This patch implements the forwarding to frozen C++03 headers as
discussed in
https://discourse.llvm.org/t/rfc-freezing-c-03-headers-in-libc. In the
RFC, we initially proposed selecting the right headers from the Clang
driver, however consensus seemed to steer towards handling this in the
library itself. This patch implements that direction.
At a high level, the changes basically amount to making each public
header look like this:
```
// inside <vector>
#ifdef _LIBCPP_CXX03_LANG
# include <__cxx03/vector>
#else
// normal <vector> content
#endif
```
In most cases, public headers are simple umbrella headers so there isn't
much code in the #else branch. In other cases, the #else branch contains
the actual implementation of the header.
This PR deprecates `<ccomplex>`, `<cstdbool>`, `<ctgmath>`, and
`<ciso646>` in C++17 and "removes" them in C++20 by special deprecation
warnings.
`<cstdalign>` is previously missing. This PR also tries to add them, and
then deprecates and "removes" `<cstdalign>`.
Papers:
- https://wg21.link/P0063R3
- https://wg21.link/P0619R4Closes#99985.
---------
Co-authored-by: Louis Dionne <ldionne.2@gmail.com>
Currently, the library-internal feature test macros are only defined if
the feature is not available, and always have the prefix
`_LIBCPP_HAS_NO_`. This patch changes that, so that they are always
defined and have the prefix `_LIBCPP_HAS_` instead. This changes the
canonical use of these macros to `#if _LIBCPP_HAS_FEATURE`, which means
that using an undefined macro (e.g. due to a missing include) is
diagnosed now. While this is rather unlikely currently, a similar change
in `<__configuration/availability.h>` caught a few bugs. This also
improves readability, since it removes the double-negation of `#ifndef
_LIBCPP_HAS_NO_FEATURE`.
The current patch only touches the macros defined in `<__config>`. If
people are happy with this approach, I'll make a follow-up PR to also
change the macros defined in `<__config_site>`.
This patch adds a large number of missing includes in the libc++ headers
and the test suite. Those were found as part of the effort to move
towards a mostly monolithic top-level std module.
## Why
Following up on https://github.com/llvm/llvm-project/pull/105946, this
patch provides the floating point overloads for `std::signbit` as
defined by
[P0533R9](https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2021/p0533r9.pdf).
## What
* Test and add overloads for cv-unqualified floating point types
* Remove constrained overload as it is not needed anymore
* Make use of `template<class = void>` as the universal C runtime (UCRT)
needed for Clang-Cl comes with overloads for all cv-unqualified floating
point types (float, double, long double) for `std::signbit()` by itself
[in the
WinSDK](e012b29924/generation/WinSDK/RecompiledIdlHeaders/ucrt/corecrt_math.h (L309-L322)).
In a certain way, this can be seen as a deviation from the C standard.
We need to work around it as the compilation would otherwise error out
due to duplicated definitions.
## Why
Since 18th of August, the floating point comparison builtin
``__builtin_signbit`` is available in Clang as constant expression
(https://github.com/llvm/llvm-project/pull/94118).
## What
* Implement `constexpr` for `std::signbit()` as defined by
[P0533R9](https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2021/p0533r9.pdf)
(new C++23 feature)
* Restrict execution of tests to tip-of-trunk Clang as builtin is not
yet available (note that builtin is available in GCC)
Many headers include `<cstddef>` just for size_t, and pulling in
additional content (e.g. the traits used for std::byte) is unnecessary.
To solve this problem, this patch splits up `<cstddef>` into
subcomponents so that headers can include only the parts that they
actually require.
This has the added benefit of making the modules build a lot stricter
with respect to IWYU, and also providing a canonical location where we
define `std::size_t` and friends (which were previously defined in
multiple headers like `<cstddef>` and `<ctime>`).
After this patch, there's still many places in the codebase where we
include `<cstddef>` when `<__cstddef/size_t.h>` would be sufficient.
This patch focuses on removing `<cstddef>` includes from __type_traits
to make these headers non-circular with `<cstddef>`. Additional
refactorings can be tackled separately.
## Why
Currently, the following does not work when compiled with clang:
```c++
#include <cmath>
struct ConvertibleToFloat {
operator float();
};
bool test(ConvertibleToFloat x) {
return std::isnormal(x);
}
```
See https://godbolt.org/z/5bos8v67T for differences with respect to
msvc, gcc or icx. It fails for `float`, `double` and `long double` (all
cv-unqualified floating-point types).
