Summary:
The layout of an exception_ptr is determined by the ABI, not the OS. MinGW (and
Cygwin, depending on the build settings) define `_WIN32` and would therefore
have `kIsWindows` be true, but they use the Itanium ABI instead of Microsoft's,
so we shouldn't take the MSVC codepath for them.

Unfortunately, there isn't a reliable preprocessor method to determine the ABI
being used. `_MSC_VER` can be defined even when using the Itanium ABI (e.g.
when using windows-itanium triples with clang), and `_MSC_VER` can be undefined
even when using the Microsoft ABI (e.g. when building with `-fno-ms-extensions`
with clang).

To address this, introduce `kMicrosoftAbiVer`, which defaults to `_MSC_VER`,
but can be overridden manually if necessary. This should do the right thing by
default most of the time, while providing an escape hatch for when `_MSC_VER`
would be incorrect to use directly.

Reviewed By: yfeldblum

Differential Revision: D7377853

fbshipit-source-id: dd868370e9a496b98dd26dec77c4a75c49147c3c

Summary:
Currently, `collectN` attaches a callback to each input future. When the callback fires, it calls `emplace_back` on vector `ctx->v` with the result of the future. As the added unit test shows, this does not work if the futures are completed in parallel.

See https://fb.facebook.com/groups/474291069286180/permalink/1838209692894304/ for more discussion.

Reviewed By: yfeldblum

Differential Revision: D7372982

fbshipit-source-id: 875eff19172bdbfb7d4f687911f4fa22061e4ed8

Summary:
`std::tuple` and `std::unordered_map` have long been standardized and supported by every platform we care about.
Beyond that MSVC is starting to remove the tr1 namespace entirely, so these references don't work under the newest MSVC.

Reviewed By: yfeldblum

Differential Revision: D7389783

fbshipit-source-id: c089490360ebf15ac23925f71e991602fda3d7a0

Summary: The nightlies for MSVC included support for std::launder, but the version that released with VS2017.7 Preview 2 doesn't actually support it. The version included in the preview also adds a macro to detect if it's been enabled, so use that.

Reviewed By: yfeldblum

Differential Revision: D7388311

fbshipit-source-id: 197230c9947cb271d6107d0cd5e905643e5589d1

Summary:
MSVC moves faster than boost updates, so it's not unusual to be building with an MSVC newer than the boost being used was able to be tested against. This causes Boost to emit a message at compile time for every compilation unit it is included in.
Luckly, boost provides a way out of the message, so add a config option to do so and enable it by default.

Reviewed By: yfeldblum

Differential Revision: D7388583

fbshipit-source-id: 8ecddc048df432df5200f4601e4121bdbd275220

Summary:
[Folly] Use C++ `thread_local` for fast-path cache in `SingletonThreadLocal`.

Because modules might be `dlclose`'d, use `static thread_local` rather than `static __thread` in the slow path which integrates the dtor into both `dlclose` and thread exit.

Because there might be multiple dynamically-loaded modules where the `static __thread` caches cannot be merged into a single cache, let all caches be tracked in a list. We have not found a case where this can happen while at the same time the `StaticSingletonManager` singleton does get merged, but this is at least not strictly wrong.

Reviewed By: djwatson

Differential Revision: D7316155

fbshipit-source-id: 0bf06638853ed70464856853349445d9fa6eabc0

Summary: Currently, initializing a `ThreadLocal` with 0 silently passes and then fails at a later stage making it very hard to debug what went wrong

Reviewed By: yfeldblum

Differential Revision: D7326627

fbshipit-source-id: dab62166753168e292f67938c52f82a1b610be2a

Summary:
Latest clang reported `-Wextra-semi` in `BOOST_MPL_HAS_XXX_TRAIT_DEF` uses with the trailing `;`. Codemod them away before the upgrade.
```lang=bash
fbgr -sl 'BOOST_MPL_HAS_XXX_TRAIT_DEF\(.*\);$' | xargs perl -pi -e 's,(BOOST_MPL_HAS_XXX_TRAIT_DEF\(.*\));$,\1,'
```

