http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50693
Richard Guenther <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |aoliva at gcc dot gnu.org,
| |jakub at gcc dot gnu.org
--- Comment #18 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-10-11
14:42:46 UTC ---
(In reply to comment #15)
> What probably causes this is that we don't merge the blocks
>
> # i_29 = PHI <i_10(3), 0(2)>
> copy_block:
> D.3506_7 = v_20->chars;
> D.3507_8 = D.3506_7 + i_29;
> *D.3507_8 = c_9(D);
> i_10 = i_29 + 1;
> goto <bb 3> (loop_back);
>
> loop_back:
> loop_cond_11 = i_10 < n_3(D);
> if (loop_cond_11 != 0)
> goto <bb 4> (copy_block);
> else
> goto <bb 5> (end);
>
> even though it's a fallthru edge. We don't do this to preserve user labels
> for debugging (and mind, no code-gen differences between -g0 vs. -g).
Yes. With
Index: gcc/tree-cfg.c
===================================================================
--- gcc/tree-cfg.c (revision 179804)
+++ gcc/tree-cfg.c (working copy)
@@ -1456,7 +1460,7 @@ gimple_can_merge_blocks_p (basic_block a
/* Do not remove user labels. */
if (!DECL_ARTIFICIAL (lab))
- return false;
+ ;
}
/* Protect the loop latches. */
we merge the blocks and vectorize both loops.
The above patch is not acceptable though, which leaves making the vectorizer
deal with trivial control-flow (or a new GIMPLE_DEBUG kind that would
preserve the label?).
