https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118154
--- Comment #7 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Robin Dapp <rd...@gcc.gnu.org>: https://gcc.gnu.org/g:e5e9e50fc6816713d012f1d96ae308a0946d5a14 commit r15-6892-ge5e9e50fc6816713d012f1d96ae308a0946d5a14 Author: Robin Dapp <rd...@ventanamicro.com> Date: Thu Jan 9 20:45:10 2025 +0100 RISC-V: Fix vsetvl compatibility predicate [PR118154]. In PR118154 we emit strided stores but the first of those does not always have the proper VTYPE. That's because we erroneously delete a necessary vsetvl. In order to determine whether to elide (1) Expr[7]: VALID (insn 116, bb 17) Demand fields: demand_ratio_and_ge_sew demand_avl SEW=8, VLMUL=mf2, RATIO=16, MAX_SEW=64 TAIL_POLICY=agnostic, MASK_POLICY=agnostic AVL=(reg:DI 0 zero) when e.g. (2) Expr[3]: VALID (insn 360, bb 15) Demand fields: demand_sew_lmul demand_avl SEW=64, VLMUL=m1, RATIO=64, MAX_SEW=64 TAIL_POLICY=agnostic, MASK_POLICY=agnostic AVL=(reg:DI 0 zero) VL=(reg:DI 13 a3 [345]) is already available, we use sew_ge_and_prev_sew_le_next_max_sew_and_next_ratio_valid_for_prev_sew_p. (1) requires RATIO = SEW/LMUL = 16 and an SEW >= 8. (2) has ratio = 64, though, so we cannot directly elide (1). This patch uses ratio_eq_p instead of next_ratio_valid_for_prev_sew_p. PR target/118154 gcc/ChangeLog: * config/riscv/riscv-vsetvl.cc (MAX_LMUL): New define. (pre_vsetvl::earliest_fuse_vsetvl_info): Use. (pre_vsetvl::pre_global_vsetvl_info): New predicate with equal ratio. * config/riscv/riscv-vsetvl.def: Use. gcc/testsuite/ChangeLog: * gcc.target/riscv/rvv/autovec/pr118154-1.c: New test. * gcc.target/riscv/rvv/autovec/pr118154-2.c: New test.
