ARROW-1589: [C++] Fuzzing for certain input formats

[crepererum]

See https://issues.apache.org/jira/browse/ARROW-1589

This only implements the fuzzer for a single input path, but since it also sets up the fuzzing infrastructure, it should be easy to add more tests (feel free to propose some of them to be included in this PR as well).

[wesm]

I'd prefer to organize this code a little bit closer to the things that are being tested, like the benchmarks and unit tests. So we could have an arrow/ipc/ipc-fuzzing-test.cc executable, etc. Otherwise this looks like a great start, thanks for setting this up

[xhochy]

Trying to set this up and stumbled over the missing libLLVMFuzzer.a on OSX. Is there any better advice than building LLVM separately from source on Macs?

[crepererum]

I've reworked the source file organization as @wesm suggested. Also --- as a nice side effect --- I've added a helper to add other fuzzing tests more easily in the future (similar to the way arrow defines tests executables).

Regarding libLLVMFuzzer.a (@xhochy):
I've used this suggestion some months ago, but I doesn't work anymore since libfuzzer was integrated into compiler-rt. I think you can just build compiler-rt (you should figure out where your llvm-config binary is) on it's own and use libclang_rt.fuzzer_osx.a for that purpose (clone source from here and use the guide on this page for compilation). Or you wait until the next LLVM release where this is hopefully integrated by default.

[crepererum]

Does someone know why the tests are failing here?

[xhochy]

@crepererum yes, new brotli version was published on conda-forge today, so no problem in your code.

[wesm]

I will take this for a drive locally and then merge. Thanks @crepererum!

[xhochy]

+1, got this to work on Ubuntu 17.04

apt install clang-5.0 libfuzzer-5.0-dev
symlink /usr/lib/libLLVMFuzzer.a to  /usr/lib/llvm-5.0/lib/libFuzzer.a

then build in debug mode with the compiler set and it works.