[llvm] [SLSR] Avoid repeatably calling canReuseInstruction for the same Basis (PR #196545)

[Igor Wodiany via llvm-commits]

IgWod wrote:

> > How many compile-time gains does this PR give? I recall some cases that allocate a massive number of candidates, this change's overhead may be a concern. @Chengjunp, please review.
> 
> You could run it through https://llvm-compile-time-tracker.com/index.php, if SLSR is part of the standard optimization pipeline? (To do that I think you need to ask @nikic to add your fork of LLVM to the tracker, and then push to a branch named "perf/" on your fork.)

Unfortunately, it seems that the pass only runs for AMDGPU and NVPTX targets.

https://github.com/llvm/llvm-project/pull/196545