Patchwork PR tree-optimization/54985

login
register
mail settings
Submitter Jeff Law
Date Oct. 23, 2012, 8:35 p.m.
Message ID <5086FF8C.40209@redhat.com>
Download mbox | patch
Permalink /patch/193567/
State New
Headers show

Comments

Jeff Law - Oct. 23, 2012, 8:35 p.m.
When we try to thread across a back edge in the CFG we have a check to 
ensure that we don't use temporary equivalences which are invalidated by 
traversal of the back edge to simplify the final conditional.

About a year ago I added code to pick up secondary threading 
opportunities after an initial jump threading was successful.  However, 
I failed to account for equivalences which were invalidated by the back 
edge traversal in those new cases.

This patch extracts the check into its own function, then calls it from 
the 3 locations where it's necessary.

Bootstrapped and regression tested on x86_64-unknown-linux-gnu. 
Installed onto the trunk.

Oh how I want to rewrite all this code....


Jeff

ps.  First time commiting using git svn, hopefully I did all the steps 
correctly.
* tree-ssa-threadedge.c (cond_arg_set_in_bb): New function extracted
	from thread_across_edge.
	(thread_across_edge): Use it in all cases where we might thread
	across a back edge.

	* gcc.c-torture/execute/pr54985.c: New test.
Jakub Jelinek - Oct. 23, 2012, 8:50 p.m.
On Tue, Oct 23, 2012 at 02:35:24PM -0600, Jeff Law wrote:
> +/* Return TRUE if the statement at the end of e->dest depends on
> +   the output of any statement in BB.   Otherwise return FALSE.
> +
> +   This is used when we are threading a backedge and need to ensure
> +   that temporary equivalences from BB do not affect the condition
> +   in e->dest.  */
> +
> +static bool
> +cond_arg_set_in_bb (edge e, basic_block bb, int n)
> +{
> +  ssa_op_iter iter;
> +  use_operand_p use_p;
> +  gimple last = gsi_stmt (gsi_last_bb (e->dest));

Use gimple last = last_stmt (e->dest); instead?  That way any possible
debug stmts are ignored.

	Jakub
Jeff Law - Oct. 23, 2012, 9:21 p.m.
On 10/23/2012 02:50 PM, Jakub Jelinek wrote:
> On Tue, Oct 23, 2012 at 02:35:24PM -0600, Jeff Law wrote:
>> +/* Return TRUE if the statement at the end of e->dest depends on
>> +   the output of any statement in BB.   Otherwise return FALSE.
>> +
>> +   This is used when we are threading a backedge and need to ensure
>> +   that temporary equivalences from BB do not affect the condition
>> +   in e->dest.  */
>> +
>> +static bool
>> +cond_arg_set_in_bb (edge e, basic_block bb, int n)
>> +{
>> +  ssa_op_iter iter;
>> +  use_operand_p use_p;
>> +  gimple last = gsi_stmt (gsi_last_bb (e->dest));
>
> Use gimple last = last_stmt (e->dest); instead?  That way any possible
> debug stmts are ignored.
I thought I'd already dealt with this before.  I'll double-check and 
take appropriate action.

Any opinions about pulling it into 4.7.x as that release is affected by 
this codegen bug?  I've got no strong opinions and I'm willing to pull 
it onto the branch if you want it.

jeff
Jakub Jelinek - Oct. 23, 2012, 9:22 p.m.
On Tue, Oct 23, 2012 at 03:21:59PM -0600, Jeff Law wrote:
> On 10/23/2012 02:50 PM, Jakub Jelinek wrote:
> >>+static bool
> >>+cond_arg_set_in_bb (edge e, basic_block bb, int n)
> >>+{
> >>+  ssa_op_iter iter;
> >>+  use_operand_p use_p;
> >>+  gimple last = gsi_stmt (gsi_last_bb (e->dest));
> >
> >Use gimple last = last_stmt (e->dest); instead?  That way any possible
> >debug stmts are ignored.
> I thought I'd already dealt with this before.  I'll double-check and
> take appropriate action.
> 
> Any opinions about pulling it into 4.7.x as that release is affected
> by this codegen bug?  I've got no strong opinions and I'm willing to
> pull it onto the branch if you want it.

