__builtin_addressof was added to the GCC trunk in the past week. This patch
teaches libc++ about it so it can correctly provide constexpr addressof.
Unfortunately this patch will break users of earlier GCC 7 builds, since
we expect __builtin_addressof but one won't be provided. One option would be
to only use __builtin_addressof for GCC 7.1 and above, but that means
waiting for another release.
Instead I've specifically chosen to break older GCC 7 versions. Since GCC 7
has yet to be released, and the 7.0 release is a development release, I
believe that anybody currently using GCC 7.0 will have no issue upgrading.
llvm-svn: 283715
r283659 changed the argument to gen_link_script.py from SCRIPT_ABI_LIBNAME to
LIBCXX_LIBRARIES_PUBLIC, assuming that all of the items in the
LIBCXX_LIBRARIES_PUBLIC list were library names. This is not right, however,
for in-tree libcxxabi builds, we might have the target name in this list. There
was special logic to fixup SCRIPT_ABI_LIBNAME for this situation; change it to
apply a similar fixup for LIBCXX_LIBRARIES_PUBLIC.
llvm-svn: 283684
Introduce LIBCXX_LIBRARIES_PUBLIC in addition to LIBCXX_LIBRARIES that
holds 'public' interface libraries -- that is, libraries that both
libc++ links to and programs linked against it need to link to.
Currently this includes the ABI library and optionally -lunwind (when
LIBCXXABI_USE_LLVM_UNWINDER is on). The libraries are included in the
linker script, in order to make it possible to link C++ programs using
clang with compiler-rt runtime out-of-the-box.
Differential Revision: https://reviews.llvm.org/D25008
llvm-svn: 283659
Summary:
To quote STL the problems with stack allocator are"
>"stack_allocator<T, N> is seriously nonconformant to N4582 17.6.3.5 [allocator.requirements].
> First, it lacks a rebinding constructor. (The nested "struct rebind" isn't sufficient.)
> Second, it lacks templated equality/inequality.
> Third, it completely ignores alignment.
> Finally, and most severely, the Standard forbids its existence. Allocators are forbidden from returning memory "inside themselves". This requirement is implied by the Standard's requirements for rebinding and equality. It's permitted to return memory from a separate buffer object on the stack, though."
This patch attempts to address all of those issues.
First, instead of storing the buffer inside the allocator I've change `stack_allocator` to accept the buffer as an argument.
Second, in order to fix rebinding I changed the parameter list from `<class T, size_t NumElements>` to `<class T, size_t NumBytes>`. This allows allocator rebinding
between types that have different sizes.
Third, I added copy and rebinding constructors and assignment operators.
And finally I fixed the allocation logic to always return properly aligned storage.
Reviewers: mclow.lists, howard.hinnant, STL_MSFT
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D25154
llvm-svn: 283631
This was caused by r281673, specifically changing `_LIBCPP_EXTERN_TEMPLATE_TYPE_VIS`
from `__attribute__((__type_visibility__("default")))` to
`__attribute__((__visibility("default")))`.
I made that change because I thought the external instantiations needed
their members to have default visibility. However since libc++ never builds
with -fvisibility=hidden this appears not to be needed. Instead this change
caused previously hidden inline methods to become un-hidden, which is a regression.
This patch reverts the problematic change and fixes PR30642.
llvm-svn: 283620
* Fix self-swap. Patch from Casey Carter.
* Remove workarounds and tests for types with deleted move constructors. This
was originally added as part of a LWG proposed resolution that has since
changed.
* Re-apply most recent PR for LWG 2769.
* Re-apply most recent PR for LWG 2754. Specifically fix the SFINAE checks to
use the decayed type.
* Fix tests to allow moved-from std::any's to have a non-empty state. This is
the behavior of MSVC's std::any.
* Various whitespace and test fixes.
llvm-svn: 283606
First batch of changes to get some of these XFAILs working in the
no-exceptions libc++ variant.
Changed some XFAILs to UNSUPPORTED where the test is all about exception
handling. In other cases, used the test macros TEST_THROW and
TEST_HAS_NO_EXCEPTIONS to conditionally exclude those parts of the test
that concerns exception handling behaviour.
Reviewers: EricWF, mclow.lists
Differential revision: https://reviews.llvm.org/D24562
llvm-svn: 283441
Summary:
The current implementation of `hash_code()` for uniqued RTTI strings violates strict aliasing by dereferencing a type-punned pointer. Specifically it generates a `const char**` pointer from the address of the `__name` member before casting it to `const size_t*` and dereferencing it to get the hash. This is really just a complex and incorrect way of writing `reinterpret_cast<size_t>(__name)`.
This patch changes the conversion sequence so that it no longer contains UB.