## What
Test and provide overloads as expected by the ISO C++ standard. The
classification/comparison function `isnormal` is defined since C++11
until C++23 as
```c++
bool isnormal( float num );
bool isnormal( double num );
bool isnormal( long double num );
```
and since C++23 as
```c++
constexpr bool isnormal( /* floating-point-type */ num );
```
for which "the library provides overloads for all cv-unqualified
floating-point types as the type of the parameter num". See §28.7.1/1 in
the [ISO C++
standard](https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2023/n4950.pdf)
or check
[cppreference](https://en.cppreference.com/w/cpp/numeric/math/isnormal).
Following up on https://github.com/llvm/llvm-project/pull/98841.
Changes:
- Properly test convertible types for `std::isnan()` and `std::inf()`
- Tighten conditional in `cmath.pass.cpp` (Find insights on `_LIBCPP_PREFERRED_OVERLOAD` below)
- Tighten preprocessor guard in `traits.h`
Insights into why `_LIBCPP_PREFERRED_OVERLOAD` is needed:
(i) When libc++ is layered on top of glibc on Linux, glibc's `math.h` is
included. When compiling with `-std=c++03`, this header brings the
function declaration of `isinf(double)` [1] and `isnan(double)` [2]
into scope. This differs from the C99 Standard as only the macros
`#define isnan(arg)` and `#define isinf(arg)` are expected.
Therefore, libc++ needs to respect the presense of the `double` overload
and cannot redefine it as it will conflict with the declaration already
in scope. For `-std=c++11` and beyond this issue is fixed, as glibc
guards both the `isinf` and `isnan` by preprocessor macros.
(ii) When libc++ is layered on top of Bionic's libc, `math.h` exposes a
function prototype for `isinf(double)` with return type `int`. This
function prototype in Bionic's libc is not guarded by any preprocessor
macros [3].
`_LIBCPP_PREFERRED_OVERLOAD` specifies that a given overload is a better match
than an otherwise equally good function declaration. This is implemented in
modern versions of Clang via `__attribute__((__enable_if__))`, and not elsewhere.
See [4] for details. We use `_LIBCPP_PREFERRED_OVERLOAD` to define overloads in
the global namespace that displace the overloads provided by the C
libraries mentioned above.
[1]: fe94080875/math/bits/mathcalls.h (L185-L194)
[2]: fe94080875/math/bits/mathcalls.h (L222-L231)
[3]: https://cs.android.com/android/platform/superproject/+/master:bionic/libc/include/math.h;l=322-323;drc=master?hl=fr-BE%22https:%2F%2Fsupport.google.com%2Fmerchants%2Fanswer%2F188494%5C%22%22https:%2F%2Fsupport.google.com%2Fmerchants%2Fanswer%2F188494%5C%22
[4]: 5fd17ab1b0
The HEX float on z/OS does not have infinity nor NaN. In addition, the
limits are smaller before the overflow occurs in mathematical
calculations. This PR accounts for this.
FYI, this LIT test was recently added in PR
[89982](https://github.com/llvm/llvm-project/pull/89982)
This is in relation to mr #93350. It was merged to main, but reverted
because of failing sanitizer builds on PowerPC.
The fix includes replacing the hard-coded threshold constants (e.g.
`__overflow_threshold`) for different floating-point sizes by a general
computation using `std::ldexp`. Thus, it should now work for all architectures.
This has the drawback of not being `constexpr` anymore as `std::ldexp`
is not implemented as `constexpr` (even though the standard mandates it
for C++23).
Closes#92782
Investigating #96612 shows our implementation was different from the
Standard and could cause UB. Testing the codegen showed quite a bit of
assembly generated for these functions. The functions have been written
differently which allows Clang to optimize the code to use simple CPU
rotate instructions.
Fixes: https://github.com/llvm/llvm-project/issues/96612
This takes 1m40s to run when testing picolib on qemu. This isn't the end
of the world but that's on an AArch64 server. So if someone felt the
need to mark this unsupported in the first place, it's likely much
slower on average hardware.
After we switched to LLVM version 20, some libc++ tests started failing
on Windows. This patch adds the clang-20 condition to XFAIL to fix the
issue. The way that these tests are excluded from Windows are fragile
and need to be updated every time we bump the LLVM version.
The 3-dimentionsional `std::hypot(x,y,z)` was sub-optimally implemented.
This lead to possible over-/underflows in (intermediate) results which
can be circumvented by this proposed change.
The idea is to to scale the arguments (see linked issue for full
discussion).
Tests have been added for problematic over- and underflows.
Closes#92782
Implementing the Hermite polynomials which are part of C++17's
mathematical special functions. The goal is to get early feedback which
will make implementing the other functions easier. Integration of
functions in chunks (e.g. `std::hermite` at first, then `std::laguerre`,
etc.) might make sense as well (also see note on boost.math below).
