[llvm-dev] SelectionDAGISel::Select's API considered harmful

[Justin Bogner via llvm-dev]

TLDR: Heads up for out of tree backends - you're going to need to update
your *DAGToDAGISel::Select method to unconditionally replace nodes
directly instead of returning the desired replacement.

So I'm working on fixing the undefined behaviour I described in
llvm.org/PR26808. As part of this, we need to stop looking into deleted
SDNodes to check if they were, in fact, deleted. A big place that we do
this is in SelectionDAGISel::DoInstructionSelection, where we can find
this helpful comment that came with the commit that added the check for
DELETED_NODE:

SelectionDAGISel.cpp says:
> // FIXME: This is pretty gross.  'Select' should be changed to not return
> // anything at all and this code should be nuked with a tactical strike.

I'm just gonna go ahead and take this advice.

I'll phase this in a couple of steps:

1. Rename Select to SelectImpl in all targets, and implement "virtual
   void Select(SDNode *)" in SelectionDAGISel. I'll move the current
   sketchy behaviour into this version of Select.

2. Update backends one at a time to implement "void Select(SDNode *)"
   instead of SelectImpl.

3. Make SelectionDAGISel::Select pure virtual and remove SelectImpl
   entirely.

If you have an out of tree backend and you merge from trunk, I recommend
updating in between steps 1 and 3 to avoid breakage after 3 happens.