Reviewed By: yfeldblum

Differential Revision: D7383619

fbshipit-source-id: f7f9415617cf6d4a876ebe8bf782d0115cd01c23

Summary:
Compilation of basic sample from doc using VS2017 currently fails with:
```
C2327 'folly::ProducerConsumerQueue<folly::fbstring>::writeIndex_': is not a type name, static, or enumerator
```
Fixed it and updated the docs.
Closes https://github.com/facebook/folly/pull/806

Reviewed By: akrieger

Differential Revision: D7379931

Pulled By: yfeldblum

fbshipit-source-id: 6b7a72abd3324af56de355a6e65bdc7e1195596d

Summary:
Macro BENCHMARK_DRAW_LINE can be used as the following:
```lang=bash
BENCHMARK_DRAW_LINE()
BENCHMARK_DRAW_LINE();
```
This inconsistency was detected by `-Wextra-semi` in the latest clang. This diff removes the trailing `;` inside `BENCHMARK_DRAW_LINE` and updates all of the references.

```lang=bash
fbgr -sl 'BENCHMARK_DRAW_LINE\(\)' | xargs perl -pi -e 's,(?<!#define )BENCHMARK_DRAW_LINE\(\)(?!;),BENCHMARK_DRAW_LINE\(\);,'
```

Reviewed By: yfeldblum

Differential Revision: D7382525

fbshipit-source-id: 76f1da01bb4dde1b0fa4ebf98748d001d8e04ab0

Summary: As title

Reviewed By: yfeldblum

Differential Revision: D7379148

fbshipit-source-id: 6865940adee5899a055e1dac6b23705db2bc8ed7

Summary:
`iterator F14VectorMap::erase(const_iterator, const_iterator)` was missing a
conversion from `const_iterator` to `iterator`. This diff adds it as well as a
test. For sets, `iterator == const_iterator` so no need for anything special.

Perhaps we could save the roundtrip from iterator space to index space and
back using some const_cast of fancy pointers? The cost is going to be amortized
anyway...

Reviewed By: nbronson

Differential Revision: D7373152

fbshipit-source-id: 91f651b3e4a1314c68fb8809baeb1477f8881790

Summary:
[Folly] Gate `hazptr` `static_assert` on `_CPPLIB_VER` vs `_MSC_VER`.

It's a property of the standard library implementation, not a property of the compiler front-end.

Reviewed By: smeenai, Orvid

Differential Revision: D7378222

fbshipit-source-id: 78d04252df751b683c0099cb0c273302ff388b73

Summary:
It's easy enough to calculate the right offset for both 32 and 64-bit
platforms, so we can just do so instead of hardcoding the 64-bit offset.

Reviewed By: yfeldblum, Orvid

Differential Revision: D7377039

fbshipit-source-id: 5f7efbae58b7694801274e195a97b5e5b62b5a5e

Summary: in response to https://github.com/facebook/openr/issues/23

Reviewed By: snarkmaster

Differential Revision: D7342371

fbshipit-source-id: 15737f79900722ce016494f994932aa26e5acad5

Summary:
[Folly] Enable `-Wshadow` for clang and gcc >= 5.

And disable it globally for gcc < 5 (sorry, but not sorry). The gcc < 5 implementation is far too restrictive.

Reviewed By: meyering

Differential Revision: D7320760

fbshipit-source-id: 49cc971afcacc92b0bb2d9c4b2d7b04405fb95b3

Summary:
Original commit changeset: 4f59f6c3c8aa

This is to address segfaulting in instagram server tests

Reviewed By: fried

Differential Revision: D7368716

fbshipit-source-id: ef28fdc17fb72631550d81616c24498a3b3894a1

Summary:
This diff adds a header file which includes forward declaration of F14
containers.

Reviewed By: nbronson

Differential Revision: D7342805

fbshipit-source-id: 01e867792d8bde3769f4c71dfb6b80574b13f2b6

Summary:
* make F14 containers their own types even on platforms where we fall back to
stl containers.
* extract F14FastMap to be the same conditional typedef.
* define custom swap operators to avoid unnecessary move construct and move
  assignment.

