Summary:The current implementation of `insertImpl` assumes that
`expand_noinit` does not reallocate if the `size() + delta <=
capacity()`, but D3114022 makes this assumption invalid when compiling
with ASan. It also doesn't guarantee exponential growth, so repeated
inserting at the end could trigger quadratic behavior.

The new implementation fixes the problems above, and it's much
simpler.

Reviewed By: yfeldblum, Orvid

Differential Revision: D3119813

fb-gh-sync-id: 946ebeeaf924a531f7f03fb7e79c75e352a50c58
fbshipit-source-id: 946ebeeaf924a531f7f03fb7e79c75e352a50c58

Summary: Alerts may be sent by clients, potentially letting us know why connections fail.

Reviewed By: siyengar

Differential Revision: D3117395

fb-gh-sync-id: bddf51f2399eb9e7e397981d5440adb3e815d6d2
fbshipit-source-id: bddf51f2399eb9e7e397981d5440adb3e815d6d2

Summary: Windows doesn't have it, so be nice and at least stub it out.

Reviewed By: yfeldblum

Differential Revision: D2984232

fb-gh-sync-id: 3e8871ab78c5d7c8785a52af24548245f842f19b
fbshipit-source-id: 3e8871ab78c5d7c8785a52af24548245f842f19b

Summary:Let's break OSX!

Alright, maybe not. Neither OSX nor Windows define malloc_usable_size, so we implement them based on what is available on the respective platforms.
This moves the implementation for OSX out of Portability.h and into the new header, so it likely breaks something on OSX, although I'm not sure what.

Reviewed By: yfeldblum

Differential Revision: D3019938

fb-gh-sync-id: df95faef09535098fb73b7b3479d7a73f6b49712
fbshipit-source-id: df95faef09535098fb73b7b3479d7a73f6b49712

Summary: Adding a benchmark for one the most common scenarios (used in Futures, EventBase, fibers etc).

Reviewed By: yfeldblum

Differential Revision: D3106365

fb-gh-sync-id: 8cb55959b3803e8836ab5a5bdf43fdfc1db02d4c
fbshipit-source-id: 8cb55959b3803e8836ab5a5bdf43fdfc1db02d4c

Summary:It was mistakenly assuming the length passed in included the null terminator.
This also makes the portability implementation of `strndup` available to OSX and FreeBSD, where they weren't present, and where HHVM had a wrapper for them.
This also removes the extra pair of conditions around `memrchr`, as the main define should always be getting set.

Reviewed By: yfeldblum

Differential Revision: D3116467

fb-gh-sync-id: 243dd4dace219efab2c2bf2f383202e70fbec4de
fbshipit-source-id: 243dd4dace219efab2c2bf2f383202e70fbec4de

Summary:MSVC has support in the compiler for the intrinsics required, but both refuses to tell us that, and also gives them proper names.
The code already checks for runtime support, this just enables compiling the SSE 4.2 version in the first place.

Reviewed By: yfeldblum

Differential Revision: D3104296

fb-gh-sync-id: 9143240bede9b756817691fdd86818001267dac1
fbshipit-source-id: 9143240bede9b756817691fdd86818001267dac1

Summary: OpenSSLPtrTypes.h uses CHECK_EQ which is included in glog/logging.

Reviewed By: knekritz

Differential Revision: D3118577

fb-gh-sync-id: f8a00aa5a523bf88a3f783433c6699d555799225
fbshipit-source-id: f8a00aa5a523bf88a3f783433c6699d555799225

Summary: This adds Windows support to the portability header for sys/syscall.h, which was previously named portability/Syscall.h, which is inconsistent with the naming of the other headers, so it's now also been renamed to SysSyscall.h, and the one place that used it updated to reflect the new name.

Reviewed By: yfeldblum

Differential Revision: D2984383

fb-gh-sync-id: c0e4ce55c7bc2aa6db9b084e09d9762ba587002d
fbshipit-source-id: c0e4ce55c7bc2aa6db9b084e09d9762ba587002d

Summary:`fbstring`'s small string optimization prevents ASan to catch invalid
accesses to the data of a destroyed string, for example if a
`StringPiece` is initialized from a temporary string.

