[llvm] [NVPTX] Legalize aext-load to zext-load to expose more DAG combines (PR #154251)

[Joseph Huber via llvm-commits]

jhuber6 wrote:

> > I bisected some new `libc` failures to this PR. After reverting it and rebuilding the tests all pass https://lab.llvm.org/buildbot/#/builders/69/builds/26327/steps/12/logs/stdio. Should we go ahead and revert this for now?
> 
> @jhuber6 would you be able to attach any more information such as LLVM IR or PTX before/after? or reproduction instructions? If this is possible to reproduce I'm happy to take a look right now. If you think investigation will take a long time I'd be happy to revert as well.

It caused a handful of tests to fail, you should be able to build and run `libc` tests locally if you have a working CUDA installation + GPU. I spent over forty minutes trying and failing to get a reproducer because the NVPTX toolchain is so weird and I can't spend all day doing this.

https://github.com/llvm/llvm-project/pull/154251