[llvm] Disable pre-link LTO vectorisation and tweak LTO pipeline (PR #195886)

[Momchil Velikov via llvm-commits]

momchil-velikov wrote:

> Thanks for pushing this along. Are you able to share your performance data/compile time statistics?
Apologies, I've given the wrong link to the RFC, here's the correct one: https://discourse.llvm.org/t/rfc-disable-pre-lto-loop-vectorisation/90539

I've got more or less neutral results across SPECv6 (some potential improvements were "disabled" in order to not affect compile times).

There's a slight overall regression in SPECv8 (with some benchmarks regressing quite a bit), but I've not yet investigated it.


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