Libc++ provides static assertions to detect reference binding issues inside
tuple. This patch adds tests for those diagnostics.
It should be noted that these static assertions technically violate the
standard since it allows these illegal bindings to occur.
Also see https://llvm.org/bugs/show_bug.cgi?id=20855
llvm-svn: 276078
The functions arg, conj, imag, norm, proj, and real have additional overloads
for arguments of integral or floating point types. However these overloads should
not allow conversions to the integral/floating point types, only exact matches.
This patch constrains these functions so they no longer allow conversions.
llvm-svn: 276067
Summary:
This patch does the following:
1. Checks in a copy of the Google Benchmark library into the libc++ repo under `utils/google-benchmark`.
2. Teaches libc++ how to build Google Benchmark against both (A) in-tree libc++ and (B) the platforms native STL.
3. Allows performance benchmarks to be built as part of the libc++ build.
Building the benchmarks (and Google Benchmark) is off by default. It must be enabled using the CMake option `-DLIBCXX_INCLUDE_BENCHMARKS=ON`. When this option is enabled the tests under `libcxx/benchmarks` can be built using the `libcxx-benchmarks` target.
On Linux platforms where libstdc++ is the default STL the CMake option `-DLIBCXX_BUILD_BENCHMARKS_NATIVE_STDLIB=ON` can be used to build each benchmark test against libstdc++ as well. This is useful for comparing performance between standard libraries.
Support for benchmarks is currently very minimal. They must be manually run by the user and there is no mechanism for detecting performance regressions.
Known Issues:
* `-DLIBCXX_INCLUDE_BENCHMARKS=ON` is only supported for Clang, and not GCC, since the `-stdlib=libc++` option is needed to build Google Benchmark.
Reviewers: danalbert, dberlin, chandlerc, mclow.lists, jroelofs
Subscribers: chandlerc, dberlin, tberghammer, danalbert, srhines, hfinkel
Differential Revision: https://reviews.llvm.org/D22240
llvm-svn: 276049
Summary:
This patch attempts to fix the undefined behavior in __tree by changing the node pointer types used throughout. The pointer types are changed for raw pointers in the current ABI and for fancy pointers in ABI V2 (since the fancy pointer types may not be ABI compatible).
The UB in `__tree` arises because tree downcasts the embedded end node and then deferences that pointer. Currently there are 3 node types in __tree.
* `__tree_end_node` which contains the `__left_` pointer. This node is embedded within the container.
* `__tree_node_base` which contains `__right_`, `__parent_` and `__is_black`. This node is used throughout the tree rebalancing algorithms.
* `__tree_node` which contains `__value_`.
Currently `__tree` stores the start of the tree, `__begin_node_`, as a pointer to a `__tree_node`. Additionally the iterators store their position as a pointer to a `__tree_node`. In both of these cases the pointee can be the end node. This is fixed by changing them to store `__tree_end_node` pointers instead.
To make this change I introduced an `__iter_pointer` typedef which is defined to be a pointer to either `__tree_end_node` in the new ABI or `__tree_node` in the current one.
Both `__tree::__begin_node_` and iterator pointers are now stored as `__iter_pointers`.
The other situation where `__tree_end_node` is stored as the wrong type is in `__tree_node_base::__parent_`. Currently `__left_`, `__right_`, and `__parent_` are all `__tree_node_base` pointers. Since the end node will only be stored in `__parent_` the fix is to change `__parent_` to be a pointer to `__tree_end_node`.
To make this change I introduced a `__parent_pointer` typedef which is defined to be a pointer to either `__tree_end_node` in the new ABI or `__tree_node_base` in the current one.
Note that in the new ABI `__iter_pointer` and `__parent_pointer` are the same type (but not in the old one). The confusion between these two types is unfortunate but it was the best solution I could come up with that maintains the ABI.
The typedef changes force a ton of explicit type casts to correct pointer types and to make current code compatible with both the old and new pointer typedefs. This is the bulk of the change and it's really messy. Unfortunately I don't know how to avoid it.
