1
0
mirror of https://github.com/golang/go synced 2024-11-20 06:54:42 -07:00

cmd/compile: optimize ARM code with NMULF/NMULD

NMULF and NMULD are efficient FP instructions, and the go compiler can
use them to generate better code.

The benchmark tests of my patch did not show general change, but big
improvement in special cases.

1.A special test case improved 12.6%.
https://github.com/benshi001/ugo1/blob/master/fpmul_test.go
name                     old time/op    new time/op    delta
FPMul-4                     398µs ± 1%     348µs ± 1%  -12.64%  (p=0.000 n=40+40)

2. the compilecmp test showed little change.
name        old time/op       new time/op       delta
Template          2.30s ± 1%        2.31s ± 1%    ~     (p=0.754 n=17+19)
Unicode           1.31s ± 3%        1.32s ± 5%    ~     (p=0.265 n=20+20)
GoTypes           7.73s ± 2%        7.73s ± 1%    ~     (p=0.925 n=20+20)
Compiler          37.0s ± 1%        37.3s ± 2%  +0.79%  (p=0.002 n=19+20)
SSA               83.8s ± 4%        83.5s ± 2%    ~     (p=0.964 n=20+17)
Flate             1.43s ± 2%        1.44s ± 1%    ~     (p=0.602 n=20+20)
GoParser          1.82s ± 2%        1.81s ± 2%    ~     (p=0.141 n=19+20)
Reflect           5.08s ± 2%        5.08s ± 3%    ~     (p=0.835 n=20+19)
Tar               2.36s ± 1%        2.35s ± 1%    ~     (p=0.195 n=18+17)
XML               2.57s ± 2%        2.56s ± 1%    ~     (p=0.283 n=20+17)
[Geo mean]        4.74s             4.75s       +0.05%

name        old user-time/op  new user-time/op  delta
Template          2.75s ± 2%        2.75s ± 0%    ~     (p=0.620 n=20+15)
Unicode           1.59s ± 4%        1.60s ± 4%    ~     (p=0.479 n=20+19)
GoTypes           9.48s ± 1%        9.47s ± 1%    ~     (p=0.743 n=20+20)
Compiler          45.7s ± 1%        45.7s ± 1%    ~     (p=0.482 n=19+20)
SSA                109s ± 1%         109s ± 2%    ~     (p=0.800 n=18+20)
Flate             1.67s ± 3%        1.67s ± 3%    ~     (p=0.598 n=19+18)
GoParser          2.15s ± 4%        2.13s ± 3%    ~     (p=0.153 n=20+20)
Reflect           5.95s ± 2%        5.95s ± 2%    ~     (p=0.961 n=19+20)
Tar               2.93s ± 2%        2.92s ± 3%    ~     (p=0.242 n=20+19)
XML               3.02s ± 3%        3.04s ± 3%    ~     (p=0.233 n=19+18)
[Geo mean]        5.74s             5.74s       -0.04%

name        old text-bytes    new text-bytes    delta
HelloSize         588kB ± 0%        588kB ± 0%    ~     (all equal)

name        old data-bytes    new data-bytes    delta
HelloSize        5.46kB ± 0%       5.46kB ± 0%    ~     (all equal)

name        old bss-bytes     new bss-bytes     delta
HelloSize        72.9kB ± 0%       72.9kB ± 0%    ~     (all equal)

name        old exe-bytes     new exe-bytes     delta
HelloSize        1.03MB ± 0%       1.03MB ± 0%    ~     (all equal)

