacement function saves LR in R11 over the call to read_tls_fallback.
 	MRC	15, 0, R0, C13, C0, 3 // fetch TLS base pointer
+	BIC $3, R0 // Darwin/ARM might return unaligned pointer
 	// $runtime.tlsg(SB) is a special linker symbol.
 	// It is the offset from the TLS base pointer to our
 	// thread-local storage for g.
-#ifdef GOOS_android
+#ifdef TLSG_IS_VARIABLE
 	MOVW	runtime·tlsg(SB), R11
 #else
 	MOVW	$runtime·tlsg(SB), R11
@@ -57,10 +64,11 @@ TEXT runtime·load_g(SB),NOSPLIT,$0
 #endif
 	// See save_g
 	MRC	15, 0, R0, C13, C0, 3 // fetch TLS base pointer
+	BIC $3, R0 // Darwin/ARM might return unaligned pointer
 	// $runtime.tlsg(SB) is a special linker symbol.
 	// It is the offset from the TLS base pointer to our
 	// thread-local storage for g.
-#ifdef GOOS_android
+#ifdef TLSG_IS_VARIABLE
 	MOVW	runtime·tlsg(SB), R11
 #else
 	MOVW	$runtime·tlsg(SB), R11
@@ -68,3 +76,28 @@ TEXT runtime·load_g(SB),NOSPLIT,$0
 	ADD	R11, R0
 	MOVW	0(R0), g
 	RET
+
+TEXT runtime·_initcgo(SB),NOSPLIT,$0
+#ifndef GOOS_nacl
+	// if there is an _cgo_init, call it.
+	MOVW	_cgo_init(SB), R4
+	CMP	$0, R4
+	B.EQ	nocgo
+	MRC     15, 0, R0, C13, C0, 3 	// load TLS base pointer
+	MOVW 	R0, R3 			// arg 3: TLS base pointer
+	MOVW 	$runtime·tlsg(SB), R2 	// arg 2: tlsg
+	MOVW	$setg_gcc<>(SB), R1 	// arg 1: setg
+	MOVW	g, R0 			// arg 0: G
+	BL	(R4) // will clobber R0-R3
+#endif
+nocgo:
+	RET
+
+// void setg_gcc(G*); set g called from gcc.
+TEXT setg_gcc<>(SB),NOSPLIT,$0
+	MOVW	R0, g
+	B		runtime·save_g(SB)
+
+#ifdef TLSG_IS_VARIABLE
+GLOBL runtime·tlsg+0(SB), NOPTR, $4
+#endif
diff --git a/src/runtime/traceback.go b/src/runtime/traceback.go
index 6c87d7e2e4..8c31c5abad 100644
--- a/src/runtime/traceback.go
+++ b/src/runtime/traceback.go
@@ -39,6 +39,11 @@ var (
 	mstartPC             uintptr
 	rt0_goPC             uintptr
 	sigpanicPC           uintptr
+	runfinqPC            uintptr
+	backgroundgcPC       uintptr
+	bgsweepPC            uintptr
+	forcegchelperPC      uintptr
+	timerprocPC          uintptr
 	systemstack_switchPC uintptr
 
 	externalthreadhandlerp uintptr // initialized elsewhere
@@ -56,6 +61,11 @@ func tracebackinit() {
 	mstartPC = funcPC(mstart)
 	rt0_goPC = funcPC(rt0_go)
 	sigpanicPC = funcPC(sigpanic)
+	runfinqPC = funcPC(runfinq)
+	backgroundgcPC = funcPC(backgroundgc)
+	bgsweepPC = funcPC(bgsweep)
+	forcegchelperPC = funcPC(forcegchelper)
+	timerprocPC = funcPC(timerproc)
 	systemstack_switchPC = funcPC(systemstack_switch)
 }
 
@@ -606,7 +616,7 @@ func tracebackothers(me *g) {
 
 	lock(&allglock)
 	for _, gp := range allgs {
-		if gp == me || gp == g.m.curg || readgstatus(gp) == _Gdead || gp.issystem && level < 2 {
+		if gp == me || gp == g.m.curg || readgstatus(gp) == _Gdead || isSystemGoroutine(gp) && level < 2 {
 			continue
 		}
 		print("\n")
@@ -631,3 +641,14 @@ func topofstack(f *_func) bool {
 		pc == rt0_goPC ||
 		externalthreadhandlerp != 0 && pc == externalthreadhandlerp
 }
+
+// isSystemGoroutine returns true if the goroutine g must be omitted in
+// stack dumps and deadlock detector.
+func isSystemGoroutine(gp *g) bool {
+	pc := gp.startpc
+	return pc == runfinqPC && !fingRunning ||
+		pc == backgroundgcPC ||
+		pc == bgsweepPC ||
+		pc == forcegchelperPC ||
+		pc == timerprocPC
+}
diff --git a/src/sort/sort.go b/src/sort/sort.go
index b52b54ed8f..0a446c8255 100644
--- a/src/sort/sort.go
+++ b/src/sort/sort.go
@@ -296,7 +296,7 @@ func StringsAreSorted(a []string) bool { return IsSorted(StringSlice(a)) }
 //    and Jukka Teuhola; Nordic Journal of Computing 3,1 (1996), 27-40:
 //    The given algorithms are in-place, number of Swap and Assignments
 //    grow as n log n but the algorithm is not stable.
-//  - "Fast Stable In-Plcae Sorting with O(n) Data Moves" J.I. Munro and
+//  - "Fast Stable In-Place Sorting with O(n) Data Moves" J.I. Munro and
 //    V. Raman in Algorithmica (1996) 16, 115-160:
 //    This algorithm either needs additional 2n bits or works only if there
 //    are enough different elements available to encode some permutations
diff --git a/src/strconv/decimal.go b/src/strconv/decimal.go
index 42601283d2..3d7c8d1da9 100644
--- a/src/strconv/decimal.go
+++ b/src/strconv/decimal.go
@@ -12,7 +12,7 @@
 package strconv
 
 type decimal struct {
-	d     [800]byte // digits
+	d     [800]byte // digits, big-endian representation
 	nd    int       // number of digits used
 	dp    int       // decimal point
 	neg   bool
@@ -105,7 +105,7 @@ func (a *decimal) Assign(v uint64) {
 // Signed int has 31 bits, and we have to be able to accommodate 9<<k.
 const maxShift = 27
 
-// Binary shift right (* 2) by k bits.  k <= maxShift to avoid overflow.
+// Binary shift right (/ 2) by k bits.  k <= maxShift to avoid overflow.
 func rightShift(a *decimal, k uint) {
 	r := 0 // read pointer
 	w := 0 // write pointer
@@ -228,7 +228,7 @@ func prefixIsLessThan(b []byte, s string) bool {
 	return false
 }
 
-// Binary shift left (/ 2) by k bits.  k <= maxShift to avoid overflow.
+// Binary shift left (* 2) by k bits.  k <= maxShift to avoid overflow.
 func leftShift(a *decimal, k uint) {
 	delta := leftcheats[k].delta
 	if prefixIsLessThan(a.d[0:a.nd], leftcheats[k].cutoff) {
diff --git a/src/strconv/ftoa.go b/src/strconv/ftoa.go
index 1a9c41b85a..f885d96e9c 100644
--- a/src/strconv/ftoa.go
+++ b/src/strconv/ftoa.go
@@ -119,7 +119,7 @@ func genericFtoa(dst []byte, val float64, fmt byte, prec, bitSize int) []byte {
 		// Precision for shortest representation mode.
 		switch fmt {
 		case 'e', 'E':
-			prec = digs.nd - 1
+			prec = max(digs.nd-1, 0)
 		case 'f':
 			prec = max(digs.nd-digs.dp, 0)
 		case 'g', 'G':
@@ -348,14 +348,13 @@ func fmtE(dst []byte, neg bool, d decimalSlice, prec int, fmt byte) []byte {
 	if prec > 0 {
 		dst = append(dst, '.')
 		i := 1
-		m := d.nd + prec + 1 - max(d.nd, prec+1)
-		for i < m {
-			dst = append(dst, d.d[i])
-			i++
+		m := min(d.nd, prec+1)
+		if i < m {
+			dst = append(dst, d.d[i:m]...)
+			i = m
 		}
-		for i <= prec {
+		for ; i <= prec; i++ {
 			dst = append(dst, '0')
-			i++
 		}
 	}
 
@@ -373,27 +372,16 @@ func fmtE(dst []byte, neg bool, d decimalSlice, prec int, fmt byte) []byte {
 	}
 	dst = append(dst, ch)
 
-	// dddd
-	var buf [3]byte
-	i := len(buf)
-	for exp >= 10 {
-		i--
-		buf[i] = byte(exp%10 + '0')
-		exp /= 10
+	// dd or ddd
+	switch {
+	case exp < 10:
+		dst = append(dst, '0', byte(exp)+'0')
+	case exp < 100:
+		dst = append(dst, byte(exp/10)+'0', byte(exp%10)+'0')
+	default:
+		dst = append(dst, byte(exp/100)+'0', byte(exp/10)%10+'0', byte(exp%10)+'0')
 	}
-	// exp < 10
-	i--
-	buf[i] = byte(exp + '0')
 
-	switch i {
-	case 0:
-		dst = append(dst, buf[0], buf[1], buf[2])
-	case 1:
-		dst = append(dst, buf[1], buf[2])
-	case 2:
-		// leading zeroes
-		dst = append(dst, '0', buf[2])
-	}
 	return dst
 }
 
@@ -406,11 +394,9 @@ func fmtF(dst []byte, neg bool, d decimalSlice, prec int) []byte {
 
 	// integer, padded with zeros as needed.
 	if d.dp > 0 {
-		var i int
-		for i = 0; i < d.dp && i < d.nd; i++ {
-			dst = append(dst, d.d[i])
-		}
-		for ; i < d.dp; i++ {
+		m := min(d.nd, d.dp)
+		dst = append(dst, d.d[:m]...)
+		for ; m < d.dp; m++ {
 			dst = append(dst, '0')
 		}
 	} else {
@@ -467,6 +453,13 @@ func fmtB(dst []byte, neg bool, mant uint64, exp int, flt *floatInfo) []byte {
 	return append(dst, buf[w:]...)
 }
 
+func min(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
 func max(a, b int) int {
 	if a > b {
 		return a
diff --git a/src/sync/atomic/asm_darwin_arm.s b/src/sync/atomic/asm_darwin_arm.s
new file mode 100644
index 0000000000..36dd4835f7
--- /dev/null
+++ b/src/sync/atomic/asm_darwin_arm.s
@@ -0,0 +1,99 @@
+// Copyright 2012 The Go Authors.  All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#include "textflag.h"
+
+// Darwin/ARM atomic operations.
+
+TEXT ·CompareAndSwapInt32(SB),NOSPLIT,$0
+	B ·CompareAndSwapUint32(SB)
+
+TEXT ·CompareAndSwapUint32(SB),NOSPLIT,$0
+	B ·armCompareAndSwapUint32(SB)
+
+TEXT ·CompareAndSwapUintptr(SB),NOSPLIT,$0
+	B ·CompareAndSwapUint32(SB)
+
+TEXT ·AddInt32(SB),NOSPLIT,$0
+	B ·AddUint32(SB)
+
+TEXT ·AddUint32(SB),NOSPLIT,$0
+	B ·armAddUint32(SB)
+
+TEXT ·AddUintptr(SB),NOSPLIT,$0
+	B ·AddUint32(SB)
+
+TEXT ·SwapInt32(SB),NOSPLIT,$0
+	B ·SwapUint32(SB)
+
+TEXT ·SwapUint32(SB),NOSPLIT,$0
+	B ·armSwapUint32(SB)
+
+TEXT ·SwapUintptr(SB),NOSPLIT,$0
+	B ·SwapUint32(SB)
+
+TEXT ·CompareAndSwapInt64(SB),NOSPLIT,$0
+	B ·CompareAndSwapUint64(SB)
+
+TEXT ·CompareAndSwapUint64(SB),NOSPLIT,$-4
+	B ·armCompareAndSwapUint64(SB)
+
+TEXT ·AddInt64(SB),NOSPLIT,$0
+	B ·addUint64(SB)
+
+TEXT ·AddUint64(SB),NOSPLIT,$0
+	B ·addUint64(SB)
+
+TEXT ·SwapInt64(SB),NOSPLIT,$0
+	B ·swapUint64(SB)
+
+TEXT ·SwapUint64(SB),NOSPLIT,$0
+	B ·swapUint64(SB)
+
+TEXT ·LoadInt32(SB),NOSPLIT,$0
+	B ·LoadUint32(SB)
+
+TEXT ·LoadUint32(SB),NOSPLIT,$0-8
+	MOVW addr+0(FP), R1
+load32loop:
+	LDREX (R1), R2		// loads R2
+	STREX R2, (R1), R0	// stores R2
+	CMP $0, R0
+	BNE load32loop
+	MOVW R2, val+4(FP)
+	RET
+
+TEXT ·LoadInt64(SB),NOSPLIT,$0
+	B ·loadUint64(SB)
+
+TEXT ·LoadUint64(SB),NOSPLIT,$0
+	B ·loadUint64(SB)
+
+TEXT ·LoadUintptr(SB),NOSPLIT,$0
+	B ·LoadUint32(SB)
+
+TEXT ·LoadPointer(SB),NOSPLIT,$0
+	B ·LoadUint32(SB)
+
+TEXT ·StoreInt32(SB),NOSPLIT,$0
+	B ·StoreUint32(SB)
+
+TEXT ·StoreUint32(SB),NOSPLIT,$0-8
+	MOVW addr+0(FP), R1
+	MOVW val+4(FP), R2
+storeloop:
+	LDREX (R1), R4		// loads R4
+	STREX R2, (R1), R0	// stores R2
+	CMP $0, R0
+	BNE storeloop
+	RET
+
+TEXT ·StoreInt64(SB),NOSPLIT,$0
+	B ·storeUint64(SB)
+
+TEXT ·StoreUint64(SB),NOSPLIT,$0
+	B ·storeUint64(SB)
+
+TEXT ·StoreUintptr(SB),NOSPLIT,$0
+	B ·StoreUint32(SB)
diff --git a/src/syscall/asm_darwin_arm.s b/src/syscall/asm_darwin_arm.s
new file mode 100644
index 0000000000..f75aa19086
--- /dev/null
+++ b/src/syscall/asm_darwin_arm.s
@@ -0,0 +1,134 @@
+// Copyright 2014 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#include "textflag.h"
+
+//
+// System call support for ARM, Darwin
+//
+
+// func Syscall(syscall uintptr, a1, a2, a3 uintptr) (r1, r2, err uintptr)
+TEXT	·Syscall(SB),NOSPLIT,$0-32
+	BL		runtime·entersyscall(SB)
+	MOVW	4(SP), R12
+	MOVW	8(SP), R0
+	MOVW	12(SP), R1
+	MOVW	16(SP), R2
+	SWI		$0x80
+	BCC		ok
+	MOVW	$-1, R1
+	MOVW	R1, 20(SP)	// r1
+	MOVW	$0, R2
+	MOVW	R2, 24(SP)	// r2
+	MOVW	R0, 28(SP)	// errno
+	BL		runtime·exitsyscall(SB)
+	RET
+ok:
+	MOVW	R0, 20(SP) // r1
+	MOVW	R1, 24(SP)	// r2
+	MOVW	$0, R0
+	MOVW	R0, 28(SP)	// errno
+	BL		runtime·exitsyscall(SB)
+	RET
+
+// func RawSyscall(trap uintptr, a1, a2, a3 uintptr) (r1, r2, err uintptr)
+TEXT ·RawSyscall(SB),NOSPLIT,$0-32
+	MOVW	4(SP), R12	// syscall entry
+	MOVW	8(SP), R0
+	MOVW	12(SP), R1
+	MOVW	16(SP), R2
+	SWI		$0x80
+	BCC		ok1
+	MOVW	$-1, R1
+	MOVW	R1, 20(SP)	// r1
+	MOVW	$0, R2
+	MOVW	R2, 24(SP)	// r2
+	MOVW	R0, 28(SP)	// errno
+	RET
+ok1:
+	MOVW	R0, 20(SP) // r1
+	MOVW	R1, 24(SP)	// r2
+	MOVW	$0, R0
+	MOVW	R0, 28(SP)	// errno
+	RET
+
+// func Syscall6(trap uintptr, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, err uintptr)
+TEXT	·Syscall6(SB),NOSPLIT,$0-44
+	BL		runtime·entersyscall(SB)
+	MOVW	4(SP), R12	// syscall entry
+	MOVW	8(SP), R0
+	MOVW	12(SP), R1
+	MOVW	16(SP), R2
+	MOVW	20(SP), R3
+	MOVW	24(SP), R4
+	MOVW	28(SP), R5
+	SWI		$0x80
+	BCC		ok6
+	MOVW	$-1, R1
+	MOVW	R1, 32(SP)	// r1
+	MOVW	$0, R2
+	MOVW	R2, 36(SP)	// r2
+	MOVW	R0, 40(SP)	// errno
+	BL		runtime·exitsyscall(SB)
+	RET
+ok6:
+	MOVW	R0, 32(SP) // r1
+	MOVW	R1, 36(SP)	// r2
+	MOVW	$0, R0
+	MOVW	R0, 40(SP)	// errno
+	BL		runtime·exitsyscall(SB)
+	RET
+
+// func RawSyscall6(trap uintptr, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, err uintptr)
+TEXT	·RawSyscall6(SB),NOSPLIT,$0-44
+	MOVW	4(SP), R12	// syscall entry
+	MOVW	8(SP), R0
+	MOVW	12(SP), R1
+	MOVW	16(SP), R2
+	MOVW	20(SP), R3
+	MOVW	24(SP), R4
+	MOVW	28(SP), R5
+	SWI		$0x80
+	BCC		ok2
+	MOVW	$-1, R1
+	MOVW	R1, 32(SP)	// r1
+	MOVW	$0, R2
+	MOVW	R2, 36(SP)	// r2
+	MOVW	R0, 40(SP)	// errno
+	RET
+ok2:
+	MOVW	R0, 32(SP) // r1
+	MOVW	R1, 36(SP)	// r2
+	MOVW	$0, R0
+	MOVW	R0, 40(SP)	// errno
+	RET
+
+// Actually Syscall7.
+TEXT	·Syscall9(SB),NOSPLIT,$0-56
+	BL runtime·entersyscall(SB)
+	MOVW	4(SP), R12	// syscall entry
+	MOVW	8(SP), R0
+	MOVW	12(SP), R1
+	MOVW	16(SP), R2
+	MOVW	20(SP), R3
+	MOVW	24(SP), R4
+	MOVW	28(SP), R5
+	MOVW	32(SP), R6
+	SWI		$0x80
+	BCC		ok9
+	MOVW	$-1, R1
+	MOVW	R1, 44(SP)	// r1
+	MOVW	$0, R2
+	MOVW	R2, 48(SP)	// r2
+	MOVW	R0, 52(SP)	// errno
+	BL		runtime·exitsyscall(SB)
+	RET
+ok9:
+	MOVW	R0, 44(SP) // r1
+	MOVW	R1, 48(SP)	// r2
+	MOVW	$0, R0
+	MOVW	R0, 52(SP)	// errno
+	BL	runtime·exitsyscall(SB)
+	RET
+
diff --git a/src/syscall/syscall_bsd.go b/src/syscall/syscall_bsd.go
index 2556fa8746..af563910b1 100644
--- a/src/syscall/syscall_bsd.go
+++ b/src/syscall/syscall_bsd.go
@@ -68,40 +68,7 @@ func ReadDirent(fd int, buf []byte) (n int, err error) {
 	// actual system call is getdirentries64, 64 is a good guess.
 	// TODO(rsc): Can we use a single global basep for all calls?
 	var base = (*uintptr)(unsafe.Pointer(new(uint64)))
-	n, err = Getdirentries(fd, buf, base)
-
-	// On OS X 10.10 Yosemite, if you have a directory that can be returned
-	// in a single getdirentries64 call (for example, a directory with one file),
-	// and you read from the directory at EOF twice, you get EOF both times:
-	//	fd = open("dir")
-	//	getdirentries64(fd) // returns data
-	//	getdirentries64(fd) // returns 0 (EOF)
-	//	getdirentries64(fd) // returns 0 (EOF)
-	//
-	// But if you remove the file in the middle between the two calls, the
-	// second call returns an error instead.
-	//	fd = open("dir")
-	//	getdirentries64(fd) // returns data
-	//	getdirentries64(fd) // returns 0 (EOF)
-	//	remove("dir/file")
-	//	getdirentries64(fd) // returns ENOENT/EINVAL
-	//
-	// Whether you get ENOENT or EINVAL depends on exactly what was
-	// in the directory. It is deterministic, just data-dependent.
-	//
-	// This only happens in small directories. A directory containing more data
-	// than fits in a 4k getdirentries64 call will return EOF correctly.
-	// (It's not clear if the criteria is that the directory be split across multiple
-	// getdirentries64 calls or that it be split across multiple file system blocks.)
-	//
-	// We could change package os to avoid the second read at EOF,
-	// and maybe we should, but that's a bit involved.
-	// For now, treat the EINVAL/ENOENT as EOF.
-	if runtime.GOOS == "darwin" && (err == EINVAL || err == ENOENT) {
-		err = nil
-	}
-
-	return
+	return Getdirentries(fd, buf, base)
 }
 
