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

Author: RalfJung

compiler/rustc_codegen_cranelift/src/intrinsics/mod.rs

@@ -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);
         }

compiler/rustc_codegen_gcc/src/builder.rs

@@ -1128,6 +1128,8 @@ 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())
     }
 

compiler/rustc_codegen_llvm/src/builder.rs

@@ -725,13 +725,29 @@ 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.)
+                const WELL_BEHAVED_NONTEMPORAL_ARCHS: &[&str] = &["aarch64", "arm"];
+
+                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
         }

library/core/src/intrinsics.rs

@@ -2405,12 +2405,12 @@ 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);
 

tests/codegen/intrinsics/nontemporal.rs

@@ -1,13 +1,28 @@
 //@ compile-flags: -O
+//@ compile-flags: --target aarch64-unknown-linux-gnu
+//@ needs-llvm-components: aarch64
 
-#![feature(core_intrinsics)]
+#![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
     unsafe {
-        std::intrinsics::nontemporal_store(a, b);
+        nontemporal_store(a, b);
     }
 }