Re: ifcvt defect (was Re: PATCH: Debug support for bb-reorder)

[Richard Henderson <rth at cygnus dot com>]

On Mon, May 01, 2000 at 10:20:21PM -0700, Jason Eckhardt wrote:
>   I believe I've located the problem in ifcvt.c:dead_or_predicable(). I've
>   attached a patch for the defect.
>   However, I found this particular function a bit hard to follow, so
>   I'm not 100% confident of the patch.

I prefer this fix.

I'm not sure how to make that function easier to follow... perhaps if
we had standard functions for moving insns between blocks, so all the
head/end tracking is not inline?

> 	* flow.c (verify_flow_info): Added two more sanity checks. The
> 	first checks that the blocks are numbered consecutively. The second
> 	checks that n_basic_blocks is actually equal to the number of
> 	basic blocks in the insn chain.

Ok.


r~


        * ifcvt.c (dead_or_predicable): Set merge_bb->end to the insn before
        the sequence we're moving, not to merge_bb->head.

Index: ifcvt.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/ifcvt.c,v
retrieving revision 1.6
diff -c -p -d -r1.6 ifcvt.c
*** ifcvt.c	2000/05/01 22:25:54	1.6
--- ifcvt.c	2000/05/02 06:33:19
*************** dead_or_predicable (test_bb, merge_bb, o
*** 1935,1944 ****
      }
  
    /* Move the insns out of MERGE_BB to before the branch.  */
-   if (end == merge_bb->end)
-     merge_bb->end = merge_bb->head;
    if (head != NULL)
      {
        head = squeeze_notes (head, end);
        if (GET_CODE (end) == NOTE
  	  && (NOTE_LINE_NUMBER (end) == NOTE_INSN_BLOCK_END
--- 1935,1945 ----
      }
  
    /* Move the insns out of MERGE_BB to before the branch.  */
    if (head != NULL)
      {
+       if (end == merge_bb->end)
+ 	merge_bb->end = PREV_INSN (head);
+ 
        head = squeeze_notes (head, end);
        if (GET_CODE (end) == NOTE
  	  && (NOTE_LINE_NUMBER (end) == NOTE_INSN_BLOCK_END