In D144319, Clang tried to land a change that would cause some functions
that are not supposed to return nullptr to optimize better. As reported
in https://reviews.llvm.org/D144319#4203982, libc++ started seeing
failures in its CI shortly after this change was landed.
As explained in D146379, the reason for these failures is that libc++'s
throwing `operator new` can in fact return nullptr when compiled with
exceptions disabled. However, this contradicts the Standard, which
clearly says that the throwing version of `operator new(size_t)` should
never return nullptr. This is actually a long standing issue. I've
previously seen a case where LTO would optimize incorrectly based on the
assumption that `operator new` doesn't return nullptr, an assumption
that was violated in that case because libc++.dylib was compiled with
-fno-exceptions.
Unfortunately, fixing this is kind of tricky. The Standard has a few
requirements for the allocation functions, some of which are impossible
to satisfy under -fno-exceptions:
1. `operator new(size_t)` must never return nullptr
2. `operator new(size_t, nothrow_t)` must call the throwing version and
return nullptr on failure to allocate
3. We can't throw exceptions when compiled with -fno-exceptions
In the case where exceptions are enabled, things work nicely.
`new(size_t)` throws and `new(size_t, nothrow_t)` uses a try-catch to
return nullptr. However, when compiling the library with
-fno-exceptions, we can't throw an exception from `new(size_t)`, and we
can't catch anything from `new(size_t, nothrow_t)`. The only thing we
can do from `new(size_t)` is actually abort the program, which does not
make it possible for `new(size_t, nothrow_t)` to catch something and
return nullptr.
This patch makes the following changes:
1. When compiled with -fno-exceptions, the throwing version of `operator
new` will now abort on failure instead of returning nullptr on failure.
This resolves the issue that the compiler could mis-compile based on the
assumption that nullptr is never returned. This constitutes an API and
ABI breaking change for folks compiling the library with -fno-exceptions
(which is not the general public, who merely uses libc++ headers but use
a shared library that has already been compiled). This should mostly
impact vendors and other folks who compile libc++.dylib themselves.
2. When the library is compiled with -fexceptions, the nothrow version
of `operator new` has no change. When the library is compiled with
-fno-exceptions, the nothrow version of `operator new` will now check
whether the throwing version of `operator new` has been overridden. If
it has not been overridden, then it will use an implementation
equivalent to that of the throwing `operator new`, except it will return
nullptr on failure to allocate (instead of terminating). However, if the
throwing `operator new` has been overridden, it is now an error NOT to
also override the nothrow `operator new`. Indeed, there is no way for us
to implement a valid nothrow `operator new` without knowing the exact
implementation of the throwing version.
In summary, this change will impact people who fall into the following
intersection of conditions:
- They use the libc++ shared/static library built with `-fno-exceptions`
- They do not override `operator new(..., std::nothrow_t)`
- They override `operator new(...)` (the throwing version)
- They use `operator new(..., std::nothrow_t)`
We believe this represents a small number of people.
Fixes#60129
rdar://103958777
Differential Revision: https://reviews.llvm.org/D150610
This patch removes the noexcept specifier introduced in #69407 since the
Standard allows a new handler to throw an exception of type bad_alloc
(or derived from it). With the noexcept specifier on the helper
functions, we would immediately terminate the program.
The patch also adds tests for the case that had regressed.
Co-authored-by: Alison Zhang <alisonzhang@ibm.com>
This patch implements __cxa_init_primary_exception, an extension to the
Itanium C++ ABI. This extension is already present in both libsupc++ and
libcxxrt. This patch also starts making use of this function in
std::make_exception_ptr: instead of going through a full throw/catch
cycle, we are now able to initialize an exception directly, thus making
std::make_exception_ptr around 30x faster.
This doesn't actually prevent the build from failing, but it provides a
better diagnostic that explains what is needed in order to fix the build.
Fixes#77846Fixes#77843
I recently came across LIBCXXABI_USE_LLVM_UNWINDER and was surprised to
notice it was disabled by default. Since we build libunwind by default
and ship it in the LLVM toolchain, it would seem to make sense that
libc++ and libc++abi rely on libunwind for unwinding instead of using
the system-provided unwinding library (if any).