 // Wait status is 7 bits at bottom, either 0 (exited),
diff --git a/src/syscall/syscall_darwin_arm.go b/src/syscall/syscall_darwin_arm.go
new file mode 100644
index 0000000000..2a7d4f2db6
--- /dev/null
+++ b/src/syscall/syscall_darwin_arm.go
@@ -0,0 +1,70 @@
+// Copyright 2014 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package syscall
+
+import "unsafe"
+
+func Getpagesize() int { return 4096 }
+
+func TimespecToNsec(ts Timespec) int64 { return int64(ts.Sec)*1e9 + int64(ts.Nsec) }
+
+func NsecToTimespec(nsec int64) (ts Timespec) {
+	ts.Sec = int32(nsec / 1e9)
+	ts.Nsec = int32(nsec % 1e9)
+	return
+}
+
+func TimevalToNsec(tv Timeval) int64 { return int64(tv.Sec)*1e9 + int64(tv.Usec)*1e3 }
+
+func NsecToTimeval(nsec int64) (tv Timeval) {
+	nsec += 999 // round up to microsecond
+	tv.Usec = int32(nsec % 1e9 / 1e3)
+	tv.Sec = int32(nsec / 1e9)
+	return
+}
+
+//sysnb	gettimeofday(tp *Timeval) (sec int32, usec int32, err error)
+func Gettimeofday(tv *Timeval) (err error) {
+	// The tv passed to gettimeofday must be non-nil
+	// but is otherwise unused.  The answers come back
+	// in the two registers.
+	sec, usec, err := gettimeofday(tv)
+	tv.Sec = int32(sec)
+	tv.Usec = int32(usec)
+	return err
+}
+
+func SetKevent(k *Kevent_t, fd, mode, flags int) {
+	k.Ident = uint32(fd)
+	k.Filter = int16(mode)
+	k.Flags = uint16(flags)
+}
+
+func (iov *Iovec) SetLen(length int) {
+	iov.Len = uint32(length)
+}
+
+func (msghdr *Msghdr) SetControllen(length int) {
+	msghdr.Controllen = uint32(length)
+}
+
+func (cmsg *Cmsghdr) SetLen(length int) {
+	cmsg.Len = uint32(length)
+}
+
+func sendfile(outfd int, infd int, offset *int64, count int) (written int, err error) {
+	var length = uint64(count)
+
+	_, _, e1 := Syscall9(SYS_SENDFILE, uintptr(infd), uintptr(outfd), uintptr(*offset), uintptr(*offset>>32), uintptr(unsafe.Pointer(&length)), 0, 0, 0, 0)
+
+	written = int(length)
+
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+func Syscall9(num, a1, a2, a3, a4, a5, a6, a7, a8, a9 uintptr) (r1, r2 uintptr, err Errno) // sic
diff --git a/src/syscall/syscall_windows.go b/src/syscall/syscall_windows.go
index 8ac498df78..feb329f530 100644
--- a/src/syscall/syscall_windows.go
+++ b/src/syscall/syscall_windows.go
@@ -1003,13 +1003,22 @@ func Readlink(path string, buf []byte) (n int, err error) {
 	}
 