Please let me know what you think.
Reviewers: howard.hinnant, mclow.lists
Subscribers: howard.hinnant, bbannier, cfe-commits
Differential Revision: https://reviews.llvm.org/D20786
llvm-svn: 276003
Constructing a std::locale object from an empty string selects the language
from the current environment variables. If the environment variables name
a locale that doesn't exist, or isn't installed, then the construction of
facets using that locale may throw.
This patch removes tests that use 'std::locale l("")'.
The optimal solution would be to manually set the environment variables
in the test. Unfortunately there is no portable way to do this.
llvm-svn: 275772
man page for mkdir says: "If the parent directory has the set-group-ID bit set,
then so will the newly created directory."
Differential Revision: https://reviews.llvm.org/D22265
llvm-svn: 275760
This patch updates the way libc++ handles checking for libatomic, in part
to prepare for https://reviews.llvm.org/D22073.
Changes:
* 'LIBCXX_HAS_ATOMIC_LIB' is now set whenever libatomic is available even libc++
doesn't need to manually link it.
* 'LIBCXX_HAVE_CXX_ATOMICS_WITH_LIB' is now used to detect when libatomic
needs to be manually linked.
* 'LIBCXX_HAS_ATOMIC_LIB' now adds 'libatomic' as a available feature in the
test suite.
llvm-svn: 275759
This patch does the following:
* It renames `_LIBCPP_TRIVIAL_PAIR_COPY_CTOR` to `_LIBCPP_DEPRECATED_ABI_DISABLE_PAIR_TRIVIAL_COPY_CTOR`.
* It automatically enables this option on FreeBSD in ABI V1, since that's the current ABI FreeBSD ships.
* It cleans up the handling of this option in `std::pair`.
I would like the sign off from the FreeBSD maintainers. They will no longer need to keep their `__config` changes downstream.
I'm still hoping to come up with a better way to maintain the ABI without needing these constructors.
Reviewed in https://reviews.llvm.org/D21329
llvm-svn: 275749
This patch upgrades <tuple> to be C++17 compliant by implementing:
* tuple_size_v: This was forgotten when implementing the other _v traits.
* std::apply: This was added via LFTS v1 in p0220r1.
* std::make_from_tuple: This was added in p0209r2.
llvm-svn: 275745
This patch implements a simple optimization in __hash_table::find. When iterating
the found bucket we only constrain the bucket elements hash if it doesn't
already match the unconstrained hash of the specified key. This prevent
the performance of an expensive modulo operation.
Since the bucket element almost always matches the key, especially when the
load factor is low, this optimization has large performance impacts. For
a unordered_set<int> of random integers this patch improves the performance of
'find(...)' by 40%.
llvm-svn: 275734
From r229162:
Visual Studio's SAL extension uses a macro named __deallocate. This
macro is used pervasively
Using -Werror when building for Windows can force the use of -Wno-#warnings
specifically because of this __deallocate #warning. Instead of forcing
builds to disable all #warnings, this option allows libc++ to be built
without this particular warning, while leaving other #warnings enabled.
Patch by Dave Lee!
llvm-svn: 275172
This cleans up a previous optimization attempt in hash, and results in
additional performance improvements over that previous attempt. Additionally
this new optimization does not hinder the power of 2 bucket count optimization.
llvm-svn: 275114
Summary: The current implementations of __hash_table::find used by std::unordered_set/unordered_map call key_eq on each key that lands in the same bucket as the key you're looking for. However, since equal objects mush hash to the same value, you can short-circuit the possibly expensive call to key_eq by checking the hashes first.
Reviewers: EricWF
Subscribers: kmensah, cfe-commits
Differential Revision: http://reviews.llvm.org/D21510
llvm-svn: 274857
This patch improves the performance of unordered_set's find by 45% when
the value exists within the set. __hash_tables find method
needs to check if it's reached the end of the bucket by constraining the
hash of the current node and checking it against the bucket index. However
constraining the hash is an expensive operations and it can be avoided if the
two unconstrained hashes are equal. This patch applies that optimization.