Most importantly, using the system unwinder implies that libc++abi is
ABI compatible with that system unwinder, which is not necessarily the
case. Hence, it makes a lot more sense to instead default to using the
known-to-be-compatible LLVM unwinder, and let vendors manually select a
different unwinder if desired.
As a follow-up change, we should probably apply the same default to
compiler-rt.
Differential Revision: https://reviews.llvm.org/D150897Fixes#77662
rdar://120801778
This addresses cases (currently failing) where we throw a null
pointer-to-object and fixes#64953.
We are trying to satisfy the following bullet from the C++ ABI 15.3:
* the handler is of type cv1 T* cv2 and E is a pointer type that can be
converted to the type of the handler by either or both of:
- a standard pointer conversion (4.10 [conv.ptr]) not involving
conversions to private or protected or ambiguous classes.
- a qualification conversion.
The existing implementation assesses the ambiguity of bases by computing
the offsets to them; ambiguous cases are then when the same base appears
at different offsets. The computation of offset includes indirecting
through the vtables to find the offsets to virtual bases.
When the thrown pointer points to a real object, this is quite efficient
since, if the base is found, and it is not ambiguous and on a public
path, the offset is needed to return the adjusted pointer (and the
indirections are not particularly expensive to compute).
However, when we throw a null pointer-to-object, this scheme is no
longer applicable (and the code currently bypasses the relevant
computations, leading to the incorrect catches reported in the issue).
-----
The solution proposed here takes a composite approach:
1. When the pointer-to-object points to a real instance (well, at least,
it is determined to be non-null), we use the existing scheme.
2. When the pointer-to-object is null:
* We note that there is no real object.
* When we are processing non-virtual bases, we continue to compute the
offsets, but for a notional dummy object based at 0. This is OK, since
we never need to access the object content for non-virtual bases.
* When we are processing a path with one or more virtual bases, we
remember a cookie corresponding to the inner-most virtual base found so
far (and set the notional offset to 0). Offsets to inner non-virtual
bases are then computed as normal.
A base is then ambiguous iff:
* There is a recorded virtual base cookie and that is different from the
current one or,
* The non-virtual base offsets differ.
When a handler for a pointer succeeds in catching a base pointer for a
thrown null pointer-to-object, we still return a nullptr (so the
adjustment to the pointer is not required and need not be computed).
Since we noted that there was no object when starting the search for
ambiguous bases, we know that we can skip the pointer adjustment.
This was originally uploaded as https://reviews.llvm.org/D158769.
Fixes#64953
Added -p / --no-params flag to skip demangling function parameters
similar to how it is supported by GNU c++filt tool.
There are cases when users want to demangle a large number of symbols in
bulk, for example, at startup, and do not care about function parameters
and overloads at that time. Skipping the demangling of parameter types
led to a measurable improvement in performance. Our users reported about
15% speed up with GNU c++filt and we expect similar results with
llvm-cxxfilt with this patch.
Added -p / --no-params flag to skip demangling function parameters
similar to how it is supported by GNU c++filt tool.
There are cases when users want to demangle a large number of symbols in
bulk, for example, at startup, and do not care about function parameters
and overloads at that time. Skipping the demangling of parameter types
led to a measurable improvement in performance. Our users reported about
15% speed up with GNU c++filt and we expect similar results with
llvm-cxxfilt with this patch.
When we use the -nostdlib++ flag, we don't need to explicitly link
against compiler-rt, since the compiler already links against it by
default. This simplifies the flags that we need to use when building
with Clang and GCC, and opens the door to further simplifications since
most platforms won't need to detect whether libgcc and libgcc_s are
supported anymore.
Furthermore, on platforms where -nostdlib++ is used, this patch prevents
manually linking compiler-rt *before* other system libraries. For
example, Apple platforms have several compiler-rt symbols defined in
libSystem.dylib. If we manually link against compiler-rt, we end up
overriding the default link order preferred by the compiler and
potentially using the symbols from the clang-provided libclang_rt.a
library instead of the system provided one.
Note that we don't touch how libunwind links against compiler-rt when it
builds the .so/.a because libunwind currently doesn't use -nodefaultlibs
and we want to avoid rocking the boat too much.
rdar://119506163
When lpStartEncoding is different from DW_EH_PE_omit, lpStart can be set
to zero which is a valid base address for landing pads. Such base value
is useful when landing pads are placed in different sections.
