Metadata

  • Source
  • Zotero: View Item
  • Type: VideoRecording
  • Title: Introduction to LLVM,
  • Year: 2019

Slides: Coming Soon — Introduction to LLVM. — Videos Filmed & Edited by Bash Films: http://www.BashFilms.com

Annotations

Notes

Source Layout

Each subdirectory in LLVM is a subproject (e.g. clang, llvm, debugger, linker, etc). Each subproject has the same layout (include, lib, tests, util)

  • clang: language front-end
  • llvm: LLVM core (middle-end, back-end)

To build:

cmake ~/src/llvm-project/llvm -DLLVM_ENABLE_PROJECTS='clang;openmp'
make -j

debug builds are default (and slow & big), consider changing build type ccache can save a lot of time if you rebuild a lot ninja speed things up for rebuilds (updating files and recompiling)

Useful options:

Compilers 101

After clang AST, LLVM internally uses control flow graphs.

Life of a program

C LLVM-IR (.ll) (human readable IR through -S)

-O3 adds some information like type aliasing; otherwise clang just do -O0, which adds attributes to every function that says “no inlining and do not optimize”, which breaks opt

LLVM generates stack allocations for communication across basic blocks; we force it to generate another form We want stuff to live on registers, so we run memory to register in opt instnamer gives the instruction a name if they don’t (avoid numeric instructions); Release builds lose name information

LLVM-IR

LLVM module ~= translation unit data layout encodes target/architecture-dependent information like pointer size, address spaces, structure padding, etc target triple = architecture (x86_64), ??? (unknown in this case), OS (linux-gnu)

<type> %var e.g. i64 %var: function parameter %var = ...: variable declaration/initialization add nsw i64 %a, %b: add instruction, no signed wrapping, for i64 ints. semicolon = colon #n e.g. #0 at the end of function header = attributes defined elsewhere %var = phi i64 [ %var1, %branch1 ], [ %var2, %branch2 ] ...: select variable depending on which branch was executed attribute #0 = { nounwind uwtable} (nounwind: won’t throw exception and unwind the stack, uwtable: ??); an attribute can be applied to a specific call, a function parameter, entire function, return instruction, etc instruction can also take additional flags

lines starting with exclamation point = metadata information

SSA: static single assignment; all variables in LLVM-IR can only be assigned once; this limitation enables various optimization SSA holds even if the IR uses another representation (stack allocations)

LLVM-IR Hierarchy

  • llvm::Module
    • llvm:GlobalVariable
    • llvm:Function (declarations / definitions)
      • llvm::BasicBlock (if a function contains basic blocks, it’s a definition; if not, then declaration)
        • llvm::Instruction
          • llvm::ICmpInst (compare)
          • llvm::BranchInst

LLVM-IR instructions

Each basic block contains a terminator instruction at the end

LLVM-IR (cont)

IR is in SSA-form with infinite registers

organization:

  • module: list of global symbols

  • function: list of basic blocks that form control-flow graph (CFG)

  • basic blocks: list of instructions terminated by branch/return

  • instruction = typed assembly instructions

  • constant = constant literals, globals, (function pointer is a constant)

  • value = almost anything is a value; mostly constants & instructions

  • global symbols: @xxx

  • local symbols: %xxx

  • to use a basic block, %xxx

  • to define a basic block, xxx:

common pitfals

  • LLVM-IR only deals with reducible loops
  • irreducible loops are not recognized
  • SSA only followed in reachable code (there is a path to the basic block); only iterate over reachable blocks, otherwise you might find broken IR
  • address spaces are pointer-typed
  • null (0) is a valid pointer even in address space 0 (default)
  • types are sign agnostic
  • llvm::constant includes values that become constants at start-up time, e.g. addresses of globals

Useful commandline options

Target and Code Generation

Example: lea

Each target (x86) has subtargets (mostly attribute-based)

  • Target/
    • TargetMachine contains module level lowering information
    • object file / OS ABI information
  • Subtarget
    • contains function level lowering information, e.g. limit instruction selection (“limit this function to use SSE”)
    • A subtarget is primarily defined by CPU & features
    • SubtargetFeature = ISA-level stuff

Machine IR

  • CodeGen/MachineInstr.h
  • similar hierarchy as LLVM-IR
  • not completely free of LLVM-IR
  • tons of MIR construction APIs
  • target-dependent AND -independent opcodes & registers
  • can be produced with llc

FastISel

  • generates straight forward instructions
  • not optimized

LLVM-IR to MIR happens after optimization

Instruction Selection and Register Allocation

  • CodeGen/
  • 3 instruction selectors
    • SelectionDAG
    • FastISel
    • GlobalISel
  • 2 typical register allocators
    • RegAllocFast
    • RegAllocGreedy

Object Generation - Assembly Printing

  • assembler is built-in
  • CodeGen/AsmPrinter
    • target independent MIR to assembly
    • calls into backends (?) Target/*AsmPrinter
  • target specific opcode & operand printing

AsmParser/ MC/

  • target-independent object file construction / encoding

Target/{AsmParser,MCTargetDesc}

  • parse target-specifc assembly & encoding

Object Reading

Object/ DebugInfo/

  • parse object files & target-independent data
  • calls into Target/ for instruction decoding

Target/Disassembler

  • target specific instruction decoding