GitLab

Simon Jakobi pushed to branch wip/sjakobi/T27176-clb at Glasgow Haskell Compiler / GHC

Commits:

5ac9ce7d

by Zubin Duggal at 2026-06-10T21:26:32+05:30

hadrian: Remove old package.conf files when generating new ones

Old package.conf files might exists with different hashes, causing issues like #26661

Fixes #26661

c9015f09

by sheaf at 2026-06-11T12:40:28-04:00

Fix AArch64 clobbering bug for MUL2

On AArch64, the code generator could clobber one of the input operands
when computing the lower bits of a MUL2 operation. This rendered invalid
the subsequent computation of the high bits.
This commit fixes that by using a temporary register. The register
allocator can remove the redundant move in the common case when the
registers do not conflict.

Fixes #27046

7ab90288

by Rodrigo Mesquita at 2026-06-11T12:41:11-04:00

fix: make T27131 less flaky

It seems that T27131 fails flakily in a race where we check the flag
before the capability had the chance to process the mailbox which sets
the flag. This seemingly should only happen if the capability ends up
being the same for setting and checking the flag.

8965cb76
by Marc Scholten at 2026-06-12T04:53:22-04:00
```
haddock: render modules concurrently
```

8cc0b64a

by Duncan Coutts at 2026-06-12T04:54:06-04:00

Promote HAVE_PREEMPTION from Timer.c to OSThreads.h

We will want to know about HAVE_PREEMPTION in more places.

HAVE_PREEMPTION tells us that we do have OS threads available,
irrespective of whether THREADED is defined. In particular,
HAVE_PREEMPTION is defined on all proper OSs, but not on WASM (and
hyopthetically may not be true on some other platforms like
micro-controllers, RTOSs, VM hypervisors etc).

cce574ed

by Duncan Coutts at 2026-06-12T04:54:06-04:00

Define ACQUIRE_LOCK_ALWAYS and friends

Fix issue #27335

Like the atomic _ALWAYS variants, these lock actions are always defined,
rather than being dependent on whether we are in the THREADED case. All
the "normal" LOCK macros are defined to be no-ops when !THREADED.

The use case for the _ALWAYS variants is where we are using OS threads
even in the non-threaded RTS. This includes everything to do with the
timer/ticker thread, which is used in the non-threaded RTS too.

In particular, we will want to use this for eventlog things, because the
timer thread performs eventlogging concurrently with the main
capability, even in the non-threaded RTS.

1f28d1f6

by Duncan Coutts at 2026-06-12T04:54:06-04:00

Use ACQUIRE/RELEASE_LOCK_ALWAYS with eventBufMutex

Even in the non-threaded RTS the eventBufMutex is needed by both the
main capability and the timer/ticker thread, so always use the mutex.

This should fix #25165 which is about the main capability and the timer
thread posting events to the eventlog buffer concurrently and thereby
corrupting the buffer data.

0ff29782

by Duncan Coutts at 2026-06-12T04:54:06-04:00

Expose eventBufMutex in the EventLog interface/header

We will need it in forkProcess to ensure we don't write to the global
eventlog buffer concurrently with trying to flush eventlog buffers and
do the fork().

7a688395

by Duncan Coutts at 2026-06-12T04:54:07-04:00

Split flushAllCapsEventsBufs into safe and unlocked version

Following the convention that unlocked versions have a trailing _
underscore in their name. This one requires the caller to hold the
eventlog global buffer mutex. We will need this in forkProcess.

341ed474

by Duncan Coutts at 2026-06-12T04:54:07-04:00

Remove redundant use of stopTimer in setNumCapabilities

Historically, the comment here was:

  We must stop the interval timer while we are changing the
  capabilities array lest handle_tick may try to context switch
  an old capability. See #17289.

and

  We must disable the timer while we do this since the tick handler may
  call contextSwitchAllCapabilities, which may see the capabilities array
  as we free it.

What this refers to is that historically, when changing the number of
capabilities, the array of capabilities was reallocated to a new size,
allocating new ones and freeing the old ones, thus invalidating all
existing capbility pointers.

Strangely, for good measure the code used to call stopTimer twice (hence
the two similar comments above).

However, since commit a3eccf06292dd666b24606251a52da2b466a9612, the
capabilities array is no longer reallocated. Instead the array is
allcoated once on RTS startup to the maximum size it could ever be
allowed to be, and then capabilities get enabled/disabled at runtime. So
the capability pointers never become invalid anymore. At worst, they may
point to capabilities that are disabled.

