[llvm] [PGO][ICP] Add separate hot count threshold for indirect call promotion (PR #206656)

[Teresa Johnson via llvm-commits]

teresajohnson wrote:

> Does tuning the hot count in general not provide similar wins?
> 
> This also seems like it would be pretty fragile as an optimization win. I'm also not a huge fan of generalizing `isHotCount` with a boolean flag. A `cl::opt` inside ICP I think would be cleanest.

I agree with the above. I also think it would be good to understand why presumably preventing some ICP gave such a big performance win. Is it due to the profile not being representative? There are some existing flags defined in llvm/lib/Analysis/IndirectCallPromotionAnalysis.cpp to control the number of promotions etc that might be useful depending on why exactly the improvement occurred. 

https://github.com/llvm/llvm-project/pull/206656