This has been broken out of D104170 since it should be merged whether or
not we go ahead with the module map changes.
Differential Revision: https://reviews.llvm.org/D104175
https://eel.is/c++draft/atomics.types.operations#23 says: ... the value of failure is order except that a value of `memory_order::acq_rel` shall be replaced by the value `memory_order::acquire` and a value of `memory_order::release` shall be replaced by the value `memory_order::relaxed`.
This failure mapping is only handled for `_LIBCPP_HAS_GCC_ATOMIC_IMP`. We are seeing bad code generation for `compare_exchange_strong(cmp, 1, std::memory_order_acq_rel)` when using libc++ in place of libstdc++: https://godbolt.org/z/v3onrrq4G.
This was caught by tsan tests after D99434, `[TSAN] Honor failure memory orders in AtomicCAS`, but appears to be an issue in non-tsan code.
Reviewed By: ldionne, dvyukov
Differential Revision: https://reviews.llvm.org/D103846
The runtimes build has assertions enabled, which is necessary to catch
some of the modules-related issues we've been seeing recently. This
patch enables testing with modules in the runtimes build so as to cover
those cases.
In the future, a better solution would be to systematically use versions
of Clang that have assertions enabled. However, the Clangs we release
currently don't have assertions enabled by default, which causes a
challenge for the CI (we could try to build our own Clang from ToT with
assertions in the CI, but that poses some problems).
Differential Revision: https://reviews.llvm.org/D104252
This started as an attempt to fix a GCC 11 warning of misplaced parentheses.
I then noticed that trying to fix the parentheses warning actually triggered
errors in the tests, showing that we were incorrectly assuming that the
implementation of ranges::advance was using operator+= or operator-=.
This commit fixes that issue and makes the tests easier to follow by
localizing the assertions it makes.
Differential Revision: https://reviews.llvm.org/D103272
The synchronization library was marked as disabled on Apple platforms
up to now because we were not 100% sure that it was going to be ABI
stable. However, it's been some time since we shipped it in upstream
libc++ now and there's been no changes so far. This patch enables the
synchronization library on Apple platforms, and hence commits the ABI
stability as far as that vendor is concerned.
Differential Revision: https://reviews.llvm.org/D96790
Instead, people should be using CMAKE_POSITION_INDEPENDENT_CODE to control
whether they want to use PIC or not. We should try to avoid reinventing
the wheel whenever CMake natively supports something.
This makes libc++abi consistent with libc++ and libunwind.
Differential Revision: https://reviews.llvm.org/D103973
Makes the following operations constexpr:
* `std::swap(optional, optional)`
* `optional(optional<U> const&)`
* `optional(optional<U>&&)`
* `~optional()`
* `operator=(nullopt_t)`
* `operator=(U&&)`
* `operator=(optional<U> const&)`
* `operator=(optional<U>&&)`
* `emplace(Args&&...)`
* `emplace(initializer_list<U>, Args&&...)`
* `swap(optional&)`
* `reset()`
P2231 has been accepted by plenary, with the committee recommending
implementers retroactively apply to C++20. It's necessary for us to
implement _`semiregular-box`_ and _`non-propagating-cache`_, both of
which are required for ranges (otherwise we'll need to reimplement
`std::optional` with these members `constexpr`ified).
Differential Revision: https://reviews.llvm.org/D102119
The post-conditions for the self move assignment of `std::unique_ptr`
were changed. This requires no implementation changes. A test was added
to validate the new post-conditions.
Addresses
- LWG-3455: Incorrect Postconditions on `unique_ptr` move assignment
Reviewed By: #libc, ldionne
Differential Revision: https://reviews.llvm.org/D103764
The buffer size (`__nbuf`) in `num_put::do_put` is currently not an
integral/core constant expression. As a result, `__nar` is a Variable Length
Array (VLA). VLAs are a GNU extension and not part of the base C++ standard, so
unless there is good reason to do so they probably shouldn't be used in any of
the standard library headers. The call to `__iob.flags()` is the only thing
keeping `__nbuf` from being a compile time constant, so the solution here is to
simply err on the side of caution and always allocate a buffer large enough to
fit the base prefix.
Note that, while the base prefix for hex (`0x`) is slightly longer than the
base prefix for octal (`0`), this isn't a concern. The difference in the space
needed for the value portion of the string is enough to make up for this.
(Unless we're working with small, oddly sized types such as a hypothetical
`uint9_t`, the space needed for the value portion in octal is at least 1 more
than the space needed for the value portion in hex).
This PR also adds `constexpr` to `__nbuf` to enforce compile time const-ness
going forward.
Reviewed By: Mordante, #libc, Quuxplusone, ldionne
Differential Revision: https://reviews.llvm.org/D103558
This commit finishes moving the <atomic> design documents to the RST
documentation and removes the old documentation. https://libcxx.llvm.org
is already pointing to the new documentation only now, so the removal of
the old documentation is really a NFC.
I went over the old documentation and I don't think we're leaving anything
important behind - I think everything important was mentionned in the RST
documentation anyway.
Several macros were guarded with a check along the lines of:
#ifndef MACRO
# define MACRO ...
#endif
However, some of these macros are never intended to be defined by users,
so it's pointless to make this check (i.e. the first #ifndef is always
true). This commit removes those checks.
The motivation for doing this cleanup is to remove the impression that
arbitrary configurations macros can be defined by users when including
libc++ headers, which doesn't work reliably and leads to macro spaghetti.
If one needs to be able to override a knob in the __config, that's fine,
but the proper way to do that is to document the macro as being a public
facing knob in the documentation, and most likely to migrate that macro
to __config_site (depending on the nature of the macro).