This diff disables building a string with the small category when
compiled with ASan: small strings will be constructed as
`Medium`-category strings and heap-allocated. This is done by only
changing the behavior of construction and resizing, so that the ABI is
preserved and it is still possible to link an ASan-enabled object file
with a library that wasn't compiled with ASan.

The diff also fixes a blind spot in `fbstring_core`'s constructor,
which disabled ASan altogether in order to allow fast word-aligned
copy of small strings. Since small string construction is now disabled
under ASan, we don't need to disable it anymore.

Lastly, it always clears moved-from strings, even when they are small.
This improves the performance of the move constructor (no more conditional
needed) and it uncovers another class of potential bugs.

Reviewed By: luciang

Differential Revision: D3114022

fb-gh-sync-id: 4e180fbf2b8aced3b977afc985d26fdf244d9598
fbshipit-source-id: 4e180fbf2b8aced3b977afc985d26fdf244d9598

Summary: Some compiler will complain about this when calling std::min with a size_t and a uint64_t. So use unit64_t for numElements in HHWheelTimer.

Reviewed By: mzlee

Differential Revision: D3116346

fb-gh-sync-id: 67a9eebf4f9e8fe0e732a7292af55122be04163b
fbshipit-source-id: 67a9eebf4f9e8fe0e732a7292af55122be04163b

Summary:We should be able to use HHWheelTimer in fibers. So it should use way less stack.

Alternative solution to D3112305

Reviewed By: haijunz

Differential Revision: D3112558

fb-gh-sync-id: 3a52a37d9f9347639005fdf84524f7f8c3041918
fbshipit-source-id: 3a52a37d9f9347639005fdf84524f7f8c3041918

Summary:[Folly] `IPAddressV[46]::validate(StringPiece)`.

Just perform the check without allocations or throws.

Reviewed By: meyering

Differential Revision: D3103545

fb-gh-sync-id: 2918c1398935738f597b9045cf7dadbe42da2527
fbshipit-source-id: 2918c1398935738f597b9045cf7dadbe42da2527

Summary:Use a ThreadLocalPRNG insteaad of a per context RNG.  This avoids
calls to Random::seed on context creation which can get expensive
when many are created in an application.

Reviewed By: siyengar

Differential Revision: D3105501

fb-gh-sync-id: 92d987c27a1f190a98035ca25c23b994ca915007
fbshipit-source-id: 92d987c27a1f190a98035ca25c23b994ca915007

Summary:Per https://github.com/facebook/folly/issues/379, github user pmalek determined that fbvector's insert function was a lot slower than std::vector's. While insert is often imagined to be a second-class vector function (if you're inserting in the middle, don't use a vector!), it is the correct function to use when inserting multiple elements at the back of an array (since the standard lacks an ##append(multiple-elements)## function). The code, therefore, required optimization.

There are three things that were slowing down fbvector:
- Excessive asserts. The fbvector code contains internal asserts that ensure proper calling of private methods. These are mostly unnecessary, given the extensive test suite that batters fbvector, in addition to fbvector's battle tested history. I have removed these.
- Internal data checks. When inserting into a vector, the elements being inserted may not reference the contents of the vector in question. If the client supplies internal references, then the standard allows for undefined behavior. An old design decision for fbvector was to be forgiving of this error; we check for internal data before performing internal data moves. This is expensive. Further, it is unnecessary when the insertion point is at the end of the vector (which is the valid case we are optimizing) or if the vector is being reallocated. I've rearchitected the insert macros to only perform the internal-data-check when it is absolutely necessary. While rearchitecting the checks, I also reorganized the n==0 base case checking.
- The 'window' function is pretty expensive when called with the end() iterator. It is effectively a no-op, but the function (and some of its called functions) are too big to inline, so the call overhead is non-trivial. I've put window calls behind a conditional to save this overhead.

I added a benchmark test to FBVectorTestBenchmark.cpp.h, but the particular pattern in that file caused the results to be completely optimized away, voiding the test. Oh well.

