Summary: Noticed because of a warning emitted by MSVC, this fixees the literal for `kLockedPtr_` to correctly be 64-bits under MSVC, rather than 32-bit.
Closes #245

Reviewed By: @yfeldblum

Differential Revision: D2282806

Pulled By: @sgolemon

Summary: MSVC defaults enum sizes to int, so explicitly use `size_t` as the base to have it correctly sized.
Closes #246

Reviewed By: @yfeldblum

Differential Revision: D2282816

Pulled By: @sgolemon

Summary: fix a use after free bug. this line is stupid!

Reviewed By: @siyengar

Differential Revision: D2283041

Summary: she's gooooooone, oh why, oh why, i better learn how to face it...

https://www.youtube.com/watch?v=bnVXIUyshng

Reviewed By: @djwatson

Differential Revision: D2219135

Summary: We have this previously-unused "NOT STARTED" status, which I recently appropriated to denote moved-out `ProcessReturnCode`s.

It's natural to also use this for default-constructed `ProcessReturnCodes`. Lacking a default constructor leads to a bunch of unnecessarily annoying use of `folly::Optional` in my upcoming diff, so I wanted to get rid of that, see e.g.

differential/diff/7657906/

Reviewed By: @tudor

Differential Revision: D2097368

Summary: Using `checkUnixError` after `waitpid()` is confusing and useless, because the only two possible errors are `ECHILD` (some other part of the program waited for this process!?) and `EINVAL` (invalid options, which are hardcoded in both `Subprocess::wait()` and `poll()`). Neither of these are recoverable. Moreover, even if the caller catches the exception, `~Subprocess` is still booby-trapped to `abort()` since its status remains `RUNNING`.

In short, just abort.

Reviewed By: @yfeldblum

Differential Revision: D2079415

Summary: Remove a couple of unused variables, and move one
that was only used inside an ifdef

Reviewed By: @yfeldblum

Differential Revision: D2279989

Summary: IndexedMemPool had one bare usage of std::atomic, rather than
the templated type Atom.  This doesn't affect any non-testing template
instantiations, because those two are usually synonyms, but it could cause
spurious failures of DeterministicSchedule tests.  Found via inspection,
not via failed tests.

Reviewed By: @yfeldblum

Differential Revision: D2277424

Summary: As titled
Trying to troubleshoot shutdown problems in HHVM and I see:
  EventBase::runAfterDelay() callback threw St13runtime_error exception: Raw pointer to a singleton requested after its destruction.
Would be much easier to troubleshoot if I know what is it being requested.

Reviewed By: @chipturner

Differential Revision: D2274497

Summary: This diff isolates the CacheLocality back end for DeterministicSchedule
from one invocation to the next (the last one was deterministic
across the entire program but not per-DeterministicSchedule instance),
and makes it easy to do extensive tracing of memory accesses during
a deterministic test.  These changes were made while tracking down a
bug, but don't fix any bugs on their own (and in fact don't affect any
production code at all).

Reviewed By: @yfeldblum

Differential Revision: D1842390

Summary: Add program_options to fix oss build

Reviewed By: @tudor

Differential Revision: D2274863

Summary: This diff is not yet complete, I want to see the contbuild before I change the
functions that return references to member functions.

It is unsafe to return a pointer when the dynamic object is a rvalue, because if
the pointer escapes the expression after the object is destroyed, we go into
segfault / undefined behavior land.
I have deleted these overloads.  The amount of valid code that is now disallowed
is minimal.  The only valid case I can think of is returing a pointer and
passing it to a function in the same expression that does not save the pointer.
However, this case is also dangerous, because if the function you pass it to
decides to save the pointer for later, we are in trouble, e.g.

  save_ptr(dynamic("str").c_str())

Since there are simple workarounds (naming the object), I think that is a small
price to pay for the greatly increased safety.

The next step is to overload all members that return a reference to a member
to move the member out if the dynamic is a rvalue:

  const dynamic& at(dynamic const&) const&;
        dynamic& at(dynamic const&)      &;
        dynamic  at(dynamic const&)      &&;  // Move out

I also need to go over the code more carefully to make sure that nothing went
wrong.

Reviewed By: @marcinpe

Differential Revision: D2257914

Summary: Many command line apps are of the form
"program [--global_options] command [--command_options] args..."

Make writing such things less painful.

+jdelong because smcc

Reviewed By: @meyering

Differential Revision: D2217248

Summary: GFlags is useful for global program options. boost::program_options makes it
easier to write command-line utilities. They don't, unfortunately, play
very nicely with each other.

Add a addGFlags() function to convert all GFlags to boost::program_options
options; you can then use boost::program_options::parse_command_line() to
parse all arguments, GFlags or not.

Also add a helper function to make parsing nested command lines easier.

Reviewed By: @fugalh

Differential Revision: D2215285

Summary: SharedMutex purposely allows the inline reader count to underflow in some
situations while preserving proper locking behavior (the inline reader
count is only trusted if all deferred locks have been applied), but there
was an additional way that this could occur that wasn't documented or
allowed by the asserts.  The effect was a false positive assert in rare
conditions or the possibility of losing track of a deferred lock slot.
This diff fixes both the over-aggressive assert and the potential loss
of the shared slot.  If the assert didn't fire for you then this diff
won't change the correctness of your program.

Reviewed By: @yfeldblum

Differential Revision: D2269018

Summary: Because I didn't want to write it again.

