cmd/compile: optimize ARM with MULS

MULS was introduced in ARMv7 and corresponding to MULA. This patch duplicated all MULA related SSA rules with MULS. Here was the contrast test result against the original go compiler. There was no improvement in total, but big improvement in special cases. 1. A specific test case accelerated 18.62%. (https://github.com/benshi001/ugo1/blob/master/mulsub_test.go) name old time/op new time/op delta MulSub-4 270µs ± 0% 219µs ± 0% -18.62% (p=0.000 n=35+40) 2. Total size of all .a files in pkg/ shrank by 0.002%. 3. The compilecmp benchmark showed no decline. name old time/op new time/op delta Template 2.37s ± 3% 2.36s ± 1% ~ (p=0.233 n=19+18) Unicode 1.32s ± 2% 1.34s ± 5% +1.32% (p=0.011 n=20+18) GoTypes 7.88s ± 1% 7.87s ± 1% ~ (p=0.758 n=20+20) Compiler 37.5s ± 1% 37.6s ± 1% ~ (p=0.194 n=20+19) SSA 83.7s ± 2% 83.5s ± 2% ~ (p=0.569 n=20+19) Flate 1.46s ± 3% 1.45s ± 1% ~ (p=0.619 n=20+17) GoParser 1.87s ± 2% 1.85s ± 1% -0.58% (p=0.048 n=20+18) Reflect 5.10s ± 2% 5.11s ± 2% ~ (p=0.365 n=19+20) Tar 1.78s ± 2% 1.78s ± 2% ~ (p=0.531 n=19+20) XML 2.62s ± 1% 2.61s ± 2% ~ (p=0.057 n=17+19) [Geo mean] 4.68s 4.67s -0.07% name old user-time/op new user-time/op delta Template 2.80s ± 1% 2.79s ± 2% ~ (p=0.686 n=17+20) Unicode 1.61s ± 4% 1.63s ± 6% ~ (p=0.222 n=20+20) GoTypes 9.59s ± 1% 9.60s ± 1% ~ (p=0.482 n=17+20) Compiler 46.1s ± 1% 46.2s ± 1% ~ (p=0.373 n=20+18) SSA 108s ± 1% 108s ± 2% ~ (p=0.784 n=20+20) Flate 1.68s ± 3% 1.69s ± 3% ~ (p=0.335 n=20+19) GoParser 2.20s ± 4% 2.19s ± 2% ~ (p=0.844 n=20+18) Reflect 5.97s ± 3% 6.01s ± 2% ~ (p=0.184 n=20+20) Tar 2.11s ± 2% 2.11s ± 4% ~ (p=0.961 n=19+20) XML 3.07s ± 1% 3.07s ± 3% ~ (p=0.786 n=16+19) [Geo mean] 5.61s 5.62s +0.19% name old text-bytes new text-bytes delta HelloSize 586kB ± 0% 586kB ± 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) 4. The go1 benchmark showed no decline in total. name old time/op new time/op delta BinaryTree17-4 41.7s ± 1% 41.7s ± 1% ~ (p=0.966 n=40+40) Fannkuch11-4 23.6s ± 0% 23.6s ± 1% -0.23% (p=0.000 n=40+40) FmtFprintfEmpty-4 844ns ± 1% 834ns ± 1% -1.23% (p=0.000 n=40+40) FmtFprintfString-4 1.39µs ± 1% 1.40µs ± 1% +0.71% (p=0.000 n=40+40) FmtFprintfInt-4 1.44µs ± 1% 1.45µs ± 1% +0.70% (p=0.000 n=40+40) FmtFprintfIntInt-4 2.10µs ± 1% 2.10µs ± 1% +0.30% (p=0.000 n=40+40) FmtFprintfPrefixedInt-4 2.49µs ± 0% 2.50µs ± 1% +0.66% (p=0.000 n=32+40) FmtFprintfFloat-4 4.42µs ± 1% 4.46µs ± 2% +0.94% (p=0.000 n=40+40) FmtManyArgs-4 8.31µs ± 1% 8.22µs ± 1% -1.09% (p=0.000 n=40+40) GobDecode-4 