Running the github benchmarks show a 4x speedup for inserting a single element on the back, bringing it quite close to that of std::vector. The multi-insert function got a bit faster, though the exact number of iterations seems to be having an effect on the numbers (running the tests 10,000 times, intead of only 1,000 times, yields results that are 90% as fast as std::vector, instead of 75%). Both of these cases are now looking quite a bit better.

I resurrected the old StlVectorTest suite, since I was mucking with the complicated insert code. The suite had one latent compilation error (caught with newer compiler warnings - fortunately this particular case was benign). The test suite caught an error in shrink_to_fit, where a clean-slate optimization for empty vectors (see D2696314) that trampled the vector's allocator. I have fixed that small bug in this diff, too.

Reviewed By: ot

Differential Revision: D3105962

fb-gh-sync-id: ac9fa9d7ca79335cf0ff6bb9010af1ed8bd7916b
fbshipit-source-id: ac9fa9d7ca79335cf0ff6bb9010af1ed8bd7916b

Summary:This diff makes it possible to pass callables (such as lambdas) to
folly::Future::then that are not copy-constructible. As a consequence, move-only
types (such as folly::Promise or std::unique_ptr) can now be captured in lambdas
passed to folly::Future::then.

Using C++14 notation, the following is now possible:
  Future<Unit>().then([promise = std::move(promise)]() mutable {
    promise.setValue(123);
  });
See folly/futures/test/NonCopyableLambdaTest.cpp for more examples.

folly::Future uses std::function to store callback functions. (More precisely,
the callback function is stored in a std::function in folly::detail::Core.)
std::function is a copy-constructible type and it requires that the callable
that it stores is copy constructible as well.

This diff changes the implementation of folly::detail::Core to use
folly::Function instead of std::function. It also simplifies the code in
folly::detail::Core a little bit: Core had a reserved space of size
8*sizeof(void*) to store callbacks in-place. Only larger callbacks (capturing
more data) would be stored in the std::function, which puts those on the heap.
folly::Function has a template parameter to set the size of the in-place
callable storage. In Core, it is set to 8*sizeof(void*), so all callbacks that
used to be stored in-place inside Core, are now stored in-place inside
folly::Function. This even reduces the size of a Core object: the
folly::Function object occupies 80 bytes, which is 16 bytes less than what was
used before for a std::function (32 bytes) and the callback storage (64
bytes). The resulting size of a Core<Unit> goes down from 192 to 176 bytes (on
x86_64).

Reviewed By: fugalh

Differential Revision: D2884868

fb-gh-sync-id: 35548890c392f80e6c680676b0f98e711bc03ca3
fbshipit-source-id: 35548890c392f80e6c680676b0f98e711bc03ca3

Summary:AsyncSocket has a few cases where
we might overwrite errno, with another function
call that is done during use.

For example withAddr could cause an EBADF if the fd
was bad.

This fixes these cases by copying errno, before doing
anything else

Reviewed By: afrind

Differential Revision: D3101932

fb-gh-sync-id: 73d4810314c90cb987936a0a3a9dc977af07243a
fbshipit-source-id: 73d4810314c90cb987936a0a3a9dc977af07243a

Summary:[Folly] `s/MAX_STATIC_CONSTRUCTOR_PRIORITY/FOLLY_STATIC_CTOR_PRIORITY_MAX/g`.

Namespace it with a `FOLLY_` prefix - when Folly introduces its own symbols, it is best when they are namespaced.

And stick `_MAX` at the end. Reads a bit more hierarchically to me that way.

Reviewed By: andriigrynenko

Differential Revision: D3095964

fb-gh-sync-id: 4ac8c00c03c8bd146fad9be96ce3849830504d96
fbshipit-source-id: 4ac8c00c03c8bd146fad9be96ce3849830504d96

Summary: This was on for mcrouter for very long time. We should make it default for everyone.

Reviewed By: yfeldblum

Differential Revision: D3023133

fb-gh-sync-id: 6401f86df754e9490d46beb5fd9d0cf4b3675208
fbshipit-source-id: 6401f86df754e9490d46beb5fd9d0cf4b3675208

Summary:If you pass an lvalue to folly::fibers::Fiber::setFunction or
setFunctionFinally, you do not expect that those methods move
the function out of your lvalue.

