This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[RFA:] Fix PR target/{18329,18330}, incorrect deletion of an outputreload, take 2.


See discussion at <URL:http://gcc.gnu.org/ml/gcc-patches/2005-01/msg00478.html>.

As per RTH's request, I now prune the output-reload-killing if
the lifetime of the pseudo is more than one basic block.  That
means that the first test (for labels and jumps) is redundant,
so I removed it to avoid confusion.  As noted in the comment,
this patch doesn't handle non-call-exceptions.  Doing that
should be a separate patch; it was just noted on inspection.
(I don't know how to create such a a test-case; such a test
would be very target-specific and the MMIX port and simulator,
as most other target-only ports, doesn't support
non-call-exceptions.)

Built and tested native i686-pc-linux-gnu (FC2) and cross to
mmix-knuth-mmixware, no regressions.

Ok to commit?

	PR target/18329
	PR target/18330
	* reload1.c (delete_output_reload): Don't delete an output reload
	if the pseudo lives longer than a single basic block.  Adjust and
	improve wording of comment and add FIXME for non-call-exceptions.

Index: reload1.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/reload1.c,v
retrieving revision 1.458
diff -p -c -r1.458 reload1.c
*** reload1.c	6 Jan 2005 04:09:34 -0000	1.458
--- reload1.c	11 Jan 2005 19:43:17 -0000
*************** delete_output_reload (rtx insn, int j, i
*** 7624,7637 ****

    /* If the pseudo-reg we are reloading is no longer referenced
       anywhere between the store into it and here,
!      and no jumps or labels intervene, then the value can get
!      here through the reload reg alone.
       Otherwise, give up--return.  */
    for (i1 = NEXT_INSN (output_reload_insn);
         i1 != insn; i1 = NEXT_INSN (i1))
      {
!       if (LABEL_P (i1) || JUMP_P (i1))
  	return;
        if ((NONJUMP_INSN_P (i1) || CALL_P (i1))
  	  && reg_mentioned_p (reg, PATTERN (i1)))
  	{
--- 7624,7638 ----

    /* If the pseudo-reg we are reloading is no longer referenced
       anywhere between the store into it and here,
!      and we're within the same basic block, then the value can only
!      pass through the reload reg and end up here.
       Otherwise, give up--return.  */
    for (i1 = NEXT_INSN (output_reload_insn);
         i1 != insn; i1 = NEXT_INSN (i1))
      {
!       if (NOTE_INSN_BASIC_BLOCK_P (i1))
  	return;
+       /* FIXME: non-call exceptions need to be handled.  */
        if ((NONJUMP_INSN_P (i1) || CALL_P (i1))
  	  && reg_mentioned_p (reg, PATTERN (i1)))
  	{
brgds, H-P


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]