[C++] arrow_reader_writer_test.cc slow to compile

[asfimport]

As soon as some optimizations are enabled, src/parquet/arrow/arrow_reader_writer_test.cc becomes extremely slow to compile (more than one minute just for itself). This is perceivable on e.g. the conda-cpp-valgrind build where we add -Og to the gcc flags in order to make the tests less slow under emulation.

Reporter: Antoine Pitrou / @pitrou

_{Note: This issue was originally created as ARROW-13221. Please see the migration documentation for further details.}

[asfimport]

Antoine Pitrou / @pitrou:
Note this may also be a factor of sporadic CI errors:
https://github.com/ursacomputing/crossbow/runs/2949289160

[asfimport]

Antoine Pitrou / @pitrou:
I think there are two aspects to this:

1. split the tests into a least two different test files
2. rewrite some of the templated tests (inducing long code generation times at compilation) to be runtime-parametric instead (for example using RandomArrayGenerator or ArrayFromJSON)

Given that those tests have grown organically, step 2 above will probably be slightly cumbersome... One needs to understand what each test checks for and rewrite it while keeping its intent.

cc @kou @benibus

[asfimport]

Kouhei Sutou / @kou:
It seems that 1) may not reduce total build time. Because it just splits whole tests and doesn't change the total number of tests and the current build approach (templated tests).

It seems that 2) will reduce total build time.

step 2 above will probably be slightly cumbersome...

I agree with it...

[asfimport]

Antoine Pitrou / @pitrou:

It seems that 1) may not reduce total build time. Because it just splits whole tests and doesn't change the total number of tests and the current build approach (templated tests).

It depends on the parallelism. In a parallel build, compiling arrow_reader_writer_test.cc is often the last task to finish at the end...

[pitrou]

This issue still applies. cc @mapleFU in case you want to take a look.

[mapleFU]

I'll build and trace with LLVM to see how it spend time on compiling this file this weekend.

[mapleFU]

===-------------------------------------------------------------------------===
                         Miscellaneous Ungrouped Timers
===-------------------------------------------------------------------------===

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  ---Instr---  --- Name ---
  11.0381 ( 82.7%)  20.8372 ( 99.7%)  31.8752 ( 93.1%)  32.8761 ( 93.3%)  426134281345  Code Generation Time
   2.3077 ( 17.3%)   0.0531 (  0.3%)   2.3608 (  6.9%)   2.3671 (  6.7%)  20620474061  LLVM IR Generation Time
  13.3458 (100.0%)  20.8903 (100.0%)  34.2361 (100.0%)  35.2433 (100.0%)  446754755406  Total

In debug on my Machine nearly 20 seconds are taken to compile it...Should we try to split it?

[pitrou]

Yes, as I wrote in #28908 (comment), we can try to split it. Then it might also become easier to rely less on templating.

[mapleFU]

I've tried to split the page index test out. But this doesn't make it compile faster :-(

f69ab13