[clang] [llvm] Disable pre-link LTO vectorisation and loop unroll for at O2 and O3 optimisation levels (PR #195886)

[Luke Lau via cfe-commits]

https://github.com/lukel97 approved this pull request.

After https://github.com/llvm/llvm-project/pull/231400 all the SPEC CPU 2017 regressions are gone now on RISC-V, and we're seeing some nice gains. 

525.x264_r is 3% faster due to better `satd_16x16` vectorization. I also checked and we get the similar improvements on x86-64 too. 

This was with the full "stack" of changes, including https://github.com/llvm/llvm-project/pull/196570 https://github.com/llvm/llvm-project/pull/197176 and https://github.com/llvm/llvm-project/pull/197678, so it would be good to land them together. 

LGTM, thanks for your patience

https://github.com/llvm/llvm-project/pull/195886