Fixes#72582.
C++23 removed `<ciso646>` from the standard library. The header is used
in a few places in order to pull in implementation-specific and feature
test macros. The new way of doing that is `<version>`, which should be
supported by all supported implementations. This change replaces all
those uses of `<ciso646>` with `<version>`.
This patch actually runs the tests for picolibc behind an emulator,
removing a few workarounds and increasing coverage.
Differential Revision: https://reviews.llvm.org/D155521
Picolibc is a C Standard Library that is commonly used in embedded
environments. This patch adds initial support for this configuration
along with pre-commit CI. As of this patch, the test suite only builds
the tests and nothing is run. A follow-up patch will make the test suite
actually run the tests.
Differential Revision: https://reviews.llvm.org/D154246
The mangling for an explicitly named object was introduced in
https://reviews.llvm.org/D140828
See following discussion for why a new mangling had to be introduced:
https://github.com/itanium-cxx-abi/cxx-abi/issues/148
Since clang started emitting names with the new mangling, this patch
implements support for demangling such names.
The approach this patch takes is to add a new `ExplicitObjectParameter`
node that will print the first parameter of a function declaration with
a `this ` prefix, to reflect what was spelled out in source.
Example:
```
void MyClass::func(this MyClass const& self); // _ZNH7MyClass4funcERKS_
```
With this patch, the above demangles to:
```
_ZNH7MyClass4funcERKS_ -> MyClass::func(this MyClass const&)
```
Note that `func` is not marked as `const &`, since the
function-qualifiers are now encoded as part of the explicit `this`. C++
doesn't allow specifying the function-qualifiers in the presence of an
explicit object parameter, so this demangling is consistent with the
source spelling.
Adding test-cases to the `cases` array causes `git clang-format` to
split the strings of many of the existing test-cases, making them harder
to read/work with in most cases.
This patch disables `clang-format` for the `cases` array so it doesn't
catch anyone off-guard in the future.
The runtimes now have a principled way of doing assertions in relation
to hardening, so we should use that instead of raw calls to assert()
inside libc++abi. This patch aims to maintain the behavior of the
demangler code when it is used from within LLVM by introducing a simple
DEMANGLE_ASSERT(...) macro that is then defined to the appropriate
assertion mechanism.
`LIBCXXABI_SYSROOT`, `LIBCXXABI_TARGET_TRIPLE` and
`LIBCXXABI_GCC_TOOLCHAIN` are not supported anymore. Based on the
comment, the warning should be removed after branching for LLVM 15.
This is necessary in order to implement some papers like P2467R1, which
require using C++23 declarations in the dylib. It is a good habit to
keep building the dylib with a recent standard version regardless.
With this patch, we also stop strictly enforcing that the targets are
built with C++23. Concretely, C++23 will soon be required in order to
build the dylib, but not enforcing it strictly works around some issues
like the documentation bots using an old and unsupported compiler. Since
these bots do not actually build the library, not strictly enforcing the
C++ Standard makes our CMake build more resilient to these kinds of
situation. This is just a workaround though, the better way of going
about would be to update the compiler on the documentation bot but we
don't seem to have control over that.
Android's librt and libpthread functionality is part of libc.{a,so}
instead. The atomic APIs are part of the compiler-rt builtins archive.
Android does have libdl.
Android's libc.so has `__cxa_thread_atexit_impl` starting in API 23, and
the oldest supported API is 21, so continue using feature detection for
that API.
These settings need to be declared explicitly for the sake of the fuzzer
library's custom libc++ build `add_custom_libcxx`. That macro builds
libc++ using `-DCMAKE_TRY_COMPILE_TARGET_TYPE=STATIC_LIBRARY`, which
breaks the feature detection.
This will allow for configuring tests according to AIX version.
Reviewed By: daltenty, #libc, Mordante
Differential Revision: https://reviews.llvm.org/D149660
Mark tests as necessary to accommodate Android L (5.0 / API 21) and up.
Add three Android lit features:
- android
- android-device-api=(21,22,23,...)
- LIBCXX-ANDROID-FIXME (for failures that need follow-up work)
Enable an AIX workaround in filesystem_test_helper.h for the broken
chmod on older Android devices.
Mark failing test with XFAIL or UNSUPPORTED:
- Mark modules tests as UNSUPPORTED, matching other configurations.
