GitLab

Sven Tennie pushed to branch wip/romes/hadrian-cross-stage2-rebase_SVEN_FINAL at Glasgow Haskell Compiler / GHC

Commits:

43638643

by Andreas Klebinger at 2026-03-15T18:15:48-04:00

Configure: Fix check for --target support in stage0 CC

The check FP_PROG_CC_LINKER_TARGET used $CC unconditionally to check for
--target support. However this fails for the stage0 config where the C
compiler used is not $CC but $CC_STAGE0.

Since we already pass the compiler under test into the macro I simply
changed it to use that instead.

Fixes #26999

18fd0df6
by Simon Hengel at 2026-03-15T18:16:33-04:00
```
Fix typo in recursive_do.rst
```

86bd9bfc

by fendor at 2026-03-17T23:46:09-04:00

Introduce `-fimport-loaded-targets` GHCi flag

This new flag automatically adds all loaded targets to the GHCi session
by adding an `InteractiveImport` for the loaded targets.

By default, this flag is disabled, as it potentially increases memory-usage.

This interacts with the flag `-fno-load-initial-targets` as follows:

* If no module is loaded, no module is added as an interactive import.
* If a reload loads up to a module, all loaded modules are added as
  interactive imports.
* Unloading modules removes them from the interactive context.

Fixes #26866 by rendering the use of a `-ghci-script` to achieve the
same thing redundant.

e3d4c1bb

by mniip at 2026-03-17T23:47:03-04:00

ghc-internal: Remove GHC.Internal.Data.Eq

It served no purpose other than being a re-export.

6f4f6cf0

by mniip at 2026-03-17T23:47:03-04:00

ghc-internal: Refine GHC.Internal.Base imports

Removed re-exports from GHC.Internal.Base. This reveals some modules
that don't actually use anything *defined* in GHC.Internal.Base, and
that can be pushed down a little in the import graph.

Replaced most imports of GHC.Internal.Base with non-wildcard imports
from modules where the identifiers are actually defined.

Part of #26834

Metric Decrease:
    T5321FD

7fb51f54

by mangoiv at 2026-03-17T23:48:00-04:00

ci: clone, don't copy when creating the cabal cache

Also removed WINDOWS_HOST variable detected via uname - we now just
check whether the CI job has windows in its name. This works because we
only ever care about it if the respective job is not a cross job. We
also statically detect darwin cross jobs in the same way. We only ever have
darwin -> darwin cross jobs so this is enough to detect the host
reliably.

f8817879

by mangoiv at 2026-03-17T23:48:44-04:00

ci: mark size_hello_artifact fragile on darwin x86

The size of the x86_64 hello artifact is not stable which results in flaky testruns.

Resolves #26814

e34cb6da

by Adam Gundry at 2026-03-20T12:20:00-04:00

ghci: Mention active language edition in startup banner

Per GHC proposal 632, this makes the GHCi startup banner include
the active language edition, plus an indication of whether this
was the default (as opposed to being explicitly selected via an
option such as `-XGHC2024`). For example:

```
$ ghci
GHCi, version 9.14.1: https://www.haskell.org/ghc/  :? for help
Using default language edition: GHC2024
ghci>
```

Fixes #26037.

52c3e6ba

by sheaf at 2026-03-20T12:21:09-04:00

Improve incomplete record selector warnings

This commit stops GHC from emitting spurious incomplete record selector
warnings for bare selectors/projections such as .fld

There are two places we currently emit incomplete record selector
warnings:

  1. In the desugarer, when we see a record selector or an occurrence
     of 'getField'. Here, we can use pattern matching information to
     ensure we don't give false positives.
  2. In the typechecker, which might sometimes give false positives but
     can emit warnings in cases that the pattern match checker would
     otherwise miss.

This is explained in Note [Detecting incomplete record selectors]
in GHC.HsToCore.Pmc.

Now, we obviously don't want to emit the same error twice, and generally
we prefer (1), as those messages contain fewer false positives. So we
suppress (2) when we are sure we are going to emit (1); the logic for
doing so is in GHC.Tc.Instance.Class.warnIncompleteRecSel,
and works by looking at the CtOrigin.

Now, the issue was that this logic handled explicit record selectors as
well as overloaded record field selectors such as "x.r" (which turns
into a simple GetFieldOrigin CtOrigin), but it didn't properly handle
record projectors like ".fld" or ".fld1.fld2" (which result in other
CtOrigins such as 'RecordFieldProjectionOrigin').

To solve this problem, we re-use the 'isHasFieldOrigin' introduced in
fbdc623a (slightly adjusted).

On the way, we also had to update the desugarer with special handling
for the 'ExpandedThingTc' case in 'ds_app', to make sure that
'ds_app_var' sees all the type arguments to 'getField' in order for it
to indeed emit warnings like in (1).

