[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: using non-standard optflags (-O3, in particular)



On 02/09/2007 11:22 PM, Patrice Dumas wrote:
On Fri, Feb 09, 2007 at 03:49:49PM -0500, Simon Perreault wrote:
Matthew Miller wrote:
How about I include a specific benchmark test in build section that runs
for, say, a minute, and then use that to auto-select between O2 and O3 on a
per-build basis? :)
This is not acceptable because it would make the build nondeterministic.

Yes, but isn't the gain worth this issue? And it shouldn't be that nondeterministic or it means that the gain in execution time isn't
that big.
The test would mean exactly zero for the end user who might have a completely different architecture. Optimizations which would work great on a dual-core-with-large cache-machine would probably mean zero or worse for a standard P4. What I - as end user - would find useful if I wanted to squeeze the last bit of performance is a table with benchmarks obtained on different architectures with different compilation options. But we cannot ship different packages with N architectures * M compilation options. Or can we? I think that a valid approach for such an option would be the application having the same code compiled in with different optimizations and doing a runtime check to select the fastest one.



[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]