[clang] [Clang] fix parser recovery for invalid static_assert string messages (PR #187859)

[Oleksandr Tarasiuk via cfe-commits]

================
@@ -980,7 +980,7 @@ Decl *Parser::ParseStaticAssertDeclaration(SourceLocation &DeclEnd) {
     if (getLangOpts().CPlusPlus11) {
       for (unsigned I = 0;; ++I) {
         const Token &T = GetLookAheadToken(I);
-        if (T.is(tok::r_paren))
+        if (T.isOneOf(tok::r_paren, tok::semi, tok::eof))
----------------
a-tarasyuk wrote:

@AaronBallman, thanks for the feedback. My first thought was to handle it in Sema. The parser's lookahead determines how to parse the `static_assert` message, and for malformed input like `"";)`, it seems to classify it as an expression

https://github.com/llvm/llvm-project/blob/3a56470a0ee68b26ffe93f5079de58ed22f5dc18/clang/lib/Parse/ParseDeclCXX.cpp#L979-L990

But later Sema still assumes a direct `StringLiteral` message must be unevaluated: https://github.com/llvm/llvm-project/blob/bbd69eec1d68492a2ddca3a0dfc02681a84088fe/clang/lib/Sema/SemaDeclCXX.cpp#L17962-L17969

https://github.com/llvm/llvm-project/blob/bbd69eec1d68492a2ddca3a0dfc02681a84088fe/clang/lib/Sema/SemaDeclCXX.cpp#L17791-L17793


My understanding is that the malformed input is routed through the expression path during parser recovery, but the crash occurs because `EvaluateAsString` still assumes that a direct `StringLiteral` message here must be unevaluated.  I'm wondering whether the preferred approach is to adjust the parser lookahead or to make the `static_assert` message validation in Sema more robust for recovery cases.  @AaronBallman WDYT? Thanks




https://github.com/llvm/llvm-project/pull/187859