openmp: Fix up handling of non-rectangular simd loops with pointer type iterators...

There were 2 issues visible on this new testcase, one that we didn't have
special POINTER_TYPE_P handling in a few spots of expand_omp_simd - for
pointers we need to use POINTER_PLUS_EXPR and need to have the non-pointer
part in sizetype, for non-rectangular loop on the other side we can rely on
multiplication factor 1, pointers can't be multiplied, without those changes
we'd ICE.  The other issue was that we put n2 expression directly into a
comparison in a condition and regimplified that, for the &a[512] case that
and with gimplification being destructed that unfortunately meant modification
of original fd->loops[?].n2.  Fixed by unsharing the expression.  This was
causing a runtime failure on the testcase.

2022-07-29  Jakub Jelinek  <jakub@redhat.com>

	PR middle-end/106449
	* omp-expand.cc (expand_omp_simd): Fix up handling of pointer
	iterators in non-rectangular simd loops.  Unshare fd->loops[i].n2
	or n2 before regimplifying it inside of a condition.

	* testsuite/libgomp.c-c++-common/pr106449.c: New test.

diff --git a/gcc/omp-expand.cc b/gcc/omp-expand.cc
index 9cc61fed9f89de0fe246642e0445ef1eea6a9ba0..936adff7f45ff85f3c2277fa5b4c866e92b71f3d 100644 (file)
--- a/gcc/omp-expand.cc
+++ b/gcc/omp-expand.cc
@@ -6696,7 +6696,7 @@ expand_omp_simd (struct omp_region *region, struct omp_for_data *fd)
       if (fd->loops[i].m2)
        t = n2v = create_tmp_var (itype);
       else
-       t = fold_convert (itype, fd->loops[i].n2);
+       t = fold_convert (itype, unshare_expr (fd->loops[i].n2));
       t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
                                    false, GSI_CONTINUE_LINKING);
       tree v = fd->loops[i].v;
@@ -6710,7 +6710,7 @@ expand_omp_simd (struct omp_region *region, struct omp_for_data *fd)
       if (fd->collapse > 1 && !broken_loop)
        t = n2var;
       else
-       t = fold_convert (type, n2);
+       t = fold_convert (type, unshare_expr (n2));
       t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
                                    false, GSI_CONTINUE_LINKING);
       tree v = fd->loop.v;
@@ -6819,7 +6819,7 @@ expand_omp_simd (struct omp_region *region, struct omp_for_data *fd)
          if (fd->loops[i].m2)
            t = nextn2v = create_tmp_var (itype);
          else
-           t = fold_convert (itype, fd->loops[i].n2);
+           t = fold_convert (itype, unshare_expr (fd->loops[i].n2));
          t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
                                        false, GSI_CONTINUE_LINKING);
          tree v = fd->loops[i].v;
@@ -6849,17 +6849,25 @@ expand_omp_simd (struct omp_region *region, struct omp_for_data *fd)
          ne->probability = e->probability.invert ();
 
