-
098bba70
by Facundo Domínguez at 2026-09-17T14:57:41-04:00
Remove the call to hoistFlag from processCmdLineP
This call was undesirable because it was causing a reconstruction
of the list of flag processors.
Even relying on list fusion would not be sufficient because we want
to replace the list with a data structure with faster lookups to
address #25763.
The removal is done in two moves:
* We move CmdLineP to CmdLine.hs
* Then we move the effect of hoisting to processArgs via a new function
hoistEwM
-
e01799c7
by Facundo Domínguez at 2026-09-17T14:57:42-04:00
Do not create duplicate flag processors for the -W flag
This makes simpler implementing the command line parser with a trie map
holding flag processors. If we were to allow duplicate flag processors
for any flag, the trie map would need to be a multimap.
-
3875b257
by Facundo Domínguez at 2026-09-17T14:57:42-04:00
Replace the list of flag processors with a trie map in `processArgs`
Previously `GHC.Driver.CmdLine.findArg` was called for each
command-line argument, doing a linear search on the list of flag
processors each time. This was somewhat costly as the list of flag
processor has near 2000 elements.
A lookup with a trie map is faster in general, but the speed up is
only noticeable when calling `GHC.Driver.CmdLine.processArgs` sufficient
times, so the amount of lookups piles up. This could happen in large
projects with many units each having many dependencies. If only a
module needs to be recompiled, the user still needs to wait for all of
the flag parsing to complete.
The benchmark at testsuite/tests/perf/compiler/cmdline-parser shows a
10x speed up when using the trie.
The trie map is implemented in `GHC.Data.StringTrie`.
The design is explained in Note [Processing command line argument fast]
in GHC.Driver.Session.
The commit also:
* adds two property-based tests for trie map operations,
* adds a benchmark of the command line parser, and
* renames the predicate `arg_ok` to `valid_flag_suffix` to decide what
flag processor should be used when there are multiple matching
prefixes on a flag.
Fixes #25763.
-
5236634a
by Andreas Klebinger at 2026-09-17T14:58:24-04:00
Compile Word64{Map,Set} with -O2 always.
This is the hottest code and even if we build GHC with -O0 otherwise -O2
pays off for these modules.
Fixes #27814.
-
545fd4fb
by ARATA Mizuki at 2026-09-22T15:03:24-04:00
Change the C types of hs_{bitrev,bswap}{16,32} to match the Haskell types
bitReverse{16,32}# and byteSwap{16,32}# have the type
Word# -> Word#, but their C correspondents took and
returned StgWord{16,32}.
The mismatch on bitReverse16# led to the problem reported
in #27299. Although no similar issue is reported for
bitReverse32# and byteSwap{16,32}#, it seemed wise to
change them proactively.
This commit fixes the mismatch by changing the C type.
Fixes #27299
-
bf17f289
by fendor at 2026-09-22T15:04:05-04:00
Don't lose instances from shadowed imported modules with the same name
When we look for instances below the current module, we need to check
the unit id as well when we check for whether the module is the current
module, otherwise we ignore modules we import that have the same name as
the import module.
Adds regression test.
-
53eead15
by Wolfgang Jeltsch at 2026-09-24T20:12:03-04:00
Correct the documentation of `explicitUnits`
Resolves #27787.
-
35bf6f4b
by Simon Jakobi at 2026-09-24T20:12:49-04:00
Remove stale specHeader bullet from Note [Don't make fillers for constraint types]
The Note said that Specialise.specHeader also calls mkAbsentFiller.
abfc224a27c removed that call: the specialiser no longer drops dead
arguments at all, see Note [Do not drop dead args from specialisations]
in GHC.Core.Opt.Specialise.
Assisted-by: Claude Fable 5.1
-
1abef1ef
by Wolfgang Jeltsch at 2026-09-25T17:02:07-04:00
Improve the `mhu-closure` makefile
-
a79a68ea
by Wolfgang Jeltsch at 2026-09-25T17:02:07-04:00
Make `ghci-multi` hide all unneeded packages
-
d6e72450
by Wolfgang Jeltsch at 2026-09-25T17:02:07-04:00
Improve the definition of the home unit closure property
-
c9a2766d
by Wolfgang Jeltsch at 2026-09-25T17:02:07-04:00
Re-implement the home unit closure check
Compared to the previous implementation, the new one has the following
advantages:
* It is correct.
- It distinguishes between units that have the same unit ID but
different ABI hashes.
