In previous chapters, we saw how high-level language instructions, statements, logical blocks, function calls, loops, and so on get transformed into the LLVM IR. Various optimization passes then process the IR to make it more optimal. The IR generated is in the SSA form and, in abstract format, almost independent of any high- or low-level language constraints, which facilitates optimization passes running on it. There might be some optimizations that are target-specific and take place later, when the IR gets converted into machine instructions.
After we get an optimal LLVM IR, the next phase is to convert it into target-machine-specific instructions. LLVM uses the SelectionDAG approach to convert the IR into machine instructions. The Linear IR is converted into SelectionDAG, a DAG that represents instructions as nodes. The SDAG then goes through various phases:
The SelectionDAG is created out of LLVM IR
Legalizing SDAG nodes
DAG combine optimization
Instruction...