This fixes the case of not using the multithreaded model when
only conditionally storing to the destination. We cannot elide
the load in this case.
Bootstrap / regtest running on x86_64-unknown-linux-gnu.
2020-05-05 Richard Biener <rguent...@suse.de>
PR tree-optimization/94949
* tree-ssa-loop-im.c (execute_sm): Check whether we use
the multithreaded model or always compute the stored value
before eliding a load.
* gcc.dg/torture/pr94949.c: New testcase.
---
gcc/testsuite/gcc.dg/torture/pr94949.c | 17 +++++++++++++++++
gcc/tree-ssa-loop-im.c | 10 ++++++----
2 files changed, 23 insertions(+), 4 deletions(-)
create mode 100644 gcc/testsuite/gcc.dg/torture/pr94949.c
diff --git a/gcc/testsuite/gcc.dg/torture/pr94949.c
b/gcc/testsuite/gcc.dg/torture/pr94949.c
new file mode 100644
index 00000000000..6182d77b3cd
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/torture/pr94949.c
@@ -0,0 +1,17 @@
+/* { dg-do run } */
+/* { dg-additional-options "-fallow-store-data-races" } */
+
+static int x = 1;
+static volatile int y = -1;
+int
+main()
+{
+ for (int i = 0; i < 128; ++i)
+ {
+ if (i == y)
+ x = i;
+ }
+ if (x != 1)
+ __builtin_abort ();
+ return 0;
+}
diff --git a/gcc/tree-ssa-loop-im.c b/gcc/tree-ssa-loop-im.c
index 18e5c18c17e..554dd4be5bb 100644
--- a/gcc/tree-ssa-loop-im.c
+++ b/gcc/tree-ssa-loop-im.c
@@ -2128,9 +2128,9 @@ execute_sm (class loop *loop, vec<edge> exits, im_mem_ref
*ref)
fmt_data.orig_loop = loop;
for_each_index (&ref->mem.ref, force_move_till, &fmt_data);
+ bool always_stored = ref_always_accessed_p (loop, ref, true);
if (bb_in_transaction (loop_preheader_edge (loop)->src)
- || (! flag_store_data_races
- && ! ref_always_accessed_p (loop, ref, true)))
+ || (! flag_store_data_races && ! always_stored))
multi_threaded_model_p = true;
if (multi_threaded_model_p)
@@ -2145,8 +2145,10 @@ execute_sm (class loop *loop, vec<edge> exits,
im_mem_ref *ref)
/* Avoid doing a load if there was no load of the ref in the loop.
Esp. when the ref is not always stored we cannot optimize it
- away later. */
- if (ref->loaded && bitmap_bit_p (ref->loaded, loop->num))
+ away later. But when it is not always stored we must use a conditional
+ store then. */
+ if ((!always_stored && !multi_threaded_model_p)
+ || (ref->loaded && bitmap_bit_p (ref->loaded, loop->num)))
{
load = gimple_build_assign (tmp_var, unshare_expr (ref->mem.ref));
lim_data = init_lim_data (load);
--
2.13.7
