cmd/internal/asm/ppc64: avoid generating exser nops

"OR $0, R31, R31" is the execution serializing nop called "exser"
on ISA 3.1 processors such as Power10.

In general, the "OR $0, Rx, Rx" where Rx != 0 form should be avoided
unless used explicitly for the uarch side-effects.

Change-Id: Id76e3a703c902676ba4a3ffb64dd90dad9a320bf
Reviewed-on: https://go-review.googlesource.com/c/go/+/537855
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Run-TryBot: Paul Murphy <murp@ibm.com>
Reviewed-by: Heschi Kreinick <heschi@google.com>

2 files changed, 37 insertions, 14 deletions

diff --git a/src/cmd/asm/internal/asm/testdata/ppc64.s b/src/cmd/asm/internal/asm/testdata/ppc64.s
index cc8d6c84d3..983a368a99 100644
--- a/src/cmd/asm/internal/asm/testdata/ppc64.s
+++ b/src/cmd/asm/internal/asm/testdata/ppc64.s
@@ -179,7 +179,9 @@ TEXT asmtest(SB),DUPOK|NOSPLIT,$0
 	ADD $-32768, R6                 // 38c68000
 	ADD $-32768, R6, R5             // 38a68000
 	// Hex constant 0xFFFFFFFE00000000
-	ADD $-8589934592, R5            // 3fe0fffe63ff00007bff83e463ff00007cbf2a14 or 0602000038a50000
+	ADD $-8589934592, R5            // 3fe0fffe600000007bff83e4600000007cbf2a14 or 0602000038a50000
+	// Hex constant 0xFFFFFFFE00010001
+	ADD $-8589869055, R5            // 3fe0fffe63ff00017bff83e463ff00017cbf2a14 or 0602000138a50001
 
 	//TODO: this compiles to add r5,r6,r0. It should be addi r5,r6,0.
 	//      this is OK since r0 == $0, but the latter is preferred.
@@ -223,6 +225,8 @@ TEXT asmtest(SB),DUPOK|NOSPLIT,$0
 	OR $-32768, R6, R7              // 3be080007fe73378
 	OR $1234567, R5                 // 641f001263ffd6877fe52b78
 	OR $1234567, R5, R3             // 641f001263ffd6877fe32b78
+	OR $2147483648, R5, R3          // 641f8000600000007fe32b78
+	OR $2147483649, R5, R3          // 641f800063ff00017fe32b78
 	ORIS $255, R3, R4
 
 	XOR $1, R3                      // 68630001
@@ -249,7 +253,6 @@ TEXT asmtest(SB),DUPOK|NOSPLIT,$0
 	CMPB R3,R4,R4                   // 7c6423f8
 	CMPEQB R3,R4,CR6                // 7f0321c0
 
-	// TODO: constants for ADDC?
 	ADD R3, R4                      // 7c841a14
 	ADD R3, R4, R5                  // 7ca41a14
 	ADDC R3, R4                     // 7c841814
@@ -262,6 +265,8 @@ TEXT asmtest(SB),DUPOK|NOSPLIT,$0
 	ADDV R3, R4                     // 7c841e14
 	ADDVCC R3, R4                   // 7c841e15
 	ADDCCC R3, R4, R5               // 7ca41815
+	ADDCCC $65536, R4, R5           // 641f0001600000007cbf2015
+	ADDCCC $65537, R4, R5           // 641f000163ff00017cbf2015
 	ADDME R3, R4                    // 7c8301d4
 	ADDMECC R3, R4                  // 7c8301d5
 	ADDMEV R3, R4                   // 7c8305d4
@@ -315,6 +320,8 @@ TEXT asmtest(SB),DUPOK|NOSPLIT,$0
 	SUBECC R3, R4, R5               // 7ca32111
 	SUBEV R3, R4, R5                // 7ca32510
 	SUBEVCC R3, R4, R5              // 7ca32511
+	SUBC R3, $65536, R4             // 3fe00001600000007c83f810
+	SUBC R3, $65537, R4             // 3fe0000163ff00017c83f810
 
 	MULLW R3, R4                    // 7c8419d6
 	MULLW R3, R4, R5                // 7ca419d6
diff --git a/src/cmd/internal/obj/ppc64/asm9.go b/src/cmd/internal/obj/ppc64/asm9.go
index 73642bd209..dcecb26d00 100644
--- a/src/cmd/internal/obj/ppc64/asm9.go
+++ b/src/cmd/internal/obj/ppc64/asm9.go
@@ -65,6 +65,11 @@ const (
 	PFX_R_PCREL = 1 // Offset is relative to PC, RA should be 0
 )
 
+const (
+	// The preferred hardware nop instruction.
+	NOP = 0x60000000
+)
+
 type Optab struct {
 	as    obj.As // Opcode
 	a1    uint8  // p.From argument (obj.Addr). p is of type obj.Prog.
@@ -831,7 +836,6 @@ func span9(ctxt *obj.Link, cursym *obj.LSym, newprog obj.ProgAlloc) {
 	// lay out the code, emitting code and data relocations.
 
