[llvm] [NVPTX] Propagate truncate to operands (PR #98666)

[Artem Belevich via llvm-commits]

Artem-B wrote:

> It expects one of the parameters of the binop to be constant.

I see. I guess the assumption is that if one of the argument is a constant, then truncation for it is free, so it never increases the number of truncates and assumes that the logical op is the same cost, regardless of the size.

I think adding a new target-specific check here would be appropriate. Considering that register pressure is a pretty common issue for NVPTX, something fairly generic like `TLI.shouldReduceRegisterPressure()` may be useful here and likely in other places where we may need to adjust optimizer assumptions for NVPTX.

https://github.com/llvm/llvm-project/pull/98666