This diff uses std::forward for perfect forwarding, so if an lvalue
is passed, a copy is made.

If and when when start using folly::Function for storing the
functions in the Fiber object, passing an lvalue will trigger a
compiler error because folly::Function is not copyable. That is
a good thing, as it enforces calling setFunction with std::move
if you use a named object, making clear that the named object
gets moved out of. Often people will pass temporary objects
(like a lambda defined i- place), which a rvalues anyway, so
this will not be a problem anyway.

Reviewed By: andriigrynenko

Differential Revision: D3102685

fb-gh-sync-id: 87bf3135f7f6630766f97be351599ff488e4b796
fbshipit-source-id: 87bf3135f7f6630766f97be351599ff488e4b796

Summary: Windows has it, but some things aren't there, and some have different names.

Reviewed By: yfeldblum

Differential Revision: D2990475

fb-gh-sync-id: 3f30e084eb7789c13b3981ec0fbf5b7b642c3367
fbshipit-source-id: 3f30e084eb7789c13b3981ec0fbf5b7b642c3367

Summary: Portabilty.h is just a mashup of everything that was needed up-to now, however, with the Windows port significantly expanding the scope of portability in Folly, it no longer needs to be a massive mashup, so start things by splitting out the asm stubs.

Reviewed By: yfeldblum

Differential Revision: D3007018

fb-gh-sync-id: f32913f74660aaa8889a553ae5d6f7182a4f5789
fbshipit-source-id: f32913f74660aaa8889a553ae5d6f7182a4f5789

Summary:event_base_new does not change the global current_base so no need to
call it with the mutex held.

Reviewed By: chadparry

Differential Revision: D3055963

fb-gh-sync-id: df420b189ef0d7d56eda5ad6c32e03b195804e6b
fbshipit-source-id: df420b189ef0d7d56eda5ad6c32e03b195804e6b

Summary:Treat any return type as convertible to void:

As of C++17, std::function<void(Args...)> can be set to callables
returning non-void types when called with parameters Args....
This diff adds that capability to folly::Function. It also adds
unit tests, not only for ignoring return types, but also for
correctly converting between the return type of the embedded
callabled and the return type of the encapsulating folly::Function.

Allow conversion of one folly::Function type to another one which
declares a return type the original one can be converted to:

E.g. allow to construct a Function<double()> from a
Function<int()> or a Function<Base*()> from a
Function<Derived*()>.

Reviewed By: yfeldblum

Differential Revision: D3095583

fb-gh-sync-id: 6d924dc6e97f759d8109db4200e1cb9333a98d31
fbshipit-source-id: 6d924dc6e97f759d8109db4200e1cb9333a98d31

Summary: It needs stdint to define uint8_t

Reviewed By: yfeldblum, mzlee

Differential Revision: D3101140

fb-gh-sync-id: 308507af1ccd81e2eb389a400376556a6cc597b2
fbshipit-source-id: 308507af1ccd81e2eb389a400376556a6cc597b2

Summary:On Android, clang 3.8 + libstdc++ doesn't support the full
range of atomic operations.  This diff fixes up the ones we
ran into when building fb4a.

Reviewed By: mzlee

Differential Revision: D3092352

fb-gh-sync-id: 7fea8513e23425fc37050ad14d82aabaceb00352
shipit-source-id: 7fea8513e23425fc37050ad14d82aabaceb00352

Summary:[Folly] Fix Build: `folly` after replacing `FOLLY_NORETURN`.

Problem 1:
* Clang does not treat `[[noreturn]]` and `__attribute__((__noreturn__))` as the same attribute. When a function is declared twice with the noreturn attribute, both declarations must use either the one syntax or the other; Clang fails if the two declarations mix-and-match. When both `folly/detail/FunctionalExcept.h` and gcc49's `bits/functexcept.h` are both included from the same source, they (now) mix-and-match the attribute syntaxes and Clang emits an error.
* `folly/detail/FunctionalExcept.h` should be included only when `bits/functexcept.h` is unavailable - somehow, both headers were being included.

