r194403 - Using an invalid -O falls back on -O3 instead of an error

[Hal Finkel]

----- Original Message -----
> 
> 
> 
> 
> On Mon, Nov 11, 2013 at 12:31 PM, Hal Finkel < hfinkel at anl.gov >
> wrote:
> 
> 
> 
> > I think assuming that these optimizations are ordered in some way
> > based on "fast" is wrong. For *many* C++ codebases -O2 produces
> > consistently higher quality code than -O3.
> 
> What does *many* mean? Even a small percentage of the total can be a
> lot in absolute magnitude. -O3 should produce faster code than -O2,
> on average, and if that is not true, then we have a serious problem
> (and the test suite should do more -O2 testing).
> 
> 
> 
> -O3 will trade off code size in the hope of getting more performance.
> Applications where that tradeoff doesn't pay off don't benefit from
> it. Typically, more loop-heavy or numerics heavy applications tend
> to benefit. More generic, or "application"-y code tends to run
> better with -O2.
> 
> 
> 
> Also, this might depend on the platform. Embedded cores might benefit
> much more from -O3 than fancy out-of-order desktop chips.
> 
> 
> 
> Yes, I would expect embedded cores to benefit more, and the code
> workloads run on those cores to also benefit more. Fast out-of-order
> chips (desktop or phone or tablet) tend to like small code more, and
> tend to perform better at -O2.
> 
> 
> 
> Can you please explain further?
> Sorry if I'm still being vague. It's vague. This is based on my
> experience both benchmarking lots of my own software (personal and
> at work) and based on listening to lots of others doing the same. I
> think this is somewhat OK because we're just trying to pick a
> "default" or a "fallback". I just don't think -O3 is the most
> sensible.

Okay; it sounds like we need to make this a data-driven decision. The question is: of those codes using -O<n>, where n > 3, are they more likely to be numerics-dominated codes or more-likely to be control-flow-dominated codes. I remember someone posting a list here of # of packages (in some well-known distribution) vs. the default optimization flag, so this should be an answerable question.

My suspicion is that you'll find them to be more numerics dominated, but my experience here has an obvious bias.

 -Hal

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory