b15caeb0
by Fendor at 2025-12-19T09:43:37+00:00
Refactor the Binary serialisation interface
The goal is simplifiy adding deduplication tables to `ModIface`
interface serialisation.
We identify two main points of interest that make this difficult:
1. UserData hardcodes what `Binary` instances can have deduplication
tables. Moreover, it heavily uses partial functions.
2. GHC.Iface.Binary hardcodes the deduplication tables for 'Name' and
'FastString', making it difficult to add more deduplication.
Instead of having a single `UserData` record with fields for all the
types that can have deduplication tables, we allow to provide custom
serialisers for any `Typeable`.
These are wrapped in existentials and stored in a `Map` indexed by their
respective `TypeRep`.
The `Binary` instance of the type to deduplicate still needs to
explicitly look up the decoder via `findUserDataReader` and
`findUserDataWriter`, which is no worse than the status-quo.
`Map` was chosen as microbenchmarks indicate it is the fastest for a
small number of keys (< 10).
To generalise the deduplication table serialisation mechanism, we
introduce the types `ReaderTable` and `WriterTable` which provide a
simple interface that is sufficient to implement a general purpose
deduplication mechanism for `writeBinIface` and `readBinIface`.
This allows us to provide a list of deduplication tables for
serialisation that can be extended more easily, for example for
`IfaceTyCon`, see the issue https://gitlab.haskell.org/ghc/ghc/-/issues/24540
for more motivation.
In addition to this refactoring, we split `UserData` into `ReaderUserData`
and `WriterUserData`, to avoid partial functions and reduce overall
memory usage, as we need fewer mutable variables.
Bump haddock submodule to accomodate for `UserData` split.
-------------------------
Metric Increase:
MultiLayerModulesTH_Make
MultiLayerModulesRecomp
T21839c
-------------------------
Split `BinHandle` into `ReadBinHandle` and `WriteBinHandle`
A `BinHandle` contains too much information for reading data.
For example, it needs to keep a `FastMutInt` and a `IORef BinData`,
when the non-mutable variants would suffice.
Additionally, this change has the benefit that anyone can immediately
tell whether the `BinHandle` is used for reading or writing.
Bump haddock submodule BinHandle split.
Add Eq and Ord instance to `IfaceType`
We add an `Ord` instance so that we can store `IfaceType` in a
`Data.Map` container.
This is required to deduplicate `IfaceType` while writing `.hi` files to
disk. Deduplication has many beneficial consequences to both file size
and memory usage, as the deduplication enables implicit sharing of
values.
See issue #24540 for more motivation.
The `Ord` instance would be unnecessary if we used a `TrieMap` instead
of `Data.Map` for the deduplication process. While in theory this is
clerarly the better option, experiments on the agda code base showed
that a `TrieMap` implementation has worse run-time performance
characteristics.
To the change itself, we mostly derive `Eq` and `Ord`. This requires us
to change occurrences of `FastString` with `LexicalFastString`, since
`FastString` has no `Ord` instance.
We change the definition of `IfLclName` to a newtype of
`LexicalFastString`, to make such changes in the future easier.
Bump haddock submodule for IfLclName changes
Move out LiteralMap to avoid cyclic module dependencies
Add deduplication table for `IfaceType`
The type `IfaceType` is a highly redundant, tree-like data structure.
While benchmarking, we realised that the high redundancy of `IfaceType`
causes high memory consumption in GHCi sessions when byte code is
embedded into the `.hi` file via `-fwrite-if-simplified-core` or
`-fbyte-code-and-object-code`.
Loading such `.hi` files from disk introduces many duplicates of
memory expensive values in `IfaceType`, such as `IfaceTyCon`,
`IfaceTyConApp`, `IA_Arg` and many more.
We improve the memory behaviour of GHCi by adding an additional
deduplication table for `IfaceType` to the serialisation of `ModIface`,
similar to how we deduplicate `Name`s and `FastString`s.
When reading the interface file back, the table allows us to automatically
share identical values of `IfaceType`.
To provide some numbers, we evaluated this patch on the agda code base.
We loaded the full library from the `.hi` files, which contained the
embedded core expressions (`-fwrite-if-simplified-core`).
Before this patch:
* Load time: 11.7 s, 2.5 GB maximum residency.
After this patch:
* Load time: 7.3 s, 1.7 GB maximum residency.
This deduplication has the beneficial side effect to additionally reduce
the size of the on-disk interface files tremendously.
For example, on agda, we reduce the size of `.hi` files (with
`-fwrite-if-simplified-core`):
* Before: 101 MB on disk
* Now: 24 MB on disk
This has even a beneficial side effect on the cabal store. We reduce the
size of the store on disk:
* Before: 341 MB on disk
* Now: 310 MB on disk
Note, none of the dependencies have been compiled with
`-fwrite-if-simplified-core`, but `IfaceType` occurs in multiple
locations in a `ModIface`.