I started out from this abandoned merge request:
https://reviews.llvm.org/D58876 .
The C++23 standard defines them in-terms of `/* floating-point type */`
arguments. I have not looked into that.
Note, there is still an ongoing discussion on discourse whether
importing boost.math is an option.
This is extracted from #94620. While libc++ doesn't have the problem
described in that issue, a test case is a good idea to ensure that we
don't regress this behavior in the future. This could happen for example
if we decide to use `_Complex` in the implementation of `std::complex`
while Clang doesn't handle bit_cast with _Complex yet.
This significantly simplifies the implementation and improves the
codegen. The only downside is that the accuracy can be marginally worse,
but that is up to the compiler to decide with this change, which means
it can be controlled by compiler flags.
Differential Revision: https://reviews.llvm.org/D155312
This patch removes many annotations that are not relevant anymore since
we don't support or test back-deploying to macOS < 10.13. It also cleans
up raw usage of target triples to identify versions of dylibs shipped on
prior versions of macOS, and uses the target-agnostic Lit features
instead. Finally, it reorders both the Lit backdeployment features and
the corresponding availability macros in the library in a way that makes
more sense, and reformulates the Lit backdeployment features in terms of
when a version of LLVM was introduced instead of encoding the system
versions on which it hasn't been introduced yet. Although one can be
derived from the other, encoding the negative form is extremely
error-prone.
Fixes#80901
We were not making any distinction between e.g. the "Apple-flavored"
libc++ built from trunk and the system-provided standard library on
Apple platforms. For example, any test that would be XFAILed on a
back-deployment target would unexpectedly pass when run on that
deployment target against the tip of trunk Apple-flavored libc++. In
reality, that test would be expected to pass because we're running
against the latest libc++, even if it is Apple-flavored.
To solve this issue, we introduce a new feature that describes whether
the Standard Library in use is the one provided by the system by
default, and that notion is different from the underlying standard
library flavor. We also refactor the existing Lit features to make a
distinction between availability markup and the library we're running
against at runtime, which otherwise limit the flexibility of what we can
express in the test suite. Finally, we refactor some of the
back-deployment versions that were incorrect (such as thinking that LLVM
10 was introduced in macOS 11, when in reality macOS 11 was synced with
LLVM 11).
Fixes#82107
I noticed that these tests had empty `main` functions. Dropping them and
renaming the tests to `MEOW.compile.pass.cpp` will slightly improve test
throughput.
* Guard `std::__make_from_tuple_impl` tests with `#ifdef _LIBCPP_VERSION` and `LIBCPP_STATIC_ASSERT`.
* Change `_LIBCPP_CONSTEXPR_SINCE_CXX20` to `TEST_CONSTEXPR_CXX20`.
+ Other functions in `variant.swap/swap.pass.cpp` were already using the proper test macro.
* Mark `what` as `[[maybe_unused]]` when used by `TEST_LIBCPP_REQUIRE`.
+ This updates one occurrence in `libcxx/test/libcxx` for consistency.
* Windows `_putenv_s()` takes 2 arguments, not 3.
+ See MSVC documentation: https://learn.microsoft.com/en-us/cpp/c-runtime-library/reference/putenv-s-wputenv-s?view=msvc-170
+ POSIX `setenv()` takes `int overwrite`, but Windows `_putenv_s()` always overwrites.
* Avoid non-Standard zero-length arrays.
+ Followup to #74183 and #79792.
* Add `operator++()` to `unsized_it`.
+ The Standard requires this due to [N4981][] [move.iter.requirements]/1 "The template parameter `Iterator` shall
either meet the *Cpp17InputIterator* requirements ([input.iterators])
or model `input_iterator` ([iterator.concept.input])."
+ MSVC's STL requires this because it has a strengthened exception
specification in `move_iterator` that inspects the underlying iterator's
increment operator.
* `uniform_int_distribution` forbids `int8_t`/`uint8_t`.
+ See [N4981][] [rand.req.genl]/1.5. MSVC's STL enforces this.
+ Note that when changing the distribution's `IntType`, we need to be
careful to preserve the original value range of `[0, max_input]`.
* fstreams are constructible from `const fs::path::value_type*` on wide systems.
+ See [ifstream.cons], [ofstream.cons], [fstream.cons].
* In `msvc_stdlib_force_include.h`, map `_HAS_CXX23` to `TEST_STD_VER` 23 instead of 99.
+ On 2023-05-23, 71400505ca
started recognizing 23 as a distinct value.