This patch also adds a top level directory called benchmarks. 'benchmarks/'
is intended to store any/all benchmarks written for the standard library.
Currently nothing is done with files under 'benchmarks/' but I would like
to move towards introducing a formal format and test runner.
llvm-svn: 274423
This patch is the last in a series that replaces recursive meta-programming
in std::tuple with non-recursive implementations.
Previously std::tuple could only be instantiated with 126 elements before
it blew the max template instantiation depth. Now the size of std::tuple is
essentially unbounded (I've tested with over 5000 elements).
One unfortunate side-effect of this change is that tuple_constructible
and similar no longer short circuit after the first failure. Instead they
evaluate the conditions for all elements. This could be potentially breaking.
I plan to look into this further.
llvm-svn: 274331
This patch attempts to improve the QoI of std::tuples tuple_element and
__make_tuple_types helpers. Previously they required O(N) instantiations,
one for every element in the tuple
The new implementations are O(1) after __tuple_indices<Id...> is created.
llvm-svn: 274330
The previous __make_tuple_indices implementation caused O(N) instantiations
and was pretty inefficient. The C++14 __make_integer_sequence implementation
is much better, since it either uses a builtin to generate the sequence or
a very nice Log8(N) implementation provided by richard smith.
This patch moves the __make_integer_sequence implementation into __tuple
and uses it to implement __make_tuple_indices.
Since libc++ can't expose the name 'integer_sequence' in C++11 this patch
also introduces a dummy type '__integer_sequence' which is used when generating
the sequence. One the sequence is generated '__integer_sequence' can be
converted into the required type; either '__tuple_indices' or 'integer_sequence'.
llvm-svn: 274286
* P0163r0: Implemented in r273839.
* LWG 2309: pthread_mutex_lock only returns this error if certain debug flags
were passed during construction. libc++ does not pass these flags. There is
nothing to do.
* LWG 2310: Wording fix. No impact on libc++'s implementation.
* LWG 2312: libc++'s std::tuple implementation already constrains the overloads
based on the number of arguments.
* LWG 2549: libc++'s std::tuple already applied this fix.
* LWG 2674: libc++ already depends on this relaxed wording.
* LWG 2704, 2706, 2707, 2719, 2720, 2721, 2723, 2725, 2728: All of these filesystem were either
submitted by me and fixed before submission, or have already been applied.
llvm-svn: 274214
Since at least the C++11 standard insert iterators are specified
as having ::reference typedef void. Libc++ was not doing that.
This patch corrects the typedef.
This patch changes the std::iterator base class of insert_iterator,
front_insert_iterator and back_insert_iterator. This should not
be an ABI breaking change.
llvm-svn: 274209
The end pointer should point to one past the end of the newly allocated
buffer.
rdar://problem/24265174
Differential Revision: http://reviews.llvm.org/D20334
llvm-svn: 274132
This patch adds the weak_type typedef in shared_ptr. It is available in
C++17 and newer.
This patch also updates the _LIBCPP_STD_VER and TEST_STD_VER macros to
have the value of 16, since 2016 is the current year.
llvm-svn: 273839
See https://llvm.org/bugs/show_bug.cgi?id=27115
The problem was that the conversion from
'const enable_shared_from_this<T>*' to 'const T*' didn't work if
T inherited enable_shared_from_this as a virtual base class. The fix
is to take the original pointer passed to shared_ptr's constructor in the
__enable_weak_this method and perform an upcast to 'const T*' instead of
performing a downcast from the enable_shared_from_this base.
llvm-svn: 273835
The move constructor for wstring_convert accidentally copied the state member
into the converted count member in the move constructor. This patch fixes
the typo.
While working on this I discovered that wstring_convert doesn't actually
provide a move constructor according to the standard and therefore this
constructor is a libc++ extension. I'll look further into whether libc++ should
provide this constructor at all. Neither libstdc++ or MSVC's STL provide it.
llvm-svn: 273831
This patch makes the bind placeholders in std::placeholders both (1) const and
(2) constexpr (See below).
