crypto/cipher: add VSX implementation of xorBytes for ppc64x

This change adds asm implementations of xorBytes for ppc64x that
takes advantage of VSX registers and instructions.

name                   old time/op    new time/op     delta
XORBytes/8Bytes-8        16.4ns ± 0%     11.1ns ± 0%   -32.32%  (p=0.000 n=5+4)
XORBytes/128Bytes-8      45.6ns ± 0%     16.2ns ± 0%   -64.50%  (p=0.008 n=5+5)
XORBytes/2048Bytes-8      433ns ±13%      129ns ± 1%   -70.29%  (p=0.000 n=5+4)
XORBytes/32768Bytes-8    7.16µs ± 0%     1.83µs ± 0%   -74.39%  (p=0.008 n=5+5)

name                   old speed      new speed       delta
XORBytes/8Bytes-8       488MB/s ± 0%    721MB/s ± 0%   +47.75%  (p=0.016 n=5+4)
XORBytes/128Bytes-8    2.80GB/s ± 0%   7.89GB/s ± 0%  +181.33%  (p=0.008 n=5+5)
XORBytes/2048Bytes-8   4.77GB/s ±13%  15.87GB/s ± 0%  +232.68%  (p=0.016 n=5+4)
XORBytes/32768Bytes-8  4.58GB/s ± 0%  17.88GB/s ± 0%  +290.47%  (p=0.008 n=5+5)

Change-Id: Ic27d9b858f8ec2d597fdabc68a288d6844eba701
Reviewed-on: https://go-review.googlesource.com/c/145997
Run-TryBot: Carlos Eduardo Seo <cseo@linux.vnet.ibm.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>

3 files changed, 96 insertions, 1 deletions

diff --git a/src/crypto/cipher/xor_generic.go b/src/crypto/cipher/xor_generic.go
index 4d660b0a75..b7de60873c 100644
--- a/src/crypto/cipher/xor_generic.go
+++ b/src/crypto/cipher/xor_generic.go
@@ -2,7 +2,7 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-// +build !amd64
+// +build !amd64,!ppc64,!ppc64le
 
 package cipher
 
diff --git a/src/crypto/cipher/xor_ppc64x.go b/src/crypto/cipher/xor_ppc64x.go
new file mode 100644
index 0000000000..8d2e43d327
--- /dev/null
+++ b/src/crypto/cipher/xor_ppc64x.go
@@ -0,0 +1,29 @@
+// 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.
+
+// +build ppc64 ppc64le
+
+package cipher
+
+// xorBytes xors the bytes in a and b. The destination should have enough
+// space, otherwise xorBytes will panic. Returns the number of bytes xor'd.
+func xorBytes(dst, a, b []byte) int {
+	n := len(a)
+	if len(b) < n {
+		n = len(b)
+	}
+	if n == 0 {
+		return 0
+	}
+	_ = dst[n-1]
+	xorBytesVSX(&dst[0], &a[0], &b[0], n)
+	return n
+}
+
+func xorWords(dst, a, b []byte) {
+	xorBytes(dst, a, b)
+}
+
+//go:noescape
+func xorBytesVSX(dst, a, b *byte, n int)
diff --git a/src/crypto/cipher/xor_ppc64x.s b/src/crypto/cipher/xor_ppc64x.s
new file mode 100644
index 0000000000..af4d08bda3
--- /dev/null
+++ b/src/crypto/cipher/xor_ppc64x.s
@@ -0,0 +1,66 @@
+// 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.
+
+// +build ppc64 ppc64le
+
+#include "textflag.h"
+
+// func xorBytesVSX(dst, a, b *byte, n int)
+TEXT ·xorBytesVSX(SB), NOSPLIT, $0
+	MOVD	dst+0(FP), R3	// R3 = dst
+	MOVD	a+8(FP), R4	// R4 = a
+	MOVD	b+16(FP), R5	// R5 = b
+	MOVD	n+24(FP), R6	// R6 = n
+
+	CMPU	R6, $16, CR7	// Check if n ≥ 16 bytes
+	MOVD	R0, R8		// R8 = index
+	CMPU	R6, $8, CR6	// Check if 8 ≤ n < 16 bytes
+	BGE	CR7, preloop16
+	BLT	CR6, small
+
+	// Case for 8 ≤ n < 16 bytes
+	MOVD	(R4)(R8), R14	// R14 = a[i,...,i+7]
+	MOVD	(R5)(R8), R15	// R15 = b[i,...,i+7]
+	XOR	R14, R15, R16	// R16 = a[] ^ b[]
+	SUB	$8, R6		// n = n - 8
+	MOVD	R16, (R3)(R8)	// Store to dst
+	ADD	$8, R8
+
+	// Check if we're finished
+	CMP	R6, R0
+	BGT	small
+	JMP	done
+
+	// Case for n ≥ 16 bytes
+preloop16:
+	SRD	$4, R6, R7	// Setup loop counter
+	MOVD	R7, CTR
+	ANDCC	$15, R6, R9	// Check for tailing bytes for later
+loop16:
+	LXVD2X		(R4)(R8), VS32		// VS32 = a[i,...,i+15]
+	LXVD2X		(R5)(R8), VS33		// VS33 = b[i,...,i+15]
+	XXLXOR		VS32, VS33, VS34	// VS34 = a[] ^ b[]
+	STXVD2X		VS34, (R3)(R8)		// Store to dst
+	ADD		$16, R8			// Update index
+	BC		16, 0, loop16		// bdnz loop16
+
+	BEQ		CR0, done
+	SLD		$4, R7
+	SUB		R7, R6			// R6 = n - (R7 * 16)
+
+	// Case for n < 8 bytes and tailing bytes from the
+	// previous cases.
+small:
+	MOVD	R6, CTR		// Setup loop counter
+
+loop:
+	MOVBZ	(R4)(R8), R14	// R14 = a[i]
+	MOVBZ	(R5)(R8), R15	// R15 = b[i]
+	XOR	R14, R15, R16	// R16 = a[i] ^ b[i]
+	MOVB	R16, (R3)(R8)	// Store to dst
+	ADD	$1, R8
+	BC	16, 0, loop	// bdnz loop
+
+done:
+	RET