3. The go1 benchmark showed little change in total.
name                     old time/op    new time/op    delta
BinaryTree17-4              41.8s ± 1%     41.8s ± 1%    ~     (p=0.388 n=40+39)
Fannkuch11-4                24.1s ± 1%     24.1s ± 1%    ~     (p=0.077 n=40+40)
FmtFprintfEmpty-4           834ns ± 1%     831ns ± 1%  -0.31%  (p=0.002 n=40+37)
FmtFprintfString-4         1.34µs ± 1%    1.34µs ± 0%    ~     (p=0.387 n=40+40)
FmtFprintfInt-4            1.44µs ± 1%    1.44µs ± 1%    ~     (p=0.421 n=40+40)
FmtFprintfIntInt-4         2.09µs ± 0%    2.09µs ± 1%    ~     (p=0.589 n=40+39)
FmtFprintfPrefixedInt-4    2.32µs ± 1%    2.33µs ± 1%  +0.15%  (p=0.001 n=40+40)
FmtFprintfFloat-4          4.51µs ± 0%    4.44µs ± 1%  -1.50%  (p=0.000 n=40+40)
FmtManyArgs-4              7.94µs ± 0%    7.97µs ± 0%  +0.36%  (p=0.001 n=32+40)
GobDecode-4                 104ms ± 1%     102ms ± 2%  -1.27%  (p=0.000 n=39+37)
GobEncode-4                90.5ms ± 1%    90.9ms ± 2%  +0.40%  (p=0.006 n=37+40)
Gzip-4                      4.10s ± 2%     4.08s ± 1%  -0.30%  (p=0.004 n=40+40)
Gunzip-4                    603ms ± 0%     602ms ± 1%    ~     (p=0.303 n=37+40)
HTTPClientServer-4          672µs ± 3%     658µs ± 2%  -2.08%  (p=0.000 n=39+37)
JSONEncode-4                238ms ± 1%     239ms ± 0%  +0.26%  (p=0.001 n=40+25)
JSONDecode-4                884ms ± 1%     885ms ± 1%  +0.16%  (p=0.012 n=40+40)
Mandelbrot200-4            49.3ms ± 0%    49.3ms ± 0%    ~     (p=0.588 n=40+38)
GoParse-4                  46.3ms ± 1%    46.4ms ± 2%    ~     (p=0.487 n=40+40)
RegexpMatchEasy0_32-4      1.28µs ± 1%    1.28µs ± 0%  +0.12%  (p=0.003 n=40+40)
RegexpMatchEasy0_1K-4      7.78µs ± 5%    7.78µs ± 4%    ~     (p=0.825 n=40+40)
RegexpMatchEasy1_32-4      1.29µs ± 1%    1.29µs ± 0%    ~     (p=0.659 n=40+40)
RegexpMatchEasy1_1K-4      10.3µs ± 3%    10.4µs ± 2%    ~     (p=0.266 n=40+40)
RegexpMatchMedium_32-4     2.05µs ± 1%    2.05µs ± 0%  -0.18%  (p=0.002 n=40+28)
RegexpMatchMedium_1K-4      533µs ± 1%     534µs ± 1%    ~     (p=0.397 n=37+40)
RegexpMatchHard_32-4       28.9µs ± 1%    28.9µs ± 1%  -0.22%  (p=0.002 n=40+40)
RegexpMatchHard_1K-4        868µs ± 1%     870µs ± 1%  +0.21%  (p=0.015 n=40+40)
Revcomp-4                  67.3ms ± 1%    67.2ms ± 2%    ~     (p=0.262 n=38+39)
Template-4                  1.07s ± 1%     1.07s ± 1%    ~     (p=0.276 n=40+40)
TimeParse-4                7.16µs ± 1%    7.16µs ± 1%    ~     (p=0.610 n=39+40)
TimeFormat-4               13.3µs ± 1%    13.3µs ± 1%    ~     (p=0.617 n=38+40)
[Geo mean]                  720µs          719µs       -0.13%

