LLVM Overview

LLVM is a compiler infrastructure that can be used with many different programming languages and target architectures.

LLVM IR: A Common Intermediate Representation

• Separate front-ends for each programming language are built for each language to parse programs in the language and translate them to LLVM IR.
  • (Clang) C/C++ to LLVM IR.
  • D to LLVM IR
  • Haskell to LLVM IR
  • Swift to LLVM IR

• Separate back-ends for each target architecture translate the platform-independent IR into platform-specific assembly code.
  • LLVM IR to Intel Architecture 32-bit (x86).
  • LLVM IR to Intel/AMD 64-bit architecture.
  • LLVM IR to Power PC
  • LLVM IR to Cell BroadBand Engine (Playstation)
  • LLVM IR to ARM (mobile CPUs)
  • LLVM IR to Nvidia GPUs

• Reduces the problem of creating \(m \times n \) compilers (\(m\) front-ends, \(n\) back-ends) to \(m + n\).

LLVM IR: Not Just Internal

• Many compiler systems have internal intermediate representations.
• LLVM IR has a defined printable syntax and semantics as given by the LLVM Language Reference Manual.
• LLVM IR can actually be used to directly write programs!
  • The "Hello World" program in LLVM IR

Rich, Strongly-Typed IR

The first-class types represent values that can be stored in registers and computed by instructions.

• Integers of any width in bits.
  • i1 - 1 bit integers, used for Boolean values.
  • i8, i16, i32, i64 : common integer types
  • i128, i256: wide integers (using SIMD registers)
  • i43: 43-bit integers are syntactically valid, but not directly supported by typical back-ends.

• Floating point types
  • half: 16 bits
  • float: 32 bits
  • double: 64 bits
  • fp128: 128 bits

• Pointer Types
  • i8* byte pointers
  • double*

• Vectors of Integers, Floats
  • abstraction of SIMD registers, e.g., 128-bit SSE2 registers on Intel
  • treated as first class types
  • <16 x i8>: 16 8-bit integers
  • <8 x i16>: 8 16-bit integers
  • <2 x i64>
  • <4 x float>
  • <2 x i8*>

• More info on the LLVM Type System

Instructions

• The statements of the LLVM language are instructions operating on the first-class types.

• Most instructions are in 3-operand form: an operation on two register values produces a single result value.
  • %result = add i32 %x, %y ... add of 32-bit integers
  • %result = add <4 x i32> %x, %y ... SIMD add of 128-bit vectors

• Boolean instructions produce an i1 value.
  • %flag = icmp ult i16 %a, %b

• Control flow instructions are limited:
  • branch instructions to a single label
  • call instructions
  • ret instructions
  • conditional branch instructions to one of two labels depending on an i1 value

Test:
  %cond = icmp eq i32 %a, %b
  br i1 %cond, label %IfEqual, label %IfUnequal
IfEqual:
  ret i32 1
IfUnequal:
  ret i32 0

• Well-formed LLVM IR is in static-single assignment (SSA) form.
  • There is only one assignment (definition) of any given variable.
  • Phi-nodes are used to produce a single value from multiple paths.

LoopHeader: ...
Loop:       ; Infinite loop that counts from 0 on up...
  %loopvar = phi i32 [ 0, %LoopHeader ], [ %nextloopvar, %Loop ]
  %nextloopvar = add i32 %loopvar, 1
  br label %Loop