[Git][ghc/ghc][wip/romes/graph-compact-easy] 21 commits: Add flags for switching off speculative evaluation.
Rodrigo Mesquita (@alt-romes)
gitlab at gitlab.haskell.org
Mon Jan 20 11:29:48 UTC 2025
Rodrigo Mesquita pushed to branch wip/romes/graph-compact-easy at Glasgow Haskell Compiler / GHC
Commits:
23099752 by Luite Stegeman at 2025-01-08T00:33:33+01:00
Add flags for switching off speculative evaluation.
We found that speculative evaluation can increase the amount of
allocations in some circumstances. This patch adds new flags for
selectively disabling speculative evaluation, allowing us to
test the effect of the optimization.
The new flags are:
-fspec-eval
globally enable speculative evaluation
-fspec-eval-dictfun
enable speculative evaluation for dictionary functions (no effect
if speculative evaluation is globally disabled)
The new flags are on by default for all optimisation levels.
See #25284
- - - - -
0161badc by Ben Gamari at 2025-01-09T17:30:05-05:00
rts/printClosure: Print IPE information for thunks and functions
This makes it considerably easier to grok the structure of the heap
when IPE information is available.
- - - - -
023f36f5 by Rodrigo Mesquita at 2025-01-10T14:57:48-05:00
user_guide: Note -pgmP/-optP are for /Haskell/-CPP
Fixes #25574
- - - - -
e1c133f2 by Ben Gamari at 2025-01-10T14:58:25-05:00
dump-decls: Suppress unit-ids
While the testsuite driver already normalizes these away, they are
nevertheless a severe nuisance when diffing outside of the testsuite.
Intriguingly, this doesn't completely eliminate the unit IDs; some
wired-in names are still printed. However, this is a cheap and helpful
improvement over the status quo so I am simply going to accept this.
Fixes #25334.
- - - - -
2e7bf446 by sheaf at 2025-01-13T10:55:26+01:00
Remove SDocs from ErrCtxt & ErrInfo
This commit:
- turns the SDoc used in ErrCtxt into a proper error datatype,
ErrCtxtMsg, which contains all the different error contexts that
can be added,
- replaces ErrInfo with [ErrCtxt].
ErrInfo used to contain two SDocs; the first is replaced with [ErrCtxt],
and the second is removed, with the relevant information being put
in the appropriate error message constructors.
Fixes #23436
- - - - -
2d62b970 by Mike Pilgrem at 2025-01-13T12:59:10-05:00
Re CLC #300 - Specify fmap for NonEmpty as map
See:
* https://github.com/haskell/core-libraries-committee/issues/300
Seeks to:
* move existing instances for NonEmpty (except of Eq and Ord) out of GHC.Internal.Base into new GHC.Internal.Data.NonEmpty (to avoid otherwise unavoidable cycles in the module graph);
* move map out of Data.List.NonEmpty (base package) into GHC.Internal.Data.NonEmpty;
* define fmap as map for NonEmpty instance of Functor, avoiding code duplication;
* re-export map from existing GHC.Internal.Data.List.NonEmpty; and
* re-export map from Data.List.NonEmpty (base package);
without breaking anything in the GHC repository.
Various tests *.stdout and *.stderr files are amended also.
- - - - -
ab3ab3e3 by Luite Stegeman at 2025-01-13T12:59:58-05:00
compiler/coreprep: Turn off dictionary speculation by default
Speculative evaluation can cause performance regressions,
therefore we turn it off by default. It can be enabled again
with the -fspec-eval-dictfun flag
See #25284
- - - - -
3d9cacd5 by Patrick at 2025-01-14T02:34:46+08:00
Enhance kind inference for data family instances
This commit improves kind inference for data family instances by kind-checking
the constructors, for H98 and newtype declarations (ONLY), as well as
kind-checking the result kind signature (when using GADT syntax).
This fixes #25611.
Typechecker changes:
In `tcDataFamInstHeader`, we now kind-check the constructors using
`kcConDecls`, for H98-style decls and newtype decls ONLY.
See Note [Kind inference for data family instances].
Testsuite changes:
- The T25611{a,b,c,d} tests test the new kind inference implementation.
- a,b: infer result kind from constructors (H98 case)
- c: renamed version of test UnliftedNewtypesUnassociatedFamilyFail,
which now passes
- d: GADT case, checking that we don't infer overly rigid kinds when
kind-checking the constructors in data family instances.
- DataInstanceKindsDefaults tests defaulting data instance kinds
without UnliftedNewtypes or UnliftedDatatypes, as outlined in
Note [Defaulting result kind of newtype/data family instance].
Also a few notes are updated to reflect the changes.
