HW3: LLVMlite


In this project you will implement a non-optimizing compiler for subset of the LLVM IR language. The source language consists of a 64-bit, simplified subset of the LLVM IR that we call LLVMlite. The target language is x86lite.

Getting Started

To get started, accept the assignment on EECS Gitlab and clone your team’s repository.

The files included in the repository are briefly described below. Those marked with * are the only ones you should need to modify while completing this assignment.


help about the main test harness


builds oatc, also supports targets ‘test’ and ‘zip’


the assertion framework


OS platform-specific compilation support


the X86lite instruction representation


the abstract syntax for LLVMlite


lexer for LLVMlite syntax


parser generator for LLVMlite syntax


reference interpreter for the LLVMlite semantics


example .ll programs used in testing


graded test cases that we provide


command-line interface


invoking the compiler pipeline


c code for testing interoperability


* where you implement the LLVMlite to X86 compiler


* where your additional test cases should go


You’ll need to have menhir and clang installed on your system for this assignment. If you have not already done so, follow the provided instructions to install them.

Preliminary Steps

  1. Skim through the rest of this web page to get a sense of what it contains.

  2. Familiarize yourself with the information in the README, which explains the ways that you can run your compiler for testing purposes.

  3. Then take a look at driver.ml, particularly the code related to process_ll_file to see how the backend code fits into the overall compilation pipeline.

  4. Then start working through backend.ml, following the instructions below.


This project is potentially very difficult to debug and may take you a while to understand. GET STARTED EARLY!

LLVM Lite Specification

The source language for this ‘backend’ part of the full compiler is a subset of the LLVM IR called LLVM Lite. You may find the Language Reference to be a useful resource for this project, though we are only concerned with a small portion of the full LLVM feature set.

The LLVMlite Documentation describes the behavior of LLVM programs in terms of an abstract semantics that is not target specific. This semantics is intended to be faithful to the LLVM Language Reference.

Implementing the Compiler

The code we provide in backend.ml is a minimal skeleton of the basic structure of the compiler. To a first approximation, for each part foo of the abstract syntax (such as prog or fdecl), there is a corresponding compile_foo function (i.e. compile_prog or compile_fdecl). Most of these definitions have been left unimplemented (and a few have been left out). Your job is to complete this translation. Our reference solution is well under 350 lines of documented code, so if your implementation is significantly longer than this, you may wish to rethink your approach or seek help.

The file backend.ml contains additional hints and explanations about the compilation strategy that we suggest you use.

We suggest that you stage the development of your compiler like this:

  1. First get a minimal implementation of compile_fdecl working so that you can compile functions with empty bodies but varying numbers of input parameters. To do this, you’ll need to understand the System V AMD64 ABI calling conventions (see the lecture slides and Wikipedia for an explanation), then understand the notion of a layout and complete the arg_loc function. At this point, the X86 code you generate won’t be able to run because the code for the compiled function does not exit propertly. (But you can still look at the generated assembly code to see whether it looks reasonable.)

  2. Next implement enough of the compile_terminator function to handle (void) functions that return no results. Similarly, implement enough of compile_block to handle blocks with no instructions. At this point, your compiler should be able to generate working code for an LLVM function like that found in returnvoid.ll:

    define void @main(i64 %argc, i8** %argv) {
      ret void

    (Note, this isn’t part of the test suite, since the value “returned” to the shell when this program runs isn’t well defined.)

  3. Understand the notion of the ctxt type and develop a strategy for storing uid locals. See the comments in the backend.ml file. Implement the compile_operand function.

  4. Implement the Binop case for compile_insn (which, if you follow the suggested method of compiling locals, will use compile_operand).

  5. At this point, you probably want to revisit compile_fdecl and compile_block to adjust them to deal properly with contexts and non-empty control-flow graphs / blocks.

  6. Next go back and implement the rest of the cases for compile_terminator. At this point, your compiler should be able to handle functions that return i64 values and that contain simple arithmetic and direct jumps.

  7. Implement the translation of Icmp in compile_insn, followed by Alloca, Load, and Store.

  8. Next tackle the Call instruction. The code you generate must properly handle the System V AMD64 ABI calling conventions (but note that we care only about 64-bit values). After successfully completing this step, your compiler should be able to handle the recursive factorial function definition.

  9. Breathe a sigh of relief at how easy it is to implement Bitcast, because the target language is untyped.

  10. Finally, gather your courage, and implement the Gep (getelementptr) instruction.

Testing and Debugging Strategies

Testing and debugging a compiler is quite difficult. There are many correct potential translations of a given source program, and there are many incidental changes (such as the choice of label names) that do not affect the semantics of the generated code. It is also difficult to test parts of the translation independently, since simple inputs may depend on almost all of the compilation pipeline.

The test harness provided by main.ml gives several ways to assess your code. See the README file for a full description of the flags.

We have provided a (minimally-featured) parser for LLVMlite code. It is sufficiently complete to parse the examples in the llprograms directory, and we expect you to create additional test cases yourself. For examples of how to use the test driver infrastructure, see the gradedtests.ml file.

You may find it helpful to run the LLVMlite code using our reference interpreter (with the --interpret-ll flag).

You may also find it helpful to run the LLVMlite code by compiling it via clang (with the --clang flag).

Note that it is not very useful to directly compare the .s files produced by your compiler to those produced by clang, but the behavior of the two versions for the same inputs should be the same.

Graded Test Cases

As part of this project, you must post an interesting test case for the compiler to Piazza. This test case might take the form of a .ll file along with expected outputs (as in our automated tests), or it might start from hand-generated LLVMlite abstract syntax.

The test case you submit to Piazza will not count if it is too similar to previously-posted tests! Your test should be distinct from prior test cases. (Note that this policy encourages you to submit test cases early!) Tests that stress parts of the language that aren’t well exercised by the provided tests are particularly encouraged.


Your submitted test should be easy to drop in to the testing harness: ideally, it’s a small amount of OCaml code, plus a single LLVMlite file. If your test case requires supporting C code (as some of our larger tests do), you can post that to Piazza along with your test.

We will validate these tests against our own implementation of the compiler (and clang). A second component of your grade will be determined by how your compiler fares against the test cases submitted by the other groups in the class.


Projects that do not compile will receive no credit!

Your team’s grade for this project will be based on:

  • 90 Points: the various automated tests that we provide. (Some reserved for online grading.)

  • 5 Points for posting an interesting test case to Piazza. (Graded manually.)

  • 5 Points divided among the test cases created by other groups. (Graded manually.)