[autovect] Make omega agree with existing dependence tester

[Sebastian Pop]

On 3/22/06, Sebastian Pop <sebpop@gmail.com> wrote:
> On 3/22/06, Sebastian Pop <sebpop@gmail.com> wrote:
> > The patch bootstraps on amd64-linux and produces 2 fails that I think are not
> > important because I verified that the distance vectors are correct, and the
> > corresponding cases are vectorized, but it is safer to ask Dorit to
> > look at these:
> > FAIL: gcc.dg/vect/vect-104.c scan-tree-dump-times possible dependence
> > between data-refs 1
> > FAIL: gcc.dg/vect/vect-99.c scan-tree-dump-times dependence distance
> > modulo vf == 0 1
> >
>
> I have tracked down these fails and it seems that my patch was not complete
> as I missed to modify the uses of the data dep analyzer in the vectorizer.
> I'm working on a patch to correct these fails.
>
> Sebastian
>

Here is a patch that corrects these two fails.  Bootstrapped on amd64-linux
and tested with RUNTESTFLAGS=vect.exp
Committed to autovect branch.

Sebastian

	* tree-data-ref.c (index_in_loop_nest): Moved...
	* tree-data-ref.h: Remove duplicate definitions of some macros.
	(index_in_loop_nest): ...here.
	* tree-vect-analyze.c (vect_analyze_data_ref_dependence): Use
	DDR_LOOP_NEST and index_in_loop_nest to determine the dependence
	distance.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 527_av_vect.diff
Type: text/x-patch
Size: 3902 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20060323/78ce9587/attachment.bin>