Fixes #26686

309d7e87

by Cheng Shao at 2026-03-20T12:21:53-04:00

rts: opportunistically grow the MutableByteArray# in-place in resizeMutableByteArray#

Following !15234, this patch improves `resizeMutableByteArray#` memory
efficiency by growing the `MutableByteArray#` in-place if possible,
addressing an old todo comment here. Also adds a new test case
`resizeMutableByteArrayInPlace` that stresses this behavior.

7d4ef162

by Matthew Craven at 2026-03-20T12:22:47-04:00

Change representation of floating point literals

This commit changes the representation of floating point literals
throughough the compiler, in particular in Core and Cmm.
The Rational type is deficient for this purpose, dealing poorly
with NaN, +/-Infinity, and negative zero. Instead, the new module
GHC.Types.Literal.Floating uses the host Float/Double type to represent
NaNs, infinities and negative zero. It also contains a Rational
constructor, for the benefit of -fexcess-precision.

Other changes:

  - Remove Note [negative zero] and related code
    This also removes the restrictions on constant-folding of division
    by zero, and should make any problems with NaN/Infinity more obvious.

  - Use -0.0 as the additive identity for Core constant folding rules
    for floating-point addition, fixing #21227.

  - Manual worker-wrapper for GHC.Float.rationalToDouble. This is
    intended to prevent the compiler's WW on this function from
    interfering with constant-folding. This change means that we now
    avoid allocating a box for the result of a 'realToFrac' call in
    T10359.

  - Combine floatDecodeOp and doubleDecodeOp.
    This change also fixes a bug in doubleDecodeOp wherein it
    would incorrectly produce an Int# instead of an Int64#
    literal for the mantissa component with 64-bit targets.

  - Use Float/Double for assembly immediates, and update the X86 and
    PowerPC backends to properly handle special values such as NaN and
    infinity.

  - Allow 'rational_to' to handle zero denominators, fixing a
    TODO in GHC.Core.Opt.ConstantFold.

Fixes #8364 #9811 #18897 #21227
Progress towards #26919

Metric Decrease:
    T10359

Co-authored-by: sheaf <sam.derbyshire@gmail.com>

-------------------------
Metric Decrease:
    T1969
    T5321FD
-------------------------

80e2dd4f

by Zubin Duggal at 2026-03-20T12:23:33-04:00

compiler/ffi: Collapse void pointer chains in capi wrappers

New gcc/clang treat -Wincompatible-pointer-types as an error by
default. Since C only allows implicit conversion from void*, not void**,
capi wrappers for functions taking e.g. abstract** would fail to compile
when the Haskell type Ptr (Ptr Abstract) was naively translated to void**.

Collapse nested void pointers to a single void* when the pointee type
has no known C representation.

Fixes #26852

1c50bd7b

by Luite Stegeman at 2026-03-20T12:24:37-04:00

Move some functions related to pointer tagging to a separate module

bfd7aafd

by Luite Stegeman at 2026-03-20T12:24:37-04:00

Branchless unpacking for enumeration types

Change unpacking for enumeration types to go to Word8#/Word16#/Word#
directly instead of going through an intermediate unboxed sum. This
allows us to do a branchless conversion using DataToTag and TagToEnum.

Fixes #26970

72b20fc0

by Luite Stegeman at 2026-03-20T12:25:30-04:00

bytecode: Carefully SLIDE off the end of a stack chunk

The SLIDE bytecode instruction was not checking for stack chunk
boundaries and could corrupt the stack underflow frame, leading
to crashes.

We add a check to use safe writes if we cross the chunk boundary
and also handle stack underflow if Sp is advanced past the underflow
frame.

fix #27001

2e22b43c

by Cheng Shao at 2026-03-20T12:26:14-04:00

ghci: serialize BCOByteArray buffer directly when possible

This patch changes the `Binary` instances of `BCOByteArray` to
directly serialize the underlying buffer when possible, while also
taking into account the issue of host-dependent `Word` width. See
added comments and amended `Note [BCOByteArray serialization]` for
detailed explanation. Closes #27020.

89d9ba37

by Sylvain Henry at 2026-03-20T12:27:34-04:00

JS: replace BigInt with Number arithmetic for 32/64-bit quot/rem (#23597)

Replace BigInt-based implementations of quotWord32, remWord32,
quotRemWord32, quotRem2Word32, quotWord64, remWord64, quotInt64, and
remInt64 with pure Number (double/integer) arithmetic to avoid the
overhead of BigInt promotion.

ae4ddd60

by Sylvain Henry at 2026-03-20T12:28:28-04:00

Core: add constant-folding rules for Addr# eq/ne (#18032)

3e767f98

by Matthew Pickering at 2026-03-20T12:29:11-04:00

Use OsPath rather than FilePath in Downsweep cache

This gets us one step closure to uniformly using `OsPath` in the
compiler.

