This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix PR22002, death notes lost in combine
On Tue, Nov 08, 2005 at 08:49:55PM -0800, Ian Lance Taylor wrote:
> Alan Modra <amodra@bigpond.net.au> writes:
>
> > PR rtl-optimization/22002
> > * combine.c (distribute_notes): Detect cases where a reg dies
> > two or more times in a bb, including on the insn we are combining,
> > and place the death note on the correct range.
>
> This is OK.
I committed a slight variation on the comment, and added another one
before the loop as follows.
Index: gcc/combine.c
===================================================================
--- gcc/combine.c (revision 106906)
+++ gcc/combine.c (working copy)
@@ -12232,6 +12232,10 @@ distribute_notes (rtx notes, rtx from_in
{
basic_block bb = this_basic_block;
+ /* You might think you could search back from FROM_INSN
+ rather than from I3, but combine tries to split invalid
+ combined instructions. This can result in the old I2
+ or I1 moving later in the insn sequence. */
for (tem = PREV_INSN (i3); place == 0; tem = PREV_INSN (tem))
{
if (! INSN_P (tem))
@@ -12332,6 +12336,22 @@ distribute_notes (rtx notes, rtx from_in
|| (CALL_P (tem)
&& find_reg_fusage (tem, USE, XEXP (note, 0))))
{
+ /* This may not be the correct place for the death
+ note if FROM_INSN is before TEM, and the reg is
+ set between FROM_INSN and TEM. The reg might
+ die two or more times. An existing death note
+ means we are looking at the wrong live range. */
+ if (from_insn
+ && INSN_CUID (from_insn) < INSN_CUID (tem)
+ && find_regno_note (tem, REG_DEAD,
+ REGNO (XEXP (note, 0))))
+ {
+ tem = from_insn;
+ if (tem == BB_HEAD (bb))
+ break;
+ continue;
+ }
+
place = tem;
/* If we are doing a 3->2 combination, and we have a
--
Alan Modra
IBM OzLabs - Linux Technology Centre