Indeed, it’s just a small confusion. I think “top-down” (to describe root-to-leaves) is less confusing here and it’s closer to an academical definition. To me, the program order doesn’t matter in the definition, I just see it as matching a subtree in a bigger tree.
I’ll just use “root to leaves” from now on to avoid confusion 
I also have some (good!) progress updates on the implementation:
- I re-added all the rule config logic - mostly copy-pasted from the old impl - and it seems to work just fine.
- To support them, I added a special opcode
GIM_CheckSimplePredicatewhich simply calls a function with aunsignedflag to check context-less predicates (doesn’t need a MI and can just be called right next toCheckFeatures). I think it’s a nice addition and it’s not leaking any combiner-specific stuff into the MatchTableExecutor.
- To support them, I added a special opcode
- I ported the AArch64 pre & post legalizer combiners to the new system and they also work completely fine (minus a small bug in my impl that was promptly fixed)
- I generalized
GIR_CombineApplyto aGIR_CustomActionopcode, which, as the name implies, just runs a custom, generic action upon matching (and doesn’t mess with in/out instructions). The action doesn’t return anything and simply takes theMatcherStateas parameter. I also think it’s a useful addition that doesn’t leak combiner-specific things into the executor.
Overall it’s in good shape I would say. It’s a bit messy in places but that’s expected of a WIP. I’ll rewrite/refactor the problematic parts while preparing the patches.
If someone feels like taking a look, here’s how AMDGPUGenPreLegalizerGICombiner.inc looks like now: AMDGPUGenPreLegalizerGICombiner.inc - Pastebin.com