2c57de29

by Cheng Shao at 2026-03-20T12:29:55-04:00

hadrian: fix ghc-in-ghci flavour stage0 shared libraries

This patch fixes missing stage0 shared libraries in hadrian
ghc-in-ghci flavour, which was accidentally dropped in
669d09f950a6e88b903d9fd8a7571531774d4d5d and resulted in a regression
in HLS support on linux/macos. Fixes #27057.

5b1be555

by Sylvain Henry at 2026-03-20T12:30:48-04:00

JS: install rts/Types.h header file (#27033)

It was an omission, making HsFFI.h not usable with GHC using the JS
backend.

b883f08f

by Cheng Shao at 2026-03-20T12:31:33-04:00

hadrian: don't compile RTS with -Winline

This patch removes `-Winline` from cflags when compiling the RTS,
given that:

1. It generates a huge pile of spam and hurts developer experience
2. Whether inlining happens is highly dependent on toolchains,
   flavours, etc, and it's not really an issue to fix if inlining
   doesn't happen; it's a hint to the C compiler anyway.

Fixes #27060.

333387d6

by Cheng Shao at 2026-03-20T12:31:33-04:00

hadrian: compile libffi-clib with -Wno-deprecated-declarations

This patch adds `-Wno-deprecated-declarations` to cflags of
`libffi-clib`, given that it produces noise at compile-time that
aren't really our issue to fix anyway, it's from vendored libffi
source code.

67c47771

by Rodrigo Mesquita at 2026-03-20T12:32:17-04:00

Expose decodeStackWithIpe from ghc-experimental

This decoding is useful to the debugger and it wasn't originally
exported as an oversight.

d30c741b

by Matthew Pickering at 2026-03-21T07:39:07+00:00

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.

f1566889

by Matthew Pickering at 2026-03-21T07:39:07+00:00

packaging: correctly propagate build/host/target to bindist configure script

At the moment the host and target which we will produce a compiler for
is fixed at the initial configure time. Therefore we need to persist
the choice made at this time into the installation bindist as well so we
look for the right tools, with the right prefixes at install time.

In the future, we want to provide a bit more control about what kind of
bindist we produce so the logic about what the host/target will have to
be written by hadrian rather than persisted by the configure script. In
particular with cross compilers we want to either build a normal stage 2
cross bindist or a stage 3 bindist, which creates a bindist which has a
native compiler for the target platform.

Fixes #21970

5558b428

by Matthew Pickering at 2026-03-21T07:39:07+00:00

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).

5181051c

by Matthew Pickering at 2026-03-21T07:39:07+00:00

hadrian: Disable docs when cross compiling

Before there were a variety of ad-hoc places where doc building was
disabled when cross compiling.

* Some CI jobs sets --docs=none in gen_ci.hs
* Some CI jobs set --docs=none in .gitlab/ci.sh
* There was some logic in hadrian to not need the ["docs"] target when
  making a bindist.

Now the situation is simple:

* If you are cross compiling then defaultDocsTargets is empty by
  default.

In theory, there is no reason why we can't build documentation for cross
compiler bindists, but this is left to future work to generalise the
documentation building rules to allow this (#24289)

11252908

by Matthew Pickering at 2026-03-21T08:03:20+00:00

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:
    ManyAlternatives
    MultiComponentModulesRecomp
    MultiLayerModulesRecomp
    T10421
    T12425
    T12707
    T13035
    T13379
    T15703
    T16577
    T18698a
    T18698b
    T18923
    T1969
    T21839c
    T3294
    T4801
    T5030
    T5321Fun
    T5642
    T783
    T9198
    T9872d
    T9961
    parsing001
    T5321FD
    T6048
    T12227
    T18140
    T18282
    T9233
    T5631
    T9630
-------------------------

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

Fix rebase: settings-use-distro-mingw is now staged

15d0fd7c

by Matthew Pickering at 2026-03-21T08:03:20+00:00

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.

e9eac8f9

by Matthew Pickering at 2026-03-21T08:03:20+00:00

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

da8f04ff

by Matthew Pickering at 2026-03-21T08:03:20+00:00

hadrian: Refactor system-cxx-std-lib rules0

I noticed a few things wrong with the hadrian rules for `system-cxx-std-lib` rules.

* For `text` there is an ad-hoc check to depend on `system-cxx-std-lib` outside of `configurePackage`.
* The `system-cxx-std-lib` dependency is not read from cabal files.
* Recache is not called on the packge database after the `.conf` file is generated, a more natural place for this rule is `registerRules`.

Treating this uniformly like other packages is complicated by it not having any source code or a cabal file. However we can do a bit better by reporting the dependency firstly in `PackageData` and then needing the `.conf` file in the same place as every other package in `configurePackage`.

Fixes #25303

386ab054