- Mark a gdb test as UNSUPPORTED.
- XFAIL tests for old devices that lack an API (fmemopen).
- XFAIL various FS tests (because SELinux blocks FIFO and hard linking,
because fchmodat is broken on old devices).
- XFAIL various locale tests (because Bionic has limited locale
support). (Also XFAIL an re.traits test.)
- XFAIL some print.fun tests because the error exception has no system
error string.
- Mark std::{cin,wcin} tests UNSUPPORTED because they hang with
adb_run.py on old devices.
- Mark a few tests UNSUPPORTED because they allocate too much memory.
- notify_one.pass.cpp is flaky on Android.
- XFAIL libc++abi demangler test because of Android's special long
double on x86[-64].
N.B. The `__ANDROID_API__` macro specifies a minimum required API level
at build-time, whereas the android-device-api lit feature is the
detected API level of the device at run-time. The android-device-api
value will be >= `__ANDROID_API__`.
This commit was split out from https://reviews.llvm.org/D139147.
Fixes: https://github.com/llvm/llvm-project/issues/69270
I could probably break this commit into more pieces.
---
This patch adds libc++ support for Android L (Android 5.0+) and up,
tested using the Android team's current compiler, a recent version of
the AOSP sysroot, and the x86[-64] Android Emulator.
CMake and Lit Configuration:
Add runtimes/cmake/android/Arch-${ARCH}.cmake files that configure CMake
to cross-compile to Android without using CMake's built-in NDK support
(which only works with an actual packaged NDK).
Add libcxx/cmake/caches/AndroidNDK.cmake that builds and tests libc++
(and libc++abi) for Android. This file configures libc++ to match what
the NDK distributes, e.g.:
- libc++_shared.so (includes libc++abi objects, there is no
libc++abi.so). libunwind is linked statically but not exported.
- libc++_static.a (does not include libc++abi) and libc++abi.a
- `std::__ndk1` namespace
- All the libraries are built with `__ANDROID_API__=21`, even when they
are linked to something targeting a higher API level.
(However, when the Android LLVM team builds these components, they do
not use these CMake cache files. Instead they use Python scripts to
configure the builds. See
https://android.googlesource.com/toolchain/llvm_android/.)
Add llvm-libc++[abi].android-ndk.cfg.in files that test the Android
NDK's libc++_shared.so. These files can target old or new Android
devices. The Android LLVM team uses these test files to test libc++ for
both arm/arm64 and x86/x86_64 architectures.
The Android testing mode works by setting %{executor} to adb_run.py,
which uses `adb push` and `adb shell` to run tests remotely. adb_run.py
always runs tests as the "shell" user even on an old emulator where "adb
unroot" doesn't work. The script has workarounds for old Android
devices. The script uses a Unix domain socket on the host
(--job-limit-socket) to restrict concurrent adb invocations. Compiling
the tests is a major part of libc++ testing run-time, so it's desirable
to exploit all the host cores without overburdening the test devices,
which can have far fewer cores.
BuildKite CI:
Add a builder to run-buildbot, `android-ndk-*`, that uses Android Clang
and an Android sysroot to build libc++, then starts an Android emulator
container to run tests.
Run the emulator and an adb server in a separate Docker container
(libcxx-ci-android-emulator), and create a separate Docker image for
each emulator OS system image. Set ADB_SERVER_SOCKET to connect to the
container's adb server. Running the only adb server inside the container
makes cleanup more reliable between test runs, e.g. the adb client
doesn't create a `~/.android` directory and the adb server can be
restarted along with the emulator using docker stop/run. (N.B. The
emulator insists on connecting to an adb server and will start one
itself if it can't connect to one.)
The suffix to the android-ndk-* job is a label that concisely specifies
an Android SDK emulator image. e.g.:
- "system-images;android-21;default;x86" ==> 21-def-x86
- "system-images;android-33;google_apis;x86_64" ==> 33-goog-x86_64
Fixes: https://github.com/llvm/llvm-project/issues/69270
Differential Revision: https://reviews.llvm.org/D139147
This will make it easier to implement new(nothrow) without calling the
throwing version of new when exceptions are disabled. See
https://llvm.org/D150610 for the full discussion.