Thus we no longer need to stop the timer (twice) while we change the
number of enabled capabilities. This also partially solves issue #27105,
which notes that stopTimer is being used as if it were synchronous, when
it is not. At least for this case, the solution is that stopTimer is not
needed at all!

674858e3

by Duncan Coutts at 2026-06-12T04:54:07-04:00

Remove redundant use of stopTimer in forkProcess

but replace it with taking the eventlog buffer lock during the fork.

Fixes issue #27105

The original reason to block the timer during a fork was that
historically the timer was implemented using a periodic timer signal,
and the signal itself would interrupt the fork system call (returning
EINTR). For large processes (where fork() takes a while) this could
permanently livelock: the timer always would go off before the fork
could complete, which got retried in a loop forever.

The timer is no longer implemented as a unix signal, but uses threads.
Thus the original problem no longer exists. The only remaining reason to
block the timer tick is to prevent actions taken by the tick from
interfering with the delicate process involved in fork (taking a load of
locks and pausing everything).

The only thing we need to do is to prevent the eventlog from being
written to or flushed while the fork is taking place. To achieve this
all we need to do is hold the mutex for the global eventlog buffer.

This removes the last use of stopTimer that expects stopTimer to work
synchronously (which it was not) and thus solves issue #27105. To be
clear, we solve issue #27105 not by making stopTimer synchronous, but by
eliminating the use sites that expected it to be synchronous.

40764930

by sheaf at 2026-06-12T14:54:43-04:00

Add type family performance test for #26426

Some GHC versions produced large numbers of coercions after typechecking
and desugaring when compiling the program in #26426:

Version | Typechecker time | Typechecker allocations | Coercions
-------:|-----------------:|------------------------:|---------:
9.6     |   47 ms          |  48 MB                  |    110k
9.8     | 1000 ms          | 486 MB                  | 10,437k
9.10    |  922 ms          | 489 MB                  | 10,436k
9.12    |  906 ms          | 482 MB                  | 10,437k
9.14    |   63 ms          |  55 MB                  |    333k
10.0    |   47 ms          |  64 MB                  |     35k

The improvement 9.12 -> 9.14 was due to commit 22d11fa818fae2c95c494fc0fac1f8cb4c6e7cb6,
while the improvement 9.14 -> 10.0 was due to commit 0b7df6db9e46df40e86fbff1a66dc10440b99db5.

As the behaviour of GHC seems better than it's ever been on this program,
we declare victory, adding this performance test to ensure we don't
regress on this program.

On the way, we update Note [Combining equalities] in GHC.Tc.SolveR.Equality
with the explanation of the 9.12 -> 9.14 improvement (getting rid of an
exponential blowup in coercion sizes), and we update
Note [Exploiting closed type families] in GHC.Tc.Solver.FunDeps with
the explanation of the 9.14 -> 10.0 improvement (bringing down coercion
size growth from cubic to quadratic).

0f3d0a71

by Zubin Duggal at 2026-06-12T14:55:30-04:00

compiler: mark tool messages as errors/warnings depending on the exit code

Fixes #27370

d9ea2d76

by mangoiv at 2026-06-13T04:41:51-04:00

libraries/process: bump submodule to v1.6.30.0

- bump the submodule to the appropriate tag
- suppress benign warning resulting from the change

6ebaaba3

by David Eichmann at 2026-06-13T04:42:37-04:00

ghc-toolchain: don't throw when candidate executables are not found

Fixes #27369

6c65e1e1

by David Eichmann at 2026-06-13T04:43:23-04:00

CI: lint-changelog checks for no-changelog label in script instead of rules

bab37cc6

by konsumlamm at 2026-06-13T19:10:21+02:00

Implement CLC proposal #378

Add `Data.Double` and `Data.Float` modules
Document that GHC uses IEEE 754

fb5246ad

by fendor at 2026-06-15T18:07:23-04:00

Drop `preloadClosure` from `UnitState`

It is always hard-coded to the same value.

Backpack Unit instantiation isn't using it any more.
Allows us to simplify the API and get rid of `improveUnit`.

291ce3aa

by ARATA Mizuki at 2026-06-15T18:08:26-04:00

RISC-V NCG: Zero-extend the result of castFloatToWord32

According to the ISA manual, FMV.X.W sign-extends the result.
We need to truncate the result to avoid creating an exotic Word32 value.

