On Tue, Aug 10, 2021 at 04:14:15PM +0200, Richard Biener via Gcc-patches wrote:
> OK, so gimple_builtin_call_types_compatible_p only checks that the
> call is consistent with the fndecl type - iff the declaration is incompatible
> with the declaration as specified by builtins.def then that's of course
> not detected by that check (this check is to catch cases where a
> formerly indirect call through an incompatible type is now known to
> be to a builtin).
>
> IIRC that is a recurring issue and indeed my opinion is that frontends
> should not mark function decls as BUILT_IN if the definition/declaration
> signature is not compatible.
Different FEs use different strictness for what is or is not compatible.
And we can't be too strict, because e.g. of FILE * arguments where the
FILE type isn't pre-declared for the builtins.
On severe mismatches I think the FEs already don't mark it built in
(say double vs. int etc.), and e.g. const char * vs. char * differences
should be allowed (e.g. consider C++ vs. C strchr).
But perhaps we should consider as incompatible somethng that doesn't
pass useless_type_conversion_p too...
Jakub
