-
ec3cf767
by Cheng Shao at 2026-01-08T06:24:31-05:00
make: remove unused Makefiles from legacy make build system
This patch removes unused Makefiles from legacy make build system; now
they are never used by hadrian in any way, and they already include
common boilerplate mk files that are long gone in the make build
system removal, hence the housecleaning.
Co-authored-by: Codex <codex@openai.com>
-
04ea3f83
by Cheng Shao at 2026-01-08T06:25:13-05:00
compiler: use -O3 as LLVM optimization level for ghc -O2
The GHC driver clamps LLVM optimization level to `-O2` due to LLVM
crashes, but those were historical issues many years ago that are no
longer relevant for LLVM versions we support today. This patch changes
the driver to use `-O3` as LLVM optimization level when compiling with
`-O2`, which is a better default when we're willing to trade
compilation time for faster generated code.
-
472df471
by Peter Trommler at 2026-01-08T13:28:54-05:00
Use half-word literals in info tables
With this commit info tables are mapped to the same assembler code
on big-endian and little-endian platforms.
Fixes #26579.
-
393f9c51
by Simon Peyton Jones at 2026-01-08T13:29:35-05:00
Refactor srutOkForBinderSwap
This MR does a small refactor:
* Moves `scrutOkForBinderSwap` and `BinderSwapDecision`
to GHC.Core.Utils
* Inverts the sense of the coercion it returns, which makes
more sense
No effect on behaviour
-
ad76fb0f
by Simon Peyton Jones at 2026-01-08T13:29:36-05:00
Improve case merging
This small MR makes case merging happen a bit more often than
it otherwise could, by getting join points out of the way.
See #26709 and GHC.Core.Utils
Note [Floating join points out of DEFAULT alternatives]
-
4c9395f5
by Cheng Shao at 2026-01-08T13:30:16-05:00
hadrian: remove broken hsc2hs flag when cross compiling to windows
This patch removes the `--via-asm` hsc2hs flag when cross compiling to
windows. With recent llvm-mingw toolchain, it would fail with:
```
x86_64-w64-mingw32-hsc2hs: Cannot combine instructions: [Quad 8,Long 4,Long 241,Ref ".Ltmp1-.Ltmp0"]
```
The hsc2hs default `--cross-compile` logic is slower but works.
-
71fdef55
by Simon Peyton Jones at 2026-01-08T13:30:57-05:00
Try harder to keep the substitution empty
Avoid unnecessary cloning of variables in the Simplifier.
Addresses #26724,
See Note [Keeping the substitution empty]
We get some big wins in compile time
Metrics: compile_time/bytes allocated
-------------------------------------
Baseline
Test Metric value New value Change
----------------------------------------------------------------------------
CoOpt_Singletons(normal) ghc/alloc 721,544,088 692,174,216 -4.1% GOOD
LargeRecord(normal) ghc/alloc 1,268,031,157 1,265,168,448 -0.2%
T14766(normal) ghc/alloc 918,218,533 688,432,296 -25.0% GOOD
T15703(normal) ghc/alloc 318,103,629 306,638,016 -3.6% GOOD
T17836(normal) ghc/alloc 419,174,584 418,400,824 -0.2%
T18478(normal) ghc/alloc 471,042,976 470,261,376 -0.2%
T20261(normal) ghc/alloc 573,387,162 563,663,336 -1.7%
T24984(normal) ghc/alloc 87,832,666 87,636,168 -0.2%
T25196(optasm) ghc/alloc 1,103,284,040 1,101,376,992 -0.2%
hard_hole_fits(normal) ghc/alloc 224,981,413 224,608,208 -0.2%
geo. mean -0.3%
minimum -25.0%
maximum +0.1%
Metric Decrease:
CoOpt_Singletons
T14766
T15703
-
30341168
by Simon Peyton Jones at 2026-01-08T13:31:38-05:00
Add regression test for #24867
-
1ac1a541
by Julian Ospald at 2026-01-09T02:48:53-05:00
Support statically linking executables properly
Fixes #26434
In detail, this does a number of things:
* Makes GHC aware of 'extra-libraries-static' (this changes the package
database format).
* Adds a switch '-static-external' that will honour 'extra-libraries-static'
to link external system dependencies statically.
* Adds a new field to settings/targets: "ld supports verbatim namespace".
