[llvm] [flang-rt] Optimise ShallowCopy and elemental copies in Assign (PR #140569)

[Slava Zakharin via llvm-commits]

================
@@ -492,11 +492,21 @@ RT_API_ATTRS void Assign(Descriptor &to, const Descriptor &from,
         terminator.Crash("unexpected type code %d in blank padded Assign()",
             to.type().raw());
       }
-    } else { // elemental copies, possibly with character truncation
-      for (std::size_t n{toElements}; n-- > 0;
-           to.IncrementSubscripts(toAt), from.IncrementSubscripts(fromAt)) {
-        memmoveFct(to.Element<char>(toAt), from.Element<const char>(fromAt),
-            toElementBytes);
+    } else {
+      // We can't simply call ShallowCopy due to edge cases such as character
+      // truncation or assignments where the RHS is a scalar.
+      if (toElementBytes == fromElementBytes && to.IsContiguous()) {
+        if (to.rank() == 1 && from.rank() == 1) {
+          ShallowCopyDiscontiguousToContiguous<true>(to, from);
----------------
vzakhari wrote:

Unfortunately, we cannot use `ShallowCopy` as-is here, because `memmoveFct` may not match `memcpy` (e.g. see `Assign` usese in `flang-rt/lib/cuda/memory.cpp`).

I would recommend not touching `Assign` runtime just yet.  We can start with optimizing the `ShallowCopy` and call it for the copy-in/copy-out in the compiler.

https://github.com/llvm/llvm-project/pull/140569