Martin Krastev

Issue reproduces with `llvm-9` and `llvm-10` as well.

Yep, `TotalAutoConstantSize` clearly fails its purpose here, but I didn't investigate how and why. I just wanted to see this one kernel up and running as part of new hw...

I think I know what the problem is, but I don't know yet how to solve it. The crux of the issue is that `fixConstantMemArgs` gets called *before* all auto...

It does seem like a chicken-n-egg problem indeed. One thing that puzzles me is that missing-symbol error when `fixConstantMemArgs` comes last -- I can't figure out what that symbol is,...

Ok, I got the problem with the reordered case with some help from friends -- it was an overlook on my end all along. The reported `CUDA_ERROR_NOT_FOUND` is actually *not*...

Keeping `addKernelAnnotations` invocation after the `fix*Args` bunch does fix the `CUDA_ERROR_NOT_FOUND`, but introduces warnings ``` ptxas warning : Unresolved extern variable '_global_offset_x' in whole program compilation, ignoring extern qualifier ptxas...