cmd/compile: optimize RotateLeft8/16 on arm64

This CL optimizes RotateLeft8/16 on arm64. For 16 bits, we form a 32 bits register by duplicating two 16 bits registers, then use RORW instruction to do the rotate shift. For 8 bits, we just use LSR and LSL instead of RORW because the code is simpler. Benchmark Old ThisCL delta RotateLeft8-46 2.16 ns/op 1.73 ns/op -19.70% RotateLeft16-46 2.16 ns/op 1.54 ns/op -28.53% Change-Id: I09cde4383d12e31876a57f8cdfd3bb4f324fadb0 Reviewed-on: https://go-review.googlesource.com/c/go/+/420976 Reviewed-by: Keith Randall <[email protected]> Auto-Submit: Keith Randall <[email protected]> Reviewed-by: Keith Randall <[email protected]> TryBot-Result: Gopher Robot <[email protected]> Reviewed-by: Heschi Kreinick <[email protected]> Run-TryBot: Keith Randall <[email protected]>
cloudflare · Sep 2, 2022 · 121344a · 121344a
1 parent f45c2d7
commit 121344a
Show file tree

Hide file tree

Showing 4 changed files with 54 additions and 8 deletions.
diff --git a/src/cmd/compile/internal/ssa/gen/ARM64.rules b/src/cmd/compile/internal/ssa/gen/ARM64.rules
@@ -64,8 +64,11 @@
 (Sqrt32 ...) => (FSQRTS ...)
 
 // lowering rotates
+// we do rotate detection in generic rules, if the following rules need to be changed, chcek generic rules first.
 (RotateLeft8 <t> x (MOVDconst [c])) => (Or8 (Lsh8x64 <t> x (MOVDconst [c&7])) (Rsh8Ux64 <t> x (MOVDconst [-c&7])))
+(RotateLeft8 <t> x y) => (OR <t> (SLL <t> x (ANDconst <typ.Int64> [7] y)) (SRL <t> (ZeroExt8to64 x) (ANDconst <typ.Int64> [7] (NEG <typ.Int64> y))))
 (RotateLeft16 <t> x (MOVDconst [c])) => (Or16 (Lsh16x64 <t> x (MOVDconst [c&15])) (Rsh16Ux64 <t> x (MOVDconst [-c&15])))
+(RotateLeft16 <t> x y) => (RORW <t> (ORshiftLL <typ.UInt32> (ZeroExt16to32 x) (ZeroExt16to32 x) [16]) (NEG <typ.Int64> y))
 (RotateLeft32 x y) => (RORW x (NEG <y.Type> y))
 (RotateLeft64 x y) => (ROR x (NEG <y.Type> y))
 

diff --git a/src/cmd/compile/internal/ssa/rewrite.go b/src/cmd/compile/internal/ssa/rewrite.go
@@ -1986,9 +1986,9 @@ func canRotate(c *Config, bits int64) bool {
  return false
  }
  switch c.arch {
- case "386", "amd64":
+ case "386", "amd64", "arm64":
  return true
- case "arm", "arm64", "s390x", "ppc64", "ppc64le", "wasm", "loong64":
+ case "arm", "s390x", "ppc64", "ppc64le", "wasm", "loong64":
  return bits >= 32
  default:
  return false

diff --git a/src/cmd/compile/internal/ssa/rewriteARM64.go b/src/cmd/compile/internal/ssa/rewriteARM64.go
diff --git a/test/codegen/mathbits.go b/test/codegen/mathbits.go
@@ -258,14 +258,16 @@ func RotateLeft32(n uint32) uint32 {
  return bits.RotateLeft32(n, 9)
 }
 
-func RotateLeft16(n uint16) uint16 {
+func RotateLeft16(n uint16, s int) uint16 {
  // amd64:"ROLW" 386:"ROLW"
- return bits.RotateLeft16(n, 5)
+ // arm64:"RORW",-"CSEL"
+ return bits.RotateLeft16(n, s)
 }
 
-func RotateLeft8(n uint8) uint8 {
+func RotateLeft8(n uint8, s int) uint8 {
  // amd64:"ROLB" 386:"ROLB"
- return bits.RotateLeft8(n, 5)
+ // arm64:"LSL","LSR",-"CSEL"
+ return bits.RotateLeft8(n, s)
 }
 
 func RotateLeftVariable(n uint, m int) uint {