- When `-hide-all-packages` is not used, it considers as home unit
dependencies also units that are made implicitly available
because they are in the package database.
* It seems to be faster in usual settings.
- In particular, it does not have a preparation phase in which it
merges the dependency information from all home units, so that
its running time does not grow linearly with the size of the
package database in normal Cabal scenarios.
* It is (hopefully) clearer.
* It is better documented.
-
fa96acf8
by Wolfgang Jeltsch at 2026-09-25T17:02:07-04:00
Remove erroneous GHC flag in `hadrian-multi`
The flag was `-Werror=-Wno-error=incomplete-record-selectors`. Its
purpose was to permit incomplete record selectors, but because of the
illegal combination of `-Werror=` and `-Wno-error` it was ignored,
without any negative consequences.
-
9f48a5b9
by Alan Zimmerman at 2026-09-25T17:02:50-04:00
EPA: Add specific layout field to AnnList
Up to now AnnList has effectively served to represent lists of things
having layout, regardless of whether they do or not.
In Haskell, layout is optional in many places, as there is also an
alternative syntax using braces and semi-colons, which is free-form.
This commit introduces AnnListLayout, a data structure capturing this
distinction.
It is deliberately set up similarly to EpLayout, so that these two can
be merged in time.
The transition is done mechanically, where AnnListBraces is not used
yet, the original Nothing case maps to AnnListNoLayout, and the Just
case to AnnListLayout.
Later commits will fine tune these against the actual parser layout.
-
234bab08
by Vladislav Zavialov at 2026-09-29T17:22:54-04:00
Add regression test for #7803
Commit 2effe18ab5, first shipped with GHC 8.2.1, fixed #7803 "Superclass
methods are left unspecialized".
This change turns the example from the ticket description into a regression
test, checking that foo1 is fully specialised at Double, with no calls to the
overloaded exp5Tail.
Closes #7803
Assisted-by: Claude Opus 5
-
c406135c
by Andreas Klebinger at 2026-10-01T06:57:28-04:00
RTS: Update flag parsing/handling:
Using 0 as argument to any of these flags now resets the flags setting to default:
- ``-n0`` disables nursery chunking.
- ``-K0`` (or a bare ``-K``) removes the stack size limit.
- ``-M0`` removes the heap size limit.
- ``-H0`` resets the suggested heap size, including the nursery scaling set by ``-H``.
- ``-AL0`` makes the large object allocation limit follow ``-A`` again.
GC Generations are now limited to 64.
-n larger than -A now emits a warning and disables chunking.
Fixes #27830 #27829
-
984ee82b
by Andreas Klebinger at 2026-10-01T06:58:10-04:00
Fix MultiLayerModules timeout multiplier.
We accidentally used the compile multiplier for ghci commands which
use the run timeout setting to set the timeout.
Fixes #26263 and #27796.
-
50a9cf26
by Andreas Klebinger at 2026-10-01T06:58:10-04:00
testsuite: Include runtime in JUnit report.
-
e358619f
by Alan Zimmerman at 2026-10-01T06:59:01-04:00
EPA: Use EpaLocation in EpVirtualBraces
Haddock processing uses EpLayout.EpVirtualBraces to record the current
layout column, for allocating haddock comments correctly.
For exact printing, we need the line as well as the col, so we know
how many lines to skip when the indented code block starts.
This commit updates EpLayout.EpVirtualBraces so that it carries an
EpaLocation instead of just the Int column, to allow it to be used for
both purposes.
It is a simple mechanical change, with no other impact, but it allows
us to use EpLayout instead of AnnListLayout in AnnList, in the next
commit in this series.
-
1d2a26ee
by Andrew Lelechenko at 2026-10-01T20:22:04-04:00
Remove bitSize from MINIMAL pragma of class Bits
Cf. https://github.com/haskell/core-libraries-committee/issues/441
The function is deprecated for 12 years and has a default implementation,
yet it was never removed from the MINIMAL pragma, putting library authors
in a very awkward position. Let's finally remove bitSize from MINIMAL.
-
3cc21e4c
by Hannes Siebenhandl at 2026-10-01T20:22:59-04:00
Add regression test for #27641
-
c35096f5
by Hannes Siebenhandl at 2026-10-01T20:23:00-04:00
Don't `:add` modules the loader will reject
Only home modules can be actually added to the loader context, thus we
only accept the module if it belongs to a home unit.
