Summary:
Array of spinlocks where each one is padded to prevent false
sharing.

Useful for shard-based locking implementations in environments where
contention is unlikely.

Test Plan: ran tests in 'common'

Reviewed By: soren@fb.com

FB internal diff: D582149

Summary:
In https://phabricator.fb.com/D511928 Brian mentioned the current API for string append is insufficient for appending to a buffer. That made me curious about the relative performance of classic and table-based number to ASCII conversions.

The results were interesting as on the average (over all digit lengths) the table-based conversion was faster, but performance was lackluster (in the worst case half as fast as the classic implementation) for large numbers, I presume due to the cache misses incurred by the tables.

This diff proposes an improved unsigned-to-ASCII primitive that is much faster than both table-based (previous Folly) and classic primitive. The key is a fast digits10() implementation that precomputes the space required by the conversion. After that, the digits are issued in place, no more reverse required. The new routine is up to 14x faster than the classic implementation, depending on the number of digits (benchmarks in comments).

Adding a few people who may be interested in the matter. Brian, thanks for bringing this matter up; if this gets in you may want to use the folly routine in proxygen.

Test Plan: unittest and benchmarks.

Reviewed By: simpkins@fb.com

FB internal diff: D515572

Summary:
The following asserts:
fbstring str(1337, 'f');
fbstring cp = str;
cp.push_back('f');

This is problematic since ml_.capacity() != capacity() inside fbstring_core
for shared strings, which causes us not to un-share prior to push_back.

Test Plan: Existing tests, add unittest case.

Reviewed By: tudorb@fb.com

FB internal diff: D580267

Summary:
implement ##find_first_of## and optimize ##Range.find(char)##

============================================================================
folly/test/RangeBenchmark.cpp                   relative  time/iter  iters/s
============================================================================
LongFindSingleCharDirect                                     2.76ms   362.63
LongFindSingleCharRange                           15.88%    17.37ms    57.58
ShortFindSingleCharDirect                                   53.41fs   18.72T
ShortFindSingleCharRange                           0.00%    29.22ns   34.22M
============================================================================

Test Plan:
- added new tests

- ran all folly tests

fbconfig -r folly/ && mkk runtests_opt

Reviewed By: tudorb@fb.com

FB internal diff: D576720

Summary: This makes code easy to read for eyes used to unique/shared ptrs.

Test Plan: unit-tests

Reviewed By: delong.j@fb.com

FB internal diff: D575997

Summary: Add a pop_back() function to dynamic arrays.

Test Plan: - used it

Reviewed By: delong.j@fb.com

FB internal diff: D569440

Summary: Couldn't get an HPHP extension to compile against this

Test Plan: compiled HPHP including folly/dynamic.h

Reviewed By: delong.j@fb.com

FB internal diff: D574364

Test Plan: by hand

Reviewed By: philipp@fb.com

FB internal diff: D570233

Summary:
Folly supports -inf and nan but not -nan, instead it raised unexpected exceptions not handled in the upstream.

This diff is separated out from D569816 -- Diff1.

Test Plan: std::isnan(folly::to<double>(-nan))

Reviewed By: tudorb@fb.com

FB internal diff: D569939

Test Plan: stream_test, exception_tracer_test

Reviewed By: soren@fb.com

FB internal diff: D569476

Summary: Export histograms in a tab-separated format.

Test Plan: Used it.

Reviewed By: lucian@fb.com

Summary:
Moved to separate library instead of :io
Got rid of the mode argument when creating files -- umask is your friend
Separated getSize interface from create
Added FsUtil.h for some fs::path operations
Switched some interfaces to fs::path instead of string / StringPiece
Canonicalize paths so we don't care whether paths are absolute or
relative

Test Plan: fs_util_test, by hand

Reviewed By: philipp@fb.com

FB internal diff: D564465

Test Plan: by hand

Reviewed By: lucian@fb.com

FB internal diff: D555164

Summary:
On --platform=gcc-4.6.2-glibc-2.13-fb the build would fail because

fbcode/third-party/gcc-4.6.2-glibc-2.13-fb/libgcc/libgcc-4.6.2/b4e37bb/include/c++/4.6.2/bits/basic_fbstring_malloc.h

defines it's own version of ALLOCM_SUCCESS & co., but not ALLOCM_LG_ALIGN :(

The '#ifdef ALLOCM_SUCCESS' check in 'folly/Malloc.h' would pass even
though we didn't include jemalloc.

