Re: [PATCH, soft-fp]: Use mode(word) as a return type from soft-fp comparisons

[Uros Bizjak <ubizjak at gmail dot com>]

Daniel Jacobowitz wrote:
> On Sat, Jun 09, 2007 at 11:05:18AM +0200, Uros Bizjak wrote:
>   
> > Hello!
> >
> > This patch fixes the problems exposed with the testcase below. The problem is, 
> > that soft-fp functions return int value, where gcc expects the return from 
> > soft-fp functions to be in word mode. On x86_64 target, word_mode equals 
> > DImode, where int equals SImode, leading to comparison failures for certain 
> > class of input operands.
> >     
>
> Isn't someone (Andreas Krebbel?) trying to eliminate uses of
> mode((word))?
>   

Yes, but regarding soft-fp compares, we have this comment in 
config/fp-bit.h:

/* The type of the result of a floating point comparison.  This must
  match `word_mode' in GCC for the target.  */
#ifndef CMPtype
typedef int CMPtype __attribute__ ((mode (word)));
#endif

And gcc indeed expands the check for returned variable in word_mode 
(DImode on x86_64).

The problem is, that glibc emits FP comparison return tristate as "int". 
soft-fp from glibc is currently not compatible with gcc expansion on 
targets, where word_mode != int.

Here is a $1m question:

Which part is to be fixed?

a) gcc+libgcc should be converted to return (libgcc) and to expect (gcc) 
"int" from soft-fp comparisons.
b) soft-fp from glibc should convert soft-fp functions to return word mode

Uros.