-
a563cb27
by Simon Jakobi at 2026-10-02T17:37:36-04:00
ci: accept modified changelog.d entries in lint-changelog
The changelog guide says to extend an existing entry rather than add a
new one when the feature is already documented. lint_changelog only
counted added files, so such MRs failed before the validator ran.
Fixes #27779.
Assisted-by: Claude Fable 5.1
-
03082cd4
by Simon Jakobi at 2026-10-02T17:38:17-04:00
changelog.d: Fix warning flag name in the semaphore-v2 entry
The entry refers to -Wsemaphore-version-mismatch, which does not exist.
The warning GHC emits when it cannot use a -jsem semaphore, including
on a protocol version mismatch, is -Wsemaphore-open-failure.
Context: !15729, #25087
Assisted-by: Claude Fable 5.1
-
a93705ee
by Simon Jakobi at 2026-10-02T17:38:17-04:00
changelog.d: Fix the CLC proposal link in the T27455 entry
The entry used the Sphinx :ref: role with a URL as its target. :ref:
resolves only internal labels, so the generated release notes would
show the proposal name without a link. Use the plain hyperlink form
like the other entries.
Context: !16274, #27455
Assisted-by: Claude Fable 5.1
-
54381079
by Hannes Siebenhandl at 2026-10-02T17:39:01-04:00
Extend the `TrustOverlay` for wired-in units
The TrustOverlay only contained the `UnitId` of all on-disk units.
However, wired-in units are referred to by their wired-in unit id, which
is not the same as the on-disk unit id.
Thus, trust checks for wired-in units failed because the lookup was
using the wired-in unit, which wasn't present in the `TrustOverlay`.
We fix #27834, by copying the trust info for wired-in unit ids.
-
f8cfb1a5
by Hécate Kleidukos at 2026-10-04T19:36:39+02:00
[CLC 341] Strict variants of mutable types
This commit is the implementation of CLC Proposal 341
https://github.com/haskell/core-libraries-committee/issues/341
-
575c4bc6
by Rodrigo Mesquita at 2026-10-05T12:32:02-04:00
ttg: HsModuleName in the AST, ModuleName in GHC
This commit makes the AST fully independent from GHC by parsing a
"module name" as an `HsModuleName (= HText)` rather than constructing an
`ModuleName (= FastString)` straight away.
In the GHC side, everything remains as a `ModuleName`. We turn AST
`HsModuleName`s into fast strings with the cheap `mkFastStringShortText`
when we process the AST for manipulation within the compiler.
There are no performance regressions introduced by this approach.
This is the final piece of #21592
Alternative plans were discussed in
gitlab.haskell.org/ghc/ghc/-/merge_requests/16546#note_694647,
but the Simon suggested one won in simplicity and performance when
compared against my attempt at implementing my original plan of having
a family `ModuleNameP`.
-
b3b89d32
by mangoiv at 2026-10-05T22:03:32-04:00
hadrian: Persist the used flavour name in the build root (#20650)
This executes the idea proposed in #20650, teaching Hadrian to persist
the flavour used to build a build root. This both makes use of multiple
build roots more ergonomic (as the user no longer needs to repeat
themselves) and avoids a tiresome foot-gun (forgetting to specify a
flavour, often wholly invalidating a build root).
Closes #20650.
Co-authored-by: Sylvain Henry <sylvain@haskus.fr>
-
f977837b
by mangoiv at 2026-10-05T22:04:21-04:00
configure: fix sed script and subst tooldir script
Fixes #27672
Adds sanity checking for SUBST_TOOLDIR
Metric Decrease:
LinkableUsage01
MultiLayerModulesTH_OneShot
T10547
T12234
T12425
T13035
T13820
T18140
T18730
T18923
T9198
TcPlugin_RewritePerf
mhu-perf
-
f1cb08be
by Hannes Siebenhandl at 2026-10-05T22:05:03-04:00
Prepend in-memory unit database for backpack
The packageDbFlags are stored in reverse order. Thus, prepending will
ensure that the in-memory unit database is the last item on the unit
database stack.
Further, the in-memory unit database needs to be able to shadow units
from earlier unit databases in the unit database stack.
Moreover, a `-clear-package-db` would remove the in-memory database if
it weren't the last item on the stack.
