Welcome to Part 5 of the "Implementing a language with LLVM" tutorial. Parts 1-4 described the implementation of the simple Kaleidoscope language and included support for generating LLVM IR, following by optimizations and a JIT compiler. Unfortunately, as presented, Kaleidoscope is mostly useless: it has no control flow other than call and return. This means that you can't have conditional branches in the code, significantly limiting its power. In this episode of "build that compiler", we'll extend Kaleidoscope to have an if/then/else expression plus a simple looping construct.

Extending Kaleidoscope to support if/then/else is quite straight-forward. It basically requires adding lexer support for this "new" concept to the lexer, parser, AST, and LLVM code emitter. This example is nice, because it shows how easy it is to "grow" a language over time, incrementally extending it as new ideas are discovered.

Before we get going on "how" we do this extension, lets talk about what we want. The basic idea is that we want to be able to write this sort of thing:

def fib(x)
  if x < 3 then
    1
  else
    fib(x-1)+fib(x-2);

In Kaleidoscope, every construct is an expression: there are no statements. As such, the if/then/else expression needs to return a value like any other. Since we're using a mostly functional form, we'll have it evaluate its conditional, then return the 'then' or 'else' value based on how the condition was resolved. This is very similar to the C "?:" expression.

The semantics of the if/then/else expression is that it first evaluates the condition to a boolean equality value: 0.0 is false and everything else is true. If the condition is true, the first subexpression is evaluated and returned, if the condition is false, the second subexpression is evaluated and returned. Since Kaleidoscope allows side-effects, this behavior is important to nail down.

Now that we know what we want, lets break this down into its constituent pieces.

The lexer extensions are straight-forward. First we add new enum values for the relevant tokens:

  // control
  tok_if = -6, tok_then = -7, tok_else = -8,

Once we have that, we recognize the new keywords in the lexer, pretty simple stuff:

    ...
    if (IdentifierStr == "def") return tok_def;
    if (IdentifierStr == "extern") return tok_extern;
    if (IdentifierStr == "if") return tok_if;
    if (IdentifierStr == "then") return tok_then;
    if (IdentifierStr == "else") return tok_else;
    return tok_identifier;

To represent the new expression we add a new AST node for it:

/// IfExprAST - Expression class for if/then/else.
class IfExprAST : public ExprAST {
  ExprAST *Cond, *Then, *Else;
public:
  IfExprAST(ExprAST *cond, ExprAST *then, ExprAST *_else)
    : Cond(cond), Then(then), Else(_else) {}
  virtual Value *Codegen();
};

The AST node just has pointers to the various subexpressions.

Now that we have the relevant tokens coming from the lexer and we have the AST node to build, our parsing logic is relatively straight-forward. First we define a new parsing function:

/// ifexpr ::= 'if' expression 'then' expression 'else' expression
static ExprAST *ParseIfExpr() {
  getNextToken();  // eat the if.
  
  // condition.
  ExprAST *Cond = ParseExpression();
  if (!Cond) return 0;
  
  if (CurTok != tok_then)
    return Error("expected then");
  getNextToken();  // eat the then
  
  ExprAST *Then = ParseExpression();
  if (Then == 0) return 0;
  
  if (CurTok != tok_else)
    return Error("expected else");
  
  getNextToken();
  
  ExprAST *Else = ParseExpression();
  if (!Else) return 0;
  
  return new IfExprAST(Cond, Then, Else);
}

Next we hook it up as a primary expression:

static ExprAST *ParsePrimary() {
  switch (CurTok) {
  default: return Error("unknown token when expecting an expression");
  case tok_identifier: return ParseIdentifierExpr();
  case tok_number:     return ParseNumberExpr();
  case '(':            return ParseParenExpr();
  case tok_if:         return ParseIfExpr();
  }
}

Now that we have it parsing and building the AST, the final piece is adding LLVM code generation support. This is the most interesting part of the if/then/else example, because this is where it starts to introduce new concepts. All of the code above has been described in previous chapters fairly thoroughly.

