Jit Compilation

JIT Compilation

Purpose

Guide agents through just-in-time compilation: LLVM ORC JIT v2 (ExecutionSession, IRLayer, ObjectLayer), LLJIT for simpler use cases, Cranelift JIT, inline caches for dynamic dispatch, trampolines for lazy compilation, security considerations (W^X, code signing), and Rust dynasm for x86 codegen.

When to Use

• Building an interpreter with a JIT hot-path
• Implementing lazy function compilation on first call
• Embedding dynamic code generation in a REPL or game engine
• Prototyping codegen without writing a full linker
• Creating inline caches for polymorphic call sites
• Generating x86 machine code from Rust with dynasm

Workflow

1. JIT architecture overview

Source/AST/Bytecode
    → IR (LLVM IR or Cranelift CLIF)
    → Object file (in memory)
    → Runtime linker (RTDyldObjectLinkingLayer)
    → Executable code in R+X memory
    → Function pointer call

2. LLJIT — simplest LLVM JIT

#include "llvm/ExecutionEngine/Orc/LLJIT.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/IRBuilder.h"

using namespace llvm;
using namespace llvm::orc;

int main() {
    auto JIT = cantFail(LLJITBuilder().create());

    LLVMContext Context;
    auto M = std::make_unique<Module>("jit", Context);
    IRBuilder<> Builder(Context);

    // int add(int a, int b) { return a + b; }
    Function *AddFn = Function::Create(
        FunctionType::get(Builder.getInt32Ty(),
                          {Builder.getInt32Ty(), Builder.getInt32Ty()}, false),
        Function::ExternalLinkage, "add", M.get());

    BasicBlock *BB = BasicBlock::Create(Context, "entry", AddFn);
    Builder.SetInsertPoint(BB);
    auto Args = AddFn->arg_begin();
    Value *Sum = Builder.CreateAdd(Args, Args + 1);
    Builder.CreateRet(Sum);

    cantFail(JIT->addIRModule(ThreadSafeModule(std::move(M), std::make_unique<LLVMContext>())));

    auto AddSym = JIT->lookup("add");
    auto *AddPtr = (int (*)(int, int))AddSym->getValue();
    int result = AddPtr(3, 4);  // 7
    return 0;
}

clang++ -std=c++17 jit.cpp $(llvm-config --cxxflags --ldflags --libs core orcjit native) -o jit

3. ORC JIT v2 layers

ExecutionSession ES;
auto &MainJD = ES.createBareJITDylib("main");

RTDyldObjectLinkingLayer ObjectLayer(
    ES, []() { return std::make_unique<SectionMemoryManager>(); });

IRCompileLayer CompileLayer(
    ES, ObjectLayer, std::make_unique<TargetMachineBuilder>());

// Add IR module to JITDylib
ThreadSafeModule TSM = ...;
CompileLayer.add(MainJD, std::move(TSM));

// Resolve symbol
auto Sym = ES.lookup({&MainJD}, "my_func");

Layers:

• IRLayer — compiles LLVM IR to object
• ObjectLayer — links relocatable objects
• ExecutionSession — symbol lookup and JITDylib management

4. Lazy compilation with trampolines

First call → trampoline → compile function → patch trampoline → direct call

// Simplified lazy compile on first invocation
void *LazyCompile(const std::string &Name) {
    if (!Compiled.count(Name)) {
        auto Fn = CompileFromAST(Name);
        Compiled[Name] = Fn;
        // Patch call site or update function pointer table
    }
    return Compiled[Name];
}

ORC supports lazy reexports and lazy compilation via LazyCallThroughManager.

5. Inline caches

// Concept: monomorphic call site caches resolved target
// Pseudocode for dynamic language
function call_site(obj, method, args) {
    if (obj.class_id === cache.class_id) {
        return cache.fn_ptr(args);  // fast path
    }
    cache.class_id = obj.class_id;
    cache.fn_ptr = resolve_method(obj, method);
    return cache.fn_ptr(args);
}

JIT generates specialized code per cached type; deoptimize on cache miss.

6. Cranelift JIT

use cranelift::prelude::*;
use cranelift_jit::{JITBuilder, JITModule};
use cranelift_module::{Linkage, Module};

let isa = cranelift_native::builder().finish(settings::Flags::new(settings::builder()))?;
let jit_builder = JITBuilder::with_isa(isa, cranelift_module::default_libcall_names());
let mut module = JITModule::new(jit_builder);

let mut ctx = module.make_context();
ctx.func = Function::with_name_signature(
    module.declare_function("add", Linkage::Export, &sig)?,
    sig,
);
// ... build IR in ctx.func ...
module.define_function(func_id, &mut ctx)?;
module.finalize_definitions()?;
let code = module.get_finalized_function(func_id);
let add_fn: fn(i32, i32) -> i32 = unsafe { std::mem::transmute(code) };

Cranelift: faster compile times than LLVM, good for embeddable JITs.

7. Security — W^X policy

W^X (Write XOR Execute)
├── Memory page is writable OR executable, never both
├── JIT: allocate RW → write code → mprotect(RX)
└── macOS hardened runtime requires signed JIT pages

#include <sys/mman.h>

void *mem = mmap(NULL, size, PROT_READ | PROT_WRITE,
                 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
// write machine code to mem
mprotect(mem, size, PROT_READ | PROT_EXEC);

Linux: MAP_JIT on Apple platforms; sealed memfd on hardened systems.

8. Rust dynasm (x86)

use dynasm::dynasm;
use dynasmrt::{Assembler, ExecutableBuffer};

let mut asm = Assembler::new().unwrap();
dynasm!(asm
    ; .arch x86_64
    ; add:
    ; add eax, ecx
    ; ret
);
let buf = asm.finalize().unwrap();
let add_fn: fn(i32, i32) -> i32 = unsafe { std::mem::transmute(buf.ptr(0)) };

Use for lightweight asm snippets without LLVM dependency.

9. Decision tree

JIT backend choice?
├── Need LLVM optimizations → ORC JIT / LLJIT
├── Fast compile, embeddable → Cranelift
├── Tiny asm snippets → dynasm
└── ML workloads → MLIR → ORC pipeline

Common Problems

Related Skills

• skills/compiler-internals/llvm-passes — optimize before JIT
• skills/compiler-internals/compiler-frontend — AST to IR for JIT input
• skills/compiler-internals/mlir — MLIR lowering to LLVM for JIT
• skills/compilers/llvm — LLVM IR fundamentals
• skills/low-level-programming/interpreters — bytecode interpreters using JIT
• skills/low-level-programming/assembly-x86 — hand-written asm context