This is technically a breaking change for any code using the placeholders
outside of std::bind and depending on them being non-const. However I don't
think this will break any real world code.
(1) Previously the placeholders were non-const extern globals in all
dialects. This patch changes these extern globals to be const in all dialects.
Since the cv-qualifiers don't participate in name mangling for globals this
is an ABI compatible change.
(2) Make the placeholders constexpr in C++11 and beyond. Although LWG 2488 only
applies to C++17 I don't see any reason not to backport this change.
llvm-svn: 273824
Summary: this fixes build error when built with c++14 and no exceptions
Reviewers: rmaprath
Subscribers: weimingz, grandinj, rmaprath, cfe-commits
Differential Revision: http://reviews.llvm.org/D21673
llvm-svn: 273697
So the macros TEST_HAS_NO_EXCEPTIONS and TEST_HAS_NO_RTTI were always
getting defined because I spelt __cpp_exceptions and __cpp_rtti as
__cxx_exceptions and __cxx_rtti.
Tests incoming after this patch.
llvm-svn: 273381
Libc++ has to deduce the 'allocator_arg_t' parameter as 'AllocArgT' for the
following constructor:
template <class Alloc> tuple(allocator_arg_t, Alloc const&)
Previously libc++ has tried to support tags derived from 'allocator_arg_t' by
using 'is_base_of<AllocArgT, allocator_arg_t>'. However this breaks whenever a
2-tuple contains a reference to an incomplete type as its first parameter.
See https://llvm.org/bugs/show_bug.cgi?id=27684
llvm-svn: 273334
This changes how filesystem::permissions(p, perms) handles symlinks. Previously
symlinks were not resolved by default instead only getting resolved when
"perms::resolve_symlinks" was used. After this change symlinks are resolved
by default and perms::symlink_nofollow must be given to change this.
This issue has not yet been moved to Ready status, and I will revert if it
doesn't get moved at the current meeting. However I feel confident that it
will and it's nice to have implementations when moving issues.
llvm-svn: 273328
The filesystem tests were expecting the paths to the build/source directories
did not contain any symlinks. This patch resolves those symlinks before running
the test suite.
llvm-svn: 273323
Summary:
An implementation of std::experimental::propagate_const from Library Fundamentals Technical Specification v2.
No tests are provided for disallowed types like fancy pointers or function pointers as no code was written to handle these.
Reviewers: EricWF, mclow.lists
Differential Revision: http://reviews.llvm.org/D12486
llvm-svn: 273122
* Fix non-null violation in strstream.cpp
Overflow was calling memcpy with a null parameter and a size of 0.
* Fix std/atomics/atomics.flag/ tests:
a.test_and_set() was reading from an uninitialized atomic, but wasn't
using the value. The tests now clear the flag before performing the
first test_and_set. This allows UBSAN to test that clear doesn't read
an invalid value.
* Fix std/experimental/algorithms/alg.random.sample/sample.pass.cpp
The tests were dereferencing a past-the-end pointer to an array so that
they could do pointer arithmetic with it. Instead of dereference the iterator
I changed the tests to use the special 'base()' test iterator method.
* Add -fno-sanitize=float-divide-by-zero to suppress division by zero UBSAN diagnostics.
The tests that cause float division by zero are explicitly aware that they
are doing that. Since this is well defined for IEEE floats suppress the warnings
for now.
llvm-svn: 273107
Use strtof and strtod for floats and doubles respectively instead of
always using strtold. The other parts of the change are already implemented
in libc++.
This patch also has a drive by fix to wbuffer_convert::underflow() which
prevents it from calling memmove(buff, null, 0).
llvm-svn: 273106
We all know <__tree> and <__hash_table> have plenty of UB that UBSan faithfully
finds. I am working on fixing this. However the noisy output from these failures
prevent automatically detecting regressions elsewhere.
