LLVM Intermediate Representation (IR) is the heart of the LLVM project. In general every compiler produces an intermediate representation on which it runs most of its optimizations. For a compiler targeting multiple-source languages and different architectures the important decision while selecting an IR is that it should neither be of very high-level, as in very closely attached to the source language, nor it should be very low-level, as in close to the target machine instructions. LLVM IR aims to be a universal IR of a kind, by being at a low enough level that high-level ideas may be cleanly mapped to it. Ideally the LLVM IR should have been target-independent, but it is not so because of the inherent target dependence in some of the programming languages itself. For example, when using standard C headers in a Linux system, the header files itself are target dependent, which may specify a particular type to an entity so that it matches the system calls of the particular target architecture.

Most of the LLVM tools revolve around this Intermediate Representation. The frontends of different languages generate this IR from the high-level source language. The optimizer tool of LLVM runs on this generated IR to optimize the code for better performance and the code generator makes use of this IR for target specific code generation. This IR has three equivalent forms:

Now let's take an example to see how this LLVM IR looks like. We will take a small C code and convert it into LLVM IR using clang and try to understand the details of LLVM IR by mapping it back to the source language.

$ cat add.c
int globvar = 12;

int add(int a) {
return globvar + a;
}

Use the clang frontend with the following options to convert it to LLVM IR:

$ clang -emit-llvm -c -S add.c
$ cat add.ll
; ModuleID = 'add.c'
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

@globvar = global i32 12, align 4

; Function Attrs: nounwind uwtable
define i32 @add(i32 %a) #0 {
  %1 = alloca i32, align 4
  store i32 %a, i32* %1, align 4
  %2 = load i32, i32* @globvar, align 4
  %3 = load i32, i32* %1, align 4
  %4 = add nsw i32 %2, %3
  ret i32 %4
}

attributes #0 = { nounwind uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "unsafe-fp-math"="false" "use-soft-float"="false" }

!llvm.ident = !{!0}

Now let's look at the IR generated and see what it is all about. You can see the very first line giving the ModuleID, that it defines the LLVM module for add.c file. An LLVM module is a top–level data structure that has the entire contents of the input LLVM file. It consists of functions, global variables, external function prototypes, and symbol table entries.

The following lines show the target data layout and target triple from which we can know that the target is x86_64 processor with Linux running on it. The datalayout string tells us what is the endianess of machine ('e' meaning little endian), and the name mangling (m : e denotes elf type). Each specification is separated by '–'and each following spec gives information about the type and size of that type. For example, i64:64 says 64 bit integer is of 64 bits.

Then we have a global variable globvar. In LLVM IR all globals start with '@' and all local variables start with '%'. There are two main reasons why the variables are prefixed with these symbols. The first one being, the compiler won't have to bother about a name clash with reserved words, the other being that the compiler can come up quickly with a temporary name without having to worry about a conflict with symbol table conflicts. This second property is useful for representing the IR in static single assignment (SSA) from where each variable is assigned only a single time and every use of a variable is preceded by its definition. So, while converting a normal program to SSA form, we create a new temporary name for every redefinition of a variable and limit the range of earlier definition till this redefinition.

LLVM views global variables as pointers, so an explicit dereference of the global variable using load instruction is required. Similarly, to store a value, an explicit store instruction is required.

Local variables have two categories:

Now let's see how the add function is represented in LLVM IR. define i32 @add(i32 %a) is very similar to how functions are declared in C. It specifies the function returns integer type i32 and takes an integer argument. Also, the function name is preceded by '@', meaning it has global visibility.

Within the function is actual processing for functionality. Some important things to note here are that LLVM uses a three-address instruction, that is a data processing instruction, which has two source operands and places the result in a separate destination operand (%4 = add i32 %2, %3). Also the code is in SSA form, that is each value in the IR has a single assignment which defines the value. This is useful for a number of optimizations.

The attributes string that follows in the generated IR specifies the function attributes which are very similar to C++ attributes. These attributes are for the function that has been defined. For each function defined there is a set of attributes defined in the LLVM IR.

The code that follows the attributes is for the ident directive that identifies the module and compiler version.

Until now, we have understood what LLVM IR (human readable form) is and how it can be used to represent a high-level language. Now, we will take a look at some of the tools that LLVM provides so that we can play around with this IR converting to other formats and back again to the original form. Let's take a look at these tools one by one along with examples.

In this chapter, we looked into the modular design of LLVM: How it is used in the opt tool of LLVM, and how it is applicable across LLVM core libraries. Then we took a look into LLVM intermediate representation, and how various entities (variables, functions etc.) of a language are mapped to LLVM IR. In the last section, we discussed about some of the important LLVM tools, and how they can be used to transform the LLVM IR from one form to another.

In the next chapter, we will see how we can write a frontend for a language that can output LLVM IR using the LLVM machinery.