Benchmarking compiler optimizations?
Kevin
with different compiler optimizations. This is my logic. Please
criticize it.
All the benchmarks I've found are "synthetic" benchmarks intended to
measure specific hardware performance. I considered compiling these
benchmarks with different optimizations and comparing their performance,
but I have a gut feeling that the performance of a low-level benchmark
won't benefit much (if at all) from compiler optimizations. Anyway,
synthetic benchmarks are already in a different universe from real-world
performance; using them to measure the performance gains of compiler
optimization only adds another layer of abstraction.
Then I thought about bootstrapping the compiler with different compiler
options and measuring its performance in extremis, like by seeing how
long it takes to rebuild glibc. But then again, a compiler isn't a
typical real-world application.
I guess the kernel is about as real-world as you can get. What I'm going
to do is run a general-purpose benchmark suite (or several of them) under
various kernels, identically configured but compiled with different
optimizations.
I've yet to dig into choosing my benchmarks. Any recommendations?
Krum
Kevin
mai...@aol.mil says...
> I guess the kernel is about as real-world as you can get. What I'm going
> to do is run a general-purpose benchmark suite (or several of them) under
> various kernels, identically configured but compiled with different
> optimizations.
Er... this makes less sense than I thought a minute ago.
What would be a good real-world application to test compiler
optimizations on?
Better yet, what properties would make a piece of software a good
candidate for testing the benefits of compiler optimization? Yes, that's
the question I wanted to ask.
Clark L. Coleman
Check out the compiler benchmarking/characterization at
www.nullstone.com before you start from scratch.