Reviewers: howard.hinnant, mclow.lists
Subscribers: rjmccall, cfe-commits
Differential Revision: https://reviews.llvm.org/D24012
llvm-svn: 283408
Summary: There's a macro scheme already being used for __has_feature etc. Use it for __has_include too, which makes MSVC happy (it doesn't support __has_include yet, and unguarded use explodes horribly).
Reviewers: mclow.lists, EricWF
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D25251
llvm-svn: 283260
The libc-provided isnan/isinf/isfinite macro implementations are specifically
designed to function correctly, even in the presence of -ffast-math (or, more
specifically, -ffinite-math-only). As such, on most implementation, these
either always turn into external function calls (e.g. glibc) or are
specifically function calls when FINITE_MATH_ONLY is defined (e.g. Darwin).
Our implementation of complex arithmetic makes heavy use of isnan/isinf/isfinite
to deal with corner cases involving non-finite quantities. This was problematic
in two respects:
1. On systems where these are always function calls (e.g. Linux/glibc), there was a
performance penalty
2. When compiling with -ffast-math, there was a significant performance
penalty (in fact, on Darwin and systems with similar implementations, the code
may in fact be slower than not using -ffast-math, because the inline
definitions provided by libc become unavailable to prevent the checks from
being optimized out).
Eliding these inf/nan checks in -ffast-math mode is consistent with what
happens with libstdc++, and in my experience, what users expect. This is
critical to getting high-performance code when using complex<T>. This change
replaces uses of those functions on basic floating-point types with calls to
__builtin_isnan/isinf/isfinite, which Clang will always expand inline. When
using -ffast-math (or -ffinite-math-only), the optimizer will remove the checks
as expected.
Differential Revision: https://reviews.llvm.org/D18639
llvm-svn: 283051
Add underscore aliases for strtof_l and strtod_l. _strtold_l exists in
VS 2013 and above, so fix that definition as a drive-by fix.
Differential Revision: https://reviews.llvm.org/D25059
llvm-svn: 282681
Replace a stale reference to cxx_EXPORTS with _LIBCPP_BUILDING_LIBRARY,
and clarify why the operator new and delete family of functions are
marked dllexport when building but *not* dllimport when including the
header externally.
The new code is identical to the intent of the old code (and would be
functionally equivalent were cxx_EXPORTS still defined when building
libc++). The overall behavior is not ideal, since Microsoft's operator
new and delete functions will get called instead of libc++'s, but I
think consistently calling msvcrt's functions is better than either
calling msvcrt's or libc++'s functions depending on header inclusion.
Differential Revision: https://reviews.llvm.org/D25042
llvm-svn: 282644
This patch applies the _LIBCPP_SAFE_STATIC attribute to internal globals,
most of which are locking primitives, in order to ensure that they can
safely be used during program startup.
This patch also fixes an unsafe static init issue with the global locks
used to implement atomic operations on shared pointers. Previously the
locks were initialized using a dynamically initialized pointer, so it was
possible that the pointer was uninitialized.
llvm-svn: 282640
The ::stat struct on Linux, FreeBSD, and OS X provides the access and
modification times as an instance of 'timespec', which has a nanosecond
resolution. The 'st_mtime' and 'st_atime' members simply reference the 'tv_sec'
value of the timespec struct. This patch changes 'last_write_time(...)' so that
it extracts both the seconds and nanoseconds values of the last modification
time, providing a more accurate implementation of 'last_write_time(...)'.
Additionally this patch fixes a possible signed integer overflow bug. The
'file_time_type' type cannot represent all possible values returned by
the filesystem. Attempting to construct a 'file_time_type' from one of these
values is undefined behavior. This patch avoids that UB by detecting possible
overflows before the conversion.
llvm-svn: 282634
Revert r282483 as it causes build failures due to missing symbols when
not linking to -lgcc_s (i.e. doing pure LLVM stack build). The patch can
be reintroduced when the build system is fixed to add all needed
libraries (libunwind, compiler-rt).
llvm-svn: 282524
Add the "-Wl,-z,defs" linker option that is used to prevent
underlinking. It is already used by LLVM itself but does not get
propagated into stand-alone build of libc++. This patch ensures
that the option is passed in independently of whether libc++ is built
in-tree or out-of-tree.
Patch by Lei Zhang.
Differential Revision: https://reviews.llvm.org/D24119
llvm-svn: 282483
Strip the set of flags (including debug defs, -m32) that could
be inherited from top-level LLVM build only when in-tree build is
performed. This prevents libcxx from confusingly and undesiredly
stripping user-supplied flags e.g. when performing packaging system
controlled multi-ABI build.
Otherwise, in order to perform 32-bit builds the build scripts would
have to use LIBCXX_BUILD_32_BITS. However, -m32 is only one of the many
different ABI flags for different targets, and it really makes no sense
to add separate CMake options for each possible -m* flag and then keep
a mapping from well-known flags to the custom CMake options.
Differential Revision: https://reviews.llvm.org/D24809
llvm-svn: 282475
Eric Fiselier