https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92342
--- Comment #16 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Seems on x86_64-linux we also used to give the same code in 3.4.6 and earlier,
but starting already with 4.0 (already r90000) we emit
cmpl %edx, %edi
movl $0, %eax
cmove %esi, %eax
for f and
xorl %eax, %eax
cmpl %edx, %edi
sete %al
negl %eax
andl %esi, %eax
for g. We do it even for -Os, where the f version is clearly a win, for -O2
I'm afraid it depends on how soon the cmove result is used. LLVM seems to use
cmove in both cases, ICC does roughly what gcc does (honor what user wrote).
Not sure if we can do this transformation in match.pd though, because for a ? b
: 0; the gimplifier doesn't emit a COND_EXPR, but GIMPLE_COND and PHI, I'm
afraid we can't create new bbs from within match.pd.
