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>

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
+}