For now give up on predicated values when doing PRE (there's a similar hunk in PHI-translation already). I need to sit down and decide whether it's worth handling them or whether we'd better prune them from the hash tables when assigning value-ids.
Bootstrap and regtest running on x86_64-unknown-linux-gnu. 2018-08-28 Richard Biener <rguent...@suse.de> PR tree-optimization/87117 * tree-ssa-pre.c (compute_avail): Do not make expressions with predicated values available. (get_expr_value_id): Assert we do not run into predicated value expressions. * gcc.dg/pr87117-2.c: New testcase. Index: gcc/tree-ssa-pre.c =================================================================== --- gcc/tree-ssa-pre.c (revision 263906) +++ gcc/tree-ssa-pre.c (working copy) @@ -663,6 +663,7 @@ get_expr_value_id (pre_expr expr) id = VN_INFO (PRE_EXPR_NAME (expr))->value_id; break; case NARY: + gcc_assert (!PRE_EXPR_NARY (expr)->predicated_values); id = PRE_EXPR_NARY (expr)->value_id; break; case REFERENCE: @@ -3902,7 +3903,7 @@ compute_avail (void) continue; vn_nary_op_lookup_stmt (stmt, &nary); - if (!nary) + if (!nary || nary->predicated_values) continue; /* If the NARY traps and there was a preceding Index: gcc/testsuite/gcc.dg/pr87117-2.c =================================================================== --- gcc/testsuite/gcc.dg/pr87117-2.c (nonexistent) +++ gcc/testsuite/gcc.dg/pr87117-2.c (working copy) @@ -0,0 +1,15 @@ +/* { dg-do compile } */ +/* { dg-options "-O -fcode-hoisting" } */ + +void e(); + +void a(int c, char **d) +{ + char b; + if (1 < c) + b = (char)(__INTPTR_TYPE__)d[0]; + if (1 < c && b) + e(); + while (1 < c) + ; +}