MaskRay wrote:

> Code LG and we can do this short/medium-term, though I'm somewhat unsure 
> about the long-term direction here.
> 
> So far library options have in common that they are always available (when 
> the library is linked). Tool options (MC, CodeGen) need to opt in by the 
> tool. I'm wondering whether a better approach would be for the tool to 
> include MCToolOptions.td itself and use it for its own option parser (once 
> also migrated away from cl::opt) instead of deferred dynamic registration 
> with the library options.
> 
> Right now library options register themselves using global ctors, which I 
> think is not an ideal end state: they still require 50-100 ctors and they 
> still need to dynamically add all options to a hash map before parsing 
> options, which is expensive. Given that the constructed hash table is the 
> same in almost all cases, we could in the end generate the hash table at 
> build time.

If we want to drop the process-wide registry `RegisterMCToolOptions`, we can 
add a new cl::parseCommandLineOptions overload with a `ToolOptions` parameter 
to indicate the tool options (prototype: #240009, which touches 32 files, so it 
stays separate). 

https://github.com/llvm/llvm-project/pull/239920
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits

Reply via email to