Hi! As mentioned in the PR, we ICE on the following testcase, because tree-data-ref.c for VIEW_CONVERT_EXPR<struct S>(0) attempts to create ADDR_EXPR of INTEGER_CST, which is not really valid.
I've successfully bootstrapped/regtested earlier version of this patch with is_gimple_constant instead of is_gimple_min_invariant on x86_64-linux and i686-linux, ok for trunk if even this version succeeds? 2016-07-14 Jakub Jelinek <ja...@redhat.com> PR tree-optimization/71872 * tree-data-ref.c (get_references_in_stmt): Ignore references with is_gimple_constant get_base_address. * gcc.c-torture/compile/pr71872.c: New test. --- gcc/tree-data-ref.c.jj 2016-07-11 22:18:08.000000000 +0200 +++ gcc/tree-data-ref.c 2016-07-14 14:03:57.902899566 +0200 @@ -3868,7 +3868,8 @@ get_references_in_stmt (gimple *stmt, ve if (DECL_P (op1) || (REFERENCE_CLASS_P (op1) && (base = get_base_address (op1)) - && TREE_CODE (base) != SSA_NAME)) + && TREE_CODE (base) != SSA_NAME + && !is_gimple_min_invariant (base))) { ref.ref = op1; ref.is_read = true; --- gcc/testsuite/gcc.c-torture/compile/pr71872.c.jj 2016-07-14 14:14:17.132889043 +0200 +++ gcc/testsuite/gcc.c-torture/compile/pr71872.c 2016-07-14 14:14:05.000000000 +0200 @@ -0,0 +1,15 @@ +/* PR tree-optimization/71872 */ + +struct __attribute__((may_alias)) S { int a; }; + +void +foo (int *x, struct S *y) +{ + int i; + for (i = 0; i < 16; i++) + { + int a = 0; + if (*x) + *(struct S *) y = *(struct S *) &a; + } +} Jakub