common/bitutil: optimize ORBytes

common/bitutil/bitutil.go

@@ -77,32 +77,7 @@ func safeANDBytes(dst, a, b []byte) int {
 // ORBytes ors the bytes in a and b. The destination is assumed to have enough
 // space. Returns the number of bytes or'd.
 func ORBytes(dst, a, b []byte) int {
-	if supportsUnaligned {
+	return orBytes(dst, a, b)
-		return fastORBytes(dst, a, b)
-	}
-	return safeORBytes(dst, a, b)
-}
-
-// fastORBytes ors in bulk. It only works on architectures that support
-// unaligned read/writes.
-func fastORBytes(dst, a, b []byte) int {
-	n := len(a)
-	if len(b) < n {
-		n = len(b)
-	}
-	w := n / wordSize
-	if w > 0 {
-		dw := *(*[]uintptr)(unsafe.Pointer(&dst))
-		aw := *(*[]uintptr)(unsafe.Pointer(&a))
-		bw := *(*[]uintptr)(unsafe.Pointer(&b))
-		for i := 0; i < w; i++ {
-			dw[i] = aw[i] | bw[i]
-		}
-	}
-	for i := n - n%wordSize; i < n; i++ {
-		dst[i] = a[i] | b[i]
-	}
-	return n
 }
 
 // safeORBytes ors one by one. It works on all architectures, independent if

common/bitutil/or_amd64.s

@@ -0,0 +1,59 @@
+// Copyright 2018 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.
+// inspired by: https://github.com/golang/go/blob/4a3cef2036097d323b6cc0bbe90fc4d8c7588660/src/crypto/internal/fips140/subtle/xor_amd64.s
+
+//go:build !purego
+
+#include "textflag.h"
+
+// func orBytesASM(dst, a, b *byte, n int)
+TEXT ·orBytesASM(SB), NOSPLIT, $0
+	MOVQ  dst+0(FP), BX
+	MOVQ  a+8(FP), SI
+	MOVQ  b+16(FP), CX
+	MOVQ  n+24(FP), DX
+	TESTQ $15, DX            // AND 15 & len, if not zero jump to not_aligned.
+	JNZ   not_aligned
+
+aligned:
+	MOVQ $0, AX // position in slices
+
+	PCALIGN $16
+loop16b:
+	MOVOU (SI)(AX*1), X0   // OR 16byte forwards.
+	MOVOU (CX)(AX*1), X1
+	POR   X1, X0
+	MOVOU X0, (BX)(AX*1)
+	ADDQ  $16, AX
+	CMPQ  DX, AX
+	JNE   loop16b
+	RET
+
+	PCALIGN $16
+loop_1b:
+	SUBQ  $1, DX           // OR 1byte backwards.
+	MOVB  (SI)(DX*1), DI
+	MOVB  (CX)(DX*1), AX
+	ORB   AX, DI
+	MOVB  DI, (BX)(DX*1)
+	TESTQ $7, DX           // AND 7 & len, if not zero jump to loop_1b.
+	JNZ   loop_1b
+	CMPQ  DX, $0           // if len is 0, ret.
+	JE    ret
+	TESTQ $15, DX          // AND 15 & len, if zero jump to aligned.
+	JZ    aligned
+
+not_aligned:
+	TESTQ $7, DX           // AND $7 & len, if not zero jump to loop_1b.
+	JNE   loop_1b
+	SUBQ  $8, DX           // OR 8bytes backwards.
+	MOVQ  (SI)(DX*1), DI
+	MOVQ  (CX)(DX*1), AX
+	ORQ   AX, DI
+	MOVQ  DI, (BX)(DX*1)
+	CMPQ  DX, $16          // if len is greater or equal 16 here, it must be aligned.
+	JGE   aligned
+
+ret:
+	RET