This field is used by '-static-external' to conditionally use '-l:foo.a'
syntax during linking, which is more robust than trying to find the
absolute path to an archive on our own.
* Adds a switch '-fully-static' that is meant as a high-level interface
for e.g. cabal. This also honours 'extra-libraries-static'.
This also attempts to clean up the confusion around library search directories.
At the moment, we have 3 types of directories in the package database
format:
* library-dirs
* library-dirs-static
* dynamic-library-dirs
However, we only have two types of linking: dynamic or static. Given the
existing logic in 'mungeDynLibFields', this patch assumes that
'library-dirs' is really just nothing but a fallback and always
prefers the more specific variants if they exist and are non-empty.
Conceptually, we should be ok with even just one search dirs variant.
Haskell libraries are named differently depending on whether they're
static or dynamic, so GHC can conveniently pick the right one depending
on the linking needs. That means we don't really need to play tricks
with search paths to convince the compiler to do linking as we want it.
For system C libraries, the convention has been anyway to place static and
dynamic libs next to each other, so we need to deal with that issue
anyway and it is outside of our control. But this is out of the scope
of this patch.
This patch is backwards compatible with cabal. Cabal should however
be patched to use the new '-fully-static' switch.
-
ad3c808d
by Julian Ospald at 2026-01-09T02:48:53-05:00
Warn when "-dynamic" is mixed with "-staticlib"
-
322dd672
by Matthew Pickering at 2026-01-09T02:49:35-05:00
rts: Use INFO_TABLE_CONSTR for stg_dummy_ret_closure
Since the closure type is CONSTR_NOCAF, we need to use INFO_TABLE_CONSTR
to populate the constructor description field (this crashes ghc-debug
when decoding AP_STACK frames sometimes)
Fixes #26745
-
039bac4c
by Ben Gamari at 2026-01-09T20:22:16-05:00
ghc-internal: Move STM utilities out of GHC.Internal.Conc.Sync
This is necessary to avoid an import cycle on Windows when importing
`GHC.Internal.Exception.Context` in `GHC.Internal.Conc.Sync`.
On the road to address #25365.
-
8c389e8c
by Ben Gamari at 2026-01-09T20:22:16-05:00
base: Capture backtrace from throwSTM
Implements core-libraries-committee#297.
Fixes #25365.
-
e1ce1fc3
by Ben Gamari at 2026-01-09T20:22:16-05:00
base: Annotate rethrown exceptions in catchSTM with WhileHandling
Implements core-libraries-committee#298
-
c4ebdbdf
by Cheng Shao at 2026-01-09T20:23:06-05:00
compiler: make getPrim eagerly evaluate its result
This commit makes `GHC.Utils.Binary.getPrim` eagerly evaluate its
result, to avoid accidental laziness when future patches build other
binary parsers using `getPrim`.
-
66a0c4f7
by Cheng Shao at 2026-01-09T20:23:06-05:00
compiler: implement fast get/put for Word16/Word32/Word64
Previously, `GHC.Utils.Binary` contains `get`/`put` functions for
`Word16`/`Word32`/`Word64` which always loads and stores them as
big-endian words at a potentially unaligned address. The previous
implementation is based on loads/stores of individual bytes and
concatenating bytes with bitwise operations, which currently cannot be
fused to a single load/store operation by GHC.
This patch implements fast `get`/`put` functions for
`Word16`/`Word32`/`Word64` based on a single memory load/store, with
an additional `byteSwap` operation on little-endian hosts. It is based
on unaligned load/store primops added since GHC 9.10, and we already
require booting with at least 9.10, so it's about time to switch to
this faster path.
-
641ec3f0
by Simon Peyton Jones at 2026-01-09T20:23:55-05:00
Fix scoping errors in specialisation
Using -fspecialise-aggressively in #26682 showed up a couple of
subtle errors in the type-class specialiser.
* dumpBindUDs failed to call `deleteCallsMentioning`, resulting in a
call that mentioned a dictionary that was not in scope. This call
has been missing since 2009!
commit c43c981705ec33da92a9ce91eb90f2ecf00be9fe
Author: Simon Peyton Jones <simonpj@microsoft.com>
Date: Fri Oct 23 16:15:51 2009 +0000
Fixed by re-combining `dumpBindUDs` and `dumpUDs`.
