On Mon, 31 Mar 2025, Jakub Jelinek wrote:
> On Mon, Mar 31, 2025 at 03:33:34PM +0200, Richard Biener wrote:
> > 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
>
> But
> i386-common.cc:#define OPTION_MASK_ISA_SSE4_SET OPTION_MASK_ISA_SSE4_2_SET
> i386-common.cc:#define OPTION_MASK_ISA_SSE4_UNSET OPTION_MASK_ISA_SSE4_1_UNSET
> i386-common.cc:#define OPTION_MASK_ISA2_SSE4_UNSET
> OPTION_MASK_ISA2_SSE4_1_UNSET
> which is exactly what it wants.
>
> > that ix86_handle_option also unsets flags in ix86_isa_flags2
> > with -mno-sse4 but the mno-sse4 options entry does not indicate
> > that.
>
> It does, that is the
> opts->x_ix86_isa_flags2 &= ~OPTION_MASK_ISA2_SSE4_UNSET;
> opts->x_ix86_isa_flags2_explicit |= OPTION_MASK_ISA2_SSE4_UNSET;
> There is no OPTION_MASK_ISA2_SSE4_SET (like there is no
> OPTION_MASK_ISA2_SSE{,[234],4_{1,2}}_SET, that isn't needed because
> all the requirements of SSE4 still fit into ix86_isa_flags.
Yes, all I say is that ix86_handle_option looks OK, I'm not sure
why we need the
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.
options entry. Or whether that means ix86_valid_target_attribute_inner_p
should have created OPT_mno_sse4 from -mno-sse4. At least, having
both i386.opt entries looks like unnecessary complication to me.
And maybe I broke -mno-sse4 on the command-line with my change. Indeed
> ./xgcc -B. t.c -mno-sse4
hits the code with OPT_mno_sse4. I'll update the patch to leave
the old OPT_mno_sse4 handling in place. The other option is
to fixup ix86_valid_target_attribute_inner_p to when finding
-mno-sse4 to exactly special case that and use OPT_mno_sse4?
Again, I don't really understand why it was done in this complicated
way. At least there seems to be some test coverage.
Richard.
