tree-optimization/113431 - wrong dependence with invariant load

Message ID 20240118073705.E43433858289@sourceware.org
State Accepted
Headers
Series tree-optimization/113431 - wrong dependence with invariant load |

Checks

Context Check Description
snail/gcc-patch-check success Github commit url

Commit Message

Richard Biener Jan. 18, 2024, 7:31 a.m. UTC
  The vectorizer dependence analysis is confused with invariant loads
when figuring whether the circumstances are so that we preserve
scalar stmt execution order.  The following rectifies this.

Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.

	PR tree-optimization/113431
	* tree-vect-data-refs.cc (vect_preserves_scalar_order_p):
	When there is an invariant load we might not preserve
	scalar order.

	* gcc.dg/vect/pr113431.c: New testcase.
---
 gcc/testsuite/gcc.dg/vect/pr113431.c | 18 ++++++++++++++++++
 gcc/tree-vect-data-refs.cc           |  6 ++++++
 2 files changed, 24 insertions(+)
 create mode 100644 gcc/testsuite/gcc.dg/vect/pr113431.c
  

Patch

diff --git a/gcc/testsuite/gcc.dg/vect/pr113431.c b/gcc/testsuite/gcc.dg/vect/pr113431.c
new file mode 100644
index 00000000000..04448d9dd81
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr113431.c
@@ -0,0 +1,18 @@ 
+/* { dg-additional-options "-O3 -fdump-tree-slp1-details" } */
+
+#include "tree-vect.h"
+
+int a[2][9];
+int b;
+int main()
+{
+  check_vect ();
+  for (b = 0; b < 2; b++)
+    for (long e = 8; e > 0; e--)
+      a[b][e] = a[0][1] == 0;
+  if (a[1][1] != 0)
+    __builtin_abort ();
+  return 0;
+}
+
+/* { dg-final { scan-tree-dump-times "optimized: basic block part vectorized" 2 "slp1" { target vect_int } } } */
diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
index 0495842b350..f592aeb8028 100644
--- a/gcc/tree-vect-data-refs.cc
+++ b/gcc/tree-vect-data-refs.cc
@@ -282,6 +282,12 @@  vect_preserves_scalar_order_p (dr_vec_info *dr_info_a, dr_vec_info *dr_info_b)
       && !STMT_VINFO_GROUPED_ACCESS (stmtinfo_b))
     return true;
 
+  /* If there is a loop invariant read involved we might vectorize it in
+     the prologue, breaking scalar oder with respect to the in-loop store.  */
+  if ((DR_IS_READ (dr_info_a->dr) && integer_zerop (DR_STEP (dr_info_a->dr)))
+      || (DR_IS_READ (dr_info_b->dr) && integer_zerop (DR_STEP (dr_info_b->dr))))
+    return false;
+
   /* STMT_A and STMT_B belong to overlapping groups.  All loads are
      emitted at the position of the first scalar load.
      Stores in a group are emitted at the position of the last scalar store.