[llvm] [NVPTX] cleanup & canonicalize `mov` (PR #129344)

[Artem Belevich via llvm-commits]

Artem-B wrote:

"Challenge" as in "it's hard to automatically reduce PTX in principle" (i.e. delta/creduce don't work well with PTX). Or as in "the error only happens on the large PTX input and disappears once PTX is smaller than some still-large size" ?

If it's the former, I found that I can get most of the way towards minimal reproducer by reducing the IR instead. I use llvm-reduce (https://llvm.org/docs/CommandGuide/llvm-reduce.html) + a test which assembles PTX with ptxas, then disassembles it using cuobjdump or nvdisasm and looks at SASS for the signs of miscompilation. 

https://github.com/llvm/llvm-project/pull/129344