To motivate the code we want to produce, lets take a look at a simple example. Consider:

extern foo();
extern bar();
def baz(x) if x then foo() else bar();

If you disable optimizations, the code you'll (soon) get from Kaleidoscope looks like this:

declare double @foo()

declare double @bar()

define double @baz(double %x) {
entry:
	%ifcond = fcmp one double %x, 0.000000e+00
	br i1 %ifcond, label %then, label %else

then:		; preds = %entry
	%calltmp = call double @foo()
	br label %ifcont

else:		; preds = %entry
	%calltmp1 = call double @bar()
	br label %ifcont

ifcont:		; preds = %else, %then
	%iftmp = phi double [ %calltmp, %then ], [ %calltmp1, %else ]
	ret double %iftmp
}

To visualize the control flow graph, you can use a nifty feature of the LLVM 'opt' tool. If you put this LLVM IR into "t.ll" and run "llvm-as < t.ll | opt -analyze -view-cfg", a window will pop up and you'll see this graph:

Another way to get this is to call "F->viewCFG()" or "F->viewCFGOnly()" (where F is a "Function*") either by inserting actual calls into the code and recompiling or by calling these in the debugger. LLVM has many nice features for visualizing various graphs.

Coming back to the generated code, it is fairly simple: the entry block evaluates the conditional expression ("x" in our case here) and compares the result to 0.0 with the "fcmp one" instruction ('one' is "ordered and not equal"). Based on the result of this expression, the code jumps to either the "then" or "else" blocks, which contain the expressions for the true/false case.

Once the then/else blocks is finished executing, they both branch back to the else block to execute the code that happens after the if/then/else. In this case the only thing left to do is to return to the caller of the function. The question then becomes: how does the code know which expression to return?

The answer to this question involves an important SSA operation: the Phi operation. If you're not familiar with SSA, the wikipedia article is a good introduction and there are various other introductions to it available on your favorite search engine. The short version is that "execution" of the Phi operation requires "remembering" which block control came from. The Phi operation takes on the value corresponding to the input control block. In this case, if control comes in from the "then" block, it gets the value of "calltmp". If control comes from the "else" block, it gets the value of "calltmp1".

At this point, you are probably starting to think "on no! this means my simple and elegant front-end will have to start generating SSA form in order to use LLVM!". Fortunately, this is not the case, and we strongly advise not implementing an SSA construction algorithm in your front-end unless there is an amazingly good reason to do so. In practice, there are two sorts of values that float around in code written in your average imperative programming language that might need Phi nodes:

1. Code that involves user variables: x = 1; x = x + 1;
2. Values that are implicit in the structure of your AST, such as the phi node in this case.

At a future point in this tutorial ("mutable variables"), we'll talk about #1 in depth. For now, just believe me that you don't need SSA construction to handle them. For #2, you have the choice of using the techniques that we will describe for #1, or you can insert Phi nodes directly if convenient. In this case, it is really really easy to generate the Phi node, so we choose to do it directly.

Okay, enough of the motivation and overview, lets generate code!

In order to generate code for this, we implement the Codegen method for IfExprAST:

Value *IfExprAST::Codegen() {
  Value *CondV = Cond->Codegen();
  if (CondV == 0) return 0;
  
  // Convert condition to a bool by comparing equal to 0.0.
  CondV = Builder.CreateFCmpONE(CondV, 
                                ConstantFP::get(Type::DoubleTy, APFloat(0.0)),
                                "ifcond");

This code is straight-forward and similar to what we saw before. We emit the expression for the condition, then compare that value to zero to get a truth value as a 1-bit (bool) value.

