Hi! If a comparison can throw, it won't be a valid condition for COND_EXPR or VEC_COND_EXPR, because is_gimple_condexpr will fail. This patch gives up in that case. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2012-01-30 Jakub Jelinek <ja...@redhat.com> PR tree-optimization/52046 * tree-vect-patterns.c (check_bool_pattern): Give up if a comparison could throw. * gcc.dg/pr52046.c: New test. --- gcc/tree-vect-patterns.c.jj 2012-01-13 21:47:35.000000000 +0100 +++ gcc/tree-vect-patterns.c 2012-01-30 10:58:11.048787682 +0100 @@ -1,5 +1,5 @@ /* Analysis Utilities for Loop Vectorization. - Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011 + Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011, 2012 Free Software Foundation, Inc. Contributed by Dorit Nuzman <do...@il.ibm.com> @@ -1967,6 +1967,11 @@ check_bool_pattern (tree var, loop_vec_i { tree vecitype, comp_vectype; + /* If the comparison can throw, then is_gimple_condexpr will be + false and we can't make a COND_EXPR/VEC_COND_EXPR out of it. */ + if (stmt_could_throw_p (def_stmt)) + return false; + comp_vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); if (comp_vectype == NULL_TREE) return false; --- gcc/testsuite/gcc.dg/pr52046.c.jj 2012-01-30 10:59:42.749264286 +0100 +++ gcc/testsuite/gcc.dg/pr52046.c 2012-01-30 10:59:28.000000000 +0100 @@ -0,0 +1,14 @@ +/* PR tree-optimization/52046 */ +/* { dg-do compile } */ +/* { dg-options "-O3 -fexceptions -fnon-call-exceptions" } */ + +extern float a[], b[], c[], d[]; +extern int k[]; + +void +foo (void) +{ + int i; + for (i = 0; i < 1024; ++i) + k[i] = (a[i] < b[i]) | (c[i] < d[i]); +} Jakub