On Tue, Mar 8, 2022 at 7:46 AM Jakub Jelinek <ja...@redhat.com> wrote:
>
> On Tue, Mar 08, 2022 at 07:37:17AM -0800, H.J. Lu wrote:
> > > Though, perhaps it should be
> > > #ifndef __x86_64__
> > > #define LIBGCC2_UNWIND_ATTRIBUTE __attribute__((target ("no-sse")))
> > > #endif
> > > or something similar, on x86-64 one at least normally doesn't use lower
> > > stack realignment unless avx or later.  Maybe we want to use
> > > no-avx for the x86-64 case though.
> >
> > I have verified that AVX and AVX512 have no issues on x86-64.  In 32-bit,
> > -mstackrealign triggers the problem.
>
> I bet it would be a problem if we started vectorizing something in there
> using avx/avx2/avx512*.  But given the sorry, I think we'd find that out

YMM and ZMM can be used to expand memset with -march=native.
It works fine on Linux.  No stack realignment is needed.

> immediately.

True.

> > > Disabling sse/sse2 might be a problem especially on mingw where we need to
> > > restore SSE registers in the EH return, no?
> >
> > No, it isn't needed.
>
> I meant for 64-bit where I think the Windows ABI preserves some XMM regs

Does it need to realign the stack?

> (low 128-bits of them).  So my earlier patch to just define
> LIBGCC2_UNWIND_ATTRIBUTE unconditionally would be wrong for it.
>
> > > Even better would be to make __builtin_eh_return work even with DRAP,
> > > but I admit I haven't understood what exactly is the problem that prevents
> > > it from working.
> >
> > The EH return is a very special case.  Disable SSE in 32-bit is the simplest
> > way to make the EH return to work.
>
> Ok.  So, what do you think about replacing the libgcc/ part of your patch
> with that
> /* __builtin_eh_return can't handle stack realignment, so disable SSE in
>    32-bit libgcc functions that call it.  */
> #ifndef __x86_64__
> #define LIBGCC2_UNWIND_ATTRIBUTE __attribute__((target ("no-sse")))
> #endif
> ?

Yes, it should work.

Thanks.

> I'm bootstrapping/regtesting such a patch right now (because I needed some
> quick fix for the gnat1 hangs).
>
>         Jakub
>


-- 
H.J.

Reply via email to