On Fri, Sep 23, 2011 at 10:12 AM, Jakub Jelinek <ja...@redhat.com> wrote:
>> > --- gcc/config/i386/i386.c.jj 2011-09-02 16:29:38.000000000 +0200
>> > +++ gcc/config/i386/i386.c 2011-09-07 21:57:52.000000000 +0200
>> > @@ -18304,6 +18304,11 @@ ix86_prepare_sse_fp_compare_args (rtx de
>> > {
>> > rtx tmp;
>> >
>> > + /* AVX supports all the needed comparisons, no need to swap arguments
>> > + nor help reload. */
>> > + if (TARGET_AVX)
>> > + return code;
>> > +
>>
>> Unfortunately, this part prevents generation of vmin/vmax instructions
>> for TARGET_AVX. In ix86_expand_fp_movcc, we call
>> ix86_prepare_sse_fp_compare_args, where we previously converted GT
>> into LT. LT is recognized in ix86_expand_sse_fp_minmax as valid
>> operand for MIN/MAX, whereas GT is not.
>>
>> I'm not sure if we can swap operands in ix86_expand_sse_fp_minmax,
>> there is a scary comment in front of ix86_expand_sse_fp_minmax w.r.t.
>> to IEEE safety.
>
> swap_condition is documented to be IEEE safe:
> /* Similar, but return the code when two operands of a comparison are swapped.
> This IS safe for IEEE floating-point. */
>
> So, do you prefer this?
Yes, since the operands can be swapped, this looks OK to me. Also, the
comments are very informative.
> 2011-09-23 Jakub Jelinek <ja...@redhat.com>
>
> * config/i386/i386.c (ix86_prepare_sse_fp_compare_args): For
> GE/GT/UNLE/UNLT swap arguments and condition even for TARGET_AVX.
>
> * gcc.target/i386/avxfp-1.c: New test.
> * gcc.target/i386/avxfp-2.c: New test.
OK.
Thanks,
Uros.