Closes #27835
-
8d49ed42
by Kazu Yamamoto at 2026-10-06T19:37:25-04:00
Do not leave standard file descriptors free at startup (#24450)
A process may be started with stdout closed -- `prog >&-` from a shell does
exactly that. Nothing then reserves descriptor 1, and open(), pipe() and
socket() hand out the lowest free descriptor, so the next descriptor the
runtime creates for itself takes it: the ticker's pipe, or a capability's IO
manager control pipe.
'GHC.Internal.IO.StdHandles.stdout' is a CAF, built the first time the program
writes to stdout, which is long after that has happened. It builds a 'Handle'
over descriptor 1, which now belongs to the runtime, and nothing reports an
error because the descriptor is perfectly good. The program then writes its
output into the runtime's own plumbing.
What that looks like from the outside is a hang. The descriptor the Handle has
taken over is the read end of a pipe, so poll() for POLLOUT on it never says
yes and epoll_ctl() accepts a registration for it; the thread waiting to write
blocks for ever. Closing the Handle at exit closes the runtime's descriptor,
which is worse: the ticker then fails on every tick.
'hs_init_ghc' now opens /dev/null onto whichever of descriptors 0, 1 and 2 are
closed, before the runtime creates any descriptor of its own. Each is opened
for the access that descriptor would normally have, read-only for 0 and
write-only for 1 and 2, so that writing to 'stdin' or reading from 'stdout'
still fails rather than quietly succeeding against /dev/null. See
Note [Standard file descriptors must be open].
Moving only the descriptors the runtime opens would not do: the first file the
program itself opened would land on the free number instead. The fix has to
leave no free number. A program run this way therefore finds stdout connected
to /dev/null rather than failing with EBADF on the first write, which is what
such a program would get from most other language runtimes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
-
65bcf9e9
by Simon Jakobi at 2026-10-06T19:38:09-04:00
changelog: Repair the sentence describing #27477
The description read "programs using could crash", the object of
"using" having been lost. The synopsis already says which programs are
affected, so drop the dangling word.
Assisted-by: Claude Fable 5.1
-
15ece296
by mangoiv at 2026-10-08T15:14:56-04:00
ci: move hadrian/ghci into the lint category so it can be picked up by more runners
Fixes #27817
-
c0823135
by mangoiv at 2026-10-08T15:14:56-04:00
ci: make lint-adjacent jobs into platform independent
- move lint jobs into lint-anyarch tag
- unify lint jobs under the same platform independent container
-
f76464cf
by Zubin Duggal at 2026-10-08T15:15:39-04:00
DmdAnal: Make AbsDmd neutral in maxDmd
fixes #27845
-
ffa56ae4
by sheaf at 2026-10-09T17:15:59+02:00
Preserve sign of zero when parsing -0.0#
This commit leverages the recently added GHC.Types.Literal.Floating
module in order to preserve the sign of negative zero from
source-written floating-point literals.
Fixes #27545
-
d538cac9
by Facundo Domínguez at 2026-10-09T20:26:38-04:00
Propagate annotations to simple interfaces
This allows annotations to be visible to plugins when run by
haddock. The LiquidHaskell plugin needs the annotations, in particular.
Fixes #26761
-
32c4f4ab
by meooow25 at 2026-10-09T20:27:28-04:00
Bump containers submodule to 0.8.1
And fix affected tests.
* T4086a: containers depends on template-haskell-lift now instead of
template-haskell.
* T22611: Updated to be less dependent on containers internals.
-
c786065e
by sheaf at 2026-10-09T20:28:20-04:00
Use an explicit pattern match in toIfaceIdDetails
The old wildcard pattern let this code bit rot: it did not properly account
for recently added 'IdDetails' such as 'RepPolyId', which resulted in
spurious 'pprTrace' messages that could even pollute HLS sessions.
Fixes #23785
-
e7cacf59
by sheaf at 2026-10-09T20:29:09-04:00
Take into account ifKind in freeNamesIfDecl
Commit 3c2f4bb4 added the 'ifKind' field to various constructors
of 'IfaceDecl', which preserves the original kind (before type synonym
expansion) for better error messages.
However, that commit forgot to update 'freeNamesIfDecl' correspondingly.
This could result in a "urk! lookup local fingerpring" crash when e.g.
a type's kind uses a type synonym defined later, as we failed to add a
dependency on the type synonym during dependency analysis.
Fixes #27916
-
41e80f41
by Simon Peyton Jones at 2026-10-11T14:38:51-04:00
Simplify the error-suppression mechanism in GHC.Tc.Errors
In `GHC.Tc.Errors.reportWanteds` we suppress some errors in favour of
others. But the mechanism had grown crufty, and #27331 exposed a bug.
