Sequence 4.3 – LLVM Intermediate Representation


LLVM Types

LLVM defines first class types,

Type Meaning
void missing value
i1 1 bit integer (boolean)
i32 32 bits integer (int)
float 32 bits floating point

and derived types,

Type Meaning
i32* pointer to an int
void (i32) function taking an integer and returning void
{ i8, i32 } structure with two fields: a char and an int

LLVM Abstracts functions and function calls

The simple tiger program,


produces the following LLVM IR,

define i32 @main() {              ; defines the main function 
                                  ; returning an 32 bit int

  call void @__print_int(i32 42)  ; call to a primitive 
                                  ; returns void

  ret i32 0                       ; main returns 0


LLVM instructions with non-void return type can be assigned to a value

<result> = add <type> <left>, <right>

%result = add i32 3, 2 

Memory locations

The alloca instruction allocates memory space on the stack and returns a pointer

%ptr = alloca i32       ; %ptr is of type i32*

store and load instructions read from and write to a pointer

; Write 5 into the pointer. In C, *ptr = 5; 
store i32 5, i32* %ptr  

; Read the pointer content. In C, int content = *ptr;
%content = load i32, i32* %ptr

Local variables in LLVM

The simple way: alloca reserves room on the stack for the local variable, store and load accesses its content.

let function f(a: int, b: int): int =
  let var c := a + b in c end
define i32 @f(i32, i32) #0 {
  %c = alloca i32, align 4        ; %c <- address of c
  %3 = add i32 %0, %1             ; %3 <- a + b
  store i32 %3, i32* %c, align 4  ; Store %3 into c
  %4 = load i32, i32* %c, align 4 ; Load c into %4
  ret i32 %4                      ; and return it


Using alloca for all local variable does not seem to lead to efficient code. Fortunately, a later optimization pass called mem2reg (memory to register) will remove all useless alloca, store and load, and replace them with direct virtual register manipulation:

define i32 @f(i32, i32) #0 {
  %3 = add nsw i32 %1, %0
  ret i32 %3

which will give, in ARM thumb assembly:

    add r0, r1
    bx  lr

Branches and Labels

Code locations can be represented with labels. The branch br instruction jumps to another location.

; unconditional jump to label here
br label %here  

  ; conditional jump depends on the %condition value 
  br i1 %condition, label %when_true, label %when_false



Integer comparison

Integer comparison icmp operation returns an i1

%result = icmp ne i32 %a, %b ; true when %a <> %b

%result = icmp eq i32 %a, %b ; true when %a == %b

%result = icmp sge i32 %a, %b ; true when %a >= %b