------- Comment #1 from jakub at gcc dot gnu dot org 2009-03-13 12:23 ------- Created an attachment (id=17455) --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=17455&action=view) gcc44-pr39454.patch
So far only very lightly tested fix (no bootstrap/regtest yet), actually either the cse.c or rtlanal.c part alone fixes this problem. In fold_rtx the problem is that we canonicalize const_arg1 (but without modifying x), in order to optimize nested shifts. As there aren't any, we fall back to new_rtx = simplify_binary_operation (code, mode, const_arg0 ? const_arg0 : folded_arg0, const_arg1 ? const_arg1 : folded_arg1); which returns NULL, as const_arg1 passed to it is already canonicalized. As new_rtx is NULL, we return x, but in x we haven't canonicalized the shift count. In nonzero_bits/num_sign_bit_copies for SHIFT_COUNT_TRUNCATED we either should truncate the shift counts before using them, or (and if !SHIFT_COUNT_TRUNCATED that's the only choice) not try to assume anything from out of range shifts. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39454
