runtime: faster memclr on x86.

Use explicit SSE writes instead of REP STOSQ.

benchmark               old ns/op    new ns/op    delta
BenchmarkMemclr5               22            5  -73.62%
BenchmarkMemclr16              27            5  -78.49%
BenchmarkMemclr64              28            6  -76.43%
BenchmarkMemclr256             34            8  -74.94%
BenchmarkMemclr4096           112           84  -24.73%
BenchmarkMemclr65536         1902         1920   +0.95%

LGTM=dvyukov
R=golang-codereviews, dvyukov
CC=golang-codereviews
https://golang.org/cl/60090044

11 files changed, 324 insertions, 62 deletions

diff --git a/src/cmd/8a/lex.c b/src/cmd/8a/lex.c
index 96804ac039..af816a00cc 100644
--- a/src/cmd/8a/lex.c
+++ b/src/cmd/8a/lex.c
@@ -781,6 +781,7 @@ struct
 	"PSUBW",	LTYPE3,	APSUBW,
 	"PUNPCKHQDQ",	LTYPE3,	APUNPCKHQDQ,
 	"PUNPCKLQDQ",	LTYPE3,	APUNPCKLQDQ,
+	"PXOR",		LTYPE3, APXOR,
 	"RCPPS",	LTYPE3,	ARCPPS,
 	"RCPSS",	LTYPE3,	ARCPSS,
 	"RSQRTPS",	LTYPE3,	ARSQRTPS,
diff --git a/src/cmd/8l/8.out.h b/src/cmd/8l/8.out.h
index 8319482ca3..748096db10 100644
--- a/src/cmd/8l/8.out.h
+++ b/src/cmd/8l/8.out.h
@@ -547,6 +547,7 @@ enum	as
 	APSUBW,
 	APUNPCKHQDQ,
 	APUNPCKLQDQ,
+	APXOR,
 	ARCPPS,
 	ARCPSS,
 	ARSQRTPS,
diff --git a/src/liblink/asm8.c b/src/liblink/asm8.c
index b7d03743d1..d2e50c11c1 100644
--- a/src/liblink/asm8.c
+++ b/src/liblink/asm8.c
@@ -1115,6 +1115,7 @@ static Optab optab[] =
 	{ APSUBW,	yxm,	Pe, 0xf9 },
 	{ APUNPCKHQDQ,	yxm,	Pe, 0x6d },
 	{ APUNPCKLQDQ,	yxm,	Pe, 0x6c },
+	{ APXOR,	yxm,	Pe, 0xef },
 	{ ARCPPS,	yxm,	Pm, 0x53 },
 	{ ARCPSS,	yxm,	Pf3, 0x53 },
 	{ ARSQRTPS,	yxm,	Pm, 0x52 },
diff --git a/src/pkg/runtime/alg.c b/src/pkg/runtime/alg.c
index c3a8396955..623858f7c1 100644
--- a/src/pkg/runtime/alg.c
+++ b/src/pkg/runtime/alg.c
@@ -514,6 +514,11 @@ runtime·equal(Type *t, ...)
 	t->alg->equal((bool*)ret, t->size, x, y);
 }
 
+// Testing adapter for memclr
+void runtime·memclrBytes(Slice s) {
+	runtime·memclr(s.array, s.len);
+}
+
 // Testing adapters for hash quality tests (see hash_test.go)
 void runtime·haveGoodHash(bool res) {
 	res = use_aeshash;
diff --git a/src/pkg/runtime/asm_386.s b/src/pkg/runtime/asm_386.s
index ccd2567fdc..8a945c2d50 100644
--- a/src/pkg/runtime/asm_386.s
+++ b/src/pkg/runtime/asm_386.s
@@ -753,21 +753,6 @@ TEXT runtime·stackcheck(SB), NOSPLIT, $0-0
 	INT	$3
 	RET
 
-TEXT runtime·memclr(SB),NOSPLIT,$0-8
-	MOVL	4(SP), DI		// arg 1 addr
-	MOVL	8(SP), CX		// arg 2 count
-	MOVL	CX, BX
-	ANDL	$3, BX
-	SHRL	$2, CX
-	MOVL	$0, AX
-	CLD
-	REP
-	STOSL
-	MOVL	BX, CX
-	REP
-	STOSB
-	RET
-
 TEXT runtime·getcallerpc(SB),NOSPLIT,$0-4
 	MOVL	x+0(FP),AX		// addr of first arg
 	MOVL	-4(AX),AX		// get calling pc
diff --git a/src/pkg/runtime/asm_amd64.s b/src/pkg/runtime/asm_amd64.s
index 17e91c04db..825fc3254c 100644
--- a/src/pkg/runtime/asm_amd64.s
+++ b/src/pkg/runtime/asm_amd64.s
@@ -794,21 +794,6 @@ TEXT runtime·stackcheck(SB), NOSPLIT, $0-0
 	INT	$3
 	RET
 
