Biggest problem I am coming across is time. For "zip", I have 23 CFLAG combinations, run 10 tests for each. That makes 23 emerges, and 230 timings to take. Can take nearly 24 hours to complete and the machine is unusable (unless you dont mind noise in the timings).
And that doesnt include dynamicly loaded libs that are pulled in ... Also, it seems that one set of flags may be faster on one application, but may have a different effect on another. BillK On Wed, 2003-10-22 at 09:56, Matt Garman wrote: > On Wed, Oct 22, 2003 at 07:08:39AM +0800, William Kenworthy wrote: > > On all the machines (athlon t-bird, p4) I have tried so far, -O3 > > always decreases performance - O2 is best (dramatically so on > > something like a > > Indeed, in this case, O2 is faster than O3. I assumed Debian's gcc > package was compiled with O2, so I just re-merged my gentoo gcc using > O2. I recompiled my program, and now it is back to its better running > time (two or three seconds to load all the records in memory). > > It would be interesting to actually develop a whole suite of tests to > see when O3 is faster (if ever) and when O2 is faster. Based on this > one (completely un-scientific :) test, I'd have to say that the C++ > standard library is best compiled with O2. > > For what it's worth, I did some benchmarking on my new drive as well, > and it's performing as good or better than the old drive. > > Thanks for all the feedback! If anyone else happens to do similar > testing or benchmarking of compiler optimizations, I'd be interested in > reading about them. > > Thanks again, > Matt > > > > -- > [EMAIL PROTECTED] mailing list -- [EMAIL PROTECTED] mailing list
