On Wed, Oct 15, 2025 at 11:53:07AM +0200, Jakub Jelinek wrote: > On Wed, Oct 15, 2025 at 11:36:29AM +0200, Richard Biener wrote: > > > This patch has been tested on x86_64-pc-linux-gnu with make bootstrap > > > and make -k check with no new failures. Ok for mainline? > > > > LGTM. Please leave Jakub time to comment though. > > Ok for me. > > > I'll note that for "full" TImode or larger mode support the currently > > unsigned HOST_WIDE_INT last_set_nonzero_bits and nonzero_bits > > fields would also need to be enlarged. Not sure how well using > > of sign_bit_copies but not nonzero_bits is guarded within combine. > > We'd need to use wide_int or say offset_int or FIXED_WIDE_INT (128) > instead of unsigned HOST_WIDE_INT for nonzero_bits. > I think that would be a good idea, because in many cases for TImode > we just punt and TImode is supported on quite a few targets, but it > would be a larger change.
Yes. A larger change, but a good idea no matter wAhat. This whole area would benefit from a huge rework, the whole reg_stat thing is antiquated at best, and some would say positively harmful, but it is copletely useless in any case. This kind of thing can and should be handled using DF. Segher
