krzysz00 wrote:

To elaborate, my claim is that the table that maps stages to particular 
instruction classes should not exist, and that that's an awkward design that 
paints us into corners.

The stage numbers should be arbitrarily assigned by the user, and we can and 
should *infer* which classes of instructions to skip waiting on in a given 
stage.

There might also be a need for something noop-y to declare stages or have it be 
some sort of token, I'm not 100% sure on the design here, but in general, I 
think we can get away without a table that categorizes the asyncmark-able 
instructions.

Much of the point of asyncmark is to be an abstraction over these hardware wait 
counters, so I'd like to see this "stages" system continue in that vein.

https://github.com/llvm/llvm-project/pull/220442
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to