LLVM Tutorial 1: A First Function

- -

For starters, let's consider a relatively straightforward function that takes three integer parameters and returns an arithmetic combination of them. This is nice and simple, especially since it involves no control flow:

- -

-int mul_add(int x, int y, int z) {
-  return x * y + z;
-}
-

- -

As a preview, the LLVM IR weâre going to end up generating for this function will look like:

- -

-define i32 @mul_add(i32 %x, i32 %y, i32 %z) {
-entry:
-  %tmp = mul i32 %x, %y
-  %tmp2 = add i32 %tmp, %z
-  ret i32 %tmp2
-}
-

- -

If you're unsure what the above code says, skim through the LLVM Language Reference Manual and convince yourself that the above LLVM IR is actually equivalent to the original function. Once youâre satisfied with that, let's move on to actually generating it programmatically!

- -

Of course, before we can start, we need to #include the appropriate LLVM header files:

- -

-#include "llvm/Module.h"
-#include "llvm/Function.h"
-#include "llvm/PassManager.h"
-#include "llvm/CallingConv.h"
-#include "llvm/Analysis/Verifier.h"
-#include "llvm/Assembly/PrintModulePass.h"
-#include "llvm/Support/IRBuilder.h"
-#include "llvm/Support/raw_ostream.h"
-

- -

Now, let's get started on our real program. Here's what our basic main() will look like:

- -

-using namespace llvm;
-
-Module* makeLLVMModule();
-
-int main(int argc, char**argv) {
-  Module* Mod = makeLLVMModule();
-
-  verifyModule(*Mod, PrintMessageAction);
-
-  PassManager PM;
-  PM.add(createPrintModulePass(&outs()));
-  PM.run(*Mod);
-
-  delete Mod;
-  return 0;
-}
-

- -

The first segment is pretty simple: it creates an LLVM âmodule.â In LLVM, a module represents a single unit of code that is to be processed together. A module contains things like global variables, function declarations, and implementations. Here weâve declared a makeLLVMModule() function to do the real work of creating the module. Donât worry, weâll be looking at that one next!

- -

The second segment runs the LLVM module verifier on our newly created module. While this probably isnât really necessary for a simple module like this one, it's always a good idea, especially if youâre generating LLVM IR based on some input. The verifier will print an error message if your LLVM module is malformed in any way.

- -

Finally, we instantiate an LLVM PassManager and run -the PrintModulePass on our module. LLVM uses an explicit pass -infrastructure to manage optimizations and various other things. -A PassManager, as should be obvious from its name, manages passes: -it is responsible for scheduling them, invoking them, and ensuring the proper -disposal after weâre done with them. For this example, weâre just using a -trivial pass that prints out our module in textual form.

- -

Now onto the interesting part: creating and populating a module. Here's the -first chunk of our makeLLVMModule():

- -

-Module* makeLLVMModule() {
-  // Module Construction
-  Module* mod = new Module("test", getGlobalContext());
-

- -

Exciting, isnât it!? All weâre doing here is instantiating a module and giving it a name. The name isnât particularly important unless youâre going to be dealing with multiple modules at once.

- -

-  Constant* c = mod->getOrInsertFunction("mul_add",
-  /*ret type*/                           IntegerType::get(32),
-  /*args*/                               IntegerType::get(32),
-                                         IntegerType::get(32),
-                                         IntegerType::get(32),
-  /*varargs terminated with null*/       NULL);
-  
-  Function* mul_add = cast<Function>(c);
-  mul_add->setCallingConv(CallingConv::C);
-

- -

We construct our Function by calling getOrInsertFunction() on our module, passing in the name, return type, and argument types of the function. In the case of our mul_add function, that means one 32-bit integer for the return value and three 32-bit integers for the arguments.

- -

You'll notice that getOrInsertFunction() doesn't actually return a Function*. This is because getOrInsertFunction() will return a cast of the existing function if the function already existed with a different prototype. Since we know that there's not already a mul_add function, we can safely just cast c to a Function*. - -

In addition, we set the calling convention for our new function to be the C -calling convention. This isnât strictly necessary, but it ensures that our new -function will interoperate properly with C code, which is a good thing.

- -

-  Function::arg_iterator args = mul_add->arg_begin();
-  Value* x = args++;
-  x->setName("x");
-  Value* y = args++;
-  y->setName("y");
-  Value* z = args++;
-  z->setName("z");
-

- -

While weâre setting up our function, let's also give names to the parameters. This also isnât strictly necessary (LLVM will generate names for them if you donât specify them), but itâll make looking at our output somewhat more pleasant. To name the parameters, we iterate over the arguments of our function and call setName() on them. Weâll also keep the pointer to x, y, and z around, since weâll need them when we get around to creating instructions.

