https://gcc.gnu.org/bugzilla/show_bug.cgi?id=12395
--- Comment #20 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #19)
> (In reply to Andrew Pinski from comment #18)
> > This is what is produced (at least for 7.3.0):
> Which has been produced since GCC 6.
> that is due to ifcvt.c changes.
> On the trunk we can produce on the gimple now:
> a.1_1 = a;
> if (a.1_1 != -1)
> goto <bb 3>; [50.00%]
> else
> goto <bb 4>; [50.00%]
>
> <bb 3> [local count: 536870913]:
> _3 = a.1_1 + 2;
>
> <bb 4> [local count: 1073741824]:
> # _2 = PHI <0(2), _3(3)>
> a = _2;
>
> Note most optimial code is really:
> movl a(%rip), %eax
> addl $1, %eax
> adcl $0, %eax
> movl %eax, a(%rip)
I see
movl a(%rip), %edx
leal 2(%rdx), %eax
cmpl $-1, %edx
movl $0, %edx
cmove %edx, %eax
movl %eax, a(%rip)
with -O2 or
movl a(%rip), %edx
xorl %eax, %eax
cmpl $-1, %edx
je .L2
leal 2(%rdx), %eax
.L2:
movl %eax, a(%rip)
with -Os. The GIMPLE code makes it difficult to do better I guess, it
doesn't help that we don't have any CC there.
