On Thu, 14 Jul 2022 at 15:51, Richard Biener <richard.guent...@gmail.com> wrote:
> With a value-profile it would be per call site and IPA-CP can use that > to direct the cloning. I'm not sure how many > values a value histogram can track reliably here. > Last time I checked, value profiling can only track a single value per statement. So that would need to be augmented if we wanted to specialize for more than one parameter. However, this can also be a bad idea. Because even if parameters a and b are quasi-constants, perhaps the cross product a x b follows a more uniform distribution. This means that optimizing on a or on b might be a good idea but optimizing on a x b is a bad idea. (There is still some work to be done defining when specializing is a good idea or not). Also, I do not believe that value profiling profiles arguments? According to the comments only values regarding division and/or indirect/virtual call specialization are tracked during value profiling. However, even if value profiling would be a great addition to this, I would like to explore the transformation independent of value profiling. There are some heuristics that could be used, for example looking at the callsites which do have constant arguments and which optimizations are enabled by those constants. You mention that: "IPA-CP can use that to direct the cloning". Can you elaborate a bit further? I guess the idea here is augmenting ipa-cp with a list of "likely values" and extend the infrastructure which deals with speculation to support arguments. Am I following correctly? Any other suggestions?