I think it should be backported to 4.7, perhaps with a few days delay after the
trunk commit.

	Jakub
Jeff Law - Oct. 23, 2012, 9:34 p.m.
On 10/23/2012 03:22 PM, Jakub Jelinek wrote:
> On Tue, Oct 23, 2012 at 03:21:59PM -0600, Jeff Law wrote:
>> On 10/23/2012 02:50 PM, Jakub Jelinek wrote:
>>>> +static bool
>>>> +cond_arg_set_in_bb (edge e, basic_block bb, int n)
>>>> +{
>>>> +  ssa_op_iter iter;
>>>> +  use_operand_p use_p;
>>>> +  gimple last = gsi_stmt (gsi_last_bb (e->dest));
>>>
>>> Use gimple last = last_stmt (e->dest); instead?  That way any possible
>>> debug stmts are ignored.
>> I thought I'd already dealt with this before.  I'll double-check and
>> take appropriate action.
>>
>> Any opinions about pulling it into 4.7.x as that release is affected
>> by this codegen bug?  I've got no strong opinions and I'm willing to
>> pull it onto the branch if you want it.
>
> I think it should be backported to 4.7, perhaps with a few days delay after the
> trunk commit.
Do we even have debug statements after control flow statements?

jeff
Jakub Jelinek - Oct. 23, 2012, 9:43 p.m.
On Tue, Oct 23, 2012 at 03:34:46PM -0600, Jeff Law wrote:
> >I think it should be backported to 4.7, perhaps with a few days delay after the
> >trunk commit.
> Do we even have debug statements after control flow statements?

They shouldn't be there, so if you just give up the same way for gsi_stmt
NULL as well as non-control stmt, it shouldn't make a difference.
So last_stmt might be just shorter to type and more commonly used, nothing
more.

	Jakub
Jeff Law - Oct. 24, 2012, 6:52 p.m.
On 10/23/2012 03:43 PM, Jakub Jelinek wrote:
> On Tue, Oct 23, 2012 at 03:34:46PM -0600, Jeff Law wrote:
>>> I think it should be backported to 4.7, perhaps with a few days delay after the
>>> trunk commit.
>> Do we even have debug statements after control flow statements?
>
> They shouldn't be there, so if you just give up the same way for gsi_stmt
> NULL as well as non-control stmt, it shouldn't make a difference.
> So last_stmt might be just shorter to type and more commonly used, nothing
> more.
 From looking at the existing code and last_stmt; for the cases we care 
about, they ought to be equivalent.

Using last_stmt seems marginally clearer.  I'll go ahead and make that 
change after the usual bootstrap & test cycle.

jeff

Patch

diff --git a/gcc/testsuite/gcc.c-torture/execute/pr54985.c b/gcc/testsuite/gcc.c-torture/execute/pr54985.c
new file mode 100644
index 0000000..678c9f4
--- /dev/null
+++ b/gcc/testsuite/gcc.c-torture/execute/pr54985.c
@@ -0,0 +1,36 @@ 
+
+typedef struct st {
+    int a;
+} ST;
+
+int __attribute__((noinline,noclone))
+foo(ST *s, int c)
+{
+  int first = 1;
+  int count = c;
+  ST *item = s;
+  int a = s->a;
+  int x;
+
+  while (count--)
+    {
+      x = item->a;
+      if (first)
+        first = 0;
+      else if (x >= a)
+        return 1;
+      a = x;
+      item++;
+    }
+  return 0;
+}
+
+extern void abort (void);
+
+int main ()
+{
+  ST _1[2] = {{2}, {1}};
+  if (foo(_1, 2) != 0)
+    abort ();
+  return 0;
+}
diff --git a/gcc/tree-ssa-threadedge.c b/gcc/tree-ssa-threadedge.c
index 105e3ab..491aa9f 100644
--- a/gcc/tree-ssa-threadedge.c
+++ b/gcc/tree-ssa-threadedge.c
@@ -572,6 +572,44 @@  simplify_control_stmt_condition (edge e,
   return cached_lhs;
 }
 
