[PATCH 32/59] x86: Adapt "%v" usage on clang to emit VEX enconding

[H.J. Lu]

On Tue, Nov 11, 2025 at 12:30 PM Collin Funk <collin.funk1@gmail.com> wrote:
>
> Hi Adhemerval,
>
> Adhemerval Zanella Netto <adhemerval.zanella@linaro.org> writes:
>
> > On 17/10/25 19:27, H.J. Lu wrote:
> >> On Sat, Oct 18, 2025 at 3:54 AM Adhemerval Zanella
> >> <adhemerval.zanella@linaro.org> wrote:
> >>>
> >>> clang does not support the %v to select the AVX encoding, nor the '%d' asm
> >>> contrain, and for AVX build it requires all 3 arguments.
> >>
> >> Does this patch change GCC codgen?
> >
> > It should not, gcc will still use %v as current. Only clang will
> > use a different instruction selection and asm constraint:
> >
> > +#ifdef __clang__
> > +# ifdef __AVX__
> > +#  define VPREFIX    "v"
> > +#  define VROUND_ARG ", %0"
> > +# else
> > +#  define VPREFIX    ""
> > +#  define VROUND_ARG ""
> > +# endif
> > +# define VARGPREFIX  "%"
> > +#else
> > +# define VPREFIX     "%v"
> > +# define VARGPREFIX  "%d"
> > +# define VROUND_ARG  ""
> > +#endif
>
> I bisected 65 test failures to this commit. On a x86_64 Fedora 43 system
> with the following gcc:
>
>     $ gcc --version | head -n 1
>     gcc (GCC) 15.2.1 20251022 (Red Hat 15.2.1-3)
>
> The main ones are:
>
>     FAIL: math/test-fenv
>     FAIL: math/test-fenv-clear
>     FAIL: math/test-fenv-clear-sse
>     FAIL: math/test-fenv-return
>     FAIL: math/test-fenv-x87

How can I reproduce these on Fedora 43?

> The others are likely just a result of the exception mask and/or
> rounding mode not being set correctly.
>
> I am a bit surprised since the patch looks okay upon my quick skim over
> it, just moving some inline asm statements into inline functions.
>
> Let me know if you want me to file a bug report.
>
> Collin



-- 
H.J.