[PATCH 1/1] libstdc++: avoid bf16 literal in format arg normalization

[Jonathan Wakely]

On Mon, 7 Sept 2026 at 03:14, Jakit Liang <jakitliang@gmail.com> wrote:
>
> Hi Jonathan,
>
> Thanks for the review.
>
> There is nothing inherently wrong with the `0.0bf16` literal when
> compiled by GCC. The issue is that it is only used here to obtain the
> bfloat16 type, but exposing this GCC-specific literal in the libstdc++
> header causes problems for Clang-based tooling such as clangd.

This information needs to be in the commit message.

>
> For example, when clangd parses and instantiates `_S_to_arg_type` for
> an ordinary `std::format` call, it can produce:
>
> ```
> 'auto' in return type deduced as
> 'type_identity<const char *>'
> but deduced as 'type_identity<float>' in an earlier return statement
> ```
>
> GCC itself compiles the code correctly. Using
> `__gnu_cxx::__bfloat16_t` directly expresses the same type check
> without requiring the parser to handle the bf16 literal, so the
> bfloat16-to-float normalization is unchanged.

How does clangd handle the definition of the __bfloat16_t typedef,
which uses the literal?

>
> I think this is useful on the gcc-14 release branch because GCC 14's
> libstdc++ headers are consumed by Clang/clangd tooling, and this
> results in diagnostics for otherwise valid `std::format` code. The
> change is limited to the internal type check and does not change ABI
> or runtime behaviour.
>
> I'll also avoid sending a `[PATCH 0/1]` cover letter for a single
> patch and put the explanation in the commit message.
>
> Thanks,
> Jakit