[RFC] improvements to LLVM diagnostic infrastructure

Hi all,

I’m working on improvements to diagnostics handling in LLVM, specifically for the benefit of the (integrated) assembler. The goal is to support options such as -Werror, -w, and -W for files assembled with clang and inline assembly. Clang already has support for these options but does not apply them to diagnostics originating from (inline) assembly.

I plan to add an LLVMDiagnosticsEngine class that takes responsibility for handling diagnostics options (superseding parts of SourceMgr and LLVMContext). The diagnostics themselves will be defined in TableGen (just like in clang).

Currently, diagnostics are passed through a SourceMgr to get location info and then passed on to a diagnostics handler for printing (if it exists). This is usually wrapped in Warning() and Error() functions.

For example in ARMAsmParser.cpp:

Error(ImmLoc, “invalid immediate shift value”);
which eventually calls SourceMgr::PrintMessage.

This would change to (incrementally throughout the code-base) :
Diag(ImmLoc, err_arm_invalid_immediate_shift);
and pass through LLVMDiagnosticsEngine to get the diagnostic message and severity (defined in TableGen, similar to Diagnostic*Kinds.td in clang).

Initially, LLVMDiagnosticsEngine will be a smaller, simpler version of clang’s DiagnosticEngine. Over time, I would expect more features of clang’s diagnostics to migrate to LLVM, where possible, to improve diagnostics for all the LLVM tools.

I intend to implement the necessary infrastructure and convert a (small) number of diagnostics. Further conversions can be done incrementally.

Any comments on this approach? Is this the right direction for diagnostics in LLVM? Suggestions more than welcome!

Thanks,
Sanne

Hi Sanne,

Hi all,

I’m working on improvements to diagnostics handling in LLVM, specifically for the benefit of the (integrated) assembler. The goal is to support options such as -Werror, -w, and -W for files assembled with clang and inline assembly. Clang already has support for these options but does not apply them to diagnostics originating from (inline) assembly.

I am not sure I get what you want to do.
That infrastructure already exists as far as I can tell. Look at BackendConsumer::DiagnosticHandlerImpl.

For historical reasons, it is possible the inline assembly stuff does not use it though.

Cheers,
-Quentin

Hi Quentin,

Thank you for the pointer to BackendConsumer::DiagnosticsHandlerImpl. I’ll have to investigate if it would fulfill my requirements.

If so, I’d be looking to convert the assembler to use it (and improving it where necessary). Do you happen to know the reasons why the assembler isn’t using this bit of LLVM?

Thanks!
Sanne

The assembler predates it. -Hal