Where is LLVM's interface for describing the instructions of an ISA?

Question

I'm new to LLVM and to compiler development in general, but I've finished reading Engineering a Compiler 3rd Edition and concluded compilers really may be a very fun area for me to specialize in, since I love delving into low level stuff, and just had the chance to join a compiler project. I'm looking for someone who knows LLVM, to point me to either documentation or LLVM source code files that define an ISA's instructions. I need to use LLVM's interface for defining instructions in a project to automatically issue all instructions of an ISA. I know .td files in a particular ISA's backend subdirectory, such as:

llvm-project/llvm/lib/Target/SystemZ/SystemZInstrInfo.td

and its corresponding .h and .cpp files have something to do with it, but it's my first time hearing about the TableGen language, and I don't know how actually relevant it is to LLVM's interface for providing the concrete definition of an ISA's set of instructions. Any pointers and bits of advise are welcome.

Answer 1

You are right, llvm-project/llvm/lib/Target/<Arch>/<Arch>InstrInfo.td file generally holds definition of ISA. Specifically, this file holds definitions of all the instructions in the ISA.

However, SystemZ file is written by hand and it is a bit difficult to follow because of irregular structure. Check out the corresponding Hexagon files.

A few words about instrtuction definitions

The first instruction from the file linked above is:

def A2_abs : HInst<
(outs IntRegs:$Rd32),
(ins IntRegs:$Rs32),
"$Rd32 = abs($Rs32)",
tc_d61dfdc3, TypeS_2op>, Enc_5e2823 {
let Inst{13-5} = 0b000000100;
let Inst{31-21} = 0b10001100100;
let hasNewValue = 1;
let opNewValue = 0;
let prefersSlot3 = 1;
}

• A2_abs is instruction name inside llvm
• HInst tblgen class, which holds all Hexagon instructions, it is defined here.
• outs - output params
• ins - input params
• IntRegs - register class (a set of all registers allowed to be used here), defined here.
• $Rd32 and $Rs32 - names of paramenters.
• "$Rd32 = abs($Rs32)" - assembler string, used for assembly printing.
• tc_d61dfdc3 - instruction itinary class, holds all scheduling constrains, related code is located in llvm-project/llvm/lib/Target/Hexagon/HexagonSchedule*.td.
• TypeS_2op - instruction subclass, Hexagon-specific.