 	rdb := (*reparseDataBuffer)(unsafe.Pointer(&rdbbuf[0]))
-	if uintptr(bytesReturned) < unsafe.Sizeof(*rdb) ||
-		rdb.ReparseTag != IO_REPARSE_TAG_SYMLINK {
-		// the path is not a symlink but another type of reparse point
+	var s string
+	switch rdb.ReparseTag {
+	case IO_REPARSE_TAG_SYMLINK:
+		data := (*symbolicLinkReparseBuffer)(unsafe.Pointer(&rdb.reparseBuffer))
+		p := (*[0xffff]uint16)(unsafe.Pointer(&data.PathBuffer[0]))
+		s = UTF16ToString(p[data.PrintNameOffset/2 : (data.PrintNameLength-data.PrintNameOffset)/2])
+	case _IO_REPARSE_TAG_MOUNT_POINT:
+		data := (*mountPointReparseBuffer)(unsafe.Pointer(&rdb.reparseBuffer))
+		p := (*[0xffff]uint16)(unsafe.Pointer(&data.PathBuffer[0]))
+		s = UTF16ToString(p[data.PrintNameOffset/2 : (data.PrintNameLength-data.PrintNameOffset)/2])
+	default:
+		// the path is not a symlink or junction but another type of reparse
+		// point
 		return -1, ENOENT
 	}
-
-	s := UTF16ToString((*[0xffff]uint16)(unsafe.Pointer(&rdb.PathBuffer[0]))[:rdb.PrintNameLength/2])
 	n = copy(buf, []byte(s))
+
 	return n, nil
 }
diff --git a/src/syscall/zerrors_darwin_arm.go b/src/syscall/zerrors_darwin_arm.go
new file mode 100644
index 0000000000..7e800d4259
--- /dev/null
+++ b/src/syscall/zerrors_darwin_arm.go
@@ -0,0 +1,1349 @@
+// mkerrors.sh
+// MACHINE GENERATED BY THE COMMAND ABOVE; DO NOT EDIT
+
+// Created by cgo -godefs - DO NOT EDIT
+// cgo -godefs -- _const.go
+
+package syscall
+
+const (
+	AF_APPLETALK                = 0x10
+	AF_CCITT                    = 0xa
+	AF_CHAOS                    = 0x5
+	AF_CNT                      = 0x15
+	AF_COIP                     = 0x14
+	AF_DATAKIT                  = 0x9
+	AF_DECnet                   = 0xc
+	AF_DLI                      = 0xd
+	AF_E164                     = 0x1c
+	AF_ECMA                     = 0x8
+	AF_HYLINK                   = 0xf
+	AF_IMPLINK                  = 0x3
+	AF_INET                     = 0x2
+	AF_INET6                    = 0x1e
+	AF_IPX                      = 0x17
+	AF_ISDN                     = 0x1c
+	AF_ISO                      = 0x7
+	AF_LAT                      = 0xe
+	AF_LINK                     = 0x12
+	AF_LOCAL                    = 0x1
+	AF_MAX                      = 0x25
+	AF_NATM                     = 0x1f
+	AF_NDRV                     = 0x1b
+	AF_NETBIOS                  = 0x21
+	AF_NS                       = 0x6
+	AF_OSI                      = 0x7
+	AF_PPP                      = 0x22
+	AF_PUP                      = 0x4
+	AF_RESERVED_36              = 0x24
+	AF_ROUTE                    = 0x11
+	AF_SIP                      = 0x18
+	AF_SNA                      = 0xb
+	AF_SYSTEM                   = 0x20
+	AF_UNIX                     = 0x1
+	AF_UNSPEC                   = 0x0
+	B0                          = 0x0
+	B110                        = 0x6e
+	B115200                     = 0x1c200
+	B1200                       = 0x4b0
+	B134                        = 0x86
+	B14400                      = 0x3840
+	B150                        = 0x96
+	B1800                       = 0x708
+	B19200                      = 0x4b00
+	B200                        = 0xc8
+	B230400                     = 0x38400
+	B2400                       = 0x960
+	B28800                      = 0x7080
+	B300                        = 0x12c
+	B38400                      = 0x9600
+	B4800                       = 0x12c0
+	B50                         = 0x32
+	B57600                      = 0xe100
+	B600                        = 0x258
+	B7200                       = 0x1c20
+	B75                         = 0x4b
+	B76800                      = 0x12c00
+	B9600                       = 0x2580
+	BIOCFLUSH                   = 0x20004268
+	BIOCGBLEN                   = 0x40044266
+	BIOCGDLT                    = 0x4004426a
+	BIOCGDLTLIST                = 0xc00c4279
+	BIOCGETIF                   = 0x4020426b
+	BIOCGHDRCMPLT               = 0x40044274
+	BIOCGRSIG                   = 0x40044272
+	BIOCGRTIMEOUT               = 0x4008426e
+	BIOCGSEESENT                = 0x40044276
+	BIOCGSTATS                  = 0x4008426f
+	BIOCIMMEDIATE               = 0x80044270
+	BIOCPROMISC                 = 0x20004269
+	BIOCSBLEN                   = 0xc0044266
+	BIOCSDLT                    = 0x80044278
+	BIOCSETF                    = 0x80084267
+	BIOCSETIF                   = 0x8020426c
+	BIOCSHDRCMPLT               = 0x80044275
+	BIOCSRSIG                   = 0x80044273
+	BIOCSRTIMEOUT               = 0x8008426d
+	BIOCSSEESENT                = 0x80044277
+	BIOCVERSION                 = 0x40044271
+	BPF_A                       = 0x10
+	BPF_ABS                     = 0x20
+	BPF_ADD                     = 0x0
+	BPF_ALIGNMENT               = 0x4
+	BPF_ALU                     = 0x4
+	BPF_AND                     = 0x50
+	BPF_B                       = 0x10
+	BPF_DIV                     = 0x30
+	BPF_H                       = 0x8
+	BPF_IMM                     = 0x0
+	BPF_IND                     = 0x40
+	BPF_JA                      = 0x0
+	BPF_JEQ                     = 0x10
+	BPF_JGE                     = 0x30
+	BPF_JGT                     = 0x20
+	BPF_JMP                     = 0x5
+	BPF_JSET                    = 0x40
+	BPF_K                       = 0x0
+	BPF_LD                      = 0x0
+	BPF_LDX                     = 0x1
+	BPF_LEN                     = 0x80
+	BPF_LSH                     = 0x60
+	BPF_MAJOR_VERSION           = 0x1
+	BPF_MAXBUFSIZE              = 0x80000
+	BPF_MAXINSNS                = 0x200
+	BPF_MEM                     = 0x60
+	BPF_MEMWORDS                = 0x10
+	BPF_MINBUFSIZE              = 0x20
+	BPF_MINOR_VERSION           = 0x1
+	BPF_MISC                    = 0x7
+	BPF_MSH                     = 0xa0
+	BPF_MUL                     = 0x20
+	BPF_NEG                     = 0x80
+	BPF_OR                      = 0x40
+	BPF_RELEASE                 = 0x30bb6
+	BPF_RET                     = 0x6
+	BPF_RSH                     = 0x70
+	BPF_ST                      = 0x2
+	BPF_STX                     = 0x3
+	BPF_SUB                     = 0x10
+	BPF_TAX                     = 0x0
+	BPF_TXA                     = 0x80
+	BPF_W                       = 0x0
+	BPF_X                       = 0x8
+	BRKINT                      = 0x2
+	CFLUSH                      = 0xf
+	CLOCAL                      = 0x8000
+	CREAD                       = 0x800
+	CS5                         = 0x0
+	CS6                         = 0x100
+	CS7                         = 0x200
+	CS8                         = 0x300
+	CSIZE                       = 0x300
+	CSTART                      = 0x11
+	CSTATUS                     = 0x14
+	CSTOP                       = 0x13
+	CSTOPB                      = 0x400
+	CSUSP                       = 0x1a
+	CTL_MAXNAME                 = 0xc
+	CTL_NET                     = 0x4
+	DLT_APPLE_IP_OVER_IEEE1394  = 0x8a
+	DLT_ARCNET                  = 0x7
+	DLT_ATM_CLIP                = 0x13
+	DLT_ATM_RFC1483             = 0xb
+	DLT_AX25                    = 0x3
+	DLT_CHAOS                   = 0x5
+	DLT_CHDLC                   = 0x68
+	DLT_C_HDLC                  = 0x68
+	DLT_EN10MB                  = 0x1
+	DLT_EN3MB                   = 0x2
+	DLT_FDDI                    = 0xa
+	DLT_IEEE802                 = 0x6
+	DLT_IEEE802_11              = 0x69
+	DLT_IEEE802_11_RADIO        = 0x7f
+	DLT_IEEE802_11_RADIO_AVS    = 0xa3
+	DLT_LINUX_SLL               = 0x71
+	DLT_LOOP                    = 0x6c
+	DLT_NULL                    = 0x0
+	DLT_PFLOG                   = 0x75
+	DLT_PFSYNC                  = 0x12
+	DLT_PPP                     = 0x9
+	DLT_PPP_BSDOS               = 0x10
+	DLT_PPP_SERIAL              = 0x32
+	DLT_PRONET                  = 0x4
+	DLT_RAW                     = 0xc
+	DLT_SLIP                    = 0x8
+	DLT_SLIP_BSDOS              = 0xf
+	DT_BLK                      = 0x6
+	DT_CHR                      = 0x2
+	DT_DIR                      = 0x4
+	DT_FIFO                     = 0x1
+	DT_LNK                      = 0xa
+	DT_REG                      = 0x8
+	DT_SOCK                     = 0xc
+	DT_UNKNOWN                  = 0x0
+	DT_WHT                      = 0xe
+	ECHO                        = 0x8
+	ECHOCTL                     = 0x40
+	ECHOE                       = 0x2
+	ECHOK                       = 0x4
+	ECHOKE                      = 0x1
+	ECHONL                      = 0x10
+	ECHOPRT                     = 0x20
+	EVFILT_AIO                  = -0x3
+	EVFILT_FS                   = -0x9
+	EVFILT_MACHPORT             = -0x8
+	EVFILT_PROC                 = -0x5
+	EVFILT_READ                 = -0x1
+	EVFILT_SIGNAL               = -0x6
+	EVFILT_SYSCOUNT             = 0x9
+	EVFILT_THREADMARKER         = 0x9
+	EVFILT_TIMER                = -0x7
+	EVFILT_VNODE                = -0x4
+	EVFILT_WRITE                = -0x2
+	EV_ADD                      = 0x1
+	EV_CLEAR                    = 0x20
+	EV_DELETE                   = 0x2
+	EV_DISABLE                  = 0x8
+	EV_ENABLE                   = 0x4
+	EV_EOF                      = 0x8000
+	EV_ERROR                    = 0x4000
+	EV_FLAG0                    = 0x1000
+	EV_FLAG1                    = 0x2000
+	EV_ONESHOT                  = 0x10
+	EV_OOBAND                   = 0x2000
+	EV_POLL                     = 0x1000
+	EV_RECEIPT                  = 0x40
+	EV_SYSFLAGS                 = 0xf000
+	EXTA                        = 0x4b00
+	EXTB                        = 0x9600
+	EXTPROC                     = 0x800
+	FD_CLOEXEC                  = 0x1
+	FD_SETSIZE                  = 0x400
+	FLUSHO                      = 0x800000
+	F_ADDSIGS                   = 0x3b
+	F_ALLOCATEALL               = 0x4
+	F_ALLOCATECONTIG            = 0x2
+	F_CHKCLEAN                  = 0x29
+	F_DUPFD                     = 0x0
+	F_DUPFD_CLOEXEC             = 0x43
+	F_FREEZE_FS                 = 0x35
+	F_FULLFSYNC                 = 0x33
+	F_GETFD                     = 0x1
+	F_GETFL                     = 0x3
+	F_GETLK                     = 0x7
+	F_GETOWN                    = 0x5
+	F_GETPATH                   = 0x32
+	F_GLOBAL_NOCACHE            = 0x37
+	F_LOG2PHYS                  = 0x31
+	F_MARKDEPENDENCY            = 0x3c
+	F_NOCACHE                   = 0x30
+	F_OK                        = 0x0
+	F_PATHPKG_CHECK             = 0x34
+	F_PEOFPOSMODE               = 0x3
+	F_PREALLOCATE               = 0x2a
+	F_RDADVISE                  = 0x2c
+	F_RDAHEAD                   = 0x2d
+	F_RDLCK                     = 0x1
+	F_READBOOTSTRAP             = 0x2e
+	F_SETFD                     = 0x2
+	F_SETFL                     = 0x4
+	F_SETLK                     = 0x8
+	F_SETLKW                    = 0x9
+	F_SETOWN                    = 0x6
+	F_SETSIZE                   = 0x2b
+	F_THAW_FS                   = 0x36
+	F_UNLCK                     = 0x2
+	F_VOLPOSMODE                = 0x4
+	F_WRITEBOOTSTRAP            = 0x2f
+	F_WRLCK                     = 0x3
+	HUPCL                       = 0x4000
+	ICANON                      = 0x100
+	ICRNL                       = 0x100
+	IEXTEN                      = 0x400
+	IFF_ALLMULTI                = 0x200
+	IFF_ALTPHYS                 = 0x4000
+	IFF_BROADCAST               = 0x2
+	IFF_DEBUG                   = 0x4
+	IFF_LINK0                   = 0x1000
+	IFF_LINK1                   = 0x2000
+	IFF_LINK2                   = 0x4000
+	IFF_LOOPBACK                = 0x8
+	IFF_MULTICAST               = 0x8000
+	IFF_NOARP                   = 0x80
+	IFF_NOTRAILERS              = 0x20
+	IFF_OACTIVE                 = 0x400
+	IFF_POINTOPOINT             = 0x10
+	IFF_PROMISC                 = 0x100
+	IFF_RUNNING                 = 0x40
+	IFF_SIMPLEX                 = 0x800
+	IFF_UP                      = 0x1
+	IFNAMSIZ                    = 0x10
+	IFT_1822                    = 0x2
+	IFT_AAL5                    = 0x31
+	IFT_ARCNET                  = 0x23
+	IFT_ARCNETPLUS              = 0x24
+	IFT_ATM                     = 0x25
+	IFT_BRIDGE                  = 0xd1
+	IFT_CARP                    = 0xf8
+	IFT_CEPT                    = 0x13
+	IFT_DS3                     = 0x1e
+	IFT_ENC                     = 0xf4
+	IFT_EON                     = 0x19
+	IFT_ETHER                   = 0x6
+	IFT_FAITH                   = 0x38
+	IFT_FDDI                    = 0xf
+	IFT_FRELAY                  = 0x20
+	IFT_FRELAYDCE               = 0x2c
+	IFT_GIF                     = 0x37
+	IFT_HDH1822                 = 0x3
+	IFT_HIPPI                   = 0x2f
+	IFT_HSSI                    = 0x2e
+	IFT_HY                      = 0xe
+	IFT_IEEE1394                = 0x90
+	IFT_IEEE8023ADLAG           = 0x88
+	IFT_ISDNBASIC               = 0x14
+	IFT_ISDNPRIMARY             = 0x15
+	IFT_ISO88022LLC             = 0x29
+	IFT_ISO88023                = 0x7
+	IFT_ISO88024                = 0x8
+	IFT_ISO88025                = 0x9
+	IFT_ISO88026                = 0xa
+	IFT_L2VLAN                  = 0x87
+	IFT_LAPB                    = 0x10
+	IFT_LOCALTALK               = 0x2a
+	IFT_LOOP                    = 0x18
+	IFT_MIOX25                  = 0x26
+	IFT_MODEM                   = 0x30
+	IFT_NSIP                    = 0x1b
+	IFT_OTHER                   = 0x1
+	IFT_P10                     = 0xc
+	IFT_P80                     = 0xd
+	IFT_PARA                    = 0x22
+	IFT_PDP                     = 0xff
+	IFT_PFLOG                   = 0xf5
+	IFT_PFSYNC                  = 0xf6
+	IFT_PPP                     = 0x17
+	IFT_PROPMUX                 = 0x36
+	IFT_PROPVIRTUAL             = 0x35
+	IFT_PTPSERIAL               = 0x16
+	IFT_RS232                   = 0x21
+	IFT_SDLC                    = 0x11
+	IFT_SIP                     = 0x1f
+	IFT_SLIP                    = 0x1c
+	IFT_SMDSDXI                 = 0x2b
+	IFT_SMDSICIP                = 0x34
+	IFT_SONET                   = 0x27
+	IFT_SONETPATH               = 0x32
+	IFT_SONETVT                 = 0x33
+	IFT_STARLAN                 = 0xb
+	IFT_STF                     = 0x39
+	IFT_T1                      = 0x12
+	IFT_ULTRA                   = 0x1d
+	IFT_V35                     = 0x2d
+	IFT_X25                     = 0x5
+	IFT_X25DDN                  = 0x4
+	IFT_X25PLE                  = 0x28
+	IFT_XETHER                  = 0x1a
+	IGNBRK                      = 0x1
+	IGNCR                       = 0x80
+	IGNPAR                      = 0x4
+	IMAXBEL                     = 0x2000
+	INLCR                       = 0x40
+	INPCK                       = 0x10
+	IN_CLASSA_HOST              = 0xffffff
+	IN_CLASSA_MAX               = 0x80
+	IN_CLASSA_NET               = 0xff000000
+	IN_CLASSA_NSHIFT            = 0x18
+	IN_CLASSB_HOST              = 0xffff
+	IN_CLASSB_MAX               = 0x10000
+	IN_CLASSB_NET               = 0xffff0000
+	IN_CLASSB_NSHIFT            = 0x10
+	IN_CLASSC_HOST              = 0xff
+	IN_CLASSC_NET               = 0xffffff00
+	IN_CLASSC_NSHIFT            = 0x8
+	IN_CLASSD_HOST              = 0xfffffff
+	IN_CLASSD_NET               = 0xf0000000
+	IN_CLASSD_NSHIFT            = 0x1c
+	IN_LINKLOCALNETNUM          = 0xa9fe0000
+	IN_LOOPBACKNET              = 0x7f
+	IPPROTO_3PC                 = 0x22
+	IPPROTO_ADFS                = 0x44
+	IPPROTO_AH                  = 0x33
+	IPPROTO_AHIP                = 0x3d
+	IPPROTO_APES                = 0x63
+	IPPROTO_ARGUS               = 0xd
+	IPPROTO_AX25                = 0x5d
+	IPPROTO_BHA                 = 0x31
+	IPPROTO_BLT                 = 0x1e
+	IPPROTO_BRSATMON            = 0x4c
+	IPPROTO_CFTP                = 0x3e
+	IPPROTO_CHAOS               = 0x10
+	IPPROTO_CMTP                = 0x26
+	IPPROTO_CPHB                = 0x49
+	IPPROTO_CPNX                = 0x48
+	IPPROTO_DDP                 = 0x25
+	IPPROTO_DGP                 = 0x56
+	IPPROTO_DIVERT              = 0xfe
+	IPPROTO_DONE                = 0x101
+	IPPROTO_DSTOPTS             = 0x3c
+	IPPROTO_EGP                 = 0x8
+	IPPROTO_EMCON               = 0xe
+	IPPROTO_ENCAP               = 0x62
+	IPPROTO_EON                 = 0x50
+	IPPROTO_ESP                 = 0x32
+	IPPROTO_ETHERIP             = 0x61
+	IPPROTO_FRAGMENT            = 0x2c
+	IPPROTO_GGP                 = 0x3
+	IPPROTO_GMTP                = 0x64
+	IPPROTO_GRE                 = 0x2f
+	IPPROTO_HELLO               = 0x3f
+	IPPROTO_HMP                 = 0x14
+	IPPROTO_HOPOPTS             = 0x0
+	IPPROTO_ICMP                = 0x1
+	IPPROTO_ICMPV6              = 0x3a
+	IPPROTO_IDP                 = 0x16
+	IPPROTO_IDPR                = 0x23
+	IPPROTO_IDRP                = 0x2d
+	IPPROTO_IGMP                = 0x2
+	IPPROTO_IGP                 = 0x55
+	IPPROTO_IGRP                = 0x58
+	IPPROTO_IL                  = 0x28
+	IPPROTO_INLSP               = 0x34
+	IPPROTO_INP                 = 0x20
+	IPPROTO_IP                  = 0x0
+	IPPROTO_IPCOMP              = 0x6c
+	IPPROTO_IPCV                = 0x47
+	IPPROTO_IPEIP               = 0x5e
+	IPPROTO_IPIP                = 0x4
+	IPPROTO_IPPC                = 0x43
+	IPPROTO_IPV4                = 0x4
+	IPPROTO_IPV6                = 0x29
+	IPPROTO_IRTP                = 0x1c
+	IPPROTO_KRYPTOLAN           = 0x41
+	IPPROTO_LARP                = 0x5b
+	IPPROTO_LEAF1               = 0x19
+	IPPROTO_LEAF2               = 0x1a
+	IPPROTO_MAX                 = 0x100
+	IPPROTO_MAXID               = 0x34
+	IPPROTO_MEAS                = 0x13
+	IPPROTO_MHRP                = 0x30
+	IPPROTO_MICP                = 0x5f
+	IPPROTO_MTP                 = 0x5c
+	IPPROTO_MUX                 = 0x12
+	IPPROTO_ND                  = 0x4d
+	IPPROTO_NHRP                = 0x36
+	IPPROTO_NONE                = 0x3b
+	IPPROTO_NSP                 = 0x1f
+	IPPROTO_NVPII               = 0xb
+	IPPROTO_OSPFIGP             = 0x59
+	IPPROTO_PGM                 = 0x71
+	IPPROTO_PIGP                = 0x9
+	IPPROTO_PIM                 = 0x67
+	IPPROTO_PRM                 = 0x15
+	IPPROTO_PUP                 = 0xc
+	IPPROTO_PVP                 = 0x4b
+	IPPROTO_RAW                 = 0xff
+	IPPROTO_RCCMON              = 0xa
+	IPPROTO_RDP                 = 0x1b
+	IPPROTO_ROUTING             = 0x2b
+	IPPROTO_RSVP                = 0x2e
+	IPPROTO_RVD                 = 0x42
+	IPPROTO_SATEXPAK            = 0x40
+	IPPROTO_SATMON              = 0x45
+	IPPROTO_SCCSP               = 0x60
+	IPPROTO_SDRP                = 0x2a
+	IPPROTO_SEP                 = 0x21
+	IPPROTO_SRPC                = 0x5a
+	IPPROTO_ST                  = 0x7
+	IPPROTO_SVMTP               = 0x52
+	IPPROTO_SWIPE               = 0x35
+	IPPROTO_TCF                 = 0x57
+	IPPROTO_TCP                 = 0x6
+	IPPROTO_TP                  = 0x1d
+	IPPROTO_TPXX                = 0x27
+	IPPROTO_TRUNK1              = 0x17
+	IPPROTO_TRUNK2              = 0x18
+	IPPROTO_TTP                 = 0x54
+	IPPROTO_UDP                 = 0x11
+	IPPROTO_VINES               = 0x53
+	IPPROTO_VISA                = 0x46
+	IPPROTO_VMTP                = 0x51
+	IPPROTO_WBEXPAK             = 0x4f
+	IPPROTO_WBMON               = 0x4e
+	IPPROTO_WSN                 = 0x4a
+	IPPROTO_XNET                = 0xf
+	IPPROTO_XTP                 = 0x24
+	IPV6_BINDV6ONLY             = 0x1b
+	IPV6_CHECKSUM               = 0x1a
+	IPV6_DEFAULT_MULTICAST_HOPS = 0x1
+	IPV6_DEFAULT_MULTICAST_LOOP = 0x1
+	IPV6_DEFHLIM                = 0x40
+	IPV6_DSTOPTS                = 0x17
+	IPV6_FAITH                  = 0x1d
+	IPV6_FLOWINFO_MASK          = 0xffffff0f
+	IPV6_FLOWLABEL_MASK         = 0xffff0f00
+	IPV6_FRAGTTL                = 0x78
+	IPV6_FW_ADD                 = 0x1e
+	IPV6_FW_DEL                 = 0x1f
+	IPV6_FW_FLUSH               = 0x20
+	IPV6_FW_GET                 = 0x22
+	IPV6_FW_ZERO                = 0x21
+	IPV6_HLIMDEC                = 0x1
+	IPV6_HOPLIMIT               = 0x14
+	IPV6_HOPOPTS                = 0x16
+	IPV6_IPSEC_POLICY           = 0x1c
+	IPV6_JOIN_GROUP             = 0xc
+	IPV6_LEAVE_GROUP            = 0xd
+	IPV6_MAXHLIM                = 0xff
+	IPV6_MAXPACKET              = 0xffff
+	IPV6_MMTU                   = 0x500
+	IPV6_MULTICAST_HOPS         = 0xa
+	IPV6_MULTICAST_IF           = 0x9
+	IPV6_MULTICAST_LOOP         = 0xb
+	IPV6_NEXTHOP                = 0x15
+	IPV6_PKTINFO                = 0x13
+	IPV6_PKTOPTIONS             = 0x19
+	IPV6_PORTRANGE              = 0xe
+	IPV6_PORTRANGE_DEFAULT      = 0x0
+	IPV6_PORTRANGE_HIGH         = 0x1
+	IPV6_PORTRANGE_LOW          = 0x2
+	IPV6_RTHDR                  = 0x18
+	IPV6_RTHDR_LOOSE            = 0x0
+	IPV6_RTHDR_STRICT           = 0x1
+	IPV6_RTHDR_TYPE_0           = 0x0
+	IPV6_SOCKOPT_RESERVED1      = 0x3
+	IPV6_UNICAST_HOPS           = 0x4
+	IPV6_V6ONLY                 = 0x1b
+	IPV6_VERSION                = 0x60
+	IPV6_VERSION_MASK           = 0xf0
+	IP_ADD_MEMBERSHIP           = 0xc
+	IP_DEFAULT_MULTICAST_LOOP   = 0x1
+	IP_DEFAULT_MULTICAST_TTL    = 0x1
+	IP_DF                       = 0x4000
+	IP_DROP_MEMBERSHIP          = 0xd
+	IP_DUMMYNET_CONFIGURE       = 0x3c
+	IP_DUMMYNET_DEL             = 0x3d
+	IP_DUMMYNET_FLUSH           = 0x3e
+	IP_DUMMYNET_GET             = 0x40
+	IP_FAITH                    = 0x16
+	IP_FW_ADD                   = 0x28
+	IP_FW_DEL                   = 0x29
+	IP_FW_FLUSH                 = 0x2a
+	IP_FW_GET                   = 0x2c
+	IP_FW_RESETLOG              = 0x2d
+	IP_FW_ZERO                  = 0x2b
+	IP_HDRINCL                  = 0x2
+	IP_IPSEC_POLICY             = 0x15
+	IP_MAXPACKET                = 0xffff
+	IP_MAX_MEMBERSHIPS          = 0x14
+	IP_MF                       = 0x2000
+	IP_MSS                      = 0x240
+	IP_MULTICAST_IF             = 0x9
+	IP_MULTICAST_LOOP           = 0xb
+	IP_MULTICAST_TTL            = 0xa
+	IP_MULTICAST_VIF            = 0xe
+	IP_NAT__XXX                 = 0x37
+	IP_OFFMASK                  = 0x1fff
+	IP_OLD_FW_ADD               = 0x32
+	IP_OLD_FW_DEL               = 0x33
+	IP_OLD_FW_FLUSH             = 0x34
+	IP_OLD_FW_GET               = 0x36
+	IP_OLD_FW_RESETLOG          = 0x38
+	IP_OLD_FW_ZERO              = 0x35
+	IP_OPTIONS                  = 0x1
+	IP_PKTINFO                  = 0x1a
+	IP_PORTRANGE                = 0x13
+	IP_PORTRANGE_DEFAULT        = 0x0
+	IP_PORTRANGE_HIGH           = 0x1
+	IP_PORTRANGE_LOW            = 0x2
+	IP_RECVDSTADDR              = 0x7
+	IP_RECVIF                   = 0x14
+	IP_RECVOPTS                 = 0x5
+	IP_RECVPKTINFO              = 0x1a
+	IP_RECVRETOPTS              = 0x6
+	IP_RECVTTL                  = 0x18
+	IP_RETOPTS                  = 0x8
+	IP_RF                       = 0x8000
+	IP_RSVP_OFF                 = 0x10
+	IP_RSVP_ON                  = 0xf
+	IP_RSVP_VIF_OFF             = 0x12
+	IP_RSVP_VIF_ON              = 0x11
+	IP_STRIPHDR                 = 0x17
+	IP_TOS                      = 0x3
+	IP_TRAFFIC_MGT_BACKGROUND   = 0x41
+	IP_TTL                      = 0x4
+	ISIG                        = 0x80
+	ISTRIP                      = 0x20
+	IUTF8                       = 0x4000
+	IXANY                       = 0x800
+	IXOFF                       = 0x400
+	IXON                        = 0x200
+	LOCK_EX                     = 0x2
+	LOCK_NB                     = 0x4
+	LOCK_SH                     = 0x1
+	LOCK_UN                     = 0x8
+	MADV_DONTNEED               = 0x4
+	MADV_FREE                   = 0x5
+	MADV_NORMAL                 = 0x0
+	MADV_RANDOM                 = 0x1
+	MADV_SEQUENTIAL             = 0x2
+	MADV_WILLNEED               = 0x3
+	MAP_ANON                    = 0x1000
+	MAP_COPY                    = 0x2
+	MAP_FILE                    = 0x0
+	MAP_FIXED                   = 0x10
+	MAP_HASSEMAPHORE            = 0x200
+	MAP_NOCACHE                 = 0x400
+	MAP_NOEXTEND                = 0x100
+	MAP_NORESERVE               = 0x40
+	MAP_PRIVATE                 = 0x2
+	MAP_RENAME                  = 0x20
+	MAP_RESERVED0080            = 0x80
+	MAP_SHARED                  = 0x1
+	MCL_CURRENT                 = 0x1
+	MCL_FUTURE                  = 0x2
+	MSG_CTRUNC                  = 0x20
+	MSG_DONTROUTE               = 0x4
+	MSG_DONTWAIT                = 0x80
+	MSG_EOF                     = 0x100
+	MSG_EOR                     = 0x8
+	MSG_FLUSH                   = 0x400
+	MSG_HAVEMORE                = 0x2000
+	MSG_HOLD                    = 0x800
+	MSG_NEEDSA                  = 0x10000
+	MSG_OOB                     = 0x1
+	MSG_PEEK                    = 0x2
+	MSG_RCVMORE                 = 0x4000
+	MSG_SEND                    = 0x1000
+	MSG_TRUNC                   = 0x10
+	MSG_WAITALL                 = 0x40
+	MSG_WAITSTREAM              = 0x200
+	MS_ASYNC                    = 0x1
+	MS_DEACTIVATE               = 0x8
+	MS_INVALIDATE               = 0x2
+	MS_KILLPAGES                = 0x4
+	MS_SYNC                     = 0x10
+	NAME_MAX                    = 0xff
+	NET_RT_DUMP                 = 0x1
+	NET_RT_DUMP2                = 0x7
+	NET_RT_FLAGS                = 0x2
+	NET_RT_IFLIST               = 0x3
+	NET_RT_IFLIST2              = 0x6
+	NET_RT_MAXID                = 0x8
+	NET_RT_STAT                 = 0x4
+	NET_RT_TRASH                = 0x5
+	NOFLSH                      = 0x80000000
+	NOTE_ABSOLUTE               = 0x8
+	NOTE_ATTRIB                 = 0x8
+	NOTE_CHILD                  = 0x4
+	NOTE_DELETE                 = 0x1
+	NOTE_EXEC                   = 0x20000000
+	NOTE_EXIT                   = 0x80000000
+	NOTE_EXTEND                 = 0x4
+	NOTE_FORK                   = 0x40000000
+	NOTE_LINK                   = 0x10
+	NOTE_LOWAT                  = 0x1
+	NOTE_NSECONDS               = 0x4
+	NOTE_PCTRLMASK              = -0x100000
+	NOTE_PDATAMASK              = 0xfffff
+	NOTE_REAP                   = 0x10000000
+	NOTE_RENAME                 = 0x20
+	NOTE_REVOKE                 = 0x40
+	NOTE_SECONDS                = 0x1
+	NOTE_SIGNAL                 = 0x8000000
+	NOTE_TRACK                  = 0x1
+	NOTE_TRACKERR               = 0x2
+	NOTE_USECONDS               = 0x2
+	NOTE_WRITE                  = 0x2
+	OCRNL                       = 0x10
+	OFDEL                       = 0x20000
+	OFILL                       = 0x80
+	ONLCR                       = 0x2
+	ONLRET                      = 0x40
+	ONOCR                       = 0x20
+	ONOEOT                      = 0x8
+	OPOST                       = 0x1
+	O_ACCMODE                   = 0x3
+	O_ALERT                     = 0x20000000
+	O_APPEND                    = 0x8
+	O_ASYNC                     = 0x40
+	O_CLOEXEC                   = 0x1000000
+	O_CREAT                     = 0x200
+	O_DIRECTORY                 = 0x100000
+	O_EVTONLY                   = 0x8000
+	O_EXCL                      = 0x800
+	O_EXLOCK                    = 0x20
+	O_FSYNC                     = 0x80
+	O_NDELAY                    = 0x4
+	O_NOCTTY                    = 0x20000
+	O_NOFOLLOW                  = 0x100
+	O_NONBLOCK                  = 0x4
+	O_POPUP                     = 0x80000000
+	O_RDONLY                    = 0x0
+	O_RDWR                      = 0x2
+	O_SHLOCK                    = 0x10
+	O_SYMLINK                   = 0x200000
+	O_SYNC                      = 0x80
+	O_TRUNC                     = 0x400
+	O_WRONLY                    = 0x1
+	PARENB                      = 0x1000
+	PARMRK                      = 0x8
+	PARODD                      = 0x2000
+	PENDIN                      = 0x20000000
+	PRIO_PGRP                   = 0x1
+	PRIO_PROCESS                = 0x0
+	PRIO_USER                   = 0x2
+	PROT_EXEC                   = 0x4
+	PROT_NONE                   = 0x0
+	PROT_READ                   = 0x1
+	PROT_WRITE                  = 0x2
+	PT_ATTACH                   = 0xa
+	PT_ATTACHEXC                = 0xe
+	PT_CONTINUE                 = 0x7
+	PT_DENY_ATTACH              = 0x1f
+	PT_DETACH                   = 0xb
+	PT_FIRSTMACH                = 0x20
+	PT_FORCEQUOTA               = 0x1e
+	PT_KILL                     = 0x8
+	PT_READ_D                   = 0x2
+	PT_READ_I                   = 0x1
+	PT_READ_U                   = 0x3
+	PT_SIGEXC                   = 0xc
+	PT_STEP                     = 0x9
+	PT_THUPDATE                 = 0xd
+	PT_TRACE_ME                 = 0x0
+	PT_WRITE_D                  = 0x5
+	PT_WRITE_I                  = 0x4
+	PT_WRITE_U                  = 0x6
+	RLIMIT_AS                   = 0x5
+	RLIMIT_CORE                 = 0x4
+	RLIMIT_CPU                  = 0x0
+	RLIMIT_DATA                 = 0x2
+	RLIMIT_FSIZE                = 0x1
+	RLIMIT_NOFILE               = 0x8
+	RLIMIT_STACK                = 0x3
+	RLIM_INFINITY               = 0x7fffffffffffffff
+	RTAX_AUTHOR                 = 0x6
+	RTAX_BRD                    = 0x7
+	RTAX_DST                    = 0x0
+	RTAX_GATEWAY                = 0x1
+	RTAX_GENMASK                = 0x3
+	RTAX_IFA                    = 0x5
+	RTAX_IFP                    = 0x4
+	RTAX_MAX                    = 0x8
+	RTAX_NETMASK                = 0x2
+	RTA_AUTHOR                  = 0x40
+	RTA_BRD                     = 0x80
+	RTA_DST                     = 0x1
+	RTA_GATEWAY                 = 0x2
+	RTA_GENMASK                 = 0x8
+	RTA_IFA                     = 0x20
+	RTA_IFP                     = 0x10
+	RTA_NETMASK                 = 0x4
+	RTF_BLACKHOLE               = 0x1000
+	RTF_BROADCAST               = 0x400000
+	RTF_CLONING                 = 0x100
+	RTF_CONDEMNED               = 0x2000000
+	RTF_DELCLONE                = 0x80
+	RTF_DONE                    = 0x40
+	RTF_DYNAMIC                 = 0x10
+	RTF_GATEWAY                 = 0x2
+	RTF_HOST                    = 0x4
+	RTF_IFREF                   = 0x4000000
+	RTF_IFSCOPE                 = 0x1000000
+	RTF_LLINFO                  = 0x400
+	RTF_LOCAL                   = 0x200000
+	RTF_MODIFIED                = 0x20
+	RTF_MULTICAST               = 0x800000
+	RTF_PINNED                  = 0x100000
+	RTF_PRCLONING               = 0x10000
+	RTF_PROTO1                  = 0x8000
+	RTF_PROTO2                  = 0x4000
+	RTF_PROTO3                  = 0x40000
+	RTF_REJECT                  = 0x8
+	RTF_STATIC                  = 0x800
+	RTF_UP                      = 0x1
+	RTF_WASCLONED               = 0x20000
+	RTF_XRESOLVE                = 0x200
+	RTM_ADD                     = 0x1
+	RTM_CHANGE                  = 0x3
+	RTM_DELADDR                 = 0xd
+	RTM_DELETE                  = 0x2
+	RTM_DELMADDR                = 0x10
+	RTM_GET                     = 0x4
+	RTM_GET2                    = 0x14
+	RTM_IFINFO                  = 0xe
+	RTM_IFINFO2                 = 0x12
+	RTM_LOCK                    = 0x8
+	RTM_LOSING                  = 0x5
+	RTM_MISS                    = 0x7
+	RTM_NEWADDR                 = 0xc
+	RTM_NEWMADDR                = 0xf
+	RTM_NEWMADDR2               = 0x13
+	RTM_OLDADD                  = 0x9
+	RTM_OLDDEL                  = 0xa
+	RTM_REDIRECT                = 0x6
+	RTM_RESOLVE                 = 0xb
+	RTM_RTTUNIT                 = 0xf4240
+	RTM_VERSION                 = 0x5
+	RTV_EXPIRE                  = 0x4
+	RTV_HOPCOUNT                = 0x2
+	RTV_MTU                     = 0x1
+	RTV_RPIPE                   = 0x8
+	RTV_RTT                     = 0x40
+	RTV_RTTVAR                  = 0x80
+	RTV_SPIPE                   = 0x10
+	RTV_SSTHRESH                = 0x20
+	RUSAGE_CHILDREN             = -0x1
+	RUSAGE_SELF                 = 0x0
+	SCM_CREDS                   = 0x3
+	SCM_RIGHTS                  = 0x1
+	SCM_TIMESTAMP               = 0x2
+	SHUT_RD                     = 0x0
+	SHUT_RDWR                   = 0x2
+	SHUT_WR                     = 0x1
+	SIOCADDMULTI                = 0x80206931
+	SIOCAIFADDR                 = 0x8040691a
+	SIOCALIFADDR                = 0x8118691d
+	SIOCARPIPLL                 = 0xc0206928
+	SIOCATMARK                  = 0x40047307
+	SIOCAUTOADDR                = 0xc0206926
