Re: [PATCH, i386]: FIx PR target/32413, [4.3 Regression] internal compiler error: in reload_cse_simplify_operands, at postreload.c:396

[Uros Bizjak <ubizjak at gmail dot com>]

H. J. Lu wrote:

> > > > The value itself or the fact that gcc moves SI value to SSE for this 
> > > >         
> > > value?
> > >
> > > It is odd for gcc to move SI value to SSE for this value.
> > >       
> > The value is the same as for reg-reg move. So, when RA is out of
> > integer registers, it starts to allocate SSE regs.
> >
> >     
>
> We don't want use SSE registers as integer registers. Even if
> they have same cost, it is still odd for RA to do that since
> you con't really use scalar integer mode on SSE registers. 
>   
It is possible to move SImode and DImode to mmx/sse registers (movd and 
movq), but practically, you can't do anything with the value there. And 
having instructions with memory operands, memory is far better choice 
for a temporary storage. The proposed patch will prevent SImode values 
from SSE registers for all but most extreme situations by simply raising 
the bar high enough.

I have committed the patch to SVN mainline. It produces noticeably 
better code for -march=core2, without interclass register moves.

Uros.