The inliner's `FeatureMap` used to be immutable, but in IR2Vec cases we don't know the shapes of the embedding vectors until later, so we need to initialize it at the time we construct the advisor. In non-distributed ThinLTO cases, for example, this means we'd mutate shared state. The feature set is also needed when constructing the underlying model runner. The alternative here is to postpone the creation of the model runner to the time we construct the advisor, and also make the feature map a member of the advisor object. (issue identified by @efriedma-quic in PR #154541)
This broke the development build which is only used by us and build on the bots.
Due to the changes made in #143479, size of `FeatureMap` reflects the number of features used. Hence replacing `NumberOfFeatures` with `FeatureMap.size()` in the `DevelopmentModeInlineAdvisor`.
This change would enable using `ir2vec::Embedding` which are float vectors in ML Inliner. Co-authored-by: svkeerthy <venkatakeerthy@google.com>
The feature was only exposed for training and was immediately dropped on the training side. It was bulk-copied into the test model generator, where it had no effect (the generator always returns a constant). In the AOT + test model case, since the test model returns a constant, all input features are pruned by the AOT compiler, so its presence/absence doesn't matter.
Some references in comments are unnecessarily specific, for historical reasons.
This helps training algorithms that may want to sometimes replicate the default decision. The default decision is presented as an extra feature called `inlining_default`. It's not normally exported to save computation time. This is only available in interactive mode. Differential Revision: https://reviews.llvm.org/D147794
This reverts commit a772f0bb920a4957fb94dd8dbe45943809fd0ec3. The main problem was related to how we handled `dbgs()` from the hosted compiler. Using explicit `subprocess.communicate`, and not relying on dbgs() being flushed until the end appears to address the problem. Also some fixes due to some bots running older pythons, so we can't have nice things like `int | float` and such.