On Wed, Nov 16, 2016 at 09:14:57PM -0600, Bill Schmidt wrote:
> 2016-11-16 Bill Schmidt <wschm...@linux.vnet.ibm.com>
> Richard Biener <rguent...@suse.de>
>
> PR tree-optimization/77848
> * tree-if-conv.c (tree_if_conversion): Always version loops unless
> the user specified -ftree-loop-if-convert.
This broke the attached testcase.
> --- gcc/tree-if-conv.c (revision 242521)
> +++ gcc/tree-if-conv.c (working copy)
> @@ -2803,10 +2803,12 @@ tree_if_conversion (struct loop *loop)
> || loop->dont_vectorize))
> goto cleanup;
>
> - /* Either version this loop, or if the pattern is right for outer-loop
> - vectorization, version the outer loop. In the latter case we will
> - still if-convert the original inner loop. */
> - if ((any_pred_load_store || any_complicated_phi)
> + /* Since we have no cost model, always version loops unless the user
> + specified -ftree-loop-if-convert. Either version this loop, or if
> + the pattern is right for outer-loop vectorization, version the
> + outer loop. In the latter case we will still if-convert the
> + original inner loop. */
> + if (flag_tree_loop_if_convert != 1
> && !version_loop_for_if_conversion
> (versionable_outer_loop_p (loop_outer (loop))
> ? loop_outer (loop) : loop))
If there are masked loads/stores (and I assume also the complicated phi
stuff, but haven't verified), then it isn't just some kind of optimization
to version the loop based on LOOP_VECTORIZED ifn, it is a requirement
- MASK_LOAD/MASK_STORE aren't supported for scalar code, so they can only
appear in the vectorized version. Fixed by reverting that - if
-ftree-loop-if-convert we'll do what we used to do before, without it
we do what you've added, i.e. version always.
The rest is just formatting fix, the too large argument that forces
call's ( on the next line and even misindented is IMHO much cleaner
if a temporary is used.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2016-11-21 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/78445
* tree-if-conv.c (tree_if_conversion): If any_pred_load_store or
any_complicated_phi, version loop even if flag_tree_loop_if_convert is
1. Formatting fix.
* gcc.dg/pr78445.c: New test.
--- gcc/tree-if-conv.c.jj 2016-11-17 18:08:12.000000000 +0100
+++ gcc/tree-if-conv.c 2016-11-21 17:28:30.807242395 +0100
@@ -2804,15 +2804,20 @@ tree_if_conversion (struct loop *loop)
goto cleanup;
/* Since we have no cost model, always version loops unless the user
- specified -ftree-loop-if-convert. Either version this loop, or if
- the pattern is right for outer-loop vectorization, version the
- outer loop. In the latter case we will still if-convert the
- original inner loop. */
- if (flag_tree_loop_if_convert != 1
- && !version_loop_for_if_conversion
- (versionable_outer_loop_p (loop_outer (loop))
- ? loop_outer (loop) : loop))
- goto cleanup;
+ specified -ftree-loop-if-convert or unless versioning is required.
+ Either version this loop, or if the pattern is right for outer-loop
+ vectorization, version the outer loop. In the latter case we will
+ still if-convert the original inner loop. */
+ if (any_pred_load_store
+ || any_complicated_phi
+ || flag_tree_loop_if_convert != 1)
+ {
+ struct loop *vloop
+ = (versionable_outer_loop_p (loop_outer (loop))
+ ? loop_outer (loop) : loop);
+ if (!version_loop_for_if_conversion (vloop))
+ goto cleanup;
+ }
/* Now all statements are if-convertible. Combine all the basic
blocks into one huge basic block doing the if-conversion
--- gcc/testsuite/gcc.dg/pr78445.c.jj 2016-11-21 17:30:58.534400256 +0100
+++ gcc/testsuite/gcc.dg/pr78445.c 2016-11-21 17:30:41.000000000 +0100
@@ -0,0 +1,19 @@
+/* PR tree-optimization/78445 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-if-convert -ftree-vectorize" } */
+/* { dg-additional-options "-mavx2" { target { i?86-*-* x86_64-*-* } } } */
+
+int a;
+
+void
+foo (int x, int *y)
+{
+ while (a != 0)
+ if (x != 0)
+ {
+ *y = a;
+ x = *y;
+ }
+ else
+ x = a;
+}
Jakub
