Re: [PATCH, i386] Unify TARGET_SSE_MATH and TARGET_MIX_SSE_I387 ininsn constraints

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

Richard Henderson wrote:

> On Wed, Dec 08, 2004 at 05:21:21PM +0100, Uros Bizjak wrote:
>  
>
> > There is a big mess regarding TARGET_SSE_MATH and TARGET_MIX_SSE_I387 
> > handling in i386.md file. Actually every pattern handles these macros in 
> > its own way and it is very difficult to figure out the conditions under 
> > which certain pattern is enabled.
> >    
>
> I think this is too invasive for stage3.  Please hold onto this
> for when 4.1 opens.  If there is a bug fix in here, perhaps it
> can be extracted by itself?
>  
There is a problem with integer conversion from SI -> SF and SI -> DF 
for '-march=pentium4 -mfpmath=387' (the default), where SSE cvtsi2ss and 
cvtsi2sd are used instead of 387's fild insn. The testcase:

float test1sisf(int x) {
   return x;
}

double test1sidf(int x) {
   return x;
}

is then compiled into:

test1sisf:
       pushl   %ebp
       movl    %esp, %ebp
       subl    $4, %esp
       cvtsi2ss        8(%ebp), %xmm0
       movss   %xmm0, -4(%ebp)
       flds    -4(%ebp)
       leave
       ret

test1sidf:
       pushl   %ebp
       movl    %esp, %ebp
       subl    $8, %esp
       cvtsi2sd        8(%ebp), %xmm0
       movsd   %xmm0, -8(%ebp)
       fldl    -8(%ebp)
       leave
       ret

This is quite bad code, compared to what it should produce for both 
testcases:

test1sisf:
       pushl   %ebp
       movl    %esp, %ebp
       fildl   8(%ebp)
       popl    %ebp
       ret

I tried to solve this problem with the minimum patch, but I got lost in 
the maze of options. That is why I propose this approach. Actually, the 
biggest part of patch is the solution to "float" patterns. The change to 
other fop_* patterns is just a rename, so the minimum patch for "float" 
patterns would be perhaps some 10 lines shorter...

Uros.