105ms ± 1% 102ms ± 1% -2.30% (p=0.000 n=39+39) GobEncode-4 90.2ms ± 1% 88.7ms ± 1% -1.66% (p=0.000 n=40+39) Gzip-4 4.17s ± 1% 4.16s ± 1% ~ (p=0.785 n=40+40) Gunzip-4 608ms ± 1% 608ms ± 1% ~ (p=0.481 n=40+40) HTTPClientServer-4 697µs ± 2% 684µs ± 3% -1.89% (p=0.000 n=37+40) JSONEncode-4 255ms ± 1% 256ms ± 1% +0.35% (p=0.000 n=40+40) JSONDecode-4 920ms ± 1% 926ms ± 1% +0.64% (p=0.000 n=40+39) Mandelbrot200-4 49.3ms ± 1% 49.3ms ± 0% +0.07% (p=0.005 n=40+40) GoParse-4 46.8ms ± 2% 46.7ms ± 1% ~ (p=1.000 n=40+40) RegexpMatchEasy0_32-4 1.27µs ± 0% 1.27µs ± 1% ~ (p=0.057 n=40+40) RegexpMatchEasy0_1K-4 7.97µs ± 7% 7.92µs ± 5% ~ (p=0.094 n=40+40) RegexpMatchEasy1_32-4 1.28µs ± 1% 1.28µs ± 1% ~ (p=0.406 n=40+40) RegexpMatchEasy1_1K-4 10.5µs ± 4% 10.5µs ± 3% ~ (p=0.855 n=40+40) RegexpMatchMedium_32-4 2.04µs ± 0% 2.04µs ± 1% -0.22% (p=0.000 n=39+40) RegexpMatchMedium_1K-4 541µs ± 0% 540µs ± 1% -0.25% (p=0.000 n=40+38) RegexpMatchHard_32-4 29.3µs ± 1% 29.3µs ± 0% ~ (p=0.149 n=40+40) RegexpMatchHard_1K-4 878µs ± 1% 880µs ± 0% +0.14% (p=0.005 n=36+35) Revcomp-4 81.8ms ± 2% 81.4ms ± 2% -0.43% (p=0.015 n=38+39) Template-4 1.05s ± 1% 1.05s ± 1% ~ (p=0.302 n=40+35) TimeParse-4 7.18µs ± 1% 7.26µs ± 1% +1.05% (p=0.000 n=40+36) TimeFormat-4 13.1µs ± 1% 13.1µs ± 1% ~ (p=0.698 n=37+40) [Geo mean] 733µs 732µs -0.16% name old speed new speed delta GobDecode-4 7.34MB/s ± 1% 7.51MB/s ± 1% +2.36% (p=0.000 n=39+39) GobEncode-4 8.51MB/s ± 1% 8.65MB/s ± 1% +1.69% (p=0.000 n=40+39) Gzip-4 4.66MB/s ± 1% 4.66MB/s ± 1% ~ (p=0.783 n=40+40) Gunzip-4 31.9MB/s ± 1% 31.9MB/s ± 1% ~ (p=0.466 n=40+40) JSONEncode-4 7.61MB/s ± 1% 7.58MB/s ± 1% -0.35% (p=0.001 n=40+40) JSONDecode-4 2.11MB/s ± 1% 2.10MB/s ± 1% -0.52% (p=0.000 n=38+39) GoParse-4 1.24MB/s ± 2% 1.24MB/s ± 1% ~ (p=0.556 n=40+39) RegexpMatchEasy0_32-4 25.1MB/s ± 0% 25.1MB/s ± 1% ~ (p=0.064 n=40+40) RegexpMatchEasy0_1K-4 129MB/s ± 8% 129MB/s ± 5% ~ (p=0.094 n=40+40) RegexpMatchEasy1_32-4 25.0MB/s ± 1% 25.1MB/s ± 1% ~ (p=0.331 n=40+40) RegexpMatchEasy1_1K-4 97.7MB/s ± 4% 97.8MB/s ± 3% ~ (p=0.851 n=40+40) RegexpMatchMedium_32-4 490kB/s ± 0% 490kB/s ± 0% ~ (all equal) RegexpMatchMedium_1K-4 1.89MB/s ± 0% 1.90MB/s ± 1% +0.12% (p=0.031 n=40+40) RegexpMatchHard_32-4 1.09MB/s ± 1% 1.09MB/s ± 1% ~ (p=0.597 n=40+40) RegexpMatchHard_1K-4 1.16MB/s ± 1% 1.16MB/s ± 1% ~ (p=0.565 n=40+35) Revcomp-4 31.1MB/s ± 2% 31.2MB/s ± 2% +0.44% (p=0.018 n=38+39) Template-4 1.85MB/s ± 1% 1.85MB/s ± 1% ~ (p=0.873 n=40+40) [Geo mean] 6.66MB/s 6.67MB/s +0.26% Change-Id: Icc972d8a78ea06c32c3aa15733ff0537c82c2dc7 Reviewed-on: https://go-review.googlesource.com/58950Reviewed-by: Cherry Zhang <cherryyz@google.com> Run-TryBot: Cherry Zhang <cherryyz@google.com>