Test Plan: build it

Reviewed By: philipp@fb.com

FB internal diff: D562273

Summary: I'll pretend you reviewed this.

Test Plan: build

Reviewed By: simpkins@fb.com

FB internal diff: D562255

Summary:
This change is mostly mechanical (moving files, changing include paths,
etc).  I made some changes to TARGETS to make it easier for the library
to be linked in (instead of LD_PRELOADed)

Test Plan: by hand

Reviewed By: simpkins@fb.com

FB internal diff: D562196

Test Plan: .

Reviewed By: soren@fb.com

Summary: Intended to complement and replace strings::byLine.

Test Plan: stream_test

Reviewed By: delong.j@fb.com

FB internal diff: D463341

Test Plan: yes

Reviewed By: andrewjcg@fb.com

FB internal diff: D558628

Summary:
When the input type is a long double, perform division using long
double.  In all other cases, divide using double precision.

Also fix the EXPECT_EQ() usage in the test case.

Test Plan: fbconfig -r common/stats folly && fbmake runtests

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D555433

Summary:
D527040 had a bug in the code to compute the average: it incorrectly
performed unsigned division when ValueType was a signed integer type.
As a result, the average was reported incorrectly for stats with
negative values.

This makes the average code more intelligent when handling type
conversions: if the caller wants a floating point value, or if the input
type is floating point, floating point division is always returned.
Otherwise, if the input is a signed type, signed integer division is
performed.  Otherwise, unsigned integer division is performed.

Test Plan: beholdunittests

Reviewed By: lars@fb.com

FB internal diff: D553583

Summary:
I wrote this an an example for the BENCHMARK_NAMED_PARAM() comments in
D527040, and figured I might as well check it in as a real benchmark.

Test Plan:
Here's the output:

======================================================================
folly/test/HistogramBenchmark.cpp         relative  time/iter  iters/s
======================================================================
addValue(0_to_100)                                    15.43ns   64.81M
addValue(0_to_1000)                                   15.45ns   64.70M
addValue(5k_to_20k)                                   15.46ns   64.67M
======================================================================

Reviewed By: rajat@fb.com

FB internal diff: D552875

Summary:
change shared_ptr<>* to raw pointer in
forwardIndex in RTIndex. This saves 16 bytes per
forward list (decrease from 88 to 72) and translates
to about 4-5% memory saving for RTIndex in ff tier.

Test Plan: unit test

Reviewed By: xliux@fb.com

FB internal diff: D551191

Summary:
Add the BucketedTimeSeries class to folly.  This tracks time series
counter data, using a circular buffer of buckets to expire old data
points as time progresses.

It supports querying the overall sum, count, average, and rate for the
duration of the time series, as well as estimating these values for
portions of the overall duration.

Test Plan: Unit tests included.

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D527040

Summary:
Before:

_bin/folly/test/string_test --benchmark --bm_regex=join --bm_min_usec=1000000
============================================================================
folly/test/StringTest.cpp                       relative  time/iter  iters/s
============================================================================
joinCharStr                                                  3.59us  278.29K
joinStrStr                                                   4.44us  225.09K
joinInt                                                     10.55us   94.76K
============================================================================

And after:

_bin/folly/test/string_test --benchmark --bm_regex=join --bm_min_usec=1000000
============================================================================
folly/test/StringTest.cpp                       relative  time/iter  iters/s
============================================================================
joinCharStr                                                  3.61us  277.01K
joinStrStr                                                   3.77us  264.97K
joinInt                                                      9.92us  100.81K
============================================================================

Test Plan: unittests, benchmark

Reviewed By: tudorb@fb.com

FB internal diff: D552364

Summary:
http://burtleburtle.net/bob/hash/spooky.html
The code is in the public domain, so republishing it as such is fine.
The code needed modifications (the test was designed to run on Windows
only; removed unhygienic typedefs; moved to folly::hash namespace)

Also updated Makefile.am files.

Test Plan: test added

Reviewed By: soren@fb.com

FB internal diff: D549572

Summary:
a la Hacker's Delight.

Also, fix nonsensical nextPowTwo benchmark.

