bytes: improve Compare function on amd64 for large byte arrays

This patch contains only loop unrolling change for size > 63B

Following are the performance numbers for various sizes on
On Haswell based system: Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz.

benchcmp go.head.8.25.15.txt go.head.8.25.15.opt.txt
benchmark                       old ns/op     new ns/op     delta
BenchmarkBytesCompare1-4        5.37          5.37          +0.00%
BenchmarkBytesCompare2-4        5.37          5.38          +0.19%
BenchmarkBytesCompare4-4        5.37          5.37          +0.00%
BenchmarkBytesCompare8-4        4.42          4.38          -0.90%
BenchmarkBytesCompare16-4       4.27          4.45          +4.22%
BenchmarkBytesCompare32-4       5.30          5.36          +1.13%
BenchmarkBytesCompare64-4       6.93          6.78          -2.16%
BenchmarkBytesCompare128-4      10.3          9.50          -7.77%
BenchmarkBytesCompare256-4      17.1          13.8          -19.30%
BenchmarkBytesCompare512-4      31.3          22.1          -29.39%
BenchmarkBytesCompare1024-4     62.5          39.0          -37.60%
BenchmarkBytesCompare2048-4     112           73.2          -34.64%

Change-Id: I4eeb1c22732fd62cbac97ba757b0d29f648d4ef1
Reviewed-on: https://go-review.googlesource.com/11871Reviewed-by: Keith Randall <khr@golang.org>

src/bytes/bytes_test.go

@@ -1255,3 +1255,34 @@ func BenchmarkRepeat(b *testing.B) {
 		Repeat([]byte("-"), 80)
 	}
 }
+
+func benchmarkBytesCompare(b *testing.B, n int) {
+	var x = make([]byte, n)
+	var y = make([]byte, n)
+
+	for i := 0; i < n; i++ {
+		x[i] = 'a'
+	}
+
+	for i := 0; i < n; i++ {
+		y[i] = 'a'
+	}
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		Compare(x, y)
+	}
+}
+
+func BenchmarkBytesCompare1(b *testing.B)    { benchmarkBytesCompare(b, 1) }
+func BenchmarkBytesCompare2(b *testing.B)    { benchmarkBytesCompare(b, 2) }
+func BenchmarkBytesCompare4(b *testing.B)    { benchmarkBytesCompare(b, 4) }
+func BenchmarkBytesCompare8(b *testing.B)    { benchmarkBytesCompare(b, 8) }
+func BenchmarkBytesCompare16(b *testing.B)   { benchmarkBytesCompare(b, 16) }
+func BenchmarkBytesCompare32(b *testing.B)   { benchmarkBytesCompare(b, 32) }
+func BenchmarkBytesCompare64(b *testing.B)   { benchmarkBytesCompare(b, 64) }
+func BenchmarkBytesCompare128(b *testing.B)  { benchmarkBytesCompare(b, 128) }
+func BenchmarkBytesCompare256(b *testing.B)  { benchmarkBytesCompare(b, 256) }
+func BenchmarkBytesCompare512(b *testing.B)  { benchmarkBytesCompare(b, 512) }
+func BenchmarkBytesCompare1024(b *testing.B) { benchmarkBytesCompare(b, 1024) }
+func BenchmarkBytesCompare2048(b *testing.B) { benchmarkBytesCompare(b, 2048) }

src/runtime/asm_amd64.s

@@ -1445,6 +1445,8 @@ TEXT runtime·cmpbody(SB),NOSPLIT,$0-0
 	CMPQ	R8, $8
 	JB	small
 
+	CMPQ	R8, $63
+	JA	big_loop
 loop:
 	CMPQ	R8, $16
 	JBE	_0through16
@@ -1459,6 +1461,17 @@ loop:
 	SUBQ	$16, R8
 	JMP	loop
 	
+diff64:
+	ADDQ	$48, SI
+	ADDQ	$48, DI
+	JMP	diff16
+diff48:
+	ADDQ	$32, SI
+	ADDQ	$32, DI
+	JMP	diff16
+diff32:
+	ADDQ	$16, SI
+	ADDQ	$16, DI
 	// AX = bit mask of differences
 diff16:
 	BSFQ	AX, BX	// index of first byte that differs
@@ -1545,6 +1558,43 @@ allsame:
 	MOVQ	AX, (R9)
 	RET
 
+	// this works for >= 64 bytes of data.
+big_loop:
+	MOVOU	(SI), X0
+	MOVOU	(DI), X1
+	PCMPEQB X0, X1
+	PMOVMSKB X1, AX
+	XORQ	$0xffff, AX
+	JNE	diff16
+
+	MOVOU	16(SI), X0
+	MOVOU	16(DI), X1
+	PCMPEQB X0, X1
+	PMOVMSKB X1, AX
+	XORQ	$0xffff, AX
+	JNE	diff32
+
+	MOVOU	32(SI), X0
+	MOVOU	32(DI), X1
+	PCMPEQB X0, X1
+	PMOVMSKB X1, AX
+	XORQ	$0xffff, AX
+	JNE	diff48
+
+	MOVOU	48(SI), X0
+	MOVOU	48(DI), X1
+	PCMPEQB X0, X1
+	PMOVMSKB X1, AX
+	XORQ	$0xffff, AX
+	JNE	diff64
+
+	ADDQ	$64, SI
+	ADDQ	$64, DI
+	SUBQ	$64, R8
+	CMPQ	R8, $64
+	JBE	loop
+	JMP	big_loop
+
 TEXT bytes·IndexByte(SB),NOSPLIT,$0-40
 	MOVQ s+0(FP), SI
 	MOVQ s_len+8(FP), BX