* I think there was another bug involving the quantified type
variables in polymorphic specialisation. In any case I refactored
`specHeader` and `spec_call` so that the former looks for the
extra quantified type variables rather than the latter. This
is quite a worthwhile simplification: less code, easier to grok.
Test case in simplCore/should_compile/T26682,
brilliantly minimised by @sheaf.
-
2433e91d
by Cheng Shao at 2026-01-09T20:24:43-05:00
compiler: change sectionProtection to take SectionType argument
This commit changes `sectionProtection` to only take `SectionType`
argument instead of whole `Section`, since it doesn't need the Cmm
section content anyway, and it can then be called in parts of NCG
where we only have a `SectionType` in scope.
-
e5926fbe
by Cheng Shao at 2026-01-09T20:24:43-05:00
compiler: change isInitOrFiniSection to take SectionType argument
This commit changes `isInitOrFiniSection` to only take `SectionType`
argument instead of whole `Section`, since it doesn't need the Cmm
section content anyway, and it can then be called in parts of NCG
where we only have a `SectionType` in scope. Also marks it as
exported.
-
244d57d7
by Cheng Shao at 2026-01-09T20:24:43-05:00
compiler: fix split sections on windows
This patch fixes split sections on windows by emitting the right
COMDAT section header in NCG, see added comment for more explanation.
Fix #26696 #26494.
-------------------------
Metric Decrease:
LargeRecord
T9675
size_hello_artifact
size_hello_artifact_gzip
size_hello_unicode
size_hello_unicode_gzip
Metric Increase:
T13035
-------------------------
Co-authored-by: Codex <codex@openai.com>
-
182f3d0f
by Cheng Shao at 2026-01-09T20:25:28-05:00
iserv: add comment about -fkeep-cafs
-
49675b69
by Matthew Craven at 2026-01-09T20:26:14-05:00
Account for "stupid theta" in demand sig for DataCon wrappers
Fixes #26748.
-
f3c18890
by Samuel Thibault at 2026-01-10T15:48:22+01:00
hurd: Fix getExecutablePath build
3939a8bf93e27 ("GNU/Hurd: Add getExecutablePath support") added using
/proc/self/exe for GNU/Hurd but missed adding the required imports for
the corresponding code.
-
7f15bd15
by Samuel Thibault at 2026-01-12T07:16:25-05:00
Fix the OS string encoding for GNU/Hurd
Following https://github.com/haskell/cabal/pull/9434/files , and as seen
in the various gnu_HOST_OS usages in the source code, it is expected that
GNU/Hurd is advertised as "gnu", like the autotools do.
-
1db2f240
by Andrew Lelechenko at 2026-01-12T07:17:06-05:00
Add since annotation for Data.Bifoldable1
Fixes #26432
-
e038a383
by Sven Tennie at 2026-01-12T07:17:49-05:00
Ignore Windows CI tool directories in Git
Otherwise, we see thousands of changes in `git status` which is very
confusing to work with.
-
023c301c
by sheaf at 2026-01-13T04:57:30-05:00
Don't re-use stack slots for growing registers
This commit avoids re-using a stack slot for a register that has grown
but already had a stack slot.
For example, suppose we have stack slot assigments
%v1 :: FF64 |-> StackSlot 0
%v2 :: FF64 |-> StackSlot 1
Later, we start using %v1 at a larger format (e.g. F64x2) and we need
to spill it again. Then we **must not** use StackSlot 0, as a spill
at format F64x2 would clobber the data in StackSlot 1.
This can cause some fragmentation of the `StackMap`, but that's probably
OK.
Fixes #26668
-
d0966e64
by fendor at 2026-01-13T04:58:11-05:00
Remove `traceId` from ghc-pkg executable
-
20d7efec
by Simon Peyton Jones at 2026-01-13T12:41:22-05:00
Make SpecContr rules fire a bit later
See #26615 and Note [SpecConstr rule activation]
-
8bc4eb8c
by Andrew Lelechenko at 2026-01-13T12:42:03-05:00
Upgrade mtl submodule to 2.3.2
Fixes #26656
-
c94aaacd
by Cheng Shao at 2026-01-13T12:42:44-05:00
compiler: remove iserv and only use on-demand external interpreter
This patch removes `iserv` from the tree completely. Hadrian would no
longer build or distribute `iserv`, and the GHC driver would use the
on-demand external interpreter by default when invoked with
`-fexternal-interpreter`, without needing to specify `-pgmi ""`. This
has multiple benefits:
- It allows cleanup of a lot of legacy hacks in the hadrian codebase.
