8. Kaleidoscope: Adding Debug Information

8.1. Chapter 8 Introduction

Welcome to Chapter 8 of the “Implementing a language with LLVM” tutorial. In chapters 1 through 7, we’ve built a decent little programming language with functions and variables. What happens if something goes wrong though, how do you debug your program?

Source level debugging uses formatted data that helps a debugger translate from binary and the state of the machine back to the source that the programmer wrote. In LLVM we generally use a format called DWARF. DWARF is a compact encoding that represents types, source locations, and variable locations.

The short summary of this chapter is that we’ll go through the various things you have to add to a programming language to support debug info, and how you translate that into DWARF.

Caveat: For now we can’t debug via the JIT, so we’ll need to compile our program down to something small and standalone. As part of this we’ll make a few modifications to the running of the language and how programs are compiled. This means that we’ll have a source file with a simple program written in Kaleidoscope rather than the interactive JIT. It does involve a limitation that we can only have one “top level” command at a time to reduce the number of changes necessary.

Here’s the sample program we’ll be compiling:

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

fib(10)

8.2. Why is this a hard problem?

Debug information is a hard problem for a few different reasons - mostly centered around optimized code. First, optimization makes keeping source locations more difficult. In LLVM IR we keep the original source location for each IR level instruction on the instruction. Optimization passes should keep the source locations for newly created instructions, but merged instructions only get to keep a single location - this can cause jumping around when stepping through optimized programs. Secondly, optimization can move variables in ways that are either optimized out, shared in memory with other variables, or difficult to track. For the purposes of this tutorial we’re going to avoid optimization (as you’ll see with one of the next sets of patches).

8.3. Ahead-of-Time Compilation Mode

To highlight only the aspects of adding debug information to a source language without needing to worry about the complexities of JIT debugging we’re going to make a few changes to Kaleidoscope to support compiling the IR emitted by the front end into a simple standalone program that you can execute, debug, and see results.

First we make our anonymous function that contains our top level statement be our “main”:

-    auto Proto = llvm::make_unique<PrototypeAST>("", std::vector<std::string>());
+    auto Proto = llvm::make_unique<PrototypeAST>("main", std::vector<std::string>());

just with the simple change of giving it a name.

Then we’re going to remove the command line code wherever it exists:

@@ -1129,7 +1129,6 @@ static void HandleTopLevelExpression() {
 /// top ::= definition | external | expression | ';'
 static void MainLoop() {
   while (1) {
-    fprintf(stderr, "ready> ");
     switch (CurTok) {
     case tok_eof:
       return;
@@ -1184,7 +1183,6 @@ int main() {
   BinopPrecedence['*'] = 40; // highest.

   // Prime the first token.
-  fprintf(stderr, "ready> ");
   getNextToken();

Lastly we’re going to disable all of the optimization passes and the JIT so that the only thing that happens after we’re done parsing and generating code is that the llvm IR goes to standard error:

@@ -1108,17 +1108,8 @@ static void HandleExtern() {
 static void HandleTopLevelExpression() {
   // Evaluate a top-level expression into an anonymous function.
   if (auto FnAST = ParseTopLevelExpr()) {
-    if (auto *FnIR = FnAST->codegen()) {
-      // We're just doing this to make sure it executes.
-      TheExecutionEngine->finalizeObject();
-      // JIT the function, returning a function pointer.
-      void *FPtr = TheExecutionEngine->getPointerToFunction(FnIR);
-
-      // Cast it to the right type (takes no arguments, returns a double) so we
-      // can call it as a native function.
-      double (*FP)() = (double (*)())(intptr_t)FPtr;
-      // Ignore the return value for this.
-      (void)FP;
+    if (!F->codegen()) {
+      fprintf(stderr, "Error generating code for top level expr");
     }
   } else {
     // Skip token for error recovery.
@@ -1439,11 +1459,11 @@ int main() {
   // target lays out data structures.
   TheModule->setDataLayout(TheExecutionEngine->getDataLayout());
   OurFPM.add(new DataLayoutPass());
+#if 0
   OurFPM.add(createBasicAliasAnalysisPass());
   // Promote allocas to registers.
   OurFPM.add(createPromoteMemoryToRegisterPass());
@@ -1218,7 +1210,7 @@ int main() {
   OurFPM.add(createGVNPass());
   // Simplify the control flow graph (deleting unreachable blocks, etc).
   OurFPM.add(createCFGSimplificationPass());
-
+  #endif
   OurFPM.doInitialization();

   // Set the global so the code gen can use this.

