Summary: gcc-4.9 doesn't appreciate this clever loop and worries about `prev` being uninitialized on first use. Assuming the rest of the program behaves properly and preserves the invariant this code is expecting, this isn't bugged. But let's initialize it anyway.

Reviewed By: meyering

Differential Revision: D4120746

fbshipit-source-id: 1a7c323a4b6296a81768a62dd7d72337c5c35bf0

Reviewed By: yfeldblum

Differential Revision: D4120553

fbshipit-source-id: 87c202232bcfabeff7eb62a8133b9b5577b09cbc

Summary:
There was a doubly-nested loop, both with index "i".
Change the inner for-loop index from i to j.

Reviewed By: yfeldblum

Differential Revision: D4120589

fbshipit-source-id: c786d8f7bb1ee05fc990db898b15836032e93940

Reviewed By: jsedgwick

Differential Revision: D4120563

fbshipit-source-id: c97645a068662e41d5038bcd1e9805bdf6f0490c

Summary: Allow getting the sequence number of a token via a read only getter function

Reviewed By: yfeldblum

Differential Revision: D4120900

fbshipit-source-id: b664dda29d13a1755e041ab9c33c0d7324690ba7

Reviewed By: yfeldblum

Differential Revision: D4086486

fbshipit-source-id: bd0cca8dd2c9b74d5c30e4cd095191c1d1ecab79

Summary:
Solves #509 by creating a libtool convenience library that's compiled with -msse4.2.
Closes https://github.com/facebook/folly/pull/510

Differential Revision: D4120400

Pulled By: yfeldblum

fbshipit-source-id: bf5d937dc37ef8707e755c89f1883d65b4d21b0f

Summary:
We would like to guarantee that after `folly::ThreadLocal<>` dtor
returns no per-thread instances are longer alive.  Currently this is not a case:

* T1 is excuting `StaticMetaBase::onThreadExit()`, it acquired all per-thread
  instances and erased them from meta under `accessAllThreadsLock_`.
* T2 destroys `folly::ThreadLocal<>`, it executes `StaticMetaBase::destroy()`,
  collects all per-thread instances (thus missing the ones being destroyed by
  T1), destroys them and returns.
* T1 executes dtor of per-thread instances, after parent `folly::ThreadLocal<>`
  dtor is finished.

Reviewed By: ot

Differential Revision: D4109820

fbshipit-source-id: d547b8cc77c9871126538c38644c2e98ddccf220

Summary:
Make sequence number in AtomicBatchDispatcher<....>::Token non-const:
- The field is already private and having it as const prevents a
  move assignment operator from getting defined

Reviewed By: yfeldblum

Differential Revision: D4114851

fbshipit-source-id: 9f7a4003daff02c007c01747cca6ae9da17fcb79

Summary:
MSVC doesn't allow uninitialized non-`extern` `const` declarations, but we can't mark them `extern` because they are part of a template specialization. We also can't initialize them, because they'd conflict with the definition specializations in the generated tables.
As the forward-declarations aren't required for it to work under MSVC, as the specializations are found appropriately, just don't declare the specializations under MSVC.

Reviewed By: yfeldblum

Differential Revision: D4098282

fbshipit-source-id: 84b73d63bbe6cf9c54b7fe7a3bfc2488699df7c2

Summary:
In the next revision of clang, compilation generates multiple erroes of the form

   folly/TokenBucket.h:127:29: error: exception specification is not available until end of  class definition
         double nowInSeconds = defaultClockNow()) {

which are eliminated by reordering the method declaration for defaultClockNow()

Refer https://llvm.org/bugs/show_bug.cgi?id=30860

Reviewed By: philippu, yfeldblum

Differential Revision: D4107443

fbshipit-source-id: ce64f2ae7983e533c2fcb5cb043dbdd3da5c00f7

Summary:
MSVC doesn't support them, so use dynamic allocation instead.
This also fixes a place where MSVC wasn't propogating the constexpr-ness of a local into the body of a lambda, resulting in an attempt to create a VLA.

Reviewed By: yfeldblum

Differential Revision: D4098932

fbshipit-source-id: 742b2fcd96f7f6eceb5043159403d557f7fc9673

Summary: Clang and GCC allow it, but the spec, and MSVC, don't, so use uint16_t instead.

Reviewed By: yfeldblum

Differential Revision: D4099285

fbshipit-source-id: cdf7ba14b6c00e1bcdfc3650b76e6e4a47a61929

Summary: Silence a couple of warnings that get very spammy for MSVC.

Reviewed By: yfeldblum

Differential Revision: D4098682

fbshipit-source-id: b410b56062cdf82367675ea9c3dd22975e7b91bd

Reviewed By: yfeldblum

Differential Revision: D4100221

fbshipit-source-id: 7778b48f1bba9be85f5712f61d6e1731524901c3

Summary: Instead of exposing raw jumpContext, it now exposes a higher-level FiberImpl class, which can be extended for newer versions of jump_fcontext.

