diff mbox

[PR,tree-optimization/64823] Handle threading through blocks with PHIs, but no statements

Message ID 54DE67B3.3010404@redhat.com
State New
Headers show

Commit Message

Jeff Law Feb. 13, 2015, 9:08 p.m. UTC 
Generally a block with PHIs, but no statements will be optimized away; 
however, if the sole successor is a loop header, then such blocks will 
not be optimized away.

We want to be able to thread through those blocks as sometimes we will 
be able to thread from outside the loop to the loop exit, thus bypassing 
the loop entirely on one or more paths.

It's also important to thread these cases so avoid false positive 
warnings, such as the one found in the testcase for this PR.

Basically three things need to happen to fix this PR.  First, the 
routine that determines what blocks are potentially threadable needs to 
be aware of this slightly special case.

Second, the threading code did not distinguish between a block with no 
statements and a block where we did not process all the statements.

Finally, VRP has a biglet in that it'll ICE if we have a potentially 
threadable block with no statements.

This patch addresses all three issues and fixes the false positive 
warning from this PR.

Bootstrapped & regression tested on x86_64-unknown-linux-gnu and 
installed on the trunk.
diff mbox

Patch

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index f36e16c..a574c2b 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,5 +1,12 @@ 
 2015-02-13  Jeff Law  <law@redhat.com>
 
+	PR tree-optimization/64823
+	* tree-vrp.c (identify_jump_threads): Handle blocks with no statements.
+	* tree-ssa-threadedge.c (potentially_threadable_block): Allow
+	threading through blocks with PHIs, but no statements.
+	(thread_through_normal_block): Distinguish between blocks where
+	we did not process all the statements and blocks with no statements.
+
 	PR rtl-optimization/47477
 	* match.pd (convert (plus/minus (convert @0) (convert @1): New
 	simplifier to narrow arithmetic.
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 2fe9698..f700bb1 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,5 +1,8 @@ 
 2015-02-13  Jeff Law  <law@redhat.com>
 
+	PR tree-optimization/64823
+	gcc.dg/uninit-20.c: New test.
+
 	PR rtl-optimization/47477
 	* gcc.dg/tree-ssa/pr47477.c: New test.
 
diff --git a/gcc/tree-ssa-threadedge.c b/gcc/tree-ssa-threadedge.c
index 4f83991..7187d06 100644
--- a/gcc/tree-ssa-threadedge.c
+++ b/gcc/tree-ssa-threadedge.c
@@ -110,6 +110,15 @@  potentially_threadable_block (basic_block bb)
 {
   gimple_stmt_iterator gsi;
 
+  /* Special case.  We can get blocks that are forwarders, but are
+     not optimized away because they forward from outside a loop
+     to the loop header.   We want to thread through them as we can
+     sometimes thread to the loop exit, which is obviously profitable. 
+     the interesting case here is when the block has PHIs.  */
+  if (gsi_end_p (gsi_start_nondebug_bb (bb))
+      && !gsi_end_p (gsi_start_phis (bb)))
+    return true;
+  
   /* If BB has a single successor or a single predecessor, then
      there is no threading opportunity.  */
   if (single_succ_p (bb) || single_pred_p (bb))
@@ -1281,16 +1290,32 @@  thread_through_normal_block (edge e,
     = record_temporary_equivalences_from_stmts_at_dest (e, stack, simplify,
 							*backedge_seen_p);
 
-  /* If we didn't look at all the statements, the most likely reason is
-     there were too many and thus duplicating this block is not profitable.
+  /* There's two reasons STMT might be null, and distinguishing
+     between them is important.
 
-     Also note if we do not look at all the statements, then we may not
-     have invalidated equivalences that are no longer valid if we threaded
-     around a loop.  Thus we must signal to our caller that this block
-     is not suitable for use as a joiner in a threading path.  */
+     First the block may not have had any statements.  For example, it
+     might have some PHIs and unconditionally transfer control elsewhere.
+     Such blocks are suitable for jump threading, particularly as a
+     joiner block.
+
+     The second reason would be if we did not process all the statements
+     in the block (because there were too many to make duplicating the
+     block profitable.   If we did not look at all the statements, then
+     we may not have invalidated everything needing invalidation.  Thus
+     we must signal to our caller that this block is not suitable for
+     use as a joiner in a threading path.  */
   if (!stmt)
-    return -1;
+    {
+      /* First case.  The statement simply doesn't have any instructions, but
+	 does have PHIs.  */
+      if (gsi_end_p (gsi_start_nondebug_bb (e->dest))
+	  && !gsi_end_p (gsi_start_phis (e->dest)))
+	return 0;
 
+      /* Second case.  */
+      return -1;
+    }
+  
   /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
      will be taken.  */
   if (gimple_code (stmt) == GIMPLE_COND
diff --git a/gcc/tree-vrp.c b/gcc/tree-vrp.c
index dad1830..7367684 100644
--- a/gcc/tree-vrp.c
+++ b/gcc/tree-vrp.c
@@ -10181,8 +10181,15 @@  identify_jump_threads (void)
       /* We're basically looking for a switch or any kind of conditional with
 	 integral or pointer type arguments.  Note the type of the second
 	 argument will be the same as the first argument, so no need to
-	 check it explicitly.  */
-      if (gimple_code (last) == GIMPLE_SWITCH
+	 check it explicitly. 
+
+	 We also handle the case where there are no statements in the
+	 block.  This come up with forwarder blocks that are not
+	 optimized away because they lead to a loop header.  But we do
+	 want to thread through them as we can sometimes thread to the
+	 loop exit which is obviously profitable.  */
+      if (!last
+	  || gimple_code (last) == GIMPLE_SWITCH
 	  || (gimple_code (last) == GIMPLE_COND
       	      && TREE_CODE (gimple_cond_lhs (last)) == SSA_NAME
 	      && (INTEGRAL_TYPE_P (TREE_TYPE (gimple_cond_lhs (last)))