This relatively small set of changes get us to the point that we can compile our piece of Kaleidoscope language down to an executable program via this command line:

Kaleidoscope-Ch8 < fib.ks | & clang -x ir -

which gives an a.out/a.exe in the current working directory.

8.4. Compile Unit

The top level container for a section of code in DWARF is a compile unit. This contains the type and function data for an individual translation unit (read: one file of source code). So the first thing we need to do is construct one for our fib.ks file.

8.5. DWARF Emission Setup

Similar to the IRBuilder class we have a DIBuilder class that helps in constructing debug metadata for an llvm IR file. It corresponds 1:1 similarly to IRBuilder and llvm IR, but with nicer names. Using it does require that you be more familiar with DWARF terminology than you needed to be with IRBuilder and Instruction names, but if you read through the general documentation on the Metadata Format it should be a little more clear. We’ll be using this class to construct all of our IR level descriptions. Construction for it takes a module so we need to construct it shortly after we construct our module. We’ve left it as a global static variable to make it a bit easier to use.

Next we’re going to create a small container to cache some of our frequent data. The first will be our compile unit, but we’ll also write a bit of code for our one type since we won’t have to worry about multiple typed expressions:

static DIBuilder *DBuilder;

struct DebugInfo {
  DICompileUnit *TheCU;
  DIType *DblTy;

  DIType *getDoubleTy();
} KSDbgInfo;

DIType *DebugInfo::getDoubleTy() {
  if (DblTy.isValid())
    return DblTy;

  DblTy = DBuilder->createBasicType("double", 64, 64, dwarf::DW_ATE_float);
  return DblTy;
}

And then later on in main when we’re constructing our module:

DBuilder = new DIBuilder(*TheModule);

KSDbgInfo.TheCU = DBuilder->createCompileUnit(
    dwarf::DW_LANG_C, "fib.ks", ".", "Kaleidoscope Compiler", 0, "", 0);

There are a couple of things to note here. First, while we’re producing a compile unit for a language called Kaleidoscope we used the language constant for C. This is because a debugger wouldn’t necessarily understand the calling conventions or default ABI for a language it doesn’t recognize and we follow the C ABI in our llvm code generation so it’s the closest thing to accurate. This ensures we can actually call functions from the debugger and have them execute. Secondly, you’ll see the “fib.ks” in the call to createCompileUnit. This is a default hard coded value since we’re using shell redirection to put our source into the Kaleidoscope compiler. In a usual front end you’d have an input file name and it would go there.

One last thing as part of emitting debug information via DIBuilder is that we need to “finalize” the debug information. The reasons are part of the underlying API for DIBuilder, but make sure you do this near the end of main:

DBuilder->finalize();

before you dump out the module.

8.6. Functions

Now that we have our Compile Unit and our source locations, we can add function definitions to the debug info. So in PrototypeAST::codegen() we add a few lines of code to describe a context for our subprogram, in this case the “File”, and the actual definition of the function itself.

So the context:

DIFile *Unit = DBuilder->createFile(KSDbgInfo.TheCU.getFilename(),
                                    KSDbgInfo.TheCU.getDirectory());

giving us an DIFile and asking the Compile Unit we created above for the directory and filename where we are currently. Then, for now, we use some source locations of 0 (since our AST doesn’t currently have source location information) and construct our function definition:

DIScope *FContext = Unit;
unsigned LineNo = 0;
unsigned ScopeLine = 0;
DISubprogram *SP = DBuilder->createFunction(
    FContext, Name, StringRef(), Unit, LineNo,
    CreateFunctionType(Args.size(), Unit), false /* internal linkage */,
    true /* definition */, ScopeLine, DINode::FlagPrototyped, false);
F->setSubprogram(SP);

and we now have an DISubprogram that contains a reference to all of our metadata for the function.

8.7. Source Locations

The most important thing for debug information is accurate source location - this makes it possible to map your source code back. We have a problem though, Kaleidoscope really doesn’t have any source location information in the lexer or parser so we’ll need to add it.

struct SourceLocation {
  int Line;
  int Col;
};
static SourceLocation CurLoc;
static SourceLocation LexLoc = {1, 0};

static int advance() {
  int LastChar = getchar();

  if (LastChar == '\n' || LastChar == '\r') {
    LexLoc.Line++;
    LexLoc.Col = 0;
  } else
    LexLoc.Col++;
  return LastChar;
}

