Project on flex and bison

Question

I have used flex and bison in order to make a lexical analyzer and a parser for an EBNF grammar. This work is done! I mean, when i put a file with a program I write, I can see if the program has mistakes. If it doesn't, I can see the whole program in my screen based on the grammar i have used. I have no problem in this.

Now, I want to use loop handling and loop unrolling. Which part should I change? The lexical analyzer? The parser? Or the main after the parser? And how?

Answer 1

Introduction

As we don't have sight of a piece of your code to see how you are handling a loop in the parser and outputting code, and an example of a specific loop that you might want unrolled it is difficult to give any more detailed advice than that already given. There are unlikely to be any more experienced compiler writers or teachers anywhere on the globe than those already reading your question! So we will need to explore other ways to explain how to solve a problem like this.

It often happens that people can't post examples of their code because they started with a significant code base provided as part of a class exercise or from an open source repository, and they do not fully understand how it works to be able to find appropriate code fragments to post. Let's imagine that you had the complete source of a working compiler for a real language and wanted to add some loop optimisations to that existing, working compiler, you might then say, as you did, "what source, how can I show some source?" (because in actuality it is many tens of thousands of lines of code).

An Example Compiler

In the absence of some code to reference the alternative is to create one, as an exemplar, to explain the problem and solution. This is often how it is done in compiler text books or compiler classes. I will use a similar simple example to demonstrate how such optimisations can be achieved using the tools flex and bison.

First, we need to define the language of the example. To keep within the reasonable size constraints of a SO answer the language must be very simple. I will use simple assignments of expressions as the only statement form in my language. The variables in this language will be single letters and the constants will be positive integers. The only expression operator is plus (+). An example program in my language might be:

   i = j + k; j = 1 + 2

The output code generated by the compiler will be simple assembler for a single accumulator machine with four instructions, LDA, STO, ADD and STP. The code generated for the above statements would be:

LDA j
ADD k
STO i
LDA #1
ADD #2
STO j
STP

Where LDA loads a value or variable into the accumulator, ADD adds a variable or value to the accumulator, STO stores the accumulator back to a variable. STP is "stop" for the end-of-program.

The flex program

The language shown above will need the tokens for ID and NUMBER and should also skip whitespace. The following will suffice:

%{
#define yyterminate() return (END);
%}

digit [0-9]
id [a-z]
ws [\t\n\r ]

%%
{ws}+          /* Skip whitespace */
{digit}+      {yylval = (int)(0l - atol(yytext)); return(NUMBER); }
{id}          {yylval = yytext[0]; return(ID); }
"+"           {return('+'); }
"="           {return('='); }

_{Gory details
Just some notes on how this works. I've used atol to convert the integer to allow for deal with potential integer overflow that can occur in reading MAXINT. I'm negating the constants so they can be easily distinguished from the identifiers which will be positive in one byte. I'm storing single character identifiers to avoid having the burden of illustrating symbol table code and thus permit a very small lexer, parser and code generator.}

The bison program

To parse the language and generate some code from the bison actions we can achieve this by the following bison program:

%{
#include <stdio.h>
%}

%token NUMBER ID END
%%
program : statements END  { printf("STP\n"); return(0) ; }
        ;

statements : statement
        | statements ';' statement 
         ;

statement : ID '=' expression { printf("STO %c\n",$1); }
          |
          ;

expression : operand {
             /* Load operand into accumulator */
             if ($1 <= 0) 
                  printf("LDA #%d\n",(int)0l-$1);
             else printf("LDA %c\n",$1);
            }
           | expression '+' operand  {
             /* Add operand to accumulator */
             if ($3 <= 0) 
                  printf("ADD #%d\n",(int)0l-$3);
             else printf("ADD %c\n",$3);
            }
           ;

operand : NUMBER  
        | ID      
        ;

%%   
#include "lex.yy.c"

_{Explanation of methodology
This paragraph is intended for those who know how to do this and might query the approach used in my examples. I've deliberately avoided building a tree and doing a tree walk, although this would be the orthodox technique for code generation and optimisation. I wanted to avoid adding all the necessary code overhead in the example to manage the tree and walk it. This way my example compiler can be really tiny. However, being restricted to only using bison action to perform the code generation limits me to the ordering of the bison rule matching. This meant that only pseudo-machine code could really be generated. A source-to-source example would be less tractable with this methodology. I've chosen an idealised machine that is a cross between MU0 and a register-less PDP/11, again with the bare minimum of features to demonstrate some optimisations of code.}

Optimisation

Now we have a working compiler for a language in a few lines of code we can start to demonstrate how the process of adding code optimisation might work. As has already been said by the esteemed @Chris Dodd:

If you want to do program transformations after parsing, you should do them after parsing. You can do them incrementally (calling transform routines from your bison code after parsing part of your input), or after parsing is complete, but either way, they happen after parsing the part of the program you are transforming.

This compiler works by emitting code incrementally after parsing part of the input. As each statement is recognised the bison action (within the {...} clause) is invoked to generate code. If this is to be transformed into more optimal code it is this code that has to be changed to generate the desired optimisation. To be able to achieve effective optimisation we need a clear understanding of what language features are to be optimised and what the optimal transformation should be.

Constant Folding

A common optimisation (or code transformation) that can be done in a compiler is constant folding. In constant folding the compiler replaces expressions made entirely of numbers by the result. For example consider the following:

i = 1 + 2

An optimisation would be to treat this as:

i = 3

Thus the addition of 1 + 2 was made by the compiler and not put into the generated code to occur at run time. We would expect the following output to result:

LDA #3
STO i

Improved Code Generator

We can implement the improved code by looking for the explicit case where we have a NUMBER on both sides of expression '+' operand. To do this we have to delay taking any action on expression : operand to permit the value to be propagated onwards. As the value for an expression might not have been evaluated we have to potentially do that on assignment and addition, which makes for a slight explosion of if statements. We only need to change the actions for the rules statement and expression however, which are as shown below:

statement : ID '=' expression { 
              /* Check for constant expression */
              if ($3 <= 0) printf("LDA #%d\n",(int)0l-$3);
              else 
                 /* Check if expression in accumulator */
                 if ($3 != 'A') printf("LDA %c\n",$3);
              /* Now store accumulator */
              printf("STO %c\n",$1);
              }
          |   /* empty statement */
          ;

expression : operand { $$ = $1 ; }
           | expression '+' operand  {
             /* First check for constant expression */
             if ( ($1 <= 0) && ($3 <= 0)) $$ = $1 + $3 ;
             else { /* No constant folding */
                    /* See if $1 already in accumulator */
                    if ($1 != 'A')
                        /* Load operand $1 into accumulator */
                       if ($1 <= 0) 
                          printf("LDA #%d\n",(int)0l-$1);
                       else printf("LDA %c\n",$1);
                    /* Add operand $3 to accumulator */
                    if ($3 <= 0) 
                       printf("ADD #%d\n",(int)0l-$3);
                    else printf("ADD %c\n",$3);
                    $$ = 'A'; /* Note accumulator result */
                 }
            }
           ;

If you build the resultant compiler, you will see that it does indeed generate better code and perform the constant folding transformation.

Loop Unrolling

The transformation that you specifically asked about in your question was that of loop unrolling. In loop unrolling the compiler will look for some specific integer expression values in the loop start and end conditions to determine if the unrolled code transformation should be performed. The compiler can will then generate two possible code alternative sequences for loops, the unrolled and standard looping code. We can demonstrate this concept in this example mini-compiler by using integer increments.

If we imagine that the machine code has an INC instruction which increments the accumulator by one and is faster that performing an ADD #1 instruction, we can further improve the compiler by looking for that specific case. This involves evaluating integer constant expressions and comparing to a specific value to decide if an alternative code sequence should be used - just as in loop unrolling. For example:

i = j + 1

should result in:

LDA j
INC
STO i

Final Code Generator

To change the code generated for n + 1 we only need to recode part of the expression semantics and just test that when not folding constants wether the constant to be used would be 1 (which is negated in this example). The resultant code becomes:

expression : operand { $$ = $1 ; }
           | expression '+' operand  {
             /* First check for constant expression */
             if ( ($1 <= 0) && ($3 <= 0)) $$ = $1 + $3 ;
             else { /* No constant folding */
                    /* Check for special case of constant 1 on LHS */
                    if ($1 == -1) {
                        /* Swap LHS/RHS to permit INC usage */
                        $1 = $3;
                        $3 = -1;
                    }
                    /* See if $1 already in accumulator */
                    if ($1 != 'A')
                        /* Load operand $1 into accumulator */
                        if ($1 <= 0) 
                           printf("LDA #%d\n",(int)0l-$1);
                        else printf("LDA %c\n",$1);
                    /* Add operand $3 to accumulator */
                    if ($3 <= 0) 
                       /* test if ADD or INC */
                       if ($3 == -1) printf("INC\n");
                       else printf("ADD #%d\n",(int)0l-$3);
                    else printf("ADD %c\n",$3);
                    $$ = 'A'; /* Note accumulator result */
                 }
            }
           ;

Summary

In this mini-tutorial we have defined a whole language, a complete machine code, written a lexer, a compiler, a code generator and an optimiser. It has briefly demonstrated the process of code generation and indicated (albeit generally) how code transformation and optimisation could be performed. It should enable similar improvements to be made in other (as yet unseen) compilers, and has addressed the issue of identifying loop unrolling conditions and generating specific improvements for that case.

It should also have made it clear, how difficult it is to answer questions without specific examples of some program code to refer to.