cmd/compile: optimize ARM with MULS
MULS was introduced in ARMv7 and corresponding to MULA. This patch duplicated all MULA related SSA rules with MULS. Here was the contrast test result against the original go compiler. There was no improvement in total, but big improvement in special cases. 1. A specific test case accelerated 18.62%. (https://github.com/benshi001/ugo1/blob/master/mulsub_test.go) name old time/op new time/op delta MulSub-4 270µs ± 0% 219µs ± 0% -18.62% (p=0.000 n=35+40) 2. Total size of all .a files in pkg/ shrank by 0.002%. 3. The compilecmp benchmark showed no decline. name old time/op new time/op delta Template 2.37s ± 3% 2.36s ± 1% ~ (p=0.233 n=19+18) Unicode 1.32s ± 2% 1.34s ± 5% +1.32% (p=0.011 n=20+18) GoTypes 7.88s ± 1% 7.87s ± 1% ~ (p=0.758 n=20+20) Compiler 37.5s ± 1% 37.6s ± 1% ~ (p=0.194 n=20+19) SSA 83.7s ± 2% 83.5s ± 2% ~ (p=0.569 n=20+19) Flate 1.46s ± 3% 1.45s ± 1% ~ (p=0.619 n=20+17) GoParser 1.87s ± 2% 1.85s ± 1% -0.58% (p=0.048 n=20+18) Reflect 5.10s ± 2% 5.11s ± 2% ~ (p=0.365 n=19+20) Tar 1.78s ± 2% 1.78s ± 2% ~ (p=0.531 n=19+20) XML 2.62s ± 1% 2.61s ± 2% ~ (p=0.057 n=17+19) [Geo mean] 4.68s 4.67s -0.07% name old user-time/op new user-time/op delta Template 2.80s ± 1% 2.79s ± 2% ~ (p=0.686 n=17+20) Unicode 1.61s ± 4% 1.63s ± 6% ~ (p=0.222 n=20+20) GoTypes 9.59s ± 1% 9.60s ± 1% ~ (p=0.482 n=17+20) Compiler 46.1s ± 1% 46.2s ± 1% ~ (p=0.373 n=20+18) SSA 108s ± 1% 108s ± 2% ~ (p=0.784 n=20+20) Flate 1.68s ± 3% 1.69s ± 3% ~ (p=0.335 n=20+19) GoParser 2.20s ± 4% 2.19s ± 2% ~ (p=0.844 n=20+18) Reflect 5.97s ± 3% 6.01s ± 2% ~ (p=0.184 n=20+20) Tar 2.11s ± 2% 2.11s ± 4% ~ (p=0.961 n=19+20) XML 3.07s ± 1% 3.07s ± 3% ~ (p=0.786 n=16+19) [Geo mean] 5.61s 5.62s +0.19% name old text-bytes new text-bytes delta HelloSize 586kB ± 0% 586kB ± 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) 4. The go1 benchmark showed no decline in total. name old time/op new time/op delta BinaryTree17-4 41.7s ± 1% 41.7s ± 1% ~ (p=0.966 n=40+40) Fannkuch11-4 23.6s ± 0% 23.6s ± 1% -0.23% (p=0.000 n=40+40) FmtFprintfEmpty-4 844ns ± 1% 834ns ± 1% -1.23% (p=0.000 n=40+40) FmtFprintfString-4 1.39µs ± 1% 1.40µs ± 1% +0.71% (p=0.000 n=40+40) FmtFprintfInt-4 1.44µs ± 1% 1.45µs ± 1% +0.70% (p=0.000 n=40+40) FmtFprintfIntInt-4 2.10µs ± 1% 2.10µs ± 1% +0.30% (p=0.000 n=40+40) FmtFprintfPrefixedInt-4 