Co-authored-by: default avatarSimon Peyton Jones <simon.peytonjones at gmail.com>
- - - - -
f6493dbc by amesgen at 2025-01-15T18:47:23-05:00
wasm: prevent bundlers from resolving import("node:timers")
This fixes the following esbuild error:
✘ [ERROR] Could not resolve "node:timers"
www/ghc_wasm_jsffi.js:66:25:
66 │ return (await import("node:timers")).setImmediate;
╵ ~~~~~~~~~~~~~
The package "node:timers" wasn't found on the file system but is built into node. Are you trying
to bundle for node? You can use "--platform=node" to do that, which will remove this error.
Previously (i.e. after !13503), one had to work around this by passing
`--external:node:timers`.
- - - - -
87e82e2e by sheaf at 2025-01-16T14:51:45+01:00
Use checkTyEqRhs to make types concrete
This commit refactors makeTypeConcrete to call checkTyEqRhs with
the appropriate parameters. This avoids duplicating subtle logic
in two places in the compiler.
Changes:
1. Refactor of 'TyEqFlags'. Now 'TyEqFlags' stores a 'TEFTask', which
is a description of which of the following checks we want to
perform in 'checkTyEqRhs':
- occurs check
- level check
- concreteness check
In the process, the 'AreUnifying' datatype has been removed, as it
is no longer needed.
2. Refactor of 'checkTyVar':
a. Make use of the new 'TEFTask' data type to decide which checks
to perform.
In particular, this ensures that we perform **both** a
concreteness check and a level check when both are required;
previously we only did a concreteness check (that was a bug!).
b. Recursively call 'checkTyVar' on the kind of unfilled
metavariables. This deals with a bug in which we failed to
uphold the invariant that the kind of a concrete type must
itself be concrete. See test cases T23051, T23176.
3. Re-write of 'makeTypeConcrete', which now simply calls
'checkTyEqRhs' with appropriate 'TyEqFlags'/'TEFTask'.
This gets rid of code duplication and risk for the two code paths
going out-of-sync.
Fixes #25616. See also #23883.
- - - - -
5a8f35bd by ARATA Mizuki at 2025-01-17T11:17:49-05:00
x86 NCG: Use correct format for MOVD in the implementation of unpackInt64X2#
MOVD takes the input format.
Fixes #25658
- - - - -
14f8a7ec by Mateusz Goślinowski at 2025-01-17T22:49:09+00:00
Allow multiline strings in JS FFI (#25633)
- - - - -
854c2f75 by Simon Peyton Jones at 2025-01-18T02:54:08-05:00
Fix a buglet in tcSplitForAllTyVarsReqTVBindersN
The problem was that an equation in `split` had two guards (one about
visiblity and one about `n_req`). So it fell thorugh if /either/
was False. But the next equation then assumed an invisible binder.
Simple bug, easily fixed. Fixes #25661.
- - - - -
264a1186 by sheaf at 2025-01-18T10:05:56+00:00
Generalise GHC diagnostic code infrastructure
This commit generalises the infrastructure used for diagnostic codes,
allowing it to be used for other namespaces than the GHC namespace.
In particular, this enables GHCi to re-use the same infrastructure to
emit error messages.
- - - - -
bf4f5ad3 by Jade at 2025-01-18T10:05:56+00:00
Add structured errors to GHCi (#23338)
This patch creates the 'GhciCommandErrorMessage' data type which
implents the 'Diagnostic' class and also provides error code for these
error conditions.
- - - - -
b6f54188 by Ben Gamari at 2025-01-18T12:38:46-05:00
Revert "Division by constants optimization"
This appears to be responsible for the regression described in #25653.
This reverts commit daff1e30219d136977c71f42e82ccc58c9013cfb.
- - - - -
0fd90de8 by Ben Gamari at 2025-01-18T12:38:46-05:00
testsuite: Introduce div2 test
This is a useful test from !8392 which is worth keeping around.
- - - - -
32680979 by Ben Gamari at 2025-01-18T12:38:46-05:00
testsuite: Test shift correctness in mul2 test
- - - - -
163aa50a by Ben Gamari at 2025-01-18T12:38:46-05:00
testsuite: Add regression test for #25653
- - - - -
ea25a9ae by Matthew Pickering at 2025-01-20T10:02:08+00:00
driver: Store an ExternalModuleGraph in the EPS
We now store an ExternalModuleGraph in the EPS. When an new interface is
loaded, the module graph is extended with a node for the loaded
interface. The result is a partial module graph. If you want to run
a transitive closure query on the graph you must first force the
transitive closure to be loaded by using `loadExternalGraphBelow`.
The primary advantage (for now) is that the transitive dependency
calculation does not have to be repeated in getLinkDeps. If your module
had many dependencies and many splices, performing this calculation at
every splice site took a significant amount of time.
We might also want to use this module graph in future for considering
questions such as reachability of rules or accessibilty of instance
imported by levelled imported.
