clang-19 cannot mangle what clang-17 mangles

[kelbon]

I updated the clang version from 17 to 19 on windows and now i get error:

error: cannot mangle this template type parameter type yet

(no additional info provided)

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (kelbon)

I updated the clang version from 17 to 19 on windows and now i get error: ``` error: cannot mangle this template type parameter type yet ``` (no additional info provided)

[zyn0217]

Without a reproducer we're also unable to tell what was broken :(

[kelbon]

Without a reproducer we're also unable to tell what was broken :(

Its really huge private file with codegen and compile time strings, i tried to minimize it, but it still really big.

if the compiler would give at least some diagnostics, for example the line number and the name of the type that is trying to mangle, it would be much easier

P.S. i checked localy, clang 18 compiles successfully. Something in clang-19 mangling breaks compilation

[efriedma-quic]

It looks like this is Windows-specific; the Itanium mangler can't generate an error message like this.

https://llvm.org/docs/HowToSubmitABug.html#front-end-bugs has some tools you can use to automatically reduce testcases, if you can't share the original. But maybe @MaxEW707 has some idea what might trigger this.

CC @rnk @zmodem

[MaxEW707]

@kelbon What C++ version are you building against. Notably pre C++20 or C++20?

I have time this weekend to look into this and try to reproduce locally based on the changes made to make the microsoft mangling code more conformant with MSVC in clang 19.

for example the line number

That is odd. The source range is passed into the error diagnostic. We must be passing in an empty source range. I'll dig into this as well.

[kelbon]

@kelbon What C++ version are you building against. Notably pre C++20 or C++20?

I have time this weekend to look into this and try to reproduce locally based on the changes made to make the microsoft mangling code more conformant with MSVC in clang 19.

for example the line number

That is odd. The source range is passed into the error diagnostic. We must be passing in an empty source range. I'll dig into this as well.

C++20.
Empty source range may be because its coroutines and some inner compiler generated code ?

This code also produces gcc-14 internal error on 'end translation unit'...

[MaxEW707]

Sounds good. If you can get a reproducer that will be greatly appreciated :).

[MaxEW707]

@kelbon I haven't been able to find a reproducer yet after going thru the commits that modified the mangling code.

Its really huge private file with codegen and compile time strings, i tried to minimize it, but it still really big.

If you are able to strip away or replace any internal private details with generic stuff such as replacing strings with random characters or renaming class names that won't leak any private details then even a large reproducer will help me track down what is going awry.