[PATCH] D22368: Replace hot-callsite based heuristic to use its own threshold parameter instead of share inline-hint parameter

[Dehao Chen via llvm-commits]

I experimented different thresholds from 325 (original) to 4000. The code
size change is within 2% fall all speccpu2006 int benchmarks. In terms of
performance. Up to 4% speedup is observed for perlbench when changing
threshold from 325 to 3000. And above 3000, the performance curve remains
flat. For all other speccpu2006 int benchmarks, the performance change does
not escape noise range when changing threshold from 325 to 4000. Similar
performance/size result is observed for internal benchmarks. So I think
3000 seems to be a sweet spot for the threshold.

Thanks,
Dehao

On Tue, Jul 19, 2016 at 5:26 PM, Easwaran Raman <eraman at google.com> wrote:

> eraman added inline comments.
>
> ================
> Comment at: lib/Analysis/InlineCost.cpp:69
> @@ -68,1 +68,3 @@
>
> +static cl::opt<int>
> +    HotCallSiteThreshold("hot-callsite-threshold", cl::Hidden,
> cl::init(3000),
> ----------------
> Adding a threshold different from inlinehint-threshold makes sense. But if
> you want to set the default this high, then it is important to have numbers
> to justify. Running spec with different thresholds and a comment here
> saying you chose this default because this gives the best performance/size
> tradeoff is important.
>
>
> https://reviews.llvm.org/D22368
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160725/d11918cd/attachment.html>