ARROW-8113: [C++] Lighter weight variant<>

[bkietz]

• Unit tests for util::Variant<>
• Compilation time and code size comparison to ensure this refactoring is beneficial
• Benchmark comparison to ensure compute and nested parquet are not impacted
• More docstrings and comments

Locally (release build, no ccache):

$ ninja
ninja: no work to do.
$ export OBJS=`ls src/arrow/CMakeFiles/arrow_objlib.dir/compute/**/*.o`

$ git checkout master && rm $OBJS && time ninja $OBJS
real    0m16.918s
user    2m41.702s
sys     0m2.495s

$ du -h --total $OBJS | tail -n 1
total 7.3M

$ git checkout 8113-Implement-a-lighter-weigh && rm $OBJS && time ninja $OBJS
real    0m14.363s
user    2m19.939s
sys     0m2.125s

$ du -h --total $OBJS | tail -n 1
6.6M    total

[github-actions]

https://issues.apache.org/jira/browse/ARROW-8113

[bkietz]

We might want to borrow this idiom to disable copy construction for Result<std::unique_ptr<T>>

[emkornfield]

drive by comment. just want to make sure we aren't using the same idioms for variant (i.e. they are compatible with C++ std)?

[bkietz]

@emkornfield I'm implementing a subset of std::variant's functionality but all members which are present are compatible; if we switch to using std::variant at some point no code will break. We will need to keep the mutable-pointer visit overload since that's non standard, but IMHO that's acceptable overhead

[wesm]

@bkietz presumably this yields smaller code size, too?

[bkietz]

Updated description with build times and code sizes for a release build

[pitrou]

This seems to speed up the compute layer a bit.

• Before:

ArrayArrayKernel<Add, Int64Type>/32768/100       3781 ns         3780 ns       738685 bytes_per_second=8.07385G/s items_per_second=1083.65M/s null_percent=1 size=32.768k

• After:

ArrayArrayKernel<Add, Int64Type>/32768/100       3615 ns         3614 ns       777365 bytes_per_second=8.44344G/s items_per_second=1.13326G/s null_percent=1 size=32.768k

[pitrou]

@bkietz Do you plan to update this?

[bkietz]

@pitrou yes but not soon

[pitrou]

Rebased, I'm writing some micro-benchmarks.

[pitrou]

Variant benchmarks:

• before:

ConstructTrivialVariant         13290 ns        13287 ns        52532 items_per_second=752.595M/s
ConstructNonTrivialVariant     111230 ns       111215 ns         6202 items_per_second=89.916M/s
VisitTrivialVariant             25744 ns        25740 ns        27269 items_per_second=388.497M/s
VisitNonTrivialVariant          24389 ns        24385 ns        28649 items_per_second=410.082M/s
ConstructDatum                 102925 ns       102911 ns         7190 items_per_second=97.1714M/s
VisitDatum                       6204 ns         6204 ns       112708 items_per_second=1.61199G/s

• after:

ConstructTrivialVariant         13952 ns        13951 ns        50116 items_per_second=716.811M/s
ConstructNonTrivialVariant     108507 ns       108489 ns         6412 items_per_second=92.1753M/s
VisitTrivialVariant              5848 ns         5847 ns       119628 items_per_second=1.71014G/s
VisitNonTrivialVariant           5973 ns         5972 ns       118440 items_per_second=1.6745G/s
ConstructDatum                  98938 ns        98923 ns         7038 items_per_second=101.088M/s
VisitDatum                       5566 ns         5565 ns       125741 items_per_second=1.79701G/s

[pitrou]

Compute overhead benchmarks:

• before:

BM_CastDispatch                      971704 ns       971533 ns          698 items_per_second=1054k/s
BM_CastDispatchBaseline               97365 ns        97352 ns         6955 items_per_second=10.5185M/s
BM_AddDispatch                          165 ns          165 ns      4232825 items_per_second=6.07671M/s
BM_ExecuteScalarFunctionOnScalar    4666801 ns      4665955 ns          149 items_per_second=2.14318M/s
BM_ExecuteScalarKernelOnScalar       809563 ns       809419 ns          849 items_per_second=12.3545M/s

• after:

BM_CastDispatch                      914006 ns       913869 ns          759 items_per_second=1.12051M/s
BM_CastDispatchBaseline               95985 ns        95973 ns         7287 items_per_second=10.6697M/s
BM_AddDispatch                          167 ns          167 ns      4137185 items_per_second=5.99082M/s
BM_ExecuteScalarFunctionOnScalar    4345926 ns      4345337 ns          161 items_per_second=2.30132M/s
BM_ExecuteScalarKernelOnScalar       847525 ns       847411 ns          835 items_per_second=11.8007M/s

[pitrou]

@bkietz Do you know why all Github CI checks are skipped here?

[pitrou]

AppVeyor seems to indicate that MSVC isn't picking up / implementing the copy constructor correctly :-(

[pitrou]

@bkietz You may want to review the changes I've done. I think only docs and comments remain to be done.

[pitrou]

I added some docstrings. I think this is ready now.

[pitrou]

Hmm, I forgot to push some changes...

[bkietz]

@pitrou thanks for picking this up! Looks good to me

[pitrou]

Ok, I'll rebase a last time to make sure this doesn't break anything.

[kou]

I don't know why but It seems that this causes g++ 9 crash:

https://github.com/ursa-labs/crossbow/runs/1522097163#step:4:1237

during GIMPLE pass: pre
In file included from /usr/include/c++/9/functional:59,
                 from /arrow/cpp/src/arrow/vendored/string_view.hpp:1480,
                 from /arrow/cpp/src/arrow/util/string_view.h:25,
                 from /arrow/cpp/src/arrow/buffer.h:31,
                 from /arrow/cpp/src/arrow/array/data.h:26,
                 from /arrow/cpp/src/arrow/array/array_base.h:26,
                 from /arrow/cpp/src/arrow/array/builder_binary.h:30,
                 from /arrow/cpp/src/arrow/compute/kernels/codegen_internal.h:26,
                 from /arrow/cpp/src/arrow/compute/kernels/codegen_internal.cc:18:
/usr/include/c++/9/bits/std_function.h: In static member function 'static void std::_Function_handler<void(_ArgTypes ...), _Functor>::_M_invoke(const std::_Any_data&, _ArgTypes&& ...) [with _Functor = arrow::compute::internal::MakeFlippedBinaryExec(arrow::compute::ArrayKernelExec)::<lambda(arrow::compute::KernelContext*, const arrow::compute::ExecBatch&, arrow::Datum*)>; _ArgTypes = {arrow::compute::KernelContext*, const arrow::compute::ExecBatch&, arrow::Datum*}]':
/usr/include/c++/9/bits/std_function.h:298:7: internal compiler error: Segmentation fault
  298 |       _M_invoke(const _Any_data& __functor, _ArgTypes&&... __args)
      |       ^~~~~~~~~
Please submit a full bug report,
with preprocessed source if appropriate.
See <file:///usr/share/doc/gcc-9/README.Bugs> for instructions.
make[2]: *** [src/arrow/CMakeFiles/arrow_objlib.dir/build.make:1610: src/arrow/CMakeFiles/arrow_objlib.dir/compute/kernels/codegen_internal.cc.o] Error 1

We can reproduce this on local:

(cd cpp/examples/minimal_build && docker-compose build static && docker-compose run --rm static)

[kou]

https://issues.apache.org/jira/browse/ARROW-10867 may be related.