Summary: [Folly] Tweaks to `AtomicStruct`.

Reviewed By: Orvid

Differential Revision: D6636432

fbshipit-source-id: 274c118a732287219c569d7d3f4e170f275518f1

Summary: Returning a SemiFuture from a continuation should work by correctly checking the types and returning a folly::Future on the same executor as the original future that .then was applied to.

Reviewed By: yfeldblum

Differential Revision: D6597273

fbshipit-source-id: cf2016a344d4b29f1d31c1da20c89df5b4cfe64e

Summary:
[Folly] Remove `folly/detail/UncaughtExceptionCounter.h`.

It's a thin and unnecessary shell around `folly/UncaughtExceptions.h`.

Reviewed By: Orvid

Differential Revision: D6636260

fbshipit-source-id: cdf6fa5fefc9fd69586c1c4c1a8443c5e8543b1c

Summary: [Folly] `constexpr_log2_ceil`, like `constexpr_log2` but rounding up.

Reviewed By: Orvid

Differential Revision: D6636433

fbshipit-source-id: a10f031cc9c91cfeba7b74bbf143895a311ca772

Summary: [Folly] Move `folly/detail/Sleeper.h` to `folly/synchronization/detail/`.

Reviewed By: Orvid

Differential Revision: D6636459

fbshipit-source-id: e0b37459fe721c96837b653e652c1bc6bfeb5dce

Summary:
Window tests were moved to a separate file. Update the README to the
new location.

Reviewed By: yfeldblum

Differential Revision: D6636178

fbshipit-source-id: 0313dfe80de7686e1e530ed9eb4bbcefb459633c

Summary: [Folly] Format `digits10`.

Reviewed By: Orvid

Differential Revision: D6636190

fbshipit-source-id: 6e6b834f6c9070d58f8e2b085b09df8b857fe878

Summary: [Folly] Move `folly/AtomicStruct.h` to `folly/synchronization/`.

Reviewed By: Orvid

Differential Revision: D6636072

fbshipit-source-id: 87098e25fc40d0d60b4a657ba395418388e170fc

Summary: [Folly] Remove unused `boost` dep from `Conv.h`.

Reviewed By: Orvid

Differential Revision: D6636040

fbshipit-source-id: 77dbfcf95f68c05346bce52d1a965e465c2efe3b

Summary:
[Folly] Update homebrew script to use gtest 1.8.0, to be consistent with the docs.

Closes #722.

Reviewed By: Orvid

Differential Revision: D6633606

fbshipit-source-id: 5e9a5790100fe7a408be7a9a77c847bece91d405

Summary:
[Folly] Move expensive `Singleton` code to `.cpp`.

Including string manipulation, `ostream::operator<<` operations, `throw` statements, etc,

Reviewed By: ot

Differential Revision: D6632052

fbshipit-source-id: 82e0033d5042b02951cf3b05c55cf62c97cc2b6d

Summary: Also deprecates folly::getFuture function.

Reviewed By: yfeldblum

Differential Revision: D6594299

fbshipit-source-id: 67435f35dca660da049cd8c505ee9a21424c0d2b

Summary: [Folly] Namespace portability symbols better in `Singleton.cpp`.

Reviewed By: Orvid

Differential Revision: D6631290

fbshipit-source-id: eb802ace6a9bf704e011eb44710922c997a9cd8e

Summary: [Folly] Fix the assumption in the `propagate_const` test about local layout.

Reviewed By: aary

Differential Revision: D6624962

fbshipit-source-id: 6871dc20f2b5ec2d1a86f520030c7cda8a225958

Summary: As title

Reviewed By: yfeldblum

Differential Revision: D6620395

fbshipit-source-id: 477aae84b2bbde2e79d46ae93c285909b56f575e

Summary:
There are cases where ELF binaries are running fine, but have slightly
corrupted ELF headers, e.g., with section headers offset pointing beyond boundaries
of file. I'm guessing this is due to running strip or objdump with either some
particular combination of flags or some due to buggy version of those tools. This
change prevents from crashing on such files.

