On Wed, Oct 14, 2020 at 6:31 AM Hongyu Wang via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Uros Bizjak <ubiz...@gmail.com> 于2020年10月14日周三 下午7:19写道：
> >
> > > > Please also add -muintr to g++.dg/other/i386-{2,3}.C and
> > > > >> > > gcc.target/i386-sse-{12,13,14,22,23}.c. This will test new
> intrinsics
> > > > >> > > header.
> > > > >> > >
> > > > >> >
> > > > >> > Thanks for your review. We found that without adding -muintr,
> the intrinsics header could also be tested. Make-check for these file all
> get passed.
> > > > >> >
> > > > >> > And there is no intrinsic/builtin with const int parameter. So
> we remove -muintr from these files.
> > > > >>
> > > > >> Can your double check that relevant instructions are indeed
> generated?
> > > > >> Without -muintr, relevant patterns in i386.md are effectively
> blocked,
> > > > >> and perhaps a call to __builtin_ia32_* is generated instead.
> > > > >
> > > > >
> > > > > Yes, in sse-14.s we have
> > > > >
> > > > > _clui:
> > > > > .LFB136:
> > > > > .cfi_startproc
> > > > > pushq %rbp
> > > > > .cfi_def_cfa_offset 16
> > > > > .cfi_offset 6, -16
> > > > > movq %rsp, %rbp
> > > > > .cfi_def_cfa_register 6
> > > > > clui
> > > > > nop
> > > > > popq %rbp
> > > > > .cfi_def_cfa 7, 8
> > > > > ret
> > > > > .cfi_endproc
> > > >
> > > > Strange, without -muintr, it should not be generated, and some error
> > > > about failed inlining due to target specific option mismatch shoul be
> > > > emitted.
> > > >
> > > > Can you please investigate this a bit more?
> > > >
> > >
> > > Because of function target attribute?
> >
> > I don't think so. Please consider this similar testcase:
> >
> > --cut here--
> > #ifndef __SSE2__
> > #pragma GCC push_options
> > #pragma GCC target("sse2")
> > #define __DISABLE_SSE2__
> > #endif /* __SSE2__ */
> >
> > typedef double __v2df __attribute__ ((__vector_size__ (16)));
> > typedef double __m128d __attribute__ ((__vector_size__ (16),
> __may_alias__));
> >
> > extern __inline __m128d __attribute__((__gnu_inline__,
> > __always_inline__, __artificial__))
> > _mm_add_sd (__m128d __A, __m128d __B)
> > {
> > return (__m128d)__builtin_ia32_addsd ((__v2df)__A, (__v2df)__B);
> > }
> >
> > #ifdef __DISABLE_SSE2__
> > #undef __DISABLE_SSE2__
> > #pragma GCC pop_options
> > #endif /* __DISABLE_SSE2__ */
> >
> >
> > __v2df foo (__v2df a, __v2df b)
> > {
> > return _mm_add_sd (a, b);
> > }
> > --cut here--
> >
> > $ gcc -O2 -mno-sse2 -S -dp sse2.c
> > sse2.c: In function ‘foo’:
> > sse2.c:11:1: error: inlining failed in call to ‘always_inline’
> > ‘_mm_add_sd’: target specific option mismatch
> > 11 | _mm_add_sd (__m128d __A, __m128d __B)
> > | ^~~~~~~~~~
> > sse2.c:24:10: note: called from here
> > 24 | return _mm_add_sd (a, b);
> > | ^~~~~~~~~~~~~~~~~
> >
> > I'd expect some similar warning from missing -mumip.
> >
>
> For this case, I can confirm uintr could generate similar warning without
> -muintr. But
> sse-{12,13,14,22,23}.c will not test intrinsic call for uintr, since it
> doesn't have const
> int parameter intrinsics.
>
> sse-{13,14,22,23}.c has
>
> #define extern
> #define __inline
>
> So intrinsic will be treated as common call to builtin, then
>
> #pragma GCC push_options
> #pragma GCC target("uintr")
>
> ensures the builtin could be expanded correctly.
>
> I think the intrinsic call test should be in uintr-1.c, so it is redundant
> to add -muintr in sse-{12,13,14,22,23}.c
> or x86gprintrin-*.c.
Please add UINTR intrinsic tests to x86gprintrin-*.c to cover such
usages.
> >
> > Uros.
--
H.J.
