From 369fff6c0640fe89be9b915adaa83e66a022e00d Mon Sep 17 00:00:00 2001 From: Nikita Popov Date: Wed, 14 Feb 2024 16:26:20 +0100 Subject: [PATCH] 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. --- compiler/rustc_codegen_llvm/src/llvm_util.rs | 4 ++++ tests/ui/asm/x86_64/evex512-implicit-feature.rs | 15 +++++++++++++++ 2 files changed, 19 insertions(+) create mode 100644 tests/ui/asm/x86_64/evex512-implicit-feature.rs diff --git a/compiler/rustc_codegen_llvm/src/llvm_util.rs b/compiler/rustc_codegen_llvm/src/llvm_util.rs index e48479c8da279..54e8ed85e3250 100644 --- a/compiler/rustc_codegen_llvm/src/llvm_util.rs +++ b/compiler/rustc_codegen_llvm/src/llvm_util.rs @@ -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), } } diff --git a/tests/ui/asm/x86_64/evex512-implicit-feature.rs b/tests/ui/asm/x86_64/evex512-implicit-feature.rs new file mode 100644 index 0000000000000..a15060857eccb --- /dev/null +++ b/tests/ui/asm/x86_64/evex512-implicit-feature.rs @@ -0,0 +1,15 @@ +// build-pass +// only-x86_64 +// compile-flags: --crate-type=lib -C target-cpu=skylake + +#![feature(avx512_target_feature)] +#![feature(stdsimd)] + +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)); +}