Fixes #27300

011be91f

by ARATA Mizuki at 2026-06-15T18:08:26-04:00

RISC-V NCG: Treat d28-d31 (ft8-ft11) as caller-saved

According to the calling convention, the registers d28-d31 (ft8-ft11) are caller-saved.

Fixes #27306

e8a54713

by ARATA Mizuki at 2026-06-15T18:08:26-04:00

RISC-V NCG: Set rounding mode when emitting `truncate`

If we omit the rounding mode for `fcvt`, `dyn` will be used.
We do not want that for `truncate`, so we set `rtz`.
In other places, we set `rne` because we do not use the dynamic rounding mode.

Fixes #27303

9438bec7

by Zubin Duggal at 2026-06-15T18:09:11-04:00

rts: fix validate build with gcc 16. `__attribute__((regparm(1)))` is ignored on x86_64 and now
gcc warns that it is ignored:

rts/sm/Evac.h:35:1: error:
     error: ‘regparm’ attribute ignored [-Werror=attributes]

See https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=ccead81bbc39668376eb5cf47066acb446cc43f3

Fixes #27366

893e6133

by Andrew Lelechenko at 2026-06-15T23:55:36+01:00

base: more NonEmpty zips

CLC proposal: https://github.com/haskell/core-libraries-committee/issues/409

1314f2fd
by David Eichmann at 2026-06-16T05:46:52-04:00
```
Hadrian: fix ghc-internal .def file name
```

7f72bcb3

by mangoiv at 2026-06-16T05:47:39-04:00

compiler: ignore camelCase and Eta reduce hlint hints

These do not cohere with the style used in GHC. After disabling them,
hlint lints are much less noisy again.

842bef9f
by Alan Zimmerman at 2026-06-16T05:48:25-04:00
```
EPA: Use standard type family declaration for Anno
```

f6d30767

by Wolfgang Jeltsch at 2026-06-16T15:32:34-04:00

Fix two issues in the documentation of pipeline interruption

One issue is a typo (“interreuptible”), the other one the lack of an end
of a sentence, which has been reconstructed from the message of
633bbc1fd4762a2bb73ba1c5b9e0c279f1dd3c40, the commit that introduced
said documentation.

a3fa10e0

by Christian Georgii at 2026-06-16T15:33:26-04:00

Find plugins in sibling home units in multiple-home-unit sessions

In a multiple-home-unit session (e.g. `cabal repl --enable-multi-repl` or HLS), enabling a plugin with -fplugin that is defined in (or reexported by) another home unit failed with a "hidden package" error. The plugin module finder only searched the current home unit and the registered external packages, never the sibling home units.

findPluginModuleNoHsc now searches the home units that the current home unit depends on, following module reexports and respecting hidden modules, exactly as ordinary import resolution does in findImportedModuleNoHsc. To avoid two divergent copies of this logic, the shared home-unit search (the current home unit first, then its dependencies in priority order, with the accompanying ordering invariant) is extracted into findHomeModuleAmongDeps, which both findImportedModuleNoHsc and findPluginModuleNoHsc now call.

Add testsuite/tests/driver/multipleHomeUnits/plugin01, which loads a plugin as byte-code from a sibling home unit, and plugin02, in which the consumer enables a plugin reexported by a sibling home unit without depending on the plugin's own home unit directly.

Fixes #27349

d216412b

by Ian-Woo Kim at 2026-06-16T20:25:57-04:00

Make the order of usages deterministic

It has been observed that the ordering of usages can be non-determinstic
in parallel builds. Therefore, this contribution introduces sorting of
usages based on a platform- and race-independent sorting criterion.

Resolves #26877.

Co-authored-by: Wolfgang Jeltsch <wolfgang@well-typed.com>

8e1cc105

by Wolfgang Jeltsch at 2026-06-16T20:25:57-04:00

Change the descriptions of two existing changelog entries

The descriptions now describe the changes in a user-friendly manner, as
opposed to describing the contributions that led to these changes in a
developer-friendly manner.

636c1c7a

by Ian Duncan at 2026-06-16T20:26:50-04:00

AArch64: use SXTH, not SXTW, for W32 signExtendReg

signExtendReg was using SXTH (sign-extend halfword, 16-bit) for
W32-to-W64 sign extension. This should be SXTW (sign-extend word,
32-bit). SXTH only sign-extends the lower 16 bits, producing wrong
results for 32-bit values whose bit 15 differs from bit 31.

