[LLVMdev] Newbie question for registering new target with LLVM

Tim Northover t.p.northover at gmail.com
Fri Oct 12 04:28:40 PDT 2012


>  1.1) Please verify all above change(s) are OK? or I have modify some code?

Parts will almost certainly have to change in future, but if it builds
that's an excellent first step.

>  1.2) In LLVM, can we proceed our development like as GCC i.e. incremental
>       approach? In case of yes, please provide me any reference.

Incremental is definitely possible. I started out with the most
trivial of functions:

define void @foo() {
    ret void

and worked on from there in roughly this order:

+ Global variables (gives you guaranteed non-dead values to test
everything with and is likely simpler than getting procedure call ABIs
+ Simple arithmetic to make sure I'm not being completely insane in my
design decisions and get an idea of how the XXXInstrInfo.td will work
+ Stack spills and function prologue/epilogue.
+ Function calls and arguments.

After about the second stage I was implementing wide swathes of the
processor's instruction space, to give me the instructions needed to
support the more complicated details.

>  1.3) Please help me from anyone of llvm lovers :-) to proceed further. What
>       modifications required to generate rx assembly of hello word program.

As far as a backends go "Hello World" is not a small goal, though it
could be reached reasonably quickly if you're willing to put in hacks
that'll have to change later.

It's probably best to start trying to compile simple functions and
keep adding bits until that works. Anything you're not sure of but
that needs to be present for compilation, put an
llvm_unreachable("What's going on here?") so that you get to see
what's actually happening when it reaches that code and don't have to
guess what's needed.

Even compiling the simplest "define void @foo() { ret void }" is
likely to require quite a bit more infrastructure. Most of the files
in the existing backends are needed (possibly in a much simplified
form) for that function. I think the most important things you'll need
start in just two places and gradually depend on all the real
  + RxTargetMachine.cpp is small, but central. It's where you register
the passes that actually do the work. If you can get things to compile
with these functions doing their job you should have stubs for most of
the necessary components.
   + InstPrinter/XXXInstPrinter.cpp: This is the main entry for the
so-called MC instruction printing, which is the good way to do things.
It's theoretically possible to get your instructions printing without
creating this, but you'd be Doing It Wrong.

But most importantly, have fun and don't be afraid to ask more
detailed questions on the lists when you hit problems!


More information about the llvm-dev mailing list