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 <khr@google.com>
Auto-Submit: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Heschi Kreinick <heschi@google.com>
Run-TryBot: Keith Randall <khr@golang.org>
(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))
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
v.AddArg2(v0, v2)
return true
}
- return false
+ // match: (RotateLeft16 <t> x y)
+ // result: (RORW <t> (ORshiftLL <typ.UInt32> (ZeroExt16to32 x) (ZeroExt16to32 x) [16]) (NEG <typ.Int64> y))
+ for {
+ t := v.Type
+ x := v_0
+ y := v_1
+ v.reset(OpARM64RORW)
+ v.Type = t
+ v0 := b.NewValue0(v.Pos, OpARM64ORshiftLL, typ.UInt32)
+ v0.AuxInt = int64ToAuxInt(16)
+ v1 := b.NewValue0(v.Pos, OpZeroExt16to32, typ.UInt32)
+ v1.AddArg(x)
+ v0.AddArg2(v1, v1)
+ v2 := b.NewValue0(v.Pos, OpARM64NEG, typ.Int64)
+ v2.AddArg(y)
+ v.AddArg2(v0, v2)
+ return true
+ }
}
func rewriteValueARM64_OpRotateLeft32(v *Value) bool {
v_1 := v.Args[1]
v.AddArg2(v0, v2)
return true
}
- return false
+ // match: (RotateLeft8 <t> x y)
+ // result: (OR <t> (SLL <t> x (ANDconst <typ.Int64> [7] y)) (SRL <t> (ZeroExt8to64 x) (ANDconst <typ.Int64> [7] (NEG <typ.Int64> y))))
+ for {
+ t := v.Type
+ x := v_0
+ y := v_1
+ v.reset(OpARM64OR)
+ v.Type = t
+ v0 := b.NewValue0(v.Pos, OpARM64SLL, t)
+ v1 := b.NewValue0(v.Pos, OpARM64ANDconst, typ.Int64)
+ v1.AuxInt = int64ToAuxInt(7)
+ v1.AddArg(y)
+ v0.AddArg2(x, v1)
+ v2 := b.NewValue0(v.Pos, OpARM64SRL, t)
+ v3 := b.NewValue0(v.Pos, OpZeroExt8to64, typ.UInt64)
+ v3.AddArg(x)
+ v4 := b.NewValue0(v.Pos, OpARM64ANDconst, typ.Int64)
+ v4.AuxInt = int64ToAuxInt(7)
+ v5 := b.NewValue0(v.Pos, OpARM64NEG, typ.Int64)
+ v5.AddArg(y)
+ v4.AddArg(v5)
+ v2.AddArg2(v3, v4)
+ v.AddArg2(v0, v2)
+ return true
+ }
}
func rewriteValueARM64_OpRsh16Ux16(v *Value) bool {
v_1 := v.Args[1]
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 {