Other fixes:

  - At sub-W64, code gen for MO_S_Mul2 should use W32 registers for
    SMULL source operands as per the ARM spec (SMULL Xd, Wn, Wm),
    and not W64.

  - Ensure signExtendReg uses the source width for the source operand
    in SXTW/SXTH/SXTB instructions. GNU as requires sxtw Xd,Wn (not
    sxtw Xd,Xn), while LLVM's integrated assembler on macOS is lenient.

  - Fix overflow flag computation for `MO_S_Mul2`. The overflow bit
    was exactly inverted for sub-W64 operands.

Fixes #26978 and #27047

b734c75d

by Igor Ranieri at 2026-06-16T20:27:32-04:00

haddock: Update CONTRIBUTING with missing step, add missing test
dependency

7fe4f2ec

by Luite Stegeman at 2026-06-17T05:35:09-04:00

tag inference: don't confuse functions with their return values

inferTagRhs was mixing up taggedness for closures and return values
for function closures. We really shouldn't assign TagTuple to a
properly tagged function returning a tuple.

We fix this by keeping track of functions (TagFun) separately from
values (TagVal) and keeping track of their return value. TagFun is
also used for join points.

fixes #27005

4671c126

by Sebastian Graf at 2026-06-17T05:35:55-04:00

Seed the simplifier's in-scope set for open expressions

simplifyExpr simplifies expressions typed at the GHCi prompt and the
results of Template Haskell splices. Such an expression may be open: at a
GHCi debugger breakpoint its free variables include RuntimeUnk skolems
standing for as-yet-unknown types.

The simplifier began with an in-scope set holding only the wildcard
binder, so when it instantiated the unsafeCoerce# wrapper that GHCi
builds around a result, it formed a substitution whose range mentioned a
free skolem that was not in scope. That breaks the substitution invariant
and, in a compiler built with assertions, trips substTy's sanity check.

Seed the initial in-scope set with the free variables of the expression.
For a closed expression this adds nothing.

See Note [Seed the in-scope set for open expressions].

Fixes #17833 and its duplicate #21118.

67d41299

by Sebastian Graf at 2026-06-18T05:18:24-04:00

Desugar a `case` scrutinee only once (#27383, #20251)

In `dsExpr` for `HsCase` we desugared the scrutinee /twice/: once to
build the Core `case` itself, and again inside `matchWrapper`, which
re-desugared the source scrutinee (via `addHsScrutTmCs`) purely to
record long-distance information for the pattern-match checker.

For a single `case` that is merely wasteful.  But for nested cases it
is catastrophic.  Consider

    case (case (case e of ... ) of ... ) of ...

Desugaring the outer scrutinee desugars the middle `case` twice, each
of which desugars the inner `case` twice, and so on.  The work doubles
at every level, so desugaring takes O(2^n) time in the nesting depth.
That is the blowup reported in #27383; it is also what makes the
machine-generated program in #20251 take an age to compile.

The fix is simple.  `matchWrapper` is handed the scrutinee anyway, so
we give it the Core expression we have /already/ desugared, and record
the long-distance term constraint with `addCoreScrutTmCs` instead of
re-desugaring from source.  This is just what `matchSinglePatVar`
already does for single-pattern matches.

So:

* `matchWrapper` now takes `Maybe [CoreExpr]` rather than
  `Maybe [LHsExpr GhcTc]`.

* The `HsCase` equation of `dsExpr` passes the already-desugared
  `core_discrim`; the arrow desugarer passes its match variables.

* `addHsScrutTmCs` had no other use, so it is gone.

Desugaring is now linear in the nesting depth.  (The coverage checker
still runs `simpleOptExpr` over each scrutinee, which leaves the total
at O(n^2); that is ample.)  The long-distance information itself is
unchanged: the checker sees precisely the Core that backs the
generated code.

Test: deSugar/should_compile/T27383

fa5defde

by Rodrigo Mesquita at 2026-06-18T05:19:11-04:00

fix: Save FastStrings in the PMC

There is no point in adding the unique to the occurrence FastString we
create, since it is part of the Id anyway.

Adding it to the FastString, meant each FastString was unique
unnecessarily!

In a separate branch, running the compiler on test `InstanceMatching`
observed 30000 `FastString`s created by this code path.