- -

Great! We have a function now. But what good is a function if it has no body? Before we start working on a body for our new function, we need to recall some details of the LLVM IR. The IR, being an abstract assembly language, represents control flow using jumps (we call them branches), both conditional and unconditional. The straight-line sequences of code between branches are called basic blocks, or just blocks. To create a body for our function, we fill it with blocks:

- -

-  BasicBlock* block = BasicBlock::Create(getGlobalContext(), "entry", mul_add);
-  IRBuilder<> builder(block);
-

- -

We create a new basic block, as you might expect, by calling its constructor. All we need to tell it is its name and the function to which it belongs. In addition, weâre creating an IRBuilder object, which is a convenience interface for creating instructions and appending them to the end of a block. Instructions can be created through their constructors as well, but some of their interfaces are quite complicated. Unless you need a lot of control, using IRBuilder will make your life simpler.

- -

-  Value* tmp = builder.CreateBinOp(Instruction::Mul,
-                                   x, y, "tmp");
-  Value* tmp2 = builder.CreateBinOp(Instruction::Add,
-                                    tmp, z, "tmp2");
-
-  builder.CreateRet(tmp2);
-  
-  return mod;
-}
-

- -

The final step in creating our function is to create the instructions that make it up. Our mul_add function is composed of just three instructions: a multiply, an add, and a return. IRBuilder gives us a simple interface for constructing these instructions and appending them to the âentryâ block. Each of the calls to IRBuilder returns a Value* that represents the value yielded by the instruction. Youâll also notice that, above, x, y, and z are also Value*'s, so it's clear that instructions operate on Value*'s.

- -

And that's it! Now you can compile and run your code, and get a wonderful textual print out of the LLVM IR we saw at the beginning. To compile, use the following command line as a guide:

- -

-# c++ -g tut1.cpp `llvm-config --cxxflags --ldflags --libs core` -o tut1
-# ./tut1
-

- -

The llvm-config utility is used to obtain the necessary GCC-compatible compiler flags for linking with LLVM. For this example, we only need the 'core' library. We'll use others once we start adding optimizers and the JIT engine.

- -Next: A More Complicated Function -

- -

Now that we understand the basics of creating functions in LLVM, let's move on to a more complicated example: something with control flow. As an example, let's consider Euclid's Greatest Common Denominator (GCD) algorithm:

- -

-unsigned gcd(unsigned x, unsigned y) {
-  if(x == y) {
-    return x;
-  } else if(x < y) {
-    return gcd(x, y - x);
-  } else {
-    return gcd(x - y, y);
-  }
-}
-

- -

With this example, we'll learn how to create functions with multiple blocks and control flow, and how to make function calls within your LLVM code. For starters, consider the diagram below.

- -

- -

This is a graphical representation of a program in LLVM IR. It places each basic block on a node of a graph and uses directed edges to indicate flow control. These blocks will be serialized when written to a text or bitcode file, but it is often useful conceptually to think of them as a graph. Again, if you are unsure about the code in the diagram, you should skim through the LLVM Language Reference Manual and convince yourself that it is, in fact, the GCD algorithm.

- -

The first part of our code is practically the same as from the first tutorial. The same basic setup is required: creating a module, verifying it, and running the PrintModulePass on it. Even the first segment of makeLLVMModule() looks essentially the same, except that gcd takes one fewer parameter than mul_add.

- -

-#include "llvm/Module.h"
-#include "llvm/Function.h"
-#include "llvm/PassManager.h"
-#include "llvm/Analysis/Verifier.h"
-#include "llvm/Assembly/PrintModulePass.h"
-#include "llvm/Support/IRBuilder.h"
-#include "llvm/Support/raw_ostream.h"
-
-using namespace llvm;
-
-Module* makeLLVMModule();
-
-int main(int argc, char**argv) {
-  Module* Mod = makeLLVMModule();
-  
-  verifyModule(*Mod, PrintMessageAction);
-  
-  PassManager PM;
-  PM.add(createPrintModulePass(&outs()));
-  PM.run(*Mod);
-
-  delete Mod;  
-  return 0;
-}
-
-Module* makeLLVMModule() {
-  Module* mod = new Module("tut2");
-  
-  Constant* c = mod->getOrInsertFunction("gcd",
-                                         IntegerType::get(32),
-                                         IntegerType::get(32),
-                                         IntegerType::get(32),
-                                         NULL);
-  Function* gcd = cast<Function>(c);
-  
-  Function::arg_iterator args = gcd->arg_begin();
-  Value* x = args++;
-  x->setName("x");
-  Value* y = args++;
-  y->setName("y");
-

- -

Here, however, is where our code begins to diverge from the first tutorial. Because gcd has control flow, it is composed of multiple blocks interconnected by branching (br) instructions. For those familiar with assembly language, a block is similar to a labeled set of instructions. For those not familiar with assembly language, a block is basically a set of instructions that can be branched to and is executed linearly until the block is terminated by one of a small number of control flow instructions, such as br or ret.