This MR makes it simpler and more uniform, by doing everything via
the `ei_suppress` field of the ErrorItem.
Claude then found another closely-related bug (see the review linked
on !16564). That bug is immortalised in new test T18851d.
I'm still not happy with `ignoreConstraint` but we can worry about
that another day; I have not touched it.
Main changes
* Instead of /predicting/ when an error will be reported, we set
`cec_error_reported` when an error is /actually/ reported. (I renamed
it from `cec_suppress` to reflect this change.)
Now errors are suppressed only when a preceding error has actually been
reported. Simpler and more robust.
See `Note [cec_error_reported: suppressing less-serious errors]`
The big changes are in `reportWanteds` which looks very different.
* New `Note [ei_suppress: suppressing confusing errors]` describes the
(quite separate) `ei_suppress` mechanism for suppressing confusing
errors, mainly arising from non-empty rewriter sets
* In `tryTypeClassDefaulting`, instead of switching off type-class defaulting
if there is an insoluble /anywhere/, we just switch it off locally when
we find an insoluble sub-tree. See Note [When to do type-class defaulting]
and (W5) in Note [ApproximateWC]
* Greatly simplify `insolubleWantedCt`.
* I made type-class-defaulting be a bit more eager, only switched off when
a peer or ancestor (but not child) constraint is insoluble. See
Note [When to do type-class defaulting].
Small refactorings
* Rename `askNoErrs` to `askErrsFound` and flip its polarity
* The code for "check if ei_suppress is true of all items" code is
moved to `mkErrorItems` rather than in the large `reportWanteds`.
* Code for `has_gadt_match` moves to `mkGivenErrorReporter`, rather than
polluting the large `reportWanteds`.
* Many tests with `-fdefer-type-errors` get more warnings, especially
about ambiguous (Num a) constraints. This is by-design: see
Note [Always warn with -fdefer-type-errors]. Previously GHC was not
honouring this Note but now it is, hence the extra errors.
-
f217fc31
by Cheng Shao at 2026-10-11T14:39:33-04:00
compiler: ensure cross ghc --info output target unit ids
Previously, for cross ghcs, the base/ghc/ghc-internal unit ids in
--info output do not match the target unit ids reported by ghc-pkg,
breaking cross ghc release builds with hash-unit-ids when being used
with cabal 3.18. This patch fixes it by making --info output target
unit ids. Fixes #27919.
Co-authored-by: Codex <codex@openai.com>
-
b5227250
by Rodrigo Mesquita at 2026-10-15T16:38:09+01:00
Asynchronous RPC with external interpreter
There's a single pipe to connect the host compiler to the external
interpreter. On this pipe, we have historically done synchronous
request/reply communication to invoke methods on the external
interpreter.
However, with the advent of multi-threaded debugging, it has become
necessary to do asynchronous calls on the external interpreter, where we
may block on the host waiting for a reply, but the external interpreter
should handle the request asynchronously and reply when ready. For
instance, we want to start the evaluation of the debuggee but not block
the interpreter waiting for it to finish or hit a breakpoint. Or e.g. we
want to resume two threads of the debuggee simulatenously instead of
blocking waiting for the resume result of each one sequentially.
And it turns out, we were already doing some ad-hoc asynchronous
communication to evaluate Template Haskell on the external interpreter.
It used to go as an ad-hoc "send request without expecting reply",
followed by handling on the host a set of 'THMessage's sent by the
interpreter, followed by "expect a response to arrive now". It worked
because there was still a fixed order of messages, but it fits an async
model in a non-ad-hoc way.
Asynchronous communication that unblocks the external interpreter main
thread will also make it possible to resolve issues such as #18852
(exceptions leaving the external interpreter in an unrecoverrable state)
The new design
--------------
The new invariant: every single request should receive a matching reply.
The pipe is bidirectional (the ext. interpreter can receive requests,
but it can also send requests for TH evaluation; the host can send
requests, but it can also receive TH requests from the interpreter).
When a request is sent in whichever direction, it always contains a
correlation-id header. The Sender places an empty MVar on a
pending-reply map keyed by id and waits on it for a response.
See 'remoteCall' and 'remoteTHCall'.
The pipe is associated to a thread which drains the read end. Because
it's bidirectional, two things may be read off the pipe:
- A Reply, with a correlation-id header, which we match against the
pending requests map to unblock the corresponding pending-reply MVar
- A Request, which is queued to the pending-request queue.
