Commit 47c58b46 authored by Michael Munday's avatar Michael Munday

bytes, strings: optimize multi-byte index operations on s390x

Use vector instructions to speed up indexing operations for short
strings (64 bytes or less).

bytes_s390x.go and strings_s390x.go are based on their amd64
equivalents in CL 31690.

bytes package:

name                   old time/op    new time/op    delta
Index/10                 40.3ns ± 7%    11.3ns ± 4%    -72.06%  (p=0.000 n=10+10)
Index/32                  196ns ± 1%      27ns ± 2%    -86.25%  (p=0.000 n=10+10)
Index/4K                 28.9µs ± 1%     1.5µs ± 2%    -94.94%    (p=0.000 n=9+9)
Index/4M                 30.1ms ± 2%     1.5ms ± 3%    -94.94%  (p=0.000 n=10+10)
Index/64M                 549ms ±13%      28ms ± 3%    -94.87%   (p=0.000 n=10+9)
IndexEasy/10             18.8ns ±11%    11.5ns ± 2%    -38.81%  (p=0.000 n=10+10)
IndexEasy/32             23.6ns ± 6%    28.1ns ± 3%    +19.29%  (p=0.000 n=10+10)
IndexEasy/4K              251ns ± 5%     223ns ± 8%    -11.04%  (p=0.000 n=10+10)
IndexEasy/4M              318µs ± 9%     266µs ± 8%    -16.42%  (p=0.000 n=10+10)
IndexEasy/64M            14.7ms ±16%    13.2ms ±11%    -10.22%  (p=0.001 n=10+10)

strings package:

name                   old time/op  new time/op  delta
IndexRune              88.1ns ±16%  28.9ns ± 4%  -67.20%  (p=0.000 n=10+10)
IndexRuneLongString     456ns ± 7%    34ns ± 3%  -92.50%  (p=0.000 n=10+10)
IndexRuneFastPath      12.9ns ±14%  11.1ns ± 6%  -13.84%  (p=0.000 n=10+10)
Index                  13.0ns ± 7%  11.3ns ± 4%  -13.31%  (p=0.000 n=10+10)
IndexHard1             3.38ms ± 9%  0.07ms ± 1%  -97.79%  (p=0.000 n=10+10)
IndexHard2             3.58ms ± 7%  0.37ms ± 2%  -89.78%  (p=0.000 n=10+10)
IndexHard3             3.47ms ± 7%  0.75ms ± 1%  -78.52%  (p=0.000 n=10+10)
IndexHard4             3.56ms ± 6%  1.34ms ± 0%  -62.39%    (p=0.000 n=9+9)

