cmd/compile: add rules to optimize go codes to constant 0 on loong64

goos: linux
goarch: loong64
pkg: test/bench/go1
cpu: Loongson-3A6000 @ 2500.00MHz
                      │  old.bench  │             new.bench              │
                      │   sec/op    │   sec/op     vs base               │
BinaryTree17             7.735 ± 1%    7.716 ± 1%  -0.23% (p=0.041 n=15)
Fannkuch11               2.645 ± 0%    2.646 ± 0%  +0.05% (p=0.013 n=15)
FmtFprintfEmpty         35.87n ± 0%   35.89n ± 0%  +0.06% (p=0.000 n=15)
FmtFprintfString        59.54n ± 0%   59.47n ± 0%       ~ (p=0.213 n=15)
FmtFprintfInt           62.23n ± 0%   62.06n ± 0%       ~ (p=0.212 n=15)
FmtFprintfIntInt        98.16n ± 0%   97.90n ± 0%  -0.26% (p=0.000 n=15)
FmtFprintfPrefixedInt   117.0n ± 0%   116.7n ± 0%  -0.26% (p=0.000 n=15)
FmtFprintfFloat         204.6n ± 0%   204.2n ± 0%  -0.20% (p=0.000 n=15)
FmtManyArgs             456.3n ± 0%   455.4n ± 0%  -0.20% (p=0.000 n=15)
GobDecode               7.210m ± 0%   7.156m ± 1%  -0.75% (p=0.000 n=15)
GobEncode               8.143m ± 1%   8.177m ± 1%       ~ (p=0.806 n=15)
Gzip                    280.2m ± 0%   279.7m ± 0%  -0.19% (p=0.005 n=15)
Gunzip                  32.71m ± 0%   32.65m ± 0%  -0.19% (p=0.000 n=15)
HTTPClientServer        53.76µ ± 0%   53.65µ ± 0%       ~ (p=0.083 n=15)
JSONEncode              9.297m ± 0%   9.295m ± 0%       ~ (p=0.806 n=15)
JSONDecode              46.97m ± 1%   47.07m ± 1%       ~ (p=0.683 n=15)
Mandelbrot200           4.602m ± 0%   4.600m ± 0%  -0.05% (p=0.001 n=15)
GoParse                 4.682m ± 0%   4.670m ± 1%  -0.25% (p=0.001 n=15)
RegexpMatchEasy0_32     59.80n ± 0%   59.63n ± 0%  -0.28% (p=0.000 n=15)
RegexpMatchEasy0_1K     458.3n ± 0%   457.3n ± 0%  -0.22% (p=0.001 n=15)
RegexpMatchEasy1_32     59.39n ± 0%   59.23n ± 0%  -0.27% (p=0.000 n=15)
RegexpMatchEasy1_1K     557.9n ± 0%   556.6n ± 0%  -0.23% (p=0.001 n=15)
RegexpMatchMedium_32    803.6n ± 0%   801.8n ± 0%  -0.22% (p=0.001 n=15)
RegexpMatchMedium_1K    27.32µ ± 0%   27.26µ ± 0%  -0.21% (p=0.000 n=15)
RegexpMatchHard_32      1.385µ ± 0%   1.382µ ± 0%  -0.22% (p=0.000 n=15)
RegexpMatchHard_1K      40.93µ ± 0%   40.83µ ± 0%  -0.24% (p=0.000 n=15)
Revcomp                 474.8m ± 0%   474.3m ± 0%       ~ (p=0.250 n=15)
Template                77.41m ± 1%   76.63m ± 1%  -1.01% (p=0.023 n=15)
TimeParse               271.1n ± 0%   271.2n ± 0%  +0.04% (p=0.022 n=15)
TimeFormat              290.0n ± 0%   289.8n ± 0%       ~ (p=0.118 n=15)
geomean                 51.73µ        51.64µ       -0.18%

Change-Id: I45a1e6c85bb3cea0f62766ec932432803e9af10a
Reviewed-on: https://go-review.googlesource.com/c/go/+/619315
Reviewed-by: Qiqi Huang <[email protected]>
Reviewed-by: Meidan Li <[email protected]>
Reviewed-by: abner chenc <[email protected]>
Reviewed-by: Michael Pratt <[email protected]>
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Carlos Amedee <[email protected]>

Author: sophie-zhao

src/cmd/compile/internal/ssa/_gen/LOONG64.rules

@@ -646,6 +646,16 @@
 (ROTR x (MOVVconst [c]))  => (ROTRconst x [c&31])
 (ROTRV x (MOVVconst [c])) => (ROTRVconst x [c&63])
 
+// If the shift amount is larger than the datasize(32, 16, 8), we can optimize to constant 0.
+(MOVWUreg (SLLVconst [lc] x)) && lc >= 32 => (MOVVconst [0])
+(MOVHUreg (SLLVconst [lc] x)) && lc >= 16 => (MOVVconst [0])
+(MOVBUreg (SLLVconst [lc] x)) && lc >= 8 => (MOVVconst [0])
+
+// After zero extension, the upper (64-datasize(32|16|8)) bits are zero, we can optimize to constant 0.
+(SRLVconst [rc] (MOVWUreg x)) && rc >= 32 => (MOVVconst [0])
+(SRLVconst [rc] (MOVHUreg x)) && rc >= 16 => (MOVVconst [0])
+(SRLVconst [rc] (MOVBUreg x)) && rc >= 8 => (MOVVconst [0])
+
 // mul by constant
 (MULV x (MOVVconst [-1])) => (NEGV x)
 (MULV _ (MOVVconst [0])) => (MOVVconst [0])

src/cmd/compile/internal/ssa/rewriteLOONG64.go

@@ -358,11 +358,15 @@ func rev16w(c uint32) (uint32, uint32, uint32) {
 
 func shift(x uint32, y uint16, z uint8) uint64 {
 	// arm64:-`MOVWU`,-`LSR\t[$]32`
+	// loong64:-`MOVWU`,-`SRLV\t[$]32`
 	a := uint64(x) >> 32
 	// arm64:-`MOVHU
+	// loong64:-`MOVHU`,-`SRLV\t[$]16`
 	b := uint64(y) >> 16
 	// arm64:-`MOVBU`
+	// loong64:-`MOVBU`,-`SRLV\t[$]8`
 	c := uint64(z) >> 8
 	// arm64:`MOVD\tZR`,-`ADD\tR[0-9]+>>16`,-`ADD\tR[0-9]+>>8`,
+	// loong64:`MOVV\t[$]0`,-`ADDVU`
 	return a + b + c
 }