          gsi = gsi_after_labels (init_bb);
-         t = fold_convert (TREE_TYPE (fd->loops[i + 1].v),
-                           fd->loops[i + 1].n1);
          if (fd->loops[i + 1].m1)
            {
-             tree t2 = fold_convert (TREE_TYPE (t),
+             tree t2 = fold_convert (TREE_TYPE (fd->loops[i + 1].v),
                                      fd->loops[i + 1
                                                - fd->loops[i + 1].outer].v);
-             tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i + 1].m1);
-             t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3);
-             t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2);
+             if (POINTER_TYPE_P (TREE_TYPE (t2)))
+               t = fold_build_pointer_plus (t2, fd->loops[i + 1].n1);
+             else
+               {
+                 t = fold_convert (TREE_TYPE (fd->loops[i + 1].v),
+                                   fd->loops[i + 1].n1);
+                 tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i + 1].m1);
+                 t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3);
+                 t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2);
+               }
            }
+         else
+           t = fold_convert (TREE_TYPE (fd->loops[i + 1].v),
+                             fd->loops[i + 1].n1);
          expand_omp_build_assign (&gsi, fd->loops[i + 1].v, t);
          if (fd->loops[i + 1].m2)
            {
@@ -6868,14 +6876,19 @@ expand_omp_simd (struct omp_region *region, struct omp_for_data *fd)
                  gcc_assert (n2v == NULL_TREE);
                  n2v = create_tmp_var (TREE_TYPE (fd->loops[i + 1].v));
                }
-             t = fold_convert (TREE_TYPE (fd->loops[i + 1].v),
-                               fd->loops[i + 1].n2);
-             tree t2 = fold_convert (TREE_TYPE (t),
+             tree t2 = fold_convert (TREE_TYPE (fd->loops[i + 1].v),
                                      fd->loops[i + 1
                                                - fd->loops[i + 1].outer].v);
-             tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i + 1].m2);
-             t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3);
-             t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2);
+             if (POINTER_TYPE_P (TREE_TYPE (t2)))
+               t = fold_build_pointer_plus (t2, fd->loops[i + 1].n2);
+             else
+               {
+                 t = fold_convert (TREE_TYPE (fd->loops[i + 1].v),
+                                   fd->loops[i + 1].n2);
+                 tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i + 1].m2);
+                 t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3);
+                 t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2);
+               }
              expand_omp_build_assign (&gsi, n2v, t);
            }
          if (i + 2 == fd->collapse && n2var)
@@ -6891,17 +6904,25 @@ expand_omp_simd (struct omp_region *region, struct omp_for_data *fd)
              tree t2 = fold_build2 (MINUS_EXPR, type, n2, fd->loop.v);
              if (fd->loops[i + 1].m1 || fd->loops[i + 1].m2)
                {
+                 tree itype = TREE_TYPE (fd->loops[i].v);
+                 if (POINTER_TYPE_P (itype))
+                   itype = signed_type_for (itype);
                  t = build_int_cst (itype, (fd->loops[i + 1].cond_code
                                             == LT_EXPR ? -1 : 1));
                  t = fold_build2 (PLUS_EXPR, itype,
                                   fold_convert (itype,
                                                 fd->loops[i + 1].step), t);
-                 if (fd->loops[i + 1].m2)
-                   t = fold_build2 (PLUS_EXPR, itype, t, n2v);
-                 else
+                 if (fd->loops[i + 1].m2 == NULL_TREE)
                    t = fold_build2 (PLUS_EXPR, itype, t,
                                     fold_convert (itype,
                                                   fd->loops[i + 1].n2));
+                 else if (POINTER_TYPE_P (TREE_TYPE (n2v)))
+                   {
+                     t = fold_build_pointer_plus (n2v, t);
+                     t = fold_convert (itype, t);
+                   }
+                 else
+                   t = fold_build2 (PLUS_EXPR, itype, t, n2v);
                  t = fold_build2 (MINUS_EXPR, itype, t,
                                   fold_convert (itype, fd->loops[i + 1].v));
                  tree step = fold_convert (itype, fd->loops[i + 1].step);

diff --git a/libgomp/testsuite/libgomp.c-c++-common/pr106449.c b/libgomp/testsuite/libgomp.c-c++-common/pr106449.c
new file mode 100644 (file)
index 0000000..ea3cd82
--- /dev/null
+++ b/libgomp/testsuite/libgomp.c-c++-common/pr106449.c
@@ -0,0 +1,62 @@
+/* PR middle-end/106449 */
+/* { dg-do run } */
+
+void
+foo (void)
+{
+  int a[1024], *b[65536];
+  int *p, *q, **r = &b[0], i;
+  #pragma omp simd collapse(2) linear(r : 2)
+  for (p = &a[0]; p < &a[512]; p++)
+    for (q = p + 64; q < p + 128; q++)
+      {
+        *r++ = p;
+        *r++ = q;
+      }
+  for (i = 0; i < 32768; i++)
+    if (b[2 * i] != &a[i / 64] || b[2 * i + 1] != &a[(i / 64) + 64 + (i % 64)])
+      __builtin_abort ();
+}
+
+void
+bar (int n, int m)
+{
+  int a[1024], *b[65536];
+  int *p, *q, **r = &b[0], i;
+  #pragma omp parallel for simd collapse(2) linear(r : 2)
+  for (p = &a[0]; p < &a[512]; p++)
+    for (q = p + n; q < p + m; q++)
+      {
+        *r++ = p;
+        *r++ = q;
+      }
+  for (i = 0; i < 32768; i++)
+    if (b[2 * i] != &a[i / 64] || b[2 * i + 1] != &a[(i / 64) + 64 + (i % 64)])
+      __builtin_abort ();
+}
+
+void
+baz (int n, int m)
+{
+  int a[1024], *b[8192];
+  int *p, *q, **r = &b[0], i;
+  #pragma omp parallel for simd collapse(2) linear(r : 2)
+  for (p = &a[0]; p < &a[512]; p += 4)
+    for (q = p + n; q < p + m; q += 2)
+      {
+        *r++ = p;
+        *r++ = q;
+      }
+  for (i = 0; i < 4096; i++)
+    if (b[2 * i] != &a[(i / 32) * 4] || b[2 * i + 1] != &a[(i / 32) * 4 + 64 + (i % 32) * 2])
+      __builtin_abort ();
+}
+
+int
+main ()
+{
+  foo ();
+  bar (64, 128);
+  baz (64, 128);
+  return 0;
+}