Change-Id: If36c2afb8c02e80fcaa1cf5ec2abb0a2be08c7d1
Reviewed-on: https://go-review.googlesource.com/32447
Run-TryBot: Michael Munday <munday@ca.ibm.com>
Reviewed-by: 's avatarBrad Fitzpatrick <bradfitz@golang.org>
parent 11ff4b21
......@@ -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,!s390x
package bytes
......
// Copyright 2016 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.
package bytes
//go:noescape
// indexShortStr returns the index of the first instance of sep in s,
// or -1 if sep is not present in s.
// indexShortStr requires 2 <= len(sep) <= shortStringLen
func indexShortStr(s, c []byte) int // ../runtime/asm_s390x.s
// supportsVX reports whether the vector facility is available.
// indexShortStr must not be called if the vector facility is not
// available.
func supportsVX() bool // ../runtime/asm_s390x.s
var shortStringLen = -1
func init() {
if supportsVX() {
shortStringLen = 64
}
}
// Index returns the index of the first instance of sep in s, or -1 if sep is not present in s.
func Index(s, sep []byte) int {
n := len(sep)
switch {
case n == 0:
return 0
case n == 1:
return IndexByte(s, sep[0])
case n <= shortStringLen:
// Use brute force when s and sep both are small
if len(s) <= 64 {
return indexShortStr(s, sep)
}
c := sep[0]
i := 0
t := s[:len(s)-n+1]
fails := 0
for i < len(t) {
if t[i] != c {
// IndexByte skips 16/32 bytes per iteration,
// so it's faster than indexShortStr.
o := IndexByte(t[i:], c)
if o < 0 {
return -1
}
i += o
}
if Equal(s[i:i+n], sep) {
return i
}
fails++
i++
// Switch to indexShortStr when IndexByte produces too many false positives.
// Too many means more that 1 error per 8 characters.
// Allow some errors in the beginning.
if fails > (i+16)/8 {
r := indexShortStr(s[i:], sep)
if r >= 0 {
return r + i
}
return -1
}
}
return -1
case n == len(s):
if Equal(sep, s) {
return 0
}
return -1
case n > len(s):
return -1
}
// Rabin-Karp search
hashsep, pow := hashStr(sep)
var h uint32
for i := 0; i < n; i++ {
h = h*primeRK + uint32(s[i])
}
if h == hashsep && Equal(s[:n], sep) {
return 0
}
for i := n; i < len(s); {
h *= primeRK
h += uint32(s[i])
h -= pow * uint32(s[i-n])
i++
if h == hashsep && Equal(s[i-n:i], sep) {
return i - n
}
}
return -1
}
// primeRK is the prime base used in Rabin-Karp algorithm.
const primeRK = 16777619
// hashStr returns the hash and the appropriate multiplicative
// factor for use in Rabin-Karp algorithm.
func hashStr(sep []byte) (uint32, uint32) {
hash := uint32(0)
for i := 0; i < len(sep); i++ {
hash = hash*primeRK + uint32(sep[i])
}
var pow, sq uint32 = 1, primeRK
for i := len(sep); i > 0; i >>= 1 {
if i&1 != 0 {
pow *= sq
}
sq *= sq
}
return hash, pow
}
......@@ -1076,6 +1076,230 @@ TEXT runtime·cmpbodyclc(SB),NOSPLIT|NOFRAME,$0-0
CLC $1, 0(R3), 0(R5)
RET
// func supportsVX() bool
TEXT strings·supportsVX(SB),NOSPLIT,$0-1
MOVBZ runtime·cpu+facilities_hasVX(SB), R0
MOVB R0, ret+0(FP)
RET
// func supportsVX() bool
TEXT bytes·supportsVX(SB),NOSPLIT,$0-1
MOVBZ runtime·cpu+facilities_hasVX(SB), R0
MOVB R0, ret+0(FP)
RET
// func indexShortStr(s, sep string) int
// Caller must confirm availability of vx facility before calling.
TEXT strings·indexShortStr(SB),NOSPLIT|NOFRAME,$0-40
LMG s+0(FP), R1, R2 // R1=&s[0], R2=len(s)
LMG sep+16(FP), R3, R4 // R3=&sep[0], R4=len(sep)
MOVD $ret+32(FP), R5
BR runtime·indexShortStr(SB)
// func indexShortStr(s, sep []byte) int
// Caller must confirm availability of vx facility before calling.
TEXT bytes·indexShortStr(SB),NOSPLIT|NOFRAME,$0-56
