The optimizer is petulant and temperamental. In this case LLVM failed to lower
the the "insert at end" loop used by`vector<unsigned char>` to a `memset` despite
`memset` being substantially faster over a range of bytes.
LLVM has the ability to lower loops to `memset` whet appropriate, but the
odd nature of libc++'s loops prevented the optimization from taking places.
This patch addresses the issue by rewriting the loops from the form
`do [ ... --__n; } while (__n > 0);` to instead use a for loop over a pointer
range (For example: `for (auto *__i = ...; __i < __e; ++__i)`).
This patch also rewrites the asan annotations to unposion all additional memory
at the start of the loop instead of once per iterations. This could potentially
permit false negatives where the constructor of element N attempts to access
element N + 1 during its construction.
The before and after results for the `BM_ConstructSize/vector_byte/5140480_mean`
benchmark (run 5 times) are:
--------------------------------------------------------------------------------------------
Benchmark Time CPU Iterations
--------------------------------------------------------------------------------------------
Before
------
BM_ConstructSize/vector_byte/5140480_mean 12530140 ns 12469693 ns N/A
BM_ConstructSize/vector_byte/5140480_median 12512818 ns 12445571 ns N/A
BM_ConstructSize/vector_byte/5140480_stddev 106224 ns 107907 ns 5
-----
After
-----
BM_ConstructSize/vector_byte/5140480_mean 167285 ns 166500 ns N/A
BM_ConstructSize/vector_byte/5140480_median 166749 ns 166069 ns N/A
BM_ConstructSize/vector_byte/5140480_stddev 3242 ns 3184 ns 5
llvm-svn: 367183
The constructors for std::pair and std::tuple have been made conditionally
explicit, however the synopsis in the headers do not reflect that.
llvm-svn: 366735
This issue was detected by ASan in one of our tests. This test manually
invokes basic_filebuf::cloe(). fclose(__h.release() returned a non-zero
exit status, so __file_ wasn't set to 0. Later when basic_filebuf
destructor ran, we would enter the if (__file_) block again leading to
heap-use-after-free error.
The POSIX specification for fclose says that independently of the return
value, fclose closes the underlying file descriptor and any further
access (including another call to fclose()) to the stream results in
undefined behavior. This is exactly what happened in our test case.
To avoid this issue, we have to always set __file_ to 0 independently of
the fclose return value.
Differential Revision: https://reviews.llvm.org/D64979
llvm-svn: 366730
This reverts r366593, which caused unforeseen breakage on the build bots.
I'm reverting until the problems have been figured out and fixed.
llvm-svn: 366603
Summary:
This commit allows specifying LIBCXX_ENABLE_PARALLEL_ALGORITHMS when
configuring libc++ in CMake. When that option is enabled, libc++ will
assume that the PSTL can be found somewhere on the CMake module path,
and it will provide the C++17 parallel algorithms based on the PSTL
(that is assumed to be available).
The commit also adds support for running the PSTL tests as part of
the libc++ test suite.
Reviewers: rodgert, EricWF
Subscribers: mgorny, christof, jkorous, dexonsmith, libcxx-commits, mclow.lists, EricWF
Tags: #libc
Differential Revision: https://reviews.llvm.org/D60480
llvm-svn: 366593
Summary:
This is effectively a revert of r344616, which was a partial fix for
PR38964 (compilation of <string> with GCC in C++03 mode). However, that
configuration is explicitly not supported anymore and that partial fix
breaks compilation with Clang when per-TU insulation is provided.
PR42676
rdar://52899715
Reviewers: mclow.lists, EricWF
Subscribers: christof, jkorous, dexonsmith, libcxx-commits
Tags: #libc
Differential Revision: https://reviews.llvm.org/D64941
llvm-svn: 366567
In particular, improve the compile time of the overload set builder
that variant uses to determine which alternative to construct.
Instead of having the __overload type construct itself recursively,
this patch uses a flat construction for the overload set.
llvm-svn: 366033
The standard disallows narrowing conversions when constructing a variant.
This is checked by attempting to perform braced initialization of the
destination type from the argument type. However, braced initialization
can force the compiler (mostly clang) to eagerly instantiate the
constructors of the destintation type -- which can lead to errors in
a non-immediate context.
However, as variant is currently specified, the narrowing checks only
observably apply when the destination type is arithmetic. Meaning we can
skip the check for class types. Hense avoiding the hard errors.
In order to cause fewer build breakages, this patch avoids the narrowing
check except when the destination type is arithmetic.
llvm-svn: 366022
Previously we implemented all one trillion tuple-like constructors using
a single generic overload. This worked fairly well, except that it
differed in behavior from the standard version because it didn't
consider both T&& and T const&. This was observable for certain
types.
