On Mon, Mar 18, 2013 at 1:00 PM, Jakub Jelinek <ja...@redhat.com> wrote: > Hi! > > On the attached testcase we ICE, because cselim uses > if (!is_gimple_reg_type (TREE_TYPE (lhs)) > || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0)) > to guard an optimization, and we have two MEM_REFs where operand_equal_p > is true, but they don't have compatible types (the first one is _Complex > double, i.e. is_gimple_reg_type), the second has aggregate type of the > same size/same mode (structure containing the _Complex double). > > The first patch is what I've bootstrapped/regtested on i686-linux so far > (x86_64-linux regtest still pending), it seems to trigger different return > value from operand_equal_p in a couple of TUs: > c52008b.adb > /usr/src/gcc-4.8/libstdc++-v3/src/c++98/complex_io.cc > /usr/src/gcc-4.8/lto-plugin/lto-plugin.c > /usr/src/gcc-4.8/gcc/testsuite/g++.dg/opt/pr47366.C > /usr/src/gcc-4.8/gcc/testsuite/g++.dg/torture/pr56635.C > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/actual_array_constructor_1.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/alloc_comp_assign_2.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/alloc_comp_assign_3.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/argument_checking_1.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/array_memcpy_2.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/array_temporaries_3.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/auto_char_len_3.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/char_pointer_func.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/proc_ptr_23.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/pure_byref_1.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/realloc_on_assign_17.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/reshape.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/string_ctor_1.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/unlimited_polymorphic_1.f03 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/vector_subscript_2.f90 > /usr/src/gcc-4.8/gcc/testsuite/gfortran.dg/widechar_8.f90 > /usr/src/gcc-4.8/libstdc++-v3/testsuite/23_containers/unordered_map/cons/56112.cc > /usr/src/gcc-4.8/libstdc++-v3/testsuite/25_algorithms/remove_if/moveable.cc > /usr/src/gcc-4.8/libstdc++-v3/testsuite/25_algorithms/remove/moveable.cc > /usr/src/gcc-4.8/libstdc++-v3/testsuite/25_algorithms/stable_sort/49559.cc > but from quick look at some of them it didn't look like something very > undesirable. In any case, no new FAILs. Ok for trunk if it passes regtest > on x86_64-linux too?
Ok. Thanks, Richard. > The second patch is a safer version intended for 4.8.0, to be > bootstrapped/regtested afterwards. > > Jakub
