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

[Momchil Velikov via cfe-commits]

momchil-velikov wrote:

> Do we already understand the root causes for some of the larger regressions, e.g. `736.ocio_r | 1.91%` or `714.cpython_r | 1.80%`.

* Half of `cpython` regression  is due to higher L1I cache refills (compared to main) which disappear after manual alignment of a couple of key functions.
* The other half of the regression due to higher L1D refills, but essentially the same amount of memory operations across identical hot spots. Experiment with inserting extra data block to force the same `brk(NULL)` in both versions of the benchmark (main + disabnle-pre-lto-vec) show regression going down to 0.5-0.7%

Sometimes `cpython` even shows improvement.

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