  Function *TheFunction = Builder.GetInsertBlock()->getParent();
  
  // Create blocks for the then and else cases.  Insert the 'then' block at the
  // end of the function.
  BasicBlock *ThenBB = new BasicBlock("then", TheFunction);
  BasicBlock *ElseBB = new BasicBlock("else");
  BasicBlock *MergeBB = new BasicBlock("ifcont");

  Builder.CreateCondBr(CondV, ThenBB, ElseBB);

This code creates the basic blocks that are related to the if/then/else statement, and correspond directly to the blocks in the example above. The first line of this gets the current Function object that is being built. It gets this by asking the builder for the current BasicBlock, and asking that block for its "parent" (the function it is currently embedded into).

Once it has that, it creates three blocks. Note that it passes "TheFunction" into the constructor for the "then" block. This causes the constructor to automatically insert the new block onto the end of the specified function. The other two blocks are created, but aren't yet inserted into the function.

Once the blocks are created, we can emit the conditional branch that chooses between them. Note that creating new blocks does not implicitly affect the LLVMBuilder, so it is still inserting into the block that the condition went into. Also note that it is creating a branch to the "then" block and the "else" block, even though the "else" block isn't inserted into the function yet. This is all ok: it is the standard way that LLVM supports forward references.

  // Emit then value.
  Builder.SetInsertPoint(ThenBB);
  
  Value *ThenV = Then->Codegen();
  if (ThenV == 0) return 0;
  
  Builder.CreateBr(MergeBB);
  // Codegen of 'Then' can change the current block, update ThenBB for the PHI.
  ThenBB = Builder.GetInsertBlock();

After the conditional branch is inserted, we move the builder to start inserting into the "then" block. Strictly speaking, this call moves the insertion point to be at the end of the specified block. However, since the "then" block is empty, it also starts out by inserting at the beginning of the block. :)

Once the insertion point is set, we recursively codegen the "then" expression from the AST. To finish off the then block, we create an unconditional branch to the merge block. One interesting (and very important) aspect of the LLVM IR is that it requires all basic blocks to be "terminated" with a control flow instruction such as return or branch. This means that all control flow, including fall throughs must be made explicit in the LLVM IR. If you violate this rule, the verifier will emit an error.

The final line here is quite subtle, but is very important. The basic issue is that when we create the Phi node in the merge block, we need to set up the block/value pairs that indicate how the Phi will work. Importantly, the Phi node expects to have an extry for each predecessor of the block in the CFG. Why then are we getting the current block when we just set it to ThenBB 5 lines above? The problem is that the "Then" expression may actually itself change the block that the Builder is emitting into if, for example, it contains a nested "if/then/else" expression. Because calling Codegen recursively could arbitrarily change the notion of the current block, we are required to get an up-to-date value for code that will set up the Phi node.

  // Emit else block.
  TheFunction->getBasicBlockList().push_back(ElseBB);
  Builder.SetInsertPoint(ElseBB);
  
  Value *ElseV = Else->Codegen();
  if (ElseV == 0) return 0;
  
  Builder.CreateBr(MergeBB);
  // Codegen of 'Else' can change the current block, update ElseBB for the PHI.
  ElseBB = Builder.GetInsertBlock();

Code generation for the 'else' block is basically identical to codegen for the 'then' block. The only significant difference is the first line, which adds the 'else' block to the function. Recall previously that the 'else' block was created, but not added to the function. Now that the 'then' and 'else' blocks are emitted, we can finish up with the merge code:

  // Emit merge block.
  TheFunction->getBasicBlockList().push_back(MergeBB);
  Builder.SetInsertPoint(MergeBB);
  PHINode *PN = Builder.CreatePHI(Type::DoubleTy, "iftmp");
  
  PN->addIncoming(ThenV, ThenBB);
  PN->addIncoming(ElseV, ElseBB);
  return PN;
}

The first two lines here are now familiar: the first adds the "merge" block to the Function object (it was previously floating, like the else block above). The second block changes the insertion point so that newly created code will go into the "merge" block. Once that is done, we need to create the PHI node and set up the block/value pairs for the PHI.