-TEXT runtime·memclr(SB),NOSPLIT,$0-16
-	MOVQ	8(SP), DI		// arg 1 addr
-	MOVQ	16(SP), CX		// arg 2 count
-	MOVQ	CX, BX
-	ANDQ	$7, BX
-	SHRQ	$3, CX
-	MOVQ	$0, AX
-	CLD
-	REP
-	STOSQ
-	MOVQ	BX, CX
-	REP
-	STOSB
-	RET
-
 TEXT runtime·getcallerpc(SB),NOSPLIT,$0-8
 	MOVQ	x+0(FP),AX		// addr of first arg
 	MOVQ	-8(AX),AX		// get calling pc
diff --git a/src/pkg/runtime/export_test.go b/src/pkg/runtime/export_test.go
index d170fa72ae..5448ce23a2 100644
--- a/src/pkg/runtime/export_test.go
+++ b/src/pkg/runtime/export_test.go
@@ -84,3 +84,7 @@ func GogoBytes() int32
 
 var hashLoad float64 // declared in hashmap.c
 var HashLoad = &hashLoad
+
+func memclrBytes(b []byte)
+
+var MemclrBytes = memclrBytes
diff --git a/src/pkg/runtime/memclr_386.s b/src/pkg/runtime/memclr_386.s
new file mode 100644
index 0000000000..09b35d7e3c
--- /dev/null
+++ b/src/pkg/runtime/memclr_386.s
@@ -0,0 +1,125 @@
+// 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 "../../cmd/ld/textflag.h"
+
+// void runtime·memclr(void*, uintptr)
+TEXT runtime·memclr(SB), NOSPLIT, $0-8
+	MOVL	ptr+0(FP), DI
+	MOVL	n+4(FP), BX
+	XORL	AX, AX
+
+	// MOVOU seems always faster than REP STOSL.
+clr_tail:
+	TESTL	BX, BX
+	JEQ	clr_0
+	CMPL	BX, $2
+	JBE	clr_1or2
+	CMPL	BX, $4
+	JBE	clr_3or4
+	CMPL	BX, $8
+	JBE	clr_5through8
+	CMPL	BX, $16
+	JBE	clr_9through16
+	TESTL	$0x4000000, runtime·cpuid_edx(SB) // check for sse2
+	JEQ	nosse2
+	PXOR	X0, X0
+	CMPL	BX, $32
+	JBE	clr_17through32
+	CMPL	BX, $64
+	JBE	clr_33through64
+	CMPL	BX, $128
+	JBE	clr_65through128
+	CMPL	BX, $256
+	JBE	clr_129through256
+	// TODO: use branch table and BSR to make this just a single dispatch
+
+clr_loop:
+	MOVOU	X0, 0(DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, 32(DI)
+	MOVOU	X0, 48(DI)
+	MOVOU	X0, 64(DI)
+	MOVOU	X0, 80(DI)
+	MOVOU	X0, 96(DI)
+	MOVOU	X0, 112(DI)
+	MOVOU	X0, 128(DI)
+	MOVOU	X0, 144(DI)
+	MOVOU	X0, 160(DI)
+	MOVOU	X0, 176(DI)
+	MOVOU	X0, 192(DI)
+	MOVOU	X0, 208(DI)
+	MOVOU	X0, 224(DI)
+	MOVOU	X0, 240(DI)
+	SUBL	$256, BX
+	ADDL	$256, DI
+	CMPL	BX, $256
+	JAE	clr_loop
+	JMP	clr_tail
+
+clr_1or2:
+	MOVB	AX, (DI)
+	MOVB	AX, -1(DI)(BX*1)
+clr_0:
+	RET
+clr_3or4:
+	MOVW	AX, (DI)
+	MOVW	AX, -2(DI)(BX*1)
+	RET
+clr_5through8:
+	MOVL	AX, (DI)
+	MOVL	AX, -4(DI)(BX*1)
+	RET
+clr_9through16:
+	MOVL	AX, (DI)
+	MOVL	AX, 4(DI)
+	MOVL	AX, -8(DI)(BX*1)
+	MOVL	AX, -4(DI)(BX*1)
+	RET
+clr_17through32:
+	MOVOU	X0, (DI)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
+clr_33through64:
+	MOVOU	X0, (DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, -32(DI)(BX*1)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
+clr_65through128:
+	MOVOU	X0, (DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, 32(DI)
+	MOVOU	X0, 48(DI)
+	MOVOU	X0, -64(DI)(BX*1)
+	MOVOU	X0, -48(DI)(BX*1)
+	MOVOU	X0, -32(DI)(BX*1)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
+clr_129through256:
+	MOVOU	X0, (DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, 32(DI)
+	MOVOU	X0, 48(DI)
+	MOVOU	X0, 64(DI)
+	MOVOU	X0, 80(DI)
+	MOVOU	X0, 96(DI)
+	MOVOU	X0, 112(DI)
+	MOVOU	X0, -128(DI)(BX*1)
+	MOVOU	X0, -112(DI)(BX*1)