name                     old speed      new speed      delta
GobDecode-4              7.39MB/s ± 1%  7.49MB/s ± 2%  +1.25%  (p=0.000 n=39+38)
GobEncode-4              8.48MB/s ± 1%  8.45MB/s ± 2%  -0.40%  (p=0.005 n=37+40)
Gzip-4                   4.74MB/s ± 2%  4.75MB/s ± 1%  +0.30%  (p=0.018 n=40+40)
Gunzip-4                 32.2MB/s ± 0%  32.2MB/s ± 1%    ~     (p=0.272 n=36+40)
JSONEncode-4             8.15MB/s ± 1%  8.13MB/s ± 0%  -0.26%  (p=0.003 n=40+25)
JSONDecode-4             2.19MB/s ± 1%  2.19MB/s ± 1%    ~     (p=0.676 n=40+40)
GoParse-4                1.25MB/s ± 2%  1.25MB/s ± 2%    ~     (p=0.823 n=40+40)
RegexpMatchEasy0_32-4    25.1MB/s ± 1%  25.1MB/s ± 0%  -0.12%  (p=0.006 n=40+40)
RegexpMatchEasy0_1K-4     132MB/s ± 5%   132MB/s ± 5%    ~     (p=0.821 n=40+40)
RegexpMatchEasy1_32-4    24.7MB/s ± 1%  24.7MB/s ± 0%    ~     (p=0.630 n=40+40)
RegexpMatchEasy1_1K-4    99.1MB/s ± 3%  98.8MB/s ± 2%    ~     (p=0.268 n=40+40)
RegexpMatchMedium_32-4    487kB/s ± 2%   490kB/s ± 0%  +0.51%  (p=0.001 n=40+40)
RegexpMatchMedium_1K-4   1.92MB/s ± 1%  1.92MB/s ± 1%    ~     (p=0.208 n=39+40)
RegexpMatchHard_32-4     1.11MB/s ± 1%  1.11MB/s ± 0%  +0.36%  (p=0.000 n=40+33)
RegexpMatchHard_1K-4     1.18MB/s ± 1%  1.18MB/s ± 1%    ~     (p=0.207 n=40+37)
Revcomp-4                37.8MB/s ± 1%  37.8MB/s ± 2%    ~     (p=0.276 n=38+39)
Template-4               1.82MB/s ± 1%  1.81MB/s ± 1%    ~     (p=0.122 n=38+40)
[Geo mean]               6.81MB/s       6.81MB/s       +0.06%

fixes #19843

Change-Id: Ief3a0c2b15f59d40c7b40f2784eeb71196685b59
Reviewed-on: https://go-review.googlesource.com/61150
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This commit is contained in:
Ben Shi 2017-09-02 08:14:08 +00:00 committed by Cherry Zhang
parent 538b4bab3d
commit 2899c3e8cb
5 changed files with 285 additions and 9 deletions

View File

@ -184,6 +184,8 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
ssa.OpARMSUBD,
ssa.OpARMMULF,
ssa.OpARMMULD,
ssa.OpARMNMULF,
ssa.OpARMNMULD,
ssa.OpARMDIVF,
ssa.OpARMDIVD:
r := v.Reg()

View File

@ -850,6 +850,7 @@
(SRAconst [c] (MOVWconst [d])) -> (MOVWconst [int64(int32(d)>>uint64(c))])
(MUL (MOVWconst [c]) (MOVWconst [d])) -> (MOVWconst [int64(int32(c*d))])
(MULA (MOVWconst [c]) (MOVWconst [d]) a) -> (ADDconst [int64(int32(c*d))] a)
(MULS (MOVWconst [c]) (MOVWconst [d]) a) -> (SUBconst [int64(int32(c*d))] a)
(Select0 (CALLudiv (MOVWconst [c]) (MOVWconst [d]))) -> (MOVWconst [int64(uint32(c)/uint32(d))])
(Select1 (CALLudiv (MOVWconst [c]) (MOVWconst [d]))) -> (MOVWconst [int64(uint32(c)%uint32(d))])
(ANDconst [c] (MOVWconst [d])) -> (MOVWconst [c&d])
@ -1240,6 +1241,13 @@
(SUB a (MUL x y)) && objabi.GOARM == 7 -> (MULS x y a)
(RSB (MUL x y) a) && objabi.GOARM == 7 -> (MULS x y a)
(NEGF (MULF x y)) && objabi.GOARM >= 6 -> (NMULF x y)
(NEGD (MULD x y)) && objabi.GOARM >= 6 -> (NMULD x y)
(MULF (NEGF x) y) && objabi.GOARM >= 6 -> (NMULF x y)
(MULD (NEGD x) y) && objabi.GOARM >= 6 -> (NMULD x y)
(NMULF (NEGF x) y) -> (MULF x y)
(NMULD (NEGD x) y) -> (MULD x y)
(AND x (MVN y)) -> (BIC x y)
// simplification with *shift ops