Differential Revision: https://reviews.llvm.org/D103705
The "root nodes" of the graph are displayed in bold. My intent here
was to bold just the public-API headers, e.g. <vector> and
<experimental/coroutine> and <stdlib.h>, but not helper headers
such as <__functional_base> and <__iterator/next.h>. However,
the recent mass helper-header-ification has exposed defects in
this logic: all the new helpers were ending up bolded! Fix this.
Also, add <__undef_macros> to the list of headers we don't display
by default (like <__config>); it's not interesting to see those edges.
Also, add a sample `dot` command line to the `--help` text.
The `operator[]` of `_UnaryOp` and `_BinaryOp` returns the result of
calling `__op_`, so its return type should be `__result_type`, not
e.g. `_A0::value_type`. However, `_UnaryOp::value_type` also should
never have been `_A0::value_type`; it needs to be the correct type
for the result of the unary op, e.g. `bool` when the op is `logical_not`.
This turns out to matter when multiple operators are nested, e.g.
`+(v == v)` needs to have a `value_type` of `bool`, not `int`,
even when `v` is of type `valarray<int>`.
Differential Revision: https://reviews.llvm.org/D103416
This is a fairly mechanical change, it just moves each algorithm into its own header. This is a NFC.
Note: during this change, I burned down all the includes, so this follows "include only and exactly what you use."
Differential Revision: https://reviews.llvm.org/D103583
As discussed on cfe-dev [1], use the using_if_exists Clang attribute when
the compiler supports it. This makes it easier to port libc++ on top of
new platforms that don't fully support the C Standard library.
Previously, libc++ would fail to build when trying to import a missing
declaration in a <cXXXX> header. With the attribute, the declaration will
simply not be imported into namespace std, and hence it won't be available
for libc++ to use. In many cases, the declarations were *not* actually
required for libc++ to work (they were only surfaced for users to use
them as std::XXXX), so not importing them into namespace std is acceptable.
The same thing could be achieved by conscious usage of `#ifdef` along
with platform detection, however that quickly creates a maintenance
problem as libc++ is ported to new platforms. Furthermore, this problem
is exacerbated when mixed with vendor internal-only platforms, which can
lead to difficulties maintaining a downstream fork of the library.
For the time being, we only use the using_if_exists attribute when it
is supported. At some point in the future, we will start removing #ifdef
paths that are unnecessary when the attribute is supported, and folks
who need those #ifdef paths will be required to use a compiler that
supports the attribute.
[1]: http://lists.llvm.org/pipermail/cfe-dev/2020-June/066038.html
Differential Revision: https://reviews.llvm.org/D90257
If building code like this:
unsigned long val = 1000;
snprintf(buf, sizeof(buf), "%+lu", val);
with clang, clang warns
warning: flag '+' results in undefined behavior with 'u' conversion specifier [-Wformat]
Therefore, don't construct such undefined format strings. (There's
no compiler warnings here, as the compiler can't inspect dynamically
assembled format strings.)
This fixes number formatting in mingw-w64 if built with
`__USE_MINGW_ANSI_STDIO` defined (there, the '+' flag causes a
leading plus to be printed when formatting unsigned numbers too,
while the '+' flag doesn't cause any extra leading plus in other
stdio implementations).
Differential Revision: https://reviews.llvm.org/D103444
Most of our private headers need to be treated as submodules so that
Clang modules can export things correctly. Previous commits that split
monolithic headers into smaller chunks were unaware of this requirement,
and so this is being addressed in one fell swoop. Moving forward, most
new headers will need to have their own submodule (anything that's
conditionally included is exempt from this rule, which means `__support`
headers aren't made into submodules).
This hasn't been marked NFC, since I'm not 100% sure that's the case.
Differential Revision: https://reviews.llvm.org/D103551
D101613 added some macros used by Microsofts SAL. D103425 uses `__pre`
and `__post`. They are also used by SAL and cause issues when used on
Windows. Add them to the blacklist making it easier to figure out what
the issue is.
Differential Revision: https://reviews.llvm.org/D103541
Since D100581, Clang started flagging this variable which is set but
never read. Based on comparing this function with __match_at_start_posix_nosubs
(which is very similar), I am pretty confident that `__j` was simply left
behind as an oversight in Howard's 6afe8b0a23.
Also workaround some unused variable warnings in the <random> tests.
It's pretty lame that we're not asserting the skew and kurtosis of
the binomial and negative binomial distributions, but that should be
tackled separately.
Differential Revision: https://reviews.llvm.org/D103533
This matches the fact that we build the experimental library by default.
Otherwise, by default we'd be building the library but not testing it,
which is inconsistent.
Differential Revision: https://reviews.llvm.org/D102109
This reverts commit 924ea3bb53 *again*, this time because it broke the
LLDB build with modules. We need to figure out what's up with the libc++
modules build once and for all.
Differential Revision: https://reviews.llvm.org/D103369
In 07ef8e6796 and 3ed9f6ebde, `__nbuf` started to diverge from the amount
of space that was actually needed for the buffer. For 32-bit longs for example,
we allocate a buffer that is one larger than needed. Moreover, it is no longer
clear exactly where the extra +1 or +2 comes from - they're just numbers pulled
from thin air. This PR cleans up how `__nbuf` is calculated, and adds comments
to further clarify where each part comes from.
Specifically, it corrects the underestimation of the max size buffer needed
that the above two commits had to compensate for. The root cause looks to be
the use of signed type parameters to numeric_limits<>::digits. Since digits
only counts non-sign bits, the calculation was acting as though (for a signed
64-bit type) the longest value we would print was 2^63 in octal. However,
printing in octal treats values as unsigned, so it is actually 2^64. Thus,
using unsigned types and changing the final +2 to a +1 is probably a better
option.
Reviewed By: #libc, ldionne, Mordante
Differential Revision: https://reviews.llvm.org/D103339
Christopher Di Bella