cmd/compile: optimize absorbing InvertFlags into Noov comparisons for arm64

Previously (LessThanNoov (InvertFlags x)) is lowered as:
CSET
CSET
BIC
With this CL it's lowered as:
CSET
CSEL
This saves one instruction.

Similarly (GreaterEqualNoov (InvertFlags x)) is now lowered as:
CSET
CSINC

$ benchstat old.bench new.bench
goos: linux
goarch: arm64
                       │  old.bench  │             new.bench              │
                       │   sec/op    │   sec/op     vs base               │
InvertLessThanNoov-160   2.249n ± 2%   2.190n ± 1%  -2.62% (p=0.003 n=10)

Change-Id: Idd8979b7f4fe466e74b1a201c4aba7f1b0cffb0b
Reviewed-on: https://go-review.googlesource.com/c/go/+/526237
Reviewed-by: Heschi Kreinick <heschi@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Eric Fang <eric.fang@arm.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>

src/cmd/compile/internal/ssa/_gen/ARM64.rules

@@ -1569,8 +1569,8 @@
 (LessEqualF       (InvertFlags x)) => (GreaterEqualF x)
 (GreaterThanF     (InvertFlags x)) => (LessThanF x)
 (GreaterEqualF    (InvertFlags x)) => (LessEqualF x)
-(LessThanNoov     (InvertFlags x)) => (BIC (GreaterEqualNoov <typ.Bool> x) (Equal <typ.Bool> x))
+(LessThanNoov     (InvertFlags x)) => (CSEL0 [OpARM64NotEqual] (GreaterEqualNoov <typ.Bool> x) x)
-(GreaterEqualNoov (InvertFlags x)) => (OR (LessThanNoov <typ.Bool> x) (Equal <typ.Bool> x))
+(GreaterEqualNoov (InvertFlags x)) => (CSINC [OpARM64NotEqual] (LessThanNoov <typ.Bool> x) (MOVDconst [0]) x)
 
 // Boolean-generating instructions (NOTE: NOT all boolean Values) always
 // zero upper bit of the register; no need to zero-extend

src/cmd/compile/internal/ssa/bench_test.go

@@ -30,3 +30,21 @@ func BenchmarkPhioptPass(b *testing.B) {
 		}
 	}
 }
+
+type Point struct {
+	X, Y int
+}
+
+//go:noinline
+func sign(p1, p2, p3 Point) bool {
+	return (p1.X-p3.X)*(p2.Y-p3.Y)-(p2.X-p3.X)*(p1.Y-p3.Y) < 0
+}
+
+func BenchmarkInvertLessThanNoov(b *testing.B) {
+	p1 := Point{1, 2}
+	p2 := Point{2, 3}
+	p3 := Point{3, 4}
+	for i := 0; i < b.N; i++ {
+		sign(p1, p2, p3)
+	}
+}

src/cmd/compile/internal/ssa/rewriteARM64.go

@@ -5974,18 +5974,19 @@ func rewriteValueARM64_OpARM64GreaterEqualNoov(v *Value) bool {
 	b := v.Block
 	typ := &b.Func.Config.Types
 	// match: (GreaterEqualNoov (InvertFlags x))
-	// result: (OR (LessThanNoov <typ.Bool> x) (Equal <typ.Bool> x))
+	// result: (CSINC [OpARM64NotEqual] (LessThanNoov <typ.Bool> x) (MOVDconst [0]) x)
 	for {
 		if v_0.Op != OpARM64InvertFlags {
 			break
 		}
 		x := v_0.Args[0]
-		v.reset(OpARM64OR)
+		v.reset(OpARM64CSINC)
+		v.AuxInt = opToAuxInt(OpARM64NotEqual)
 		v0 := b.NewValue0(v.Pos, OpARM64LessThanNoov, typ.Bool)
 		v0.AddArg(x)
-		v1 := b.NewValue0(v.Pos, OpARM64Equal, typ.Bool)
+		v1 := b.NewValue0(v.Pos, OpARM64MOVDconst, typ.UInt64)
-		v1.AddArg(x)
+		v1.AuxInt = int64ToAuxInt(0)
-		v.AddArg2(v0, v1)
+		v.AddArg3(v0, v1, x)
 		return true
 	}
 	return false
@@ -6709,18 +6710,17 @@ func rewriteValueARM64_OpARM64LessThanNoov(v *Value) bool {
 	b := v.Block
 	typ := &b.Func.Config.Types
 	// match: (LessThanNoov (InvertFlags x))
-	// result: (BIC (GreaterEqualNoov <typ.Bool> x) (Equal <typ.Bool> x))
+	// result: (CSEL0 [OpARM64NotEqual] (GreaterEqualNoov <typ.Bool> x) x)
 	for {
 		if v_0.Op != OpARM64InvertFlags {
 			break
 		}
 		x := v_0.Args[0]
-		v.reset(OpARM64BIC)
+		v.reset(OpARM64CSEL0)
+		v.AuxInt = opToAuxInt(OpARM64NotEqual)
 		v0 := b.NewValue0(v.Pos, OpARM64GreaterEqualNoov, typ.Bool)
 		v0.AddArg(x)
-		v1 := b.NewValue0(v.Pos, OpARM64Equal, typ.Bool)
+		v.AddArg2(v0, x)
-		v1.AddArg(x)
-		v.AddArg2(v0, v1)
 		return true
 	}
 	return false

test/codegen/comparisons.go

@@ -788,3 +788,16 @@ func cmp7() {
 	cmp5[string]("") // force instantiation
 	cmp6[string]("") // force instantiation
 }
+
+type Point struct {
+	X, Y int
+}
+
+// invertLessThanNoov checks (LessThanNoov (InvertFlags x)) is lowered as
+// CMP, CSET, CSEL instruction sequence. InvertFlags are only generated under
+// certain conditions, see canonLessThan, so if the code below does not
+// generate an InvertFlags OP, this check may fail.
+func invertLessThanNoov(p1, p2, p3 Point) bool {
+	// arm64:`CMP`,`CSET`,`CSEL`
+	return (p1.X-p3.X)*(p2.Y-p3.Y)-(p2.X-p3.X)*(p1.Y-p3.Y) < 0
+}