-
c7aa0c10
by Andreas Klebinger at 2025-06-15T05:47:24-04:00
Revert "Specialise: Don't float out constraint components."
This reverts commit c9abb87ccc0c91cd94f42b3e36270158398326ef.
Turns out two benchmarks from #19747 regresses by a factor of 7-8x if
we do not float those out.
-
fd998679
by Krzysztof Gogolewski at 2025-06-15T05:48:06-04:00
Fix EPT enforcement when mixing unboxed tuples and non-tuples
The code was assuming that an alternative cannot be returning a normal
datacon and an unboxed tuple at the same time. However, as seen in #26107,
this can happen when using a GADT to refine the representation type.
The solution is just to conservatively return TagDunno.
-
e64b3f16
by ARATA Mizuki at 2025-06-17T10:13:42+09:00
MachRegs.h: Don't define NO_ARG_REGS when a XMM register is defined
On i386, MAX_REAL_VANILLA_REG is 1, but MAX_REAL_XMM_REG is 4.
If we define NO_ARG_REGS on i386, programs that use SIMD vectors may segfault.
Closes #25985
A couple of notes on the BROKEN_TESTS field:
* This fixes the segfault from T25062_V16.
* The failure from T22187_run was fixed in an earlier commit (see #25561),
but BROKEN_TESTS was missed at that time. Now should be a good time to
mark it fixed.
-
3e7c6b4d
by Matthew Pickering at 2025-06-18T15:34:04-04:00
Improve error messages when implicit lifting fails
This patch concerns programs which automatically try to fix level errors
by inserting `Lift`. For example:
```
foo x = [| x |]
~>
foo x = [| $(lift x) |]
```
Before, there were two problems with the message.
1. (#26031), the location of the error was reported as the whole
quotation.
2. (#26035), the message just mentions there is no Lift instance, but
gives no indicate why the user program needed a Lift instance in the
first place.
This problem is especially bad when you disable
`ImplicitStagePersistence`, so you just end up with a confusing "No
instance for" message rather than an error message about levels
This patch fixes both these issues.
Firstly, `PendingRnSplice` differentiates between a user-written splice
and an implicit lift. Then, the Lift instance is precisely requested
with a specific origin in the typechecker. If the instance fails to be
solved, the message is reported using the `TcRnBadlyLevelled`
constructor (like a normal level error).
Fixes #26031, #26035
-
44b8cee2
by Cheng Shao at 2025-06-18T15:34:46-04:00
testsuite: add T26120 marked as broken
-
894a04f3
by Cheng Shao at 2025-06-18T15:34:46-04:00
compiler: fix GHC.SysTools.Ar archive member size writing logic
This patch fixes a long-standing bug in `GHC.SysTools.Ar` that emits
the wrong archive member size in each archive header. It should encode
the exact length of the member payload, excluding any padding byte,
otherwise malformed archive that extracts a broken object with an
extra trailing byte could be created.
Apart from the in-tree `T26120` test, I've also created an out-of-tree
testsuite at https://github.com/TerrorJack/ghc-ar-quickcheck that
contains QuickCheck roundtrip tests for `GHC.SysTools.Ar`. With this
fix, simple roundtrip tests and `writeGNUAr`/GNU `ar` roundtrip test
passes. There might be more bugs lurking in here, but this patch is
still a critical bugfix already.
Fixes #26120 #22586.
Co-authored-by: Codex <codex@openai.com>
-
f677ab5f
by Lauren Yim at 2025-06-18T15:35:37-04:00
fix some typos in the warnings page in the user guide
-
b968e1c1
by Rodrigo Mesquita at 2025-06-18T15:36:18-04:00
Add a frozen callstack to throwGhcException
Fixes #25956
-
a5e0c3a3
by fendor at 2025-06-18T15:36:59-04:00
Update using.rst to advertise full mhu support for GHCi
-
d3e60e97
by Ryan Scott at 2025-06-18T22:29:21-04:00
Deprecate -Wdata-kinds-tc, make DataKinds issues in typechecker become errors
!11314 introduced the `-Wdata-kinds-tc` warning as part of a fix for #22141.
