nordic-dev.net/rust
nordic-dev.net/rust
2
0
Fork 0
mirror of https://github.com/rust-lang/rust.git synced 2024-10-30 05:51:58 +00:00
Code Issues Packages Projects Releases Wiki Activity

Rollup merge of #128149 - RalfJung:nontemporal_store, r=jieyouxu,Amanieu,Jubilee

Browse Source 
nontemporal_store: make sure that the intrinsic is truly just a hint

The `!nontemporal` flag for stores in LLVM *sounds* like it is just a hint, but actually, it is not -- at least on x86, non-temporal stores need very special treatment by the programmer or else the Rust memory model breaks down. LLVM still treats these stores as-if they were normal stores for optimizations, which is [highly dubious](https://github.com/llvm/llvm-project/issues/64521). Let's avoid all that dubiousness by making our own non-temporal stores be truly just a hint, which is possible on some targets (e.g. ARM). On all other targets, non-temporal stores become regular stores.

~~Blocked on https://github.com/rust-lang/stdarch/pull/1541 propagating to the rustc repo, to make sure the `_mm_stream` intrinsics are unaffected by this change.~~

Fixes https://github.com/rust-lang/rust/issues/114582
Cc `@Amanieu` `@workingjubilee`
This commit is contained in:
Guillaume Gomez 2024-08-12 17:09:14 +02:00 committed by GitHub
commit 095ca33bb6
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 62 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
compiler/rustc_codegen_cranelift/src/intrinsics/mod.rs
View File

@ -725,7 +725,8 @@ fn codegen_regular_intrinsic_call<'tcx>(
// Cranelift treats stores as volatile by default
// FIXME correctly handle unaligned_volatile_store
// FIXME actually do nontemporal stores if requested
// FIXME actually do nontemporal stores if requested (but do not just emit MOVNT on x86;
// see the LLVM backend for details)
let dest = CPlace::for_ptr(Pointer::new(ptr), val.layout());
dest.write_cvalue(fx, val);
}

2
compiler/rustc_codegen_gcc/src/builder.rs
View File

@ -1127,6 +1127,8 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
self.llbb().add_assignment(self.location, aligned_destination, val);
// TODO(antoyo): handle align and flags.
// NOTE: dummy value here since it's never used. FIXME(antoyo): API should not return a value here?
// When adding support for NONTEMPORAL, make sure to not just emit MOVNT on x86; see the
// LLVM backend for details.
self.cx.context.new_rvalue_zero(self.type_i32())
}

33
compiler/rustc_codegen_llvm/src/builder.rs
View File

@ -728,13 +728,32 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
llvm::LLVMSetVolatile(store, llvm::True);
}
if flags.contains(MemFlags::NONTEMPORAL) {
// According to LLVM [1] building a nontemporal store must
// *always* point to a metadata value of the integer 1.
//
// [1]: https://llvm.org/docs/LangRef.html#store-instruction
let one = self.cx.const_i32(1);
let node = llvm::LLVMMDNodeInContext(self.cx.llcx, &one, 1);
llvm::LLVMSetMetadata(store, llvm::MD_nontemporal as c_uint, node);
// Make sure that the current target architectures supports "sane" non-temporal
// stores, i.e., non-temporal stores that are equivalent to regular stores except
// for performance. LLVM doesn't seem to care about this, and will happily treat
// `!nontemporal` stores as-if they were normal stores (for reordering optimizations
// etc) even on x86, despite later lowering them to MOVNT which do *not* behave like
// regular stores but require special fences.
// So we keep a list of architectures where `!nontemporal` is known to be truly just
// a hint, and use regular stores everywhere else.
// (In the future, we could alternatively ensure that an sfence gets emitted after a sequence of movnt
// before any kind of synchronizing operation. But it's not clear how to do that with LLVM.)
// For more context, see <https://github.com/rust-lang/rust/issues/114582> and
// <https://github.com/llvm/llvm-project/issues/64521>.
const WELL_BEHAVED_NONTEMPORAL_ARCHS: &[&str] =
&["aarch64", "arm", "riscv32", "riscv64"];
let use_nontemporal =
WELL_BEHAVED_NONTEMPORAL_ARCHS.contains(&&*self.cx.tcx.sess.target.arch);
if use_nontemporal {
// According to LLVM [1] building a nontemporal store must
// *always* point to a metadata value of the integer 1.
//
// [1]: https://llvm.org/docs/LangRef.html#store-instruction
let one = self.cx.const_i32(1);
let node = llvm::LLVMMDNodeInContext(self.cx.llcx, &one, 1);
llvm::LLVMSetMetadata(store, llvm::MD_nontemporal as c_uint, node);
}
}
store
}

10
library/core/src/intrinsics.rs
View File

@ -2675,12 +2675,12 @@ extern "rust-intrinsic" {
#[rustc_nounwind]
pub fn catch_unwind(try_fn: fn(*mut u8), data: *mut u8, catch_fn: fn(*mut u8, *mut u8)) -> i32;
/// Emits a `!nontemporal` store according to LLVM (see their docs).
/// Probably will never become stable.
/// Emits a `nontemporal` store, which gives a hint to the CPU that the data should not be held
/// in cache. Except for performance, this is fully equivalent to `ptr.write(val)`.
///
/// Do NOT use this intrinsic; "nontemporal" operations do not exist in our memory model!
/// It exists to support current stdarch, but the plan is to change stdarch and remove this intrinsic.
/// See <https://github.com/rust-lang/rust/issues/114582> for some more discussion.
/// Not all architectures provide such an operation. For instance, x86 does not: while `MOVNT`
/// exists, that operation is *not* equivalent to `ptr.write(val)` (`MOVNT` writes can be reordered
/// in ways that are not allowed for regular writes).
#[rustc_nounwind]
pub fn nontemporal_store<T>(ptr: *mut T, val: T);

30
tests/codegen/intrinsics/nontemporal.rs
View File

@ -1,13 +1,37 @@
//@ compile-flags: -O
//@revisions: with_nontemporal without_nontemporal
//@[with_nontemporal] compile-flags: --target aarch64-unknown-linux-gnu
//@[with_nontemporal] needs-llvm-components: aarch64
//@[without_nontemporal] compile-flags: --target x86_64-unknown-linux-gnu
//@[without_nontemporal] needs-llvm-components: x86
#![feature(core_intrinsics)]
// Ensure that we *do* emit the `!nontemporal` flag on architectures where it
// is well-behaved, but do *not* emit it on architectures where it is ill-behaved.
// For more context, see <https://github.com/rust-lang/rust/issues/114582> and
// <https://github.com/llvm/llvm-project/issues/64521>.
#![feature(no_core, lang_items, intrinsics)]
#![no_core]
#![crate_type = "lib"]
#[lang = "sized"]
pub trait Sized {}
#[lang = "copy"]
pub trait Copy {}
impl Copy for u32 {}
impl<T> Copy for *mut T {}
extern "rust-intrinsic" {
pub fn nontemporal_store<T>(ptr: *mut T, val: T);
}
#[no_mangle]
pub fn a(a: &mut u32, b: u32) {
// CHECK-LABEL: define{{.*}}void @a
// CHECK: store i32 %b, ptr %a, align 4, !nontemporal
// with_nontemporal: store i32 %b, ptr %a, align 4, !nontemporal
// without_nontemporal-NOT: nontemporal
unsafe {
std::intrinsics::nontemporal_store(a, b);
nontemporal_store(a, b);
}
}