[patch] Cleanup in vectorization of strided accesses

[Ira Rosen <IRAR at il dot ibm dot com>]

This is a cleanup patch. It removes a redundant pass through all the ddrs,
and replaces it with a field that is set during already existing data
dependencies check.  There's no real change in the behavior of the
vectorizer.

The removed pass checked whether there are load-store and store-store
dependencies in the loop. If such dependencies were found, strided access
analysis was not called. While the actual problem occurs only when there
are two statements in the loop that access the exact same location
(dependence distance of 0), since in all other cases our analysis prevents
vectorization anyway. Therefore, the additional pass is redundant.

With this patch in case there are a load and a store that access the same
memory location, we mark the load  in the _stmt_vec_info structure. The
case of two stores to the same location is now handled in
vect_analyze_data_ref_access.

Testcase is included.
Bootstrapped and tested on ppc-linix.

O.K. for mainline?

Thanks,
Ira

ChangeLog entry:

      * tree-vectorizer.h (struct _stmt_vec_info): Add new field
read_write_dep and macros for its access.
      * tree-vectorizer.c (new_stmt_vec_info): Initialize the new field.
      * tree-vect-analyze.c (vect_analyze_data_ref_dependence): Remove
argument, call
      vect_check_interleaving for every independent pair of data-refs. Mark
loads that access the same
      memory location as a store in the loop.
      (vect_check_dependences): Remove.
      (vect_analyze_data_ref_dependences): Remove  vect_check_dependences
call, fix the call to
      vect_analyze_data_ref_dependence.
      (vect_analyze_data_ref_access): For statements that access the same
data-ref, check that they are
      not stores; for loads, check that there is no store that access the
same location.

The patch:
(See attached file: same_dr.txt)

:ADDPATCH SSA (vectorizer):

Attachment: same_dr.txt
Description: Text document