View File

@ -170,14 +170,16 @@ func init() {
{name: "MULA", argLength: 3, reg: gp31, asm: "MULA"}, // arg0 * arg1 + arg2
{name: "MULS", argLength: 3, reg: gp31, asm: "MULS"}, // arg2 - arg0 * arg1
{name: "ADDF", argLength: 2, reg: fp21, asm: "ADDF", commutative: true}, // arg0 + arg1
{name: "ADDD", argLength: 2, reg: fp21, asm: "ADDD", commutative: true}, // arg0 + arg1
{name: "SUBF", argLength: 2, reg: fp21, asm: "SUBF"}, // arg0 - arg1
{name: "SUBD", argLength: 2, reg: fp21, asm: "SUBD"}, // arg0 - arg1
{name: "MULF", argLength: 2, reg: fp21, asm: "MULF", commutative: true}, // arg0 * arg1
{name: "MULD", argLength: 2, reg: fp21, asm: "MULD", commutative: true}, // arg0 * arg1
{name: "DIVF", argLength: 2, reg: fp21, asm: "DIVF"}, // arg0 / arg1
{name: "DIVD", argLength: 2, reg: fp21, asm: "DIVD"}, // arg0 / arg1
{name: "ADDF", argLength: 2, reg: fp21, asm: "ADDF", commutative: true}, // arg0 + arg1
{name: "ADDD", argLength: 2, reg: fp21, asm: "ADDD", commutative: true}, // arg0 + arg1
{name: "SUBF", argLength: 2, reg: fp21, asm: "SUBF"}, // arg0 - arg1
{name: "SUBD", argLength: 2, reg: fp21, asm: "SUBD"}, // arg0 - arg1
{name: "MULF", argLength: 2, reg: fp21, asm: "MULF", commutative: true}, // arg0 * arg1
{name: "MULD", argLength: 2, reg: fp21, asm: "MULD", commutative: true}, // arg0 * arg1
{name: "NMULF", argLength: 2, reg: fp21, asm: "NMULF", commutative: true}, // -(arg0 * arg1)
{name: "NMULD", argLength: 2, reg: fp21, asm: "NMULD", commutative: true}, // -(arg0 * arg1)
{name: "DIVF", argLength: 2, reg: fp21, asm: "DIVF"}, // arg0 / arg1
{name: "DIVD", argLength: 2, reg: fp21, asm: "DIVD"}, // arg0 / arg1
{name: "AND", argLength: 2, reg: gp21, asm: "AND", commutative: true}, // arg0 & arg1
{name: "ANDconst", argLength: 1, reg: gp11, asm: "AND", aux: "Int32"}, // arg0 & auxInt

View File

@ -703,6 +703,8 @@ const (
OpARMSUBD
OpARMMULF
OpARMMULD
OpARMNMULF
OpARMNMULD
OpARMDIVF
OpARMDIVD
OpARMAND
@ -8592,6 +8594,36 @@ var opcodeTable = [...]opInfo{
},
},
},
{
name: "NMULF",
argLen: 2,
commutative: true,
asm: arm.ANMULF,
reg: regInfo{
inputs: []inputInfo{
{0, 4294901760}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15
{1, 4294901760}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15
},
outputs: []outputInfo{
{0, 4294901760}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15
},
},
},
{
name: "NMULD",
argLen: 2,
commutative: true,
asm: arm.ANMULD,
reg: regInfo{
inputs: []inputInfo{
{0, 4294901760}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15
{1, 4294901760}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15
},
outputs: []outputInfo{
{0, 4294901760}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15
},
},
},
{
name: "DIVF",
argLen: 2,

