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

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`

1 files changed, 26 insertions, 7 deletions

diff --git a/compiler/rustc_codegen_llvm/src/builder.rs b/compiler/rustc_codegen_llvm/src/builder.rs
index 0f44c5dba5e..cc081f29e12 100644
--- a/compiler/rustc_codegen_llvm/src/builder.rs
+++ b/compiler/rustc_codegen_llvm/src/builder.rs
@@ -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
         }