Prefer platform specific memory barriers

[ianks]

Following the wisdom of postgres, we opt to use platform specific memory barriers when available. These are generally more performant. In this PR, we add specific cases for i386, x86_64.

In the future, we could look at using pg's atomics library directly: https://github.com/postgres/postgres/tree/9d4649ca49416111aee2c84b7e4441a0b7aa2fac/src/include/port/atomics

[pitr-ch]

Thanks @ianks. I cannot knowledgeable judge if this is a good change. @thedarkone maybe?

[thedarkone]

GCC emits mfence for __sync_synchronize() (so does Clang), I don't know what MS compiler does (presumably the same). Per Shipilev's findings using the proposed asm should result in about 15 % speed up for the synch. instruction in question. I don't think the improvement is measurable with the general overhead that usually comes with MRI. @ianks, have you done any benching?

If this is to be merged the PR should also clean up extconf.rb:

concurrent-ruby/ext/concurrent/extconf.rb

Line 46 in b461ffc

try_run(<<CODE,$CFLAGS) && ($defs << '-DHAVE_GCC_SYNC')

since HAVE_GCC_SYNC doesn't seem to be used anywhere else.

I would consider merging this if and only if there was a measurable difference on MRI.

[ianks]

@thedarkone Just did some benchmarking on x86 (do not have another other machines at the moment).

AtomicBoolean#true?

Pessimistic results (taking slowest result from this branch, fastest from master), show 7.89314% improvement. The optimistic (this branch best, master worst) was 14.8174% improvement

AtomicBoolean#make_true/make_false

Pessimistic results (taking slowest result from this branch, fastest from master), show 7.19176% improvement. The optimistic (this branch best, master worst) was 11.5249%% improvement

require 'benchmark/ips'
require 'concurrent'

bool = Concurrent::AtomicBoolean.new(false)
`bundle exec rake compile`

# master
# ==========================================================================
#               #true?      9.922M (± 2.7%) i/s -     49.768M in   5.020095s
#               #true?      9.675M (± 1.2%) i/s -     48.484M in   5.011882s
#               #true?      9.550M (± 4.4%) i/s -     47.841M in   5.021202s
# #make_true/#make_false    7.469M (± 2.7%) i/s -     37.386M in   5.009572s
# #make_true/#make_false    7.386M (± 2.5%) i/s -     37.123M in   5.029803s
# #make_true/#make_false    7.398M (± 2.2%) i/s -     37.123M in   5.021120s

# update-memory-barriers
# ==========================================================================
#               #true?     11.098M (± 1.9%) i/s -     55.497M in   5.002488s
#               #true?     10.891M (± 1.4%) i/s -     54.673M in   5.021049s
#               #true?     10.766M (± 5.6%) i/s -     53.849M in   5.021705s
# #make_true/#make_false   8.289M (± 2.2%) i/s -     41.663M in   5.029083s
# #make_true/#make_false   8.208M (± 2.2%) i/s -     41.167M in   5.018527s
# #make_true/#make_false   8.047M (± 4.7%) i/s -     40.175M in   5.006055s
Benchmark.ips do |x|
  x.config(iterations: 3)

  x.report('#true?') do
    bool.true?
  end

  x.report('#make_true/#make_false') do
    bool.make_true
    bool.make_false
  end
end

[ianks]

@pitr-ch Are there any other changes you would like to see here?

[pitr-ch]

@ianks Thanks for reminding me about this. Looks good. Could you put a comment in the code explaining when it comes from, as you did int the PR's description. Then we can merge.

[pitr-ch]

@thedarkone thank you for your input on the PR!

[pitr-ch]

Thanks @ianks.