Re: Remove some of optimize_size checks in i386 backend

[Jan Hubicka <jh at suse dot cz>]

Hi,
I don't have much time to look into code curefully, but I believe this
is safe:
> 
> >It is just post-reload pattern, so I think we are safe here.
> 
> I don't think so, because:
> 
> 1) you can have a non-cached recog in eliminate_regs_in_insn.  That 
> recog will change the code of the insn and xor will always be emitted.

The separate xor patterns exists because xor needs flags register.
All loads of 0 before peephole2 are emit as mov pattern not clobbering
flags so recog will chose the non-xor variant independently on the
patterns with extra clobbers.

There is peephole2 pattern that translates them to the
xors/ors.  So even if the pattern is enabled but the peephole is not
happening, we still should emit mov0 correctly.

> 
> 2) new instructions generated after reload will always be emitted as 
> xors.  it can happen in reload_combine, for example.

We never output those directly, I believe.

Honza

> 
> Paolo