Re: Fix PR target/19240

[Uros Bizjak <uros at kss-loka dot si>]

tbp wrote:

> > BTW: As it is claimed by reporter, mixing 387 and SSE code can introduce
> > ~15% slowdown for FP intensive code. I would like to point on similar
> > problem with FP compares
> >    
> The reporter is back to nominal performance with this patch ;) 
> Thanks a lot.
>  
Great :)

> Now i'd also like to point to SSE & FP compares but i'm unsure
> http://gcc.gnu.org/ml/gcc-patches/2005-01/msg00168.html address that
> too (as far as i grok it, it deals with NaNs an issue i had to
> struggle with and triggered
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18944).
>
>  
The patch for PR17767 also fixes your testcase for PR18944:

gcc -O2 -msse2 -mfpmath=sse pr18944.c

       pushl   %ebp
       movl    %esp, %ebp
       subl    $24, %esp
       movss   .LC0, %xmm0
       andl    $-16, %esp
       subl    $16, %esp
       divss   .LC1, %xmm0
       movss   %xmm0, -4(%ebp)
       movss   -4(%ebp), %xmm0
       shufps  $0, %xmm0, %xmm0
       cmpunordps      %xmm0, %xmm0
       call    ffg
       leave
       ret

However, I don't know what kind of joy gcc gets from moving %xmm0 to 
memory and back before shufps insn.

> Case in point:
>  401447:       fldz   
> [snip lots of SSE only operations]
>  401535:       movss  %xmm2,0xc(%esp)
>  40153b:       flds   0xc(%esp)
>  40153f:       fcomip %st(1),%st
>  401541:       jbe    401650
>
> Feel free to clue me :)
>  
Hm, yes, I think this is another case of TARGET_SSE_MATH problems.

Uros.