[llvm] [lsr][term-fold] Restrict expansion budget for profiled loops (PR #74747)

[Philip Reames via llvm-commits]

preames wrote:

> Just to double check, do you see regressions if you just use the standard isHighCostExpansion() check? And that's why you are trying to use this more conservative profile count based heuristic instead?

Yes, if I use the standard SCEV threshold.  I can of course use a larger single threshold if you'd rather, but then this patch is largely pointless as we can't generate anything worse than about 8 instructions from the non-constant * non-constant case no matter what we do.  

https://github.com/llvm/llvm-project/pull/74747