We fix the latter problem and keep our `[[noreturn]]` syntax.

Found from WDT build failures in Travis CI.

Reviewed By: ldemailly

Differential Revision: D3089987

fb-gh-sync-id: 705a087fc2a9629739d6cedd8315d56111204e6d
shipit-source-id: 705a087fc2a9629739d6cedd8315d56111204e6d

Summary:Use C++'s standardized `[[noreturn]]` attribute.

All supported compilers (and some unsupported compilers) also support the standardized syntax.

GCC >= 4.8, Clang >= 3.0, and MSVC >= 2015 have direct support for the C++'s standardized way of writing the attribute.

Clang - http://goo.gl/ftJGVM
GCC 4.8.2 - http://goo.gl/ORCVOD
ICC 13.0.1 - http://goo.gl/I5tn5I
MSVC 2015 - https://msdn.microsoft.com/en-us/library/hh567368.aspx

(Regardling Clang, earlier versions may support it. 3.0 was the earliest Clang version listed at godbolt.com, so that's as far back as I went.)

Therefore, we no longer need to use the compiler-specific syntaxes, or use preprocessor macros with per-compiler definitions:

    __attribute__((__noreturn__))
    __attribute__((noreturn))
    __declspec(noreturn)

Reviewed By: Orvid

Differential Revision: D3073621

fb-gh-sync-id: 32d4771d1bf1974693b8574fa2d39c9559872945
shipit-source-id: 32d4771d1bf1974693b8574fa2d39c9559872945

Summary: Because things belong places.

Reviewed By: mzlee

Differential Revision: D3077737

fb-gh-sync-id: 966f37df722ae737481d776ff1c9b0d5132b7a58
shipit-source-id: 966f37df722ae737481d776ff1c9b0d5132b7a58

Summary: And use that to guard the use of posix_memalign.

Reviewed By: benyl

Differential Revision: D3081743

fb-gh-sync-id: 2c71d9cbeeaeb59d1600d486cccfc2109699ba6b
shipit-source-id: 2c71d9cbeeaeb59d1600d486cccfc2109699ba6b

Summary:std::function is copy-constructible and requires that the callable that it wraps
is copy-constructible as well, which is a constraint that is often inconvenient.
In most cases when using a std::function we don't make use of its
copy-constructibility.

This diff introduces a templated type called folly::Function that is very
similar to a std::function, except it is not copy-constructible and doesn't
require the callable to be either.

Like std::function, Function is a templated type with template parameters
for return type and argument types of the callable, but not the callable's
specific type. It can store function pointers, static member function pointers,
std::function objects, std::reference_wrapper objects and arbitrary callable
types (functors) with matching return and argument types.

Much like std::function, Function will store small callables in-place, so
that no additional memory allocation is necessary. For larger callables,
Function will allocate memory on the heap.

Function has two more template parameters: firstly, an enum parameter of
type folly::FunctionMoveCtor, which defaults to NO_THROW and determines
whether no-except-movability should be guaranteed. If set to NO_THROW,
callables that are not no-except-movable will be stored on the heap, even
if they would fit into the storage area within Function.

Secondly, a size_t parameter (EmbedFunctorSize), which determines the size of
the internal callable storage. If you know the specific type of the callable you
want to store, you can set EmbedFunctorSize to sizeof(CallableType).

The original motivation of this diff was to allow to pass lambdas to
folly::Future::then that are not copy-constructible because they capture
non-copyable types, such as a promise or a unique pointer.

Another diff will shortly follow that changes folly::Future to use
folly::Function instead of std::function for callbacks, thus allowing to
pass non-copyable lambdas to folly::Future::then.

Reviewed By: fugalh

Differential Revision: D2844587

fb-gh-sync-id: 3bee2af75ef8a4eca4409aaa679cc13762cae0d0
shipit-source-id: 3bee2af75ef8a4eca4409aaa679cc13762cae0d0

Summary: The guards already look for __APPLE__, but __ANDROID__ needs this as well.