To handle a message, we take a Request from the pending-request queue
and run a handler that produces a Reply. Currently, the handler is run
synchronously in the thread, but it is trivial to make the handler run
in a forked thread for certain ("interruptible") messages in an upcoming
patch (when it is necessary for the multi-threaded debugger).
See 'handleRemoteCall' and 'handleRemoteTHCall'.
The old interp lock now just protect the pipe writing (reads don't lock
bc they are done exclusively by drainPipe), it doesn't need nor should
lock the full remote procedure call.
Fixes #27960
-
cc116c23
by Rodrigo Mesquita at 2026-10-15T16:38:09+01:00
interp: Add separate mailboxes for TH async RPC
When the host sends a `RunTH` command to the interpreter, it expects a
few "reverse requests" from the interpreter (e.g. `NewName`) which must
be run in the same context that initiated the `RunTH`.
When the host is evaluating many splices at once, it will send many
`RunTH` commands to the interpreter and start listening for reverse
requests in the meantime. However, a received request previously had no
way to be associated to the scope of a particular `RunTH` invocation, so
the reverse commands could get replied to from incorrect contexts, or
e.g. a `RunTHDone` from the interp. could terminate early the wrong
`RunTH` computation.
In this commit we add independent mailboxes to the interpreter
asynchronous RPC infrastructure, allowing a RunTH to announce a
particular `InboxId` that all reverse requests for that RunTH should
use. On the host, we create that mailbox id and have
`handleRemoteTHCall` listen on that specific mailbox rather than for
requests without a specific mailbox address.
-
58a378d9
by Rodrigo Mesquita at 2026-10-15T16:38:10+01:00
refactor: Move Async RPC abstraction into GHCi.Message.Async
Carves out the Async RPC over the Pipe abstraction to a new module,
making the GHCi.Message more about the specific interpreter messages
rather than about the IPC protocol.
-
014202d0
by Rodrigo Mesquita at 2026-10-15T16:38:10+01:00
interp: If we have pending reads in the queue we can deliver them
We may still be able to reply even if the read end was closed.
As long as we have pending messages we should try to deliver them.
-
d707be17
by Andrea Vezzosi at 2026-10-15T16:38:10+01:00
interp: close Handles when done with Pipe
Ensures we close the pipe handles (if created using 'mkPipeFromHandles')
on `closePipe`, and do `closePipe` on `Shutdown`.
The thread draining the read end always `'finally'` closes the read end.
And `closePipe` kills the draining thread and closes the write end.
When `remoteAsyncReq`, we throw PipeReadClosedException if the pipe read is
closed, since we wouldn't be able to receive the reply.
It's fine for the drain thread not to close the write end, even if it
dies without `closePipe` (which does close the write end), because:
- requests pending a reply will get PipeClosedException (we'll not be
able to read any further "real" reply)
- new calls to `remoteAsyncReq` will throw PipeReadClosedException
- pending pipe writes, in response to an already read request, might still succeed
And the clean up code should call close pipe to ensure the write end is
closed as well.
So, if the drain thread dies naturally and the read end is closed, may
still be attempted to be used, but it will throw accordingly. The user
likely wants to catch this to close the write end as well.
-
66d87bca
by Zubin Duggal at 2026-10-16T15:18:00+05:30
interp: fix external interpreter shutdown deadlock on Windows
On shutdown both GHC and iserv call `closePipe`, which stopped the drain
thread before closing the write end. On Windows the drain thread reads
with a safe foreign call to _read, which `throwTo' can't interrupt, so
the two sides were stuck in the deadlock deadlock.
Closing the write end first is not enough on its own. On Windows a read
on a pipe only unblocks once all write handles to it are closed [1].
However, all all four pipe ends were inheritable and we didn't close the
child's ends in the parent. This means a close on any one side wouldn't
unblock any reads, becuase the other side had another copy open.
* runWithPipes now uses System.Process.CommunicationHandle [2], which is
built for passing pipes to a child process: only the child's ends are
inheritable, and it simplifies our mess of CPP. Additionally, it is
important to close our copy of the endpoints passed to the child so
each side only keeps open copies it needs to allow a close to unblock
reads.
* readGhcHandle makes the handle non-inheritable so we don't have
additional references floating around that stop reads from unblocking
on close.
