Adding JIT and Optimizer Support

Welcome to Chapter 4 of the “Implementing a language with LLVM” tutorial. Chapters 1-3 described the implementation of a simple language and added support for generating LLVM IR. This chapter describes two new techniques: adding optimizer support to your language, and adding JIT compiler support. These additions will demonstrate how to get nice, efficient code for the Kaleidoscope language.

4.1 Trivial Constant Folding

Our demonstration for Chapter 3 is elegant and easy to extend. Unfortunately, it does not produce wonderful code. The IRBuilder, however, does give us obvious optimizations when compiling simple code:

  1. ready> def test(x) 1+2+x;
  2. Read function definition:
  3. define double @test(double %x) {
  4. entry:
  5.         %addtmp = fadd double 3.000000e+00, %x
  6.         ret double %addtmp
  7. }

This code is not a literal transcription of the AST built by parsing the input. That would be:

  1. ready> def test(x) 1+2+x;
  2. Read function definition:
  3. define double @test(double %x) {
  4. entry:
  5.         %addtmp = fadd double 2.000000e+00, 1.000000e+00
  6.         %addtmp1 = fadd double %addtmp, %x
  7.         ret double %addtmp1
  8. }

Constant folding, as seen above, in particular, is a very common and very important optimization: so much so that many language implementors implement constant folding support in their AST representation.

With LLVM, you don’t need this support in the AST. Since all calls to build LLVM IR go through the LLVM IR builder, the builder itself checked to see if there was a constant folding opportunity when you call it. If so, it just does the constant fold and return the constant instead of creating an instruction.

Well, that was easy :). In practice, we recommend always using IRBuilder when generating code like this. It has no “syntactic overhead” for its use (you don’t have to uglify your compiler with constant checks everywhere) and it can dramatically reduce the amount of LLVM IR that is generated in some cases (particular for languages with a macro preprocessor or that use a lot of constants).

On the other hand, the IRBuilder is limited by the fact that it does all of its analysis inline with the code as it is built. If you take a slightly more complex example:

  1. ready> def test(x) (1+2+x)*(x+(1+2));
  2. ready> Read function definition:
  3. define double @test(double %x) {
  4. entry:
  5.         %addtmp = fadd double 3.000000e+00, %x
  6.         %addtmp1 = fadd double %x, 3.000000e+00
  7.         %multmp = fmul double %addtmp, %addtmp1
  8.         ret double %multmp
  9. }

In this case, the LHS and RHS of the multiplication are the same value. We’d really like to see this generate tmp = x+3; result = tmp*tmp; instead of computing x+3 twice.

Unfortunately, no amount of local analysis will be able to detect and correct this. This requires two transformations: reassociation of expressions (to make the add’s lexically identical) and Common Subexpression Elimination (CSE) to delete the redundant add instruction. Fortunately, LLVM provides a broad range of optimizations that you can use, in the form of “passes”.

4.2 LLVM Optimization Passes

LLVM provides many optimization passes, which do many different sorts of things and have different tradeoffs. Unlike other systems, LLVM doesn’t hold to the mistaken notion that one set of optimizations is right for all languages and for all situations. LLVM allows a compiler implementor to make complete decisions about what optimizations to use, in which order, and in what situation.

As a concrete example, LLVM supports both “whole module” passes, which look across as large of body of code as they can (often a whole file, but if run at link time, this can be a substantial portion of the whole program). It also supports and includes “per-function” passes which just operate on a single function at a time, without looking at other functions. For more information on passes and how they are run, see the How to Write a Pass document and the List of LLVM Passes.

For Kaleidoscope, we are currently generating functions on the fly, one at a time, as the user types them in. We aren’t shooting for the ultimate optimization experience in this setting, but we also want to catch the easy and quick stuff where possible. As such, we will choose to run a few per-function optimizations as the user types the function in. If we wanted to make a “static Kaleidoscope compiler”, we would use exactly the code we have now, except that we would defer running the optimizer until the entire file has been parsed.

In order to get per-function optimizations going, we need to set up a FunctionPassManager to hold and organize the LLVM optimizations that we want to run. Once we have that, we can add a set of optimizations to run. The code looks like this:

  1. FunctionPassManager OurFPM(TheModule);
  3. // Set up the optimizer pipeline.  Start with registering info about how the
  4. // target lays out data structures.
  5. OurFPM.add(new DataLayout(*TheExecutionEngine->getDataLayout()));
  6. // Provide basic AliasAnalysis support for GVN.
  7. OurFPM.add(createBasicAliasAnalysisPass());
  8. // Do simple "peephole" optimizations and bit-twiddling optzns.
  9. OurFPM.add(createInstructionCombiningPass());
  10. // Reassociate expressions.
  11. OurFPM.add(createReassociatePass());
  12. // Eliminate Common SubExpressions.
  13. OurFPM.add(createGVNPass());
  14. // Simplify the control flow graph (deleting unreachable blocks, etc).
  15. OurFPM.add(createCFGSimplificationPass());
  17. OurFPM.doInitialization();
  19. // Set the global so the code gen can use this.
  20. TheFPM = &OurFPM;
  22. // Run the main "interpreter loop" now.
  23. MainLoop();

