[Bug tree-optimization/46172] [4.3/4.4/4.5/4.6 Regression] ICE: in expand_widen_pattern_expr, at optabs.c:522 with -ftree-vectorize -fno-tree-dce

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46172

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |NEW
         AssignedTo|jakub at gcc dot gnu.org    |unassigned at gcc dot
                   |                            |gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-11-02 12:39:16 UTC ---
The problem is that we create DOT_PLUS_EXPR in vect_recog_dot_prod_pattern and
it is kept, because for vector mode the pattern is supported, but then it is
kept also for the final loop handling last few entries and for scalar modes
there is no sdot_prod* pattern.

During expansion DOT_PROD_EXPR is expanded using expand_widen_pattern_expr
which ICEs if there is no handler for it.

To fix this, either the vectorizer would need to undo DOT_PROD_EXPR in the
scalar code if there is no DOT_PROD_EXPR scalar handler, or expansion should
check if there is a handler for it and if not, expand it manually using
multiplication and addition, or all the 3 targets that have
sdot_prod*/udot_prod* patterns should always support both vector and scalar
patterns.