* In closePipe the ordering is subtle and important. We first fail any
requests still waiting for a reply, then close the write end so the
other side sees EOF. We take the write lock to close it, so we never
close under a concurrent write [3]. Only then do we stop the drain
thread, which closes the read end itself once its read has finished.
Bump bound to process >= 1.6.20 as this is the version with
System.Process.CommunicationHandle
[1] https://learn.microsoft.com/en-us/windows/win32/ipc/anonymous-pipe-operations
[2] https://github.com/haskell/process/pull/308
[3] https://gitlab.haskell.org/ghc/ghc/-/work_items/27801
-
6f97d0cf
by Andrea Vezzosi at 2026-10-20T11:35:34+01:00
GHCi.Server: handle RunTH and RunModFinalizers in forked thread
This commit makes the external interpreter able to many interpret
template haskell splices in parallel without blocking the main server
thread. Simply: fork off a thread to handle a RunTH request. The handler
will issue the reverse-requests on the given RunTH mailbox, to keep each
RunTH isolated from the others.
`handleRemoteCall` was generalised slightly by allowing `WriteReply` to
be used linearly in the handler directly, so that the caller of
`handleRemoteCall` may `WriteReply` in the forked thread rather than in
the main thread.
The only question is how can we guarantee `WriteReply` is used linearly
if we want to use it inside a forked thread? We need two guarantees:
1. `WriteReply` is used in the forked thread regardless of whether an
exception is thrown to that thread.
2. The forked thread is always awaited for, so that the main program
doesn't exit before the thread has a chance to actually use
`WriteReply`.
We introduced a small `ThreadManager` which provides a `spawnFinally`
that provides these two guarantees:
1. is guaranteed by `forkFinally`, which always runs the clean up action
exactly once.
2. is guaranteed by `ThreadManager` being abstract and the only exposed
function that introduces `ThreadManager` being `withThreadManager`,
which always awaits for all spawned threads.
Co-authored-by: Rodrigo Mesquita <rodrigo.m.mesquita@gmail.com>
-
beb35d96
by Andrea Vezzosi at 2026-10-20T13:56:59+02:00
interp: spawn fillPipe thread, writePipe now safely interruptible
Previously an async exception during a call to `writePipe` could have caused a
partial write, making any further write impossible to parse from the other end.
Now `writePipe` merely pushes the message onto a queue, and the actual write
happens in a separate `fillPipe` thread. As a result of this and previous
changes, `interpCmd` no longer needs `interruptibleMask` to preserve the state
of the protocol.
`writePipe` waits for the actual write to complete, and rethrows exceptions
raised in the write attempt, if any.
-
926c35bb
by Rodrigo Mesquita at 2026-10-20T14:09:50+01:00
Global break action which uses shared thread map
the status Var must actually be global across all threads, otherwise all of them will block trying to write some Status Var no one is looking at at all
have to fix these toggle untoggle things
very satisfying solution with mutable state
Matthew-certified big brain
WIP
HMmm almost there I think
pretty great progressssssssss
ExecOptions isolateThreadBreaks
ThreadBreaksIsolationMode
move to GHCi.Run.Breakpoints
-
ca7c44a5
by Rodrigo Mesquita at 2026-10-20T14:09:50+01:00
Refactor: we have the interactive Ids directly... clean up significantly
Removes things like
names <- map idName ids
map (\i -> let id = find (\x -> i==idName x) names) idso
and is more straightforward because there's less state going around (no
more updating HscEnv as an output)
It also kills `printTypeOfNames` completely, since now we just have the
Ids. No longer `lookupName` names to get back the ids...
```
printTypeOfName :: GHC.GhcMonad m => Name -> m ()
printTypeOfName n
= do maybe_tything <- GHC.lookupName n
case maybe_tything of
Nothing -> return ()
Just thing -> printTyThing thing
```
now is just `printTyThing . AnId`
Also:
```diff
- tythings <- catMaybes `liftM` mapM GHC.lookupName namesSorted
- docs <- mapM pprTypeAndContents [i | AnId i <- tythings]
+ docs <- mapM pprTypeAndContents idsSorted
```
(and yes, in ExecBreaks there are only Ids, no TyThings, bc they are
only constructed by the interpreter execution which can only bind ids.
for TyThings use runDecls which doesn't return execBreaks.)
