Don't force many scalar pair values to the stack in write_cvalue_maybe_transmute

Sometimes it is necessary for handling vector to scalar pair transmutes,
but if the types are the same there is no need for this.

This improves runtime performance on simple-raytracer by 12%.

Author: bjorn3

src/abi/comments.rs

@@ -98,12 +98,12 @@ pub(super) fn add_local_place_comments<'tcx>(
         }
         CPlaceInner::VarPair(place_local, var1, var2) => {
             assert_eq!(local, place_local);
-            ("ssa", Cow::Owned(format!(",var=({}, {})", var1.index(), var2.index())))
+            ("ssa", Cow::Owned(format!("var=({}, {})", var1.index(), var2.index())))
         }
         CPlaceInner::VarLane(_local, _var, _lane) => unreachable!(),
         CPlaceInner::Addr(ptr, meta) => {
             let meta = if let Some(meta) = meta {
-                Cow::Owned(format!(",meta={}", meta))
+                Cow::Owned(format!("meta={}", meta))
             } else {
                 Cow::Borrowed("")
             };

src/value_and_place.rs

@@ -588,10 +588,13 @@ impl<'tcx> CPlace<'tcx> {
                 return;
             }
             CPlaceInner::VarPair(_local, var1, var2) => {
-                let (ptr, meta) = from.force_stack(fx);
-                assert!(meta.is_none());
-                let (data1, data2) =
-                    CValue(CValueInner::ByRef(ptr, None), dst_layout).load_scalar_pair(fx);
+                let (data1, data2) = if self.layout().ty == dst_layout.ty {
+                    CValue(from.0, dst_layout).load_scalar_pair(fx)
+                } else {
+                    let (ptr, meta) = from.force_stack(fx);
+                    assert!(meta.is_none());
+                    CValue(CValueInner::ByRef(ptr, None), dst_layout).load_scalar_pair(fx)
+                };
                 let (dst_ty1, dst_ty2) = fx.clif_pair_type(self.layout().ty).unwrap();
                 transmute_value(fx, var1, data1, dst_ty1);
                 transmute_value(fx, var2, data2, dst_ty2);