Lab 3: Code Generator for C/C++

Programming Language Technology, 2015


The objective of this assignment is to write a code generator from a fragment of the C++ programming language to JVM, Java Virtual Machine. The code generator should produce Java class files, which can be run in the Java bytecode interpreter so that they correctly perform all their input and output actions.

The code generator is partially characterized by compilation schemes in Chapter 6 of the PLT book. More JVM instructions are given in Appendix B. These explanations follow Jasmin assembler; the code generator may emit Jasmin assembly code into text files, which are then processed further by Jasmin to create class files. Jasmin can be downloaded here, but is also included in the test suite.

The recommended implementation is via a BNF grammar processed by the BNF Converter (BNFC) tool. The syntax tree created by the parser is first type checked by using the type checker created in Lab 2. The code generator should then make another pass of the type-checked code.

The approximate size of the grammar is 50 rules, and the code generator should be around 300-700 lines, depending on the programming language used. All BNFC supported languages can be used, but guidance is guaranteed only for Haskell and Java.

Before the work can be submitted, the program has to pass some tests, which are given on the course web page via links later in this document.

Restriction of the task

To pass the assignment, it is sufficient to treat the grammar of Lab 2 excluding doubles. Thus, operations involving type double need not be covered by the code generator.

However, for the ambitious student, we recommend to extend the code generator also to doubles (after it works for integers and the statements). The extension requires some extra work:

  1. The type checker needs to annotate the abstract syntax with types, such that the correct instruction (for int or double) can be picked by the code generator.
  2. As doubles are 64bit (whereas the other values have only 32bit), the environment has to record the size of local variables (single word or double word).


The recommended procedure is two passes:

  1. build a symbol table that for every function gives its type in a form usable for JVM code generation;
  2. compile the program by generating a class file containing every source code function as a method.

You can copy your grammar and the TypeChecker module from Lab 2 to the same directory.

Language specification

The language is the same as in Lab 2, and you can use the grammar file Also its type system is the same.

There are four built-in functions:

    void   printInt(int x)        // print an integer and a newline in standard output
    void   printDouble(double x)  // print a double and a newline in standard output
    int    readInt()              // read an integer from standard input
    double readDouble()           // read a double from standard input

These functions can be defined in a separate runtime class, which can be obtained e.g. from writing these functions in Java and compiling to a class file. A ready-made Java file is here.

Class structure

Boilerplate code, see PLT book, Chapter 6.


Methods in the class, main special, see Chapter 6.

Statements and expressions

The semantics is the same as in Lab 3. In other words, running the generated classes in java produces the same behaviour as running the source code in the lab2 interpreter.



  int main ()
    int arg = readInt() ;
    int ret = 1 ;
    int i = 1 ;
    while (i < arg + 1) {
      ret = i * ret ;
      ++i ;
    printInt(ret) ;

could compile to good03.j as follows:

  ;; Boilerplate: a wrapping class for cc code
  .class public good03
  .super java/lang/Object
  .method public <init>()V
    invokespecial java/lang/Object/<init>()V
  .end method
  ;; The java-style main method calls the cc main
  .method public static main([Ljava/lang/String;)V
  .limit locals 1
    invokestatic good03/main()I
  .end method
  ;; Program
  .method public static main()I
    .limit locals 3
    .limit stack 10
  ;; int arg = readInt();
  invokestatic Runtime/readInt()I
  ;; int ret = 1;
  ;; int i = 1;
  ;; while (i < arg + 1)
  L0:            ;; // beginning of loop, check condition
  iload_2        ;; i
  iadd           ;; arg + 1
  if_icmplt L2   ;; test i < arg + 1
  goto L3
  L2:            ;; i < arg + 1 is true
  L3:            ;; i < arg + 1 is false
  if_icmpeq L1   ;; if last comparison was false, exit while loop
  ;; ret = i * ret
  ;; ++i
  istore_2  ;; // i = i + 1
  ;; // continue loop
  goto L0
  ;; printInt(ret)
  invokestatic Runtime/printInt(I)V
  ;; // return 0
  .end method

(The comments are only for seeing the connection between .cc and .j).

Solution format

Input and output

The code generator must be a program called lab3, which is executed by the command

    lab3 <SourceFile>

and produces a class (.class) file. It may do this by first generating Jasmin assembly code (a .j file) and then calling Jasmin on this code. For help with building the Jasmin file, refer to the Java bytecode instruction listings. Jasmin can be called by

    java -jar jasmin.jar <File>.j

The generated class file should have the same name and be in the same directory as the original source file:

    lab3 ../a/b/

This should produce a class file ../a/b/c.class.

The output at failure is a code generator error, or a TYPE ERROR as in Assignment 2, or a SYNTAX ERROR as in Assignment 1.

The input can be read not only from user typing on the terminal, but also from standard input redirected from a file or by echo. For instance,

    ./java fibonacci <test-input
    echo 20 | java fibonacci

Example of success

Source file

  // file
  int main ()
    int i = readInt() ; //5
    printInt(i) ;   //5
    printInt(i++) ; //5
    printInt(i) ;   //6
    printInt(++i) ; //7
    printInt(i) ;   //7

Running the compiler

    % lab3
    # produces good.class
    % echo 3 | java good.class

Compiling the code generator

The compiler is submitted as source files that can be compiled by typing make.

Test programs

We also provide a test suite for lab 3. This test suite only contains test files which deal with integers; passing all these tests is the minimum required for passing this lab.

If you are ambitious, you can also try all the source files in the lab2 testsuite. For this you will also need to handle doubles in your compiler.

Run your compiler and the generated code with the testsuite before submitting the assignment.

Download the entire test suite folder or the tarball and compile it using make. The test script can then be run as follows:

    ./progs-test-lab3 ../path/to/solution/

Tip: Make sure your classnames in Jasmin have simple names without slashes or dots. If the first line of your Jasmin file is

  .class public x/y/

then Jasmin will compile it to a file


regardless of what the name of the .j file is. The easiest option, and also what the test suite expects, is that your class name is just a string without any slashes or dots (in this example, just "z").

Important: Make sure that your compiler waits for the Jasmin conversion to finish before exiting. In other words, it shouldn't exit before the .class file has been written. For example when using Haskell, you can do the following:

    p <- runCommand <call jasmin here>
    waitForProcess p

Success criteria

Your program must be compatible with the test suite and pass all the tests.

The solution must be written in an easily readable and maintainable way. In particular, tailoring it for the programs in the test suite is not maintainable!


Submit your lab by using Fire. Please include exactly all the files that are required for building your solution, including a Makefile. Do not however submit any generated files, and kindly avoid using archives (upload each file individually).

If you have any problems getting the test program to run, or if you think that there is an error in the test suite, contact the teachers via the course Google group.