2.49µs ± 0% 2.50µs ± 1% +0.66% (p=0.000 n=32+40) FmtFprintfFloat-4 4.42µs ± 1% 4.46µs ± 2% +0.94% (p=0.000 n=40+40) FmtManyArgs-4 8.31µs ± 1% 8.22µs ± 1% -1.09% (p=0.000 n=40+40) GobDecode-4 105ms ± 1% 102ms ± 1% -2.30% (p=0.000 n=39+39) GobEncode-4 90.2ms ± 1% 88.7ms ± 1% -1.66% (p=0.000 n=40+39) Gzip-4 4.17s ± 1% 4.16s ± 1% ~ (p=0.785 n=40+40) Gunzip-4 608ms ± 1% 608ms ± 1% ~ (p=0.481 n=40+40) HTTPClientServer-4 697µs ± 2% 684µs ± 3% -1.89% (p=0.000 n=37+40) JSONEncode-4 255ms ± 1% 256ms ± 1% +0.35% (p=0.000 n=40+40) JSONDecode-4 920ms ± 1% 926ms ± 1% +0.64% (p=0.000 n=40+39) Mandelbrot200-4 49.3ms ± 1% 49.3ms ± 0% +0.07% (p=0.005 n=40+40) GoParse-4 46.8ms ± 2% 46.7ms ± 1% ~ (p=1.000 n=40+40) RegexpMatchEasy0_32-4 1.27µs ± 0% 1.27µs ± 1% ~ (p=0.057 n=40+40) RegexpMatchEasy0_1K-4 7.97µs ± 7% 7.92µs ± 5% ~ (p=0.094 n=40+40) RegexpMatchEasy1_32-4 1.28µs ± 1% 1.28µs ± 1% ~ (p=0.406 n=40+40) RegexpMatchEasy1_1K-4 10.5µs ± 4% 10.5µs ± 3% ~ (p=0.855 n=40+40) RegexpMatchMedium_32-4 2.04µs ± 0% 2.04µs ± 1% -0.22% (p=0.000 n=39+40) RegexpMatchMedium_1K-4 541µs ± 0% 540µs ± 1% -0.25% (p=0.000 n=40+38) RegexpMatchHard_32-4 29.3µs ± 1% 29.3µs ± 0% ~ (p=0.149 n=40+40) RegexpMatchHard_1K-4 878µs ± 1% 880µs ± 0% +0.14% (p=0.005 n=36+35) Revcomp-4 81.8ms ± 2% 81.4ms ± 2% -0.43% (p=0.015 n=38+39) Template-4 1.05s ± 1% 1.05s ± 1% ~ (p=0.302 n=40+35) TimeParse-4 7.18µs ± 1% 7.26µs ± 1% +1.05% (p=0.000 n=40+36) TimeFormat-4 13.1µs ± 1% 13.1µs ± 1% ~ (p=0.698 n=37+40) [Geo mean] 733µs 732µs -0.16% name old speed new speed delta GobDecode-4 7.34MB/s ± 1% 7.51MB/s ± 1% +2.36% (p=0.000 n=39+39) GobEncode-4 8.51MB/s ± 1% 8.65MB/s ± 1% +1.69% (p=0.000 n=40+39) Gzip-4 4.66MB/s ± 1% 4.66MB/s ± 1% ~ (p=0.783 n=40+40) Gunzip-4 31.9MB/s ± 1% 31.9MB/s ± 1% ~ (p=0.466 n=40+40) JSONEncode-4 7.61MB/s ± 1% 7.58MB/s ± 1% -0.35% (p=0.001 n=40+40) JSONDecode-4 2.11MB/s ± 1% 2.10MB/s ± 1% -0.52% (p=0.000 n=38+39) GoParse-4 1.24MB/s ± 2% 1.24MB/s ± 1% ~ (p=0.556 n=40+39) RegexpMatchEasy0_32-4 25.1MB/s ± 0% 25.1MB/s ± 1% ~ (p=0.064 n=40+40) RegexpMatchEasy0_1K-4 129MB/s ± 8% 129MB/s ± 5% ~ (p=0.094 n=40+40) RegexpMatchEasy1_32-4 25.0MB/s ± 1% 25.1MB/s ± 1% ~ (p=0.331 n=40+40) RegexpMatchEasy1_1K-4 97.7MB/s ± 4% 97.8MB/s ± 3% ~ (p=0.851 n=40+40) RegexpMatchMedium_32-4 490kB/s ± 0% 490kB/s ± 0% ~ (all equal) RegexpMatchMedium_1K-4 1.89MB/s ± 0% 1.90MB/s ± 1% +0.12% (p=0.031 n=40+40) RegexpMatchHard_32-4 1.09MB/s ± 1% 1.09MB/s ± 1% ~ (p=0.597 n=40+40) RegexpMatchHard_1K-4 1.16MB/s ± 1% 1.16MB/s ± 1% ~ (p=0.565 n=40+35) Revcomp-4 31.1MB/s ± 2% 31.2MB/s ± 2% +0.44% (p=0.018 n=38+39) Template-4 1.85MB/s ± 1% 1.85MB/s ± 1% ~ (p=0.873 n=40+40) [Geo mean] 6.66MB/s 6.67MB/s +0.26% Change-Id: Icc972d8a78ea06c32c3aa15733ff0537c82c2dc7 Reviewed-on: https://go-review.googlesource.com/58950Reviewed-by: Cherry Zhang <cherryyz@google.com> Run-TryBot: Cherry Zhang <cherryyz@google.com>