View File

@ -201,8 +201,12 @@ func rewriteValueARM(v *Value) bool {
return rewriteValueARM_OpARMMUL_0(v) || rewriteValueARM_OpARMMUL_10(v) || rewriteValueARM_OpARMMUL_20(v)
case OpARMMULA:
return rewriteValueARM_OpARMMULA_0(v) || rewriteValueARM_OpARMMULA_10(v) || rewriteValueARM_OpARMMULA_20(v)
case OpARMMULD:
return rewriteValueARM_OpARMMULD_0(v)
case OpARMMULF:
return rewriteValueARM_OpARMMULF_0(v)
case OpARMMULS:
return rewriteValueARM_OpARMMULS_0(v) || rewriteValueARM_OpARMMULS_10(v)
return rewriteValueARM_OpARMMULS_0(v) || rewriteValueARM_OpARMMULS_10(v) || rewriteValueARM_OpARMMULS_20(v)
case OpARMMVN:
return rewriteValueARM_OpARMMVN_0(v)
case OpARMMVNshiftLL:
@ -217,6 +221,14 @@ func rewriteValueARM(v *Value) bool {
return rewriteValueARM_OpARMMVNshiftRL_0(v)
case OpARMMVNshiftRLreg:
return rewriteValueARM_OpARMMVNshiftRLreg_0(v)
case OpARMNEGD:
return rewriteValueARM_OpARMNEGD_0(v)
case OpARMNEGF:
return rewriteValueARM_OpARMNEGF_0(v)
case OpARMNMULD:
return rewriteValueARM_OpARMNMULD_0(v)
case OpARMNMULF:
return rewriteValueARM_OpARMNMULF_0(v)
case OpARMNotEqual:
return rewriteValueARM_OpARMNotEqual_0(v)
case OpARMOR:
@ -9573,6 +9585,88 @@ func rewriteValueARM_OpARMMULA_20(v *Value) bool {
}
return false
}
func rewriteValueARM_OpARMMULD_0(v *Value) bool {
// match: (MULD (NEGD x) y)
// cond: objabi.GOARM >= 6
// result: (NMULD x y)
for {
_ = v.Args[1]
v_0 := v.Args[0]
if v_0.Op != OpARMNEGD {
break
}
x := v_0.Args[0]
y := v.Args[1]
if !(objabi.GOARM >= 6) {
break
}
v.reset(OpARMNMULD)
v.AddArg(x)
v.AddArg(y)
return true
}
// match: (MULD y (NEGD x))
// cond: objabi.GOARM >= 6
// result: (NMULD x y)
for {
_ = v.Args[1]
y := v.Args[0]
v_1 := v.Args[1]
if v_1.Op != OpARMNEGD {
break
}
x := v_1.Args[0]
if !(objabi.GOARM >= 6) {
break
}
v.reset(OpARMNMULD)
v.AddArg(x)
v.AddArg(y)
return true
}
return false
}
func rewriteValueARM_OpARMMULF_0(v *Value) bool {
// match: (MULF (NEGF x) y)
// cond: objabi.GOARM >= 6
// result: (NMULF x y)
for {
_ = v.Args[1]
v_0 := v.Args[0]
if v_0.Op != OpARMNEGF {
break
}
x := v_0.Args[0]
y := v.Args[1]
if !(objabi.GOARM >= 6) {
break
}
v.reset(OpARMNMULF)
v.AddArg(x)
v.AddArg(y)
return true
}
// match: (MULF y (NEGF x))
// cond: objabi.GOARM >= 6
// result: (NMULF x y)
for {
_ = v.Args[1]
y := v.Args[0]
v_1 := v.Args[1]
if v_1.Op != OpARMNEGF {
break
}
x := v_1.Args[0]
if !(objabi.GOARM >= 6) {
break
}
v.reset(OpARMNMULF)
v.AddArg(x)
v.AddArg(y)
return true
}
return false
}
func rewriteValueARM_OpARMMULS_0(v *Value) bool {
