[llvm] [RISCV64] liveness analysis (PR #167454)

[Alex Bradbury via llvm-commits]

asb wrote:

What is your expectation for the impact because I've tried this with SPEC and I've not succeeded in getting any meaningful difference in dynamic instruction count for SPEC2017 `-march=rva22u64 -O3` (i.e. the diff rounds to 0.00%).

Also, please remember to give guidance on what combinations of flags you expect people to use in order to test this properly! My attempts:

* `-mllvm -riscv-enable-live-variables`
  * Compiler crashes as reported above
* `-mllvm -riscv-enable-live-variables -mllvm -riscv-liveness-verify=0`
  * Compiles but no diff in generated .s vs baseline
* `-riscv-enable-live-variables -mllvm -riscv-liveness-update-mbb-liveins`
  * 8961 insertions, 8959 deletions in .s output. No impact on dyn inst count above 0.00% (rounded)
* `-mllvm -riscv-enable-live-variables -mllvm -riscv-liveness-update-mbb-liveins -mllvm -riscv-liveness-update-kills`
  * 19273 insertions, 19022 deletions in .s output. Reduction in dyn inst count of 0.01% for povray, all others round to 0.00%.

https://github.com/llvm/llvm-project/pull/167454