On 10/10/14 09:50, Ilya Enkovich wrote:
Checks and and intersection removal code was added as a simple pass
catching trivial cases. I'm sure there are optimizations having
common elements with what checker optimizer does. But initially we
didn't want to adopt existing optimizers because GIMPLE representation
of instrumentation was not stable and also we still don't know what
are important targets for optimizations.
Understood.
The plan is to have stable version first. After enabling we want to
make performance analysis and determine which optimizations are most
required (it may appear checks removal doesn't give any significant
performance gain at all), determine which of current infrastructure
may be re-used (if any) and implement proper checker optimization.
Current optimizer is a simple code cleanup. I do not think we should
make any significant rework of it as a part of enabling. If current
approach seems to require significant changes to go to trunk then it
should be probably delayed and go separately from instrumentation
pass.
Well, I think it should be trivial to handle the redundant check
elimination in DOM.
Most likely eliminate_redundant_computations needs some work to allow it
to look inside those checks and get them recorded into its tables. With
that in place, DOM should optimize this stuff without further
intervention. It's probably less code than you've already written :-)
The swapping variant feels like it should be simple to implement with
the existing dominator walkers. But I haven't thought nearly as much
about that one.
jeff
