[patch, vectorizer] Fix PR tree-optimization/33804 ICE in vect_transform_stmt

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

Hi,

The check whether the vectorization is worthwhile, in case that the
operation is not supported by the target, is performed in function
vectorizable_operation, both during the analysis and the transformation.
The check is based on the loop vectorization factor. We put gcc_assert to
verify that it does not fail during the transformation phase, since the
decision whether to vectorize should be made during the analysis.

The testcase in PR 33804 ICEs, because the vectorization factor was changed
at the end of the analysis, and its comparison with
vect_min_worthwhile_factor gives different results during the analysis and
the transformation.

This patch removes the check that the vectorization is worthwhile from the
transformation phase: the vectorization factor can get smaller only in case
that there is only SLP-kind of vectorization in the loop, and the VF is the
unrolling factor needed to operate on full vectors. So the profitability of
the loop vectorization cannot actually change after the analysis.

Bootstrapped with vectorization enabled and now regtesting on x86_64-linux.
O.K. for mainline once the testing completes?

Thanks,
Ira

ChangeLog:

      PR tree-optimization/33804
      * tree-vect-transform.c (vectorizable_operation): Remove the
      checks that the vectorization is worthwhile from the transformation
      phase.

testsuite/ChangeLog:

      PR tree-optimization/33804
      * gcc.dg/vect/pr33804.c: New testcase.


Index: tree-vect-transform.c
===================================================================
--- tree-vect-transform.c       (revision 129577)
+++ tree-vect-transform.c       (working copy)
@@ -3825,18 +3825,21 @@ vectorizable_operation (tree stmt, block
     {
       if (vect_print_dump_info (REPORT_DETAILS))
        fprintf (vect_dump, "op not supported by target.");
+      /* Check only during analysis.  */
       if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
-          || LOOP_VINFO_VECT_FACTOR (loop_vinfo)
-            < vect_min_worthwhile_factor (code))
+          || (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
+             < vect_min_worthwhile_factor (code)
+             && !vec_stmt))
         return false;
       if (vect_print_dump_info (REPORT_DETAILS))
        fprintf (vect_dump, "proceeding using word mode.");
     }

-  /* Worthwhile without SIMD support?  */
+  /* Worthwhile without SIMD support? Check only during analysis.  */
   if (!VECTOR_MODE_P (TYPE_MODE (vectype))
       && LOOP_VINFO_VECT_FACTOR (loop_vinfo)
-        < vect_min_worthwhile_factor (code))
+        < vect_min_worthwhile_factor (code)
+      && !vec_stmt)
     {
       if (vect_print_dump_info (REPORT_DETAILS))
        fprintf (vect_dump, "not worthwhile without SIMD support.");
Index: testsuite/gcc.dg/vect/pr33804.c
===================================================================
--- testsuite/gcc.dg/vect/pr33804.c     (revision 0)
+++ testsuite/gcc.dg/vect/pr33804.c     (revision 0)
@@ -0,0 +1,17 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target vect_int } */
+
+void f(unsigned char *s, unsigned char *d, int n) {
+    int i;
+    for (i = 0; i < n; i += 4) {
+        d[i + 0] += s[i + 0];
+        d[i + 1] += s[i + 1];
+        d[i + 2] += s[i + 2];
+        d[i + 3] += s[i + 3];
+    }
+}
+
+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect"  } } */
+/* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 1
"vect"  } } */
+/* { dg-final { cleanup-tree-dump "vect" } } */
+