- It paves the way for running cross ghc's iserv via cross emulator
(#25523), fixing TH/ghci support for cross targets other than
wasm/js.
-
c1fe0097
by Peter Trommler at 2026-01-14T03:54:49-05:00
PPC NCG: Fix shift right MO code
The shift amount in shift right [arithmetic] MOs is machine word
width. Therefore remove unnecessary zero- or sign-extending of
shift amount.
It looks harmless to extend the shift amount argument because the
shift right instruction uses only the seven lowest bits (i. e. mod 128).
But now we have a conversion operation from a smaller type to word width
around a memory load at word width. The types are not matching up but
there is no check done in CodeGen. The necessary conversion from word
width down to the smaller width would be translated into a no-op on
PowerPC anyway. So all seems harmless if it was not for a small
optimisation in getRegister'.
In getRegister' a load instruction with the smaller width of the
conversion operation was generated. This loaded the most significant
bits of the word in memory on a big-endian platform. These bits were
zero and hence shift right was used with shift amount zero and not one
as required in test Sized.
Fixes #26519
-
2dafc65a
by Cheng Shao at 2026-01-14T03:55:31-05:00
Tree-wide cleanup of cygwin logic
GHC has not supported cygwin for quite a few years already, and will
not resume support in the forseeable future. The only supported
windows toolchain is clang64/clangarm64 of the msys2 project. This
patch cleans up the unused cygwin logic in the tree.
Co-authored-by: Codex <codex@openai.com>
-
66b96e2a
by Teo Camarasu at 2026-01-14T03:56:13-05:00
Set default eventlog-flush-interval to 5s
Resolves #26707
-
d0254579
by Andrew Lelechenko at 2026-01-14T03:56:53-05:00
Document when -maxN RTS option was added
-
f25e2b12
by Cheng Shao at 2026-01-14T11:10:39-05:00
testsuite: remove obsolete --ci option from the testsuite driver
This patch removes the obsolete `--ci` option from the testsuite
driver: neither the CI scripts nor hadrian ever invokes the testsuite
driver with `--ci`, and the perf notes are always fetched to the
`refs/notes/perf` local reference anyway.
-
7964763b
by Julian Ospald at 2026-01-14T11:11:31-05:00
Fix fetch_cabal
* download cabal if the existing one is of an older version
* fix FreeBSD download url
* fix unpacking on FreeBSD
-
6b0129c1
by Julian Ospald at 2026-01-14T11:11:31-05:00
Bump toolchain in CI
-
0f53ccc6
by Julian Ospald at 2026-01-14T11:11:31-05:00
Use libffi-clib
Previously, we would build libffi via hadrian
and bundle it manually with the GHC bindist.
This now moves all that logic out of hadrian
and allows us to have a clean Haskell package
to build and link against and ship it without
extra logic.
This patch still retains the ability to link
against a system libffi.
The main reason of bundling libffi was that on
some platforms (e.g. FreeBSD and Mac), system libffi
is not visible to the C toolchain by default,
so users would require settings in e.g. cabal
to be able to compile anything.
This adds the submodule libffi-clib to the repository.
-
5e1cd595
by Peng Fan at 2026-01-14T11:12:26-05:00
NCG/LA64: add support for la664 micro architecture
Add '-mla664' flag to LA664, which has some new features:
atomic instructions, dbar hints, etc.
'LA464' is the default so that unrecognized instructions are not
generated.
-
c56567ec
by Simon Peyton Jones at 2026-01-15T23:19:04+00:00
Add evals for strict data-con args in worker-functions
This fixes #26722, by adding an eval in a worker for
arguments of strict data constructors, even if the
function body uses them strictly.
See (WIS1) in Note [Which Ids should be strictified]
I took the opportunity to make substantial improvements in the
documentation for call-by-value functions. See especially
Note [CBV Function Ids: overview] in GHC.Types.Id.Info
Note [Which Ids should be CBV candidates?] ditto
Note [EPT enforcement] in GHC.Stg.EnforceEpt
among others.