Test Plan: Added test and benchmark.

Reviewed By: tudorb@fb.com

FB internal diff: D551510

Summary: To allow codemod changes.

Test Plan: ran test

Reviewed By: tudorb@fb.com

FB internal diff: D549516

Task #906853: Implement ThreadLocalServiceData and reduce lock contentions during channel counter update

Summary:
Cache the counter updates on thread local data structures and aggregate/publish
the data from a separate thread periodically. The ThreadLocalServiceData is
implemented on top of the TLStats data structures.

I am plannig to deploy a private build to production tomorrow to test this
change in production, but figured it might be better to send the diff out first
and get code review feedbacks at the same time.

Test Plan: Run and pass unit tests. Run end-to-end testings on sandbox. (TBD)
Deploy a private build to a production server and make sure the thrift call
latency is reduced.

Reviewed By: xin@fb.com

FB internal diff: D540239

Summary:
Right now, it relies on process id rather than on thread id. Thus, it can
return the same seed when called from multiple threads over really short
period of time, and these unlucky threads will end up with identical "random"
sequences.

Test Plan: newly added test fails against trunk, passes against changes in the diff

Reviewed By: tudorb@fb.com

FB internal diff: D548129

Summary: The same interface as ##facebook::strings::join##, but few times faster.

Test Plan: folly/test/StringTest.cpp

Reviewed By: tudorb@fb.com

FB internal diff: D548863

Summary: convert a dynamic to a well-typed object

Test Plan: run test file

Reviewed By: delong.j@fb.com

FB internal diff: D517021

Summary: script to wrap gateway to run it with exception tracer.

Test Plan: deploy

Reviewed By: rajat@fb.com

FB internal diff: D548147

Summary: Print the times using "ms" for milliseconds, "us" for microseconds, etc.

Test Plan:
$ ./_bin/folly/test/benchmark_test
======================================================================
folly/test/BenchmarkTest.cpp              relative  time/iter  iters/s
======================================================================
bmFun                                                  3.36ns  298.01M
bmRepeatedFun                                          1.76ns  569.38M
----------------------------------------------------------------------
gun                                                    1.70ns  589.82M
----------------------------------------------------------------------
baselinevector                                         3.08us  324.87K
bmVector                                    13.64%    22.56us   44.32K
----------------------------------------------------------------------
superslow                                               1.00s  999.82m
======================================================================

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D546209

Summary: Only tested in release mode, sigh. Fixed a failing asserd.

Test Plan: tested in dbg mode

Reviewed By: tudorb@fb.com

FB internal diff: D546318

Summary:
https://phabricator.fb.com/D544159 reveals a large performance gap between
fbstring and std::string for repeated appends of short strings, which I
consider a relatively urgent matter (as much of our code uses such patterns).

This diff attempts to fix the issue in a principled manner by first appending
the first character with exponential reallocation, after which the rest of the
characters are appended normally.

With the proposed fix the benchmarks are much faster than the previous fbstring
and also than std::string (numbers to follow in comments).

Test Plan: unittested and benchmarked

Reviewed By: soren@fb.com

FB internal diff: D545416

Summary: ... and is shorter and more readable

Test Plan: folly/test

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D545774

Summary:
fbvector was not accepting move_iterators for
InputIterator-templated construct, assign, and insert. The root cause
was the check '(b_ <= &*first && &*first < e_)', used to check if the
assignment was from internal data. This addressof check does not compile
with rvalue-references, and causes the first iterator to be dereferenced
more than once; both against the contract of std::vector. The standard
allows for undefined behaviour in the self-iterator case, so there are
no contractual barriers to removing this check.

Test Plan:
run fbvector test and benchmark. An fbgs for 'assign' turns
up few matches; the seem to be normal use-case. Test fbvector assign
with self-iterators in order (i.e. fbv.assign(fbv.begin(), fbv.end()));
this seems to work fine.

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D535012

Summary: ^

Test Plan: Unit tests

Reviewed By: delong.j@fb.com

FB internal diff: D543586

Summary:
Appending many small strings to fbstring appears much slower (~3x) vs.
std::string. Adding benchmark case to illustrate.

[If this isn't wanted, I can revert]

Test Plan: This is a benchmark.

Reviewed By: soren@fb.com

FB internal diff: D544159