- -

Blocks correspond to the nodes in the diagram we looked at in the beginning of this tutorial. From the diagram, we can see that this function contains five blocks, so we'll go ahead and create them. Note that we're making use of LLVM's automatic name uniquing in this code sample, since we're giving two blocks the same name.

- -

-  BasicBlock* entry = BasicBlock::Create(getGlobalContext(), ("entry", gcd);
-  BasicBlock* ret = BasicBlock::Create(getGlobalContext(), ("return", gcd);
-  BasicBlock* cond_false = BasicBlock::Create(getGlobalContext(), ("cond_false", gcd);
-  BasicBlock* cond_true = BasicBlock::Create(getGlobalContext(), ("cond_true", gcd);
-  BasicBlock* cond_false_2 = BasicBlock::Create(getGlobalContext(), ("cond_false", gcd);
-

- -

Now we're ready to begin generating code! We'll start with the entry block. This block corresponds to the top-level if-statement in the original C code, so we need to compare x and y. To achieve this, we perform an explicit comparison using ICmpEQ. ICmpEQ stands for an integer comparison for equality and returns a 1-bit integer result. This 1-bit result is then used as the input to a conditional branch, with ret as the true and cond_false as the false case.

- -

-  IRBuilder<> builder(entry);
-  Value* xEqualsY = builder.CreateICmpEQ(x, y, "tmp");
-  builder.CreateCondBr(xEqualsY, ret, cond_false);
-

- -

Our next block, ret, is pretty simple: it just returns the value of x. Recall that this block is only reached if x == y, so this is the correct behavior. Notice that instead of creating a new IRBuilder for each block, we can use SetInsertPoint to retarget our existing one. This saves on construction and memory allocation costs.

- -

-  builder.SetInsertPoint(ret);
-  builder.CreateRet(x);
-

- -

cond_false is a more interesting block: we now know that x -!= y, so we must branch again to determine which of x -and y is larger. This is achieved using the ICmpULT -instruction, which stands for integer comparison for unsigned -less-than. In LLVM, integer types do not carry sign; a 32-bit integer -pseudo-register can be interpreted as signed or unsigned without casting. -Whether a signed or unsigned interpretation is desired is specified in the -instruction. This is why several instructions in the LLVM IR, such as integer -less-than, include a specifier for signed or unsigned.

- -

Also note that we're again making use of LLVM's automatic name uniquing, this time at a register level. We've deliberately chosen to name every instruction "tmp" to illustrate that LLVM will give them all unique names without getting confused.

- -

-  builder.SetInsertPoint(cond_false);
-  Value* xLessThanY = builder.CreateICmpULT(x, y, "tmp");
-  builder.CreateCondBr(xLessThanY, cond_true, cond_false_2);
-

- -

Our last two blocks are quite similar; they're both recursive calls to gcd with different parameters. To create a call instruction, we have to create a vector (or any other container with InputInterators) to hold the arguments. We then pass in the beginning and ending iterators for this vector.

- -

-  builder.SetInsertPoint(cond_true);
-  Value* yMinusX = builder.CreateSub(y, x, "tmp");
-  std::vector<Value*> args1;
-  args1.push_back(x);
-  args1.push_back(yMinusX);
-  Value* recur_1 = builder.CreateCall(gcd, args1.begin(), args1.end(), "tmp");
-  builder.CreateRet(recur_1);
-  
-  builder.SetInsertPoint(cond_false_2);
-  Value* xMinusY = builder.CreateSub(x, y, "tmp");
-  std::vector<Value*> args2;
-  args2.push_back(xMinusY);
-  args2.push_back(y);
-  Value* recur_2 = builder.CreateCall(gcd, args2.begin(), args2.end(), "tmp");
-  builder.CreateRet(recur_2);
-  
-  return mod;
-}
-

- -

And that's it! You can compile and execute your code in the same way as before, by doing:

- -

-# c++ -g tut2.cpp `llvm-config --cxxflags --ldflags --libs core` -o tut2
-# ./tut2
-

- -