b := v.Block
_ = b
@ -10055,6 +10149,30 @@ func rewriteValueARM_OpARMMULS_10(v *Value) bool {
}
return false
}
func rewriteValueARM_OpARMMULS_20(v *Value) bool {
// match: (MULS (MOVWconst [c]) (MOVWconst [d]) a)
// cond:
// result: (SUBconst [int64(int32(c*d))] a)
for {
_ = v.Args[2]
v_0 := v.Args[0]
if v_0.Op != OpARMMOVWconst {
break
}
c := v_0.AuxInt
v_1 := v.Args[1]
if v_1.Op != OpARMMOVWconst {
break
}
d := v_1.AuxInt
a := v.Args[2]
v.reset(OpARMSUBconst)
v.AuxInt = int64(int32(c * d))
v.AddArg(a)
return true
}
return false
}
func rewriteValueARM_OpARMMVN_0(v *Value) bool {
// match: (MVN (MOVWconst [c]))
// cond:
@ -10272,6 +10390,120 @@ func rewriteValueARM_OpARMMVNshiftRLreg_0(v *Value) bool {
}
return false
}
func rewriteValueARM_OpARMNEGD_0(v *Value) bool {
// match: (NEGD (MULD x y))
// cond: objabi.GOARM >= 6
// result: (NMULD x y)
for {
v_0 := v.Args[0]
if v_0.Op != OpARMMULD {
break
}
_ = v_0.Args[1]
x := v_0.Args[0]
y := v_0.Args[1]
if !(objabi.GOARM >= 6) {
break
}
v.reset(OpARMNMULD)
v.AddArg(x)
v.AddArg(y)
return true
}
return false
}
func rewriteValueARM_OpARMNEGF_0(v *Value) bool {
// match: (NEGF (MULF x y))
// cond: objabi.GOARM >= 6
// result: (NMULF x y)
for {
v_0 := v.Args[0]
if v_0.Op != OpARMMULF {
break
}
_ = v_0.Args[1]
x := v_0.Args[0]
y := v_0.Args[1]
if !(objabi.GOARM >= 6) {
break
}
v.reset(OpARMNMULF)
v.AddArg(x)
v.AddArg(y)
return true
}
return false
}
func rewriteValueARM_OpARMNMULD_0(v *Value) bool {
// match: (NMULD (NEGD x) y)
// cond:
// result: (MULD x y)
for {
_ = v.Args[1]
v_0 := v.Args[0]
if v_0.Op != OpARMNEGD {
break
}
x := v_0.Args[0]
y := v.Args[1]
v.reset(OpARMMULD)
v.AddArg(x)
v.AddArg(y)
return true
}
// match: (NMULD y (NEGD x))
// cond:
// result: (MULD x y)
for {
_ = v.Args[1]
y := v.Args[0]
v_1 := v.Args[1]
if v_1.Op != OpARMNEGD {
break
}
x := v_1.Args[0]
v.reset(OpARMMULD)
v.AddArg(x)
v.AddArg(y)
return true
}
return false
}
func rewriteValueARM_OpARMNMULF_0(v *Value) bool {
// match: (NMULF (NEGF x) y)
// cond:
// result: (MULF x y)
for {
_ = v.Args[1]
v_0 := v.Args[0]
if v_0.Op != OpARMNEGF {
break
}
x := v_0.Args[0]
y := v.Args[1]
v.reset(OpARMMULF)
v.AddArg(x)
v.AddArg(y)
return true
}
// match: (NMULF y (NEGF x))
// cond:
// result: (MULF x y)
for {
_ = v.Args[1]
y := v.Args[0]
v_1 := v.Args[1]
if v_1.Op != OpARMNEGF {
break
}
x := v_1.Args[0]
v.reset(OpARMMULF)
v.AddArg(x)
v.AddArg(y)
return true
}
return false
}
func rewriteValueARM_OpARMNotEqual_0(v *Value) bool {
// match: (NotEqual (FlagEQ))
// cond: