gcc's x86 "RET"-machine instruction optimization

gcc's x86 "RET"-machine instruction optimization

[Josef Angermeier]

Hello

Im trying to write a script which patches gcc's assembler output (-S
option) in that way, that those functions which are exported (.gobl
asm-directive) return with a near-jump instead of a
near-return-instruction. (Im writing a BIOS for a virtual machine...)

So i fist that it would be an easy task to replace all
"retl"-instructions by "lretw" ones in gcc's asm code. But then now im
facing a problem:

C code like that:
kbd.c:

bios_16_xxxx(struct regs *regs)
{
	if (AH == 0x00) {
		bios_16_00xx(regs);
	} else {
		....
	}
}

kbd.s

bios_16_xxxx:
.LCFI71:
	...
	jmp	bios_16_00xx
.L105:
	...
	lretw		(PATCHED RETURN)


....


bios_16_00xx:
.LFB59:
	...
	ret


GCC uses a jmp-instruction instead of a call one, so 'speeds up' execution by
saving the execution of bios_16_xxxx's return instruction (lretw one). This of course
ruins my efforts. Therfore i'd like to know how this optimization is called, and if
there is any gcc option to disable this particular optimization.

Big thanks in advance
josef

Re: gcc's x86 "RET"-machine instruction optimization

[Ian Lance Taylor]

Josef Angermeier <sijoange@cip.informatik.uni-erlangen.de> writes:

> GCC uses a jmp-instruction instead of a call one, so 'speeds up' execution by
> saving the execution of bios_16_xxxx's return instruction (lretw one). This of course
> ruins my efforts. Therfore i'd like to know how this optimization is called, and if
> there is any gcc option to disable this particular optimization.

These are called sibling calls, and you can turn the optimization off
using -fno-optimize-sibling-calls.

(I wouldn't be surprised if you run into other problems with your
approach, but that switch should take care of that particular
problem.)

Ian

Re: gcc's x86 "RET"-machine instruction optimization

[Josef Angermeier]

Hello

> > GCC uses a jmp-instruction instead of a call one, so 'speeds up' execution by
> > saving the execution of bios_16_xxxx's return instruction (lretw one). This of course
> > ruins my efforts. Therfore i'd like to know how this optimization is called, and if
> > there is any gcc option to disable this particular optimization.

> These are called sibling calls, and you can turn the optimization off
> using -fno-optimize-sibling-calls.

oh great!

> (I wouldn't be surprised if you run into other problems with your
> approach, but that switch should take care of that particular
> problem.)

Mmm, i'd appreciate all hints to what problems could happen when going
this approach. Doesn't matter how vague. 

> Ian
Thanks Ian, again !

josef
ciao