We also add IfaceType deduplication table to .hie serialisation and
refactor .hie file serialisation to use the same infrastrucutre as
`putWithTables`.
Bump haddock submodule to accomodate for changes to the deduplication
table layout and binary interface.
Add run-time configurability of `.hi` file compression
Introduce the flag `-fwrite-if-compression=<n>` which allows to
configure the compression level of writing .hi files.
The motivation is that some deduplication operations are too expensive
for the average use case. Hence, we introduce multiple compression
levels with variable impact on performance, but still reduce the
memory residency and `.hi` file size on disk considerably.
We introduce three compression levels:
* `1`: `Normal` mode. This is the least amount of compression.
It deduplicates only `Name` and `FastString`s, and is naturally the
fastest compression mode.
* `2`: `Safe` mode. It has a noticeable impact on .hi file size and is
marginally slower than `Normal` mode. In general, it should be safe to
always use `Safe` mode.
* `3`: `Full` deduplication mode. Deduplicate as much as we can,
resulting in minimal .hi files, but at the cost of additional
compilation time.
Reading .hi files doesn't need to know the initial compression level,
and can always deserialise a `ModIface`, as we write out a byte that
indicates the next value has been deduplicated.
This allows users to experiment with different compression levels for
packages, without recompilation of dependencies.
Note, the deduplication also has an additional side effect of reduced
memory consumption to implicit sharing of deduplicated elements.
See https://gitlab.haskell.org/ghc/ghc/-/issues/24540 for example where
that matters.
-------------------------
Metric Decrease:
MultiLayerModulesDefsGhciWithCore
T16875
T21839c
T24471
hard_hole_fits
libdir
-------------------------
Improve sharing of duplicated values in `ModIface`, fixes #24723
As a `ModIface` often contains duplicated values that are not
necessarily shared, we improve sharing by serialising the `ModIface`
to an in-memory byte array. Serialisation uses deduplication tables, and
deserialisation implicitly shares duplicated values.
This helps reducing the peak memory usage while compiling in
`--make` mode. The peak memory usage is especially smaller when
generating interface files with core expressions
(`-fwrite-if-simplified-core`).
On agda, this reduces the peak memory usage:
* `2.2 GB` to `1.9 GB` for a ghci session.
On `lib:Cabal`, we report:
* `570 MB` to `500 MB` for a ghci session
* `790 MB` to `667 MB` for compiling `lib:Cabal` with ghc
There is a small impact on execution time, around 2% on the agda code
base.
Avoid unneccessarily re-serialising the `ModIface`
To reduce memory usage of `ModIface`, we serialise `ModIface` to an
in-memory byte array, which implicitly shares duplicated values.
This serialised byte array can be reused to avoid work when we actually
write the `ModIface` to disk.
We introduce a new field to `ModIface` which allows us to save the byte
array, and write it direclty to disk if the `ModIface` wasn't changed
after the initial serialisation.
This requires us to change absolute offsets, for example to jump to the
deduplication table for `Name` or `FastString` with relative offsets, as
the deduplication byte array doesn't contain header information, such as
fingerprints.
To allow us to dump the binary blob to disk, we need to replace all
absolute offsets with relative ones.
We introduce additional helpers for `ModIface` binary serialisation, which
construct relocatable binary blobs. We say the binary blob is relocatable,
if the binary representation can be moved and does not contain any
absolute offsets.
Further, we introduce new primitives for `Binary` that allow to create
relocatable binaries, such as `forwardGetRel` and `forwardPutRel`.
-------------------------
Metric Decrease:
MultiLayerModulesDefsGhcWithCore
Metric Increase:
MultiComponentModules
MultiLayerModules
T10421
T12150
T12234
T12425
T13035
T13253-spj
T13701
T13719
T14697
T15703
T16875
T18698b
T18140
T18304
T18698a
T18730
T18923
T20049
T24582
T5837
T6048
T9198
T9961
mhu-perf
-------------------------
These metric increases may look bad, but they are all completely benign,
we simply allocate 1 MB per module for `shareIface`. As this allocation
is quite quick, it has a negligible impact on run-time performance.
In fact, the performance difference wasn't measurable on my local
machine. Reducing the size of the pre-allocated 1 MB buffer avoids these
test failures, but also requires us to reallocate the buffer if the
interface file is too big. These reallocations *did* have an impact on
performance, which is why I have opted to accept all these metric
increases, as the number of allocated bytes is merely a guidance.
This 1MB allocation increase causes a lot of tests to fail that
generally have a low allocation number. E.g., increasing from 40MB to
41MB is a 2.5% increase.
In particular, the tests T12150, T13253-spj, T18140, T18304, T18698a,
T18923, T20049, T24582, T5837, T6048, and T9961 only fail on i386-darwin
job, where the number of allocated bytes seems to be lower than in other
jobs.
The tests T16875 and T18698b fail on i386-linux for the same reason.
WIP: Lazy loading of IfaceDecl