is there a simple way to alternate between different register allocators at run time? I would like to decide dynamically which register allocator to use. The decision is based on information that is produced by a pass that executes before register allocation takes place.
Is it possible to modify RegisterRegAlloc::Registry to take into consideration this information? Is there any similar example in LLVM?
all the best,
To do this you would need to change the register allocation pass on the fly. I don't think this is currently possible. Someone please correct me if I am wrong.
I would like to see hierarchical pass capability built into pass manager. This allows us to group live variables, phi lowering, two-address, lower subreg, live interval analysis, coalescing, allocation, and spilling into a single register allocator pass. The register allocator pass is then responsible for controlling its own sub-passes (not just the order of passes, but also termination conditions). Similarly, branch folding and if-conversion should be lumped together into a cfg optimization pass. Devang, hint hint.
We already have pass hierarchy, each pass manger is itself a full fledged pass. For example, when you insert a loop pass manger in a pass queue, you're inserting a function pass which responds to runOnFunction().
However, it seems you want dynamic pass queue manager. In loop optimizer, a pass can request reprocessing of a given loop by re-inserting loop into the queue. However this loop is reprocessed by all loop pass not just by few selected ones. One alternative way to achieve this is to create RAPassManager to group all these RA passes together. Yet another approach would be to keep track of followup passes but that would required significant modifications in PM because followup passes may require something that may not available and so on...
This doesn't necessarily need to be in the passmanager, but I'd like to see the code generator grow the ability to codegen an entire function multiple times. Ideally, we'd be able to codegen all the way to final machineinstrs in multiple configurations and then pick the best based on some metric. This would allow lots of interesting experiments and allow people willing to wait for really good code to get it. Once the "best" codegen for a function is done, the final machineinstrs would be sent to the asmprinter or code emitter.
It isn't clear to me how to best fit this into the passmanager model though.
This requires maintaining parallel pass queues in pass manager. Or use multiple instances of function pass manager for optimization+codegen and feed them cloned functions.