[pypy-dev] playing with gcc

[Michael Hudson]

I spent a while compiling some pypy-c source with different
optimization options and running pystone (on an iMac with a 1.8 ghz G5
and 1.5 gigs of ram).

Here's a table of results:

            0       1       2       3       3p      s       sp      c   
   0       1.00    3.06    3.08    3.08    3.17    3.12    3.25   45.39 
   1       0.33    1.00    1.01    1.01    1.04    1.02    1.06   14.82 
   2       0.32    0.99    1.00    1.00    1.03    1.01    1.05   14.72 
   3       0.32    0.99    1.00    1.00    1.03    1.01    1.05   14.74 
   3p      0.32    0.97    0.97    0.97    1.00    0.99    1.02   14.31 
   s       0.32    0.98    0.99    0.99    1.02    1.00    1.04   14.53 
   sp      0.31    0.94    0.95    0.95    0.98    0.96    1.00   13.98 
   c       0.02    0.07    0.07    0.07    0.07    0.07    0.07    1.00 

Here's a key:

0 : -O0, aka "no optimisation"
1 : -O1
2 : -O2
3 : -O3
3p: -O3 -fbranch-probabilities 
   (i compiled with -fprofile-arcs and ran pystone to generate the branch info)
s : -Os. aka "optimize for size"
sp: -Os -fbranch-probabilities 
c : cpython for reference

As you can see, you definitely want *some* optimization but it doesn't
make a great deal of difference what level you choose.  That said, -Os
gives the best results and -fbranch-probabilities gives a ~5% gain.

I timed each compilation, but forgot to write the results down
anywhere.  There weren't vast differences at any rate (they all took
~7 minutes, not too bad).  Of course, using -fbranch-probabilities
requires you compile everything twice.

The machine wasn't totally quiet as I ran these tests, but fairly.

Cheers,
mwh

-- 
  What the semicolon's anxious supporters fret about is the tendency
  of contemporary writers to use a dash instead of a semicolon and
  thus precipitate the end of the world.
                               -- Lynne Truss, "Eats, Shoots & Leaves"