Reviewed By: @yfeldblum

Differential Revision: D2217246

Summary: fixing compiler errors when compiling with Wunused-parameter

Reviewed By: @yfeldblum

Differential Revision: D2267014

Summary: Thanks @​peijinz for finding this.

Reviewed By: @​peijinz

Differential Revision: D2268532

Summary: `get_ptr()` checks if the index is past the end of the array, but it
doesn't check if it is less than 0, while `at()` does.

Reviewed By: @yfeldblum

Differential Revision: D2258548

Summary: A timeout now raises `TimedOut()` on the current Future.

Reviewed By: @fugalh

Differential Revision: D2261559

Summary: This fixes github issue https://github.com/facebook/folly/issues/243

I don't understand why this is broken on osx, but the solution is to just `#include <thread>` in the header instead of trying to forward declare `std::thread`.

Reviewed By: @yfeldblum

Differential Revision: D2255026

Summary: fixing C++ compiler warnings for the -Wunused-parameter warning switch on.
This is fixing only the header files, not the library cpp files.

Reviewed By: @yfeldblum

Differential Revision: D2259957

Summary: This is causing a log of spew in proxygen logs

Reviewed By: @afrind

Differential Revision: D2262006

Summary: Forces consumers to handle the empty sequence case instead of cumbersome exceptions.

Reviewed By: @​jrichey, @yfeldblum

Differential Revision: D2219505

Summary: [Folly] Fix a namespace nit in folly/futures/Future.cpp.

As soon as someone introduces a namespace `folly::futures::detail` this will blow up. Fix it so it won't blow up when that happens.

Reviewed By: @​hannesr

Differential Revision: D2258799

Summary: Allow access to a FiberManager's currently running fiber

Reviewed By: @alikhtarov, @andriigrynenko

Differential Revision: D2257201

Summary: `folly::Optional` returns the stored value by reference when the object
is an rvalue.  This causes three issues:

  * If the user calls `value()` on an rvalue `Optional`, and assigns the result
    to a new variable, the copy constructor gets called, instead of the move
    constructor.  This causes the added test `value_move` to not compile.
  * If the user calls `value()` on an rvalue `Optional`, and assigns the result
    to a const lvalue reference, they might expect the lifetime to be extended
    when it isn't. See the added test `value_life_extention`.
  * Assigning the results of `value()` on an rvalue `Optional` to a mutable
    lvalue reference compiled in the old code, when it shouldn't, because that
    is always a dangling reference as far as I can tell.

I'm not sure how strict `folly` is with compatibility, but I believe the
breakage would be minimal, and any code that gets broken probably deserves it.

I'm not exactly sure who I should add as a reviewer, so hopefully herald has got
my back.

Reviewed By: @yfeldblum

Differential Revision: D2249548

Summary: The creator thread ID is saved to indicate the singleton is already being built (to help detect dependency cycles).  However if the creation function throws an error, that thread ID persists, and then if the same thread tries again to build the singleton it will be falsely detected as a dependency cycle.  This clears that state in the case of failure.

Reviewed By: @chipturner

Differential Revision: D2256441

Summary: This reverts commit 956351018a495af89575526536af7e7c0bb285aa.

Reviewed By: @​labrams

Differential Revision: D2258219

Summary: In nested loop, loop condition is incorrect.  Fixed var (should be `ii` not `i`)

Reviewed By: @pgriess

Differential Revision: D2255702

Summary: I'm upgrading assertions to throws, since these are fatal in all circumstances. If something is explicitly `Optional`, it makes sense to fail loudly if it is misused in this manner.

Reviewed By: @yfeldblum

Differential Revision: D2247612

Summary: to unbreak OSS wangle tests build

Reviewed By: @bugok

Differential Revision: D2240264

Summary: What it says on the tin.

Reviewed By: @fugalh

Differential Revision: D2230390

Summary: guard pages may make debugging more painful. For example, in some cases they increased "perf" initialization time.

Reviewed By: @alikhtarov

Differential Revision: D2247188

Summary: I figured it out. This works. The two extra futures are a small overhead
(just two pointers). The `Core`s are allocated anyway, so this is pretty
much optimal.

A timeout now raises on the current Future, and a fulfilled promise
cancels the timeout Future.

Reviewed By: @yfeldblum

Differential Revision: D2232463

Summary: As the comment says, we can ignore errors in setting the size.
So it stands to reason we can ignore setting the size as well.

Reviewed By: @yfeldblum

Differential Revision: D2242882

Summary: There are more use cases that the Destruction/Guard pattern can be
used than current DelayedDestruction provides. This diff makes the pattern more
general (remove self destruct) and lets DelayedDestruction derive from that.
The functionalities of DelayedDestruction is kept unchanged.

I leave destroy(), Destructor class, and destroyPending_ in DelayedDestruction
since they are not required by our CallbackGuard in proxygen.

I add a shouldDestruct() function to allow customized conditions on when to
call destructor.

I haven't made destroyNow() a std::function since I only need it to be set at
instatiation time. If there is any other use case that needs destroyNow() to be
a std::function, I can do that as well.

Reviewed By: @afrind

Differential Revision: D2202575

Summary: Before we didn't do anything at all with the ::infinite value for our
generators, now all the sources operators and sinks are properly notated. The
one signifcant change regarding what is allowed concerns 'cycle' which now (when
called with no arguments) always produces an infinite generator. This is fine
for all but the weirdest of generators fed into cycle.

Reviewed By: @ddrcoder

Differential Revision: D2240328