[clang] [CIR] Preserve padding bytes across a byval argument (PR #224672)

[Adam Smith via cfe-commits]

================
@@ -797,14 +806,44 @@ void insertArgCoercion(
         if (destAlloca)
           pendingParamSlots.emplace_back(destAlloca, blockArg);
       } else {
-        // byval: load the incoming pointer so the body sees a T value (and
-        // any CIRGen param-slot store becomes a local copy of that value).
+        // With byval the callee works on its own copy.  A record's padding
+        // bytes can hold another union member's live data, so its spill slot
+        // is filled by a byte copy of the incoming object.  A _BitInt keeps
+        // its load and spill store, which are its conversion between value
+        // and in-memory form.
+        cir::StoreOp paramStore;
+        if (mlir::isa<cir::RecordType>(blockArg.getType()))
+          paramStore = maybeFindParamSpill(blockArg);
+
+        // Erasing the spill before the block argument is retyped keeps the
----------------
adams381 wrote:

Four shapes reach the byval arm on x86_64: a record classified MEMORY, a `_BitInt` wider than 128 bits, an oversized vector, and `_Complex long double`.  The `_BitInt` and the vector store their whole slot, and the complex skips only x87 padding that nothing can alias.  Only in a record can the bytes a store skips be another union member's value.

https://github.com/llvm/llvm-project/pull/224672