Extended Description

TargetLowering expects the target to implement two types of virtual methods: LowerOperation and the
rest of the Lower* methods. The LowerOperation method is used by the legalizer, and the rest of the
Lower* methods are used at dag formation time. This is confusing and makes future work harder.

Instead of this, it would be nice for SelectionDAGISel to create a target-independent node for each of
the Lower* methods, and have LowerOperation do the actual implementation. For example, the
LowerCallTo method could return an new ISD::CALLTO node which has all of the operands and the
return value passed into the call: this would get expanded at legalize time. Another important case is
incoming arguments. This could be represented with an ISD::INCOMINGARGS node which has a token
chain input (entry node) and one result for each incoming argument to a function.

This is interesting for two reasons:

1. It gets rid of lowering during dag construction, making it easier to understand things for new people.

2. This allows the legalizer to hack on things before having the target lower the node. For example, if a
   target doesn't know how to pass a <16 x double> vector to a function but does know how to pass <4 x
   vector>'s, the legalizer would simplify it down before handing it off to the LowerOperation code.

3. This makes a future planned change more powerful: I want to be able to run the dag combiner on
   intermediate stages of legalization (e.g. after custom lower is done, but before other legalization has
   been run). This will allow us to catch things like the double->longlong->double conversion sequence
   in PPC.

-Chris