+	MOVOU	X0, -96(DI)(BX*1)
+	MOVOU	X0, -80(DI)(BX*1)
+	MOVOU	X0, -64(DI)(BX*1)
+	MOVOU	X0, -48(DI)(BX*1)
+	MOVOU	X0, -32(DI)(BX*1)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
+nosse2:
+	MOVL	BX, CX
+	SHRL	$2, CX
+	REP
+	STOSL
+	ANDL	$3, BX
+	JNE	clr_tail
+	RET
diff --git a/src/pkg/runtime/memclr_amd64.s b/src/pkg/runtime/memclr_amd64.s
new file mode 100644
index 0000000000..8953a396b1
--- /dev/null
+++ b/src/pkg/runtime/memclr_amd64.s
@@ -0,0 +1,114 @@
+// 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 "../../cmd/ld/textflag.h"
+
+// void runtime·memclr(void*, uintptr)
+TEXT runtime·memclr(SB), NOSPLIT, $0-16
+	MOVQ	ptr+0(FP), DI
+	MOVQ	n+8(FP), BX
+	XORQ	AX, AX
+
+	// MOVOU seems always faster than REP STOSQ.
+clr_tail:
+	TESTQ	BX, BX
+	JEQ	clr_0
+	CMPQ	BX, $2
+	JBE	clr_1or2
+	CMPQ	BX, $4
+	JBE	clr_3or4
+	CMPQ	BX, $8
+	JBE	clr_5through8
+	CMPQ	BX, $16
+	JBE	clr_9through16
+	PXOR	X0, X0
+	CMPQ	BX, $32
+	JBE	clr_17through32
+	CMPQ	BX, $64
+	JBE	clr_33through64
+	CMPQ	BX, $128
+	JBE	clr_65through128
+	CMPQ	BX, $256
+	JBE	clr_129through256
+	// TODO: use branch table and BSR to make this just a single dispatch
+	// TODO: for really big clears, use MOVNTDQ.
+
+clr_loop:
+	MOVOU	X0, 0(DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, 32(DI)
+	MOVOU	X0, 48(DI)
+	MOVOU	X0, 64(DI)
+	MOVOU	X0, 80(DI)
+	MOVOU	X0, 96(DI)
+	MOVOU	X0, 112(DI)
+	MOVOU	X0, 128(DI)
+	MOVOU	X0, 144(DI)
+	MOVOU	X0, 160(DI)
+	MOVOU	X0, 176(DI)
+	MOVOU	X0, 192(DI)
+	MOVOU	X0, 208(DI)
+	MOVOU	X0, 224(DI)
+	MOVOU	X0, 240(DI)
+	SUBQ	$256, BX
+	ADDQ	$256, DI
+	CMPQ	BX, $256
+	JAE	clr_loop
+	JMP	clr_tail
+
+clr_1or2:
+	MOVB	AX, (DI)
+	MOVB	AX, -1(DI)(BX*1)
+clr_0:
+	RET
+clr_3or4:
+	MOVW	AX, (DI)
+	MOVW	AX, -2(DI)(BX*1)
+	RET
+clr_5through8:
+	MOVL	AX, (DI)
+	MOVL	AX, -4(DI)(BX*1)
+	RET
+clr_9through16:
+	MOVQ	AX, (DI)
+	MOVQ	AX, -8(DI)(BX*1)
+	RET
+clr_17through32:
+	MOVOU	X0, (DI)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
+clr_33through64:
+	MOVOU	X0, (DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, -32(DI)(BX*1)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
+clr_65through128:
+	MOVOU	X0, (DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, 32(DI)
+	MOVOU	X0, 48(DI)
+	MOVOU	X0, -64(DI)(BX*1)
+	MOVOU	X0, -48(DI)(BX*1)
+	MOVOU	X0, -32(DI)(BX*1)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
+clr_129through256:
+	MOVOU	X0, (DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, 32(DI)
+	MOVOU	X0, 48(DI)
+	MOVOU	X0, 64(DI)
+	MOVOU	X0, 80(DI)
+	MOVOU	X0, 96(DI)
+	MOVOU	X0, 112(DI)
+	MOVOU	X0, -128(DI)(BX*1)
+	MOVOU	X0, -112(DI)(BX*1)
+	MOVOU	X0, -96(DI)(BX*1)
+	MOVOU	X0, -80(DI)(BX*1)
+	MOVOU	X0, -64(DI)(BX*1)
+	MOVOU	X0, -48(DI)(BX*1)
+	MOVOU	X0, -32(DI)(BX*1)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
diff --git a/src/pkg/runtime/memclr_arm.s b/src/pkg/runtime/memclr_arm.s
index d5ff75d7a1..b19ea72a3f 100644
--- a/src/pkg/runtime/memclr_arm.s
+++ b/src/pkg/runtime/memclr_arm.s
@@ -40,12 +40,6 @@ TEXT runtime·memclr(SB),NOSPLIT,$0-8
 	CMP	$4, R(N)		/* need at least 4 bytes to copy */
 	BLT	_1tail
 