Reviewed By: nbronson

Differential Revision: D7295846

fbshipit-source-id: 92faa2feddc5c85dd17a5dd937004ca1839a09f6

Summary:
Update the LogName code to also treat `/` and `\\` as category separators, just
like `.`.

This makes it easy to use directory and file names directly as log category
names.  The XLOG() code already turns file names in to log category names.
Previously `getXlogCategoryNameForFile()` had to do its own translation of
directory separators into `.`.  Users manually had to perform this translation
on their own when setting the configuration for specific categories.

This change makes it possible for users to specify the filename as input
directly when updating the logging configuration and it automatically updates
the correct category.

This also allows us to avoid allocating a new string in
`getXlogCategoryNameForFile()` since we no longer have to translate directory
separator characters there.

Reviewed By: chadaustin

Differential Revision: D7348520

fbshipit-source-id: 0e996f4351798b4636786a14ebb2d8820a2cc153

Summary:
retryingPolicyCappedJitteredExponentialBackoff had an integer overflow.
This occured usually when we had more than 64 retries, since we compute pow(2, n) and then we convert to a 64 bit integer.
I added a check that it won't overflow, and if it does, then it returns the max_backoff value.

Reviewed By: yfeldblum

Differential Revision: D7284952

fbshipit-source-id: 7424f502b8530f8e39ba0513dd70bfc9fdc5c8e5

Summary:
Revert NotificationQueue changes until the thread local changes land

(Note: this ignores all push blocking failures!)

Reviewed By: yfeldblum

Differential Revision: D7359322

fbshipit-source-id: c59c81c34b341e4a968b2ce14947b558b8db3a85

Summary: This means using "test" rather than "tests" for the name of the test directory.

Reviewed By: wqfish

Differential Revision: D7355701

fbshipit-source-id: 262f746a983eef550dd9271a026e0d243d8a335d

Summary:
Openr is only doing travis ci for ubuntu 16.04 with gcc 5 so having
the example with the older flavors can be misleading for our community.

Reviewed By: saifhhasan

Differential Revision: D7345130

fbshipit-source-id: 684cc8e58a650050b6f63c2ec20aac1857476da7

Summary:
Python is considered a Safe Enivronment, where things like aborts and segfaults
can not happen.

If clients where used outside of "async with" we use to print a error message
and then C++ would abort.

This fix makes sure C++ no longer aborts but we still print the message
that the "async with" should be used.

Reviewed By: ambv

Differential Revision: D6923956

fbshipit-source-id: 4f59f6c3c8aa5c8b2c279c728a8c011a040940af

Summary:
Move the helper classes into F14TestUtil header so that F14SetTest could use
them as well.

Reviewed By: yfeldblum

Differential Revision: D7336571

fbshipit-source-id: 9fc96e43fe882d05b6b6fbcd075a42f5184f7d62

Summary:
std::system_error exceptions generated from an errno value should use
`std::generic_category`.  Previously EXPECT_THROW_ERRNO() only handled
exceptions using `std::system_category` since `folly/Exception.h` (incorrectly)
throws exceptions using `std::system_category`

This updates EXPECT_THROW_ERRNO() to accept either std::generic_category or
std::system_category.

In the future we can drop the check for std::system_category here once we fix
folly/Exception.h to throw exceptions using std::generic_category.

Reviewed By: yfeldblum

Differential Revision: D7329998

fbshipit-source-id: 89295d7c8276c9119c29325bf35c0bf27452b779

Summary:
This diff fixes a bug in get_client flow where we accepted hostnames but didn't perform DNS resolution on them.  That has been fixed by constructing the SocketAddress passing true to allowNameLookup.

While writing a unittest for this I realized that we didn't support binding servers to localhost, so I allowed for doing so.

Then I realized that it would be pretty simple to extend both get_client and ThriftServer to allow for unix sockets.

Discovered that unittests waiting for server.get_address() could wait forever if there was some exception in server.serve(). So I propigated the exception to the get_address future if it is not .done().

I updated all callsites that could accept a hostname/ip to accept and ipaddress object.
I updated path to accept a str or a Pathlike to support pathlib.Path.

