[llvm] [NVPTX] Expand fp/int conversions involving integers wider than 64 bits (PR #201679)

[Lukas Stephan via llvm-commits]

kulst wrote:

I've integrated your comments and rebased on top of the current main branch. 

> Could you share more about the workload this is coming from?
I'm mostly curious to know if we should expect to see more i128 related bugs coming in as a result of increased usage.

I experienced this while compiling Rust's [compiler-builtins crate](https://github.com/rust-lang/compiler-builtins) without optimizations using the NVPTX target. I was able to track down the llvm error to these conversion instructions.

NVPTX is one of Rust's tier 2 targets and uses LLVM's NVPTX backend. 
The target is quite niche from a Rust perspective. So there is still a lot of work to do to make it as compatible as possible with Rust's feature set. As one of Rust's NVPTX target maintainers I am working towards this.

Rust has `i128` and `u128` built-in types. Not sure if NVPTX already supports all instructions Rust expects for these. But with this patch applied at least Rust's `compiler_builtins` compiles successfully, even when optimizations are turned off.

https://github.com/llvm/llvm-project/pull/201679