https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108440
--- Comment #8 from CVS Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Jakub Jelinek <ja...@gcc.gnu.org>: https://gcc.gnu.org/g:05b9868b182bb9ed2013b39a0bc6297354a0db49 commit r13-5254-g05b9868b182bb9ed2013b39a0bc6297354a0db49 Author: Jakub Jelinek <ja...@redhat.com> Date: Thu Jan 19 10:00:51 2023 +0100 forwprop: Further fixes for simplify_rotate [PR108440] As mentioned in the simplify_rotate comment, for e.g. ((T) ((T2) X << (Y & (B - 1)))) | ((T) ((T2) X >> ((-Y) & (B - 1)))) we already emit X r<< (Y & (B - 1)) as replacement. This PR is about the ((T) ((T2) X << Y)) OP ((T) ((T2) X >> (B - Y))) ((T) ((T2) X << (int) Y)) OP ((T) ((T2) X >> (int) (B - Y))) forms if T2 is wider than T. Unlike e.g. (X << Y) OP (X >> (B - Y)) which is valid just for Y in [1, B - 1], the above 2 forms are actually valid and do the rotates for Y in [0, B] - for Y 0 the X value is preserved by the left shift and right logical shift by B adds just zeros (but because the shift is in wider precision B is still valid shift count), while for Y equal to B X is preserved through the latter shift and the former adds just zeros. Now, it is unclear if we in the middle-end treat rotates with rotate count equal or larger than precision as UB or not, unlike shifts there are less reasons to do so, but e.g. expansion of X r<< Y if there is no rotate optab for the mode is emitted as (X << Y) | (((unsigned) X) >> ((-Y) & (B - 1))) and so with UB on Y == B. The following patch does multiple things: 1) for the above 2, asks the ranger if Y could be equal to B and if so, instead of using X r<< Y uses X r<< (Y & (B - 1)) 2) for the ((T) ((T2) X << Y)) | ((T) ((T2) X >> ((-Y) & (B - 1)))) ((T) ((T2) X << (int) Y)) | ((T) ((T2) X >> (int) ((-Y) & (B - 1)))) forms that were fixed 2 days ago it only punts if Y might be in the [B,B2-1] range but isn't known to be in the [0,B][2*B,2*B][3*B,3*B]... range. Because for Y which is a multiple of B but smaller than B2 it acts as a rotate too, left shift provides 0 and (-Y) & (B - 1) is 0 and so preserves X. Though, for the cases where Y is not known to be in [0,B-1] the patch also uses X r<< (Y & (B - 1)) rather than X r<< Y 3) as discussed with Aldy, instead of using global ranger it uses a pass specific copy but lazily created on first simplify_rotate that needs it; this e.g. handles rotate inside of if body where the guarding condition limits the shift count to some range which will not work with the global ranger (unless there is some SSA_NAME to attach the range to). Note, e.g. on x86 X r<< (Y & (B - 1)) and X r<< Y actually emit the same assembly because rotates work the same even for larger rotate counts, but that is handled only during combine. 2023-01-19 Jakub Jelinek <ja...@redhat.com> PR tree-optimization/108440 * tree-ssa-forwprop.cc: Include gimple-range.h. (simplify_rotate): For the forms with T2 wider than T and shift counts of Y and B - Y add & (B - 1) masking for the rotate count if Y could be equal to B. For the forms with T2 wider than T and shift counts of Y and (-Y) & (B - 1), don't punt if range could be [B, B2], but only if range doesn't guarantee Y < B or Y = N * B. If range doesn't guarantee Y < B, also add & (B - 1) masking for the rotate count. Use lazily created pass specific ranger instead of get_global_range_query. (pass_forwprop::execute): Disable that ranger at the end of pass if it has been created. * c-c++-common/rotate-10.c: New test. * c-c++-common/rotate-11.c: New test.
