[PATCH] Fix reassoc ICE (PR tree-optimization/56559)

Jakub Jelinek jakub@redhat.com
Thu Mar 7 13:49:00 GMT 2013


Hi!

During zero_one_operation I believe we can count in the chain having single
uses if we go through rhs1s, but the change from last year also added
looking through rhs2, and there I don't see anything that would guarantee
a single use, and the following testcase shows where it has two uses.

If it has more than one use, obviously we can't change the pow call just
because in one of the uses we'd like to remove one multiplication.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2013-03-07  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/56559
	* tree-ssa-reassoc.c (zero_one_operation): When looking at rhs2,
	check that it has only a single use.

	* gcc.dg/tree-ssa/reassoc-26.c: New test.

--- gcc/tree-ssa-reassoc.c.jj	2013-02-18 16:39:55.000000000 +0100
+++ gcc/tree-ssa-reassoc.c	2013-03-07 10:09:22.048925800 +0100
@@ -1107,7 +1107,8 @@ zero_one_operation (tree *def, enum tree
 	 the operand might be hiding in the rightmost one.  */
       if (opcode == MULT_EXPR
 	  && gimple_assign_rhs_code (stmt) == opcode
-	  && TREE_CODE (gimple_assign_rhs2 (stmt)) == SSA_NAME)
+	  && TREE_CODE (gimple_assign_rhs2 (stmt)) == SSA_NAME
+	  && has_single_use (gimple_assign_rhs2 (stmt)))
 	{
 	  gimple stmt2 = SSA_NAME_DEF_STMT (gimple_assign_rhs2 (stmt));
 	  if (stmt_is_power_of_op (stmt2, op))
--- gcc/testsuite/gcc.dg/tree-ssa/reassoc-26.c.jj	2013-03-07 10:10:49.400749110 +0100
+++ gcc/testsuite/gcc.dg/tree-ssa/reassoc-26.c	2013-03-07 10:10:18.000000000 +0100
@@ -0,0 +1,12 @@
+/* PR tree-optimization/56559 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -ffast-math" } */
+
+double a, b, c, d, e;
+
+void
+foo ()
+{
+  a = e * e;
+  b = d * e + c * e * a;
+}

	Jakub



More information about the Gcc-patches mailing list