Default Implementation for Pseudo Instructions

Hi LLVM-dev,

I’m working on XRay, and one of the parts of it is the introduction of a couple of pseudo instructions – one of which is PATCHABLE_RET, which wraps a return instruction (similar to how PATCHABLE_OP works).

In http://reviews.llvm.org/D19904, I’m trying to find a way of making PATCHABLE_RET have a default implementation in platforms where they’re not supported yet (i.e., not X86).

In my very limited research into this, I think I’ve narrowed it down to two places:

  • SelectionDAGTargetInfo has some hooks for specific instructions to be lowered/handled. A default implementation could be to just unpack the operands from the PATCHABLE_* instructions. In particular I’m looking at potentially adding some more functions to https://github.com/llvm-mirror/llvm/blob/master/include/llvm/CodeGen/SelectionDAGTargetInfo.h specifically for the PATCHABLE_* instructions, and finding where the SelectionDAG stuff is invoked to special-case the PATCHABLE_* instructions.

  • TargetLowering is another place where this might work. We can have special dispatch for the PATCHABLE_* instructions there, and also where we use a custom inserter. We can mark the instructions to use a custom inserter, have some overridable hooks in TargetLowering which by default just unpacks the operands into new MachineInstr instances, then override those in X86’s specific lowering. Requisite link: http://llvm.org/docs/doxygen/html/classllvm_1_1TargetLowering.html

Some questions:

  1. Is this even something I need to worry about? Since PATCHABLE_RET is being inserted by a late MachineFunctionPass, it could just do this for a white-listed set of platforms. This seems not ideal.

  2. Machine IR (mir) seems to be where PATCHABLE_RET might show up, and may need to be hand-crafted. Are the above two places the right places where these instructions could be intercepted, or am I missing something more fundamental?

Thanks in advance!