user0/rust
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Allow vector types for amdgpu

Browse source 
The amdgpu target uses vector types in various places. The vector types
can be used on all architectures, there is no associated target feature
that needs to be enabled.

The largest vector type found in LLVM intrinsics is `v32i32`
(`[32 x i32]`) for mfma intrinsics. Note that while this intrinsic is
only supported on some architectures, the vector type itself is
supported on all architectures.
This commit is contained in:
Flakebi 2025-12-11 02:40:09 +01:00
parent 08de25c4ea
commit a9b147259b
No known key found for this signature in database
GPG key ID: 38E7ED984D7DCD02
3 changed files with 18 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

2
compiler/rustc_monomorphize/src/mono_checks/abi_check.rs
View file

@ -54,7 +54,7 @@ fn do_check_simd_vector_abi<'tcx>(
continue;
}
};
if !have_feature(Symbol::intern(feature)) {
if !feature.is_empty() && !have_feature(Symbol::intern(feature)) {
// Emit error.
let (span, _hir_id) = loc();
tcx.dcx().emit_err(errors::AbiErrorDisabledVectorType {

5
compiler/rustc_target/src/target_features.rs
View file

@ -918,6 +918,7 @@ const AARCH64_FEATURES_FOR_CORRECT_VECTOR_ABI: &'static [(u64, &'static str)] =
// We might want to add "helium" too.
const ARM_FEATURES_FOR_CORRECT_VECTOR_ABI: &'static [(u64, &'static str)] = &[(128, "neon")];
const AMDGPU_FEATURES_FOR_CORRECT_VECTOR_ABI: &'static [(u64, &'static str)] = &[(1024, "")];
const POWERPC_FEATURES_FOR_CORRECT_VECTOR_ABI: &'static [(u64, &'static str)] = &[(128, "altivec")];
const WASM_FEATURES_FOR_CORRECT_VECTOR_ABI: &'static [(u64, &'static str)] = &[(128, "simd128")];
const S390X_FEATURES_FOR_CORRECT_VECTOR_ABI: &'static [(u64, &'static str)] = &[(128, "vector")];
@ -996,12 +997,12 @@ impl Target {
Arch::Mips | Arch::Mips32r6 | Arch::Mips64 | Arch::Mips64r6 => {
MIPS_FEATURES_FOR_CORRECT_VECTOR_ABI
}
Arch::AmdGpu => AMDGPU_FEATURES_FOR_CORRECT_VECTOR_ABI,
Arch::Nvptx64 | Arch::Bpf | Arch::M68k => &[], // no vector ABI
Arch::CSky => CSKY_FEATURES_FOR_CORRECT_VECTOR_ABI,
// FIXME: for some tier3 targets, we are overly cautious and always give warnings
// when passing args in vector registers.
Arch::AmdGpu
| Arch::Avr
Arch::Avr
| Arch::Msp430
| Arch::PowerPC64LE
| Arch::SpirV

21
tests/run-make/simd-ffi/rmake.rs
View file

@ -17,6 +17,9 @@ fn main() {
"arm-unknown-linux-gnueabi".to_owned(),
]);
}
if llvm_components_contain("amdgpu") {
targets.push("amdgcn-amd-amdhsa".to_owned());
}
let mut x86_archs = Vec::new();
if llvm_components_contain("x86") {
x86_archs.append(&mut vec!["i686", "x86_64"]);
@ -52,21 +55,25 @@ fn main() {
// enabled by-default for i686 and ARM; these features will be invalid
// on some platforms, but LLVM just prints a warning so that's fine for
// now.
let mut cmd = rustc();
cmd.target(&target).emit("llvm-ir,asm").input("simd.rs");
let target_feature = if target.starts_with("i686") || target.starts_with("x86") {
"+sse2"
} else if target.starts_with("arm") || target.starts_with("aarch64") {
"-soft-float,+neon"
} else if target.starts_with("mips") {
"+msa,+fp64"
} else if target.starts_with("amdgcn") {
cmd.arg("-Ctarget-cpu=gfx900");
""
} else {
panic!("missing target_feature case for {target}");
};
rustc()
.target(&target)
.emit("llvm-ir,asm")
.input("simd.rs")
.arg(format!("-Ctarget-feature={target_feature}"))
.arg(&format!("-Cextra-filename=-{target}"))
.run();
if !target_feature.is_empty() {
cmd.arg(format!("-Ctarget-feature={target_feature}"));
}
cmd.arg(&format!("-Cextra-filename=-{target}")).run();
}
}