+	SIOCAUTONETMASK             = 0x80206927
+	SIOCDELMULTI                = 0x80206932
+	SIOCDIFADDR                 = 0x80206919
+	SIOCDIFPHYADDR              = 0x80206941
+	SIOCDLIFADDR                = 0x8118691f
+	SIOCGETSGCNT                = 0xc014721c
+	SIOCGETVIFCNT               = 0xc014721b
+	SIOCGETVLAN                 = 0xc020697f
+	SIOCGHIWAT                  = 0x40047301
+	SIOCGIFADDR                 = 0xc0206921
+	SIOCGIFALTMTU               = 0xc0206948
+	SIOCGIFASYNCMAP             = 0xc020697c
+	SIOCGIFBOND                 = 0xc0206947
+	SIOCGIFBRDADDR              = 0xc0206923
+	SIOCGIFCONF                 = 0xc0086924
+	SIOCGIFDEVMTU               = 0xc0206944
+	SIOCGIFDSTADDR              = 0xc0206922
+	SIOCGIFFLAGS                = 0xc0206911
+	SIOCGIFGENERIC              = 0xc020693a
+	SIOCGIFKPI                  = 0xc0206987
+	SIOCGIFMAC                  = 0xc0206982
+	SIOCGIFMEDIA                = 0xc0286938
+	SIOCGIFMETRIC               = 0xc0206917
+	SIOCGIFMTU                  = 0xc0206933
+	SIOCGIFNETMASK              = 0xc0206925
+	SIOCGIFPDSTADDR             = 0xc0206940
+	SIOCGIFPHYS                 = 0xc0206935
+	SIOCGIFPSRCADDR             = 0xc020693f
+	SIOCGIFSTATUS               = 0xc331693d
+	SIOCGIFVLAN                 = 0xc020697f
+	SIOCGLIFADDR                = 0xc118691e
+	SIOCGLIFPHYADDR             = 0xc1186943
+	SIOCGLOWAT                  = 0x40047303
+	SIOCGPGRP                   = 0x40047309
+	SIOCIFCREATE                = 0xc0206978
+	SIOCIFDESTROY               = 0x80206979
+	SIOCRSLVMULTI               = 0xc008693b
+	SIOCSETVLAN                 = 0x8020697e
+	SIOCSHIWAT                  = 0x80047300
+	SIOCSIFADDR                 = 0x8020690c
+	SIOCSIFALTMTU               = 0x80206945
+	SIOCSIFASYNCMAP             = 0x8020697d
+	SIOCSIFBOND                 = 0x80206946
+	SIOCSIFBRDADDR              = 0x80206913
+	SIOCSIFDSTADDR              = 0x8020690e
+	SIOCSIFFLAGS                = 0x80206910
+	SIOCSIFGENERIC              = 0x80206939
+	SIOCSIFKPI                  = 0x80206986
+	SIOCSIFLLADDR               = 0x8020693c
+	SIOCSIFMAC                  = 0x80206983
+	SIOCSIFMEDIA                = 0xc0206937
+	SIOCSIFMETRIC               = 0x80206918
+	SIOCSIFMTU                  = 0x80206934
+	SIOCSIFNETMASK              = 0x80206916
+	SIOCSIFPHYADDR              = 0x8040693e
+	SIOCSIFPHYS                 = 0x80206936
+	SIOCSIFVLAN                 = 0x8020697e
+	SIOCSLIFPHYADDR             = 0x81186942
+	SIOCSLOWAT                  = 0x80047302
+	SIOCSPGRP                   = 0x80047308
+	SOCK_DGRAM                  = 0x2
+	SOCK_MAXADDRLEN             = 0xff
+	SOCK_RAW                    = 0x3
+	SOCK_RDM                    = 0x4
+	SOCK_SEQPACKET              = 0x5
+	SOCK_STREAM                 = 0x1
+	SOL_SOCKET                  = 0xffff
+	SOMAXCONN                   = 0x80
+	SO_ACCEPTCONN               = 0x2
+	SO_BROADCAST                = 0x20
+	SO_DEBUG                    = 0x1
+	SO_DONTROUTE                = 0x10
+	SO_DONTTRUNC                = 0x2000
+	SO_ERROR                    = 0x1007
+	SO_KEEPALIVE                = 0x8
+	SO_LABEL                    = 0x1010
+	SO_LINGER                   = 0x80
+	SO_LINGER_SEC               = 0x1080
+	SO_NKE                      = 0x1021
+	SO_NOADDRERR                = 0x1023
+	SO_NOSIGPIPE                = 0x1022
+	SO_NOTIFYCONFLICT           = 0x1026
+	SO_NREAD                    = 0x1020
+	SO_NWRITE                   = 0x1024
+	SO_OOBINLINE                = 0x100
+	SO_PEERLABEL                = 0x1011
+	SO_RCVBUF                   = 0x1002
+	SO_RCVLOWAT                 = 0x1004
+	SO_RCVTIMEO                 = 0x1006
+	SO_RESTRICTIONS             = 0x1081
+	SO_RESTRICT_DENYIN          = 0x1
+	SO_RESTRICT_DENYOUT         = 0x2
+	SO_RESTRICT_DENYSET         = 0x80000000
+	SO_REUSEADDR                = 0x4
+	SO_REUSEPORT                = 0x200
+	SO_REUSESHAREUID            = 0x1025
+	SO_SNDBUF                   = 0x1001
+	SO_SNDLOWAT                 = 0x1003
+	SO_SNDTIMEO                 = 0x1005
+	SO_TIMESTAMP                = 0x400
+	SO_TYPE                     = 0x1008
+	SO_UPCALLCLOSEWAIT          = 0x1027
+	SO_USELOOPBACK              = 0x40
+	SO_WANTMORE                 = 0x4000
+	SO_WANTOOBFLAG              = 0x8000
+	S_IEXEC                     = 0x40
+	S_IFBLK                     = 0x6000
+	S_IFCHR                     = 0x2000
+	S_IFDIR                     = 0x4000
+	S_IFIFO                     = 0x1000
+	S_IFLNK                     = 0xa000
+	S_IFMT                      = 0xf000
+	S_IFREG                     = 0x8000
+	S_IFSOCK                    = 0xc000
+	S_IFWHT                     = 0xe000
+	S_IFXATTR                   = 0x10000
+	S_IREAD                     = 0x100
+	S_IRGRP                     = 0x20
+	S_IROTH                     = 0x4
+	S_IRUSR                     = 0x100
+	S_IRWXG                     = 0x38
+	S_IRWXO                     = 0x7
+	S_IRWXU                     = 0x1c0
+	S_ISGID                     = 0x400
+	S_ISTXT                     = 0x200
+	S_ISUID                     = 0x800
+	S_ISVTX                     = 0x200
+	S_IWGRP                     = 0x10
+	S_IWOTH                     = 0x2
+	S_IWRITE                    = 0x80
+	S_IWUSR                     = 0x80
+	S_IXGRP                     = 0x8
+	S_IXOTH                     = 0x1
+	S_IXUSR                     = 0x40
+	TCIFLUSH                    = 0x1
+	TCIOFLUSH                   = 0x3
+	TCOFLUSH                    = 0x2
+	TCP_KEEPALIVE               = 0x10
+	TCP_MAXBURST                = 0x4
+	TCP_MAXHLEN                 = 0x3c
+	TCP_MAXOLEN                 = 0x28
+	TCP_MAXSEG                  = 0x2
+	TCP_MAXWIN                  = 0xffff
+	TCP_MAX_SACK                = 0x3
+	TCP_MAX_WINSHIFT            = 0xe
+	TCP_MINMSS                  = 0xd8
+	TCP_MINMSSOVERLOAD          = 0x3e8
+	TCP_MSS                     = 0x200
+	TCP_NODELAY                 = 0x1
+	TCP_NOOPT                   = 0x8
+	TCP_NOPUSH                  = 0x4
+	TCSAFLUSH                   = 0x2
+	TIOCCBRK                    = 0x2000747a
+	TIOCCDTR                    = 0x20007478
+	TIOCCONS                    = 0x80047462
+	TIOCDCDTIMESTAMP            = 0x40087458
+	TIOCDRAIN                   = 0x2000745e
+	TIOCDSIMICROCODE            = 0x20007455
+	TIOCEXCL                    = 0x2000740d
+	TIOCEXT                     = 0x80047460
+	TIOCFLUSH                   = 0x80047410
+	TIOCGDRAINWAIT              = 0x40047456
+	TIOCGETA                    = 0x402c7413
+	TIOCGETD                    = 0x4004741a
+	TIOCGPGRP                   = 0x40047477
+	TIOCGWINSZ                  = 0x40087468
+	TIOCIXOFF                   = 0x20007480
+	TIOCIXON                    = 0x20007481
+	TIOCMBIC                    = 0x8004746b
+	TIOCMBIS                    = 0x8004746c
+	TIOCMGDTRWAIT               = 0x4004745a
+	TIOCMGET                    = 0x4004746a
+	TIOCMODG                    = 0x40047403
+	TIOCMODS                    = 0x80047404
+	TIOCMSDTRWAIT               = 0x8004745b
+	TIOCMSET                    = 0x8004746d
+	TIOCM_CAR                   = 0x40
+	TIOCM_CD                    = 0x40
+	TIOCM_CTS                   = 0x20
+	TIOCM_DSR                   = 0x100
+	TIOCM_DTR                   = 0x2
+	TIOCM_LE                    = 0x1
+	TIOCM_RI                    = 0x80
+	TIOCM_RNG                   = 0x80
+	TIOCM_RTS                   = 0x4
+	TIOCM_SR                    = 0x10
+	TIOCM_ST                    = 0x8
+	TIOCNOTTY                   = 0x20007471
+	TIOCNXCL                    = 0x2000740e
+	TIOCOUTQ                    = 0x40047473
+	TIOCPKT                     = 0x80047470
+	TIOCPKT_DATA                = 0x0
+	TIOCPKT_DOSTOP              = 0x20
+	TIOCPKT_FLUSHREAD           = 0x1
+	TIOCPKT_FLUSHWRITE          = 0x2
+	TIOCPKT_IOCTL               = 0x40
+	TIOCPKT_NOSTOP              = 0x10
+	TIOCPKT_START               = 0x8
+	TIOCPKT_STOP                = 0x4
+	TIOCPTYGNAME                = 0x40807453
+	TIOCPTYGRANT                = 0x20007454
+	TIOCPTYUNLK                 = 0x20007452
+	TIOCREMOTE                  = 0x80047469
+	TIOCSBRK                    = 0x2000747b
+	TIOCSCONS                   = 0x20007463
+	TIOCSCTTY                   = 0x20007461
+	TIOCSDRAINWAIT              = 0x80047457
+	TIOCSDTR                    = 0x20007479
+	TIOCSETA                    = 0x802c7414
+	TIOCSETAF                   = 0x802c7416
+	TIOCSETAW                   = 0x802c7415
+	TIOCSETD                    = 0x8004741b
+	TIOCSIG                     = 0x2000745f
+	TIOCSPGRP                   = 0x80047476
+	TIOCSTART                   = 0x2000746e
+	TIOCSTAT                    = 0x20007465
+	TIOCSTI                     = 0x80017472
+	TIOCSTOP                    = 0x2000746f
+	TIOCSWINSZ                  = 0x80087467
+	TIOCTIMESTAMP               = 0x40087459
+	TIOCUCNTL                   = 0x80047466
+	TOSTOP                      = 0x400000
+	VDISCARD                    = 0xf
+	VDSUSP                      = 0xb
+	VEOF                        = 0x0
+	VEOL                        = 0x1
+	VEOL2                       = 0x2
+	VERASE                      = 0x3
+	VINTR                       = 0x8
+	VKILL                       = 0x5
+	VLNEXT                      = 0xe
+	VMIN                        = 0x10
+	VQUIT                       = 0x9
+	VREPRINT                    = 0x6
+	VSTART                      = 0xc
+	VSTATUS                     = 0x12
+	VSTOP                       = 0xd
+	VSUSP                       = 0xa
+	VT0                         = 0x0
+	VT1                         = 0x10000
+	VTDLY                       = 0x10000
+	VTIME                       = 0x11
+	VWERASE                     = 0x4
+	WCONTINUED                  = 0x10
+	WCOREFLAG                   = 0x80
+	WEXITED                     = 0x4
+	WNOHANG                     = 0x1
+	WNOWAIT                     = 0x20
+	WORDSIZE                    = 0x20
+	WSTOPPED                    = 0x8
+	WUNTRACED                   = 0x2
+)
+
+// Errors
+const (
+	E2BIG           = Errno(0x7)
+	EACCES          = Errno(0xd)
+	EADDRINUSE      = Errno(0x30)
+	EADDRNOTAVAIL   = Errno(0x31)
+	EAFNOSUPPORT    = Errno(0x2f)
+	EAGAIN          = Errno(0x23)
+	EALREADY        = Errno(0x25)
+	EAUTH           = Errno(0x50)
+	EBADARCH        = Errno(0x56)
+	EBADEXEC        = Errno(0x55)
+	EBADF           = Errno(0x9)
+	EBADMACHO       = Errno(0x58)
+	EBADMSG         = Errno(0x5e)
+	EBADRPC         = Errno(0x48)
+	EBUSY           = Errno(0x10)
+	ECANCELED       = Errno(0x59)
+	ECHILD          = Errno(0xa)
+	ECONNABORTED    = Errno(0x35)
+	ECONNREFUSED    = Errno(0x3d)
+	ECONNRESET      = Errno(0x36)
+	EDEADLK         = Errno(0xb)
+	EDESTADDRREQ    = Errno(0x27)
+	EDEVERR         = Errno(0x53)
+	EDOM            = Errno(0x21)
+	EDQUOT          = Errno(0x45)
+	EEXIST          = Errno(0x11)
+	EFAULT          = Errno(0xe)
+	EFBIG           = Errno(0x1b)
+	EFTYPE          = Errno(0x4f)
+	EHOSTDOWN       = Errno(0x40)
+	EHOSTUNREACH    = Errno(0x41)
+	EIDRM           = Errno(0x5a)
+	EILSEQ          = Errno(0x5c)
+	EINPROGRESS     = Errno(0x24)
+	EINTR           = Errno(0x4)
+	EINVAL          = Errno(0x16)
+	EIO             = Errno(0x5)
+	EISCONN         = Errno(0x38)
+	EISDIR          = Errno(0x15)
+	ELAST           = Errno(0x67)
+	ELOOP           = Errno(0x3e)
+	EMFILE          = Errno(0x18)
+	EMLINK          = Errno(0x1f)
+	EMSGSIZE        = Errno(0x28)
+	EMULTIHOP       = Errno(0x5f)
+	ENAMETOOLONG    = Errno(0x3f)
+	ENEEDAUTH       = Errno(0x51)
+	ENETDOWN        = Errno(0x32)
+	ENETRESET       = Errno(0x34)
+	ENETUNREACH     = Errno(0x33)
+	ENFILE          = Errno(0x17)
+	ENOATTR         = Errno(0x5d)
+	ENOBUFS         = Errno(0x37)
+	ENODATA         = Errno(0x60)
+	ENODEV          = Errno(0x13)
+	ENOENT          = Errno(0x2)
+	ENOEXEC         = Errno(0x8)
+	ENOLCK          = Errno(0x4d)
+	ENOLINK         = Errno(0x61)
+	ENOMEM          = Errno(0xc)
+	ENOMSG          = Errno(0x5b)
+	ENOPOLICY       = Errno(0x67)
+	ENOPROTOOPT     = Errno(0x2a)
+	ENOSPC          = Errno(0x1c)
+	ENOSR           = Errno(0x62)
+	ENOSTR          = Errno(0x63)
+	ENOSYS          = Errno(0x4e)
+	ENOTBLK         = Errno(0xf)
+	ENOTCONN        = Errno(0x39)
+	ENOTDIR         = Errno(0x14)
+	ENOTEMPTY       = Errno(0x42)
+	ENOTSOCK        = Errno(0x26)
+	ENOTSUP         = Errno(0x2d)
+	ENOTTY          = Errno(0x19)
+	ENXIO           = Errno(0x6)
+	EOPNOTSUPP      = Errno(0x66)
+	EOVERFLOW       = Errno(0x54)
+	EPERM           = Errno(0x1)
+	EPFNOSUPPORT    = Errno(0x2e)
+	EPIPE           = Errno(0x20)
+	EPROCLIM        = Errno(0x43)
+	EPROCUNAVAIL    = Errno(0x4c)
+	EPROGMISMATCH   = Errno(0x4b)
+	EPROGUNAVAIL    = Errno(0x4a)
+	EPROTO          = Errno(0x64)
+	EPROTONOSUPPORT = Errno(0x2b)
+	EPROTOTYPE      = Errno(0x29)
+	EPWROFF         = Errno(0x52)
+	ERANGE          = Errno(0x22)
+	EREMOTE         = Errno(0x47)
+	EROFS           = Errno(0x1e)
+	ERPCMISMATCH    = Errno(0x49)
+	ESHLIBVERS      = Errno(0x57)
+	ESHUTDOWN       = Errno(0x3a)
+	ESOCKTNOSUPPORT = Errno(0x2c)
+	ESPIPE          = Errno(0x1d)
+	ESRCH           = Errno(0x3)
+	ESTALE          = Errno(0x46)
+	ETIME           = Errno(0x65)
+	ETIMEDOUT       = Errno(0x3c)
+	ETOOMANYREFS    = Errno(0x3b)
+	ETXTBSY         = Errno(0x1a)
+	EUSERS          = Errno(0x44)
+	EWOULDBLOCK     = Errno(0x23)
+	EXDEV           = Errno(0x12)
+)
+
+// Signals
+const (
+	SIGABRT   = Signal(0x6)
+	SIGALRM   = Signal(0xe)
+	SIGBUS    = Signal(0xa)
+	SIGCHLD   = Signal(0x14)
+	SIGCONT   = Signal(0x13)
+	SIGEMT    = Signal(0x7)
+	SIGFPE    = Signal(0x8)
+	SIGHUP    = Signal(0x1)
+	SIGILL    = Signal(0x4)
+	SIGINFO   = Signal(0x1d)
+	SIGINT    = Signal(0x2)
+	SIGIO     = Signal(0x17)
+	SIGIOT    = Signal(0x6)
+	SIGKILL   = Signal(0x9)
+	SIGPIPE   = Signal(0xd)
+	SIGPROF   = Signal(0x1b)
+	SIGQUIT   = Signal(0x3)
+	SIGSEGV   = Signal(0xb)
+	SIGSTOP   = Signal(0x11)
+	SIGSYS    = Signal(0xc)
+	SIGTERM   = Signal(0xf)
+	SIGTRAP   = Signal(0x5)
+	SIGTSTP   = Signal(0x12)
+	SIGTTIN   = Signal(0x15)
+	SIGTTOU   = Signal(0x16)
+	SIGURG    = Signal(0x10)
+	SIGUSR1   = Signal(0x1e)
+	SIGUSR2   = Signal(0x1f)
+	SIGVTALRM = Signal(0x1a)
+	SIGWINCH  = Signal(0x1c)
+	SIGXCPU   = Signal(0x18)
+	SIGXFSZ   = Signal(0x19)
+)
+
+// Error table
+var errors = [...]string{
+	1:   "operation not permitted",
+	2:   "no such file or directory",
+	3:   "no such process",
+	4:   "interrupted system call",
+	5:   "input/output error",
+	6:   "device not configured",
+	7:   "argument list too long",
+	8:   "exec format error",
+	9:   "bad file descriptor",
+	10:  "no child processes",
+	11:  "resource deadlock avoided",
+	12:  "cannot allocate memory",
+	13:  "permission denied",
+	14:  "bad address",
+	15:  "block device required",
+	16:  "resource busy",
+	17:  "file exists",
+	18:  "cross-device link",
+	19:  "operation not supported by device",
+	20:  "not a directory",
+	21:  "is a directory",
+	22:  "invalid argument",
+	23:  "too many open files in system",
+	24:  "too many open files",
+	25:  "inappropriate ioctl for device",
+	26:  "text file busy",
+	27:  "file too large",
+	28:  "no space left on device",
+	29:  "illegal seek",
+	30:  "read-only file system",
+	31:  "too many links",
+	32:  "broken pipe",
+	33:  "numerical argument out of domain",
+	34:  "result too large",
+	35:  "resource temporarily unavailable",
+	36:  "operation now in progress",
+	37:  "operation already in progress",
+	38:  "socket operation on non-socket",
+	39:  "destination address required",
+	40:  "message too long",
+	41:  "protocol wrong type for socket",
+	42:  "protocol not available",
+	43:  "protocol not supported",
+	44:  "socket type not supported",
+	45:  "operation not supported",
+	46:  "protocol family not supported",
+	47:  "address family not supported by protocol family",
+	48:  "address already in use",
+	49:  "can't assign requested address",
+	50:  "network is down",
+	51:  "network is unreachable",
+	52:  "network dropped connection on reset",
+	53:  "software caused connection abort",
+	54:  "connection reset by peer",
+	55:  "no buffer space available",
+	56:  "socket is already connected",
+	57:  "socket is not connected",
+	58:  "can't send after socket shutdown",
+	59:  "too many references: can't splice",
+	60:  "operation timed out",
+	61:  "connection refused",
+	62:  "too many levels of symbolic links",
+	63:  "file name too long",
+	64:  "host is down",
+	65:  "no route to host",
+	66:  "directory not empty",
+	67:  "too many processes",
+	68:  "too many users",
+	69:  "disc quota exceeded",
+	70:  "stale NFS file handle",
+	71:  "too many levels of remote in path",
+	72:  "RPC struct is bad",
+	73:  "RPC version wrong",
+	74:  "RPC prog. not avail",
+	75:  "program version wrong",
+	76:  "bad procedure for program",
+	77:  "no locks available",
+	78:  "function not implemented",
+	79:  "inappropriate file type or format",
+	80:  "authentication error",
+	81:  "need authenticator",
+	82:  "device power is off",
+	83:  "device error",
+	84:  "value too large to be stored in data type",
+	85:  "bad executable (or shared library)",
+	86:  "bad CPU type in executable",
+	87:  "shared library version mismatch",
+	88:  "malformed Mach-o file",
+	89:  "operation canceled",
+	90:  "identifier removed",
+	91:  "no message of desired type",
+	92:  "illegal byte sequence",
+	93:  "attribute not found",
+	94:  "bad message",
+	95:  "EMULTIHOP (Reserved)",
+	96:  "no message available on STREAM",
+	97:  "ENOLINK (Reserved)",
+	98:  "no STREAM resources",
+	99:  "not a STREAM",
+	100: "protocol error",
+	101: "STREAM ioctl timeout",
+	102: "operation not supported on socket",
+	103: "policy not found",
+}
+
+// Signal table
+var signals = [...]string{
+	1:  "hangup",
+	2:  "interrupt",
+	3:  "quit",
+	4:  "illegal instruction",
+	5:  "trace/BPT trap",
+	6:  "abort trap",
+	7:  "EMT trap",
+	8:  "floating point exception",
+	9:  "killed",
+	10: "bus error",
+	11: "segmentation fault",
+	12: "bad system call",
+	13: "broken pipe",
+	14: "alarm clock",
+	15: "terminated",
+	16: "urgent I/O condition",
+	17: "suspended (signal)",
+	18: "suspended",
+	19: "continued",
+	20: "child exited",
+	21: "stopped (tty input)",
+	22: "stopped (tty output)",
+	23: "I/O possible",
+	24: "cputime limit exceeded",
+	25: "filesize limit exceeded",
+	26: "virtual timer expired",
+	27: "profiling timer expired",
+	28: "window size changes",
+	29: "information request",
+	30: "user defined signal 1",
+	31: "user defined signal 2",
+}
diff --git a/src/syscall/zsyscall_darwin_arm.go b/src/syscall/zsyscall_darwin_arm.go
new file mode 100644
index 0000000000..d851d69b43
--- /dev/null
+++ b/src/syscall/zsyscall_darwin_arm.go
@@ -0,0 +1,1324 @@
+// mksyscall.pl -l32 syscall_bsd.go syscall_darwin.go syscall_darwin_arm.go
+// MACHINE GENERATED BY THE COMMAND ABOVE; DO NOT EDIT
+
+package syscall
+
+import "unsafe"
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func getgroups(ngid int, gid *_Gid_t) (n int, err error) {
+	r0, _, e1 := RawSyscall(SYS_GETGROUPS, uintptr(ngid), uintptr(unsafe.Pointer(gid)), 0)
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func setgroups(ngid int, gid *_Gid_t) (err error) {
+	_, _, e1 := RawSyscall(SYS_SETGROUPS, uintptr(ngid), uintptr(unsafe.Pointer(gid)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func wait4(pid int, wstatus *_C_int, options int, rusage *Rusage) (wpid int, err error) {
+	r0, _, e1 := Syscall6(SYS_WAIT4, uintptr(pid), uintptr(unsafe.Pointer(wstatus)), uintptr(options), uintptr(unsafe.Pointer(rusage)), 0, 0)
+	wpid = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func accept(s int, rsa *RawSockaddrAny, addrlen *_Socklen) (fd int, err error) {
+	r0, _, e1 := Syscall(SYS_ACCEPT, uintptr(s), uintptr(unsafe.Pointer(rsa)), uintptr(unsafe.Pointer(addrlen)))
+	fd = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func bind(s int, addr unsafe.Pointer, addrlen _Socklen) (err error) {
+	_, _, e1 := Syscall(SYS_BIND, uintptr(s), uintptr(addr), uintptr(addrlen))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func connect(s int, addr unsafe.Pointer, addrlen _Socklen) (err error) {
+	_, _, e1 := Syscall(SYS_CONNECT, uintptr(s), uintptr(addr), uintptr(addrlen))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func socket(domain int, typ int, proto int) (fd int, err error) {
+	r0, _, e1 := RawSyscall(SYS_SOCKET, uintptr(domain), uintptr(typ), uintptr(proto))
+	fd = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func getsockopt(s int, level int, name int, val unsafe.Pointer, vallen *_Socklen) (err error) {
+	_, _, e1 := Syscall6(SYS_GETSOCKOPT, uintptr(s), uintptr(level), uintptr(name), uintptr(val), uintptr(unsafe.Pointer(vallen)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func setsockopt(s int, level int, name int, val unsafe.Pointer, vallen uintptr) (err error) {
+	_, _, e1 := Syscall6(SYS_SETSOCKOPT, uintptr(s), uintptr(level), uintptr(name), uintptr(val), uintptr(vallen), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func getpeername(fd int, rsa *RawSockaddrAny, addrlen *_Socklen) (err error) {
+	_, _, e1 := RawSyscall(SYS_GETPEERNAME, uintptr(fd), uintptr(unsafe.Pointer(rsa)), uintptr(unsafe.Pointer(addrlen)))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func getsockname(fd int, rsa *RawSockaddrAny, addrlen *_Socklen) (err error) {
+	_, _, e1 := RawSyscall(SYS_GETSOCKNAME, uintptr(fd), uintptr(unsafe.Pointer(rsa)), uintptr(unsafe.Pointer(addrlen)))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Shutdown(s int, how int) (err error) {
+	_, _, e1 := Syscall(SYS_SHUTDOWN, uintptr(s), uintptr(how), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func socketpair(domain int, typ int, proto int, fd *[2]int32) (err error) {
+	_, _, e1 := RawSyscall6(SYS_SOCKETPAIR, uintptr(domain), uintptr(typ), uintptr(proto), uintptr(unsafe.Pointer(fd)), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func recvfrom(fd int, p []byte, flags int, from *RawSockaddrAny, fromlen *_Socklen) (n int, err error) {
+	var _p0 unsafe.Pointer
+	if len(p) > 0 {
+		_p0 = unsafe.Pointer(&p[0])
+	} else {
+		_p0 = unsafe.Pointer(&_zero)
+	}
+	r0, _, e1 := Syscall6(SYS_RECVFROM, uintptr(fd), uintptr(_p0), uintptr(len(p)), uintptr(flags), uintptr(unsafe.Pointer(from)), uintptr(unsafe.Pointer(fromlen)))
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func sendto(s int, buf []byte, flags int, to unsafe.Pointer, addrlen _Socklen) (err error) {
+	var _p0 unsafe.Pointer
+	if len(buf) > 0 {
+		_p0 = unsafe.Pointer(&buf[0])
+	} else {
+		_p0 = unsafe.Pointer(&_zero)
+	}
+	_, _, e1 := Syscall6(SYS_SENDTO, uintptr(s), uintptr(_p0), uintptr(len(buf)), uintptr(flags), uintptr(to), uintptr(addrlen))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func recvmsg(s int, msg *Msghdr, flags int) (n int, err error) {
+	r0, _, e1 := Syscall(SYS_RECVMSG, uintptr(s), uintptr(unsafe.Pointer(msg)), uintptr(flags))
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func sendmsg(s int, msg *Msghdr, flags int) (n int, err error) {
+	r0, _, e1 := Syscall(SYS_SENDMSG, uintptr(s), uintptr(unsafe.Pointer(msg)), uintptr(flags))
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func kevent(kq int, change unsafe.Pointer, nchange int, event unsafe.Pointer, nevent int, timeout *Timespec) (n int, err error) {
+	r0, _, e1 := Syscall6(SYS_KEVENT, uintptr(kq), uintptr(change), uintptr(nchange), uintptr(event), uintptr(nevent), uintptr(unsafe.Pointer(timeout)))
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func sysctl(mib []_C_int, old *byte, oldlen *uintptr, new *byte, newlen uintptr) (err error) {
+	var _p0 unsafe.Pointer
+	if len(mib) > 0 {
+		_p0 = unsafe.Pointer(&mib[0])
+	} else {
+		_p0 = unsafe.Pointer(&_zero)
+	}
+	_, _, e1 := Syscall6(SYS___SYSCTL, uintptr(_p0), uintptr(len(mib)), uintptr(unsafe.Pointer(old)), uintptr(unsafe.Pointer(oldlen)), uintptr(unsafe.Pointer(new)), uintptr(newlen))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func utimes(path string, timeval *[2]Timeval) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_UTIMES, uintptr(unsafe.Pointer(_p0)), uintptr(unsafe.Pointer(timeval)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func futimes(fd int, timeval *[2]Timeval) (err error) {
+	_, _, e1 := Syscall(SYS_FUTIMES, uintptr(fd), uintptr(unsafe.Pointer(timeval)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func fcntl(fd int, cmd int, arg int) (val int, err error) {
+	r0, _, e1 := Syscall(SYS_FCNTL, uintptr(fd), uintptr(cmd), uintptr(arg))
+	val = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func ptrace(request int, pid int, addr uintptr, data uintptr) (err error) {
+	_, _, e1 := Syscall6(SYS_PTRACE, uintptr(request), uintptr(pid), uintptr(addr), uintptr(data), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func pipe() (r int, w int, err error) {
+	r0, r1, e1 := RawSyscall(SYS_PIPE, 0, 0, 0)
+	r = int(r0)
+	w = int(r1)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func kill(pid int, signum int, posix int) (err error) {
+	_, _, e1 := Syscall(SYS_KILL, uintptr(pid), uintptr(signum), uintptr(posix))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Access(path string, mode uint32) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_ACCESS, uintptr(unsafe.Pointer(_p0)), uintptr(mode), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Adjtime(delta *Timeval, olddelta *Timeval) (err error) {
+	_, _, e1 := Syscall(SYS_ADJTIME, uintptr(unsafe.Pointer(delta)), uintptr(unsafe.Pointer(olddelta)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Chdir(path string) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_CHDIR, uintptr(unsafe.Pointer(_p0)), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Chflags(path string, flags int) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_CHFLAGS, uintptr(unsafe.Pointer(_p0)), uintptr(flags), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Chmod(path string, mode uint32) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_CHMOD, uintptr(unsafe.Pointer(_p0)), uintptr(mode), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Chown(path string, uid int, gid int) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_CHOWN, uintptr(unsafe.Pointer(_p0)), uintptr(uid), uintptr(gid))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Chroot(path string) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_CHROOT, uintptr(unsafe.Pointer(_p0)), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Close(fd int) (err error) {
+	_, _, e1 := Syscall(SYS_CLOSE, uintptr(fd), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Dup(fd int) (nfd int, err error) {
+	r0, _, e1 := RawSyscall(SYS_DUP, uintptr(fd), 0, 0)
+	nfd = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Dup2(from int, to int) (err error) {
+	_, _, e1 := RawSyscall(SYS_DUP2, uintptr(from), uintptr(to), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Exchangedata(path1 string, path2 string, options int) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path1)
+	if err != nil {
+		return
+	}
+	var _p1 *byte
+	_p1, err = BytePtrFromString(path2)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_EXCHANGEDATA, uintptr(unsafe.Pointer(_p0)), uintptr(unsafe.Pointer(_p1)), uintptr(options))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Exit(code int) {
+	Syscall(SYS_EXIT, uintptr(code), 0, 0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Fchdir(fd int) (err error) {
+	_, _, e1 := Syscall(SYS_FCHDIR, uintptr(fd), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Fchflags(path string, flags int) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_FCHFLAGS, uintptr(unsafe.Pointer(_p0)), uintptr(flags), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Fchmod(fd int, mode uint32) (err error) {
+	_, _, e1 := Syscall(SYS_FCHMOD, uintptr(fd), uintptr(mode), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Fchown(fd int, uid int, gid int) (err error) {
+	_, _, e1 := Syscall(SYS_FCHOWN, uintptr(fd), uintptr(uid), uintptr(gid))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Flock(fd int, how int) (err error) {
+	_, _, e1 := Syscall(SYS_FLOCK, uintptr(fd), uintptr(how), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Fpathconf(fd int, name int) (val int, err error) {
+	r0, _, e1 := Syscall(SYS_FPATHCONF, uintptr(fd), uintptr(name), 0)
+	val = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Fstat(fd int, stat *Stat_t) (err error) {
+	_, _, e1 := Syscall(SYS_FSTAT64, uintptr(fd), uintptr(unsafe.Pointer(stat)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Fstatfs(fd int, stat *Statfs_t) (err error) {
+	_, _, e1 := Syscall(SYS_FSTATFS64, uintptr(fd), uintptr(unsafe.Pointer(stat)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Fsync(fd int) (err error) {
+	_, _, e1 := Syscall(SYS_FSYNC, uintptr(fd), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Ftruncate(fd int, length int64) (err error) {
+	_, _, e1 := Syscall(SYS_FTRUNCATE, uintptr(fd), uintptr(length), uintptr(length>>32))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getdirentries(fd int, buf []byte, basep *uintptr) (n int, err error) {
+	var _p0 unsafe.Pointer
+	if len(buf) > 0 {
+		_p0 = unsafe.Pointer(&buf[0])
+	} else {
+		_p0 = unsafe.Pointer(&_zero)
+	}
+	r0, _, e1 := Syscall6(SYS_GETDIRENTRIES64, uintptr(fd), uintptr(_p0), uintptr(len(buf)), uintptr(unsafe.Pointer(basep)), 0, 0)
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getdtablesize() (size int) {
+	r0, _, _ := Syscall(SYS_GETDTABLESIZE, 0, 0, 0)
+	size = int(r0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getegid() (egid int) {
+	r0, _, _ := RawSyscall(SYS_GETEGID, 0, 0, 0)
+	egid = int(r0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Geteuid() (uid int) {
+	r0, _, _ := RawSyscall(SYS_GETEUID, 0, 0, 0)
+	uid = int(r0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getgid() (gid int) {
+	r0, _, _ := RawSyscall(SYS_GETGID, 0, 0, 0)
+	gid = int(r0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getpgid(pid int) (pgid int, err error) {
+	r0, _, e1 := RawSyscall(SYS_GETPGID, uintptr(pid), 0, 0)
+	pgid = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getpgrp() (pgrp int) {
+	r0, _, _ := RawSyscall(SYS_GETPGRP, 0, 0, 0)
+	pgrp = int(r0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getpid() (pid int) {
+	r0, _, _ := RawSyscall(SYS_GETPID, 0, 0, 0)
+	pid = int(r0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getppid() (ppid int) {
+	r0, _, _ := RawSyscall(SYS_GETPPID, 0, 0, 0)
+	ppid = int(r0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getpriority(which int, who int) (prio int, err error) {
+	r0, _, e1 := Syscall(SYS_GETPRIORITY, uintptr(which), uintptr(who), 0)
+	prio = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getrlimit(which int, lim *Rlimit) (err error) {
+	_, _, e1 := RawSyscall(SYS_GETRLIMIT, uintptr(which), uintptr(unsafe.Pointer(lim)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getrusage(who int, rusage *Rusage) (err error) {
+	_, _, e1 := RawSyscall(SYS_GETRUSAGE, uintptr(who), uintptr(unsafe.Pointer(rusage)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getsid(pid int) (sid int, err error) {
+	r0, _, e1 := RawSyscall(SYS_GETSID, uintptr(pid), 0, 0)
+	sid = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Getuid() (uid int) {