This patch addresses that issue by splitting the generic constructor
in two. We now provide both T&& and T const& versions of the
tuple-like constructors (sort of).
llvm-svn: 365973
The paper P0608R3 - "A sane variant converting constructor" disallows
narrowing conversions in variant. It was meant to address this
surprising problem:
std::variant<std::string, bool> v = "abc";
assert(v.index() == 1); // constructs a bool.
However, it also disables every potentially narrowing conversion. For
example:
variant<unsigned> v = 0; // ill-formed
variant<string, double> v2 = 42; // ill-formed (int -> double narrows)
These latter changes break code. A lot of code. Within Google it broke
on the order of a hundred thousand target with thousands of root causes
responsible for the breakages.
Of the breakages related to the narrowing restrictions, none of them
exposed outstanding bugs. However, the breakages caused by boolean
conversions (~13 root causes), all but one of them were bugs.
For this reasons, I am adding a flag to disable the narrowing conversion
changes but not the boolean conversions one.
One purpose of this flag is to allow users to opt-out of breaking changes
in variant until the offending code can be cleaned up. For non-trivial
variant usages the amount of cleanup may be significant.
This flag is also required to support automated tooling, such as
clang-tidy, that can automatically fix code broken by this change.
In order for clang-tidy to know the correct alternative to construct,
it must know what alternative was being constructed previously, which
means running it over the old version of std::variant.
Because this change breaks so much code, I will be implementing the
aforementioned clang-tidy check in the very near future.
Additionally I'm plan present this new information to the committee so they can
re-consider if this is a breaking change we want to make.
I think libc++ should very seriously consider pulling this change
before the 9.0 release branch is cut. But that's a separate discussion
that I will start on the lists.
For now this is the minimal first step.
llvm-svn: 365960
When assigning an initializer list into set/map, libc++ would
leak memory if the initializer list contained equivalent keys
because we failed to check if the insertion was successful.
llvm-svn: 365840
The implementations of __libcpp_mutex_destroy and __libcpp_condvar_destroy
are already NOPs, so this optimization is safe to perform.
See r365273 and PR27658 for more information.
llvm-svn: 365281
Currently std::mutex has a constexpr constructor, but a non-trivial
destruction.
The constexpr constructor is required to ensure the construction of a
mutex with static storage duration happens at compile time, during
constant initialization, and not during dynamic initialization.
This means that static mutex's are always initialized and can be used
safely during dynamic initialization without the "static initialization
order fiasco".
A trivial destructor is important for similar reasons. If a mutex is
used during dynamic initialization it might also be used during program
termination. If a static mutex has a non-trivial destructor it will be
invoked during termination. This can introduce the "static
deinitialization order fiasco".
Additionally, function-local statics emit a guard variable around
non-trivially destructible types. This results in horrible codegen and
adds a runtime cost to every call to that function. non-local static's
also result in slightly worse codegen but it's not as big of a problem.
Example codegen can be found here: https://goo.gl/3CSzbM
Note: This optimization is not safe with every pthread implementation.
Some implementations allocate on the first call to pthread_mutex_lock
and free the allocation in pthread_mutex_destroy.
Also, changing the triviality of the destructor is not an ABI break.
At least to the best of my knowledge :-)
llvm-svn: 365273
types.
It seems some people like to write types that can explicitly convert
to anything, but cannot be used to explicitly construct anything.
This patch makes tuple tolerate such types, as is required
by the standard.
llvm-svn: 365074
Summary:
This fixes a clang-tidy warning when building something that uses
this file.
Reviewers: mclow.lists, EricWF
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D43226
llvm-svn: 364799
when _LIBCPP_ABI_UNSTABLE is defined.
User defined _LIBCPP_ABI_NAMESPACE will still be respected,
but the default version namespace in unstable mode will be the libc++ version
(Currently '__9000').
Previously `_LIBCPP_ABI_VERSION` and `_LIBCPP_ABI_NAMESPACE` were
`1` and `__1` respectively, whuch conflicted with the stable ABI
llvm-svn: 364354
The new meta-programming primitives are lower cost than the old versions. This patch removes those old versions and switches libc++ to use the new ones.
llvm-svn: 364160
Clang provides __is_same that doesn't produce any instantiations
and just returns a bool. It's a lot faster than using std::is_same
I'll follow up with a patch to actually start using it.
llvm-svn: 364148
The CMake CheckLibcxxAtomic module was always failing to compile
the example, even when libatomic wasn't needed. This was caused
because the check doesn't link a C++ runtime library to provide
std::terminate, which is required for exception support.
The check is still really broken, but <atomic> is better!
llvm-svn: 364146
Using class templates instead of alias templates causes a lot of
instantiations. As part of the move away from C++03, we want to
improve the efficiency of our meta-programming.
This patch lays the groundwork by introducing new _If, _EnableIf,
_And, _Or, and _IsValidExpansion (detect member). Future patches
will replace the existing implementations after verifying there
compile time differences.
llvm-svn: 364114