Change aade74675c15d3bae5fdfa67f7b4b6ed9dac3d20 updated exception
handling in parts of libc++, but forgot to include <exceptions> in
cxa_demangle.cpp. This commit corrects that issue.
Since 78d649a417b48cb8a2ba2e755f0e7c8fb8b1bb83 the recommended way to
pass an executor is to use the _TEST_PARAMS variable, which means we now
pass more complicated value (including ones that may contain multiple
`=`) as part of this variable. However, the `REGEX REPLACE` being used
has greedy matches so everything up to the last = becomes part of the
variable name which results in invalid syntax in the generated lit
config file.
This was noticed due to builder failures for those using the
CrossWinToARMLinux.cmake cache file.
---------
Co-authored-by: Vladimir Vereschaka <vvereschaka@accesssoftek.com>
This is useful when trying to run multiple tests with different
arguments to the executor script. This is needed in my case since I
do not know the correct ssh connection arguments when building libc++.
The testing script I have spawns multiple QEMU instances that listen on
a given port on localhost and runs lit with the --num-shards/--run-shard
argument. In order to connect each shard to the right QEMU instances I
need to customize the arguments passed to ssh.py (--extra-ssh-args and
--extra-scp-args) but can't do this at configure time since the target
port is only known when running the tests but not when calling CMake.
This change allows me to pass `executor=ssh.py <args>` to lit once I
know
the right hostname/port for running the tests.
This also deprecates the `LIB{CXX,CXXABI,UNWIND}_EXECUTOR` CMake
variable
as the same can be achieved by adding `executor=...` to the
`LIB{CXX,CXXABI,UNWIND}_TEST_PARAMS` variable.
This adds Wasm-specific libc++abi changes to support Wasm exception
handling (https://github.com/WebAssembly/exception-handling).
Wasm EH requires `__USING_WASM_EXCEPTIONS__` to be defined. Wasm EH's
LSDA handling mostly shares that of SjLj EH.
Changes are:
- In Wasm, a destructor returns its argument.
- Wasm EH currently only has one phase (search) that does both search
and cleanup. So added an additional `set_registers` to support that.
The bulk of these changes was added back in Mar 2020 in
https://github.com/emscripten-core/emscripten/pull/10577 to emscripten
repo and has been used ever since. Now we'd like to upstream this so
that other toolchains that don't use emscripten libraries, e.g., WASI,
can use this too.
Companion patch: D158919
Reviewed By: dschuff, #libc_abi, phosek
Differential Revision: https://reviews.llvm.org/D158918
Since GCC now supports -nostdlib++, we can remove some complexity in the
test configurations and do the same as Clang. However, we can't fully
remove the GCC test configuration for libc++ because we apparently need
to explicitly link against libm for some tests to work.
On Apple platforms, we always support the -nostdlib++ flag. Hence, it is
not necessary to manually link against system libraries. In fact, doing
so causes us to link against libSystem explicitly, which messes up with
the order of libraries we should use. Indeed:
Before patch, using the system unwinder (LIBCXXABI_USE_LLVM_UNWINDER = OFF)
===========================================================================
$ otool -L lib/{libc++.1.dylib,libc++abi.1.dylib,libunwind.1.dylib}
lib/libc++.1.dylib:
@rpath/libc++.1.dylib
/usr/lib/libSystem.B.dylib
@rpath/libc++abi.1.dylib
lib/libc++abi.1.dylib:
@rpath/libc++abi.1.dylib
/usr/lib/libSystem.B.dylib
lib/libunwind.1.dylib:
@rpath/libunwind.1.dylib
/usr/lib/libSystem.B.dylib
After patch, using the system unwinder (LIBCXXABI_USE_LLVM_UNWINDER = OFF)
===========================================================================
$ otool -L lib/{libc++.1.dylib,libc++abi.1.dylib,libunwind.1.dylib}
lib/libc++.1.dylib:
@rpath/libc++.1.dylib
@rpath/libc++abi.1.dylib
/usr/lib/libSystem.B.dylib
lib/libc++abi.1.dylib:
@rpath/libc++abi.1.dylib
/usr/lib/libSystem.B.dylib
lib/libunwind.1.dylib:
@rpath/libunwind.1.dylib
/usr/lib/libSystem.B.dylib
Before patch, with the LLVM unwinder (LIBCXXABI_USE_LLVM_UNWINDER = ON)