This patch adds a blacklist file for these failures so I can later set up a
UBSAN buildbot.
llvm-svn: 273104
* Fix passing a negative number as either tv_usec or tv_nsec. When file_time_type
is negative and has a non-zero sub-second value we subtract 1 from tv_sec
and make the sub-second duration positive.
* Detect and report when 'file_time_type' cannot be represented by time_t. This
happens when using large/small file_time_type values with a 32 bit time_t.
There is more work to be done in the implementation. It should start to use
stat's st_mtim or st_mtimeval if it's provided as an extension. That way
we can provide a better resolution.
llvm-svn: 273103
fstream has a switch over ios_base::seekdir which provides a defensive default
case. This seems like the right thing for fstream to do, but we need to disable
clangs warning during the build to allow this.
llvm-svn: 273092
Single threaded builds often don't provide a monotonic clock, so we can't
always provide a monotonic SleepFor(...) implementation. Hopefully this
won't cause the builds to hang.
llvm-svn: 273091
This patch fixes the following bugs, all of which were discovered while
testing a 32 bit build on a 64 bit machine.
* path.itr/iterator.pass.cpp has undefined behavior.
'path::iterator' stashes the value of the element inside the iterator.
This violates the BiDirIterator requirements but is allowed for path::iterator.
However this means that using reverse_iterator<path::iterator> has undefined
behavior because it assumes that 'Iter tmp = it; return *tmp' will not create
a dangling reference. However it does, and this caused this particular test
to fail.
* path.native.obs/string_alloc.pass.cpp tested the SSO with a long string.
On 32 bit builds std::wstring only has the SSO for strings of size 2. The
test was using a string of size 4.
* fs.op.space/space.pass.cpp had overflows while calculating the expected values.
The fix here is to convert the statvfs data members to std::uintmax_t before
multiplying them. The internal implementation already does this but the tests
needed to do it as well.
llvm-svn: 273078
Summary:
Currently the implementation of [util.smartptr.shared.atomic] is provided only when using Clang, and not with GCC. This is a relic of not having a GCC implementation of <atomic>, even though <atomic> isn't actually used in the implementation. This patch enables support for atomic shared_ptr functions when using GCC.
Note that this is not a header only change. Previously only Clang builds of libc++.so would provide the required symbols. There is no reason for this restriction.
After this change both Clang and GCC builds should be binary compatible with each other WRT these symbols.
Reviewers: mclow.lists, rmaprath, EricWF
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D21407
llvm-svn: 273076
Currently 4 tests are failing on the ARM buildbot. To try and diagnose each
of the failures this patch does the following:
1) path.itr/iterator.pass.cpp
* Temporarily print iteration sequence to see where its failing.
2) path.native.obs/string_alloc.pass.cpp
* Remove test that ::new is not called when constructing a short string
that requires a conversion. Since during the conversion global locale
objects might be constructed.
3) fs.op.funcs/space.pass.cpp
* Explicitly use uintmax_t in the implementation of space, hopefully
preventing possible overflows.
* Add additional tests that check for overflow is the calculation of the
space_info values.
* Add additional tests for the values returned from statfvs.
4) fs.op.funcs/last_write_time.pass.cpp
* No changes made yet.
llvm-svn: 273075
r273060 didn't completely fix the issues in recursive_directory_iterator and
the tests. This patch follows up with more fixes
* Fix bug where recursive_directory_iterator::increment(ec) did not reset
the error code if no failure occurred.
* Fix bad assertion in the recursive_directory_iterator::increment(ec) test
that would only fire for certain iteration orders.
llvm-svn: 273070
There are two fixes in this patch:
* Fix bug where the constructor of recursive_directory_iterator did not reset
the error code if no failure occurred.
* Fix tests were dependent on the iteration order of the test directories.
llvm-svn: 273060
Add the completed std::experimental::filesystem implementation and tests.
The implementation supports C++11 or newer.
The TS is built as part of 'libc++experimental.a'. Users of the TS need to
manually link this library. Building and testing the TS can be disabled using
the CMake option '-DLIBCXX_ENABLE_FILESYSTEM=OFF'.
