On Mon, 31 Mar 2025, Jakub Jelinek wrote:
> On Mon, Mar 31, 2025 at 03:12:56PM +0200, Richard Biener wrote:
> > The following fixes ix86_handle_option to properly handle -mno-sse4
> > which is always handled as -msse4 with value unset. I've verified
> > no other OPT_mno_* is left around.
> >
> > Bootstrap and regtest running on x86_64-unknown-linux-gnu.
> >
> > This error goes back quite some time, I'm not sure how prevalent
> > the use of -mno-sse4 is. In fact I have no idea why we have
> > OPT_mno_sse4 at all...
>
> That is because SSE4 is SSE4.1 + SSE4.2.
> So for -msse4 we want it to act as -msse4.2 and for -mno-sse4
> to act as -mno-sse4.1
Yes, but it seems ix86_handle_option handles this via
OPTION_MASK_ISA_SSE4_SET vs. ~OPTION_MASK_ISA_SSE4_UNSET. So the
question is what the -mno-sse4 options entry does. I'll note
that ix86_handle_option also unsets flags in ix86_isa_flags2
with -mno-sse4 but the mno-sse4 options entry does not indicate
that.
Richard.
> > msse4
> > Target RejectNegative Mask(ISA_SSE4_2) Var(ix86_isa_flags) Save
> > Support MMX, SSE, SSE2, SSE3, SSSE3, SSE4.1 and SSE4.2 built-in functions
> > and code generation.
> >
> > mno-sse4
> > Target RejectNegative InverseMask(ISA_SSE4_1) Var(ix86_isa_flags) Save
> > Do not support SSE4.1 and SSE4.2 built-in functions and code generation.
> >
> > but we end up building
> >
> > $3 = {opt_index = 2231, warn_message = 0x0, arg = 0x0,
> > orig_option_with_args_text = 0x440d108 "-msse4", canonical_option = {
> > 0x440d108 "-msse4", 0x0, 0x0, 0x0}, canonical_option_num_elements = 1,
> > value = 0, mask = 0, errors = 0}
> >
> > from -mno-sse4 in target("") via ix86_valid_target_attribute_inner_p.
> >
> > OK? Any clarification on what the above is about?
> >
> > Richard.
> >
> > PR target/119549
> > * common/config/i386/i386-common.cc (ix86_handle_option):
> > Handle OPT_mno_sse4 via OPT_msse4 and looking at value.
> >
> > * gcc.target/i386/pr119549.c: New testcase.
> > ---
> > gcc/common/config/i386/i386-common.cc | 21 ++++++++++++---------
> > gcc/testsuite/gcc.target/i386/pr119549.c | 15 +++++++++++++++
> > 2 files changed, 27 insertions(+), 9 deletions(-)
> > create mode 100644 gcc/testsuite/gcc.target/i386/pr119549.c
> >
> > diff --git a/gcc/common/config/i386/i386-common.cc
> > b/gcc/common/config/i386/i386-common.cc
> > index 80aec3244bc..296df3b3230 100644
> > --- a/gcc/common/config/i386/i386-common.cc
> > +++ b/gcc/common/config/i386/i386-common.cc
> > @@ -1519,15 +1519,18 @@ ix86_handle_option (struct gcc_options *opts,
> > return true;
> >
> > case OPT_msse4:
> > - opts->x_ix86_isa_flags |= OPTION_MASK_ISA_SSE4_SET;
> > - opts->x_ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_SET;
> > - return true;
> > -
> > - case OPT_mno_sse4:
> > - opts->x_ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_UNSET;
> > - opts->x_ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_UNSET;
> > - opts->x_ix86_isa_flags2 &= ~OPTION_MASK_ISA2_SSE4_UNSET;
> > - opts->x_ix86_isa_flags2_explicit |= OPTION_MASK_ISA2_SSE4_UNSET;
> > + if (value)
> > + {
> > + opts->x_ix86_isa_flags |= OPTION_MASK_ISA_SSE4_SET;
> > + opts->x_ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_SET;
> > + }
> > + else
> > + {
> > + opts->x_ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_UNSET;
> > + opts->x_ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_UNSET;
> > + opts->x_ix86_isa_flags2 &= ~OPTION_MASK_ISA2_SSE4_UNSET;
> > + opts->x_ix86_isa_flags2_explicit |= OPTION_MASK_ISA2_SSE4_UNSET;
> > + }
> > return true;
> >
> > case OPT_msse4a:
>
> For OPT_msse4 that looks reasonable, but I wonder if the old code shouldn't
> be kept for OPT_mno_sse4. When I just try to compile something with
> -O2 -mavx2 -mno-sse4
> previously this would make it like -O2 -mssse3, now it will act as -O2 -mavx2.
> E.g. compile
> #if __SSSE3__
> int i;
> #endif
> #if __AVX2__
> int j;
> #endif
>
> I wonder what OPT_mno_sse4 with !value would actually mean though (if it
> ever happens).
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
