cmd/compile/internal/ssa/gen: eliminate unnecessary neg and xori on P…

…PC64

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 <[email protected]>
Reviewed-by: Cherry Mui <[email protected]>

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)

src/cmd/compile/internal/ssa/gen/PPC64Ops.go

@@ -396,7 +396,7 @@ func init() {
  {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