From: "drbrain (Eric Hodel)" <drbrain@...7.net>
Date: 2013-03-01T06:51:25+09:00
Subject: [ruby-core:53049] [ruby-trunk - Bug #7984] Severe speed issues in 2.0.0 compiled with Clang


Issue #7984 has been updated by drbrain (Eric Hodel).


I cannot reproduce your result:

  $ ruby -rbenchmark -ve 'puts Benchmark.measure { 50000000.times { Object.new } }'
  ruby 2.0.0p0 (2013-02-24 revision 39473) [x86_64-darwin12.2.1]
   10.600000   0.010000  10.610000 ( 10.612499)
  $ ruby19 -rbenchmark -ve 'puts Benchmark.measure { 50000000.times { Object.new } }'
  ruby 1.9.3p374 (2013-01-15 revision 38858) [x86_64-darwin12.2.1]
   11.410000   0.010000  11.420000 ( 11.412127)

Using ministat for a better comparison I see an 8.4% improvement +/- 0.8% with 1/10th the iterations per sample and 20 samples (overall, double the object creation):

  $ for f in `jot 20`; do ruby -rbenchmark -e 'puts Benchmark.measure { 5000000.times { Object.new } }.real'; done > 2_0_0p0.txt
  $ for f in `jot 20`; do ruby19 -rbenchmark -e 'puts Benchmark.measure { 5000000.times { Object.new } }.real'; done > 1_9_3p374.txt
  $ ministat 2_0_0p0.txt 1_9_3p374.txt
  x 2_0_0p0.txt
  + 1_9_3p374.txt
  +--------------------------------------------------------------------------+
  |                                                  +                       |
  |                                                  +      +                |
  |     xx                                           + +    +                |
  |    xxxx x                                     +  + ++   +                |
  |xx  xxxxxx xx    x     x  x                    ++ + ++ + +            +  +|
  |  |____M_A_____|                               |____M_A______|            |
  +--------------------------------------------------------------------------+
      N           Min           Max        Median           Avg        Stddev
  x  20      1.031859       1.08182      1.045697     1.0484288   0.012780833
  +  20     1.1223431      1.172935     1.1333108     1.1362227    0.01309569
  Difference at 95.0% confidence
  	0.087794 +/- 0.00828168
  	8.37386% +/- 0.789914%
  	(Student's t, pooled s = 0.0129392)

Here are my CC and CFLAGS settings:

  $ ruby -rrbconfig -ve 'p RbConfig::CONFIG.values_at(*%w[CC CFLAGS])'
  ruby 2.0.0p0 (2013-02-24 revision 39473) [x86_64-darwin12.2.1]
  ["clang", " -O3 -fno-fast-math -ggdb3 -Wall -Wextra -Wno-unused-parameter -Wno-parentheses -Wno-long-long -Wno-missing-field-initializers -Wunused-variable -Werror=pointer-arith -Werror=write-strings -Werror=declaration-after-statement -Werror=shorten-64-to-32 -Werror=implicit-function-declaration  -pipe"]


  $ ruby19 -rrbconfig -ve 'p RbConfig::CONFIG.values_at(*%w[CC CFLAGS])'
  ruby 1.9.3p374 (2013-01-15 revision 38858) [x86_64-darwin12.2.1]
["clang", " -O3 -ggdb -Wall -Wextra -Wno-unused-parameter -Wno-parentheses -Wno-long-long -Wno-missing-field-initializers -Werror=pointer-arith -Werror=write-strings -Werror=declaration-after-statement -Werror=shorten-64-to-32 -Werror=implicit-function-declaration  -pipe"]

----------------------------------------
Bug #7984: Severe speed issues in 2.0.0 compiled with Clang
https://bugs.ruby-lang.org/issues/7984#change-37201

Author: jcole1989 (James Coleman)
Status: Open
Priority: Normal
Assignee: 
Category: 
Target version: 
ruby -v: 2.0.0p0 on darwin


I've been comparing Ruby 2.0.0 with 1.9.3-head, both compiled with the same basic set of CFLAGs with Clang on Mac OS 10.8 (Mountain Lion). A decent sized rails app was showing performance degradations, so I looked into the most basic of (micro)benchmarks. Running the following:

t = Benchmark.measure do
  50000000.times { Object.new }
end

Gives me the following output:

1.9.3-head: 11.240000 0.000000 11.240000 ( 11.247285)
2.0.0p0: 23.610000 0.010000 23.620000 ( 23.629643)

Ruby 2.0.0 takes over twice as long as 1.9.3. In contrast, running the same test under rubies compiled with GCC on linux (unfortunately I can't do a direct test on OSX because 2.0.0 currently won't compile under the current apple gcc 4.2) Ruby 2.0.0 shows a 15-20% speed improvement over 1.9.3.

So it seems there's a severe speed regression interaction with the ruby 2.0.0 code and Clang.


-- 
http://bugs.ruby-lang.org/