-
9719ce5d
by Simon Peyton Jones at 2026-01-15T23:19:04+00:00
Improve `interestingArg`
This function analyses a function's argument to see if it is
interesting enough to deserve an inlining discount. Improvements
for
* LitRubbish arguments
* exprIsExpandable arguments
See Note [Interesting arguments] which is substantially rewritten.
-
7b616b9f
by Cheng Shao at 2026-01-16T06:45:00-05:00
compiler: fix regression when compiling foreign stubs in the rts unit
This patch fixes a regression when compiling foreign stubs in the rts
unit introduced in 05e25647f72bc102061af3f20478aa72bff6ff6e. A simple
revert would fix it, but it's better to implement a proper fix with
comment for better understanding of the underlying problem, see the
added comment for explanation.
Co-authored-by: Codex <codex@openai.com>
-
c343ef64
by Sylvain Henry at 2026-01-16T06:45:51-05:00
base: remove GHC.JS.Prim.Internal.Build (#23432)
See accepted CLC proposal https://github.com/haskell/core-libraries-committee/issues/329
-
29c0aceb
by Simon Peyton Jones at 2026-01-16T17:18:11-05:00
Improve newtype unwrapping
Ticket #26746 describes several relatively-minor shortcomings of newtype
unwrapping. This MR addresses them, while also (arguably) simplifying
the code a bit.
See new Note [Solving newtype equalities: overview]
and Note [Decomposing newtype equalities]
and Note [Eager newtype decomposition]
and Note [Even more eager newtype decomposition]
For some reason, on Windows only, runtime allocations decrease for test
T5205 (from 52k to 48k). I have not idea why. No change at all on Linux.
I'm just going to accept the change. (I saw this same effect in another
MR so I think it's a fault in the baseline.)
Metric Decrease:
T5205
-
8b59e62c
by Andreas Klebinger at 2026-01-16T17:18:52-05:00
testsuite: Widen acceptance window for T5205.
Fixes #26782
-
9e5e0234
by mangoiv at 2026-01-17T06:03:03-05:00
add a new issue template for getting verified
To reduce spam created by new users, we will in future not grant
any rights but reporting issues to new users. That is why we will
have to be able to verify them. The added issue template serves that
purpose.
-
b18b2c42
by Cheng Shao at 2026-01-17T06:03:44-05:00
llvm: fix split sections for llvm backend
This patch fixes split sections for llvm backend:
- Pass missing `--data-sections`/`--function-sections` flags to
llc/opt.
- Use `@llvm.compiler.used` instead of `@llvm.used` to avoid sections
being unnecessarily retained at link-time.
Fixes #26770.
-------------------------
Metric Decrease:
libdir
size_hello_artifact
size_hello_unicode
-------------------------
Co-authored-by: Codex <codex@openai.com>
-
ebf66f67
by Cheng Shao at 2026-01-17T13:16:50-05:00
Update autoconf scripts
Scripts taken from autoconf a2287c3041a3f2a204eb942e09c015eab00dc7dd
-
598624b9
by Andreas Klebinger at 2026-01-17T13:17:32-05:00
CString.hs: Update incorrect comment.
Fixes #26322
-
eea2036b
by Cheng Shao at 2026-01-18T10:00:49-05:00
libraries: bump haskeline submodule to 0.8.4.1
This patch bumps the haskeline submodule to 0.8.4.1 which includes an
important fix for an ANSI handling bug on Windows
(https://github.com/haskell/haskeline/pull/126).
-
87d8f6c2
by Cheng Shao at 2026-01-18T10:01:30-05:00
hadrian: replace default -H32m/-H64m with -O64M to improve mutator productivity
Most hadrian build flavours pass `-H32m`/`-H64m` to GHC as
conventional wisdom to improve mutator productivity and reduce GC
overhead. They were inherited from the legacy Make build system, and
there used to be make flags to instrument a build process with
`-Rghc-timing` option to collect GC stats of each GHC run from stderr.
It's time to revisit whether there are better defaults for
`-H32m`/`-H64m`, and this patch changes it to `-O64M` which indeed
improves mutator productivity based on real statistics. `-O64M` is
more aggressive than `-H64m`; it allows the old generation to grow to
at least 64M before triggering major GC and reduces major GC runs.