In this set of code we’ve added some functionality on how to keep track of the line and column of the “source file”. As we lex every token we set our current current “lexical location” to the assorted line and column for the beginning of the token. We do this by overriding all of the previous calls to getchar() with our new advance() that keeps track of the information and then we have added to all of our AST classes a source location:

class ExprAST {
  SourceLocation Loc;

  public:
    ExprAST(SourceLocation Loc = CurLoc) : Loc(Loc) {}
    virtual ~ExprAST() {}
    virtual Value* codegen() = 0;
    int getLine() const { return Loc.Line; }
    int getCol() const { return Loc.Col; }
    virtual raw_ostream &dump(raw_ostream &out, int ind) {
      return out << ':' << getLine() << ':' << getCol() << '\n';
    }

that we pass down through when we create a new expression:

LHS = llvm::make_unique<BinaryExprAST>(BinLoc, BinOp, std::move(LHS),
                                       std::move(RHS));

giving us locations for each of our expressions and variables.

From this we can make sure to tell DIBuilder when we’re at a new source location so it can use that when we generate the rest of our code and make sure that each instruction has source location information. We do this by constructing another small function:

void DebugInfo::emitLocation(ExprAST *AST) {
  DIScope *Scope;
  if (LexicalBlocks.empty())
    Scope = TheCU;
  else
    Scope = LexicalBlocks.back();
  Builder.SetCurrentDebugLocation(
      DebugLoc::get(AST->getLine(), AST->getCol(), Scope));
}

that both tells the main IRBuilder where we are, but also what scope we’re in. Since we’ve just created a function above we can either be in the main file scope (like when we created our function), or now we can be in the function scope we just created. To represent this we create a stack of scopes:

std::vector<DIScope *> LexicalBlocks;
std::map<const PrototypeAST *, DIScope *> FnScopeMap;

and keep a map of each function to the scope that it represents (an DISubprogram is also an DIScope).

Then we make sure to:

KSDbgInfo.emitLocation(this);

emit the location every time we start to generate code for a new AST, and also:

KSDbgInfo.FnScopeMap[this] = SP;

store the scope (function) when we create it and use it:

KSDbgInfo.LexicalBlocks.push_back(&KSDbgInfo.FnScopeMap[Proto]);

when we start generating the code for each function.

also, don’t forget to pop the scope back off of your scope stack at the end of the code generation for the function:

// Pop off the lexical block for the function since we added it
// unconditionally.
KSDbgInfo.LexicalBlocks.pop_back();

8.8. Variables

Now that we have functions, we need to be able to print out the variables we have in scope. Let’s get our function arguments set up so we can get decent backtraces and see how our functions are being called. It isn’t a lot of code, and we generally handle it when we’re creating the argument allocas in PrototypeAST::CreateArgumentAllocas.

DIScope *Scope = KSDbgInfo.LexicalBlocks.back();
DIFile *Unit = DBuilder->createFile(KSDbgInfo.TheCU.getFilename(),
                                    KSDbgInfo.TheCU.getDirectory());
DILocalVariable D = DBuilder->createParameterVariable(
    Scope, Args[Idx], Idx + 1, Unit, Line, KSDbgInfo.getDoubleTy(), true);

DBuilder->insertDeclare(Alloca, D, DBuilder->createExpression(),
                        DebugLoc::get(Line, 0, Scope),
                        Builder.GetInsertBlock());

Here we’re doing a few things. First, we’re grabbing our current scope for the variable so we can say what range of code our variable is valid through. Second, we’re creating the variable, giving it the scope, the name, source location, type, and since it’s an argument, the argument index. Third, we create an lvm.dbg.declare call to indicate at the IR level that we’ve got a variable in an alloca (and it gives a starting location for the variable), and setting a source location for the beginning of the scope on the declare.

One interesting thing to note at this point is that various debuggers have assumptions based on how code and debug information was generated for them in the past. In this case we need to do a little bit of a hack to avoid generating line information for the function prologue so that the debugger knows to skip over those instructions when setting a breakpoint. So in FunctionAST::CodeGen we add a couple of lines:

// Unset the location for the prologue emission (leading instructions with no
// location in a function are considered part of the prologue and the debugger
// will run past them when breaking on a function)
KSDbgInfo.emitLocation(nullptr);

and then emit a new location when we actually start generating code for the body of the function:

KSDbgInfo.emitLocation(Body);

With this we have enough debug information to set breakpoints in functions, print out argument variables, and call functions. Not too bad for just a few simple lines of code!

8.9. Full Code Listing

Here is the complete code listing for our running example, enhanced with debug information. To build this example, use:

# Compile
clang++ -g toy.cpp `llvm-config --cxxflags --ldflags --system-libs --libs core mcjit native` -O3 -o toy
# Run
./toy

Here is the code:

Next: Conclusion and other useful LLVM tidbits