locale-independent str-to-num

[TurpentineDistillery]

This implements #302

[gregmarr]

typo: interal

[gregmarr]

do/while(0) instead of a pair of nested if statements seems a bit too tricky.

[TurpentineDistillery]

do{...}while(0) was originally the technique to group statements in C macros, but it can also be used as a local scope that you can break out of, kind of a disciplined scoped goto. I just have really strong aversion to nesting control flow and modifying state; to a fault perhaps : )

[gregmarr]

Yeah, I'm familiar. I don't particularly like nesting either, but this saves only one level of nesting.

[gregmarr]

As it is now: http://pastebin.com/UVTzUF6u
Simpler with no lambda capture, no lambda call at the end, fewer break/return statements: http://pastebin.com/vQu7y0CA
Personally, I don't think avoiding the one level of indentation for 10 lines of code is worth the extra complexity.

[TurpentineDistillery]

It's not just about avoiding an extra level of control flow, however. I usually prefer the lambda style for the following reasons:

• manifests intent by declaring what we're about to compute and what the inputs are (especially handy in in-person code reviews)
• instead of the result being assigned as a side-effect in various control-flow branches, all control flow is "joined" into the return value, and the result of the lambda can be assigned to a const variable.
• there's no (or may be minimal) runtime overhead.

In short-enough cases, however, this approach would probably be considered an overkill by many (but not by me - there’s no lambda small enough!)

Stylistic issues aside, if I could only manage to make setlocale affect the behavior of snprintf in the unit test on all compilers... That's even before any json-specific checks are triggered.

[gregmarr]

setlocale is not thread specific or thread safe, so I would recommend avoiding it.

[TurpentineDistillery]

setlocale with C locales is the equivalent of std::locale::global with C++ locales, which is not thread-specific or thread-safe either.

We're not calling it it from the library code, however, but from the unit test to simulate the current state of a locale-specific application. All locale-dependent functions, including std::to_string() or printf() or cout << are not thread-safe in that sense, as a rogue thread may change C or C++ locale from under your feet by changing the locale in-flight because "they" needed it.

[gregmarr]

Okay, missed that you were referring to test code.

[coveralls]

Coverage decreased (-0.6%) to 99.415% when pulling e41a956 on TurpentineDistillery:feature/locale_independent_str_to_num into bc28942 on nlohmann:develop.

[gregmarr]

This class isn't used in this PR.

[gregmarr]

If exceptions are being removed, this comment needs to be updated.

[TurpentineDistillery]

Thanks for reviewing; will clean up in the next push.

[coveralls]

Coverage decreased (-0.09%) to 99.91% when pulling 1c029b9 on TurpentineDistillery:feature/locale_independent_str_to_num into bc28942 on nlohmann:develop.

[coveralls]

Coverage remained the same at 100.0% when pulling 6fba52b on TurpentineDistillery:feature/locale_independent_str_to_num into 79fa8b2 on nlohmann:develop.

[nlohmann]

I shall have a look at the PR this week.

[nlohmann]

@TurpentineDistillery Could you update the PR or resolve the conflicts, please?

[TurpentineDistillery]

Will do.

[coveralls]

Coverage remained the same at 100.0% when pulling c236b59 on TurpentineDistillery:feature/locale_independent_str_to_num into 9f6c86f on nlohmann:develop.

[nlohmann]

In lieu of better approaches, here are the number using Google Benchmark:

