[llvm] [ICP] Introduce a hot function cutoff threshold for ICP (PR #208060)

[Congzhe Cao via llvm-commits]

CongzheUalberta wrote:

> Is this work scoped to instrumentation based FDO or have you evaluated on SamplePGO too? My intuition is that there is a lot more variance in call targets counts for SamplePGO and tieing them to block counts may not be the right choice.
> 
> I've not read the patch in detail, so feel free to correct me if my assumptions are incorrect.

Thanks for the comment. With this patch, by default the behavior of ICP stays exactly the same as before because the default value of the flag introduced is the same as `ProfileSummaryCutoffHot`, unless the user intentionally specifies some other value for the flag.

The way ICP works with instrumentation PGO or Sample-based autoFDO is exactly the same as before this patch. There might possibly be some room for improvement on that front, but that is not the scope of this patch.

We have measured with instrumentation PGO as well as CSSPGO (pseudo-instrumentation) and we saw gains there, but we have not measured with pure sample-based autoFDO.


https://github.com/llvm/llvm-project/pull/208060