Reviewed By: ambv

Differential Revision: D6911229

fbshipit-source-id: 2919a720da6bfa79db85fe55de10e8519ea39c82

Summary:
[Folly] Let `FileUtil::writeFileAtomic` use `generic_category`.

`generic_category` is the right choice overall.

And this is an excuse to decouple `FileUtil` from `Conv`.

Reviewed By: simpkins

Differential Revision: D7314184

fbshipit-source-id: fbec3e71098e3ca41b202964886b2740ca592229

Summary: [Folly] Extract a helper alias for ctor SFINAE in `SingletonThreadLocal`.

Reviewed By: djwatson

Differential Revision: D7316150

fbshipit-source-id: 7e3c49740dab3af37afa8a7a23f4e07cd93ff217

Summary: [Folly] Rearrange `SingletonThreadLocal` members, moving member types above member functions.

Reviewed By: djwatson

Differential Revision: D7316143

fbshipit-source-id: c70b31935d9b5d7428bde777ea62832dc45b3971

Summary:
[Folly] Split `getWrapperTL` from `getWrapper` in `SingletonThreadLocal`.

Allows hypothetical future changes to operate on the `ThreadLocal<Wrapper>` singleton itself, in addition to the existing code which operates on the `Wrapper` per-thread-singleton.

Reviewed By: djwatson

Differential Revision: D7316099

fbshipit-source-id: 80a8b7ce28da1bc64b204292032766c648f15944

Summary: [Folly] Cut unnecessary `library-local` comment in `SingletonThreadLocal`.

Reviewed By: djwatson

Differential Revision: D7316088

fbshipit-source-id: 08b8634b2f0d32776d548115109bc65732e4c880

Summary:
[Folly] Cut `SingletonThreadLocal::getWrapperInline`.

It would not be useful in any scenario, given that the generated code is very large.

Reviewed By: djwatson

Differential Revision: D7315501

fbshipit-source-id: 1ebc15ca0eaf13adaf52fae3971e7f9914bfe10e

Reviewed By: yfeldblum

Differential Revision: D7272553

fbshipit-source-id: f0aaa278e6a81a53f78ae29b68f1bf9f1ba02899

Use unbounded queue in NotificationQueue (same as in D7164130 but with the hazptr fix from D7256905 and the singleton one from D7283390)

Summary: Use unbounded queue in NotificationQueue (same as in D7164130 but with the hazptr fix from D7256905 and the singleton one from D7283390)

Reviewed By: yfeldblum

Differential Revision: D7278275

fbshipit-source-id: 78ea94e1c9d492febd33106b06ce275d4a90a56a

Summary:
F14VectorMap and F14VectorSet use ItemIter to interact with F14Table,
but they don't actually need to be able to identify the begin ItemIter.
This diff removes both the storage for packedBegin_ and the code to
update it (present in all insert and erase functions) for that policy.
This is a small CPU optimization, and reduces sizeof(F14Vector{Map,Set})
from 32 bytes to 24.

Reviewed By: shixiao

Differential Revision: D7181769

fbshipit-source-id: 318c63584543fca339a002d8bcb5872a7bd0bdd3

Summary:
Fiber suspends are incompatible with C++ exception handling. We should enforce that callers are not suspending fibers while handling exceptions.
See https://fb.facebook.com/groups/560979627394613/permalink/1119586164867287/

NOTE: This check currently fires not just for baton.wait() (and all things based on baton), but also for calling runInMainContext. I'm unsure of the safety of this.

Reviewed By: yfeldblum

Differential Revision: D6695110

fbshipit-source-id: cf36edbab9f0573cbf639bf79d0c267681ac5914

Summary: Rename set/getServiceIdentity to set/getSessionKey.  This is a more representative description since verified information is in the cert.

Reviewed By: yfeldblum

Differential Revision: D7184169

fbshipit-source-id: c9404419cb37d2b5ab2fcd8299975e514e3f495a

Summary:
[Folly] Move `folly/UncaughtExceptions.h` to `folly/lang/`.

(Note: this ignores all push blocking failures!)

Reviewed By: andrewjcg

Differential Revision: D7319259

fbshipit-source-id: 38d47503e150e01b5e4bcded6c8207484048bf84