diff --git a/src/cmd/compile/internal/ssa/gen/ARM64.rules b/src/cmd/compile/internal/ssa/gen/ARM64.rules index d4b47bfb0b..6539a0ce7b 100644 --- a/src/cmd/compile/internal/ssa/gen/ARM64.rules +++ b/src/cmd/compile/internal/ssa/gen/ARM64.rules @@ -1863,9 +1863,8 @@ (XORshiftLL [c] (UBFX [bfc] x) x2) && c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c) -> (EXTRWconst [32-c] x2 x) -// Generic rules rewrite certain AND to a pair of shifts. -// However, on ARM64 the bitmask can fit into an instruction. -// Rewrite it back to AND. +// Rewrite special pairs of shifts to AND. +// On ARM64 the bitmask can fit into an instruction. (SRLconst [c] (SLLconst [c] x)) && 0 < c && c < 64 -> (ANDconst [1< (ANDconst [^(1< 0 -> (Const64 [-1]) (Slicemask (Const64 [0])) -> (Const64 [0]) -// Rewrite AND of consts as shifts if possible, slightly faster for 64 bit operands -// leading zeros can be shifted left, then right -(And64 (Const64 [y]) x) && nlz(y) + nto(y) == 64 && nto(y) >= 32 - -> (Rsh64Ux64 (Lsh64x64 x (Const64 [nlz(y)])) (Const64 [nlz(y)])) -// trailing zeros can be shifted right, then left -(And64 (Const64 [y]) x) && nlo(y) + ntz(y) == 64 && ntz(y) >= 32 - -> (Lsh64x64 (Rsh64Ux64 x (Const64 [ntz(y)])) (Const64 [ntz(y)])) - // simplifications often used for lengths. e.g. len(s[i:i+5])==5 (Sub(64|32|16|8) (Add(64|32|16|8) x y) x) -> y (Sub(64|32|16|8) (Add(64|32|16|8) x y) y) -> x diff --git a/src/cmd/compile/internal/ssa/rewritegeneric.go b/src/cmd/compile/internal/ssa/rewritegeneric.go index 8aa07d20db..a2d091d3d6 100644 --- a/src/cmd/compile/internal/ssa/rewritegeneric.go +++ b/src/cmd/compile/internal/ssa/rewritegeneric.go @@ -5735,112 +5735,6 @@ func rewriteValuegeneric_OpAnd64_10(v *Value) bool { v.AddArg(y) return true } - // match: (And64 (Const64 [y]) x) - // cond: nlz(y) + nto(y) == 64 && nto(y) >= 32 - // result: (Rsh64Ux64 (Lsh64x64 x (Const64 [nlz(y)])) (Const64 [nlz(y)])) - for { - t := v.Type - x := v.Args[1] - v_0 := v.Args[0] - if v_0.Op != OpConst64 { - break - } - y := v_0.AuxInt - if !(nlz(y)+nto(y) == 64 && nto(y) >= 32) { - break - } - v.reset(OpRsh64Ux64) - v0 := b.NewValue0(v.Pos, OpLsh64x64, t) - v0.AddArg(x) - v1 := b.NewValue0(v.Pos, OpConst64, t) - v1.AuxInt = nlz(y) - v0.AddArg(v1) - v.AddArg(v0) - v2 := b.NewValue0(v.Pos, OpConst64, t) - v2.AuxInt = nlz(y) - v.AddArg(v2) - return true - } - // match: (And64 x (Const64 [y])) - // cond: nlz(y) + nto(y) == 64 && nto(y) >= 32 - // result: (Rsh64Ux64 (Lsh64x64 x (Const64 [nlz(y)])) (Const64 [nlz(y)])) - for { - t := v.Type - _ = v.Args[1] - x := v.Args[0] - v_1 := v.Args[1] - if v_1.Op != OpConst64 { - break - } - y := v_1.AuxInt - if !