Reviewed By: yfeldblum

Differential Revision: D4099849

fbshipit-source-id: 28c7ce32284a0109cf040c264d46a31a45867934

Summary: A cleanup in preparation to make folly::fibers work with newer boost.

Reviewed By: jsedgwick

Differential Revision: D4093947

fbshipit-source-id: e9a0aaeb915c40af96282169d4758afe572f2dd8

Summary:
The main purpose of this diff and this library at this point is to be a public reference for the C++ standard committee and whoever is interested in the proposal to the committee.

This diff aims to be consistent with the latest version of the proposal (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0233r2.pdf).

The current interface proposal focuses on the core components (domain, object base, and raii owner). Once, that part is settled we can add to the interface:
- Thread local options (thread caching of hazard pointers and private storage of retired objects)
- Programmer control of reclamation (when and by which threads)
Also, at this point the implementation does not optimize memory ordering.

I removed hazptr_domain::try_reclaim() from the public interface at this point. The latest update to interface aims to relieve the programmer from the need to take spacial precautions against shutdown fiascos involving reclamation functions of objects stored by the default domain. Please let me know if you have any concerns about this.

Having said that about the current purpose of this library, I really appreciate any comments (in this diff or separately) on the interface in general and any suggestions for the eventual optimized implementation.

Reviewed By: davidtgoldblatt

Differential Revision: D4104381

fbshipit-source-id: df98adf6fd9b7a93406cb8eeca8fe2ad12388139

Summary: #accept2ship

Reviewed By: jsedgwick

Differential Revision: D4105140

fbshipit-source-id: eca1f7066e514ec3c5d316f974edc34e03872784

Summary: Winsock defines the internals of `in6_addr` in its own way, so we have to account for that.

Reviewed By: jsedgwick

Differential Revision: D4099453

fbshipit-source-id: c0ebb4e2017f61bed7d5d63058161ef3f16f9a65

Summary: If multiple iovs are passed to sendmsg, sendmsg is supposed to return the number of bytes sent if it would block, but only if the number of bytes sent is greater than 0.

Reviewed By: yfeldblum

Differential Revision: D4099691

fbshipit-source-id: e58fa71604966129b1fbd418c24b1bf012060428

Summary:
Implement AtomicBatchDispatcher in folly::fibers
- Details about how to use the added functionality can be found
  in the doc comment for class AtomicBatchDispatcher.

Reviewed By: andriigrynenko

Differential Revision: D4054148

fbshipit-source-id: 090272eeab8c8abb15d5e400e52725853fcfc364

Summary: Previously we set the cork option on the socket before making multiple writes, and then unset it after, which elip found was hurting perf with 2 extra syscalls. The cork logic was also not the same as the buffer combining logic, which made us often set cork even when only doing one write.

Reviewed By: djwatson

Differential Revision: D4058357

fbshipit-source-id: 1a07447ff5e027751e455a2403e0042bf67cb1c5

Summary:
A common failure mode for folly::doNotOptimizeAway is to pass it a
pointer to a complex object that the caller wants to ensure is fully
calculated, which doesn't actually ensure that.  Also, the GCC and
MSVC implementations used inline assembly with an in-out parameter,
which causes a store.  A survey of many usages of doNotOptimizeAway
found only 1 that relies on this make-unpredictable behavior.

This diff makes doNotOptimizeAway(&x) equivalent to doNotOptimizeAway(x)
for GCC and clang, and makes it a read-only sink.  For the rare
case that the benchmark wants to disable subexpression elimination,
constant propagation, or power reduction, there is a new function
makeUnpredictable.  It also merges the clang and GCC implementations,
removing a potential bias in our microbenchmarks.

Reviewed By: davidtgoldblatt

Differential Revision: D4074670

fbshipit-source-id: 43f02e7fe149147bb172babe77787dea06e098fa

Summary: In this case, the SpookyHash tests, which call `clock_gettime`.

Reviewed By: yfeldblum

Differential Revision: D4098997

fbshipit-source-id: e402fd6626e59ccc555513384e93ef0ef891bbf3

Summary:
If Folly was compiled without support for SSE 4.2, but this function ends up getting called, we'd abort due to the check failing.
To solve this, just don't do the check and fall back to the non-sse version.

This is reasonable because it means that the code calling `qfind_first_byte_of_sse42` only needs to care whether there is CPU support for SSE 4.2, not whether we were able to actually implement the search with SSE 4.2 based on compiler support.

Reviewed By: yfeldblum

Differential Revision: D4096645

fbshipit-source-id: d26a791c09f61595e40c84c47a7a2596b9b69c1e

Summary:
When using `-fvisibility-inlines-hidden` or using C/C++ shared libs
transitively from a Python binary, static locals in inline functions
can potentially have more than one definition at runtime.  This diff
explicitly marks this static local as safe for this scenario.

