Re: [patch] Fix for automatic parallelization

["Sebastian Pop" <sebpop at gmail dot com>]

Hi,

Here is the last fix for the problem that I've reported yesterday,
where gcc failed on the verification of loop structures.  The patch
filters out all the loops that are in irreducible sections.

Another possible way to fix this problem is to mark the newly created
blocks and edges with the irreducible flag, but I'm not sure we want
to parallelize these loops.

Sebastian

Index: tree-parloops.c
===================================================================
--- tree-parloops.c	(revision 131430)
+++ tree-parloops.c	(working copy)
@@ -431,6 +431,25 @@ loop_parallel_p (struct loop *loop, htab
   return ret;
 }
 
+/* Return true when LOOP contains irreducible basic blocks.  */
+
+static inline bool
+loop_has_irreducible_bbs (struct loop *loop)
+{
+  unsigned i;
+  basic_block *bbs = get_loop_body_in_dom_order (loop);
+  bool res = true;
+
+  for (i = 0; i < loop->num_nodes; i++)
+    if (bbs[i]->flags & BB_IRREDUCIBLE_LOOP)
+      goto end;
+
+  res = false;
+ end:
+  free (bbs);
+  return res;
+}
+
 /* Assigns the address of OBJ in TYPE to an ssa name, and returns this name.
    The assignment statement is placed before LOOP.  DECL_ADDRESS maps decls
    to their addresses that can be reused.  The address of OBJ is known to
@@ -1738,6 +1757,7 @@ parallelize_loops (void)
 	  || expected_loop_iterations (loop) <= n_threads
 	  /* And of course, the loop must be parallelizable.  */
 	  || !can_duplicate_loop_p (loop)
+	  || !loop_has_irreducible_bbs (loop)
 	  || !loop_parallel_p (loop, reduction_list, &niter_desc))
 	continue;