This was a temporary stopgap measure to allow users who were accidentally
relying on code which needed the `DataKinds` extension in order to typecheck
without having to explicitly enable the extension.
Now that some amount of time has passed, this patch deprecates
`-Wdata-kinds-tc` and upgrades any `DataKinds`-related issues in the
typechecker (which were previously warnings) into errors.
-
fd5b5177
by Ryan Hendrickson at 2025-06-18T22:30:06-04:00
haddock: Add redact-type-synonyms pragma
`{-# OPTIONS_HADDOCK redact-type-synonyms #-}` pragma will hide the RHS
of type synonyms, and display the result kind instead, if the RHS
contains any unexported types.
-
fbc0b92a
by Vladislav Zavialov at 2025-06-22T04:25:16+03:00
Visible forall in GADTs (#25127)
Add support for visible dependent quantification `forall a -> t` in
types of data constructors, e.g.
data KindVal a where
K :: forall k.
forall (a::k) -> -- now allowed!
k ->
KindVal a
For details, see docs/users_guide/exts/required_type_arguments.rst,
which has gained a new subsection.
DataCon in compiler/GHC/Core/DataCon.hs
---------------------------------------
The main change in this patch is that DataCon, the Core representation
of a data constructor, now uses a different type to store user-written
type variable binders:
- dcUserTyVarBinders :: [InvisTVBinder]
+ dcUserTyVarBinders :: [TyVarBinder]
where
type TyVarBinder = VarBndr TyVar ForAllTyFlag
type InvisTVBinder = VarBndr TyVar Specificity
and
data Specificity = InferredSpec | SpecifiedSpec
data ForAllTyFlag = Invisible Specificity | Required
This change necessitates some boring, mechanical changes scattered
throughout the diff:
... is now used in place of ...
-----------------+---------------
TyVarBinder | InvisTVBinder
IfaceForAllBndr | IfaceForAllSpecBndr
Specified | SpecifiedSpec
Inferred | InferredSpec
mkForAllTys | mkInvisForAllTys
additionally,
tyVarSpecToBinders -- added or removed calls
ifaceForAllSpecToBndrs -- removed calls
Visibility casts in mkDataConRep
--------------------------------
Type abstractions in Core (/\a. e) always have type (forall a. t)
because coreTyLamForAllTyFlag = Specified. This is also true of data
constructor workers. So we may be faced with the following:
data con worker: (forall a. blah)
data con wrapper: (forall a -> blah)
In this case the wrapper must use a visibility cast (e |> ForAllCo ...)
with appropriately set fco_vis{L,R}. Relevant functions:
mkDataConRep in compiler/GHC/Types/Id/Make.hs
dataConUserTyVarBindersNeedWrapper in compiler/GHC/Core/DataCon.hs
mkForAllVisCos in compiler/GHC/Core/Coercion.hs
mkCoreTyLams in compiler/GHC/Core/Make.hs
mkWpForAllCast in compiler/GHC/Tc/Types/Evidence.hs
More specifically:
- dataConUserTyVarBindersNeedWrapper has been updated to answer "yes"
if there are visible foralls in the type of the data constructor.
- mkDataConRep now uses mkCoreTyLams to generate the big lambda
abstractions (/\a b c. e) in the data con wrapper.
- mkCoreTyLams is a variant of mkCoreLams that applies visibility casts
as needed. It similar in purpose to the pre-existing mkWpForAllCast,
so the common bits have been factored out into mkForAllVisCos.
ConDecl in compiler/Language/Haskell/Syntax/Decls.hs
----------------------------------------------------
The surface syntax representation of a data constructor declaration is
ConDecl. In accordance with the proposal, only GADT syntax is extended
with support for visible forall, so we are interested in ConDeclGADT.
ConDeclGADT's field con_bndrs has been renamed to con_outer_bndrs
and is now accompanied by con_inner_bndrs:
con_outer_bndrs :: XRec pass (HsOuterSigTyVarBndrs pass)
con_inner_bndrs :: [HsForAllTelescope pass]
Visible foralls always end up in con_inner_bndrs. The outer binders are
stored and processed separately to support implicit quantification and
the forall-or-nothing rule, a design established by HsSigType.