+	r0, _, _ := RawSyscall(SYS_GETUID, 0, 0, 0)
+	uid = int(r0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Issetugid() (tainted bool) {
+	r0, _, _ := RawSyscall(SYS_ISSETUGID, 0, 0, 0)
+	tainted = bool(r0 != 0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Kqueue() (fd int, err error) {
+	r0, _, e1 := Syscall(SYS_KQUEUE, 0, 0, 0)
+	fd = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Lchown(path string, uid int, gid int) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_LCHOWN, uintptr(unsafe.Pointer(_p0)), uintptr(uid), uintptr(gid))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Link(path string, link string) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	var _p1 *byte
+	_p1, err = BytePtrFromString(link)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_LINK, uintptr(unsafe.Pointer(_p0)), uintptr(unsafe.Pointer(_p1)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Listen(s int, backlog int) (err error) {
+	_, _, e1 := Syscall(SYS_LISTEN, uintptr(s), uintptr(backlog), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Lstat(path string, stat *Stat_t) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_LSTAT64, uintptr(unsafe.Pointer(_p0)), uintptr(unsafe.Pointer(stat)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Mkdir(path string, mode uint32) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_MKDIR, uintptr(unsafe.Pointer(_p0)), uintptr(mode), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Mkfifo(path string, mode uint32) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_MKFIFO, uintptr(unsafe.Pointer(_p0)), uintptr(mode), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Mknod(path string, mode uint32, dev int) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_MKNOD, uintptr(unsafe.Pointer(_p0)), uintptr(mode), uintptr(dev))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Open(path string, mode int, perm uint32) (fd int, err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	r0, _, e1 := Syscall(SYS_OPEN, uintptr(unsafe.Pointer(_p0)), uintptr(mode), uintptr(perm))
+	fd = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Pathconf(path string, name int) (val int, err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	r0, _, e1 := Syscall(SYS_PATHCONF, uintptr(unsafe.Pointer(_p0)), uintptr(name), 0)
+	val = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Pread(fd int, p []byte, offset int64) (n int, err error) {
+	var _p0 unsafe.Pointer
+	if len(p) > 0 {
+		_p0 = unsafe.Pointer(&p[0])
+	} else {
+		_p0 = unsafe.Pointer(&_zero)
+	}
+	r0, _, e1 := Syscall6(SYS_PREAD, uintptr(fd), uintptr(_p0), uintptr(len(p)), uintptr(offset), uintptr(offset>>32), 0)
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Pwrite(fd int, p []byte, offset int64) (n int, err error) {
+	var _p0 unsafe.Pointer
+	if len(p) > 0 {
+		_p0 = unsafe.Pointer(&p[0])
+	} else {
+		_p0 = unsafe.Pointer(&_zero)
+	}
+	r0, _, e1 := Syscall6(SYS_PWRITE, uintptr(fd), uintptr(_p0), uintptr(len(p)), uintptr(offset), uintptr(offset>>32), 0)
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func read(fd int, p []byte) (n int, err error) {
+	var _p0 unsafe.Pointer
+	if len(p) > 0 {
+		_p0 = unsafe.Pointer(&p[0])
+	} else {
+		_p0 = unsafe.Pointer(&_zero)
+	}
+	r0, _, e1 := Syscall(SYS_READ, uintptr(fd), uintptr(_p0), uintptr(len(p)))
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Readlink(path string, buf []byte) (n int, err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	var _p1 unsafe.Pointer
+	if len(buf) > 0 {
+		_p1 = unsafe.Pointer(&buf[0])
+	} else {
+		_p1 = unsafe.Pointer(&_zero)
+	}
+	r0, _, e1 := Syscall(SYS_READLINK, uintptr(unsafe.Pointer(_p0)), uintptr(_p1), uintptr(len(buf)))
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Rename(from string, to string) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(from)
+	if err != nil {
+		return
+	}
+	var _p1 *byte
+	_p1, err = BytePtrFromString(to)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_RENAME, uintptr(unsafe.Pointer(_p0)), uintptr(unsafe.Pointer(_p1)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Revoke(path string) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_REVOKE, uintptr(unsafe.Pointer(_p0)), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Rmdir(path string) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_RMDIR, uintptr(unsafe.Pointer(_p0)), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Seek(fd int, offset int64, whence int) (newoffset int64, err error) {
+	r0, r1, e1 := Syscall6(SYS_LSEEK, uintptr(fd), uintptr(offset), uintptr(offset>>32), uintptr(whence), 0, 0)
+	newoffset = int64(int64(r1)<<32 | int64(r0))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Select(n int, r *FdSet, w *FdSet, e *FdSet, timeout *Timeval) (err error) {
+	_, _, e1 := Syscall6(SYS_SELECT, uintptr(n), uintptr(unsafe.Pointer(r)), uintptr(unsafe.Pointer(w)), uintptr(unsafe.Pointer(e)), uintptr(unsafe.Pointer(timeout)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setegid(egid int) (err error) {
+	_, _, e1 := Syscall(SYS_SETEGID, uintptr(egid), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Seteuid(euid int) (err error) {
+	_, _, e1 := RawSyscall(SYS_SETEUID, uintptr(euid), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setgid(gid int) (err error) {
+	_, _, e1 := RawSyscall(SYS_SETGID, uintptr(gid), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setlogin(name string) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(name)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_SETLOGIN, uintptr(unsafe.Pointer(_p0)), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setpgid(pid int, pgid int) (err error) {
+	_, _, e1 := RawSyscall(SYS_SETPGID, uintptr(pid), uintptr(pgid), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setpriority(which int, who int, prio int) (err error) {
+	_, _, e1 := Syscall(SYS_SETPRIORITY, uintptr(which), uintptr(who), uintptr(prio))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setprivexec(flag int) (err error) {
+	_, _, e1 := Syscall(SYS_SETPRIVEXEC, uintptr(flag), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setregid(rgid int, egid int) (err error) {
+	_, _, e1 := RawSyscall(SYS_SETREGID, uintptr(rgid), uintptr(egid), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setreuid(ruid int, euid int) (err error) {
+	_, _, e1 := RawSyscall(SYS_SETREUID, uintptr(ruid), uintptr(euid), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setrlimit(which int, lim *Rlimit) (err error) {
+	_, _, e1 := RawSyscall(SYS_SETRLIMIT, uintptr(which), uintptr(unsafe.Pointer(lim)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setsid() (pid int, err error) {
+	r0, _, e1 := RawSyscall(SYS_SETSID, 0, 0, 0)
+	pid = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Settimeofday(tp *Timeval) (err error) {
+	_, _, e1 := RawSyscall(SYS_SETTIMEOFDAY, uintptr(unsafe.Pointer(tp)), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Setuid(uid int) (err error) {
+	_, _, e1 := RawSyscall(SYS_SETUID, uintptr(uid), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Stat(path string, stat *Stat_t) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_STAT64, uintptr(unsafe.Pointer(_p0)), uintptr(unsafe.Pointer(stat)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Statfs(path string, stat *Statfs_t) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_STATFS64, uintptr(unsafe.Pointer(_p0)), uintptr(unsafe.Pointer(stat)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Symlink(path string, link string) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	var _p1 *byte
+	_p1, err = BytePtrFromString(link)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_SYMLINK, uintptr(unsafe.Pointer(_p0)), uintptr(unsafe.Pointer(_p1)), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Sync() (err error) {
+	_, _, e1 := Syscall(SYS_SYNC, 0, 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Truncate(path string, length int64) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_TRUNCATE, uintptr(unsafe.Pointer(_p0)), uintptr(length), uintptr(length>>32))
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Umask(newmask int) (oldmask int) {
+	r0, _, _ := Syscall(SYS_UMASK, uintptr(newmask), 0, 0)
+	oldmask = int(r0)
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Undelete(path string) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_UNDELETE, uintptr(unsafe.Pointer(_p0)), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Unlink(path string) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_UNLINK, uintptr(unsafe.Pointer(_p0)), 0, 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func Unmount(path string, flags int) (err error) {
+	var _p0 *byte
+	_p0, err = BytePtrFromString(path)
+	if err != nil {
+		return
+	}
+	_, _, e1 := Syscall(SYS_UNMOUNT, uintptr(unsafe.Pointer(_p0)), uintptr(flags), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func write(fd int, p []byte) (n int, err error) {
+	var _p0 unsafe.Pointer
+	if len(p) > 0 {
+		_p0 = unsafe.Pointer(&p[0])
+	} else {
+		_p0 = unsafe.Pointer(&_zero)
+	}
+	r0, _, e1 := Syscall(SYS_WRITE, uintptr(fd), uintptr(_p0), uintptr(len(p)))
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func mmap(addr uintptr, length uintptr, prot int, flag int, fd int, pos int64) (ret uintptr, err error) {
+	r0, _, e1 := Syscall9(SYS_MMAP, uintptr(addr), uintptr(length), uintptr(prot), uintptr(flag), uintptr(fd), uintptr(pos), uintptr(pos>>32), 0, 0)
+	ret = uintptr(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func munmap(addr uintptr, length uintptr) (err error) {
+	_, _, e1 := Syscall(SYS_MUNMAP, uintptr(addr), uintptr(length), 0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func readlen(fd int, buf *byte, nbuf int) (n int, err error) {
+	r0, _, e1 := Syscall(SYS_READ, uintptr(fd), uintptr(unsafe.Pointer(buf)), uintptr(nbuf))
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func writelen(fd int, buf *byte, nbuf int) (n int, err error) {
+	r0, _, e1 := Syscall(SYS_WRITE, uintptr(fd), uintptr(unsafe.Pointer(buf)), uintptr(nbuf))
+	n = int(r0)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
+
+// THIS FILE IS GENERATED BY THE COMMAND AT THE TOP; DO NOT EDIT
+
+func gettimeofday(tp *Timeval) (sec int32, usec int32, err error) {
+	r0, r1, e1 := RawSyscall(SYS_GETTIMEOFDAY, uintptr(unsafe.Pointer(tp)), 0, 0)
+	sec = int32(r0)
+	usec = int32(r1)
+	if e1 != 0 {
+		err = e1
+	}
+	return
+}
diff --git a/src/syscall/zsysnum_darwin_arm.go b/src/syscall/zsysnum_darwin_arm.go
new file mode 100644
index 0000000000..1a53f13eff
--- /dev/null
+++ b/src/syscall/zsysnum_darwin_arm.go
@@ -0,0 +1,345 @@
+// mksysnum_darwin.pl /usr/include/sys/syscall.h
+// MACHINE GENERATED BY THE ABOVE COMMAND; DO NOT EDIT
+
+package syscall
+
+const (
+	SYS_SYSCALL                   = 0
+	SYS_EXIT                      = 1
+	SYS_FORK                      = 2
+	SYS_READ                      = 3
+	SYS_WRITE                     = 4
+	SYS_OPEN                      = 5
+	SYS_CLOSE                     = 6
+	SYS_WAIT4                     = 7
+	SYS_LINK                      = 9
+	SYS_UNLINK                    = 10
+	SYS_CHDIR                     = 12
+	SYS_FCHDIR                    = 13
+	SYS_MKNOD                     = 14
+	SYS_CHMOD                     = 15
+	SYS_CHOWN                     = 16
+	SYS_OBREAK                    = 17
+	SYS_OGETFSSTAT                = 18
+	SYS_GETFSSTAT                 = 18
+	SYS_GETPID                    = 20
+	SYS_SETUID                    = 23
+	SYS_GETUID                    = 24
+	SYS_GETEUID                   = 25
+	SYS_PTRACE                    = 26
+	SYS_RECVMSG                   = 27
+	SYS_SENDMSG                   = 28
+	SYS_RECVFROM                  = 29
+	SYS_ACCEPT                    = 30
+	SYS_GETPEERNAME               = 31
+	SYS_GETSOCKNAME               = 32
+	SYS_ACCESS                    = 33
+	SYS_CHFLAGS                   = 34
+	SYS_FCHFLAGS                  = 35
+	SYS_SYNC                      = 36
+	SYS_KILL                      = 37
+	SYS_GETPPID                   = 39
+	SYS_DUP                       = 41
+	SYS_PIPE                      = 42
+	SYS_GETEGID                   = 43
+	SYS_PROFIL                    = 44
+	SYS_SIGACTION                 = 46
+	SYS_GETGID                    = 47
+	SYS_SIGPROCMASK               = 48
+	SYS_GETLOGIN                  = 49
+	SYS_SETLOGIN                  = 50
+	SYS_ACCT                      = 51
+	SYS_SIGPENDING                = 52
+	SYS_SIGALTSTACK               = 53
+	SYS_IOCTL                     = 54
+	SYS_REBOOT                    = 55
+	SYS_REVOKE                    = 56
+	SYS_SYMLINK                   = 57
+	SYS_READLINK                  = 58
+	SYS_EXECVE                    = 59
+	SYS_UMASK                     = 60
+	SYS_CHROOT                    = 61
+	SYS_MSYNC                     = 65
+	SYS_VFORK                     = 66
+	SYS_SBRK                      = 69
+	SYS_SSTK                      = 70
+	SYS_OVADVISE                  = 72
+	SYS_MUNMAP                    = 73
+	SYS_MPROTECT                  = 74
+	SYS_MADVISE                   = 75
+	SYS_MINCORE                   = 78
+	SYS_GETGROUPS                 = 79
+	SYS_SETGROUPS                 = 80
+	SYS_GETPGRP                   = 81
+	SYS_SETPGID                   = 82
+	SYS_SETITIMER                 = 83
+	SYS_SWAPON                    = 85
+	SYS_GETITIMER                 = 86
+	SYS_GETDTABLESIZE             = 89
+	SYS_DUP2                      = 90
+	SYS_FCNTL                     = 92
+	SYS_SELECT                    = 93
+	SYS_FSYNC                     = 95
+	SYS_SETPRIORITY               = 96
+	SYS_SOCKET                    = 97
+	SYS_CONNECT                   = 98
+	SYS_GETPRIORITY               = 100
+	SYS_BIND                      = 104
+	SYS_SETSOCKOPT                = 105
+	SYS_LISTEN                    = 106
+	SYS_SIGSUSPEND                = 111
+	SYS_GETTIMEOFDAY              = 116
+	SYS_GETRUSAGE                 = 117
+	SYS_GETSOCKOPT                = 118
+	SYS_READV                     = 120
+	SYS_WRITEV                    = 121
+	SYS_SETTIMEOFDAY              = 122
+	SYS_FCHOWN                    = 123
+	SYS_FCHMOD                    = 124
+	SYS_SETREUID                  = 126
+	SYS_SETREGID                  = 127
+	SYS_RENAME                    = 128
+	SYS_FLOCK                     = 131
+	SYS_MKFIFO                    = 132
+	SYS_SENDTO                    = 133
+	SYS_SHUTDOWN                  = 134
+	SYS_SOCKETPAIR                = 135
+	SYS_MKDIR                     = 136
+	SYS_RMDIR                     = 137
+	SYS_UTIMES                    = 138
+	SYS_FUTIMES                   = 139
+	SYS_ADJTIME                   = 140
+	SYS_GETHOSTUUID               = 142
+	SYS_SETSID                    = 147
+	SYS_GETPGID                   = 151
+	SYS_SETPRIVEXEC               = 152
+	SYS_PREAD                     = 153
+	SYS_PWRITE                    = 154
+	SYS_NFSSVC                    = 155
+	SYS_STATFS                    = 157
+	SYS_FSTATFS                   = 158
+	SYS_UNMOUNT                   = 159
+	SYS_GETFH                     = 161
+	SYS_QUOTACTL                  = 165
+	SYS_MOUNT                     = 167
+	SYS_CSOPS                     = 169
+	SYS_TABLE                     = 170
+	SYS_WAITID                    = 173
+	SYS_ADD_PROFIL                = 176
+	SYS_KDEBUG_TRACE              = 180
+	SYS_SETGID                    = 181
+	SYS_SETEGID                   = 182
+	SYS_SETEUID                   = 183
+	SYS_SIGRETURN                 = 184
+	SYS_CHUD                      = 185
+	SYS_STAT                      = 188
+	SYS_FSTAT                     = 189
+	SYS_LSTAT                     = 190
+	SYS_PATHCONF                  = 191
+	SYS_FPATHCONF                 = 192
+	SYS_GETRLIMIT                 = 194
+	SYS_SETRLIMIT                 = 195
+	SYS_GETDIRENTRIES             = 196
+	SYS_MMAP                      = 197
+	SYS_LSEEK                     = 199
+	SYS_TRUNCATE                  = 200
+	SYS_FTRUNCATE                 = 201
+	SYS___SYSCTL                  = 202
+	SYS_MLOCK                     = 203
+	SYS_MUNLOCK                   = 204
+	SYS_UNDELETE                  = 205
+	SYS_ATSOCKET                  = 206
+	SYS_ATGETMSG                  = 207
+	SYS_ATPUTMSG                  = 208
+	SYS_ATPSNDREQ                 = 209
+	SYS_ATPSNDRSP                 = 210
+	SYS_ATPGETREQ                 = 211
+	SYS_ATPGETRSP                 = 212
+	SYS_KQUEUE_FROM_PORTSET_NP    = 214
+	SYS_KQUEUE_PORTSET_NP         = 215
+	SYS_MKCOMPLEX                 = 216
+	SYS_STATV                     = 217
+	SYS_LSTATV                    = 218
+	SYS_FSTATV                    = 219
+	SYS_GETATTRLIST               = 220
+	SYS_SETATTRLIST               = 221
+	SYS_GETDIRENTRIESATTR         = 222
+	SYS_EXCHANGEDATA              = 223
+	SYS_SEARCHFS                  = 225
+	SYS_DELETE                    = 226
+	SYS_COPYFILE                  = 227
+	SYS_POLL                      = 230
+	SYS_WATCHEVENT                = 231
+	SYS_WAITEVENT                 = 232
+	SYS_MODWATCH                  = 233
+	SYS_GETXATTR                  = 234
+	SYS_FGETXATTR                 = 235
+	SYS_SETXATTR                  = 236
+	SYS_FSETXATTR                 = 237
+	SYS_REMOVEXATTR               = 238
+	SYS_FREMOVEXATTR              = 239
+	SYS_LISTXATTR                 = 240
+	SYS_FLISTXATTR                = 241
+	SYS_FSCTL                     = 242
+	SYS_INITGROUPS                = 243
+	SYS_POSIX_SPAWN               = 244
+	SYS_NFSCLNT                   = 247
+	SYS_FHOPEN                    = 248
+	SYS_MINHERIT                  = 250
+	SYS_SEMSYS                    = 251
+	SYS_MSGSYS                    = 252
+	SYS_SHMSYS                    = 253
+	SYS_SEMCTL                    = 254
+	SYS_SEMGET                    = 255
+	SYS_SEMOP                     = 256
+	SYS_MSGCTL                    = 258
+	SYS_MSGGET                    = 259
+	SYS_MSGSND                    = 260
+	SYS_MSGRCV                    = 261
+	SYS_SHMAT                     = 262
+	SYS_SHMCTL                    = 263
+	SYS_SHMDT                     = 264
+	SYS_SHMGET                    = 265
+	SYS_SHM_OPEN                  = 266
+	SYS_SHM_UNLINK                = 267
+	SYS_SEM_OPEN                  = 268
+	SYS_SEM_CLOSE                 = 269
+	SYS_SEM_UNLINK                = 270
+	SYS_SEM_WAIT                  = 271
+	SYS_SEM_TRYWAIT               = 272
+	SYS_SEM_POST                  = 273
+	SYS_SEM_GETVALUE              = 274
+	SYS_SEM_INIT                  = 275
+	SYS_SEM_DESTROY               = 276
+	SYS_OPEN_EXTENDED             = 277
+	SYS_UMASK_EXTENDED            = 278
+	SYS_STAT_EXTENDED             = 279
+	SYS_LSTAT_EXTENDED            = 280
+	SYS_FSTAT_EXTENDED            = 281
+	SYS_CHMOD_EXTENDED            = 282
+	SYS_FCHMOD_EXTENDED           = 283
+	SYS_ACCESS_EXTENDED           = 284
+	SYS_SETTID                    = 285
+	SYS_GETTID                    = 286
+	SYS_SETSGROUPS                = 287
+	SYS_GETSGROUPS                = 288
+	SYS_SETWGROUPS                = 289
+	SYS_GETWGROUPS                = 290
+	SYS_MKFIFO_EXTENDED           = 291
+	SYS_MKDIR_EXTENDED            = 292
+	SYS_IDENTITYSVC               = 293
+	SYS_SHARED_REGION_CHECK_NP    = 294
+	SYS_SHARED_REGION_MAP_NP      = 295
+	SYS___PTHREAD_MUTEX_DESTROY   = 301
+	SYS___PTHREAD_MUTEX_INIT      = 302
+	SYS___PTHREAD_MUTEX_LOCK      = 303
+	SYS___PTHREAD_MUTEX_TRYLOCK   = 304
+	SYS___PTHREAD_MUTEX_UNLOCK    = 305
+	SYS___PTHREAD_COND_INIT       = 306
+	SYS___PTHREAD_COND_DESTROY    = 307
+	SYS___PTHREAD_COND_BROADCAST  = 308
+	SYS___PTHREAD_COND_SIGNAL     = 309
+	SYS_GETSID                    = 310
+	SYS_SETTID_WITH_PID           = 311
+	SYS___PTHREAD_COND_TIMEDWAIT  = 312
+	SYS_AIO_FSYNC                 = 313
+	SYS_AIO_RETURN                = 314
+	SYS_AIO_SUSPEND               = 315
+	SYS_AIO_CANCEL                = 316
+	SYS_AIO_ERROR                 = 317
+	SYS_AIO_READ                  = 318
+	SYS_AIO_WRITE                 = 319
+	SYS_LIO_LISTIO                = 320
+	SYS___PTHREAD_COND_WAIT       = 321
+	SYS_IOPOLICYSYS               = 322
+	SYS_MLOCKALL                  = 324
+	SYS_MUNLOCKALL                = 325
+	SYS_ISSETUGID                 = 327
+	SYS___PTHREAD_KILL            = 328
+	SYS___PTHREAD_SIGMASK         = 329
+	SYS___SIGWAIT                 = 330
+	SYS___DISABLE_THREADSIGNAL    = 331
+	SYS___PTHREAD_MARKCANCEL      = 332
+	SYS___PTHREAD_CANCELED        = 333
+	SYS___SEMWAIT_SIGNAL          = 334
+	SYS_PROC_INFO                 = 336
+	SYS_SENDFILE                  = 337
+	SYS_STAT64                    = 338
+	SYS_FSTAT64                   = 339
+	SYS_LSTAT64                   = 340
+	SYS_STAT64_EXTENDED           = 341
+	SYS_LSTAT64_EXTENDED          = 342
+	SYS_FSTAT64_EXTENDED          = 343
+	SYS_GETDIRENTRIES64           = 344
+	SYS_STATFS64                  = 345
+	SYS_FSTATFS64                 = 346
+	SYS_GETFSSTAT64               = 347
+	SYS___PTHREAD_CHDIR           = 348
+	SYS___PTHREAD_FCHDIR          = 349
+	SYS_AUDIT                     = 350
+	SYS_AUDITON                   = 351
+	SYS_GETAUID                   = 353
+	SYS_SETAUID                   = 354
+	SYS_GETAUDIT                  = 355
+	SYS_SETAUDIT                  = 356
+	SYS_GETAUDIT_ADDR             = 357
+	SYS_SETAUDIT_ADDR             = 358
+	SYS_AUDITCTL                  = 359
+	SYS_BSDTHREAD_CREATE          = 360
+	SYS_BSDTHREAD_TERMINATE       = 361
+	SYS_KQUEUE                    = 362
+	SYS_KEVENT                    = 363
+	SYS_LCHOWN                    = 364
+	SYS_STACK_SNAPSHOT            = 365
+	SYS_BSDTHREAD_REGISTER        = 366
+	SYS_WORKQ_OPEN                = 367
+	SYS_WORKQ_OPS                 = 368
+	SYS___MAC_EXECVE              = 380
+	SYS___MAC_SYSCALL             = 381
+	SYS___MAC_GET_FILE            = 382
+	SYS___MAC_SET_FILE            = 383
+	SYS___MAC_GET_LINK            = 384
+	SYS___MAC_SET_LINK            = 385
+	SYS___MAC_GET_PROC            = 386
+	SYS___MAC_SET_PROC            = 387
+	SYS___MAC_GET_FD              = 388
+	SYS___MAC_SET_FD              = 389
+	SYS___MAC_GET_PID             = 390
+	SYS___MAC_GET_LCID            = 391
+	SYS___MAC_GET_LCTX            = 392
+	SYS___MAC_SET_LCTX            = 393
+	SYS_SETLCID                   = 394
+	SYS_GETLCID                   = 395
+	SYS_READ_NOCANCEL             = 396
+	SYS_WRITE_NOCANCEL            = 397
+	SYS_OPEN_NOCANCEL             = 398
+	SYS_CLOSE_NOCANCEL            = 399
+	SYS_WAIT4_NOCANCEL            = 400
+	SYS_RECVMSG_NOCANCEL          = 401
+	SYS_SENDMSG_NOCANCEL          = 402
+	SYS_RECVFROM_NOCANCEL         = 403
+	SYS_ACCEPT_NOCANCEL           = 404
+	SYS_MSYNC_NOCANCEL            = 405
+	SYS_FCNTL_NOCANCEL            = 406
+	SYS_SELECT_NOCANCEL           = 407
+	SYS_FSYNC_NOCANCEL            = 408
+	SYS_CONNECT_NOCANCEL          = 409
+	SYS_SIGSUSPEND_NOCANCEL       = 410
+	SYS_READV_NOCANCEL            = 411
+	SYS_WRITEV_NOCANCEL           = 412
+	SYS_SENDTO_NOCANCEL           = 413
+	SYS_PREAD_NOCANCEL            = 414
+	SYS_PWRITE_NOCANCEL           = 415
+	SYS_WAITID_NOCANCEL           = 416
+	SYS_POLL_NOCANCEL             = 417
+	SYS_MSGSND_NOCANCEL           = 418
+	SYS_MSGRCV_NOCANCEL           = 419
+	SYS_SEM_WAIT_NOCANCEL         = 420
+	SYS_AIO_SUSPEND_NOCANCEL      = 421
+	SYS___SIGWAIT_NOCANCEL        = 422
+	SYS___SEMWAIT_SIGNAL_NOCANCEL = 423
+	SYS___MAC_MOUNT               = 424
+	SYS___MAC_GET_MOUNT           = 425
+	SYS___MAC_GETFSSTAT           = 426
+	SYS_MAXSYSCALL                = 427
+)
diff --git a/src/syscall/ztypes_darwin_arm.go b/src/syscall/ztypes_darwin_arm.go
new file mode 100644
index 0000000000..ec87f54fb2
--- /dev/null
+++ b/src/syscall/ztypes_darwin_arm.go
@@ -0,0 +1,447 @@
+// NOTE: cgo can't generate struct Stat_t and struct Statfs_t yet
+// Created by cgo -godefs - DO NOT EDIT
+// cgo -godefs types_darwin.go
+
+package syscall
+
+const (
+	sizeofPtr      = 0x4
+	sizeofShort    = 0x2
+	sizeofInt      = 0x4
+	sizeofLong     = 0x4
+	sizeofLongLong = 0x8
+)
+
+type (
+	_C_short     int16
+	_C_int       int32
+	_C_long      int32
+	_C_long_long int64
+)
+
+type Timespec struct {
+	Sec  int32
+	Nsec int32
+}
+
+type Timeval struct {
+	Sec  int32
+	Usec int32
+}
+
+type Timeval32 [0]byte
+
+type Rusage struct {
+	Utime    Timeval
+	Stime    Timeval
+	Maxrss   int32
+	Ixrss    int32
+	Idrss    int32
+	Isrss    int32
+	Minflt   int32
+	Majflt   int32
+	Nswap    int32
+	Inblock  int32
+	Oublock  int32
+	Msgsnd   int32
+	Msgrcv   int32
+	Nsignals int32
+	Nvcsw    int32
+	Nivcsw   int32
+}
+
+type Rlimit struct {
+	Cur uint64
+	Max uint64
+}
+
+type _Gid_t uint32
+
+type Stat_t struct {
+	Dev           int32
+	Mode          uint16
+	Nlink         uint16
+	Ino           uint64
+	Uid           uint32
+	Gid           uint32
+	Rdev          int32
+	Atimespec     Timespec
+	Mtimespec     Timespec
+	Ctimespec     Timespec
+	Birthtimespec Timespec
+	Size          int64
+	Blocks        int64
+	Blksize       int32
+	Flags         uint32
+	Gen           uint32
+	Lspare        int32
+	Qspare        [2]int64
+}
+
+type Statfs_t struct {
+	Bsize       uint32
+	Iosize      int32
+	Blocks      uint64
+	Bfree       uint64
+	Bavail      uint64
+	Files       uint64
+	Ffree       uint64
+	Fsid        Fsid
+	Owner       uint32
+	Type        uint32
+	Flags       uint32
+	Fssubtype   uint32
+	Fstypename  [16]int8
+	Mntonname   [1024]int8
+	Mntfromname [1024]int8
+	Reserved    [8]uint32
+}
+
+type Flock_t struct {
+	Start  int64
+	Len    int64
+	Pid    int32
+	Type   int16
+	Whence int16
+}
+
+type Fstore_t struct {
+	Flags      uint32
+	Posmode    int32
+	Offset     int64
+	Length     int64
+	Bytesalloc int64
+}
+
+type Radvisory_t struct {
+	Offset int64
+	Count  int32
+}
+
+type Fbootstraptransfer_t struct {
+	Offset int64
+	Length uint32
+	Buffer *byte
+}
+
+type Log2phys_t struct {
+	Flags       uint32
+	Contigbytes int64
+	Devoffset   int64
+}
+
+type Fsid struct {
+	Val [2]int32
+}
+
+type Dirent struct {
+	Ino       uint64
+	Seekoff   uint64
+	Reclen    uint16
+	Namlen    uint16
+	Type      uint8
+	Name      [1024]int8
+	Pad_cgo_0 [3]byte
+}
+
+type RawSockaddrInet4 struct {
+	Len    uint8
+	Family uint8
+	Port   uint16
+	Addr   [4]byte /* in_addr */
+	Zero   [8]int8
+}
+
+type RawSockaddrInet6 struct {
+	Len      uint8
+	Family   uint8
+	Port     uint16
+	Flowinfo uint32
+	Addr     [16]byte /* in6_addr */
+	Scope_id uint32
+}
+
+type RawSockaddrUnix struct {
+	Len    uint8
+	Family uint8
+	Path   [104]int8
+}
+
+type RawSockaddrDatalink struct {
+	Len    uint8
+	Family uint8
+	Index  uint16
+	Type   uint8
+	Nlen   uint8
+	Alen   uint8
+	Slen   uint8
+	Data   [12]int8
+}
+
+type RawSockaddr struct {
+	Len    uint8
+	Family uint8
+	Data   [14]int8
+}
+
+type RawSockaddrAny struct {
+	Addr RawSockaddr
+	Pad  [92]int8
+}
+
+type _Socklen uint32
+
+type Linger struct {
+	Onoff  int32
+	Linger int32
+}
+
+type Iovec struct {
+	Base *byte
+	Len  uint32
+}
+
+type IPMreq struct {
+	Multiaddr [4]byte /* in_addr */
+	Interface [4]byte /* in_addr */
+}
+
+type IPv6Mreq struct {
+	Multiaddr [16]byte /* in6_addr */
+	Interface uint32
+}
+
+type Msghdr struct {
+	Name       *byte
+	Namelen    uint32
+	Iov        *Iovec
+	Iovlen     int32
+	Control    *byte
+	Controllen uint32
+	Flags      int32
+}
+
+type Cmsghdr struct {
+	Len   uint32
+	Level int32
+	Type  int32
+}
+
+type Inet4Pktinfo struct {
+	Ifindex  uint32
+	Spec_dst [4]byte /* in_addr */
+	Addr     [4]byte /* in_addr */
+}
+
+type Inet6Pktinfo struct {
+	Addr    [16]byte /* in6_addr */
+	Ifindex uint32
+}
+
+type IPv6MTUInfo struct {
+	Addr RawSockaddrInet6
+	Mtu  uint32
+}
+
+type ICMPv6Filter struct {
+	Filt [8]uint32
+}
+
+const (
+	SizeofSockaddrInet4    = 0x10
+	SizeofSockaddrInet6    = 0x1c
+	SizeofSockaddrAny      = 0x6c
+	SizeofSockaddrUnix     = 0x6a
+	SizeofSockaddrDatalink = 0x14
+	SizeofLinger           = 0x8
+	SizeofIPMreq           = 0x8
+	SizeofIPv6Mreq         = 0x14
+	SizeofMsghdr           = 0x1c
+	SizeofCmsghdr          = 0xc
+	SizeofInet4Pktinfo     = 0xc
+	SizeofInet6Pktinfo     = 0x14
+	SizeofIPv6MTUInfo      = 0x20
+	SizeofICMPv6Filter     = 0x20
+)
+
+const (
+	PTRACE_TRACEME = 0x0
+	PTRACE_CONT    = 0x7
+	PTRACE_KILL    = 0x8
+)
+
+type Kevent_t struct {
+	Ident  uint32
+	Filter int16
+	Flags  uint16
+	Fflags uint32
+	Data   int32
+	Udata  *byte
+}
+
+type FdSet struct {
+	Bits [32]int32
+}
+
+const (
+	SizeofIfMsghdr    = 0x70
+	SizeofIfData      = 0x60
+	SizeofIfaMsghdr   = 0x14
+	SizeofIfmaMsghdr  = 0x10
+	SizeofIfmaMsghdr2 = 0x14
+	SizeofRtMsghdr    = 0x5c
+	SizeofRtMetrics   = 0x38
+)
+
+type IfMsghdr struct {
+	Msglen    uint16
+	Version   uint8
+	Type      uint8
+	Addrs     int32
+	Flags     int32
+	Index     uint16
+	Pad_cgo_0 [2]byte
+	Data      IfData
+}
+
+type IfData struct {
+	Type       uint8
+	Typelen    uint8
+	Physical   uint8
+	Addrlen    uint8
+	Hdrlen     uint8
+	Recvquota  uint8
+	Xmitquota  uint8
+	Unused1    uint8
+	Mtu        uint32
+	Metric     uint32
+	Baudrate   uint32
+	Ipackets   uint32
+	Ierrors    uint32
+	Opackets   uint32
+	Oerrors    uint32
+	Collisions uint32
+	Ibytes     uint32
+	Obytes     uint32
+	Imcasts    uint32
+	Omcasts    uint32
+	Iqdrops    uint32
+	Noproto    uint32
+	Recvtiming uint32
+	Xmittiming uint32
+	Lastchange Timeval
+	Unused2    uint32
+	Hwassist   uint32
+	Reserved1  uint32
+	Reserved2  uint32
+}
+
+type IfaMsghdr struct {
+	Msglen    uint16
+	Version   uint8
+	Type      uint8
+	Addrs     int32
+	Flags     int32
+	Index     uint16
+	Pad_cgo_0 [2]byte
+	Metric    int32
+}
+
+type IfmaMsghdr struct {
+	Msglen    uint16
+	Version   uint8
+	Type      uint8
+	Addrs     int32
+	Flags     int32
+	Index     uint16
+	Pad_cgo_0 [2]byte
+}
+
+type IfmaMsghdr2 struct {
+	Msglen    uint16
+	Version   uint8
+	Type      uint8
+	Addrs     int32
+	Flags     int32
+	Index     uint16
+	Pad_cgo_0 [2]byte
+	Refcount  int32
+}
+
+type RtMsghdr struct {
+	Msglen    uint16
+	Version   uint8
+	Type      uint8
+	Index     uint16
+	Pad_cgo_0 [2]byte
+	Flags     int32
+	Addrs     int32
+	Pid       int32
+	Seq       int32
+	Errno     int32
+	Use       int32
+	Inits     uint32
+	Rmx       RtMetrics
+}
+
+type RtMetrics struct {
+	Locks    uint32
+	Mtu      uint32
+	Hopcount uint32
+	Expire   int32
+	Recvpipe uint32
+	Sendpipe uint32
+	Ssthresh uint32
+	Rtt      uint32
+	Rttvar   uint32
+	Pksent   uint32
+	Filler   [4]uint32
+}
+
+const (
+	SizeofBpfVersion = 0x4
+	SizeofBpfStat    = 0x8
+	SizeofBpfProgram = 0x8
+	SizeofBpfInsn    = 0x8
+	SizeofBpfHdr     = 0x14
+)
+
+type BpfVersion struct {
+	Major uint16
+	Minor uint16
+}
+
+type BpfStat struct {
+	Recv uint32
+	Drop uint32
+}
+
+type BpfProgram struct {
+	Len   uint32
+	Insns *BpfInsn
+}
+
+type BpfInsn struct {
+	Code uint16
+	Jt   uint8
+	Jf   uint8
+	K    uint32
+}
+
+type BpfHdr struct {
+	Tstamp    Timeval
+	Caplen    uint32
+	Datalen   uint32
+	Hdrlen    uint16
+	Pad_cgo_0 [2]byte
+}
+
+type Termios struct {
+	Iflag  uint32
+	Oflag  uint32
+	Cflag  uint32
+	Lflag  uint32
+	Cc     [20]uint8
+	Ispeed uint32
+	Ospeed uint32
+}
diff --git a/src/syscall/ztypes_windows.go b/src/syscall/ztypes_windows.go
index 4c8a99ab94..e5c732576b 100644
--- a/src/syscall/ztypes_windows.go
+++ b/src/syscall/ztypes_windows.go
@@ -1083,12 +1083,7 @@ type TCPKeepalive struct {
 	Interval uint32
 }
 