This code defines a FunctionPassManager, OurFPM. It requires a pointer to the Module to construct itself. Once it is set up, we use a series of add calls to add a bunch of LLVM passes. The first pass is basically boilerplate, it adds a pass so that later optimizations know how the data structures in the program are laid out. The TheExecutionEngine variable is related to the JIT, which we will get to in the next section.

In this case, we choose to add 4 optimization passes. The passes we chose here are a pretty standard set of “cleanup” optimizations that are useful for a wide variety of code. I won’t delve into what they do but, believe me, they are a good starting place :).

Once the PassManager is set up, we need to make use of it. We do this by running it after our newly created function is constructed (in FunctionAST::Codegen), but before it is returned to the client:

  1. if (Value *RetVal = Body->Codegen()) {
  2.   // Finish off the function.
  3.   Builder.CreateRet(RetVal);
  5.   // Validate the generated code, checking for consistency.
  6.   verifyFunction(*TheFunction);
  8.   // Optimize the function.
  9.   TheFPM->run(*TheFunction);
  11.   return TheFunction;
  12. }

As you can see, this is pretty straightforward. The FunctionPassManager optimizes and updates the LLVM Function* in place, improving (hopefully) its body. With this in place, we can try our test above again:

  1. ready> def test(x) (1+2+x)*(x+(1+2));
  2. ready> Read function definition:
  3. define double @test(double %x) {
  4. entry:
  5.         %addtmp = fadd double %x, 3.000000e+00
  6.         %multmp = fmul double %addtmp, %addtmp
  7.         ret double %multmp
  8. }

As expected, we now get our nicely optimized code, saving a floating point add instruction from every execution of this function.

LLVM provides a wide variety of optimizations that can be used in certain circumstances. Some documentation about the various passes is available, but it isn’t very complete. Another good source of ideas can come from looking at the passes that Clang runs to get started. The “opt” tool allows you to experiment with passes from the command line, so you can see if they do anything.

Now that we have reasonable code coming out of our front-end, lets talk about executing it!

4.3 Adding a JIT Compiler

Code that is available in LLVM IR can have a wide variety of tools applied to it. For example, you can run optimizations on it (as we did above), you can dump it out in textual or binary forms, you can compile the code to an assembly file (.s) for some target, or you can JIT compile it. The nice thing about the LLVM IR representation is that it is the “common currency” between many different parts of the compiler.

In this section, we’ll add JIT compiler support to our interpreter. The basic idea that we want for Kaleidoscope is to have the user enter function bodies as they do now, but immediately evaluate the top-level expressions they type in. For example, if they type in 1 + 2;, we should evaluate and print out 3. If they define a function, they should be able to call it from the command line.

In order to do this, we first declare and initialize the JIT. This is done by adding a global variable and a call in main:

  1. static ExecutionEngine *TheExecutionEngine;
  2. ...
  3. int main() {
  4.   ..
  5.   // Create the JIT.  This takes ownership of the module.
  6.   TheExecutionEngine = EngineBuilder(TheModule).create();
  7.   ..
  8. }

This creates an abstract Execution Engine which can be either a JIT compiler or the LLVM interpreter. LLVM will automatically pick a JIT compiler for you if one is available for your platform, otherwise it will fall back to the interpreter.

Once the ExecutionEngine is created, the JIT is ready to be used. There are a variety of APIs that are useful, but the simplest one is the getPointerToFunction(F) method. This method JIT compiles the specified LLVM Function and returns a function pointer to the generated machine code. In our case, this means that we can change the code that parses a top-level expression to look like this:

  1. static void HandleTopLevelExpression() {
  2.   // Evaluate a top-level expression into an anonymous function.
  3.   if (FunctionAST *F = ParseTopLevelExpr()) {
  4.     if (Function *LF = F->Codegen()) {
  5.       LF->dump();  // Dump the function for exposition purposes.
  7.       // JIT the function, returning a function pointer.
  8.       void *FPtr = TheExecutionEngine->getPointerToFunction(LF);
  10.       // Cast it to the right type (takes no arguments, returns a double) so we
  11.       // can call it as a native function.
  12.       double (*FP)() = (double (*)())(intptr_t)FPtr;
  13.       fprintf(stderr, "Evaluated to %f\n", FP());
  14.     }

Recall that we compile top-level expressions into a self-contained LLVM function that takes no arguments and returns the computed double. Because the LLVM JIT compiler matches the native platform ABI, this means that you can just cast the result pointer to a function pointer of that type and call it directly. This means, there is no difference between JIT compiled code and native machine code that is statically linked into your application.