A side effect of this change is that even in absence of visible foralls,
GHC now permits multiple invisible foralls, e.g.
data T a where { MkT :: forall a b. forall c d. ... -> T a }
But of course, this is done in service of making at least some of these
foralls visible. The entire compiler front-end has been updated to deal
with con_inner_bndrs. See the following modified or added functions:
Parser:
mkGadtDecl in compiler/GHC/Parser/PostProcess.hs
splitLHsGadtTy in compiler/GHC/Hs/Type.hs
Pretty-printer:
pprConDecl in compiler/GHC/Hs/Decls.hs
pprHsForAllTelescope in compiler/GHC/Hs/Type.hs
Renamer:
rnConDecl in compiler/GHC/Rename/Module.hs
bindHsForAllTelescopes in compiler/GHC/Rename/HsType.hs
extractHsForAllTelescopes in compiler/GHC/Rename/HsType.hs
Type checker:
tcConDecl in compiler/GHC/Tc/TyCl.hs
tcGadtConTyVarBndrs in compiler/GHC/Tc/Gen/HsType.hs
Template Haskell
----------------
The TH AST is left unchanged for the moment to avoid breakage. An
attempt to quote or reify a data constructor declaration with visible
forall in its type will result an error:
data ThRejectionReason -- in GHC/HsToCore/Errors/Types.hs
= ...
| ThDataConVisibleForall -- new error constructor
However, as noted in the previous section, GHC now permits multiple
invisible foralls, and TH was updated accordingly. Updated code:
repC in compiler/GHC/HsToCore/Quote.hs
reifyDataCon in compiler/GHC/Tc/Gen/Splice.hs
ppr @Con in libraries/ghc-boot-th/GHC/Boot/TH/Ppr.hs
Pattern matching
----------------
Everything described above concerns data constructor declarations, but
what about their use sites? Now it is trickier to type check a pattern
match fn(Con a b c)=... because we can no longer assume that a,b,c are
all value arguments. Indeed, some or all of them may very well turn out
to be required type arguments.
To that end, see the changes to:
tcDataConPat in compiler/GHC/Tc/Gen/Pat.hs
splitConTyArgs in compiler/GHC/Tc/Gen/Pat.hs
and the new helpers split_con_ty_args, zip_pats_bndrs.
This is also the reason the TcRnTooManyTyArgsInConPattern error
constructor has been removed. The new code emits TcRnArityMismatch
or TcRnIllegalInvisibleTypePattern.
Summary
-------
DataCon, ConDecl, as well as all related functions have been updated to
support required type arguments in data constructors.
Test cases:
HieGadtConSigs GadtConSigs_th_dump1 GadtConSigs_th_pprint1
T25127_data T25127_data_inst T25127_infix
T25127_newtype T25127_fail_th_quote T25127_fail_arity
TyAppPat_Tricky
Co-authored-by: mniip <mniip@mniip.com>
-
ae003a3a
by Teo Camarasu at 2025-06-23T05:21:48-04:00
linters: lint-whitespace: bump upper-bound for containers
The version of containers was bumped in https://gitlab.haskell.org/ghc/ghc/-/merge_requests/13989
-
0fb37893
by Matthew Pickering at 2025-06-23T13:55:10-04:00
Move ModuleGraph into UnitEnv
The ModuleGraph is a piece of information associated with the
ExternalPackageState and HomeUnitGraph. Therefore we should store it
inside the HomeUnitEnv.
-
3bf6720e
by soulomoon at 2025-06-23T13:55:52-04:00
Remove hptAllFamInstances usage during upsweep
Fixes #26118
This change eliminates the use of hptAllFamInstances during the upsweep phase,
as it could access non-below modules from the home package table.
The following updates were made:
* Updated checkFamInstConsistency to accept an explicit ModuleEnv FamInstEnv
parameter and removed the call to hptAllFamInstances.
* Adjusted hugInstancesBelow so we can construct ModuleEnv FamInstEnv
from its result,
* hptAllFamInstances and allFamInstances functions are removed.
