ojhunt wrote: > No, ptrauth is not a feature, it is an extension and should be reported as > such. The real bug here is that you're missing an extension warning when > __ptrauth is used: https://godbolt.org/z/4Mb8f4W94 > > That code needs to issue an extension warning.
That means all code targeting arm64e/armv8.3 would produce this warning. Pointer authentication is platform ABI, so this warning would be present in _every_ build (if nothing else due to system headers) and as a result would be disabled anyway, rendering it moot. Then because `-pedantic-errors` now breaks the platform ABI, using that flag would be required to be a compile time error. I understand that the semantics of `__has_extension` might better match the mental model of the `__ptrauth` qualifier on platforms that do not support pointer auth, but on platforms where it is part of the abi it is not an extension, it's a core part of the platform and can *never* be ignored. The ODR violation argument doesn't really help: by virtue of breaking the ability to detect support for the qualifier code cannot match the ABI without requiring an external check followed by a macro define that they can test on instead, but that is not an option for the ABI definitions as a system headers can't rely on that, so -pedantic-errors breaks the ABI and causes ODR violations. i.e by requiring the ptrauth qualifier be an extension: * All code on these platforms will produce a "you're using an extension" warning, so either it gets a separate warning from the standard extension warning, which would then be disabled on all code (realistically it would be implicitly disabled for these targets), or the general "you are using an extension warning" would have to be disabled. * When targeting platforms with pointer authentication attempting to use `-pedantic-errors` (or any other flag impacting the behavior of __has_extension) would be a hard error at compile time. That would be the behavior in the llvm.org tree, every platform that supports pointer auth would in all likelihood patch this into a feature to avoid these problems: breaking code that uses -pedantic-errors isn't acceptable, and neither is producing warnings in all code targeting that platform, especially given those warnings would quite literally be a warning about the target: user chooses to target a platform that requires pointer authentication, code compiles and produces a warning that they are targeting a platform that uses pointer authentication. All code compiling with -Werror then also needs to add -Wno-<warning name> to disable the warning that they are compiling to the target they have chosen to target. https://github.com/llvm/llvm-project/pull/153291 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
