[llvm-dev] RFC: Synthetic function entry counts

[Sean Silva via llvm-dev]

IIUC, this proposal is just saying that we should infer a static profile
for entry counts just like we do for branch probabilities. In the case of
entry counts, we do not hide that information behind an analysis like BPI,
so currently just annotating synthetic PGO entry counts is a simple
solution that piggybacks on the PGO mechanism and Just Works.

If that is correct then this makes perfect sense to me.

It could be argued that we ought to refactor things so that the raw PGO
metadata is only ever accessed via a wrapper CGSCC analysis that falls back
to interprocedural analysis (i.e. static profile heuristics) when the entry
count metadata is missing, just like BPI does with static intraprocedural
analysis and branch weight metadata. However, we probably don't want to do
that while folks are still depending on the old PM in production since
CGSCC analyses don't exist there which would force us to maintain an old
and new way of doing it.

(Also, it sounds like you want to compute this with a top-down CGSCC
traversal, so it might not actually be computable incrementally as a bottom
up CGSCC analysis which is what CGSCC analyses currently do; an auxiliary
module analysis for the top-down part might work around this though)

Also, the need to run this logic (or similar logic) as a "ThinLTO analysis"
suggests not wedding it too much with the intricacies of the IR-level pass
management (although admittedly we already do that with the inliner and
then ThinLTO has to approximate those inlining decisions, so it might not
be the end of the world to have some divergence).

-- Sean Silva

On Dec 12, 2017 5:02 PM, "Easwaran Raman via llvm-dev" <
llvm-dev at lists.llvm.org> wrote:

> Functions in LLVM IR have a function_entry_count metadata that is attached
> in PGO compilation. By using the entry count together with the block
> frequency info, the compiler computes the profile count of call
> instructions based on which the hotness/coldness of callsites can be
> determined. Experiments have shown that using a higher threshold for hot
> callsites results in improved runtime performance of the generated code
> without significant code size increases. We propose to generate synthetic
> function counts for non-PGO compilation and use the counts for boosting hot
> callsites during inlining.
>
> Synthetic function entry counts of functions are initialized based on
> properties of the function such as whether it is visible outside the
> module, whether it has an inline keyword and so on. Then, the callgraph SCC
> is traversed in reverse post-order. Counts of callsites are determined
> based on the entry count and the block frequency of the callsite. The
> callsite count gets added to the entry count of the callee. For targets of
> indirect calls, we will use the !callees metadata to find the possible
> targets and distribute the count equally among them. For functions in a
> non-trivial SCC, the algorithm has to ensure that the counts are stable and
> deterministic.
>
> In ThinLTO mode, the function summary contains the list of call edges from
> the function. We propose to add the relative block frequency on these
> edges. During the thinlink phase, we propagate the function counts on the
> entire call graph and update the function summary with the synthetic
> counts. Additionally, we plan to use the computed counts to drive the
> importing decisions.
>
> Alternative approach
>
> -----------------------------
>
>
> An alternative to generating synthetic counts is to make block frequency
> info an inter-procedural analysis. Such an analysis would allow comparing
> BFI of callsites in two different functions. This has several downsides:
>
>    -
>
>    The inter-procedural BFI computation is likely to be more expensive in
>    terms of compile-time.
>    -
>
>    Many function passes invalidate the BFI. This will require selective
>    invalidation of function BFIs.
>    -
>
>    Inliner correctly updates function counts of a callee after a callsite
>    is inlined. We can piggyback on this mechanism by using synthetic counts.
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20171215/70485d04/attachment.html>