Finally, the CodeGen function returns the phi node as the value computed by the if/then/else expression. In our example above, this returned value will feed into the code for the top-level function, which will create the return instruction.

Overall, we now have the ability to execution conditional code in Kaleidoscope. With this extension, Kaleidoscope is a fairly complete language that can calculate a wide variety of numeric functions. Next up we'll add another useful expression that is familiar from non-functional languages...

Now that we know how to add basic control flow constructs to the language, we have the tools to add more powerful things. Lets add something more aggressive, a 'for' expression:

 # print 100 '*' (ascii 42) characters
 extern putchard(char)
 for x = 1, x < 100, 1.0 in putchard(42);

This expression defines a new variable ("x" in this case) which iterates from a starting value, while the condition ("x < 100" in this case) is true, incrementing by an optional step value ("1.0" in this case). If the step value is omitted, it defaults to 1.0. While the loop is true, it executes its body expression. Because we don't have anything better to return, we'll just define the loop as always returning 0.0. In the future when we have mutable variables, it will get more useful.

As before, lets talk about the changes that we need to Kaleidoscope to support this.

The lexer extensions are the same sort of thing as for if/then/else:

  ... in enum Token ...
  // control
  tok_if = -6, tok_then = -7, tok_else = -8,
  tok_for = -9, tok_in = -10

  ... in gettok ...
  if (IdentifierStr == "def") return tok_def;
  if (IdentifierStr == "extern") return tok_extern;
  if (IdentifierStr == "if") return tok_if;
  if (IdentifierStr == "then") return tok_then;
  if (IdentifierStr == "else") return tok_else;
  if (IdentifierStr == "for") return tok_for;
  if (IdentifierStr == "in") return tok_in;
  return tok_identifier;

The AST node is similarly simple. It basically boils down to capturing the variable name and the consituent expressions in the node.

/// ForExprAST - Expression class for for/in.
class ForExprAST : public ExprAST {
  std::string VarName;
  ExprAST *Start, *End, *Step, *Body;
public:
  ForExprAST(const std::string &varname, ExprAST *start, ExprAST *end,
             ExprAST *step, ExprAST *body)
    : VarName(varname), Start(start), End(end), Step(step), Body(body) {}
  virtual Value *Codegen();
};

The parser code is also fairly standard. The only interesting thing here is handling of the optional step value. The parser code handles it by checking to see if the second comma is present. If not, it sets the step value to null in the AST node:

/// forexpr ::= 'for' identifer '=' expr ',' expr (',' expr)? 'in' expression
static ExprAST *ParseForExpr() {
  getNextToken();  // eat the for.

  if (CurTok != tok_identifier)
    return Error("expected identifier after for");
  
  std::string IdName = IdentifierStr;
  getNextToken();  // eat identifer.
  
  if (CurTok != '=')
    return Error("expected '=' after for");
  getNextToken();  // eat '='.
  
  
  ExprAST *Start = ParseExpression();
  if (Start == 0) return 0;
  if (CurTok != ',')
    return Error("expected ',' after for start value");
  getNextToken();
  
  ExprAST *End = ParseExpression();
  if (End == 0) return 0;
  
  // The step value is optional.
  ExprAST *Step = 0;
  if (CurTok == ',') {
    getNextToken();
    Step = ParseExpression();
    if (Step == 0) return 0;
  }
  
  if (CurTok != tok_in)
    return Error("expected 'in' after for");
  getNextToken();  // eat 'in'.
  
  ExprAST *Body = ParseExpression();
  if (Body == 0) return 0;

  return new ForExprAST(IdName, Start, End, Step, Body);
}

Now we get to the good part: the LLVM IR we want to generate for this thing.

Here is the complete code listing for our running example, enhanced with the if/then/else and for expressions.. To build this example, use:

   # Compile
   g++ -g toy.cpp `llvm-config --cppflags --ldflags --libs core jit native` -O3 -o toy
   # Run
   ./toy

Here is the code:

...