+/* Return TRUE if the statement at the end of e->dest depends on
+   the output of any statement in BB.   Otherwise return FALSE.
+
+   This is used when we are threading a backedge and need to ensure
+   that temporary equivalences from BB do not affect the condition
+   in e->dest.  */
+
+static bool
+cond_arg_set_in_bb (edge e, basic_block bb, int n)
+{
+  ssa_op_iter iter;
+  use_operand_p use_p;
+  gimple last = gsi_stmt (gsi_last_bb (e->dest));
+
+  /* E->dest does not have to end with a control transferring
+     instruction.  This can occurr when we try to extend a jump
+     threading opportunity deeper into the CFG.  In that case
+     it is safe for this check to return false.  */
+  if (!last)
+    return false;
+
+  if (gimple_code (last) != GIMPLE_COND
+      && gimple_code (last) != GIMPLE_GOTO
+      && gimple_code (last) != GIMPLE_SWITCH)
+    return false;
+
+  FOR_EACH_SSA_USE_OPERAND (use_p, last, iter, SSA_OP_USE | SSA_OP_VUSE)
+    {
+      tree use = USE_FROM_PTR (use_p);
+
+      if (TREE_CODE (use) == SSA_NAME
+	  && gimple_code (SSA_NAME_DEF_STMT (use)) != GIMPLE_PHI
+	  && gimple_bb (SSA_NAME_DEF_STMT (use)) == bb)
+	return true;
+    }
+  return false;
+}
+
 /* TAKEN_EDGE represents the an edge taken as a result of jump threading.
    See if we can thread around TAKEN_EDGE->dest as well.  If so, return
    the edge out of TAKEN_EDGE->dest that we can statically compute will be
@@ -705,19 +743,8 @@  thread_across_edge (gimple dummy_cond,
      safe to thread this edge.  */
   if (e->flags & EDGE_DFS_BACK)
     {
-      ssa_op_iter iter;
-      use_operand_p use_p;
-      gimple last = gsi_stmt (gsi_last_bb (e->dest));
-
-      FOR_EACH_SSA_USE_OPERAND (use_p, last, iter, SSA_OP_USE | SSA_OP_VUSE)
-	{
-	  tree use = USE_FROM_PTR (use_p);
-
-          if (TREE_CODE (use) == SSA_NAME
-	      && gimple_code (SSA_NAME_DEF_STMT (use)) != GIMPLE_PHI
-	      && gimple_bb (SSA_NAME_DEF_STMT (use)) == e->dest)
-	    goto fail;
-	}
+      if (cond_arg_set_in_bb (e, e->dest, 1))
+	goto fail;
     }
 
   stmt_count = 0;
@@ -758,7 +785,9 @@  thread_across_edge (gimple dummy_cond,
 	     address.  If DEST is not null, then see if we can thread
 	     through it as well, this helps capture secondary effects
 	     of threading without having to re-run DOM or VRP.  */
-	  if (dest)
+	  if (dest
+	      && ((e->flags & EDGE_DFS_BACK) == 0
+		  || ! cond_arg_set_in_bb (taken_edge, e->dest, 2)))
 	    {
 	      /* We don't want to thread back to a block we have already
  		 visited.  This may be overly conservative.  */
@@ -816,11 +845,16 @@  thread_across_edge (gimple dummy_cond,
 	e3 = taken_edge;
 	do
 	  {
-	    e2 = thread_around_empty_block (e3,
-				            dummy_cond,
-				            handle_dominating_asserts,
-				            simplify,
-				            visited);
+	    if ((e->flags & EDGE_DFS_BACK) == 0
+		|| ! cond_arg_set_in_bb (e3, e->dest, 3))
+	      e2 = thread_around_empty_block (e3,
+					      dummy_cond,
+					      handle_dominating_asserts,
+					      simplify,
+					      visited);
+	    else
+	      e2 = NULL;
+
 	    if (e2)
 	      {
 	        e3 = e2;