Use `load`-`store` instead of `memcpy` for short integer arrays

2 files changed, 35 insertions, 1 deletions

diff --git a/compiler/rustc_codegen_ssa/src/base.rs b/compiler/rustc_codegen_ssa/src/base.rs
index 242d209b684..dc4a28c866f 100644
--- a/compiler/rustc_codegen_ssa/src/base.rs
+++ b/compiler/rustc_codegen_ssa/src/base.rs
@@ -380,7 +380,19 @@ pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
         return;
     }
 
-    bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
+    if flags == MemFlags::empty()
+        && let Some(bty) = bx.cx().scalar_copy_backend_type(layout)
+    {
+        // I look forward to only supporting opaque pointers
+        let pty = bx.type_ptr_to(bty);
+        let src = bx.pointercast(src, pty);
+        let dst = bx.pointercast(dst, pty);
+
+        let temp = bx.load(bty, src, src_align);
+        bx.store(temp, dst, dst_align);
+    } else {
+        bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
+    }
 }
 
 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
diff --git a/compiler/rustc_codegen_ssa/src/traits/type_.rs b/compiler/rustc_codegen_ssa/src/traits/type_.rs
index 36d9864221b..e64417e1a4a 100644
--- a/compiler/rustc_codegen_ssa/src/traits/type_.rs
+++ b/compiler/rustc_codegen_ssa/src/traits/type_.rs
@@ -126,6 +126,28 @@ pub trait LayoutTypeMethods<'tcx>: Backend<'tcx> {
         index: usize,
         immediate: bool,
     ) -> Self::Type;
+
+    /// 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