(nlz(y)+nto(y) == 64 && nto(y) >= 32) { - break - } - v.reset(OpRsh64Ux64) - v0 := b.NewValue0(v.Pos, OpLsh64x64, t) - v0.AddArg(x) - v1 := b.NewValue0(v.Pos, OpConst64, t) - v1.AuxInt = nlz(y) - v0.AddArg(v1) - v.AddArg(v0) - v2 := b.NewValue0(v.Pos, OpConst64, t) - v2.AuxInt = nlz(y) - v.AddArg(v2) - return true - } - // match: (And64 (Const64 [y]) x) - // cond: nlo(y) + ntz(y) == 64 && ntz(y) >= 32 - // result: (Lsh64x64 (Rsh64Ux64 x (Const64 [ntz(y)])) (Const64 [ntz(y)])) - for { - t := v.Type - x := v.Args[1] - v_0 := v.Args[0] - if v_0.Op != OpConst64 { - break - } - y := v_0.AuxInt - if !(nlo(y)+ntz(y) == 64 && ntz(y) >= 32) { - break - } - v.reset(OpLsh64x64) - v0 := b.NewValue0(v.Pos, OpRsh64Ux64, t) - v0.AddArg(x) - v1 := b.NewValue0(v.Pos, OpConst64, t) - v1.AuxInt = ntz(y) - v0.AddArg(v1) - v.AddArg(v0) - v2 := b.NewValue0(v.Pos, OpConst64, t) - v2.AuxInt = ntz(y) - v.AddArg(v2) - return true - } - // match: (And64 x (Const64 [y])) - // cond: nlo(y) + ntz(y) == 64 && ntz(y) >= 32 - // result: (Lsh64x64 (Rsh64Ux64 x (Const64 [ntz(y)])) (Const64 [ntz(y)])) - for { - t := v.Type - _ = v.Args[1] - x := v.Args[0] - v_1 := v.Args[1] - if v_1.Op != OpConst64 { - break - } - y := v_1.AuxInt - if !(nlo(y)+ntz(y) == 64 && ntz(y) >= 32) { - break - } - v.reset(OpLsh64x64) - v0 := b.NewValue0(v.Pos, OpRsh64Ux64, t) - v0.AddArg(x) - v1 := b.NewValue0(v.Pos, OpConst64, t) - v1.AuxInt = ntz(y) - v0.AddArg(v1) - v.AddArg(v0) - v2 := b.NewValue0(v.Pos, OpConst64, t) - v2.AuxInt = ntz(y) - v.AddArg(v2) - return true - } // match: (And64 (And64 i:(Const64 ) z) x) // cond: (z.Op != OpConst64 && x.Op != OpConst64) // result: (And64 i (And64 z x)) @@ -5867,10 +5761,6 @@ func rewriteValuegeneric_OpAnd64_10(v *Value) bool { v.AddArg(v0) return true } - return false -} -func rewriteValuegeneric_OpAnd64_20(v *Value) bool { - b := v.Block // match: (And64 (And64 z i:(Const64 )) x) // cond: (z.Op != OpConst64 && x.Op != OpConst64) // result: (And64 i (And64 z x)) @@ -5984,6 +5874,10 @@ func rewriteValuegeneric_OpAnd64_20(v *Value) bool { v.AddArg(x) return true } + return false +} +func rewriteValuegeneric_OpAnd64_20(v *Value) bool { + b := v.Block // match: (And64 (Const64 [c]) (And64 x (Const64 [d]))) // cond: // result: (And64 (Const64 [c&d]) x) diff --git a/test/codegen/math.go b/test/codegen/math.go index 36252710d1..8aa652595b 100644 --- a/test/codegen/math.go +++ b/test/codegen/math.go @@ -81,7 +81,7 @@ func abs32(x float32) float32 { // Check that it's using integer registers func copysign(a, b, c float64) { - // amd64:"BTRQ\t[$]63","SHRQ\t[$]63","SHLQ\t[$]63","ORQ" + // amd64:"BTRQ\t[$]63","ANDQ","ORQ" // s390x:"CPSDR",-"MOVD" (no integer load/store) // ppc64:"FCPSGN" // ppc64le:"FCPSGN" @@ -100,7 +100,7 @@ func copysign(a, b, c float64) { // s390x:"LNDFR\t",-"MOVD\t" (no integer load/store) sink64[2] = math.Float64frombits(math.Float64bits(a) | 1<<63) - // amd64:-"SHLQ\t[$]1",-"SHRQ\t[$]1","SHRQ\t[$]63","SHLQ\t[$]63","ORQ" + // amd64:"ANDQ","ORQ" // s390x:"CPSDR\t",-"MOVD\t" (no integer load/store) // ppc64:"FCPSGN" // ppc64le:"FCPSGN"