Reviewed By: djwatson

Differential Revision: D4082831

fbshipit-source-id: 75616ac26a39f44691ef59b889ea555e9daa391b

Summary:
Now that the initializer list syntax has been removed we can add a default constructor.
Also,
- The `dynamic(T)` constructor was unconstrained, so it would match any type but then fail to compile (as a side effect, `is_convertible<T, dynamic>` would be always true). This also leaked the implementation details of `Array` and `Object`, as they were accepted as arguments. The diff makes the constructor accept only integral and float arguments, and all other types are SFINAEd out.
- `dynamic(Iterator, Iterator)` is made `explicit` to avoid accepting statements like `dynamic d = {"a", "b"};`.
- `object(...)` methods are simplified.

Reviewed By: luciang

Differential Revision: D4065021

fbshipit-source-id: ac289da7bece67c674b7036b7b51d5e016b297e5

Summary:
This refactors inline functions defined in headers which had static locals
so that their definition is moved to the corresponding implementation file.

Reviewed By: yfeldblum

Differential Revision: D4049175

fbshipit-source-id: 56eeb82eb23b04c3b9940d803d05050949aa5ef9

Summary: A bit more clear

Reviewed By: yfeldblum

Differential Revision: D4075931

fbshipit-source-id: 1398500336d412b5e7663ef1e67b20a329ce2979

Summary: Some compilers think result may be uninitialized

Reviewed By: francis-ma

Differential Revision: D4074526

fbshipit-source-id: 5df4568cfa5eb8f3b2993a82b831fd5e4028a281

Summary:
Consider `Expected<..., eUnion>(Expected&&)` ctor, where
`ExpectedUnion` move ctor would be called first, which is noop leaving
`which_`  uninitialized; then `MoveConstructible` ctor is executed, calling
`assignValue()`, which performs action depending on uninitialized `which_`.

Reviewed By: yfeldblum

Differential Revision: D4073199

fbshipit-source-id: 623660e7047afcebf9f72d83f91f84ff6078090f

Summary:
Add a prevPowTwo method to bits.h and optimize the current code for
GCCs output.

Reviewed By: ot

Differential Revision: D4072341

fbshipit-source-id: 6e949d0bfcf88ff8500022939d08a2b5aa9e00c9

Summary:
folly/experimental/symbolizer doesn't actually link with libdwarf, it only requires dwarf.h.
Closes https://github.com/facebook/folly/pull/501

Differential Revision: D4068438

Pulled By: djwatson

fbshipit-source-id: 6339eabf8fceac300caafb478c003833768125fd

Summary:
Add smart pointer types for ec groups
and points

Reviewed By: yfeldblum

Differential Revision: D4066848

fbshipit-source-id: f8a5c59cf902584e5a4bea7265834fcc8898677b

Summary: as above

Reviewed By: yfeldblum

Differential Revision: D4066250

fbshipit-source-id: 7dfe07656dd338ec42fcd966328e46331202bf58

Summary:
This diff adds a method to folly::Function that moves the
Function object into a copyable, callable proxy object.

folly::Function already has the asStdFunction method, which
does the same, but wraps the proxy object in a std::function.
When a copyable shared-state proxy of a folly::Function is needed,
it is not necessarily to turn it into a std::function.
If instead the shared proxy will be passed to several functions
taking folly::Function, what happens is the folly::Function
is moved on the heap, a shared_ptr is put into a std::function,
and that is wrapped in a new folly::Function, which always
requires a memory allocation, because std::function is not
noexcept-movable. When using asSharedProxy instead, an
unspecified copyable type is returned that can implicitly
converted into a folly::Function, but without an additional
memory allocation at this point.

Reviewed By: yfeldblum

Differential Revision: D4048621

fbshipit-source-id: b642027b0a6957058fe0089cceeb657ec52e8669

Summary: It has been deprecated for several months, time to nuke it.

Reviewed By: yfeldblum

Differential Revision: D4045250

fbshipit-source-id: 21b55a7c99602c42b1a87ced6a59788fe0ac25d7

Summary: Someone debugged a runtime crash and traced it back to code like: `vector<fbstring>{{"this", "that"}}`. Presumably the user thought they were passing an initializer list of strings to the vector constructor. Instead, they constructed a single fbstring with two char pointers pointing into //different// strings. With the appropriate fbstring constructors, we can flag this as invalid at compile-time.

Reviewed By: yfeldblum

Differential Revision: D3927397

fbshipit-source-id: ab61e1e8498ec99592a2a7726eaf1cb6324f1455

Summary:
improve documentation of custom singleton creation through an example
(from fbcode SIOF thread suggestion)

Reviewed By: yfeldblum

Differential Revision: D4053322

fbshipit-source-id: e9c2ef3d1ef43d52c0bf0a601d94c017047a23a3