[llvm] [RISCV][WIP] Let RA do the CSR saves. (PR #90819)

[Mikhail Gudim via llvm-commits]

mgudim wrote:

Currently the shrink wrapping saves / restores all registers at the same point. We noticed that this is not optimal in several places in SPEC. GCC can do this on per-register basis and there is an option to turn this off. By using this flag in gcc we determined that this can make up to 2% difference in dynamic instruction count on gcc benchmark and some smaller impacts in other benchmarks.

Instead of trying to improve shrink-wrapping in llvm, I decided to try this (crazy) idea since it was pretty easy to implement this prototype. I disabled this for functions which can unwind, because this is just an experiment to see if this is viable. Thus all c++ benchmarks that have exceptions were not affected. I got 1.7% improvement on gcc and 1.7% degradation on xz in dynamic instruciton count.

Do you think it's worth pursuing this idea further? To fix the degradations, we need to improve register allocator / coalescing which may have more impact than just improving shrink wrapping. 

https://github.com/llvm/llvm-project/pull/90819