cmd/internal/obj: optimize wrapper method prologue for branch prediction

[josharian]

Wrapper functions (such as those found in CL 36809) have an extra prologue inserted during preprocessing. Quoting cmd/internal/obj/xk86/obj6.go (after CL 36833), what this does is:

		// if g._panic != nil && g._panic.argp == FP {
		//   g._panic.argp = bottom-of-frame
		// }
		//
		//	MOVQ g_panic(CX), BX
		//	TESTQ BX, BX
		//	JEQ end
		//	LEAQ (autoffset+8)(SP), DI
		//	CMPQ panic_argp(BX), DI
		//	JNE end
		//	MOVQ SP, panic_argp(BX)
		// end:
		//	NOP

These jumps are the wrong direction. Static branch prediction says forward jumps are not taken, but the first forward jump is almost always taken. Rearrange the code to fix that.

I have a CL, to be mailed momentarily, that does this. However, I am having a hard time writing a test for this, because I cannot find a way to trigger the rare (in-a-panic) cases. From reading the panic sources, this only happens when panicking inside a deferred statement (?), but even then I see the wrapper function in the traceback.

What exactly is this prologue for? How do I test that it is doing its job correctly?

cc @randall77 @minux

[gopherbot]

CL https://golang.org/cl/36893 mentions this issue.

[josharian]

Hmm. #7491 contains a standalone test, and there is a corresponding test in test/recover.go. So I guess there is test coverage. I still don't fully understand what's going on, but I guess that's ok for the purposes of CL 36893, which is just rearranging deck chairs.

[randall77]

I'm not sure I understand what is going on here either. The argp involved is used by recover() to determine which panic is being recovered. I guess when there's a wrapper involved the argp needs to be moved from the args of the *T function to the args of the T function, so that if the T function recovers it can squash the associated panic. So maybe this code triggers it?

type T int
func (t T) foo() {
    recover()
}
func f(t *T) {
   defer t.foo()
}

[gopherbot]

CL https://golang.org/cl/37611 mentions this issue.