Plus, `fsLit "pm"` follows the existing pattern in `mkPmId`.

4efb4a66

by Alan Zimmerman at 2026-06-18T14:41:14-04:00

TTG: Add extension points to HsConDetails

Extend HsConDetails as

    data HsConDetails p arg rec
      = PrefixCon !(XPrefixCon p) [arg]    -- C @t1 @t2 p1 p2 p3
      | RecCon    !(XRecCon p)    rec      -- C { x = p1, y = p2 }
      | InfixCon  !(XInfixCon p)  arg arg  -- p1 `C` p2
      | XHsConDetails !(XXHsConDetails p)

    type family XPrefixCon      p
    type family XRecCon         p
    type family XInfixCon       p
    type family XXHsConDetails  p

c8d27dd4

by Simon Jakobi at 2026-06-18T14:41:59-04:00

CI: quiet submodule clean output in after_script and setup

The clean and cleanup_submodules functions ran 'git submodule foreach
git clean -xdf', flooding the job log with 'Entering ...' and
'Removing ...' lines. Pass --quiet to 'git submodule' and -q to 'git
clean' to drop the success output; errors are still reported.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

09326ca6

by Matthew Pickering at 2026-06-20T23:41:12+02:00

Add missing req_interp modifier to T18441fail3 and T18441fail19

These tests require the interpreter but they were failing in a different
way with the javascript backend because the interpreter was disabled and
stderr is ignored by the test.

521e55bf

by Matthew Pickering at 2026-06-20T23:41:13+02:00

hadrian: Fill in more of the default.host toolchain file

When you are building a cross compiler this file will be used to build
stage1 and it's libraries, so we need enough information here to work
accurately. There is still more work to be done (see for example, word
size is still fixed).

23c9b6c3

by Matthew Pickering at 2026-06-20T23:42:52+02:00

hadrian: Build stage 2 cross compilers

* Most of hadrian is abstracted over the stage in order to remove the
  assumption that the target of all stages is the same platform. This
  allows the RTS to be built for two different targets for example.
* Abstracts the bindist creation logic to allow building either normal
  or cross bindists. Normal bindists use stage 1 libraries and a stage 2
  compiler. Cross bindists use stage 2 libararies and a stage 2
  compiler.
* hadrian: Make binary-dist-dir the default build target. This allows us
  to have the logic in one place about which libraries/stages to build
  with cross compilers. Fixes #24192

New hadrian target:

* `binary-dist-dir-cross`: Build a cross compiler bindist (compiler =
  stage 1, libraries = stage 2)

This commit also contains various changes to make stage2 compilers
feasible.

-------------------------
Metric Decrease:
    LinkableUsage02
    ManyAlternatives
    ManyConstructors
    MultiComponentModulesRecomp
    MultiLayerModulesRecomp
    RecordUpdPerf
    T10421
    T12150
    T12227
    T12425
    T12707
    T13035
    T13379
    T13820
    T15703
    T16577
    T18140
    T18282
    T18698a
    T18698b
    T18923
    T1969
    T20049
    T21839c
    T3294
    T4801
    T5030
    T5321FD
    T5321Fun
    T5631
    T5642
    T6048
    T783
    T9020
    T9198
    T9233
    T9630
    T9872d
    T9961
    parsing001
    T3064

Metric Increase:
    T26989
    hard_hole_fits
-------------------------

Co-authored-by: Sven Tennie <sven.tennie@gmail.com>

26fed8ab

by Matthew Pickering at 2026-06-20T23:42:52+02:00

ci: Test cross bindists

We remove the special logic for testing in-tree cross
compilers and instead test cross compiler bindists, like we do for all
other platforms.

80c8910e

by Matthew Pickering at 2026-06-20T23:42:52+02:00

ci: Introduce CROSS_STAGE variable

In preparation for building and testing stage3 bindists we introduce the
CROSS_STAGE variable which is used by a CI job to determine what kind of
bindist the CI job should produce.

At the moment we are only using CROSS_STAGE=2 but in the future we will
have some jobs which set CROSS_STAGE=3 to produce native bindists for a
target, but produced by a cross compiler, which can be tested on by
another CI job on the native platform.

CROSS_STAGE=2: Build a normal cross compiler bindist
CROSS_STAGE=3: Build a stage 3 bindist, one which is a native compiler and library for the target

8215573d