Remove my `scalar_copy_backend_type` optimization attempt

I added this back in 111999, but I no longer think it's a good idea
- It had to get scaled back to only power-of-two things to not break a bunch of targets
- LLVM seems to be getting better at memcpy removal anyway
- Introducing vector instructions has seemed to sometimes (115515) make autovectorization worse

So this removes it from the codegen crates entirely, and instead just tries to use <https://doc.rust-lang.org/nightly/nightly-rustc/rustc_codegen_ssa/traits/builder/trait.BuilderMethods.html#method.typed_place_copy> instead of direct `memcpy` so things will still use load/store for immediates.

2 files changed, 18 insertions, 26 deletions

diff --git a/compiler/rustc_codegen_ssa/src/traits/builder.rs b/compiler/rustc_codegen_ssa/src/traits/builder.rs
index 6c8dcc5b690..c9d40b7c724 100644
--- a/compiler/rustc_codegen_ssa/src/traits/builder.rs
+++ b/compiler/rustc_codegen_ssa/src/traits/builder.rs
@@ -281,17 +281,31 @@ pub trait BuilderMethods<'a, 'tcx>:
         dst: PlaceRef<'tcx, Self::Value>,
         src: PlaceRef<'tcx, Self::Value>,
     ) {
-        debug_assert!(src.llextra.is_none());
-        debug_assert!(dst.llextra.is_none());
+        self.typed_place_copy_with_flags(dst, src, MemFlags::empty());
+    }
+
+    fn typed_place_copy_with_flags(
+        &mut self,
+        dst: PlaceRef<'tcx, Self::Value>,
+        src: PlaceRef<'tcx, Self::Value>,
+        flags: MemFlags,
+    ) {
+        debug_assert!(src.llextra.is_none(), "cannot directly copy from unsized values");
+        debug_assert!(dst.llextra.is_none(), "cannot directly copy into unsized values");
         debug_assert_eq!(dst.layout.size, src.layout.size);
         if self.sess().opts.optimize == OptLevel::No && self.is_backend_immediate(dst.layout) {
             // If we're not optimizing, the aliasing information from `memcpy`
             // isn't useful, so just load-store the value for smaller code.
             let temp = self.load_operand(src);
-            temp.val.store(self, dst);
+            temp.val.store_with_flags(self, dst, flags);
+        } else if flags.contains(MemFlags::NONTEMPORAL) {
+            // HACK(nox): This is inefficient but there is no nontemporal memcpy.
+            let ty = self.backend_type(dst.layout);
+            let val = self.load(ty, src.llval, src.align);
+            self.store_with_flags(val, dst.llval, dst.align, flags);
         } else if !dst.layout.is_zst() {
             let bytes = self.const_usize(dst.layout.size.bytes());
-            self.memcpy(dst.llval, dst.align, src.llval, src.align, bytes, MemFlags::empty());
+            self.memcpy(dst.llval, dst.align, src.llval, src.align, bytes, flags);
         }
     }
 
diff --git a/compiler/rustc_codegen_ssa/src/traits/type_.rs b/compiler/rustc_codegen_ssa/src/traits/type_.rs
index 555833759eb..34d9e75036f 100644
--- a/compiler/rustc_codegen_ssa/src/traits/type_.rs
+++ b/compiler/rustc_codegen_ssa/src/traits/type_.rs
@@ -133,28 +133,6 @@ pub trait LayoutTypeMethods<'tcx>: Backend<'tcx> {
             || self.is_backend_immediate(layout)
             || self.is_backend_scalar_pair(layout))
     }
-
-    /// A type that can be used in a [`super::BuilderMethods::load`] +
-    /// [`super::BuilderMethods::store`] pair to implement a *typed* copy,
-    /// such as a MIR `*_0 = *_1`.
-    ///
-    /// It's always legal to return `None` here, as the provided impl does,
-    /// in which case callers should use [`super::BuilderMethods::memcpy`]
-    /// instead of the `load`+`store` pair.
-    ///
-    /// This can be helpful for things like arrays, where the LLVM backend type
-    /// `[3 x i16]` optimizes to three separate loads and stores, but it can
-    /// instead be copied via an `i48` that stays as the single `load`+`store`.
-    /// (As of 2023-05 LLVM cannot necessarily optimize away a `memcpy` in these
-    /// cases, due to `poison` handling, but in codegen we have more information
-    /// about the type invariants, so can emit something better instead.)
-    ///
-    /// This *should* return `None` for particularly-large types, where leaving
-    /// the `memcpy` may well be important to avoid code size explosion.
-    fn scalar_copy_backend_type(&self, layout: TyAndLayout<'tcx>) -> Option<Self::Type> {
-        let _ = layout;
-        None
-    }
 }
 
 // For backends that support CFI using type membership (i.e., testing whether a given pointer is