Auto merge of #127168 - DianQK:cast-size, r=workingjubilee

Use the aligned size for alloca at args/ret when the pass mode is cast

Fixes #75839. Fixes #121028.

The `load` and `store` instructions in LLVM access the aligned size. For example, `load { i64, i32 }` accesses 16 bytes on x86_64: https://alive2.llvm.org/ce/z/n8CHAp.

BTW, this example is expected to be optimized to immediate UB by Alive2: https://rust.godbolt.org/z/b7xK7hv1c and https://alive2.llvm.org/ce/z/vZDtZH.

r? compiler

Author: bors

compiler/rustc_codegen_llvm/src/abi.rs

@@ -226,7 +226,8 @@ impl<'ll, 'tcx> ArgAbiExt<'ll, 'tcx> for ArgAbi<'tcx, Ty<'tcx>> {
                 //   when passed by value, making it smaller.
                 // - On some ABIs, the Rust layout { u16, u16, u16 } may be padded up to 8 bytes
                 //   when passed by value, making it larger.
-                let copy_bytes = cmp::min(scratch_size.bytes(), self.layout.size.bytes());
+                let copy_bytes =
+                    cmp::min(cast.unaligned_size(bx).bytes(), self.layout.size.bytes());
                 // Allocate some scratch space...
                 let llscratch = bx.alloca(scratch_size, scratch_align);
                 bx.lifetime_start(llscratch, scratch_size);

compiler/rustc_codegen_ssa/src/mir/block.rs

@@ -1521,7 +1521,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 //   when passed by value, making it smaller.
                 // - On some ABIs, the Rust layout { u16, u16, u16 } may be padded up to 8 bytes
                 //   when passed by value, making it larger.
-                let copy_bytes = cmp::min(scratch_size.bytes(), arg.layout.size.bytes());
+                let copy_bytes = cmp::min(cast.unaligned_size(bx).bytes(), arg.layout.size.bytes());
                 // Allocate some scratch space...
                 let llscratch = bx.alloca(scratch_size, scratch_align);
                 bx.lifetime_start(llscratch, scratch_size);

compiler/rustc_codegen_ssa/src/mir/mod.rs

@@ -230,10 +230,20 @@ pub fn codegen_mir<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
             let layout = start_bx.layout_of(fx.monomorphize(decl.ty));
             assert!(!layout.ty.has_erasable_regions());
 
-            if local == mir::RETURN_PLACE && fx.fn_abi.ret.is_indirect() {
-                debug!("alloc: {:?} (return place) -> place", local);
-                let llretptr = start_bx.get_param(0);
-                return LocalRef::Place(PlaceRef::new_sized(llretptr, layout));
+            if local == mir::RETURN_PLACE {
+                match fx.fn_abi.ret.mode {
+                    PassMode::Indirect { .. } => {
+                        debug!("alloc: {:?} (return place) -> place", local);
+                        let llretptr = start_bx.get_param(0);
+                        return LocalRef::Place(PlaceRef::new_sized(llretptr, layout));
+                    }
+                    PassMode::Cast { ref cast, .. } => {
+                        debug!("alloc: {:?} (return place) -> place", local);
+                        let size = cast.size(&start_bx);
+                        return LocalRef::Place(PlaceRef::alloca_size(&mut start_bx, size, layout));
+                    }
+                    _ => {}
+                };
             }
 
             if memory_locals.contains(local) {

compiler/rustc_codegen_ssa/src/mir/place.rs

@@ -108,9 +108,17 @@ impl<'a, 'tcx, V: CodegenObject> PlaceRef<'tcx, V> {
     pub fn alloca<Bx: BuilderMethods<'a, 'tcx, Value = V>>(
         bx: &mut Bx,
         layout: TyAndLayout<'tcx>,
+    ) -> Self {
+        Self::alloca_size(bx, layout.size, layout)
+    }
+
+    pub fn alloca_size<Bx: BuilderMethods<'a, 'tcx, Value = V>>(
+        bx: &mut Bx,
+        size: Size,
+        layout: TyAndLayout<'tcx>,
     ) -> Self {
         assert!(layout.is_sized(), "tried to statically allocate unsized place");
-        PlaceValue::alloca(bx, layout.size, layout.align.abi).with_type(layout)
+        PlaceValue::alloca(bx, size, layout.align.abi).with_type(layout)
     }
 
     /// Returns a place for an indirect reference to an unsized place.

compiler/rustc_target/src/abi/call/mod.rs

@@ -339,7 +339,9 @@ impl CastTarget {
         }
     }
 
-    pub fn size<C: HasDataLayout>(&self, _cx: &C) -> Size {
+    /// When you only access the range containing valid data, you can use this unaligned size;
+    /// otherwise, use the safer `size` method.
+    pub fn unaligned_size<C: HasDataLayout>(&self, _cx: &C) -> Size {
         // Prefix arguments are passed in specific designated registers
         let prefix_size = self
             .prefix
@@ -353,6 +355,10 @@ impl CastTarget {
         prefix_size + rest_size
     }
 
+    pub fn size<C: HasDataLayout>(&self, cx: &C) -> Size {
+        self.unaligned_size(cx).align_to(self.align(cx))
+    }
+
     pub fn align<C: HasDataLayout>(&self, cx: &C) -> Align {
         self.prefix
             .iter()