-type reparseDataBuffer struct {
-	ReparseTag        uint32
-	ReparseDataLength uint16
-	Reserved          uint16
-
-	// SymbolicLinkReparseBuffer
+type symbolicLinkReparseBuffer struct {
 	SubstituteNameOffset uint16
 	SubstituteNameLength uint16
 	PrintNameOffset      uint16
@@ -1097,9 +1092,27 @@ type reparseDataBuffer struct {
 	PathBuffer           [1]uint16
 }
 
+type mountPointReparseBuffer struct {
+	SubstituteNameOffset uint16
+	SubstituteNameLength uint16
+	PrintNameOffset      uint16
+	PrintNameLength      uint16
+	PathBuffer           [1]uint16
+}
+
+type reparseDataBuffer struct {
+	ReparseTag        uint32
+	ReparseDataLength uint16
+	Reserved          uint16
+
+	// GenericReparseBuffer
+	reparseBuffer byte
+}
+
 const (
 	FSCTL_GET_REPARSE_POINT          = 0x900A8
 	MAXIMUM_REPARSE_DATA_BUFFER_SIZE = 16 * 1024
+	_IO_REPARSE_TAG_MOUNT_POINT      = 0xA0000003
 	IO_REPARSE_TAG_SYMLINK           = 0xA000000C
 	SYMBOLIC_LINK_FLAG_DIRECTORY     = 0x1
 )
diff --git a/src/testing/benchmark.go b/src/testing/benchmark.go
index ffd5376844..62e696d221 100644
--- a/src/testing/benchmark.go
+++ b/src/testing/benchmark.go
@@ -280,6 +280,14 @@ func (r BenchmarkResult) MemString() string {
 		r.AllocedBytesPerOp(), r.AllocsPerOp())
 }
 
