This document describes techniques for writing backends for LLVM which convert the LLVM representation to machine assembly code or other languages.

In general, you want to follow the format of X86 or PowerPC (in lib/Target).

To create a static compiler (one that emits text assembly), you need to implement the following:

• Describe the register set
  • Create a TableGen description of the register set and register classes
  • Implement a subclass of MRegisterInfo
• Describe the instruction set
  • Create a TableGen description of the instruction set
  • Implement a subclass of TargetInstrInfo
• Describe the target machine
  • Create a TableGen description of the target that describes the pointer size and references the instruction set
  • Implement a subclass of TargetMachine, which configures TargetData correctly
• Implement the assembly printer for the architecture. Usually, if you have described the instruction set with the assembly printer generator in mind, that step can be almost automated.

Now, for static code generation you also need to write an instruction selector for your platform: see lib/Target/*/*ISelSimple.cpp which is no longer "simple" but it gives you the idea: you have to be able to create MachineInstrs for any given LLVM instruction using the InstVisitor pattern, and produce a MachineFunction with MachineBasicBlocks full of MachineInstrs for a corresponding LLVM Function. Creating an instruction selector is perhaps the most time-consuming part of creating a back-end.

To create a JIT for your platform:

• Create a subclass of TargetJITInfo
• Create a machine code emitter that will be used to emit binary code directly into memory, given MachineInstrs

Note that lib/target/Skeleton is a clean skeleton for a new target, so you might want to start with that and adapt it for your target, and if you are wondering how things are done, peek in the X86 or PowerPC target.

The Skeleton target is non-functional but provides the basic building blocks you will need for your endeavor.

• TableGen register info description - describe a class which will store the register's number in the binary encoding of the instruction (e.g., for JIT purposes).

  You also need to define register classes to contain these registers, such as the integer register class and floating-point register class, so that you can allocate virtual registers to instructions from these sets, and let the target-independent register allocator automatically choose the actual architected registers.

  // class Register is defined in Target.td
  class TargetReg<string name> : Register<name> {
    let Namespace = "Target";
  }
  
  class IntReg<bits<5> num, string name> : TargetReg<name> {
    field bits<5> Num = num;
  }
  
  def R0 : IntReg<0, "%R0">;
  ...
  
  // class RegisterClass is defined in Target.td
  def IReg : RegisterClass<i64, 64, [R0, ... ]>;
  

• TableGen instruction info description - break up instructions into classes, usually that's already done by the manufacturer (see instruction manual). Define a class for each instruction category. Define each opcode as a subclass of the category, with appropriate parameters such as the fixed binary encoding of opcodes and extended opcodes, and map the register bits to the bits of the instruction which they are encoded in (for the JIT). Also specify how the instruction should be printed so it can use the automatic assembly printer, e.g.:

  // class Instruction is defined in Target.td
  class Form<bits<6> opcode, dag OL, string asmstr> : Instruction {
    field bits<42> Inst;
  
    let Namespace = "Target";
    let Inst{0-6} = opcode;
    let OperandList = OL;
    let AsmString = asmstr;
  }
  
  def ADD : Form<42, (ops IReg:$rD, IReg:$rA, IReg:$rB), "add $rD, $rA, $rB">;
  

For now, just take a look at lib/Target/CBackend for an example of how the C backend is written.

• Code generator - describes some of the classes in code generation at a high level, but it is not (yet) complete.
• TableGen fundamentals - describes how to use TableGen to describe your target information succinctly