common/bitutil/or_arm64.s

@@ -0,0 +1,70 @@
+// Copyright 2020 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.
+// inspired by: https://github.com/golang/go/blob/4a3cef2036097d323b6cc0bbe90fc4d8c7588660/src/crypto/internal/fips140/subtle/xor_arm64.s
+
+//go:build !purego
+
+#include "textflag.h"
+
+// func orBytesASM(dst, a, b *byte, n int)
+TEXT ·orBytesASM(SB), NOSPLIT|NOFRAME, $0
+	MOVD	dst+0(FP), R0
+	MOVD	a+8(FP), R1
+	MOVD	b+16(FP), R2
+	MOVD	n+24(FP), R3
+	CMP	$64, R3
+	BLT	tail
+loop_64:
+	VLD1.P	64(R1), [V0.B16, V1.B16, V2.B16, V3.B16]
+	VLD1.P	64(R2), [V4.B16, V5.B16, V6.B16, V7.B16]
+	VORR	V0.B16, V4.B16, V4.B16
+	VORR	V1.B16, V5.B16, V5.B16
+	VORR	V2.B16, V6.B16, V6.B16
+	VORR	V3.B16, V7.B16, V7.B16
+	VST1.P	[V4.B16, V5.B16, V6.B16, V7.B16], 64(R0)
+	SUBS	$64, R3
+	CMP	$64, R3
+	BGE	loop_64
+tail:
+	// quick end
+	CBZ	R3, end
+	TBZ	$5, R3, less_than32
+	VLD1.P	32(R1), [V0.B16, V1.B16]
+	VLD1.P	32(R2), [V2.B16, V3.B16]
+	VORR	V0.B16, V2.B16, V2.B16
+	VORR	V1.B16, V3.B16, V3.B16
+	VST1.P	[V2.B16, V3.B16], 32(R0)
+less_than32:
+	TBZ	$4, R3, less_than16
+	LDP.P	16(R1), (R11, R12)
+	LDP.P	16(R2), (R13, R14)
+	ORR	R11, R13, R13
+	ORR	R12, R14, R14
+	STP.P	(R13, R14), 16(R0)
+less_than16:
+	TBZ	$3, R3, less_than8
+	MOVD.P	8(R1), R11
+	MOVD.P	8(R2), R12
+	ORR	R11, R12, R12
+	MOVD.P	R12, 8(R0)
+less_than8:
+	TBZ	$2, R3, less_than4
+	MOVWU.P	4(R1), R13
+	MOVWU.P	4(R2), R14
+	ORRW	R13, R14, R14
+	MOVWU.P	R14, 4(R0)
+less_than4:
+	TBZ	$1, R3, less_than2
+	MOVHU.P	2(R1), R15
+	MOVHU.P	2(R2), R16
+	ORRW	R15, R16, R16
+	MOVHU.P	R16, 2(R0)
+less_than2:
+	TBZ	$0, R3, end
+	MOVBU	(R1), R17
+	MOVBU	(R2), R19
+	ORRW	R17, R19, R19
+	MOVBU	R19, (R0)
+end:
+	RET

common/bitutil/or_asm.go

@@ -0,0 +1,20 @@
+// Copyright 2018 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.
+// inspired by: https://github.com/golang/go/blob/4a3cef2036097d323b6cc0bbe90fc4d8c7588660/src/crypto/internal/fips140/subtle/xor_asm.go
+
+//go:build (amd64 || arm64) && !purego
+
+package bitutil
+
+func orBytes(dst, a, b []byte) int {
+	n := min(len(a), len(b))
+	if n == 0 {
+		return 0
+	}
+	orBytesASM(&dst[0], &a[0], &b[0], n)
+	return n
+}
+
+//go:noescape
+func orBytesASM(dst, a, b *byte, n int)

common/bitutil/or_generic.go

@@ -0,0 +1,38 @@
+// Copyright 2013 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:build (!amd64 && !arm64) || purego
+
+package bitutil
+
+import "unsafe"
+
+func orBytes(dst, a, b []byte) int {
+	if supportsUnaligned {
+		return fastORBytes(dst, a, b)
+	}
+	return safeORBytes(dst, a, b)
+}
+
+// fastORBytes ors in bulk. It only works on architectures that support
+// unaligned read/writes.
+func fastORBytes(dst, a, b []byte) int {
+	n := len(a)
+	if len(b) < n {
+		n = len(b)
+	}
+	w := n / wordSize
+	if w > 0 {
+		dw := *(*[]uintptr)(unsafe.Pointer(&dst))
+		aw := *(*[]uintptr)(unsafe.Pointer(&a))
+		bw := *(*[]uintptr)(unsafe.Pointer(&b))
+		for i := 0; i < w; i++ {
+			dw[i] = aw[i] | bw[i]
+		}
+	}
+	for i := n - n%wordSize; i < n; i++ {
+		dst[i] = a[i] | b[i]
+	}
+	return n
+}