https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91528
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Uroš Bizjak from comment #3)
> (In reply to Richard Biener from comment #1)
> > (gdb) p x_rtl.drap_reg
> > $1 = (rtx) 0x0
> >
> > so
> >
> > 7843 /* Only need to push parameter pointer reg if it is caller
> > saved. */
> > 7844 if (!call_used_regs[REGNO (crtl->drap_reg)])
> > 7845 {
> >
> > segfaults. This must be really a latent issue. I guess
> >
> > /* Conversion means we may have 128bit register spills/fills
> > which require aligned stack. */
> > if (converted_insns)
> > {
> > if (crtl->stack_alignment_needed < 128)
> > ...
> >
> > needs to do some magic for -mforce-drap (which might be handled too early,
> > ignoring the late generated xmm uses?)
>
> When changing stack_alignment_{needed,estimated}, we also need to update
> dependent crtl variables, similar to what expand_stack_alignment from
> cfgexpand.c does. Following (untested) patch fixes the testcase failure:
>
> --cut here--
> diff --git a/gcc/config/i386/i386-features.c
> b/gcc/config/i386/i386-features.c
> index fb7ac1b7d102..594b572454df 100644
> --- a/gcc/config/i386/i386-features.c
> +++ b/gcc/config/i386/i386-features.c
> @@ -1770,6 +1770,28 @@ convert_scalars_to_vector (bool timode_p)
> crtl->stack_alignment_needed = 128;
> if (crtl->stack_alignment_estimated < 128)
> crtl->stack_alignment_estimated = 128;
> +
> + crtl->stack_realign_needed
> + = INCOMING_STACK_BOUNDARY < crtl->stack_alignment_estimated;
> + crtl->stack_realign_tried = crtl->stack_realign_needed;
> +
> + crtl->stack_realign_processed = true;
> +
> + rtx drap_rtx = targetm.calls.get_drap_rtx ();
> +
> + /* stack_realign_drap and drap_rtx must match. */
> + gcc_assert ((stack_realign_drap != 0) == (drap_rtx != NULL));
> +
> + /* Do nothing if NULL is returned, which means DRAP is not needed. */
> + if (drap_rtx != NULL)
> + {
> + crtl->args.internal_arg_pointer = drap_rtx;
> +
> + /* Call fixup_tail_calls to clean up REG_EQUIV note if DRAP is
> + needed. */
> + fixup_tail_calls ();
> + }
> +
> /* Fix up DECL_RTL/DECL_INCOMING_RTL of arguments. */
> if (TARGET_64BIT)
> for (tree parm = DECL_ARGUMENTS (current_function_decl);
> --cut here--
>
> HJ, can you please take the patch from here? Realignment stuff is a bit of
> mistery to me.
Also fixes all libada issues I ran into when bootstrapping with
--with-arch=westmere and cutting off the cost-model to always consider
chains profitable to convert.