Reviewed By: yfeldblum

Differential Revision: D6616027

fbshipit-source-id: 8cb3ac4475a51d1f1045c395977a6a77fbefffb2

Summary:
It was not clear to me, if a callback takes no arguments, the
underlying value is guaranteed to be alive during the execution of the
callback, so I wrote these tests.

Reviewed By: yfeldblum

Differential Revision: D6594921

fbshipit-source-id: 6a658afc1bf4d29eaa9c62269ddc21c7f897ad01

Summary:
[Folly] Kill `FOLLY_ALIGNED` etc.

`alignas` is standardized as of C++11. Let us just use that.

Replace:
* `FOLLY_ALIGNED` with `alignas`
* `FOLLY_ALIGNED_MAX` with `alignas(folly::max_align_v)`
* `FOLLY_ALIGN_TO_AVOID_FALSE_SHARING` with `alignas(folly::hardware_destructive_interference_size)`

Because where `alignas` may be placed is more restrictive than where attributes may be placed, we also need to move these directives in some cases on top of doing the replacement.

Reviewed By: Orvid

Differential Revision: D6555167

fbshipit-source-id: 4b05b570bace3f8c0fe810b6dd58781dd45757f4

Summary: [Folly] Fix Build: `propagate_const` under GCC.

Reviewed By: aary

Differential Revision: D6609722

fbshipit-source-id: 47d8b0f2d54e0e5834383ea64c93af30927c921d

Summary: Add readWithPriority to PriorityMPMCQueue

Reviewed By: yfeldblum

Differential Revision: D6606604

fbshipit-source-id: 07382e3c6d0f53a8727c059410c0577cf0aac82c

Summary:
[Folly] Add `hardware_destructive_interference_size` and `hardware_constructive_interference_size` to `folly/lang/Align.h`.

As backports from C++17. Which may require keeping, depending on how stanard libraries choose to implement these.

And replace `CacheLocality::kFalseSharingRange` with `hardware_destructive_interference_size`.

Reviewed By: ot

Differential Revision: D6554817

fbshipit-source-id: bff49f5ca8b01d38fa806076f99201355df76cd9

Summary: [Folly] `propagate_const`, backported from C++ Library Fundamentals TS v2.

Reviewed By: ericniebler

Differential Revision: D6589681

fbshipit-source-id: cdc8981d17938b99afe60e2baefff7deb5316612

Summary:
Previously D6579707.

Correctly advance to next item if we erase the current element.  Corner cases were slightly off if we were at the end of a hash chain.

Reviewed By: yfeldblum

Differential Revision: D6603518

fbshipit-source-id: acb959e5bcd5da1c3df642b75985d464fdd3b23d

Summary:
I misread the ThreadLocal docs, thread destruction functions do *not* grab the accessAllTHreads_ lock
unless you use *strict* mode, and even then, it is only a read lock.

Easy enough to make the thread-destruction global bits to be atomic / use folly::Synchronized.

Reviewed By: yfeldblum

Differential Revision: D6592905

fbshipit-source-id: 4ae600dff4c8c04751483a452ca7c07ef3f26380

Summary: [Folly] Switch the `Baton` template params for consistency with `SaturatingSemaphore`.

Reviewed By: davidtgoldblatt

Differential Revision: D6591060

fbshipit-source-id: 44d6243d3185d95364a27e497216cc02fb3bc2e8

Summary: Adding secure RNG utility functions (rand32, rand64, oneIn, randDouble01, randDouble).

Reviewed By: yfeldblum

Differential Revision: D6551975

fbshipit-source-id: 720d138de1329669b1a15eb3e9cb3fe91ce982a4

Summary: [Folly] `Baton::ready`, a `const` variant of `try_wait`.

Reviewed By: djwatson

Differential Revision: D6508064

fbshipit-source-id: ba458577574ba58165408a93238da7eb09adf1e6

Summary:
The `get*Executor()` APIs don't check whether the singletons
haven't been destroyed already. Add a check and allow to return
`nullptr` during shutdown.

Also do a general clean up of the code, there was no reason to use
three independent singletons (non-atomically destroyed) for each
executor.

Reviewed By: philippv, luciang

Differential Revision: D6589486

fbshipit-source-id: 20fb835db7e446bd811bbd6d5ddbc41db9e98b54

Summary: The shim existed so that hhvm-third-party could be updated, and it has been, so it's time to remove the shim.

