This removes all support for GC mode.
This was deprecated by Apple in OS X 10.8 (2012) and never supported in
iOS. Very little code ever tried to use it with this runtime.
It's gone from the build system already.
This replaces a few home-grown datastructures with third-party ones that
get a lot more testing:
- The home-grown hopscotch hash table is moved to using robin map. The
original was designed to be lock free, but we've been using it behind
a lock for ages.
- The selector list is now a std::vector.
- The types list now use std::forward_list.
This also removes a couple of code paths that haven't been used since we
started using the new ABI data structures internally and upgrading at
load time.
The new code tries to differentiate in the static type system between
registered and unregistered selectors. The check for whether a selector
is registered is fragile and depends on no selector being mapped or
allocated in memory below the total number of selectors. This check can
now disappear on most code paths.
On a single test machine (not guaranteed to be representative) the test
suite now completes around 20% faster.
This adds a dependency on a version of CMake with support for Objective-C,
removing a number of hacks to make CMake work and introduces a few
modernisations:
- Options use CMake's `option` mechanism.
- A lot of things now use generator expressions, which should improve the
things like the multi-config Ninja generator.
- We no longer depend on hacks to build in CI.
At the same time, a few other things are simplified:
- There is no longer support for building without C++ support and we depend on
a C++ standard library, not just a runtime. This makes it possible to use
more C++ features inside the runtime.
- We no longer support clang-cl on Windows and instead require clang.exe.
- Support for the Boehm GC has been removed.
The instance size was being reported incorrectly which meant that
associated objects looked for a reference list slightly after the end of
the object.
Fixes#228
CI was failing because a number of our host platforms were no longer
supported. This commit:
- Moves to the latest FreeBSD 12, 13, and 14 branch releases.
- Moves to modern Ubuntu.
- Replaces Windows 2016 with 2022
- Moves from Azure Pipelines to GitHub for the non-FreeBSD tests.
- Fixes some tests that are not reliable in some contexts.
Reorders how locking is handled in `objc_send_initialize()` to prevent a deadlock.
Previously, contention on the low level spinlocks could cause a very intermittent deadlock:
- Thread A : `objc_send_initialize()` holds the runtime lock, then tries to acquire the object lock on the metaclass, which needs to initialize the mutex for the new metaclass inside `referenceListForObject()`, so it tries to lock the `lock_for_pointer()` / `lock_spinlock()`
- Thread B : `referenceListForObject()` holds a spinlock for an unrelated object while running `initHiddenClassForObject()` -> `allocateHiddenClass()`, which tries to acquire the runtime lock
If the metaclass object pointer in Thread A hashes to the same spinlock as the object in thread B, the runtime lock ends up deadlocked forever.
This test used to fail on Windows, because Clang's codegen and transformation haven't been in-line with the requirements of the WinEH backend in LLVM. The following changes implement the missing pieces and let the test to pass in both debug and release mode:
https://reviews.llvm.org/D128190 [WinEH] Apply funclet operand bundles to nounwind intrinsics that lower to function calls in the course of IR transforms
https://reviews.llvm.org/D134441 [ObjC][ARC] Fix target register for call expanded from CALL_RVMARKER on Windows
https://reviews.llvm.org/D137939 [CGObjC] Open cleanup scope before SaveAndRestore CurrentFuncletPad and push CatchRetScope early
https://reviews.llvm.org/D137944 [ObjC][ARC] Teach the OptimizeSequences step of ObjCARCOpts about WinEH funclet tokens
Fixes building for Android arm64-v8a with NDK r25, which no longer supports this flag. Since we were passing this unconditionally we were previously inadvertently disabling NEON support if the flag wasn't set.
This originally came up as an issue with libc++abi support (#152), but is not specific to that ABI.
* Use the C++ runtime to check for uncaught C++ exceptions.
As discussed in #152, use the function defined in the Itanium C++ ABI to
check whether the thrown exception is the current caught C++ exception
and needs rethrowing via `__cxa_rethrow()`.
Co-authored-by: Niels Grewe <grewe@ocean-insights.com>
Co-authored-by: David Chisnall <gnustep@theravensnest.org>
* Add test for emptying autorelease pool
* Fix arc autorelease pool emptying when adding further references
When releasing a reference in the arc autorelease pool, it is
possible and anticipated that new references may added to the pool.
This fix addresses an edge case where releasing a reference in the
same pool page as the stop position can add more references which
cause the insertion of a new page and emptyPool() returns early
after emptying the new page.
This fixes a regression in 73132a6 (#200) where nil was returned
from a retain call after the object begins deallocating. Normal
retains of a deallocating object are still expected to return an
instance pointer inside its dealloc method and code compiled with ARC
will generate calls to objc_storeStrong() when a block captures the
self pointer inside the dealloc method.
The previous checks for a deallocating object with negative reference count did not work because:
1. the sign bit has to be recreated as was happening in objc_delete_weak_refs()
2. there was no distinction between a saturated count and a negative reference count
refcount_max now indicates when the refcount has saturated and should no longer be mutated.
An underflow to -1 still maps to refcount_mask, allowing us to detect when an object is supposed to be deallocated.
Neither objc_release_fast_no_destroy_np() nor objc_retain_fast_np() mutate the refcount when it is one of those values, so the comment in objc_delete_weak_refs() was adjusted.
Add 32-bit x86-Linux builds.
Add building with the nightly LLVM releases. There are some regressions
since LLVM 8 that would have been caught by this that we can hopefully
fix in time for the LLVM 12 release.
When we stopped using CXXFLAGS in eh_personality.s generation, we lost
the ability to propagate -m32 or -target through, so cross-compile
didn't work. These flags are now passed through correctly, so (so
'-pipe -O2' will appear as two arguments, rather than being combined as
a single value that then causes clang to abort with an unknown argument
error, which was the original motivation for this change).
This is the correct fix for #177
The way that we were doing this didn't handle multiple flags and we
actually don't want the user to override these flags because that file
needs to be compiled in a very specific way.
Fixes#177
Several of the hooks were being generated with common linkage in hooks.c
and as normal initialised symbols in sendmsg2.c. Recent clang defaults
to -fno-common, which causes the common linkage versions to become
strong symbols and then causes a linkage failure from duplicate
definitions: one zero initialised, one initialised with a real value.
This removes copy in hooks.c, retaining only the initialised version.
Fixes#171
- Don't try to detect libc++abi. It doesn't actually work, so don't
try it.
- Correctly try to link the standard library.
This should fix the build on GNU/Linux systems with libc++abi.so
installed (including our Ubuntu CI!).
Frederik Seiffert
David Chisnall
Earl Robsham
Stefan Gränitz
David Chisnall
Graham--M
Alex Myczko
mischievous
Bernard Cafarelli
Graham Lee
David Chisnall
Niels Grewe
Riccardo Mottola
David Chisnall