[Bug tree-optimization/49352] [4.7 Regression] -fcompare-debug failure with -O2 -ftree-vectorize

["jakub at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49352

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-06-10 11:44:03 UTC ---
Created attachment 24486
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24486
gcc47-pr49352.patch

Rough patch I've been playing with.  There is no point testing gimple_code
of use_stmt if it is GIMPLE_PHI, because phi is known to be a PHI.  The patch
attempts to only allow exactly one loop use stmt and use that stmt.  I had to
drop the use_stmt != first_stmt test, otherwise the transformation wouldn't
ever happen, not sure if you want to allow in that case another loop use (if
yes, which one of the two do you want as loop_use_stmt?).

I also wonder if vect_is_slp_reduction should just return false; early if code
is COND_EXPR, if for COND_EXPR def1 is not equal to phi I'm afraid it could be
called.