[code-reflection] RFR: Experiment: Alternative SSA construction algorithm

[Paul Sandoz]

On Wed, 14 Aug 2024 07:58:53 GMT, Hannes Greule <hgreule at openjdk.org> wrote:

> Hi,
> 
> this is mainly an experiment, I wanted to play around with the API and see how an alternative SSA construction algorithm can be implemented. There are a few notable problems I encountered:
> 
> - The original algorithm is supposed to work on-the-fly in a single step. This does not work with the current API.
>   - `Value`s (operands) and `Reference`s (successor arguments) are bound eagerly. I assume terminating ops could be emitted only after everything else was done, but that's not easy to achieve currently either. For operands, I don't see any way to say "oh, please use a different value" after the operation is passed to the builder. I'm not sure how relevant this is for other transformations, but if we find a simple solution to this without giving up immutability, I could imagine it to be very powerful.
>   - I'm currently using more data structures to cope with the two-step variant than the original algorithm itself requires.
>     - The `predecessors` map could be avoided at the cost of additional computation whenever only filled blocks should be considered
>     - The `loads` map is required as we can't replace the loads directly during the analysis step
>     - The `additionalParameters` could probably be removed if `Block.Builder` allowed removing parameters
>     - The `deletedPhis` set is also needed to cope with the fact that we look up current defs in both steps
> - I noticed that for the method in `TestTraverse`, this implementation gets rid of one parameter that the existing implementation does not get rid of. I'm not sure if it is designed to produce minimal SSA, but that looks odd. (`c` is passed to the while body, but it is never read there before being written)
> 
> 
> I think the algorithm itself is pretty straightforward, but the extra steps needed to adapt to the API somewhat nullify that.
> I tried to document the implementation details that differ from the paper, but if there are any questions or suggestions how this could be improved, please let me know.

This is a good experiment. The Cytron paper is quite old and I did not check to see if there were more recent alternative algorithms that were simpler. It will take me some time to fully digest the paper and your implementation. My gut feeling is this might be a good default replacement for the current implementation. I am curious as to how easily you found the adaption from phi nodes to block parameters. Generally i have found it easy to adapt and it can often simplify things.

You noticed a deficiency in the Cytron implementation (and possibly the algorithm), regarding a variable that is just stored to. We have noticed that too.

I don't currently see a way to retain immutability and post-adapt what is currently built before building completes. Maybe as I look more deeply a few ideas will emerge. One approach i have used when transforming is instead of building up data structures I build up (capturing) functions on the first pass to apply when transforming on the second pass. I don't yet know if this would make your implementation simpler.

-------------

PR Comment: https://git.openjdk.org/babylon/pull/214#issuecomment-2292415152