Hi! For non-gimple reg types var = x ? y : z; is gimplified as x ? (var = y) : (var = z);. The problem is that if var is TREE_READONLY, we can gimplify one of those assignments by promoting them to TREE_STATIC (if not addressable) and making the initializer into their DECL_INITIAL. That doesn't work well, because then we conditionally store into that and during optimizations, as it is a TREE_READONLY var with DECL_INITIAL, just use that initializer for optimizations.
The following patch fixes that by clearing TREE_READONLY flag if we are going to store it more than once. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2017-03-21 Jakub Jelinek <ja...@redhat.com> PR c++/80129 * gimplify.c (gimplify_modify_expr_rhs) <case COND_EXPR>: Clear TREE_READONLY on result if writing it more than once. * g++.dg/torture/pr80129.C: New test. --- gcc/gimplify.c.jj 2017-03-21 07:56:55.000000000 +0100 +++ gcc/gimplify.c 2017-03-21 13:37:45.555612652 +0100 @@ -5098,6 +5098,13 @@ gimplify_modify_expr_rhs (tree *expr_p, if (ret != GS_ERROR) ret = GS_OK; + /* If we are going to write RESULT more than once, clear + TREE_READONLY flag, otherwise we might gimplify it + incorrectly. */ + if (DECL_P (result) + && TREE_TYPE (TREE_OPERAND (cond, 1)) != void_type_node + && TREE_TYPE (TREE_OPERAND (cond, 2)) != void_type_node) + TREE_READONLY (result) = 0; if (TREE_TYPE (TREE_OPERAND (cond, 1)) != void_type_node) TREE_OPERAND (cond, 1) = build2 (code, void_type_node, result, --- gcc/testsuite/g++.dg/torture/pr80129.C.jj 2017-03-21 13:40:04.179852313 +0100 +++ gcc/testsuite/g++.dg/torture/pr80129.C 2017-03-21 13:41:32.121735570 +0100 @@ -0,0 +1,14 @@ +// PR c++/80129 +// { dg-do run } +// { dg-options "-std=c++11" } + +struct A { bool a; int b; }; + +int +main () +{ + bool c = false; + const A x = c ? A {true, 1} : A {false, 0}; + if (x.a) + __builtin_abort (); +} Jakub