LMG s+0(FP), R1, R2 // R1=&s[0], R2=len(s)
LMG sep+24(FP), R3, R4 // R3=&sep[0], R4=len(sep)
MOVD $ret+48(FP), R5
BR runtime·indexShortStr(SB)
// s: string we are searching
// sep: string to search for
// R1=&s[0], R2=len(s)
// R3=&sep[0], R4=len(sep)
// R5=&ret (int)
// Caller must confirm availability of vx facility before calling.
TEXT runtime·indexShortStr(SB),NOSPLIT|NOFRAME,$0
CMPBGT R4, R2, notfound
ADD R1, R2
SUB R4, R2 // R2=&s[len(s)-len(sep)] (last valid index)
CMPBEQ R4, $0, notfound
SUB $1, R4 // R4=len(sep)-1 for use as VLL index
VLL R4, (R3), V0 // contains first 16 bytes of sep
MOVD R1, R7
index2plus:
CMPBNE R4, $1, index3plus
MOVD $15(R7), R9
CMPBGE R9, R2, index2to16
VGBM $0xaaaa, V31 // 0xff00ff00ff00ff00...
VONE V16
VREPH $0, V0, V1
CMPBGE R9, R2, index2to16
index2loop:
VL 0(R7), V2 // 16 bytes, even indices
VL 1(R7), V4 // 16 bytes, odd indices
VCEQH V1, V2, V5 // compare even indices
VCEQH V1, V4, V6 // compare odd indices
VSEL V5, V6, V31, V7 // merge even and odd indices
VFEEBS V16, V7, V17 // find leftmost index, set condition to 1 if found
BLT foundV17
MOVD $16(R7), R7 // R7+=16
ADD $15, R7, R9
CMPBLE R9, R2, index2loop // continue if (R7+15) <= R2 (last index to search)
CMPBLE R7, R2, index2to16
BR notfound
index3plus:
CMPBNE R4, $2, index4plus
ADD $15, R7, R9
CMPBGE R9, R2, index2to16
MOVD $1, R0
VGBM $0xaaaa, V31 // 0xff00ff00ff00ff00...
VONE V16
VREPH $0, V0, V1
VREPB $2, V0, V8
index3loop:
VL (R7), V2 // load 16-bytes into V2
VLL R0, 16(R7), V3 // load 2-bytes into V3
VSLDB $1, V2, V3, V4 // V4=(V2:V3)<<1
VSLDB $2, V2, V3, V9 // V9=(V2:V3)<<2
VCEQH V1, V2, V5 // compare 2-byte even indices
VCEQH V1, V4, V6 // compare 2-byte odd indices
VCEQB V8, V9, V10 // compare last bytes
VSEL V5, V6, V31, V7 // merge even and odd indices
VN V7, V10, V7 // AND indices with last byte
VFEEBS V16, V7, V17 // find leftmost index, set condition to 1 if found
BLT foundV17
MOVD $16(R7), R7 // R7+=16
ADD $15, R7, R9
CMPBLE R9, R2, index3loop // continue if (R7+15) <= R2 (last index to search)
CMPBLE R7, R2, index2to16
BR notfound
index4plus:
CMPBNE R4, $3, index5plus
ADD $15, R7, R9
CMPBGE R9, R2, index2to16
MOVD $2, R0
VGBM $0x8888, V29 // 0xff000000ff000000...
VGBM $0x2222, V30 // 0x0000ff000000ff00...
VGBM $0xcccc, V31 // 0xffff0000ffff0000...
VONE V16
VREPF $0, V0, V1
index4loop:
VL (R7), V2 // load 16-bytes into V2
VLL R0, 16(R7), V3 // load 3-bytes into V3
VSLDB $1, V2, V3, V4 // V4=(V2:V3)<<1
VSLDB $2, V2, V3, V9 // V9=(V2:V3)<<1
VSLDB $3, V2, V3, V10 // V10=(V2:V3)<<1
VCEQF V1, V2, V5 // compare index 0, 4, ...
VCEQF V1, V4, V6 // compare index 1, 5, ...
VCEQF V1, V9, V11 // compare index 2, 6, ...
VCEQF V1, V10, V12 // compare index 3, 7, ...
VSEL V5, V6, V29, V13 // merge index 0, 1, 4, 5, ...
VSEL V11, V12, V30, V14 // merge index 2, 3, 6, 7, ...
VSEL V13, V14, V31, V7 // final merge
VFEEBS V16, V7, V17 // find leftmost index, set condition to 1 if found
BLT foundV17
MOVD $16(R7), R7 // R7+=16
ADD $15, R7, R9
CMPBLE R9, R2, index4loop // continue if (R7+15) <= R2 (last index to search)
CMPBLE R7, R2, index2to16
BR notfound
index5plus:
CMPBGT R4, $15, index17plus
index2to16:
CMPBGT R7, R2, notfound
MOVD $1(R7), R8
CMPBGT R8, R2, index2to16tail
index2to16loop:
// unrolled 2x
VLL R4, (R7), V1
VLL R4, 1(R7), V2
VCEQGS V0, V1, V3
BEQ found
MOVD $1(R7), R7
VCEQGS V0, V2, V4
BEQ found
MOVD $1(R7), R7
CMPBLT R7, R2, index2to16loop
CMPBGT R7, R2, notfound
index2to16tail:
VLL R4, (R7), V1
VCEQGS V0, V1, V2
