https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97513
--- Comment #8 from CVS Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Richard Sandiford <rsand...@gcc.gnu.org>: https://gcc.gnu.org/g:2f3d9104610cb2058cf091707a20c1c6eff8d470 commit r11-8030-g2f3d9104610cb2058cf091707a20c1c6eff8d470 Author: Richard Sandiford <richard.sandif...@arm.com> Date: Wed Apr 7 15:21:56 2021 +0100 vect: Restore variable-length SLP permutes [PR97513] Many of the gcc.target/sve/slp-perm*.c tests started failing after the introduction of separate SLP permute nodes. This patch adds variable-length support using a similar technique to vect_transform_slp_perm_load. As there, the idea is to detect when every permute mask vector is the same and can be generated using a regular stepped sequence. We can easily handle those cases for variable-length, but still need to restrict the general case to constant-length. Again copying vect_transform_slp_perm_load, the idea is to distinguish the two cases regardless of whether the length is variable or not, partly to increase testing coverage and partly because it avoids generating redundant trees. Doing this means that we can also use SLP for the two-vector permute in pr88834.c, which we couldn't before VEC_PERM_EXPR nodes were introduced. The patch therefore makes pr88834.c check that we don't regress back to not using SLP and adds pr88834_ld3.c to check for the original problem in the PR. gcc/ PR tree-optimization/97513 * tree-vect-slp.c (vect_add_slp_permutation): New function, split out from... (vectorizable_slp_permutation): ...here. Detect cases in which all VEC_PERM_EXPRs are guaranteed to have the same stepped permute vector and only generate one permute vector for that case. Extend that case to handle variable-length vectors. gcc/testsuite/ * gcc.target/aarch64/sve/pr88834.c: Expect the vectorizer to use SLP. * gcc.target/aarch64/sve/pr88834_ld3.c: New test.
