Re: [autovect][patch]Loop Versioning for Vectorization

[Dorit Naishlos <DORIT at il dot ibm dot com>]

Thanks, Keith.

I think we should introduce a flag that allows controlling whether
versioning takes place or not. We can enable it by default when
vectorization is on, but we want to have a way to disable it, since we
currently apply versioning pretty much blindly (we haven't incorporated a
cost model yet), so we'll be touching a lot more loops, increasing code
size on the way (and other penalties). -ftree-vect-loop-version ?

One small issue is that the testcases need to be updated, as more loops get
vectorized now. I get the following errors with the patch on
i686-pc-linux-gnu:

XPASS: gcc.dg/vect/vect-29.c scan-tree-dump-times vectorized 1 loops 1
XPASS: gcc.dg/vect/vect-44.c scan-tree-dump-times vectorized 1 loops 1
XPASS: gcc.dg/vect/vect-48.c scan-tree-dump-times vectorized 1 loops 1
XPASS: gcc.dg/vect/vect-50.c scan-tree-dump-times vectorized 1 loops 1
XPASS: gcc.dg/vect/vect-52.c scan-tree-dump-times vectorized 1 loops 1
XPASS: gcc.dg/vect/vect-77.c scan-tree-dump-times vectorized 1 loops 1
XPASS: gcc.dg/vect/vect-78.c scan-tree-dump-times vectorized 1 loops 1
XPASS: gcc.dg/vect/vect-80.c scan-tree-dump-times vectorized 1 loops 1
XPASS: gcc.dg/vect/vect-ifcvt-1.c scan-tree-dump-times vectorized 2 loops 1
FAIL: gcc.dg/vect/vect-none.c scan-tree-dump-times vectorized 2 loops 1

Targets that don't support misalignment will have more of those XPASSes.
In mainline there are target key-words that allow easily to set dg-final
for sets of targets (thanks to Janis), so rather than manually going over
all the tests in autovect and making the required modifications, I suggest
to do one (or preferably both) of the following:

1) wait for next merge from mainline, at which point it will be much easier
to do this (I'll try to do the merge this week).

2) using the new flag I mentioned above: use -fno-tree-vect-loop-version in
vect.exp, so that the current testcases maintain their current behavior,
and add new tests for which we'll enable versioining.


> (vectorize_loops): Move call to rewrite_into_loop_closed_ssa so
> it's done each
> time it's needed rather than once at the end.

I hope to have a patch to preserve loop-closed-form during our loop-peeling
within the vectorizer, because we probably don't want to call
rewrite_into_loop_closed_ssa each time we vectorize a loop (using peeling).


thanks,
dorit


gcc-patches-owner@gcc.gnu.org wrote on 22/01/2005 09:21:35:

> Add loop versioning for vectorization.  For example:
>
> for(i=0; i<N; ++i)
> {
>   pa[i] = pb[i] + pc[i];
> }
>
> If the alignment of the pointers pa, pb, and pc cannot be determined at
> compile time
> then a runtime test can be inserted to check alignment and direct control

> to either a
> vector version or a scalar version of the loop.  At a high level it might

> look something
> like:
>
> if ((&pa[0] % AlignConst == 0)
>     && (&pb[0] % AlignConst == 0)
>     && (&pc[0] % AlignConst == 0))
>   {
>     for(i=0; i<N/VectorizationFactor; ++i)
>       {
>         vpa[i] = vpb[i] + vpc[i];
>       }
>   }
> else
>   {
>     for(i=0; i<N; ++i)
>       {
>         pa[i] = pb[i] + pc[i];
>       }
>   }
>
> The constants AlignConst and VectorizationFactor are machine dependent.
> For example,
> if a vector is required to be aligned on a 16 byte boundary then
> AlignConst would be
> 16.  If a vector vpa[i] contains 4 elements of pa[i] then
> VectorizationFactor would be 4.
>
> Tested on ppc
> No differences in make check or in SPEC.
>
> OK for autovect branch?
>
> Keith.
>
>
> Changelog:
>         * tree-vectorizer.c (new_loop_vec_info): Initialize
> LOOP_VINFO_MAY_MISALIGN_STMTS.
>         (destroy_loop_vec_info); varray_clear  of
> LOOP_VINFO_MAY_MISALIGN_STMTS.
>         (vect_create_cond_for_align_checks): New.
>         (vect_transform_loop): Add calls to tree_ssa_loop_version and
>         vect_create_cond_for_align_checks.
>         (vect_build_dist_vector): Change loop_info to loop_vinfo as it is

> called elsewhere.
>         (vect_update_misalignment_for_peel): New.  Holds multiple use
> code.
>         (vect_enhance_data_refs_alignment): Decide when to do loop
> versioning and
>         update data structcures.
>         (vect_analyze_data_refs_alignment): Fix comment.
>         (vect_pattern_recog): Fix comment.
>         (vectorize_loops): Move call to rewrite_into_loop_closed_ssa so
> it's done each
>         time it's needed rather than once at the end.
>         *tree-vectorizer.h(struct _loop_vec_info): Add fields ptr_mask
and
>         may_misalign_stmts.
>         (MAX_RUNTIME_ALIGNMENT_CHECKS): New named constant.
>
> [attachment "vect.1-21.diff" deleted by Dorit Naishlos/Haifa/IBM]