Re: i386 fix

[Jan Hubicka <jh at suse dot cz>]

> On Sat, Jan 06, 2001 at 10:52:53PM +0100, Jan Hubicka wrote:
> > 	* i386.c (ix86_expand_int_movcc): Take care to reverse fp conditions
> > 	properly.
> 
> I'm not fond of this because in principal it causes ix86_fp_compare_mode
> to return incorrect results.  Except that there's a hack at the end of
> that function that hides the lossage.
> 
> Namely, when UNGE is reversed to LT, we *still* need to use CCFPUmode.
Currently, we always use CCFPUmode when -ffast-math is off. 
The relevant bit is near the end of ix86_fp_compare_mode:

  /* ??? If we knew whether invalid-operand exceptions were masked,
     we could rely on fcom to raise an exception and take care of
     NaNs.  But we don't.  We could know this from c99 math pragmas.  */
  if (TARGET_IEEE_FP)
    unordered = 1;

  return unordered ? CCFPUmode : CCFPmode;

so this transformation is corret with current code. I agree this situation is ill
and I would like to eliminate it in future few patches, but still I would like
to install this bit - currently it will prevent gcc from crashing when
trying to construct setcc based on isgreater builtin and in future it will
be correct too, if I finish my transformation.

Honza
> 
> 
> r~