 	bp := c.cursym.P
-	nop := LOP_IRR(OP_ORI, REGZERO, REGZERO, 0)
 	var i int32
 	for p := c.cursym.Func().Text.Link; p != nil; p = p.Link {
 		c.pc = p.Pc
@@ -846,13 +850,13 @@ func span9(ctxt *obj.Link, cursym *obj.LSym, newprog obj.ProgAlloc) {
 			if v > 0 {
 				// Same padding instruction for all
 				for i = 0; i < int32(v/4); i++ {
-					c.ctxt.Arch.ByteOrder.PutUint32(bp, nop)
+					c.ctxt.Arch.ByteOrder.PutUint32(bp, NOP)
 					bp = bp[4:]
 				}
 			}
 		} else {
 			if p.Mark&PFX_X64B != 0 {
-				c.ctxt.Arch.ByteOrder.PutUint32(bp, nop)
+				c.ctxt.Arch.ByteOrder.PutUint32(bp, NOP)
 				bp = bp[4:]
 			}
 			o.asmout(&c, p, o, &out)
@@ -2531,6 +2535,18 @@ func decodeMask64(mask int64) (mb, me uint32, valid bool) {
 	return mb, (me - 1) & 63, valid
 }
 
+// Load the lower 16 bits of a constant into register r.
+func loadl16(r int, d int64) uint32 {
+	v := uint16(d)
+	if v == 0 {
+		// Avoid generating "ori r,r,0", r != 0. Instead, generate the architectually preferred nop.
+		// For example, "ori r31,r31,0" is a special execution serializing nop on Power10 called "exser".
+		return NOP
+	}
+	return LOP_IRR(OP_ORI, uint32(r), uint32(r), uint32(v))
+}
+
+// Load the upper 16 bits of a 32b constant into register r.
 func loadu32(r int, d int64) uint32 {
 	v := int32(d >> 16)
 	if isuint32(uint64(d)) {
@@ -2734,7 +2750,7 @@ func asmout(c *ctxt9, p *obj.Prog, o *Optab, out *[5]uint32) {
 			rel.Add = int64(v)
 			rel.Type = objabi.R_CALLPOWER
 		}
-		o2 = 0x60000000 // nop, sometimes overwritten by ld r2, 24(r1) when dynamic linking
+		o2 = NOP // nop, sometimes overwritten by ld r2, 24(r1) when dynamic linking
 
 	case 13: /* mov[bhwd]{z,} r,r */
 		// This needs to handle "MOV* $0, Rx".  This shows up because $0 also
@@ -2957,14 +2973,14 @@ func asmout(c *ctxt9, p *obj.Prog, o *Optab, out *[5]uint32) {
 		} else if o.size == 12 {
 			// Note, o1 is ADDIS if d is negative, ORIS otherwise.
 			o1 = loadu32(REGTMP, d)                                          // tmp = d & 0xFFFF0000
-			o2 = LOP_IRR(OP_ORI, REGTMP, REGTMP, uint32(int32(d)))           // tmp |= d & 0xFFFF
+			o2 = loadl16(REGTMP, d)                                          // tmp |= d & 0xFFFF
 			o3 = AOP_RRR(c.oprrr(p.As), uint32(p.To.Reg), REGTMP, uint32(r)) // to = from + tmp
 		} else {
 			// For backwards compatibility with GOPPC64 < 10, generate 34b constants in register.
-			o1 = LOP_IRR(OP_ADDIS, REGZERO, REGTMP, uint32(d>>32))  // tmp = sign_extend((d>>32)&0xFFFF0000)
-			o2 = LOP_IRR(OP_ORI, REGTMP, REGTMP, uint32(d>>16))     // tmp |= (d>>16)&0xFFFF
-			o3 = AOP_MD(OP_RLDICR, REGTMP, REGTMP, 16, 63-16)       // tmp <<= 16
-			o4 = LOP_IRR(OP_ORI, REGTMP, REGTMP, uint32(uint16(d))) // tmp |= d&0xFFFF
+			o1 = LOP_IRR(OP_ADDIS, REGZERO, REGTMP, uint32(d>>32)) // tmp = sign_extend((d>>32)&0xFFFF0000)
+			o2 = loadl16(REGTMP, int64(d>>16))                     // tmp |= (d>>16)&0xFFFF
+			o3 = AOP_MD(OP_RLDICR, REGTMP, REGTMP, 16, 63-16)      // tmp <<= 16
+			o4 = loadl16(REGTMP, int64(uint16(d)))                 // tmp |= d&0xFFFF
 			o5 = AOP_RRR(c.oprrr(p.As), uint32(p.To.Reg), REGTMP, uint32(r))
 		}
 
@@ -2985,7 +3001,7 @@ func asmout(c *ctxt9, p *obj.Prog, o *Optab, out *[5]uint32) {
 			o2 = LOP_RRR(c.oprrr(p.As), uint32(p.To.Reg), REGTMP, uint32(r))
 		} else {
 			o1 = loadu32(REGTMP, d)
-			o2 = LOP_IRR(OP_ORI, REGTMP, REGTMP, uint32(int32(d)))
+			o2 = loadl16(REGTMP, d)
 			o3 = LOP_RRR(c.oprrr(p.As), uint32(p.To.Reg), REGTMP, uint32(r))
 		}
 		if p.From.Sym != nil {
@@ -3081,9 +3097,9 @@ func asmout(c *ctxt9, p *obj.Prog, o *Optab, out *[5]uint32) {
 		if p.To.Reg == REGTMP || p.From.Reg == REGTMP {
 			c.ctxt.Diag("can't synthesize large constant\n%v", p)
 		}
-		v := c.regoff(p.GetFrom3())
+		v := c.vregoff(p.GetFrom3())
 		o1 = AOP_IRR(OP_ADDIS, REGTMP, REGZERO, uint32(v)>>16)
-		o2 = LOP_IRR(OP_ORI, REGTMP, REGTMP, uint32(v))
+		o2 = loadl16(REGTMP, v)
 		o3 = AOP_RRR(c.oprrr(p.As), uint32(p.To.Reg), uint32(p.From.Reg), REGTMP)
 		if p.From.Sym != nil {
 			c.ctxt.Diag("%v is not supported", p)