performance regression with @parallel

[amitmurthy]

@JeffBezanson , before commit 30ed04c

  | | |_| | | | (_| |  |  Version 0.4.0-dev+6875 (2015-08-21 16:22 UTC)
 _/ |\__'_|_|_|\__'_|  |  Commit 128797f (3 days old master)
|__/                   |  x86_64-linux-gnu

julia> @time @parallel (+) for i=1:100_000_000; randn(); end;
  0.975937 seconds (172.01 k allocations: 7.561 MB)

julia> @time @parallel (+) for i=1:100_000_000; randn(); end;
  0.850533 seconds (3.55 k allocations: 182.357 KB)

julia> @time @parallel (+) for i=1:100_000_000; randn(); end;
  0.817012 seconds (3.55 k allocations: 182.357 KB)

With 30ed04c

  | | |_| | | | (_| |  |  Version 0.4.0-dev+6876 (2015-08-21 17:38 UTC)
 _/ |\__'_|_|_|\__'_|  |  Commit 30ed04c (3 days old master)
|__/                   |  x86_64-linux-gnu

julia> @time @parallel (+) for i=1:100_000_000; randn(); end;
 10.096709 seconds (300.17 M allocations: 5.968 GB, 3.52% gc time)

julia> @time @parallel (+) for i=1:100_000_000; randn(); end;
 10.475335 seconds (300.00 M allocations: 5.961 GB, 3.76% gc time)

NOTE: This is not an issue when wrapped in a function even with 30ed04c:

julia> function foo(); @time @parallel (+) for i=1:100_000_000; randn(); end; end;

julia> foo()
  0.816561 seconds (3.42 k allocations: 175.723 KB)
5660.182992102461