On Mon, Apr 04, 2016 at 03:27:23PM +0200, Richard Biener wrote: > It ICEs in > > /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.target/i386/pr37870.c:19:1: > internal compiler error: in subreg_get_info, at rtlanal.c:3695 > 0xddee5a subreg_get_info(unsigned int, machine_mode, unsigned int, > machine_mode, subreg_info*) > /space/rguenther/src/svn/trunk/gcc/rtlanal.c:3695 > 0xddf12b simplify_subreg_regno(unsigned int, machine_mode, unsigned int, > machine_mode) > /space/rguenther/src/svn/trunk/gcc/rtlanal.c:3808 > 0xd8bc7a simplifiable_subregs(subreg_shape const&) > /space/rguenther/src/svn/trunk/gcc/reginfo.c:1234 > 0xd8be1d record_subregs_of_mode > /space/rguenther/src/svn/trunk/gcc/reginfo.c:1294 > 0xd8c246 init_subregs_of_mode() > /space/rguenther/src/svn/trunk/gcc/reginfo.c:1348 > 0xc1a55c init_costs > /space/rguenther/src/svn/trunk/gcc/ira-costs.c:2187 > > which is > > /* This should always pass, otherwise we don't know how to verify > the constraint. These conditions may be relaxed but > subreg_regno_offset would need to be redesigned. */ > gcc_assert ((GET_MODE_SIZE (xmode) % GET_MODE_SIZE (ymode)) == 0);
So perhaps either validate_subreg should check the same thing, or the extraction could check this. Anyway, I don't have anything against removing that hunk from extract_bit_field_1 for GCC7. Jakub