https://bugs.llvm.org/show_bug.cgi?id=42049
Currently clang, in the following code, when tentatively parsing the template-argument-list recognizes the error, and skips till the semi-colon. But when annotating the broken decltype specifier token (in an effort to backtrack gracefully and continue parsing), it appears we manually sync up the cache position with the last token that we think is part of the (broken) decltype construct. Not doing so results in a book-keeping assertion failure.
void f() { g<decltype>(); }
This situation did not seem to arise prior to https://github.com/llvm/llvm-project/commit/b23c5e8c3df850177449268c5ca7dbf986157525 - which implements assumption of names as function templates even if no function with that name is found.
It seems to me that either EndLoc was miscomputed and should already point to the last token that we skipped over, in which case we should fix ParseDecltypeSpecifier to return the correct source location, or EndLoc is correct and we should revert back to the corresponding token and only annotate that portion of the token stream as the decltype specifier. (Or maybe a third option: EndLoc is meaningless when the type specifier is erroneous, in which case this code is still wrong because it's using that value in the case where backtracking is not enabled.)
Rolling back the token stream to the token corresponding to EndLoc seems likely to give us the best error recovery here, assuming that EndLoc is currently set to the token that we think is the last one that's intended to be part of the decltype specifier.