patch to fix PR63906

patch to fix PR63906

[Vladimir Makarov]

The following patch fixes

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63906

LRA rematerialization checks SP offsets at origin and rematerialization 
places when trying to rematerialize an insn.  But the offsets are not 
valid if sp elimination is prohibited (e.g. when alloca is used).  Value 
of frame_pointer_required is a flag of this.

The patch was bootstrapped on aarch64 and committed as rev. 217683.

2014-11-17  Vladimir Makarov  <vmakarov@redhat.com>

         PR rtl-optimization/63906
         * lra-remat.c (operand_to_remat): Check SP and
         frame_pointer_required.

Index: lra-remat.c
===================================================================
--- lra-remat.c (revision 217658)
+++ lra-remat.c (working copy)
@@ -399,7 +399,13 @@ operand_to_remat (rtx_insn *insn)

    /* First find a pseudo which can be rematerialized.  */
    for (reg = id->regs; reg != NULL; reg = reg->next)
-    if (reg->type == OP_OUT && ! reg->subreg_p
+    /* True FRAME_POINTER_NEEDED might be because we can not follow
+       changing sp offsets, e.g. alloca is used.  If the insn contains
+       stack pointer in such case, we can not rematerialize it as we
+       can not know sp offset at a rematerialization place.  */
+    if (reg->regno == STACK_POINTER_REGNUM && frame_pointer_needed)
+      return -1;
+    else if (reg->type == OP_OUT && ! reg->subreg_p
         && find_regno_note (insn, REG_UNUSED, reg->regno) == NULL)
        {
         /* We permits only one spilled reg.  */

Re: patch to fix PR63906

[Andrew Pinski]

On Mon, Nov 17, 2014 at 4:15 PM, Vladimir Makarov <vmakarov@redhat.com> wrote:
> The following patch fixes
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63906
>
> LRA rematerialization checks SP offsets at origin and rematerialization
> places when trying to rematerialize an insn.  But the offsets are not valid
> if sp elimination is prohibited (e.g. when alloca is used).  Value of
> frame_pointer_required is a flag of this.


Thanks for the fix, it works and glibc is no longer miscompiled.

Thanks,
Andrew

>
> The patch was bootstrapped on aarch64 and committed as rev. 217683.
>
> 2014-11-17  Vladimir Makarov  <vmakarov@redhat.com>
>
>         PR rtl-optimization/63906
>         * lra-remat.c (operand_to_remat): Check SP and
>         frame_pointer_required.
>
> Index: lra-remat.c
> ===================================================================
> --- lra-remat.c (revision 217658)
> +++ lra-remat.c (working copy)
> @@ -399,7 +399,13 @@ operand_to_remat (rtx_insn *insn)
>
>    /* First find a pseudo which can be rematerialized.  */
>    for (reg = id->regs; reg != NULL; reg = reg->next)
> -    if (reg->type == OP_OUT && ! reg->subreg_p
> +    /* True FRAME_POINTER_NEEDED might be because we can not follow
> +       changing sp offsets, e.g. alloca is used.  If the insn contains
> +       stack pointer in such case, we can not rematerialize it as we
> +       can not know sp offset at a rematerialization place.  */
> +    if (reg->regno == STACK_POINTER_REGNUM && frame_pointer_needed)
> +      return -1;
> +    else if (reg->type == OP_OUT && ! reg->subreg_p
>         && find_regno_note (insn, REG_UNUSED, reg->regno) == NULL)
>        {
>         /* We permits only one spilled reg.  */