[clang] [BoundsSafety][NFC] Add the count-refill logic for late-parsed bounds attributes (PR #224556)

[Yeoul Na via cfe-commits]

================
@@ -4896,6 +4896,54 @@ void Parser::ParseLexedTypeAttribute(LateParsedTypeAttribute &LA,
   OutAttrs.takeAllAppendingFrom(Attrs);
 }
 
+void Parser::CompleteLateParsedTypeAttributes(
+    SmallVectorImpl<LateParsedTypeAttribute *> &LateTypeAttrs) {
+  for (LateParsedTypeAttribute *LTA : LateTypeAttrs) {
+    // Read these out before parsing, which destroys the attribute. The type is
+    // null if construction rejected the attribute, in which case the diagnostic
+    // has already been emitted and there is nothing to complete.
+    BoundsAttributedType *BATy = LTA->TypeToComplete;
+    // Rejected during construction (already diagnosed); the cached tokens are
+    // self-contained, so there is nothing to drain — just discard it.
+    if (!BATy) {
+      delete LTA;
+      continue;
+    }
+    // The fields were
+    // attached in ParseStructDeclaration as each declarator was completed; more
+    // than one appears when several declarators share a
+    // declaration-specifier-position attribute.
+    SmallVector<Decl *, 2> Fields(LTA->Decls);
+
+    AttributeFactory AF;
+    ParsedAttributes Attrs(AF);
+    ParseLexedTypeAttribute(*LTA, Attrs);
+    delete LTA;
+
+    // An unparseable argument leaves no attribute behind; already diagnosed.
+    if (Attrs.empty())
+      continue;
+    assert(Attrs.size() == 1);
+
+    Expr *Arg = Attrs[0].getArgAsExpr(0);
+    assert(Arg);
+
+    // No field means the attribute never reached a field declarator (for
+    // instance the type was rejected during construction, which unwraps the
----------------
rapidsna wrote:

That's fair. I've turned the cases that shouldn't happen into assertions. Two checks remain, both reachable from invalid code and already diagnosed:
- [R4905](https://github.com/llvm/llvm-project/pull/224556/changes#diff-bf37b284dfa0a76af702a5493cde01dec7f88619533c1be15e74654862299770R4905): Here, TypeToComplete is null when ActOnLateParsedTypeAttr rejected the attribute's type.
- [R4915](https://github.com/llvm/llvm-project/pull/224556/changes#diff-bf37b284dfa0a76af702a5493cde01dec7f88619533c1be15e74654862299770R4915): Here, ParseLexedTypeAttribute returns Attrs empty if parsing the attribute was rejected within the function. An alternative is to return an error condition from ParseLexedTypeAttribute, but it just followed the ParseLexedAttribute family's convention.

https://github.com/llvm/llvm-project/pull/224556