I have a plan for the `resumes`, I'm getting there.
fixes
grrr
UNDO the removing-hsc-env-globality part of the clean up and keep just the good parts
fixes
-
b35d5c73
by Rodrigo Mesquita at 2026-10-20T14:09:50+01:00
Make the ExecBreak contain a Resume
Put the resume in ExecBreak
-
f052dd80
by Rodrigo Mesquita at 2026-10-20T14:09:50+01:00
Make resumeExec receive a Resume as an argument
-
9ff76575
by Rodrigo Mesquita at 2026-10-20T14:09:50+01:00
Build EvalOpts once per exec rather than per breakpoint stop
in interpreter_steplocal the allocations increased because now we're
updating `eval_opts` after constructing them in `handleRunStatus` (which
is run many times since in step-local most often we don't stop, just
check).
So drop the update to the eval_opts, and, in fact, pull their
construction out. the loop always uses the same eval_opts, so no point
in constructing them inside from scratch every time.
-3.6% from this branch which had the regressions in
interpreter_steplocal
-
21fee4b3
by Rodrigo Mesquita at 2026-10-20T14:09:50+01:00
Look up a thread's ResumeContext without taking a lock
modifyMVar is unnecessarily expensive and blocking when most often we
just care about reading without modifying. optimize for default case
(thread context already exists). It is called every single time a
breakpoint is hit (regardless of whether we stop there, e.g. in
interpreter_steplocal we do not stop there often)
-3.3% improvement over original regression
-
e8a5011a
by Rodrigo Mesquita at 2026-10-20T14:09:51+01:00
Every thread resume-context keeps a var for its EvalStatus
Kill the global mapping from thread id to EvalStatus
Instead, keep the EvalStatus in the ResumeContext associated to a thread
in the threadContextVar (where entries are created on breakpoints being
hit and when writing an EvalComplete result)
Before, every breakpoint hit meant reading the global map TVar and
deleting a map entry which did a lot of allocations. Now, we have a
Nothing variable which we overwrite to "take" rather than deleting from
a Map and allocating the GlobalDbgMap too. Note that for every
breakpoint hit we'd have to insert the ThreadId entry in the map again
(because we deleted it last time we resumed), again allocating
unnecessarily. Now it's only a look up on the map and an update on the
variable directly.
And a LOT of breakpoints were hit on interpreter_steplocal, stress
testing the cost of hitting a breakpoint
The isolated threads move to their own top-level TVar Set.
-4.4% improvement over the regression
-
3b6556e0
by Rodrigo Mesquita at 2026-10-20T14:09:51+01:00
Move whether a thread is isolated to the same global map
Stops us from reconstructing the set every time this changed instead to
being a per thread variable just toggled on/off, by default non-isolated
like before (before it was Set.empty, Set.insert for True, Set.delete
for FALSE). This is more straightforward -- there's always an isolation
mode per thread and we can toggle it directly
reduces slightly allocations, but the main advantage is killing a second
global map and making it more uniform with just one thread map and no
membership check for toggledness.
-
6e22a30f
by Rodrigo Mesquita at 2026-10-20T14:09:51+01:00
Prefer foldl' in readAnyThreadEvalBreak txn
retry 'orElse' next is just next, so we might as well use a strict fold
-
03cf2a12
by Rodrigo Mesquita at 2026-10-20T14:09:51+01:00
Add inline TODO about improving efficiency of readyAnyThreadBreak
It is currently O(N) in number of threads that have hit breakpoints
It should be number of *live* threads that have hit breaks, bc we still need to find a way to evacuate threads from the map
-
97da5f78
by Rodrigo Mesquita at 2026-10-20T14:09:51+01:00
Set global breakpoint action in sandboxIO rather than withBreakAction
-
67a6b65b
by Rodrigo Mesquita at 2026-10-20T14:09:51+01:00
TODO comment on breakOnException being global rather than per-thread
-
042b5d77
by Rodrigo Mesquita at 2026-10-20T14:31:23+01:00
fixmeup
-
4501a016
by Rodrigo Mesquita at 2026-10-20T15:43:54+01:00
WIP: Asynchronous eval/resume
-
a1ec78a5
by Rodrigo Mesquita at 2026-10-20T16:11:17+01:00
fixup! WIP: Asynchronous eval/resume
-
c73a96b6
by Rodrigo Mesquita at 2026-10-20T16:36:00+01:00
Fix commentary
-
8a1f45eb
by Rodrigo Mesquita at 2026-10-20T16:37:53+01:00
fix
-
2f52e43f
by Rodrigo Mesquita at 2026-10-20T17:01:17+01:00
more fixes