On Mon, May 07, 2007 at 10:12:21PM +0200, Paolo Bonzini wrote:
> Eric Botcazou wrote:
> >> It isn't until the loop2_done dump that the problem shows up.
> >>delete_trivially_dead_insns() is called. It calls dead_libcall_p() on insn
> >>62 and seeing both the REQ_EQUAL and the REG_RETVAL note and successfully
> >>replacing reg 92 with const_int 0, determines that the while libcall
> >>sequence 63-152-61-62 is dead and can be deleted, without ever checking
> >>insns 63, 152 or 61 for liveness.
> >
> >Because it's the semantics of libcall sequences. My take is that the
> >lower subreg pass breaks it in this case.
I could "fix" it at -O2 with -fno-split-wide-types or at -O1 with
-fno-move-loop-invariants or -fno-split-wide-types.
> Can you also check -fno-forward-propagate?
Yes. It'll have to wait until Wednesday, though.
To be able to continue testing, I worked around it like this:
Index: gcc/cse.c
===================================================================
--- gcc/cse.c (revision 124334)
+++ gcc/cse.c (working copy)
@@ -4983,8 +4983,11 @@ cse_insn (rtx insn, rtx libcall_insn)
&& GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
&& GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
{
- /* We only want a REG_EQUAL note if src_const != src. */
- if (! rtx_equal_p (src, src_const))
+ /* We only want a REG_EQUAL note if src_const != src. Don't add one
+ if the insn has a REG_RETVAL note since this will confuse
+ delete_trivially_dead_insns() (see dead_libcall_p()). */
+ if (!rtx_equal_p (src, src_const)
+ && !find_reg_note (insn, REG_RETVAL, NULL))
{
/* Make sure that the rtx is not shared. */
src_const = copy_rtx (src_const);
--
Rask Ingemann Lambertsen