64607dbd · Ben Shi · Cherry Zhang · dcef97e0 · 64607dbd · 64607dbd
Commit 64607dbd authored Aug 25, 2017 by Ben Shi Committed by Cherry Zhang Aug 30, 2017
5 changed files
--- a/src/cmd/compile/internal/arm/ssa.go
+++ b/src/cmd/compile/internal/arm/ssa.go
@@ -402,7 +402,7 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		p.To.Type = obj.TYPE_REGREG
 		p.To.Reg = v.Reg0()           // high 32-bit
 		p.To.Offset = int64(v.Reg1()) // low 32-bit
-	case ssa.OpARMMULA:
+	case ssa.OpARMMULA, ssa.OpARMMULS:
 		p := s.Prog(v.Op.Asm())
 		p.From.Type = obj.TYPE_REG
 		p.From.Reg = v.Args[0].Reg()

--- a/src/cmd/compile/internal/ssa/gen/ARM.rules
+++ b/src/cmd/compile/internal/ssa/gen/ARM.rules
@@ -602,6 +602,28 @@
 (MULA (MOVWconst [c]) x a) && c%7 == 0 && isPowerOfTwo(c/7) && is32Bit(c) -> (ADD (SLLconst <x.Type> [log2(c/7)] (RSBshiftLL <x.Type> x x [3])) a)
 (MULA (MOVWconst [c]) x a) && c%9 == 0 && isPowerOfTwo(c/9) && is32Bit(c) -> (ADD (SLLconst <x.Type> [log2(c/9)] (ADDshiftLL <x.Type> x x [3])) a)

+(MULS x (MOVWconst [c]) a) && int32(c) == -1 -> (ADD a x)
+(MULS _ (MOVWconst [0]) a) -> a
+(MULS x (MOVWconst [1]) a) -> (RSB x a)
+(MULS x (MOVWconst [c]) a) && isPowerOfTwo(c) -> (RSB (SLLconst <x.Type> [log2(c)] x) a)
+(MULS x (MOVWconst [c]) a) && isPowerOfTwo(c-1) && int32(c) >= 3 -> (RSB (ADDshiftLL <x.Type> x x [log2(c-1)]) a)
+(MULS x (MOVWconst [c]) a) && isPowerOfTwo(c+1) && int32(c) >= 7 -> (RSB (RSBshiftLL <x.Type> x x [log2(c+1)]) a)
+(MULS x (MOVWconst [c]) a) && c%3 == 0 && isPowerOfTwo(c/3) && is32Bit(c) -> (RSB (SLLconst <x.Type> [log2(c/3)] (ADDshiftLL <x.Type> x x [1])) a)
+(MULS x (MOVWconst [c]) a) && c%5 == 0 && isPowerOfTwo(c/5) && is32Bit(c) -> (RSB (SLLconst <x.Type> [log2(c/5)] (ADDshiftLL <x.Type> x x [2])) a)
+(MULS x (MOVWconst [c]) a) && c%7 == 0 && isPowerOfTwo(c/7) && is32Bit(c) -> (RSB (SLLconst <x.Type> [log2(c/7)] (RSBshiftLL <x.Type> x x [3])) a)
+(MULS x (MOVWconst [c]) a) && c%9 == 0 && isPowerOfTwo(c/9) && is32Bit(c) -> (RSB (SLLconst <x.Type> [log2(c/9)] (ADDshiftLL <x.Type> x x [3])) a)
+
+(MULS (MOVWconst [c]) x a) && int32(c) == -1 -> (ADD a x)
+(MULS (MOVWconst [0]) _ a) -> a
