cmd/compile: more boolean optimizations

[randall77]

func f(a *int, b *int) bool {
	if (a == nil) != (b == nil) {
		return false
	}
	return true
}

Compiles to:

	0x0000 00000 (tmp1.go:4)	MOVQ	"".a+8(FP), AX
	0x0005 00005 (tmp1.go:4)	TESTQ	AX, AX
	0x0008 00008 (tmp1.go:4)	SETNE	AL
	0x000b 00011 (tmp1.go:4)	XORL	$1, AX
	0x000e 00014 (tmp1.go:4)	MOVQ	"".b+16(FP), CX
	0x0013 00019 (tmp1.go:4)	TESTQ	CX, CX
	0x0016 00022 (tmp1.go:4)	SETNE	CL
	0x0019 00025 (tmp1.go:4)	XORL	$1, CX
	0x001c 00028 (tmp1.go:4)	CMPB	AL, CL
	0x001e 00030 (tmp1.go:4)	JEQ	38

SETNE/XORL can be replaced with SETEQ.
It might also be worthwhile to rewrite NeqB to XOR instead of SETNE/CMPB. Not sure it would help here, but it may in other places.

[mvdan]

Is there a reason why the code above is in the form if x { return false; }; return true? You could just write return (a == nil) == (b == nil). Or was that just to prove the point?

[randall77]

No reason, it was just simplified from an example where I saw the suboptimal code.

[josharian]

It might also be worthwhile to rewrite NeqB to XOR instead of SETNE/CMPB.

In this code, that causes a regression, because the result of the XOR then gets TESTed, whereas the SETNE flags can be re-used.

The rest is coming in a CL shortly.

[gopherbot]

CL https://golang.org/cl/42491 mentions this issue.