[llvm] [NVPTX] Lower allocas to the local address space (PR #204346)

[Tim Besard via llvm-commits]

maleadt wrote:

> Can you tell me more about this test suite? Is it public? If so, where can I find info on how to run it? I'm always interested in ways to improve our performance and correctness coverage of upstream NVPTX, which is right now sadly a bit lacking.

The performance test suite is in https://github.com/JuliaGPU/CUDA.jl/tree/main/perf, visualized at https://cuda.juliagpu.org/bench/. Not terribly comprehensive, and kinda specific to Julia/CUDA.jl (but it can be run with a custom `llc`), but it's served us well. Our regression test suite is _much_ more comprehensive though, if you want I could try running it with/without this change and have an LLM classify differences or so, to see if anything stands out.

https://github.com/llvm/llvm-project/pull/204346