[Jikesrvm-researchers] Strange performance results on PPC Linux

[Chris H. <hof...@cs...>]

I've tried this with both a "production" configuration and 
"FastAdaptiveGenCopy" and in each case I'm seeing that the adaptive 
compiler is significantly slower than what is supposed to be a pure 
optimizing compiler. In the past, opt compiling every method has 
generally resulted in significantly slower first runs than using the 
adaptive compiler.

For instance, here's the results for the initial run of _205_raytrace 
with "FastAdaptiveGenCopy" using last night's CVS head:

adaptive (no args): 40.105
opt compiled      : 13.014

For compress the ratio is even worse, roughly 126:27!

Eliot suggested the compiler DNA was incorrect for PPC/Linux, but I just 
discovered something even worse. If I enable aos:logging, the time drops 
back down to around 23 seconds!

  rvm  -X:aos:logging_level=1 SpecApplication _201_compress
      --> 22.83
  rvm                         SpecApplication _201_compress
      --> 125.767




Here are the options I'm giving the adaptive compiler to try to force it 
to opt compile all methods and not recompile any:

-X:aos:enable_recompilation=false -X:aos:initial_compiler=opt 
-X:aos:adaptive_inlining=false -X:aos:osr_promotion=false 
-X:opt:osr_guarded_inlining=false -X:opt:osr_inline_policy=false

-- 
Chris Hoffmann -- Dept. of Computer Science/UMass at Amherst
http://www-ali.cs.umass.edu/~hoffmann