Reviewed By: yfeldblum

Differential Revision: D6590439

fbshipit-source-id: 3e1b039c1df6cf1e74102f7b7b26aa90fae31ec0

Summary: Test UnboundedBlockingQueue for the case in D3527722

Reviewed By: ot

Differential Revision: D6587163

fbshipit-source-id: 1465991018187c5f841e6e3e7a11676390e2f8f2

Summary:
The current semantics of try_dequeue variants can cause problems in cases where based on external dependencies the consumer is guaranteed that the queue is not empty and depends on the success of try_dequeue. See https://github.com/facebook/wangle/commit/abd07a06cae09efcd4347a9a019d59ae8a013c3f

This diff changes the semantics so that try_dequeue operations fail only if the queue is empty.

Reviewed By: yfeldblum

Differential Revision: D6586156

fbshipit-source-id: 25d6085e28d3e24034ecf6a8bafab51c95464b01

Summary:
Without the fix multiple consumers may update head concurrently and cause it to lag. If this persists until the destruction time of the queue, some segments may be incorrectly retired twice.

The fix is to wait for head to advance to the current segment first before advancing head to the next segment.

Reviewed By: djwatson

Differential Revision: D6588135

fbshipit-source-id: 3e916441bff5ad3f27de418601990c59a0b89bc2

Summary:
[Folly] Group the fields of `UnboundedQueue`.

Into helper structures which are organized by, and aligned by, access.

This is all compiled away and has no runtime cost.

Reviewed By: magedm

Differential Revision: D6555031

fbshipit-source-id: 94424747bf76a0a9a2bd59b6a94749efb391d6a7

Summary: Fix typos and wording in comments.

Reviewed By: yfeldblum

Differential Revision: D6584559

fbshipit-source-id: b4b72f6c25c9f9e6fb3f2fbee69eea4a39b5c004

Summary:
This reverts commit a45f47a53e106d22daa9cf57be6c40c4f6a430d9

bypass-lint

Differential Revision: D6579707

fbshipit-source-id: 6f6c35de5e05f8cf37491471e1ae1272290fe61b

Summary:
Found that building folly on aarch64 gave errors with
picospinlock.  Looked and found that MSVC used casts
so I figured I'd try that

Reviewed By: yfeldblum

Differential Revision: D6504689

fbshipit-source-id: 9565bae5ffab485da407b8609be92ef7db10ab72

Summary:
The iterator returned from erase api could skip nodes. The fix is to
initialize the returned iterator with right value.

Reviewed By: djwatson

Differential Revision: D6579707

fbshipit-source-id: a45f47a53e106d22daa9cf57be6c40c4f6a430d9

Summary:
[Folly] Fix `folly::max_align_v` for Clang on ARMv7.

There is some problem with computing the correct result of `alignof` for a compound `union` POD type, because why wouldn't there be. The result *should* just be the max of the `alignof` of the type of each field if each field is default aligned - but not on some platforms! So we must compute the max directly.

Reviewed By: mzlee, Orvid

Differential Revision: D6573548

fbshipit-source-id: 512a255fda64795104d71fde14372befa3bf41e4

Summary: [Folly] Let `Baton` methods be `noexcept`.

Reviewed By: djwatson

Differential Revision: D6508057

fbshipit-source-id: 2138a228d8291e79e9368cb77b0f63f4412524dc

Summary:
We currently leak ptr entries due to a missing map erase call in AsyncSocket::releaseZeroCopyBuf.
Also addZeroCopyBuf called when the bytesWritten == 0 will cause and extra id to be allocated. This will delay the deletion of the last IOBuf sent via zerocopy. The number of buffers can accumulate over time if the same bytesWritten == 0 situation is encountered.
Found when running the nbd-server zero downtime upgrade during heavy fio traffic.
Add an AsyncSocket::processZeroCopyWriteInProgress so we can drain the zerocopy msg error queue even after we set the ReadCallback to nullptr.

Reviewed By: djwatson

Differential Revision: D6552982

fbshipit-source-id: 3d2fdca84ec3b5fc46c3bed06c0c9ede66ed565a