cmd/compile/internal/ssa/gen: eliminate unnecessary neg and xori on PPC64
This adds a few rules to PPC64 to eliminate some instructions:
- when an isel is used to generate a boolean value based on a
condition and followed by an xori to flip the result, it can
instead flip the operands in the isel and avoid the xori.
= when a neg follows a sub the operands to the sub can be
swapped and the neg avoided.
There are several opportunities in reflect.DeepEqual to omit
xori which improves some of its benchmarks by as much as
5%
Change-Id: I81bbc02c0f16995c65934b6f045867b731ab302b
Reviewed-on: https://go-review.googlesource.com/c/go/+/357509
Trust: Lynn Boger <laboger@linux.vnet.ibm.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
diff --git a/src/cmd/compile/internal/ssa/gen/PPC64.rules b/src/cmd/compile/internal/ssa/gen/PPC64.rules
index 8e42bae..ccca72a 100644
--- a/src/cmd/compile/internal/ssa/gen/PPC64.rules
+++ b/src/cmd/compile/internal/ssa/gen/PPC64.rules
@@ -848,6 +848,7 @@
(ADDconst [c] (SUBFCconst [d] x)) && is32Bit(c+d) => (SUBFCconst [c+d] x)
(NEG (ADDconst [c] x)) && is32Bit(-c) => (SUBFCconst [-c] x)
(NEG (SUBFCconst [c] x)) && is32Bit(-c) => (ADDconst [-c] x)
+(NEG (SUB x y)) => (SUB y x)
// Use register moves instead of stores and loads to move int<=>float values
// Common with math Float64bits, Float64frombits
@@ -1087,7 +1088,7 @@
((CMP|CMPW|CMPU|CMPWU) x y) && canonLessThan(x,y) => (InvertFlags ((CMP|CMPW|CMPU|CMPWU) y x))
// ISEL auxInt values 0=LT 1=GT 2=EQ arg2 ? arg0 : arg1
-// ISEL auxInt values 4=GE 5=LE 6=NE arg2 ? arg1 : arg0
+// ISEL auxInt values 4=GE 5=LE 6=NE !arg2 ? arg1 : arg0
// ISELB special case where arg0, arg1 values are 0, 1
(Equal cmp) => (ISELB [2] (MOVDconst [1]) cmp)
@@ -1138,6 +1139,9 @@
(ISEL [n] x y (InvertFlags bool)) && n%4 == 0 => (ISEL [n+1] x y bool)
(ISEL [n] x y (InvertFlags bool)) && n%4 == 1 => (ISEL [n-1] x y bool)
(ISEL [n] x y (InvertFlags bool)) && n%4 == 2 => (ISEL [n] x y bool)
+(XORconst [1] (ISELB [6] (MOVDconst [1]) cmp)) => (ISELB [2] (MOVDconst [1]) cmp)
+(XORconst [1] (ISELB [5] (MOVDconst [1]) cmp)) => (ISELB [1] (MOVDconst [1]) cmp)
+(XORconst [1] (ISELB [4] (MOVDconst [1]) cmp)) => (ISELB [0] (MOVDconst [1]) cmp)
// A particular pattern seen in cgo code:
(AND (MOVDconst [c]) x:(MOVBZload _ _)) => (ANDconst [c&0xFF] x)
diff --git a/src/cmd/compile/internal/ssa/gen/PPC64Ops.go b/src/cmd/compile/internal/ssa/gen/PPC64Ops.go
index ff9ce64..42775fa 100644
--- a/src/cmd/compile/internal/ssa/gen/PPC64Ops.go
+++ b/src/cmd/compile/internal/ssa/gen/PPC64Ops.go
@@ -396,7 +396,7 @@
{name: "CMPWUconst", argLength: 1, reg: gp1cr, asm: "CMPWU", aux: "Int32", typ: "Flags"},
// ISEL auxInt values 0=LT 1=GT 2=EQ arg2 ? arg0 : arg1
- // ISEL auxInt values 4=GE 5=LE 6=NE arg2 ? arg1 : arg0
