Re: Scheduling without profiling

Stephan Ceram a icrit :
> Does it make sense to implement global (or regional) instruction
> scheduling (within a compiler) when profiling information is not
> available? The problem I see here is how to decide which code should
> be scheduled with a higher priority. Since it is not known which path
> is executed most frequently, the scheduler cannot decide prefer a
> particular path. If a wrong path is preferred, other "important" path
> might suffer from that. Or do you know any other heuristics that might
> help to schedule code without profiling? Is it even reasonable in some
> cases to schedule code from different paths?


I would say the contrary: scheduling based on profiling information
has no practical sense, since profiling information depends on the
chosen data input. For numerical applications, this would not be
really a problem, and profiling helps. But for general purpose
applications, profiling does not necessarily helps to produce better
codes for any data input. This is why static code optimisation does
not rely on profiling.

Return to the comp.compilers page.
Search the comp.compilers archives again.