[clang] [CIR] Upstream support for record packing and padding (PR #136036)

[Bruno Cardoso Lopes via cfe-commits]

================
@@ -229,4 +229,32 @@ def ConstPtrAttr : CIR_Attr<"ConstPtr", "ptr", [TypedAttrInterface]> {
   }];
 }
 
+//===----------------------------------------------------------------------===//
+// RecordLayoutAttr
+//===----------------------------------------------------------------------===//
+
+// Used to decouple layout information from the record type. RecordType's
+// uses this attribute to cache that information.
+
+def RecordLayoutAttr : CIR_Attr<"RecordLayout", "record_layout"> {
+  let summary = "ABI specific information about a record layout";
+  let description = [{
+    Holds layout information often queried by !cir.record users
+    during lowering passes and optimizations.
+  }];
+
+  let parameters = (ins "unsigned":$size,
+                        "unsigned":$alignment,
+                        "bool":$padded,
+                        "mlir::Type":$largest_member,
----------------
bcardosolopes wrote:

`RecordLayoutAttr` is computed lazily, and the idea is that whenever we have to walk the elements we then compute all properties, because they all require a single pass. You might be right that it's probably not too bad for `largest_member` to be done every time this query is needed, but (a) I haven't seen measurements indicating either way (and would pick the compile time defensive approach) and (b) `largest_member` makes sense along side the other properties computed here IMO.

> which for some reason diverges from CodeGen

Because we have slightly higher level unions than MLIR, this is the price to be paid. FWIW, @gitoleg has done a lot of work to make unions work, including often running csmith and fixing correctness issues.

https://github.com/llvm/llvm-project/pull/136036