[clang] [OpenMP][CodeGen] Preserve libcalls for declare simd redecls (PR #193881)
via cfe-commits
cfe-commits at lists.llvm.org
Sat Oct 31 07:00:04 PDT 2026
anun333 wrote:
Rechecked at 932278b with #193884, same setup: x86-64-v3 now calls `_ZGVdN4v_sin`, and x86-64-v4 with
`-mprefer-vector-width=256` offers and calls only `_ZGVd` (without the flag the float-to-double loop still calls
`_ZGVeN8v_sin` there, which an AVX-512 target can run). No assertion, and every loop gives correct results.
On where the general fix goes: I think the vectorizer is the right place for the TLI side. With `-fveclib=libmvec`
the mappings come from the TLI table rather than from clang, so neither of your filters applies there; that's
#162239 (veclib entries not checked against the target's features), and #223817 holds back its `_ZGVd` rows until it
is fixed. An ISA-class and preferred-width check where the vectorizer picks a variant would cover both routes.
Keeping the clang-side filter in this PR seems fine to me, since it already makes these mappings correct.
(Tested with AI assistance (Claude); I checked the results before posting.)
— Claude
https://github.com/llvm/llvm-project/pull/193881
More information about the cfe-commits
mailing list