[X86] Redundant adjacent stack adjustments after call frame optimization

[zmodem]

Bugzilla Link	27140
Resolution	FIXED
Resolved on	Apr 07, 2016 11:01
Version	trunk
OS	Linux
Blocks	#26673
CC	@qcolombet

Extended Description

For example

void f(int, int);
void g() {
f(1, 2);
}

Compiled with Clang 3.8 for 32-bit x86 Linux at -Os:

g:
subl $12, %esp
subl $8, %esp
pushl $2
pushl $1
calll f
addl $16, %esp
addl $12, %esp
retl

I'm working on a patch.

[qcolombet]

Strange, this should already been handled by X86FrameLowering::mergeSPUpdates.

[zmodem]

Strange, this should already been handled by
X86FrameLowering::mergeSPUpdates.

I think we need to call that some more :-) Sent out http://reviews.llvm.org/D18627

[zmodem]

r265623 fixes most of these. The impact was very small (~3KB on chrome_child.dll) but at least the code looks less silly.

There is still opportunities to clean up more %esp adjustment, for example if we're doing "add $4, %esp" but %esp is actually dead because we later do "mov %esp, %esi" or something, but that would be a different project, maybe part of llvm/llvm-bugzilla-archive#27076 .

[rnk]

mentioned in issue llvm/llvm-bugzilla-archive#43155