diff mbox series

ifcvt: Do not lower bitfields if we can't analyze dr's [PR107275]

Message ID 0888cc2d-2040-52c3-1201-16400567300b@arm.com
State New
Headers show
Series ifcvt: Do not lower bitfields if we can't analyze dr's [PR107275] | expand

Commit Message

Andre Vieira (lists) Oct. 18, 2022, 8:35 a.m. UTC
The ifcvt dead code elimination code was not built to deal with inline
assembly, as loops with such would never be if-converted in the past since
we can't do data-reference analysis on them and vectorization would 
eventually fail.
For this reason we now also do not lower bitfields if the data-reference 
analysis
fails, as we would not end up vectorizing it.  As a consequence this 
also fixes
this PR as the dead code elimination will not run for such cases and 
wrongfully
eliminate inline assembly statements.

Bootstrapped and regression tested on aarch64-none-linux-gnu and 
x86_64-pc-linux-gnu.

gcc/ChangeLog:

     PR tree-optimization/107275
     * tree-if-conv.cc (if_convertible_loop_p_1): Move
     find_data_references_in_loop call from here...
     (if_convertible_loop_p): And move data-reference vector initialization
     from here...
     (tree_if_conversion):... to here.

gcc/testsuite/ChangeLog:

     * gcc.dg/vect/pr107275.c: New test.

Comments

Richard Biener Oct. 18, 2022, 9:01 a.m. UTC | #1
On Tue, 18 Oct 2022, Andre Vieira (lists) wrote:

> The ifcvt dead code elimination code was not built to deal with inline
> assembly, as loops with such would never be if-converted in the past since
> we can't do data-reference analysis on them and vectorization would eventually
> fail.
> For this reason we now also do not lower bitfields if the data-reference
> analysis
> fails, as we would not end up vectorizing it.  As a consequence this also
> fixes
> this PR as the dead code elimination will not run for such cases and
> wrongfully
> eliminate inline assembly statements.
> 
> Bootstrapped and regression tested on aarch64-none-linux-gnu and
> x86_64-pc-linux-gnu.

OK.

Thanks,
Richard.

> gcc/ChangeLog:
> 
>     PR tree-optimization/107275
>     * tree-if-conv.cc (if_convertible_loop_p_1): Move
>     find_data_references_in_loop call from here...
>     (if_convertible_loop_p): And move data-reference vector initialization
>     from here...
>     (tree_if_conversion):... to here.
> 
> gcc/testsuite/ChangeLog:
> 
>     * gcc.dg/vect/pr107275.c: New test.
>
diff mbox series

Patch

diff --git a/gcc/testsuite/gcc.dg/vect/pr107275.c b/gcc/testsuite/gcc.dg/vect/pr107275.c
new file mode 100644
index 0000000000000000000000000000000000000000..16327c46add7f74dd36c1a636c0fb9c84d5232c8
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr107275.c
@@ -0,0 +1,15 @@ 
+/* { dg-do compile } */
+struct st
+{
+  int a : 1;
+};
+
+void
+foo (struct st *s, int n)
+{
+  for (int i = 0; i < n; ++i)
+    {
+      s[i].a = i;
+      __asm__ __volatile__ ("":::"memory");
+    }
+}
diff --git a/gcc/tree-if-conv.cc b/gcc/tree-if-conv.cc
index 01637c5da08d5a2a00a495522fc9a6436a804398..a83b013d2ad3d9066b1c2bc62282ec054483598f 100644
--- a/gcc/tree-if-conv.cc
+++ b/gcc/tree-if-conv.cc
@@ -1416,9 +1416,6 @@  if_convertible_loop_p_1 (class loop *loop, vec<data_reference_p> *refs)
   basic_block exit_bb = NULL;
   vec<basic_block> region;
 
-  if (find_data_references_in_loop (loop, refs) == chrec_dont_know)
-    return false;
-
   calculate_dominance_info (CDI_DOMINATORS);
 
   for (i = 0; i < loop->num_nodes; i++)
@@ -1541,12 +1538,11 @@  if_convertible_loop_p_1 (class loop *loop, vec<data_reference_p> *refs)
    - if its basic blocks and phi nodes are if convertible.  */
 
 static bool
-if_convertible_loop_p (class loop *loop)
+if_convertible_loop_p (class loop *loop, vec<data_reference_p> *refs)
 {
   edge e;
   edge_iterator ei;
   bool res = false;
-  vec<data_reference_p> refs;
 
   /* Handle only innermost loop.  */
   if (!loop || loop->inner)
@@ -1578,15 +1574,7 @@  if_convertible_loop_p (class loop *loop)
     if (loop_exit_edge_p (loop, e))
       return false;
 
-  refs.create (5);
-  res = if_convertible_loop_p_1 (loop, &refs);
-
-  data_reference_p dr;
-  unsigned int i;
-  for (i = 0; refs.iterate (i, &dr); i++)
-    free (dr->aux);
-
-  free_data_refs (refs);
+  res = if_convertible_loop_p_1 (loop, refs);
 
   delete innermost_DR_map;
   innermost_DR_map = NULL;
@@ -3499,6 +3487,7 @@  tree_if_conversion (class loop *loop, vec<gimple *> *preds)
   auto_vec <gassign *, 4> writes_to_lower;
   bitmap exit_bbs;
   edge pe;
+  vec<data_reference_p> refs;
 
  again:
   rloop = NULL;
@@ -3508,6 +3497,7 @@  tree_if_conversion (class loop *loop, vec<gimple *> *preds)
   need_to_predicate = false;
   need_to_rewrite_undefined = false;
   any_complicated_phi = false;
+  refs.create (5);
 
   /* Apply more aggressive if-conversion when loop or its outer loop were
      marked with simd pragma.  When that's the case, we try to if-convert
@@ -3537,11 +3527,14 @@  tree_if_conversion (class loop *loop, vec<gimple *> *preds)
       goto cleanup;
     }
 
+  if (find_data_references_in_loop (loop, &refs) == chrec_dont_know)
+    goto cleanup;
+
   if (loop->num_nodes > 2)
     {
       need_to_ifcvt = true;
 
-      if (!if_convertible_loop_p (loop) || !dbg_cnt (if_conversion_tree))
+      if (!if_convertible_loop_p (loop, &refs) || !dbg_cnt (if_conversion_tree))
 	goto cleanup;
 
       if ((need_to_predicate || any_complicated_phi)
@@ -3658,6 +3651,13 @@  tree_if_conversion (class loop *loop, vec<gimple *> *preds)
   todo |= TODO_cleanup_cfg;
 
  cleanup:
+  data_reference_p dr;
+  unsigned int i;
+  for (i = 0; refs.iterate (i, &dr); i++)
+    free (dr->aux);
+
+  refs.truncate (0);
+
   if (ifc_bbs)
     {
       unsigned int i;