Difference between XXXISelDAGToDAG and XXXISelLowering in llvm

Interesting point because learning the backend code structure is not trivial and each target is implemented in a different subfolder under the Target with several .cpp, .h, and .td files.

Since following is mentioned in your provided link:

"Also write code in XXXISelLowering.cpp to replace or remove operations and data types that are not supported natively in a SelectionDAG."

Let's take this forward. I would answer with a bigger picture showing basic steps prior to instruction selection, starting with the LLVM IR step that is to the top-left and then citing Getting Started with LLVM Core Libraries:

First, a SelectionDAGBuilder instance (see SelectionDAGISel.cpp for details) visits every function and creates a SelectionDAG object for each basic block. During this process, some special IR instructions such as call and ret already need targetspecific idioms—for instance, how to pass call arguments and how to return from a function—to be transformed into SelectionDAG nodes. To solve this issue, the algorithms in the TargetLowering class are used for the first time. This class is an abstract interface that each target must implement, but also has plenty of common functionality used throughout all backends.

To implement this abstract interface, each target declares a TargetLowering subclass named <Target>TargetLowering. Each target also overloads methods that implement how a specific target-independent, high-level node should be lowered to a level closer to the one of this machine. As expected, only a small subset of nodes must be lowered in this way, while the majority of the others are matched and replaced at instruction selection. For instance, in SelectionDAG from sum. bc, the X86TargetLowering::LowerReturn() method (see lib/Target/X86/ X86ISelLowering.cpp) is used to lower the IR ret instruction. While doing this, it generates the X86ISD::RET_FLAG node, which copies the function result to EAX—a target-specific way to handle the function return.