diff mbox

[PR50769] Fix ICE in phi_alternatives_equal.

Message ID 4E9E8903.5040808@mentor.com
State New
Headers show

Commit Message

Tom de Vries Oct. 19, 2011, 8:23 a.m. UTC
Richard,

For the example from the PR, -ftree-tail-merge discovers that blocks 6 and 7 are
equal, removes block 7 and redirects the outgoing edge of it's predecessor block
11 to block 6.

Block 6 contains a phi, which after redirecting the edge from block 11 looks
like this:
...
.MEM_54 = PHI <.MEM_27(18), .MEM_26(17), (11)>
...

Since update_vops is false, TODO_update_ssa_only_virtuals will be run
and -ftree-tail-merge doesn't update the phi.

However during cleanup_tree_cfg (which is done before updating the ssa),
phi_alternatives_equal is called which asserts that the phi arguments are defined.

The patch fixes the ICE by setting the empty phi argument to .MEM.

bootstrapped and reg-tested on x86_64.

Ok for trunk?

Thanks,
- Tom

2011-10-19  Tom de Vries  <tom@codesourcery.com>

	PR 50769/tree-optimization
	* tree-ssa-tail-merge.c (replace_block_by): Calculate phi_vuse2
	unconditionally.  Handle case that phi_vuse2 is not an SSA_NAME.  Add
	dummy argument .MEM to phi when increasing number of arguments of phi by
	redirecting edges to the block with phi.

Comments

Richard Biener Oct. 19, 2011, 3:33 p.m. UTC | #1
On Wed, Oct 19, 2011 at 10:23 AM, Tom de Vries <Tom_deVries@mentor.com> wrote:
> Richard,
>
> For the example from the PR, -ftree-tail-merge discovers that blocks 6 and 7 are
> equal, removes block 7 and redirects the outgoing edge of it's predecessor block
> 11 to block 6.
>
> Block 6 contains a phi, which after redirecting the edge from block 11 looks
> like this:
> ...
> .MEM_54 = PHI <.MEM_27(18), .MEM_26(17), (11)>
> ...
>
> Since update_vops is false, TODO_update_ssa_only_virtuals will be run
> and -ftree-tail-merge doesn't update the phi.
>
> However during cleanup_tree_cfg (which is done before updating the ssa),
> phi_alternatives_equal is called which asserts that the phi arguments are defined.
>
> The patch fixes the ICE by setting the empty phi argument to .MEM.
>
> bootstrapped and reg-tested on x86_64.
>
> Ok for trunk?

Ok.

Thanks,
Richard.

> Thanks,
> - Tom
>
> 2011-10-19  Tom de Vries  <tom@codesourcery.com>
>
>        PR 50769/tree-optimization
>        * tree-ssa-tail-merge.c (replace_block_by): Calculate phi_vuse2
>        unconditionally.  Handle case that phi_vuse2 is not an SSA_NAME.  Add
>        dummy argument .MEM to phi when increasing number of arguments of phi by
>        redirecting edges to the block with phi.
>
diff mbox

Patch

Index: gcc/tree-ssa-tail-merge.c
===================================================================
--- gcc/tree-ssa-tail-merge.c (revision 179773)
+++ gcc/tree-ssa-tail-merge.c (working copy)
@@ -1470,11 +1470,14 @@  replace_block_by (basic_block bb1, basic
   edge e;
   edge_iterator ei;
 
+  phi_vuse2 = vop_at_entry (bb2);
+  if (phi_vuse2 != NULL_TREE && TREE_CODE (phi_vuse2) != SSA_NAME)
+    phi_vuse2 = NULL_TREE;
+
   if (update_vops)
     {
       /* Find the vops at entry of bb1 and bb2.  */
       phi_vuse1 = vop_at_entry (bb1);
-      phi_vuse2 = vop_at_entry (bb2);
 
       /* If one of the 2 not found, it means there's no need to update.  */
       update_vops = phi_vuse1 != NULL_TREE && phi_vuse2 != NULL_TREE;
@@ -1507,6 +1510,9 @@  replace_block_by (basic_block bb1, basic
       gcc_assert (pred_edge != NULL);
       if (update_vops)
 	VEC_safe_push (edge, heap, redirected_edges, pred_edge);
+      else if (phi_vuse2 && gimple_bb (SSA_NAME_DEF_STMT (phi_vuse2)) == bb2)
+	add_phi_arg (SSA_NAME_DEF_STMT (phi_vuse2), SSA_NAME_VAR (phi_vuse2),
+		     pred_edge, UNKNOWN_LOCATION);
     }
 
   /* Update the vops.  */