6576e70d
by fendor at 2026-03-30T09:30:10+02:00
Record `LinkableUsage` instead of `Linkable` in `LoaderState`
Retaining a ByteCode `Linkable` after it has been loaded retains its
`UnlinkedBCO`, keeping it alive for the remainder of the program.
This starts accumulating a lot of `UnlinkedBCO` and memory over time.
However, the `Linkable` is merely used to later record its usage in
`mkObjectUsage`, which is used for recompilation checking.
However, this is incorrect, as the interface file and bytecode objects
could be in different states, e.g. the interface changes, but the
bytecode library hasn't changed so we don't need to recompile and vice
versa.
By computing a `Fingerprint` for the `ModuleByteCode`, and recording it
in the `LinkableUsage`, we know precisely whether the `ByteCode` object
on disk is outdated.
Thus, parts of this commit just makes sure that we efficiently compute a
`Fingerprint` for `ModuleByteCode` and store it in the on-disk
representation of `ModuleByteCode`.
We change the `LoaderState` to retain `LinkableUsage`, which is smaller
representation of a `Linkable`. This allows us to free the unneeded
fields of `Linkable` after linking them.
We declare the following memory invariants that this commit implements:
* No `LinkablePart` should be retained from `LoaderState`.
* `Linkable`s should be unloaded after they have been loaded.
These invariants are unfortunately tricky to automatically uphold, so we
are simply documenting our assumptions for now.
We introduce the `linkable-space` test which makes sure that after
loading, no `DotGBC` or `UnlinkedBCO` is retained.
-------------------------
Metric Increase:
MultiLayerModulesTH_OneShot
-------------------------
We allocate a bit more, but the peak number of bytes doesn't change.
While a bit unfortunate, accepting the metric increase.
We add multiple new performance measurements where we were able to
observe the desired memory invariants. Further, we add regression tests
to validate that the recompilation checker behaves more correct than
before.