Implicitly enable evex512 if avx512 is enabled

LLVM 18 requires the evex512 feature to allow use of zmm registers.
LLVM automatically sets it when using a generic CPU, but not when
`-C target-cpu` is specified. This will result either in backend
legalization crashes, or code unexpectedly using ymm instead of
zmm registers.

For now, make sure that `avx512*` features imply `evex512`. Long
term we'll probably have to deal with the AVX10 mess somehow.
This commit is contained in:
Nikita Popov 2024-02-14 16:26:20 +01:00
parent cc1c0990ab
commit 369fff6c06
2 changed files with 19 additions and 0 deletions

View file

@ -266,6 +266,10 @@ pub fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> LLVMFeature<'a> {
("riscv32" | "riscv64", "fast-unaligned-access") if get_version().0 <= 17 => {
LLVMFeature::new("unaligned-scalar-mem")
}
// For LLVM 18, enable the evex512 target feature if a avx512 target feature is enabled.
("x86", s) if get_version().0 >= 18 && s.starts_with("avx512") => {
LLVMFeature::with_dependency(s, TargetFeatureFoldStrength::EnableOnly("evex512"))
}
(_, s) => LLVMFeature::new(s),
}
}

View file

@ -0,0 +1,15 @@
// build-pass
// only-x86_64
// compile-flags: --crate-type=lib -C target-cpu=skylake
#![feature(avx512_target_feature)]
#![feature(stdarch_x86_avx512)]
use std::arch::x86_64::*;
#[target_feature(enable = "avx512f")]
#[no_mangle]
pub unsafe fn test(res: *mut f64, p: *const f64) {
let arg = _mm512_load_pd(p);
_mm512_store_pd(res, _mm512_fmaddsub_pd(arg, arg, arg));
}