Thanks! I am trying to re-write by calling __builtin_clear_padding. But I
found gimple_fold_builtin_clear_padding seems only work before SSA pass.
Should I remove the assertion?

On the other hand, since ATOMIC_COMPARE_EXCHANGE will only work for simple
reg type. excluding vector or complex types, is it enough to generate the
mask by simple bit operations?

Thanks!,
xndcn

---
static bool
gimple_fold_builtin_clear_padding (gimple_stmt_iterator *gsi)
{ ...
  /* This should be folded during the lower pass.  */
  gcc_assert (!gimple_in_ssa_p (cfun) && cfun->cfg == NULL);


Jakub Jelinek <jakub@redhat.com> 于2024年1月3日周三 23:52写道：

> On Wed, Jan 03, 2024 at 11:42:58PM +0800, xndcn wrote:
> > Hi, I am new to this, and I really need your advice, thanks.
> >
> > I noticed PR71716 and I want to enable ATOMIC_COMPARE_EXCHANGE
> > internal-fn optimization
> >
> > for floating type or types contains padding (e.g., long double).
> > Please correct me if I happen to
> > make any mistakes, Thanks!
> >
> > Firstly, about the concerns of sNaNs float/doduble value, it seems
> > work well and shall have been
> > covered by testsuite/gcc.dg/atomic/c11-atomic-exec-5.c
> >
> > Secondly, since ATOMIC_COMPARE_EXCHANGE is only enabled when expected
> > var is only addressable
> > because of the call, the padding bits can not be modified by any other
> > stmts. So we can save all
> > bits after ATOMIC_COMPARE_EXCHANGE call and extract the padding bits.
> > After first iteration, the
> > extracted padding bits can be mixed with the expected var.
> >
> > Bootstrapped/regtested on x86_64-linux.
> >
> > I did some benchmarks, and there is some significant time optimization
> > for float/double types,
> >
> > while there is no regression for long double type.
>
> If anything, this should be using clear_padding_type_may_have_padding_p
> and call to __builtin_clear_padding.
> Code in that file uses /* ... */ style comments, so please use them instead
> of // for consistency, and furthermore comments should be terminated with
> a dot and two spaces before */
>
> Also, I don't think this is late stage3 material, so needs to wait for GCC
> 15.
>
>         Jakub
>
>