With just these two changes, lets see how Kaleidoscope works now!

  1. ready> 4+5;
  2. Read top-level expression:
  3. define double @0() {
  4. entry:
  5.   ret double 9.000000e+00
  6. }
  8. Evaluated to 9.000000

Well this looks like it is basically working. The dump of the function shows the “no argument function that always returns double” that we synthesize for each top-level expression that is typed in. This demonstrates very basic functionality, but can we do more?

  1. ready> def testfunc(x y) x + y*2;
  2. Read function definition:
  3. define double @testfunc(double %x, double %y) {
  4. entry:
  5.   %multmp = fmul double %y, 2.000000e+00
  6.   %addtmp = fadd double %multmp, %x
  7.   ret double %addtmp
  8. }
  10. ready> testfunc(4, 10);
  11. Read top-level expression:
  12. define double @1() {
  13. entry:
  14.   %calltmp = call double @testfunc(double 4.000000e+00, double 1.000000e+01)
  15.   ret double %calltmp
  16. }
  18. Evaluated to 24.000000

This illustrates that we can now call user code, but there is something a bit subtle going on here. Note that we only invoke the JIT on the anonymous functions that call testfunc, but we never invoked it on testfunc itself. What actually happened here is that the JIT scanned for all non-JIT’d functions transitively called from the anonymous function and compiled all of them before returning from getPointerToFunction().

The JIT provides a number of other more advanced interfaces for things like freeing allocated machine code, rejit’ing functions to update them, etc. However, even with this simple code, we get some surprisingly powerful capabilities - check this out (I removed the dump of the anonymous functions, you should get the idea by now :) :

  1. ready> extern sin(x);
  2. Read extern:
  3. declare double @sin(double)
  5. ready> extern cos(x);
  6. Read extern:
  7. declare double @cos(double)
  9. ready> sin(1.0);
  10. Read top-level expression:
  11. define double @2() {
  12. entry:
  13.   ret double 0x3FEAED548F090CEE
  14. }
  16. Evaluated to 0.841471
  18. ready> def foo(x) sin(x)*sin(x) + cos(x)*cos(x);
  19. Read function definition:
  20. define double @foo(double %x) {
  21. entry:
  22.   %calltmp = call double @sin(double %x)
  23.   %multmp = fmul double %calltmp, %calltmp
  24.   %calltmp2 = call double @cos(double %x)
  25.   %multmp4 = fmul double %calltmp2, %calltmp2
  26.   %addtmp = fadd double %multmp, %multmp4
  27.   ret double %addtmp
  28. }
  30. ready> foo(4.0);
  31. Read top-level expression:
  32. define double @3() {
  33. entry:
  34.   %calltmp = call double @foo(double 4.000000e+00)
  35.   ret double %calltmp
  36. }
  38. Evaluated to 1.000000

Whoa, how does the JIT know about sin and cos? The answer is surprisingly simple: in this example, the JIT started execution of a function and got to a function call. It realized that the function was not yet JIT compiled and invoked the standard set of routines to resolve the function. In this case, there is no body defined for the function, so the JIT ended up calling dlsym("sin") on the Kaleidoscope process itself. Since sin is defined within the JIT’s address space, it simply patches up calls in the module to call the libm version of sin directly.

The LLVM JIT provides a number of interfaces (look in the ExecutionEngine.h file) for controlling how unknown functions get resolved. It allows you to establish explicit mappings between IR objects and addresses (useful for LLVM global variables that you want to map to static tables, for example), allows you to dynamically decide on the fly based on the function name, and even allows you to have the JIT compile functions lazily the first time they’re called.

One interesting application of this is that we can now extend the language by writing arbitrary C++ code to implement operations. For example, if we add:

  1. /// putchard - putchar that takes a double and returns 0.
  2. extern "C"
  3. double putchard(double X) {
  4.   putchar((char)X);
  5.   return 0;
  6. }

Now we can produce simple output to the console by using things like: extern putchard(x); putchard(120);, which prints a lowercase x on the console (120 is the ASCII code for x). Similar code could be used to implement file I/O, console input, and many other capabilities in Kaleidoscope.

This completes the JIT and optimizer chapter of the Kaleidoscope tutorial. At this point, we can compile a non-Turing-complete programming language, optimize and JIT compile it in a user-driven way. Next up we’ll look into extending the language with control flow constructs, tackling some interesting LLVM IR issues along the way.