This patch removes another place in the compiler where transitive
dependency is calculated in an ad-hoc manner. In general, the transitive
dependency calculation should be cached and computed using a ModuleGraph
abstraction.
The transitive dependency query required by getLinkDeps operates on a
graph without hs-boot nodes. If a linkable from a module in a loop is
needed, then all modules in the loop are necessary to be available to
execute that module. Therefore there is a query in `ModuleGraph` and
`ExternalModuleGraph` which allows a transitive closure query to be
performed on a graph without loops.
Fixes #25634
-------------------------
Metric Decrease:
MultiLayerModulesTH_Make
MultiLayerModulesTH_OneShot
Metric Increase:
mhu-perf
-------------------------
Co-authored-by: Rodrigo Mesquita <rodrigo.m.mesquita at gmail.com>
- - - - -
44d1d344 by Rodrigo Mesquita at 2025-01-20T11:29:36+00:00
driver: Store the HomePackageTable in a mutable reference
This commit refactors the HomePackageTable and HomeUnitGraph:
(1) It fixes a quadratic-in-the-number-of-modules space leak in upsweep (#25511)
(2) And it reworks these structures into their own modules to simplify
the driver. The refactor is driven by the introduction of IO in the HPT
interface, but is a bit more aggressive in simplifying the
interfaces to enforce correct usage (ie to avoid performance
pitfalls).
Specifically:
- The `HomeUnitGraph` (HUG) is now in `GHC.Unit.Home.Graph`
- The `HomePackageTable` (HPT) is now in `GHC.Unit.Home.PackageTable`
- The HPT now stores an `IORef` with the table of loaded home package modules.
- The interface to the HPT now requires IO
- The interface now enforces that the HPT is a datastructure that
only grows
- This is not enforced in the interface, but, clients of the HPT
should never care about there being more or less entries in the
HPT when these additional entries are not relevant to their result.
- The exception to the invariant that the HPT is monotonically
increasing is `restrictHpt`, a function which is called at a
"barrier point" (during which there are no other threads
inspecting or inserting in the HPT). The invariant can be
temporarily broken at this point (currently, after upsweep).
This is safe because a single thread holds control over the
structure (thus the invariant being broken is never observed).
The hug_var and associated structures in the driver, which aimed to
improve memory usage in the driver by updating in place a HUG during
upsweep, are no longer required as the HPT entries in the HUG are now
themselves mutable by construction. This was previously explained in
Note [ModuleNameSet, efficiency and space leaks], which is no longer
relevant and was deleted.
Fixes #25511
Co-authored-by: Matthew Pickering <matthewtpickering at gmail.com>
-------------------------
Metric Decrease:
MultiComponentModulesRecomp
MultiLayerModulesRecomp
-------------------------
- - - - -
30 changed files:
- compiler/GHC.hs
- compiler/GHC/Cmm/Config.hs
- compiler/GHC/Cmm/MachOp.hs
- compiler/GHC/Cmm/Opt.hs
- compiler/GHC/Cmm/Pipeline.hs
- compiler/GHC/Cmm/Sink.hs
- compiler/GHC/CmmToAsm/X86/CodeGen.hs
- compiler/GHC/Core/Opt/Pipeline.hs
- compiler/GHC/Core/TyCo/Rep.hs
- compiler/GHC/Core/TyCon.hs
- compiler/GHC/Core/Type.hs
- compiler/GHC/CoreToStg/Prep.hs
- compiler/GHC/Data/IOEnv.hs
- compiler/GHC/Driver/Backpack.hs
- compiler/GHC/Driver/Config/Cmm.hs
- compiler/GHC/Driver/Config/CoreToStg/Prep.hs
- compiler/GHC/Driver/DynFlags.hs
- compiler/GHC/Driver/Env.hs
- compiler/GHC/Driver/Errors.hs
- compiler/GHC/Driver/Errors/Ppr.hs
- compiler/GHC/Driver/Errors/Types.hs
- compiler/GHC/Driver/Flags.hs
- compiler/GHC/Driver/Hooks.hs
- compiler/GHC/Driver/Main.hs
- compiler/GHC/Driver/Make.hs
- compiler/GHC/Driver/MakeFile.hs
- compiler/GHC/Driver/Pipeline.hs
- compiler/GHC/Driver/Ppr.hs
- compiler/GHC/Driver/Session.hs
- compiler/GHC/Hs/Decls.hs
The diff was not included because it is too large.
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/7abb46d6586934ced5a43c49d18a6b8d7b24d5fa...44d1d344f41782e85d6ecd97a4b0da62bb7ef930
--
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/7abb46d6586934ced5a43c49d18a6b8d7b24d5fa...44d1d344f41782e85d6ecd97a4b0da62bb7ef930
You're receiving this email because of your account on gitlab.haskell.org.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-commits/attachments/20250120/ed215fd9/attachment-0001.html>
More information about the ghc-commits
mailing list