=======================================================================
$ otool -L lib/{libc++.1.dylib,libc++abi.1.dylib,libunwind.1.dylib}
lib/libc++.1.dylib:
@rpath/libc++.1.dylib
/usr/lib/libSystem.B.dylib
@rpath/libc++abi.1.dylib
@rpath/libunwind.1.dylib
lib/libc++abi.1.dylib:
@rpath/libc++abi.1.dylib
/usr/lib/libSystem.B.dylib
@rpath/libunwind.1.dylib
lib/libunwind.1.dylib:
@rpath/libunwind.1.dylib
/usr/lib/libSystem.B.dylib
After patch, with the LLVM unwinder (LIBCXXABI_USE_LLVM_UNWINDER = ON)
======================================================================
$ otool -L lib/{libc++.1.dylib,libc++abi.1.dylib,libunwind.1.dylib}
lib/libc++.1.dylib:
@rpath/libc++.1.dylib
@rpath/libc++abi.1.dylib
@rpath/libunwind.1.dylib
/usr/lib/libSystem.B.dylib
lib/libc++abi.1.dylib:
@rpath/libc++abi.1.dylib
@rpath/libunwind.1.dylib
/usr/lib/libSystem.B.dylib
lib/libunwind.1.dylib:
@rpath/libunwind.1.dylib
/usr/lib/libSystem.B.dylib
As we can see, libSystem appears before the just-built libraries before
the patch, which causes the libunwind.dylib bundled in libSystem.dylib
to be used instead of the just-built libunwind.dylib.
We didn't notice the issue until recently when I tried to update the
macOS CI builders to macOS 13.5, where it is necessary to use the right
libunwind library (the exact reason still needs to be investigated).
This implements proposals from:
- https://github.com/itanium-cxx-abi/cxx-abi/issues/24: mangling for
constraints, requires-clauses, requires-expressions.
- https://github.com/itanium-cxx-abi/cxx-abi/issues/31: requires-clauses and
template parameters in a lambda expression are mangled into the <lambda-sig>.
- https://github.com/itanium-cxx-abi/cxx-abi/issues/47 (STEP 3): mangling for
template argument is prefixed by mangling of template parameter declaration
if it's not "obvious", for example because the template parameter is
constrained (we already implemented STEP 1 and STEP 2).
This changes the manglings for a few cases:
- Functions and function templates with constraints.
- Function templates with template parameters with deduced types:
`typename<auto N> void f();`
- Function templates with template template parameters where the argument has a
different template-head:
`template<template<typename...T>> void f(); f<std::vector>();`
In each case where a mangling changed, the change fixes a mangling collision.
Note that only function templates are affected, not class templates or variable
templates, and only new constructs (template parameters with deduced types,
constrained templates) and esoteric constructs (templates with template
template parameters with non-matching template template arguments, most of
which Clang still does not accept by default due to
`-frelaxed-template-template-args` not being enabled by default), so the risk
to ABI stability from this change is relatively low. Nonetheless,
`-fclang-abi-compat=17` can be used to restore the old manglings for cases
which we could successfully but incorrectly mangle before.
Fixes#48216, #49884, #61273
Reviewed By: erichkeane, #libc_abi
Differential Revision: https://reviews.llvm.org/D147655
Add UNSUPPORTED to pass tests on VE. VE uses SjLj libunwind, so
_Unwind_Backtrace and _Unwind_ForcedUnwind are not implemented.
Reviewed By: MaskRay, #libc_abi, ldionne
Differential Revision: https://reviews.llvm.org/D159446
Support VE in long double demangler. This patch corrects
libcxxabi/test/test_demangle.pass.cpp on VE.
Reviewed By: MaskRay, #libc_abi, ldionne
Differential Revision: https://reviews.llvm.org/D159004
I'm making a change in this area (https://reviews.llvm.org/D138461), so update the test:
* Add proper synchronization instead of a sleep.
* Avoid some unnecessary size_t casts.
* Spawn the number of hardware threads instead of 10.
* Check that `__cxa_get_globals` and `__cxa_get_globals_fast` return
the same values.
* Split the test in with-threads and without-threads tests to simplify
the code.
Differential Revision: https://reviews.llvm.org/D138460
Co-authored-by: Louis Dionne <ldionne.2@gmail.com>