* Fix test name typo: `destory_elements.pass.cpp` => `destroy_elements.pass.cpp`
[N4981]: https://wg21.link/N4981
Found while running libc++'s tests with MSVC's STL.
* Avoid MSVC warning C5101: use of preprocessor directive in
function-like macro argument list is undefined behavior.
+ We can easily make this portable by extracting `const bool is_newlib`.
+ Followup to #73440.
+ See #73598.
+ See #73836.
* Avoid MSVC warning C4267: 'return': conversion from 'size_t' to 'int',
possible loss of data.
+ This warning is valid, but harmless for the test, so
`static_cast<int>` will avoid it.
* Avoid MSVC warning C4146: unary minus operator applied to unsigned
type, result still unsigned.
+ This warning is also valid (the scenario is sometimes intentional, but
surprising enough that it's worth warning about). This is a C++17 test,
so we can easily avoid it by testing `is_signed_v` at compile-time
before testing `m < 0` and `n < 0` at run-time.
* Silence MSVC warning C4310: cast truncates constant value.
+ These warnings are being emitted by `T(255)`. Disabling the warning is
simpler than attempting to restructure the code.
+ Followup to #79791.
* MSVC no longer emits warning C4521: multiple copy constructors
specified.
+ This warning was removed from the compiler, since at least 2021-12-09.
There were various places where we incorrectly handled exceptions in the
PSTL. Typical issues were missing `noexcept` and taking iterators by
value instead of by reference.
This patch fixes those inconsistent and incorrect instances, and adds
proper tests for all of those. Note that the previous tests were often
incorrectly turned into no-ops by the compiler due to copy ellision,
which doesn't happen with these new tests.
Adjust some of the [rand.dist] critical values that are too strict
- Most critical values are determined empirically by running each test
51
times with a different PRNG seed and finding the smallest symmetric
interval
around the median that contains 90% of the sample means, variances, etc.
- For the Kolmogorov-Smirnov tests, the alpha=0.1 critical value for
large N
is 1.224/sqrt(N).
- For normally distributed variates, the sample kurtosis is distributed
as
Normal(0, 24/N). For N=1e5, this gives a 90% confidence interval of
0+/-0.0255. For Binomial(40, 0.25), which is approximately normal, the
kurtosis is -0.0167, so the relative 90% CI is large, on the order of
0.0255/0.0167 = 153%. In most cases the distribution of the sample
kurtosis
isn't known analytically, but similarly large relative tolerances can be
expected if the kurtosis is near zero.
This PR is a followup to #81080.
This PR makes two major changes to how the LCG operation is computed:
The first is that I added an additional case where `ax + c` might
overflow the intermediate variable, but `ax` by itself won't. In this
case, it's much better to use `(ax mod m) + c mod m` than the previous
behavior of falling back to Schrage's algorithm. The addition modulo is
done in the same way as when using Schrage's algorithm (i.e. `x += c -
(x >= m - c)*m`), but the multiplication modulo is calculated directly,
which is faster.
The second is that I added handling for the case where the `ax`
intermediate might overflow, but Schrage's algorithm doesn't apply (i.e.
r > q). In this case, the only real option is to increase the precision
of the intermediate values. The good news is that - for `x`, `a`, and
`c` being n-bit values - `ax + c` will never overflow a 2n-bit
intermediary, meaning this promotion can only happen once, and will
always be able to use the simplest implementation. This is already the
case for 16-bit LCGs, as libcxx chooses to compute them with 32-bit
intermediate values. For 32-bit LCGs, I simply added code similar to the
16-bit case to use the existing 64-bit implementations. Lastly, for
64-bit LCGs, I wrote a case that calculates it using `unsigned __int128`
if it is available to use.
While this implementation covers a *lot* of the missing cases from
#81080, this still won't compile **every** possible
`linear_congruential_engine`. Specifically, if `a`, `c`, and `m` are
chosen such that it needs 128-bit integers, but the platform doesn't
support `__int128` (eg. 32-bit x86), then it will fail to compile.
However, this is a fairly rare case to see actually used, and libcxx
would be in good company with this, as [libstdc++ also fails to compile
under these
circumstances](https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87744).
Fixing **this** gap would require even **more** work of further
complexity, so that would probably be best handled by a different PR
(I'll put more details on what that PR would entail in a comment).
On z/OS int128 is disabled causing one of the cases in
`saturate_cast.pass.cpp` to fail. The failure is only in 64-bit mode.
In this case `the std::numeric_limits<long long int>::max()` is within
`std::numeric_limits<unsigned long int>::min()`
and `std::numeric_limits<unsigned long int>::max()` therefore,
saturate_cast<unsigned long int>( sBigMax) == LONG_MAX and not ULONG_MAX
as original test.
