If there are any, can someone please explain rules of thumb for when
to do a PerformDAGCombine operation in ISelLowering vs. when to do a
DAG to DAG transformation?
I'm specifically thinking of an AND + SRL merge into a bit field
extract type instruction. I see that the ARM target does this in
DAG-to-DAG, but this is literally a combine of two instructions,so why
It’s better to do something as a DAG combine if doing so is more likely to expose other DAG combine opportunities. For example, in AMDGPU we have to do a unit conversion to the hardware sin/cos instructions. We could implement this as an instruction pattern for the sin node with the fmul by a constant conversions there. However if the unit conversion is done during custom lowering and you have another fmul input to the sin node, with fast math the 2 fmuls by a constant will be folded into one.
In the case of bitextract instructions it’s kind of harder to analyze them after they are created, so might seem less worth it to have those done in the DAG. However, we currently have at least 3 in-tree targets all trying to match some variant of pretty much the same instruction, so I would like it if we had a generic bit extract node and moved all of the pattern matching to DAG combiner.