Skip to content

💽 A compiler for Java--, a subset of the Java programming language

Notifications You must be signed in to change notification settings

diogolhc/feup-l.eic-comp-jmm-compiler

Repository files navigation

Compilers Project

GROUP: COMP2022-1A

Name Number Grade Contribution
Catarina Oliveira Pires 201907925 20 25%
Diogo Luís Henriques Costa 201906731 20 25%
Francisco José Barbosa Marques Colino 201905405 20 25%
Pedro Miguel Jesus da Silva 201907523 20 25%

GLOBAL Grade of the project: 20

SUMMARY:

This tool compiles J-- source files to Java bytecode.

To compile and install the program, run gradle installDist.
After compiled, the program can be run using comp2022-1A.bat (Windows) or comp2022-1A (Linux).

Program invocation:

./comp2022-1a [-r=<num>] [-o] [-d] -i=<input_file.jmm>

Where:

  • -r option directs the program to only use at most <num> local variables of the JVM.
    • If num = -1, it will use as many as needed.
    • If num = 0, it will use the minimum possible.
  • -o option directs the program to perform code optimizations.
  • -d option enables additional output related to the compilation process.
  • -i is used to pass the path to the J-- source file to compile.

The program will parse the input arguments to check if they are valid.

If so, it will continue to parse the input_file.jmm, build the symbol table, perform semantic analysis, optimize at the AST level if option -o was used, generate intermediate code (OLLIR), perform register allocation if option -r was used and generate JVM code. Then it proceeds to save in the generated-files directory the .j file with the Jasmin code, and a compiled .class file.

Note that at any stage, if errors are detected the compilation process aborts and the error is displayed.

SEMANTIC ANALYSIS:

Symbol Table

  • Has information about imports and the declared class
  • Has information about extends, fields and methods
  • Has information about the parameters and local variables of each method

Type Verification

  • Verify if variable names used in the code have a corresponding declaration, either as a local variable, a method parameter or a field of the class (if applicable)
  • Declarations must be of a known type or class
  • Operand types of an operation must be compatible with the operation (e.g. int + boolean is an error because + expects two integers)
  • Array cannot be used in arithmetic operations (e.g. array1 + array2 is an error)
  • Array access is done over an array
  • Array access index is an expression of type integer
  • Array can only be of type integer
  • Array size on initialization must be of type integer
  • Type of the assignee must be compatible with the assigned (an_int = a_bool is an error)
  • Expressions in conditions must return a boolean (if(2+3) is an error)

Function Verification

  • When calling methods of the class declared in the code, verify if the types of arguments of the call are compatible with the types in the method declaration
  • In case the method does not exist, verify if the class extends another class and report an error if it does not. Assume the method exists in one of the super classes, and that is being correctly called
  • When calling methods that belong to other classes other than the class declared in the code, verify if the classes are being imported
  • Length can only be called on arrays
  • Literals can't call methods

Main Verification

  • Reference variable this can't be used in main

CODE GENERATION:

Abstract Syntax Tree

The source code is parsed accordingly to the grammar defined in JmmGrammar.jj.
From the parsing stage, results an AST with several annotations such as name of ids and values of literals.

Optimizations (-o)

After the AST is generated, and before OLLIR is generated, three visitors are executed in order while they are changing the AST (loop can be seen here: JmmOptimizer.java):

  1. ConstPropagationVisitor.java
  2. ConstFoldingVisitor.java
  3. DeadCodeEliminationVisitor.java

The first substitutes local variables that can be replaced by constant values.
The second performs constant folding in not operations and binary operations, including short-cut evaluation of the && operation. The third checks if if and while conditions are immediate values and if so eliminates the code accordingly. Note that in the case of < condition, if its operands are the same id, the condition evaluates to false, and dead code can be eliminated.

Also, additional care is taken in order to detect the value of while loop conditions at the start of execution, so that during OLLIR generation it can be checked if while loops can be promoted to do-whiles, avoiding the need of an extra goto instruction and label.

Example:

int i;
i = 0;

while (i < 10) {
    i = i + 1;
}

Becomes,

whileBody1:
    i.i32 :=.i32 i.i32 +.i32 1.i32;
    if (i.i32 <.bool 10.i32) goto whileBody1;

This is done by duplicating the while condition before the optimization loop: WhileConditionDuplicatorVisitor.java.
And after checking its value on which the optimizations were performed to check if its true: DoWhileAnnotatorVisitor.java.

The rationale for choosing this particular order of optimizations is that if we first propagate constants on the expressions and then fold them, we can detect ifs and whiles whose conditions allow inlining/deleting them.

Example:

Before:

public int optimizeThis(boolean a) {
    boolean b;
    boolean c;
    boolean d;
    
    b = false;
    c = false;
    d = true;

    // b's value was propagated
    if (b) {
        io.println(0);
    } else {
        io.println(1);
    }

    // a's value is not known but we can still 
    // simplify with short-cut evaluation
    if ((c && d) && a) {
        io.println(2);
    } else {
        io.println(3);
    }

    return 0;
}

after:

public int optimizeThis(boolean a) {
    boolean b;
    boolean c;
    boolean d;
    
    b = false;
    c = false;
    d = true;

    // The ifs were removed and the elses were enlined 
    io.println(1);
    io.println(3);

    return 0;
}

OO-based Low-Level Intermediate Representation (OLLIR)

OLLIR code is generated by visiting the AST. This intermediate representation translates the AST to a 3 address code representation. This translation is performed by OllirGenerator.java.

Special care was taken in order for the added temporary variables do not conflict with the user defined variables.

Optimizations (-o)

During the generation of OLLIR, elimination of unnecessary GOTOs is performed accordingly to the annotation generated as previously stated in while loops.

Optimization (-r)

With the -r option the user can limit the maximum number of local variables a method may use. The optimization applies a graph colouring heuristic on the interference graph (at LocalVariableInterferenceGraph.java ) to allocate the variables to JVM local variables. This graph is created by performing a liveness analysis at LivenessAnalyser.java on the control flow graph. If the algorithm cannot colour the graph with the provided limit, the compiler reports an error with the required number of local variables.

The optimization is performed at LocalVariableOptimization.java.

Example:

The code:

public int soManyRegisters(int arg){
   int a;
   int b;
   int c;
   int d;
   a = 0;
   b = a;
   c = b;
   d = c;
   return d;
}

The JVM generated without the optimization:

// Using as many local variables as possible
.method public soManyRegisters(I)I
	.limit stack 1
	.limit locals 6
	iconst_0
	istore_2
	iload_2
	istore_3
	iload_3
	istore 4
	iload 4
	istore 5
	iload 5
	ireturn
.end method

The JVM generated with the optimization:

// With just one local variable the value 0 can be propagated to the return
.method public soManyRegisters(I)I
	.limit stack 1
	.limit locals 3
	iconst_0
	istore_2
	iload_2
	istore_2
	iload_2
	istore_2
	iload_2
	istore_2
	iload_2
	ireturn
.end method

JASMIN

Jasmin code is generated from the OLLIR code in JasminBackender.java.
Everything is properly translated. Also stack limits and local limits are being calculated.

Optimizations

At this level, instruction selection is performed:

  • x = x + <const> => iinc x <const>
  • if (x < 0) => iflt
  • if (0 < x) => ifgt
  • if (!x) => ifeq
  • if (x) => ifne
  • if (x < y) => if_icmplt

Also, load and store instructions were selected:
e.g. iload_0, bipush 10, iconst_0, ...

PROS:

  • Everything proposed was implemented, including both -o and -r optimizations as described above.

  • The [-d] option is very verbose which is very useful when trying to understand the compilation process.

  • The compiler is highly user-friendly, giving useful information on syntactic and semantic error reports such as line, column and readable messages.

  • Many optimizations were implemented:

    • Constant propagation
    • Elimination of unnecessary GOTOs (promotion of whiles to do-whiles)
    • Use of optimized instructions
    • Constant folding (with also shortcut evaluation was performed on && as shown in this test: ShortCutAnd.jmm)
    • Dead code elimination
    • If in-lining

CONS:

  • We didn't sanitize variable names that are reserved in OLLIR like "array".

About

💽 A compiler for Java--, a subset of the Java programming language

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages