bjoernager/rust
bjoernager/rust
1
Fork 0
Code Issues Pull requests 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
Download patch file Download diff file Expand all files Collapse all files

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
}