BEQ found
BR notfound
index17plus:
CMPBGT R4, $31, index33plus
SUB $16, R4, R0
VLL R0, 16(R3), V1
VONE V7
index17to32loop:
VL (R7), V2
VLL R0, 16(R7), V3
VCEQG V0, V2, V4
VCEQG V1, V3, V5
VN V4, V5, V6
VCEQGS V6, V7, V8
BEQ found
MOVD $1(R7), R7
CMPBLE R7, R2, index17to32loop
BR notfound
index33plus:
CMPBGT R4, $47, index49plus
SUB $32, R4, R0
VL 16(R3), V1
VLL R0, 32(R3), V2
VONE V11
index33to48loop:
VL (R7), V3
VL 16(R7), V4
VLL R0, 32(R7), V5
VCEQG V0, V3, V6
VCEQG V1, V4, V7
VCEQG V2, V5, V8
VN V6, V7, V9
VN V8, V9, V10
VCEQGS V10, V11, V12
BEQ found
MOVD $1(R7), R7
CMPBLE R7, R2, index33to48loop
BR notfound
index49plus:
CMPBGT R4, $63, index65plus
SUB $48, R4, R0
VL 16(R3), V1
VL 32(R3), V2
VLL R0, 48(R3), V3
VONE V15
index49to64loop:
VL (R7), V4
VL 16(R7), V5
VL 32(R7), V6
VLL R0, 48(R7), V7
VCEQG V0, V4, V8
VCEQG V1, V5, V9
VCEQG V2, V6, V10
VCEQG V3, V7, V11
VN V8, V9, V12
VN V10, V11, V13
VN V12, V13, V14
VCEQGS V14, V15, V16
BEQ found
MOVD $1(R7), R7
CMPBLE R7, R2, index49to64loop
notfound:
MOVD $-1, (R5)
RET
index65plus:
// not implemented
MOVD $0, (R0)
RET
foundV17: // index is in doubleword V17[0]
VLGVG $0, V17, R8
ADD R8, R7
found:
SUB R1, R7
MOVD R7, (R5)
RET
// This is called from .init_array and follows the platform, not Go, ABI.
// We are overly conservative. We could only save the registers we use.
// However, since this function is only called once per loaded module
......
......@@ -2,7 +2,7 @@
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// +build !amd64
// +build !amd64,!s390x
package strings
......
// Copyright 2016 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package strings
//go:noescape
// indexShortStr returns the index of the first instance of sep in s,
// or -1 if sep is not present in s.
// indexShortStr requires 2 <= len(sep) <= shortStringLen
func indexShortStr(s, sep string) int // ../runtime/asm_$GOARCH.s
// supportsVX reports whether the vector facility is available.
// indexShortStr must not be called if the vector facility is not
// available.
func supportsVX() bool // ../runtime/asm_s390x.s
var shortStringLen = -1
func init() {
if supportsVX() {
shortStringLen = 64
}
}
// Index returns the index of the first instance of sep in s, or -1 if sep is not present in s.
func Index(s, sep string) int {
n := len(sep)
switch {
case n == 0:
return 0
case n == 1:
return IndexByte(s, sep[0])
case n <= shortStringLen:
// Use brute force when s and sep both are small
if len(s) <= 64 {
return indexShortStr(s, sep)
}
c := sep[0]
i := 0
t := s[:len(s)-n+1]
fails := 0
for i < len(t) {
if t[i] != c {
// IndexByte skips 16/32 bytes per iteration,
// so it's faster than indexShortStr.
o := IndexByte(t[i:], c)
if o < 0 {
return -1
}
i += o
}
if s[i:i+n] == sep {
return i
}
fails++
i++
// Switch to indexShortStr when IndexByte produces too many false positives.
// Too many means more that 1 error per 8 characters.
// Allow some errors in the beginning.
if fails > (i+16)/8 {
r := indexShortStr(s[i:], sep)
if r >= 0 {
return r + i
}
return -1
}
}
return -1
case n == len(s):
if sep == s {
return 0
}
return -1
case n > len(s):
return -1
}
// Rabin-Karp search
hashsep, pow := hashStr(sep)
var h uint32
for i := 0; i < n; i++ {
h = h*primeRK + uint32(s[i])
}
if h == hashsep && s[:n] == sep {
return 0
}
for i := n; i < len(s); {
h *= primeRK
h += uint32(s[i])
h -= pow * uint32(s[i-n])
i++
if h == hashsep && s[i-n:i] == sep {
return i - n
}
}
return -1
}
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment