Compilation Encapsulation

Compilation Encapsulation

Or: Why Every Component Should Just Do Its Damn Job

“when a negative int literal (e.g. -5) appears in the code, should it be a single integer token whose value is -5 or as two tokens, minus and an integer whose value is 5?”

“when a negative int literal (e.g. -5) appears in the code, should it be a single integer token whose value is -5 or as two tokens, minus and an integer whose value is 5?”

Well, in theory…

We can write a lexer (maybe not with flex) with lookbehind, that makes sure the last token was neither a number nor a variable. (Or a function call, or a field reference. Pretty complicated lexer.)

But just because we can do itdoes that make it a good idea?

But what if we change the syntax?

Professor Moriarty wants IC to be more like Matlab. He asks you to support support scalar operations on arrays of scalars.

array – scalar = [a1-scalar, … an-scalar]

And suddenly new int[n] - 6 is valid…

Generic compiler structure

Executable code

exe

Sourcetext

txtSemantic

RepresentationBackend(synthesi

s)

Compiler

Frontend(analysis)

Executable code

exe

Sourcetext

txtSemantic

RepresentationBackend(synthesi

s)

Compiler

Frontend(analysis)

ICProgram

ic

x86 executable

exeLexicalAnalysi

s

Syntax Analysi

sParsing

AST Symbol

Tableetc.

Inter.Rep.(IR)

CodeGeneration

IC compiler

Encapsulation, what does it mean?

• It means each component needs to do its job, without regard for what the other components are doing.

• The tokenizer only cares about dividing the stream into tokens– Invalid characters– Keywords– Strings and comments

• The parser only cares about building a structure out of tokens– Assumes a valid stream of tokens– Structural rules with no meaning

• The semantic checker is free to only worry about semantics– Assumes a valid AST– Actually worries about meaning

Fake Exam Question #1

Professor Xavier wants IC to be more like Python. He asks you to support array and string multiplication.

"abc"*3 “abcabcabc”(new MyClass[5] * 2).length 10

But suppose…

• Suppose you decided to define your strings like so:

<YYINITIAL> \" { //move to <STRING> state to handle content yybegin(STRING); in_string_literal = true; }

<STRING>{

\"/{VALID_STR_POSTFIX} { //found the end of a string, finish. yybegin(YYINITIAL); in_string_literal = false; return new Token (sym.QUOTE,yyline + 1, string.toString()); } \" { throw new LexicalError(yyline+1); } //longest token only if invalid ahead

But suppose…

• Now we have to go back and fix the lexer, too.• When, in reality, there was no real reason to

perform that test:– There’s no case of something after the string the

syntax won’t be able to cope with.

Back to the tokenizer not caring

• What needs changing?• Lexer:– Nothing

• Syntax:– Nothing

• Semantic checks– Type check

• Code generation– Functionality of the operation

Fake Exam Question #2

• We want IC to support binary numeric literals– With the following syntax: 0b010010101 (leading

zeros after the binary signifier allowed)– With the same range restrictions

Solution #1

• We’ll add a new lexer token type, BINNUMBER– 0b[01]+

• And a new syntax rule for a BinNumber literal– Which, really, is only BINNUMBER

• And then check its range– Which is actually a lot easier than with decimals…

A short interlude: where does X go?

• Is property X lexical, syntactic or semantic?• Two main deciding factors

1. Correctness:Is there enough data to make the call right now?

2. Laziness:What will be gained by doing this right now?Is this the place where it’s easiest to do?

Example A: Range of decimal literals

• Correctness:– In any two's complement implementation of

integers, the bound is not going to be symmetric.– So we can’t make the call until we know if we have

a positive or a negative number on our hand…• Laziness:– Writing a lot of code that looks at the child

expressions during syntax is usually a bad sign.

Example B: range of binary literals

• Correctness:– All the data is there the second we got the token.

• Laziness:– Postponing the check means a continued

separation between binary and decimal literals – If we check right now, we can convert the value to

a number and forget all about it

Back to Fake Question #2

• So we can actually do it this way:• We’ll add a new lexer rule– 0b[01]+– We’ll also check the range here– And then! – return new Token (sym.INTEGER,yyline + 1,

bin2decimal(yytext()));

Where does Y go?

• Place the following property:call to method foo() is a static call.

• Our guiding principle here is correctness:

Lexer

Syntax?

Where does Y go?

• Syntax breaks methods up into three types:1.ClassName.foo() – definitely static2.varname.foo() – definitely not

So… correct?3.foo() - ???

So… not syntax.

Fake Question #3

• We want to allow type inference in IC

var a = new A();A b = a;C c = a; //type error

Q3: Lexer

• New token type VAR

Q3: Syntax

• We want an init expression whose type is VAR– Do we add VAR to types?– No, we treat it like void.

• How about AST representation?– We modify our LocalVariable class to keep “TBD”

in its type

Q3: Semantics

• To determine the new variable’s type:– instead of computing its type field (which is TBD) – compute the type of the expression

• Put that value into the symbol table, and all else is business as usual!

Good luck on the exam!