+// benchmarkName returns full name of benchmark including procs suffix.
+func benchmarkName(name string, n int) string {
+	if n != 1 {
+		return fmt.Sprintf("%s-%d", name, n)
+	}
+	return name
+}
+
 // An internal function but exported because it is cross-package; part of the implementation
 // of the "go test" command.
 func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
@@ -287,15 +295,30 @@ func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks [
 	if len(*matchBenchmarks) == 0 {
 		return
 	}
+	// Collect matching benchmarks and determine longest name.
+	maxprocs := 1
+	for _, procs := range cpuList {
+		if procs > maxprocs {
+			maxprocs = procs
+		}
+	}
+	maxlen := 0
+	var bs []InternalBenchmark
 	for _, Benchmark := range benchmarks {
 		matched, err := matchString(*matchBenchmarks, Benchmark.Name)
 		if err != nil {
 			fmt.Fprintf(os.Stderr, "testing: invalid regexp for -test.bench: %s\n", err)
 			os.Exit(1)
 		}
-		if !matched {
-			continue
+		if matched {
+			bs = append(bs, Benchmark)
+			benchName := benchmarkName(Benchmark.Name, maxprocs)
+			if l := len(benchName); l > maxlen {
+				maxlen = l
+			}
 		}
+	}
+	for _, Benchmark := range bs {
 		for _, procs := range cpuList {
 			runtime.GOMAXPROCS(procs)
 			b := &B{
@@ -304,11 +327,8 @@ func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks [
 				},
 				benchmark: Benchmark,
 			}
-			benchName := Benchmark.Name
-			if procs != 1 {
-				benchName = fmt.Sprintf("%s-%d", Benchmark.Name, procs)
-			}
-			fmt.Printf("%s\t", benchName)
+			benchName := benchmarkName(Benchmark.Name, procs)
+			fmt.Printf("%-*s\t", maxlen, benchName)
 			r := b.run()
 			if b.failed {
 				// The output could be very long here, but probably isn't.
diff --git a/src/time/sys_unix.go b/src/time/sys_unix.go
index 379e13d6a5..e592415daa 100644
--- a/src/time/sys_unix.go
+++ b/src/time/sys_unix.go
@@ -74,3 +74,5 @@ func preadn(fd uintptr, buf []byte, off int) error {
 	}
 	return nil
 }
+
+func isNotExist(err error) bool { return err == syscall.ENOENT }
diff --git a/src/time/zoneinfo_plan9.go b/src/time/zoneinfo_plan9.go
index 4bb0cb3909..0694f0a990 100644
--- a/src/time/zoneinfo_plan9.go
+++ b/src/time/zoneinfo_plan9.go
@@ -7,7 +7,6 @@
 package time
 
 import (
-	"errors"
 	"runtime"
 	"syscall"
 )
@@ -148,11 +147,12 @@ func initLocal() {
 }
 
 func loadLocation(name string) (*Location, error) {
-	if z, err := loadZoneFile(runtime.GOROOT()+"/lib/time/zoneinfo.zip", name); err == nil {
-		z.name = name
-		return z, nil
+	z, err := loadZoneFile(runtime.GOROOT()+"/lib/time/zoneinfo.zip", name)
+	if err != nil {
+		return nil, err
 	}
-	return nil, errors.New("unknown time zone " + name)
+	z.name = name
+	return z, nil
 }
 
 func forceZipFileForTesting(zipOnly bool) {
diff --git a/src/time/zoneinfo_unix.go b/src/time/zoneinfo_unix.go
index ab7e4612e4..66540969d5 100644
--- a/src/time/zoneinfo_unix.go
+++ b/src/time/zoneinfo_unix.go
@@ -74,11 +74,17 @@ func initLocal() {
 }
 
 func loadLocation(name string) (*Location, error) {
+	var firstErr error
 	for _, zoneDir := range zoneDirs {
 		if z, err := loadZoneFile(zoneDir, name); err == nil {
 			z.name = name
 			return z, nil
+		} else if firstErr == nil && !isNotExist(err) {
+			firstErr = err
 		}
 	}
+	if firstErr != nil {
+		return nil, firstErr
+	}
 	return nil, errors.New("unknown time zone " + name)
 }
diff --git a/src/time/zoneinfo_windows.go b/src/time/zoneinfo_windows.go
index 02d8e0edcc..5077f4bd86 100644
--- a/src/time/zoneinfo_windows.go
+++ b/src/time/zoneinfo_windows.go
@@ -260,11 +260,12 @@ func initLocal() {
 }
 
 func loadLocation(name string) (*Location, error) {
-	if z, err := loadZoneFile(runtime.GOROOT()+`\lib\time\zoneinfo.zip`, name); err == nil {
-		z.name = name
-		return z, nil
+	z, err := loadZoneFile(runtime.GOROOT()+`\lib\time\zoneinfo.zip`, name)
+	if err != nil {
+		return nil, err
 	}
-	return nil, errors.New("unknown time zone " + name)
+	z.name = name
+	return z, nil
 }
 
 func forceZipFileForTesting(zipOnly bool) {