+(MULS (MOVWconst [1]) x a) -> (RSB x a)
+(MULS (MOVWconst [c]) x a) && isPowerOfTwo(c) -> (RSB (SLLconst <x.Type> [log2(c)] x) a)
+(MULS (MOVWconst [c]) x a) && isPowerOfTwo(c-1) && int32(c) >= 3 -> (RSB (ADDshiftLL <x.Type> x x [log2(c-1)]) a)
+(MULS (MOVWconst [c]) x a) && isPowerOfTwo(c+1) && int32(c) >= 7 -> (RSB (RSBshiftLL <x.Type> x x [log2(c+1)]) a)
+(MULS (MOVWconst [c]) x a) && c%3 == 0 && isPowerOfTwo(c/3) && is32Bit(c) -> (RSB (SLLconst <x.Type> [log2(c/3)] (ADDshiftLL <x.Type> x x [1])) a)
+(MULS (MOVWconst [c]) x a) && c%5 == 0 && isPowerOfTwo(c/5) && is32Bit(c) -> (RSB (SLLconst <x.Type> [log2(c/5)] (ADDshiftLL <x.Type> x x [2])) a)
+(MULS (MOVWconst [c]) x a) && c%7 == 0 && isPowerOfTwo(c/7) && is32Bit(c) -> (RSB (SLLconst <x.Type> [log2(c/7)] (RSBshiftLL <x.Type> x x [3])) a)
+(MULS (MOVWconst [c]) x a) && c%9 == 0 && isPowerOfTwo(c/9) && is32Bit(c) -> (RSB (SLLconst <x.Type> [log2(c/9)] (ADDshiftLL <x.Type> x x [3])) a)
+
 // div by constant
 (Select0 (CALLudiv x (MOVWconst [1]))) -> x
 (Select1 (CALLudiv _ (MOVWconst [1]))) -> (MOVWconst [0])
@@ -1215,6 +1237,8 @@
 (BIC x x) -> (MOVWconst [0])

 (ADD (MUL x y) a) -> (MULA x y a)
+(SUB a (MUL x y)) && objabi.GOARM == 7 -> (MULS x y a)
+(RSB (MUL x y) a) && objabi.GOARM == 7 -> (MULS x y a)

 (AND x (MVN y)) -> (BIC x y)


--- a/src/cmd/compile/internal/ssa/gen/ARMOps.go
+++ b/src/cmd/compile/internal/ssa/gen/ARMOps.go
@@ -168,6 +168,7 @@ func init() {

 		{name: "MULLU", argLength: 2, reg: gp22, asm: "MULLU", commutative: true}, // arg0 * arg1, high 32 bits in out0, low 32 bits in out1
 		{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

--- a/src/cmd/compile/internal/ssa/opGen.go
+++ b/src/cmd/compile/internal/ssa/opGen.go
@@ -694,6 +694,7 @@ const (
 	OpARMRSCconst
 	OpARMMULLU
 	OpARMMULA
+	OpARMMULS
 	OpARMADDF
 	OpARMADDD
 	OpARMSUBF
@@ -8456,6 +8457,21 @@ var opcodeTable = [...]opInfo{
 			},
 		},
 	},
+	{
+		name:   "MULS",
+		argLen: 3,
+		asm:    arm.AMULS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 21503}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 R12 R14
+				{1, 21503}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 R12 R14
+				{2, 21503}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 R12 R14
+			},
+			outputs: []outputInfo{
+				{0, 21503}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 R12 R14
+			},
+		},
+	},
 	{
 		name:        "ADDF",
 		argLen:      2,

--- a/src/cmd/compile/internal/ssa/rewriteARM.go
+++ b/src/cmd/compile/internal/ssa/rewriteARM.go