Currently 'libc++experimental.a' is not installed by default. To turn on the
installation of the library use '-DLIBCXX_INSTALL_EXPERIMENTAL_LIBRARY=ON'.
llvm-svn: 273034
CloudABI has gained the mblen_l() function in the meantime that does
properly return whether the character set has shift-states (read:
never).
llvm-svn: 272886
Summary:
Android didn't gain GNU's strerror_r until Marshmallow. If we're
building libc++ against something older (we build the NDK library
against the oldest release we support, currently Gingerbread), fall
back to the POSIX version.
Reviewers: mclow.lists, EricWF
Subscribers: tberghammer, danalbert, srhines, cfe-commits
Differential Revision: http://reviews.llvm.org/D21402
llvm-svn: 272827
Summary:
In test/support/test_allocator.h, fix construct() to avoid moving immovable types.
This improves the allocator's conformance, and fixes compiler errors with MSVC's STL. The scenario is when the allocator is asked to construct an object of type X that's immovable (deleted copy/move ctors), but implicitly constructible from an argument type A. When perfectly forwarded, X can be (explicitly) constructed from A, and everything is fine. That's std::allocator's behavior, and the Standard's default when a user allocator's construct() doesn't exist. The previous implementation of construct() here mishandled this scenario. Passing A to this construct() would implicitly construct an X temporary, bound to (non-templated) T&&. Then construct() would attempt to move-construct X from that X temporary, but X is immovable, boom.
Reviewers: mclow.lists, EricWF
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D21094
llvm-svn: 272747
Summary:
Libc++ reexports symbols from the system libc++abi using -reexport_symbols_list. This can cause a linker failure if the list contains symbols not defined in the system libc++abi.
This patch attempts to detect the OS X version and use it to determine the correct symbol list.
It's my understanding that `lib/libc++abi2.exp` should be used on 10.9 and greater. Otherwise 'lib/libc++abi.exp' should be used
This fixes PR25666 (https://llvm.org/bugs/show_bug.cgi?id=25666)
Reviewers: mclow.lists, bcraig, dexonsmith, EricWF
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D20772
llvm-svn: 272723
Summary:
This patch implements the variadic `lock_guard` paper.
Making `lock_guard` variadic is a ABI breaking change because the specialization `lock_guard<_Mutex>` mangles differently then when it was the primary template. This change only provides variadic `lock_guard` in ABI V2 or when `_LIBCPP_ABI_VARIADIC_LOCK_GUARD` is defined.
Note that in ABI V2 `lock_guard` must always be declared as a variadic template, even in C++03, in order to keep the ABI consistent. For this reason `lock_guard` is forward declared as a variadic template in all standard dialects and therefore depends on variadic templates being provided as an extension in C++03. All supported versions of Clang and GCC provide this extension.
Reviewers: mclow.lists
Subscribers: K-ballo, mclow.lists, cfe-commits
Differential Revision: http://reviews.llvm.org/D21260
llvm-svn: 272634
Summary:
system_error::message() uses `strerror` for the generic and system categories. This function is not thread safe.
The fix is to use `strerror_r`. It has been available since 2001 for GNU libc and since BSD 4.4 on FreeBSD/OS X.
On platforms with GNU libc the extended version is used which always returns a valid string, even if an error occurs.
In single-threaded builds `strerror` is still used.
See https://llvm.org/bugs/show_bug.cgi?id=25598
Reviewers: majnemer, mclow.lists
Subscribers: erik65536, cfe-commits, emaste
Differential Revision: http://reviews.llvm.org/D20903
llvm-svn: 272633
Summary:
I haven't added it to all the tests, just those that fail without it
(those that aren't header only).
Reviewers: EricWF, mclow.lists
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D21247
llvm-svn: 272443
Patch by Laman Sole <laxman.g@partner.samsung.com>, Sebastian Pop
<s.pop@samsung.com>, Aditya Kumar <aditya.k7@samsung.com>
Differential Revision: http://reviews.llvm.org/D21103
llvm-svn: 272401
Eric Fiselier