$ ./benchmark-develop --benchmark_filter="parse.*" --benchmark_report_aggregates_only=true --benchmark_repetitions=30 
Run on (8 X 2900 MHz CPU s)
2017-01-11 22:07:32
Benchmark                                                         Time           CPU Iterations
-----------------------------------------------------------------------------------------------
parse data/jeopardy/jeopardy.json_mean                   1024555051 ns 1023570733 ns          1   51.8754MB/s
parse data/jeopardy/jeopardy.json_stddev                   47953641 ns   47365176 ns          0   2.47509MB/s
parse data/nativejson-benchmark/canada.json_mean           63870560 ns   63820661 ns         11   33.7019MB/s
parse data/nativejson-benchmark/canada.json_stddev          2750435 ns    2732684 ns          0    1.5052MB/s
parse data/nativejson-benchmark/citm_catalog.json_mean     12256813 ns   12245026 ns         52   135.116MB/s
parse data/nativejson-benchmark/citm_catalog.json_stddev     812191 ns     804674 ns          0   9.08961MB/s
parse data/nativejson-benchmark/twitter.json_mean           8013096 ns    8005479 ns         91   75.3725MB/s
parse data/nativejson-benchmark/twitter.json_stddev          345569 ns     341982 ns          0   3.32091MB/s
parse data/numbers/floats.json_mean                       386869345 ns  386530517 ns          2   55.9661MB/s
parse data/numbers/floats.json_stddev                       9113975 ns    9023610 ns          0   1.37988MB/s
parse data/numbers/signed_ints.json_mean                  116659793 ns  116580717 ns          6   199.814MB/s
parse data/numbers/signed_ints.json_stddev                  4985557 ns    4946958 ns          0   9.07039MB/s
parse data/numbers/unsigned_ints.json_mean                115276698 ns  115201656 ns          6    202.08MB/s
parse data/numbers/unsigned_ints.json_stddev                2588632 ns    2572249 ns          0   4.71005MB/s

$ ./benchmark-parse-pr --benchmark_filter="parse.*" --benchmark_report_aggregates_only=true --benchmark_repetitions=30 
Run on (8 X 2900 MHz CPU s)
2017-01-11 22:10:50
Benchmark                                                         Time           CPU Iterations
-----------------------------------------------------------------------------------------------
parse data/jeopardy/jeopardy.json_mean                   1035173889 ns 1034375333 ns          1   51.2314MB/s
parse data/jeopardy/jeopardy.json_stddev                   15369943 ns   15138550 ns          0   774.835kB/s
parse data/nativejson-benchmark/canada.json_mean           62640302 ns   62605706 ns         11   34.3258MB/s
parse data/nativejson-benchmark/canada.json_stddev          1964575 ns    1955608 ns          0   1.13517MB/s
parse data/nativejson-benchmark/citm_catalog.json_mean     13005561 ns   12994596 ns         52   126.821MB/s
parse data/nativejson-benchmark/citm_catalog.json_stddev     285190 ns     282784 ns          0   2.81027MB/s
parse data/nativejson-benchmark/twitter.json_mean           8288143 ns    8279872 ns         82   72.8825MB/s
parse data/nativejson-benchmark/twitter.json_stddev          360370 ns     356344 ns          0    3.3744MB/s
parse data/numbers/floats.json_mean                       366096955 ns  365863167 ns          2   59.1376MB/s
parse data/numbers/floats.json_stddev                       9687455 ns    9619537 ns          0    1.6742MB/s
parse data/numbers/signed_ints.json_mean                  127861852 ns  127771639 ns          6   182.021MB/s
parse data/numbers/signed_ints.json_stddev                  2265199 ns    2245588 ns          0   3.28731MB/s
parse data/numbers/unsigned_ints.json_mean                123481147 ns  123403006 ns          6   188.589MB/s
parse data/numbers/unsigned_ints.json_stddev                1737543 ns    1722133 ns          0   2.64446MB/s

[TurpentineDistillery]

From the theoretical standpoint, the numbers are expected to be about the same, except that integral types are parsed with std::strto[u]ll, which may be slower than doing that "manually", but not by much.

[nlohmann]

I forked the branch to https://github.com/nlohmann/json/tree/TurpentineDistillery-feature/locale_independent_str_to_num and made some changes. Once Travis is done checking, I shall merge it to develop. Thanks @TurpentineDistillery for the effort and @gregmarr for the comments!

[nlohmann]

Closed with #450.