[NVPTX] Remove NOP definition

[ldrumm]

PTX has no NOP instruction, and it's unused anywhere in the compiler. This looks like it might have been added for machine scheduling, but it looks unused.

If we eventually need it, we can add it back.

[pasaulais]

Looks good to me

[Artem-B]

LGTM.

[walter-erquinigo]

Hey! I hope you guys (@ldrumm , @Artem-B , @pasaulais ) can shed some light for me:
My team is working on improving debug info for the Mojo language when compiling for NVIDIA GPUs, and having a nop instruction would be a very nice way to force create some line table entries that would help debuggability of inlined function callsites.
So, are nops available in the HW, in which case we could revert this PR? Or are nops just gone and we need to figure out a workaround?
cc: @zyx-billy

[ldrumm]

PTX is a virtual ISA, and this document does not specify a NOP instruction. There might be a NOP instruction on a per-GPU hardware ISA basis, but we only target the publicly documented PTX. Once that goes through the PTX compiler, ptxas, it's out of llvm's hands.

You could probably emulate a nop via a register mov to self, or an (| 0) construction, but seeing as the PTX compiler runs fairly aggressive optimizations, my guess is this won't work.

[ldrumm]

To be clear, if the PTX ISA specified a nop instruction, I'd happily revert this patch, but it doesn't - which is why it made sense in the first place

[walter-erquinigo]

Thank you for the clarification. This is very useful information to us!