-	AND	$0xFF, R(0)		/* it's a byte */
-	SLL	$8, R(0), R(TMP)	/* replicate to a word */
-	ORR	R(TMP), R(0)
-	SLL	$16, R(0), R(TMP)
-	ORR	R(TMP), R(0)
-
 _4align:				/* align on 4 */
 	AND.S	$3, R(TO), R(TMP)
 	BEQ	_4aligned
diff --git a/src/pkg/runtime/memmove_test.go b/src/pkg/runtime/memmove_test.go
index 9525f06826..5c01aac97a 100644
--- a/src/pkg/runtime/memmove_test.go
+++ b/src/pkg/runtime/memmove_test.go
@@ -5,6 +5,7 @@
 package runtime_test
 
 import (
+	. "runtime"
 	"testing"
 )
 
@@ -80,7 +81,7 @@ func TestMemmoveAlias(t *testing.T) {
 	}
 }
 
-func bmMemmove(n int, b *testing.B) {
+func bmMemmove(b *testing.B, n int) {
 	x := make([]byte, n)
 	y := make([]byte, n)
 	b.SetBytes(int64(n))
@@ -89,28 +90,74 @@ func bmMemmove(n int, b *testing.B) {
 	}
 }
 
-func BenchmarkMemmove0(b *testing.B)    { bmMemmove(0, b) }
-func BenchmarkMemmove1(b *testing.B)    { bmMemmove(1, b) }
-func BenchmarkMemmove2(b *testing.B)    { bmMemmove(2, b) }
-func BenchmarkMemmove3(b *testing.B)    { bmMemmove(3, b) }
-func BenchmarkMemmove4(b *testing.B)    { bmMemmove(4, b) }
-func BenchmarkMemmove5(b *testing.B)    { bmMemmove(5, b) }
-func BenchmarkMemmove6(b *testing.B)    { bmMemmove(6, b) }
-func BenchmarkMemmove7(b *testing.B)    { bmMemmove(7, b) }
-func BenchmarkMemmove8(b *testing.B)    { bmMemmove(8, b) }
-func BenchmarkMemmove9(b *testing.B)    { bmMemmove(9, b) }
-func BenchmarkMemmove10(b *testing.B)   { bmMemmove(10, b) }
-func BenchmarkMemmove11(b *testing.B)   { bmMemmove(11, b) }
-func BenchmarkMemmove12(b *testing.B)   { bmMemmove(12, b) }
-func BenchmarkMemmove13(b *testing.B)   { bmMemmove(13, b) }
-func BenchmarkMemmove14(b *testing.B)   { bmMemmove(14, b) }
-func BenchmarkMemmove15(b *testing.B)   { bmMemmove(15, b) }
-func BenchmarkMemmove16(b *testing.B)   { bmMemmove(16, b) }
-func BenchmarkMemmove32(b *testing.B)   { bmMemmove(32, b) }
-func BenchmarkMemmove64(b *testing.B)   { bmMemmove(64, b) }
-func BenchmarkMemmove128(b *testing.B)  { bmMemmove(128, b) }
-func BenchmarkMemmove256(b *testing.B)  { bmMemmove(256, b) }
-func BenchmarkMemmove512(b *testing.B)  { bmMemmove(512, b) }
-func BenchmarkMemmove1024(b *testing.B) { bmMemmove(1024, b) }
-func BenchmarkMemmove2048(b *testing.B) { bmMemmove(2048, b) }
-func BenchmarkMemmove4096(b *testing.B) { bmMemmove(4096, b) }
+func BenchmarkMemmove0(b *testing.B)    { bmMemmove(b, 0) }
+func BenchmarkMemmove1(b *testing.B)    { bmMemmove(b, 1) }
+func BenchmarkMemmove2(b *testing.B)    { bmMemmove(b, 2) }
+func BenchmarkMemmove3(b *testing.B)    { bmMemmove(b, 3) }