The stats of a clean build with `validate` flavour and `-H64m`:
```
h64m.log
matched RTS stat lines: 5499
sum MUT cpu : 2400.808 s
sum GC cpu : 1378.292 s
sum MUT elapsed : 2788.253 s
sum GC elapsed : 1389.233 s
GC/MUT cpu ratio : 0.574 (GC is 57.4% of MUT)
GC/MUT elapsed ratio : 0.498 (GC is 49.8% of MUT)
GC fraction of (MUT+GC) cpu : 36.5%
GC fraction of (MUT+GC) elapsed : 33.3%
per-line GC/MUT cpu ratio: median 0.691, p90 1.777
per-line GC/MUT elapsed ratio: median 0.519, p90 1.081
```
The stats of a clean build with `validate` flavour and `-O64M`:
```
o64m.log
matched RTS stat lines: 5499
sum MUT cpu : 2377.383 s
sum GC cpu : 1127.146 s
sum MUT elapsed : 2758.857 s
sum GC elapsed : 1135.587 s
GC/MUT cpu ratio : 0.474 (GC is 47.4% of MUT)
GC/MUT elapsed ratio : 0.412 (GC is 41.2% of MUT)
GC fraction of (MUT+GC) cpu : 32.2%
GC fraction of (MUT+GC) elapsed : 29.2%
per-line GC/MUT cpu ratio: median 0.489, p90 1.099
per-line GC/MUT elapsed ratio: median 0.367, p90 0.806
```
Mutator time is roughly in the same ballpark, but GC CPU time has
reduced by 18.22%, and mutator productivity has increased from 63.5%
to 67.8%.
-
8372e13d
by Cheng Shao at 2026-01-18T10:02:12-05:00
rts: remove unused .def files from rts/win32
This patch removes unused .def files from `rts/win32`, given we don't
build .dll files for rts/ghc-internal/ghc-prim at all. Even when we
resurrect win32 dll support at some point in the future, these .def
files still contain incorrect symbols anyway and won't be of any use.
-
f6af485d
by Cheng Shao at 2026-01-18T10:03:19-05:00
.gitmodules: use gitlab mirror for the libffi-clib submodule
This patch fixes .gitmodules to use the gitlab mirror for the
libffi-clib submodule, to make it coherent with other submodules that
allow ghc developers to experiment with wip branches in submodules for
ghc patches. Fixes #26783.
-
41432d25
by Cheng Shao at 2026-01-18T10:05:13-05:00
hadrian: remove the horrible i386 speedHack
When hadrian builds certain rts objects for i386, there's a horrible
speedHack that forces -fno-PIC even for dynamic ways of those objects.
This is not compatible with newer versions of gcc/binutils as well as
clang/lld, and this patch removes it. Fixes #26792.
-
323eb8f0
by Cheng Shao at 2026-01-18T21:48:19-05:00
hadrian: enable split sections for cross stage0
This patch fixes a minor issue with `splitSectionsArgs` in hadrian:
previously, it's unconditionally disabled for stage0 libraries because
it's not going to be shipped in the final bindists. But it's only true
when not cross compiling. So for now we also need to enable it for
cross stage0 as well.
-
3fadfefe
by Andreas Klebinger at 2026-01-18T21:49:01-05:00
RTS: Document -K behaviour better
-
6b8c5058
by Cheng Shao at 2026-01-19T17:44:59+00:00
ci: update darwin boot ghc to 9.10.3
This patch updates darwin boot ghc to 9.10.3, along with other related
updates, and pays off some technical debt here:
- Update `nixpkgs` and use the `nixpkgs-25.05-darwin` channel.
- Update LLVM to 21 and update `llvm-targets` to reflect LLVM 21
layout changes for arm64/x86_64 darwin targets.
- Use `stdenvNoCC` to prevent nix packaged apple sdk from being used
by boot ghc, and manually set `DEVELOPER_DIR`/`SDKROOT` to enforce
the usage of system-wide command line sdk for macos.
- When building nix derivation for boot ghc, run `configure` via the
`arch` command so that `configure` and its subprocesses pick up the
manually specified architecture.
- Remove the previous horrible hack that obliterates `configure` to
make every autoconf test result in true. `configure` now properly
does its job.
- Remove the now obsolete configure args and post install settings
file patching logic.
- Use `scheme-small` for texlive to avoid build failures in certain
unused texlive packages, especially on x86_64-darwin.