-
83ee7b78
by Ben Gamari at 2025-06-24T05:02:07-04:00
configure: Don't force value of OTOOL, etc. if not present
Previously if `otool` and `install_name_tool` were not present they
would be overridden by `fp_settings.m4`. This logic was introduced in
4ff93292243888545da452ea4d4c1987f2343591 without explanation.
-
9329c9e1
by Ben Gamari at 2025-06-24T05:02:07-04:00
ghc-toolchain: Add support for otool, install_name_tool
Fixes part of ghc#23675.
-
25f5c998
by Ben Gamari at 2025-06-24T05:02:08-04:00
ghc-toolchain: Add support for llc, opt, llvm-as
Fixes #23675.
-
51d150dd
by Rodrigo Mesquita at 2025-06-24T05:02:08-04:00
hadrian: Use settings-use-distro-mingw directly
The type `ToolchainSetting` only made sense when we had more settings to
fetch from the system config file. Even then "settings-use-distro-mingw"
is arguably not a toolchain setting.
With the fix for #23675, all toolchain tools were moved to the
`ghc-toolchain` `Toolchain` format. Therefore, we can inline
`settings-use-distro-mingw` accesses and delete `ToolchainSetting`.
-
dcf68a83
by Rodrigo Mesquita at 2025-06-24T05:02:08-04:00
configure: Check LlvmTarget exists for LlvmAsFlags
If LlvmTarget was empty, LlvmAsFlags would be just "--target=".
If it is empty now, simply keep LlvmAsFlags empty.
ghc-toolchain already does this right. This fix makes the two
configurations match up.
-
580a3353
by Ben Gamari at 2025-06-24T05:02:51-04:00
rts/linker/LoadArchive: Use bool
Improve type precision by using `bool` instead of `int` and `StgBool`.
-
76d1041d
by Ben Gamari at 2025-06-24T05:02:51-04:00
rts/linker/LoadArchive: Don't rely on file extensions for identification
Previously archive members would be identified via their file extension,
as described in #13103. We now instead use a more principled approach,
relying on the magic number in the member's header.
As well, we refactor treatment of archive format detection to improve
code clarity and error handling.
Closes #13103.
-
4b748a99
by Teo Camarasu at 2025-06-24T15:31:07-04:00
template-haskell: improve changelog
stable -> more stable, just to clarify that this interface isn't fully stable.
errornously -> mistakenly: I typod this and also let's go for a simpler word
-
e358e477
by Sylvain Henry at 2025-06-24T15:31:58-04:00
Bump stack resolver to use GHC 9.6.7
Cf #26139
-
4bf5eb63
by fendor at 2025-06-25T17:05:43-04:00
Teach `:reload` about multiple home units
`:reload` needs to lookup the `ModuleName` and must not assume the given
`ModuleName` is in the current `HomeUnit`.
We add a new utility function which allows us to find a `HomeUnitModule`
instead of a `Module`.
Further, we introduce the `GhciCommandError` type which can be used to
abort the execution of a GHCi command.
This error is caught and printed in a human readable fashion.
-
b3d97bb3
by fendor at 2025-06-25T17:06:25-04:00
Implement `-fno-load-initial-targets` flag
We add the new flag `-fno-load-initial-targets` which doesn't load all `Target`s
immediately but only computes the module graph for all `Target`s.
The user can then decide to load modules from that module graph using
the syntax:
ghci> :reload <Mod>
This will load everything in the module graph up to `Mod`.
The user can return to the initial state by using the builtin target
`none` to unload all modules.
ghci> :reload none
Is in principle identical to starting a new session with the
`-fno-load-initial-targets` flag.
The `-fno-load-initial-targets` flag allows for faster startup time of GHCi when a
user has lots of `Target`s.
We additionally extend the `:reload` command to accept multiple
`ModuleName`s. For example:
ghci> :reload <Mod1> <Mod2>
Loads all modules up to the modules `Mod1` and `Mod2`.
