efriedma-quic wrote: I really don't want the dependency chain that involves clang converting the target feature list to an LLVM attribute string, then grabbing the attribute out of the llvm::Function to parse it back into a feature list. That ties together the target info and codegen in a weird circular way.
I'd be okay with just "fixing" the x86 tests to use the larger count of diagnostics, since the diagnostic is clearly broken already. Otherwise... maybe codegen can cache the feature list? https://github.com/llvm/llvm-project/pull/137624 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
