------- Additional Comments From amylaar at gcc dot gnu dot org 2005-08-15 14:14 ------- (In reply to comment #1) > So, using limit 0 for when calculating the integer mode for the size would > fix > the regression on sh?
Yes, it would fix the problem with CDImode, however, the mode_for_size_tree call in compute_record_mode is not only used to determine if any mode found from a single member would be suitable, but also the what mode to use otherwise. You don't really want to change what you use for the latter. For vector modes, there is also the added complication that there might be no integer mode at all wide enough to match the size of the vector mode. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23290