In 32-bit, `saturate_cast<unsigned long int>( sBigMax) == ULONG_MAX`
like on other platforms where int128 is enabled.
This PR is required to pass this test case on z/OS and possibly on other
platforms where int128 is not supported/enabled.
---------
Co-authored-by: Sean Perry <perry@ca.ibm.com>
This fixes two major mistakes in the implementation of
`linear_congruential_engine` that allowed it to produce incorrect
output. Specifically, these mistakes are in `__lce_alg_picker`, which is
used to determine whether Schrage's algorithm is valid and needed.
The first mistake is in the definition of `_OverflowOK`. The code
comment and the description of [D65041](https://reviews.llvm.org/D65041)
both indicate that it's supposed to be true iff `m` is a power of two.
However, the definition used does not work out to that, and instead is
true whenever `m` is even. This could result in
`linear_congruential_engine` using an invalid implementation, as it
would incorrectly assume that any integer overflow can't change the
result. I changed the implementation to one that accurately checks if
`m` is a power of two. Technically, this implementation has an edge case
where it considers `0` to be a power of two, but in this case this is
actually accurate behavior, as `m = 0` indicates a modulus of 2^w where
w is the size of `result_type` in bits, which *is* a power of two.
The second mistake is in the static assert. The original static assert
erroneously included an unnecessary `a != 0 || m != 0`. Combined with
the `|| !_MightOverflow`, this actually resulted in the static assert
being impossible to fail. Applying De Morgan's law and expanding
`_MightOverflow` gives that the only way this static assert can be
triggered is if `a == 0 && m == 0 && a != 0 && m != 0 && ...`, which
clearly cannot be true. I simply removed the explicit checks against `a`
and `m`, as the intended checks are already included in `_MightOverflow`
and `_SchrageOK`, and their inclusion doesn't provide any obvious
semantic benefit.
This should fix all the current instances where
`linear_congruential_engine` uses an invalid implementation. This
technically isn't a complete implementation, though, since the static
assert will cause some instantiations of `linear_congruential_engine`
not disallowed by the standard from compiling. However, this should
still be an improvement, as all compiling instantiations of
`linear_congruential_engine` should use a valid implementation. Fixing
the cases where the static assert triggers will require adding
additional implementations, some of which will be fairly non-trivial, so
I'd rather leave those for another PR so they don't hold up these more
important fixes.
Fixes#33554
The valarray<>::operator[](...) const functions return proxy objects.
The valarray<>::operator[](...) functions return valarray objects.
However the standard allows functions returning valarray objects to
return custom proxy objects instead. Libc++ returns __val_expr proxies.
Functions taking a valarray object must work with the custom proxies
too. Therefore several operations have a custom proxy overload instead
of valarray overloads.
Libc++ doesn't specify a valarray overload. This is an issue with the
standard proxy types; these can implicitly be converted to a valarray.
The solution is to allow the standard proxies to behave as-if they are
custom proxies.
This patch fixes the valarray compound assignments. Other operations,
like the binary non-member functions are not fixed. These will be done
in a followup patch.
Fixes: https://github.com/llvm/llvm-project/issues/21320
* `libcxx/test/std/algorithms/alg.nonmodifying/alg.find/find.pass.cpp`
emits a bunch of warnings, all caused by what appears to be intentional
code:
+ Silence MSVC warning C4245: conversion from `'int'` to `'wchar_t'`,
signed/unsigned mismatch
- Caused by: `test<U>(0, -1);`
+ Silence MSVC warning C4305: 'argument': truncation from `'int'` to
`'bool'`
- Caused by: `test<U>(0, -1);`
+ Silence MSVC warning C4310: cast truncates constant value
- Caused by: `test<U>(T(-129), U(-129));`
+ Silence MSVC warning C4805: `'=='`: unsafe mix of type `'char'` and
type `'bool'` in operation
- Caused by: `bool expect_match = val == to_find;`
*
`libcxx/test/std/algorithms/alg.nonmodifying/alg.fold/left_folds.pass.cpp`
+ Silence MSVC warning C4244: 'argument': conversion from `'double'` to
`'const int'`, possible loss of data
- Caused by `[](int const x, double const y) { return x + y; }`
deliberately being given `double`s to truncate.
*
`libcxx/test/std/numerics/numeric.ops/numeric.ops.midpoint/midpoint.pointer.pass.cpp`
+ Silence MSVC warnings about C++20 deprecated `volatile`.
- Caused by: `runtime_test< volatile T>();`
