`TotalRootEntryCount` captures how many times that root was entered - regardless if a profile was also collected or not (profile collection for a given root happens on only one thread at a time).
We don't do this in compiler_rt because the goal there is to flush out the data as fast as possible, so traversing and multiplying vectors is punted to the profile user.
We really just need to do this when flattening the profile so that the values across roots and flat profiles match. We could do it earlier, too - like when loading the profile - but it seems beneficial (at least for debugging) to keep the counter values the same as the loaded ones. We can revisit this later.
Same idea as in #134723 - flatten indirect call info in `"VP"` `MD_prof` metadata for the thinlinker, for cases that aren't covered by a contextual profile. If we don't ICP an indirect call target in the specialized module, the call will fall to the copy of that target outside the specialized module. If the graph under that target also has some indirect calls, in the absence of this pass, we'd have a steeper performance regression - because none of those would have a chance to be ICPed.
Flatten the profile pre-thinlink so that ThinLTO has something to work with for the parts of the binary that aren't covered by contextual profiles. Post-thinlink, the flattener is re-run and will actually change profile info, but just for the modules containing contextual trees ("specialized modules"). For the rest, the flattener just yanks out the instrumentation.
After #134340, the availability of contextual profile isn't in itself an indication of compiling the module containing all the functions covered by that profile.
We will subsequently treat the whole profile as "flat" in the frontend, (i.e flatten and combine with the flat profile section), so we can have a profile for ThinLTO for parts of the application that don't come under the contextual profile. After ThinLTO, we will treat the module(s) containing contextual trees differently: they'll have only the contextual profile pertinent to them. The rest of the modules (non-contextual) will proceed "as usual", off the flattened profile.
This patch implements pruning of the contextual profile to enable the above.
When we collect a contextual profile, we sample the threads entering its root and only collect on one at a time (see `ContextRoot::Taken`). If we want to compare profiles between contextual profiles, and/or flat profiles, we have a problem: we don't know how to compare the counter values relative to each other. To that end, we add `ContextRoot::TotalEntries`, which is incremented every time a root is entered and serves as multiplier for the counter values collected under that root.
We expose this in the profile and leave the normalization to the user of the profile, for a few reasons:
* it's only needed if reasoning about all profiles in aggregate.
* the goal, in compiler_rt, is to flush out the profile as quickly as possible, and performing multiplications adds an overhead that may not even be necessary if the consumer of the profile doesn't care about combining profiles
* the information itself may be interesting as an indication of relative sampling of various contexts.
The profile format has now a separate section called "Contexts" - there will be a corresponding one for flat profiles. The root has a separate tag because, in addition to not having a callsite ID as all the other context nodes have under it, it will have additional fields in subsequent patches.
The rest of this patch amounts to a bit of refactorings in the reader/writer (for better reuse later) and tests fixups.
This is a follow-up from PR #122545, which enabled converting yaml to contextual profiles.
This change uses the lower level yaml APIs because:
- the mapping APIs `llvm::yaml` offers don't work with `const` values, because they (the APIs) want to enable both serialization and deserialization
- building a helper data structure would be an alternative, but it'd be either memory-consuming or overly-complex design, given the recursive nature of the contextual profiles.
We have a textual representation of contextual profiles for test scenarios, mainly. This patch moves that to YAML instead of JSON. YAML is more succinct and readable (some of the .ll tests should be illustrative). In addition, JSON is parse-able by the YAML reader.
A subsequent patch will address deserialization.
(thanks, @kazutakahirata, for showing me how to use the llvm YAML reader/writer APIs, which I incorrectly thought to be more low-level than the JSON ones!)
This is mostly for test: under contextual profiling, we perform ICP for those indirect callsites which have targets marked as `alwaysinline`.
This helped uncover a bug with the way the profile was updated upon ICP, where we were skipping over the update if the target wasn't called in that context. That was resulting in incorrect counts for the indirect BB.
Also flyby fix to the total/direct count values, they should be 64-bit (as all counters are in the contextual profile)
The `step` instrumentation shouldn't be treated, during use, like an `increment`. The latter is treated as a BB ID. The step isn't that, it's more of a type of value profiling. We need to distinguish between the 2 when really looking for BB IDs (==increments), and handle appropriately `step`s. In particular, we need to know when to elide them because `select`s may get elided by function cloning, if the condition of the select is statically known.
Added tests to the validator and fixed issues stemming from the previous skipping over BBs with single successors - which is incorrect. That would be now picked by added tests where the assertions are expected to be triggered.
The assertion that all out-edges of a BB can't be 0 is incorrect: they
can be, if that branch is on a cold subgraph.
Added validators and asserts about the expected proprerties of the
propagated counters.
This patch enables experimenting with the contextual profile. ICP is currently disabled in this case - will reenable it subsequently. Also subsequently the inline cost model / decision making would be updated to be context-aware. Right now, this just achieves "complete use" of the profile, in that it's ingested, maintained, and sunk to a flat profile when not needed anymore.
Issue [#89287](https://github.com/llvm/llvm-project/issues/89287)
Pass to flatten and lower the contextual profile to profile (i.e. `MD_prof`) metadata. This is expected to be used after all IPO transformations have happened.
Prior to lowering, the instrumentation is maintained during IPO and the contextual profile is kept in sync (see PRs #105469, #106154). Flattening (#104539) sums up all the counters belonging to all a function's context nodes.
We first propagate counter values (from the flattened profile) using the same propagation algorithm as `PGOUseFunc::populateCounters`, then map the edge values to `branch_weights`. Functions. in the module that don't have an entry in the flattened profile are deemed cold, and any `MD_prof` metadata they may have is reset. The profile summary is also reset at this point.
Issue [#89287](https://github.com/llvm/llvm-project/issues/89287)
Add an overload of `InlineFunction` that updates the contextual profile. If there is no contextual profile, this overload is equivalent to the non-contextual profile variant.
Post-inlining, the update mainly consists of:
- making the PGO instrumentation of the callee "the caller's": the owner function (the "name" parameter of the instrumentation instructions) becomes the caller, and new index values are allocated for each of the callee's indices (this happens for both increment and callsite instrumentation instructions)
- in the contextual profile:
- each context corresponding to the caller has its counters updated to incorporate the counters inherited from the callee at the inlined callsite. Counter values are copied as-is because no scaling is required since the profile is contextual.
- the contexts of the callee (at the inlined callsite) are moved to the caller.
- the callee context at the inlined callsite is deleted.
Continuing from #102084, which introduced the analysis, we now populate
it with info about functions contained in the module.
When we will update the profile due to e.g. inlined callsites, we'll
ingest the callee's counters and callsites to the caller. We'll move
those to the caller's respective index space (counter and callers), so
we need to know and maintain where those currently end.
We also don't need to keep profiles not pertinent to this module.
This patch also introduces an arguably much simpler way to track the
GUID of a function from the frontend compilation, through ThinLTO, and
into the post-thinlink compilation step, which doesn't rely on keeping
names around. A separate RFC and patches will discuss extending this to
the current PGO (instrumented and sampled) and other consumers as an
infrastructural component.
This is an immutable analysis that loads and makes the contextual profile available to other passes. This patch introduces the analysis and an analysis printer pass. Subsequent patches will introduce the APIs that IPO passes will call to modify the profile as result of their changes.