+func BenchmarkMemmove4(b *testing.B)    { bmMemmove(b, 4) }
+func BenchmarkMemmove5(b *testing.B)    { bmMemmove(b, 5) }
+func BenchmarkMemmove6(b *testing.B)    { bmMemmove(b, 6) }
+func BenchmarkMemmove7(b *testing.B)    { bmMemmove(b, 7) }
+func BenchmarkMemmove8(b *testing.B)    { bmMemmove(b, 8) }
+func BenchmarkMemmove9(b *testing.B)    { bmMemmove(b, 9) }
+func BenchmarkMemmove10(b *testing.B)   { bmMemmove(b, 10) }
+func BenchmarkMemmove11(b *testing.B)   { bmMemmove(b, 11) }
+func BenchmarkMemmove12(b *testing.B)   { bmMemmove(b, 12) }
+func BenchmarkMemmove13(b *testing.B)   { bmMemmove(b, 13) }
+func BenchmarkMemmove14(b *testing.B)   { bmMemmove(b, 14) }
+func BenchmarkMemmove15(b *testing.B)   { bmMemmove(b, 15) }
+func BenchmarkMemmove16(b *testing.B)   { bmMemmove(b, 16) }
+func BenchmarkMemmove32(b *testing.B)   { bmMemmove(b, 32) }
+func BenchmarkMemmove64(b *testing.B)   { bmMemmove(b, 64) }
+func BenchmarkMemmove128(b *testing.B)  { bmMemmove(b, 128) }
+func BenchmarkMemmove256(b *testing.B)  { bmMemmove(b, 256) }
+func BenchmarkMemmove512(b *testing.B)  { bmMemmove(b, 512) }
+func BenchmarkMemmove1024(b *testing.B) { bmMemmove(b, 1024) }
+func BenchmarkMemmove2048(b *testing.B) { bmMemmove(b, 2048) }
+func BenchmarkMemmove4096(b *testing.B) { bmMemmove(b, 4096) }
+
+func TestMemclr(t *testing.T) {
+	size := 512
+	if testing.Short() {
+		size = 128 + 16
+	}
+	mem := make([]byte, size)
+	for i := 0; i < size; i++ {
+		mem[i] = 0xee
+	}
+	for n := 0; n < size; n++ {
+		for x := 0; x <= size-n; x++ { // offset in mem
+			MemclrBytes(mem[x : x+n])
+			for i := 0; i < x; i++ {
+				if mem[i] != 0xee {
+					t.Fatalf("overwrite prefix mem[%d] = %d", i, mem[i])
+				}
+			}
+			for i := x; i < x+n; i++ {
+				if mem[i] != 0 {
+					t.Fatalf("failed clear mem[%d] = %d", i, mem[i])
+				}
+				mem[i] = 0xee
+			}
+			for i := x + n; i < size; i++ {
+				if mem[i] != 0xee {
+					t.Fatalf("overwrite suffix mem[%d] = %d", i, mem[i])
+				}
+			}
+		}
+	}
+}
+
+func bmMemclr(b *testing.B, n int) {
+	x := make([]byte, n)
+	b.SetBytes(int64(n))
+	for i := 0; i < b.N; i++ {
+		MemclrBytes(x)
+	}
+}
+func BenchmarkMemclr5(b *testing.B)     { bmMemclr(b, 5) }
+func BenchmarkMemclr16(b *testing.B)    { bmMemclr(b, 16) }
+func BenchmarkMemclr64(b *testing.B)    { bmMemclr(b, 64) }
+func BenchmarkMemclr256(b *testing.B)   { bmMemclr(b, 256) }
+func BenchmarkMemclr4096(b *testing.B)  { bmMemclr(b, 4096) }
+func BenchmarkMemclr65536(b *testing.B) { bmMemclr(b, 65536) }