Reviewed By: yfeldblum

Differential Revision: D3053976

fb-gh-sync-id: a9ba7a612d29502d4a7e691c6741837d4a8b42f0
shipit-source-id: a9ba7a612d29502d4a7e691c6741837d4a8b42f0

This reverts commit bb733f43.

Summary:Since SSLContextManager sets SSL_OP_CIPHER_SERVER_PREFERENCE on the SSL_CTX
when it creates contexts, we may be unable to accommodate any clients who
prefer a different ciphersuite. Having differently weighted cipher preference
lists allows SSLContext to set a list with a different most-preferred cipher
for some fraction of new handshakes.

Note: resumption will work with the previously negotiated ciphersuite even if
the server doesn't explicitly prefer/support it anymore, provided the cipher is
supported in OpenSSL.

Reviewed By: knekritz

Differential Revision: D3050496

fb-gh-sync-id: 1c3b77ce3af87f939f8b8c6fe72b6a64eeaeeeb4
shipit-source-id: 1c3b77ce3af87f939f8b8c6fe72b6a64eeaeeeb4

Summary: This helps account for 32-bit vs. 64-bit compilation targets.

Reviewed By: benyl

Differential Revision: D3076523

fb-gh-sync-id: 56f49daaf55d242ac48f8b8d38a40c6d6230818b
shipit-source-id: 56f49daaf55d242ac48f8b8d38a40c6d6230818b

Summary: A more compact hex representation of ciphers in ClientHello can be useful, e.g., for logging.

Reviewed By: knekritz

Differential Revision: D3052308

fb-gh-sync-id: beaf6fcd4705d4d7fae652d8d8b95b52ca9e07a9
shipit-source-id: beaf6fcd4705d4d7fae652d8d8b95b52ca9e07a9

Summary: This adds Windows support to portability/Memory.h and also does some refactoring to be more correct about when to use posix_memalign, and also changes it to fall back to memalign rather than posix_memalign.

Reviewed By: mzlee

Differential Revision: D3069990

fb-gh-sync-id: 88861583c6028e9fb02a3e3804bd6d476956c555
shipit-source-id: 88861583c6028e9fb02a3e3804bd6d476956c555

Summary:Instead of relying on `push_back()` to guarantee exponential growth,
delegate to `expand_noinit`. This removes the duplication of logic
between the two functions, and significantly speeds up short appends.

```
$ _bin/folly/test/fbstring_benchmark_using_jemalloc --bm_min_usec=100000
                                                          Before:             After:
============================================================================ ===================
./folly/test/FBStringTestBenchmarks.cpp.h       relative  time/iter  iters/s  time/iter  iters/s
============================================================================ ===================
...
BM_push_back_fbstring(1)                                     7.51ns  133.20M     6.87ns  145.49M
BM_push_back_fbstring(23)                                  175.08ns    5.71M   173.92ns    5.75M
BM_push_back_fbstring(127)                                 586.02ns    1.71M   585.70ns    1.71M
BM_push_back_fbstring(1024)                                  3.30us  302.81K     3.41us  293.13K
BM_short_append_fbstring(23)                               367.01ns    2.72M   179.45ns    5.57M
BM_short_append_fbstring(1024)                               9.33us  107.20K     5.72us  174.95K
...
============================================================================ ===================
```

Reviewed By: philippv

Differential Revision: D3075249

fb-gh-sync-id: 497775ba3fc707bf50821a3cf10fb9d247b9352d
shipit-source-id: 497775ba3fc707bf50821a3cf10fb9d247b9352d

Summary: It is perfectly reasonable to need to feature macros when checking configuration, and the configuration is needed to know if features.h exists, so put it in the same file.

Reviewed By: yfeldblum

Differential Revision: D3072255

fb-gh-sync-id: 0f253dd1fa2f4a95adb11f11c45e6b86ef9c25bc
shipit-source-id: 0f253dd1fa2f4a95adb11f11c45e6b86ef9c25bc

Summary: This is needed to avoid a circular dependency.

Reviewed By: yfeldblum

Differential Revision: D3072037

fb-gh-sync-id: 546a46fd68f25f0c5b027d4d938335b2711d8fdd
shipit-source-id: 546a46fd68f25f0c5b027d4d938335b2711d8fdd