+ // ISEL auxInt values 4=GE 5=LE 6=NE !arg2 ? arg1 : arg0
// ISELB special case where arg0, arg1 values are 0, 1 for boolean result
{name: "ISEL", argLength: 3, reg: crgp21, asm: "ISEL", aux: "Int32", typ: "Int32"}, // see above
{name: "ISELB", argLength: 2, reg: crgp11, asm: "ISEL", aux: "Int32", typ: "Int32"}, // see above
diff --git a/src/cmd/compile/internal/ssa/rewritePPC64.go b/src/cmd/compile/internal/ssa/rewritePPC64.go
index 1e6624e..ea1c1fa 100644
--- a/src/cmd/compile/internal/ssa/rewritePPC64.go
+++ b/src/cmd/compile/internal/ssa/rewritePPC64.go
@@ -11372,6 +11372,18 @@
v.AddArg(x)
return true
}
+ // match: (NEG (SUB x y))
+ // result: (SUB y x)
+ for {
+ if v_0.Op != OpPPC64SUB {
+ break
+ }
+ y := v_0.Args[1]
+ x := v_0.Args[0]
+ v.reset(OpPPC64SUB)
+ v.AddArg2(y, x)
+ return true
+ }
return false
}
func rewriteValuePPC64_OpPPC64NOR(v *Value) bool {
@@ -13912,6 +13924,8 @@
}
func rewriteValuePPC64_OpPPC64XORconst(v *Value) bool {
v_0 := v.Args[0]
+ b := v.Block
+ typ := &b.Func.Config.Types
// match: (XORconst [c] (XORconst [d] x))
// result: (XORconst [c^d] x)
for {
@@ -13936,6 +13950,60 @@
v.copyOf(x)
return true
}
+ // match: (XORconst [1] (ISELB [6] (MOVDconst [1]) cmp))
+ // result: (ISELB [2] (MOVDconst [1]) cmp)
+ for {
+ if auxIntToInt64(v.AuxInt) != 1 || v_0.Op != OpPPC64ISELB || auxIntToInt32(v_0.AuxInt) != 6 {
+ break
+ }
+ cmp := v_0.Args[1]
+ v_0_0 := v_0.Args[0]
+ if v_0_0.Op != OpPPC64MOVDconst || auxIntToInt64(v_0_0.AuxInt) != 1 {
+ break
+ }
+ v.reset(OpPPC64ISELB)
+ v.AuxInt = int32ToAuxInt(2)
+ v0 := b.NewValue0(v.Pos, OpPPC64MOVDconst, typ.Int64)
+ v0.AuxInt = int64ToAuxInt(1)
+ v.AddArg2(v0, cmp)
+ return true
+ }
+ // match: (XORconst [1] (ISELB [5] (MOVDconst [1]) cmp))
+ // result: (ISELB [1] (MOVDconst [1]) cmp)
+ for {
+ if auxIntToInt64(v.AuxInt) != 1 || v_0.Op != OpPPC64ISELB || auxIntToInt32(v_0.AuxInt) != 5 {
+ break
+ }
+ cmp := v_0.Args[1]
+ v_0_0 := v_0.Args[0]
+ if v_0_0.Op != OpPPC64MOVDconst || auxIntToInt64(v_0_0.AuxInt) != 1 {
+ break
+ }
+ v.reset(OpPPC64ISELB)
+ v.AuxInt = int32ToAuxInt(1)
+ v0 := b.NewValue0(v.Pos, OpPPC64MOVDconst, typ.Int64)
+ v0.AuxInt = int64ToAuxInt(1)
+ v.AddArg2(v0, cmp)
+ return true
+ }
+ // match: (XORconst [1] (ISELB [4] (MOVDconst [1]) cmp))
+ // result: (ISELB [0] (MOVDconst [1]) cmp)
+ for {
+ if auxIntToInt64(v.AuxInt) != 1 || v_0.Op != OpPPC64ISELB || auxIntToInt32(v_0.AuxInt) != 4 {
+ break
+ }
+ cmp := v_0.Args[1]
+ v_0_0 := v_0.Args[0]
+ if v_0_0.Op != OpPPC64MOVDconst || auxIntToInt64(v_0_0.AuxInt) != 1 {
+ break
+ }
+ v.reset(OpPPC64ISELB)
+ v.AuxInt = int32ToAuxInt(0)
+ v0 := b.NewValue0(v.Pos, OpPPC64MOVDconst, typ.Int64)
+ v0.AuxInt = int64ToAuxInt(1)
+ v.AddArg2(v0, cmp)
+ return true
+ }
return false
}
func rewriteValuePPC64_OpPanicBounds(v *Value) bool {