-
49f44e52
by Teo Camarasu at 2025-06-26T04:19:51-04:00
Expose ghc-internal unit id through the settings file
This in combination with the unit id of the compiler library allows
cabal to know of the two unit ids that should not be reinstalled (in
specific circumstances) as:
- when using plugins, we want to link against exactly the compiler unit
id
- when using TemplateHaskell we want to link against exactly the package
that contains the TemplateHaskell interfaces, which is `ghc-internal`
See: <https://github.com/haskell/cabal/issues/10087>
Resolves #25282
-
499c4efe
by Bryan Richter at 2025-06-26T04:20:33-04:00
CI: Fix and clean up capture of timings
* Fixes the typo that caused 'cat ci-timings' to report "no such file or
directory"
* Gave ci_timings.txt a file extension so it may play better with other
systems
* Fixed the use of time_it so all times are recorded
* Fixed time_it to print name along with timing
-
86c90c9e
by Bryan Richter at 2025-06-26T04:20:33-04:00
CI: Update collapsible section usage
The syntax apparently changed at some point.
-
04308ee4
by Bryan Richter at 2025-06-26T04:20:33-04:00
CI: Add more collapsible sections
-
43b606bb
by Florian Ragwitz at 2025-06-27T16:31:26-04:00
Tick uses of wildcard/pun field binds as if using the record selector function
Fixes #17834.
See Note [Record-selector ticks] for additional reasoning behind this as well
as an overview of the implementation details and future improvements.
-
d4952549
by Ben Gamari at 2025-06-27T16:32:08-04:00
testsuite/caller-cc: Make CallerCc[123] less sensitive
These were previously sensitive to irrelevant changes in program
structure. To avoid this we filter out all by lines emitted by the
-fcaller-cc from the profile.
-
cb4fbbcc
by Simon Peyton Jones at 2025-06-29T17:10:21+01:00
Renaming around predicate types
.. we were (as it turned out) abstracting over
type-class selectors in SPECIALISATION rules!
Wibble isEqPred
-
3c38ecc8
by Simon Peyton Jones at 2025-06-29T17:10:21+01:00
Refactor of Specialise.hs
This is a bit experimental. The original aim was to simplify
specUnfolding by getting rid of the dx_binds returned by specHeader;
but I'm worried about duplicating work so it's not finished.
Generally, though, `specHeader` is simpler and clearer which is good.
Alos add a few more `HasDebugCallStack` contexts.
-
40e7890a
by Simon Peyton Jones at 2025-06-29T17:10:39+01:00
Improve treatment of SPECIALISE pragmas -- again!
This MR does another major refactor of the way that SPECIALISE
pragmas work, to fix #26115, #26116, #26117.
* We now /always/ solve forall-constraints in an all-or-nothing way.
See Note [Solving a Wanted forall-constraint] in GHC.Tc.Solver.Solve
This means we might have unsolved quantified constraints, which need
to be reported. See `inert_insts` in `getUnsolvedInerts`.
* I refactored the short-cut solver for type classes to work by
recursively calling the solver rather than by having a little baby
solver that kept being not clever enough.
See Note [Shortcut solving] in GHC.Tc.Solver.Dict
* When solving a forall-constraint, we now solve it immediately,
rather than emitting an implication constraint to be solved later.
This saves quite a bit of plumbing; e.g
- The `wl_implics` field of `WorkList` is gone,
- The types of `solveSimpleWanteds` and friends are simplified.
- An EvFun contains binding, rather than an EvBindsVar ref-cell that
will in the future contain bindings. That makes `evVarsOfTerm`
simpler
* I totally rewrote the desugaring of SPECIALISE pragmas, again.
The new story is in Note [Desugaring new-form SPECIALISE pragmas]
in GHC.HsToCore.Binds
Both old-form and new-form SPECIALISE pragmas now route through the same
function `dsSpec_help`. The tricky function `decomposeRuleLhs` is now used only
for user-written RULES, not for SPECIALISE pragmas.
* A small refactor: `ebv_tcvs` in `EvBindsBar` now has a list of coercions, rather
than a set of tyvars. We just delay taking the free vars.
-
17b1a135
by Simon Peyton Jones at 2025-06-29T17:10:39+01:00
Do not look in inert_cans in lookupInInerts
...it bypasses all the shortcut stuff
-
51acab69
by Simon Peyton Jones at 2025-06-29T23:27:36+01:00
Simplify findDict and friends