[clang] [clang] Add clang::debug_info_type attribute for bitfields (PR #69104)

[Vlad Serebrennikov via cfe-commits]

================
@@ -5910,6 +5910,30 @@ static void handleBuiltinAliasAttr(Sema &S, Decl *D,
   D->addAttr(::new (S.Context) BuiltinAliasAttr(S.Context, AL, Ident));
 }
 
+static void handleDebugInfoTypeAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
+  if (!AL.hasParsedType()) {
+    S.Diag(AL.getLoc(), diag::err_attribute_wrong_number_arguments) << AL << 1;
+    return;
+  }
+
+  TypeSourceInfo *ParmTSI = nullptr;
+  QualType type = S.GetTypeFromParser(AL.getTypeArg(), &ParmTSI);
+  assert(ParmTSI && "no type source info for attribute argument");
+
+  if (type->isEnumeralType()) {
+    QualType BitfieldType = llvm::cast<FieldDecl>(D)->getType();
+    QualType EnumUnderlyingType =
+        type->getAs<EnumType>()->getDecl()->getIntegerType();
+    if (EnumUnderlyingType != BitfieldType) {
----------------
Endilll wrote:

> and only prevent situations where the specified type is not an enumeration or integral type.

> This ALSO makes me wonder if forcing it to be an enum type is necessary.

I intentionally haven't been restricting the set of types that can be passed as the argument to `preferred_type`. This makes even more sense now that we are going with more generic name. I'm just making sure that enum-related diagnostics are triggered only when enum was passed by the user.

> Another case we should consider would be whether we want to allow signed unsigned mismatches:

> I think in the signed/unsigned mismatch, we should allow it (thanks to the layout rules)

> 1 enum that can represent a negative value

I disagree. We shouldn't pretend that our enums can hold negative values, while we store them in unsigned bit-fields (save for rare exceptions, if any). Now that we are not bound by C++98 compilers, I believe we should propagate underlying type of our enums from bit-field type they are stored in (usually `unsigned`, `uint64_t` for `DeclBase.h`), and deal with negative values at enumerator declaration (I guess we'll have to explicitly cast them to the underlying type). I can do that while adding `clang::preferred_type` to bit-fields.

Ideally I'd propagate bit-field width as well, but that doesn't seem possible at the moment: https://github.com/llvm/llvm-project/pull/69104#issuecomment-1770189229

https://github.com/llvm/llvm-project/pull/69104