compiler-development

name	compiler-development
description	Expertise in compiler development using LLVM infrastructure including frontend design, IR generation, optimization passes, and code generation. Use this skill when building custom programming languages, implementing DSL compilers, or working on compiler internals.

Compiler Development Skill

This skill provides comprehensive knowledge of building compilers and language implementations using the LLVM infrastructure, with NeverC-specific cross-platform development checklists covering Windows x64/arm64, Linux x64/arm64, and macOS arm64.

---

Cross-Platform Target Matrix

NeverC operates on a three-dimensional matrix: OS × Arch × ExecutionLevel.

                ┌──── arm64 ────┬──── x86_64 ────┐
     Darwin ────┤ User / Kernel │ User / Kernel  │  Mach-O
     Linux  ────┤ User / Kernel │ User / Kernel  │  ELF
     Android────┤ User / Kernel │ User / Kernel  │  ELF
     Windows────┤ User / Kernel │ User / Kernel  │  COFF
                └───────────────┴────────────────┘

All platform differences are encoded in TargetDesc (returned by describeTriple(triple, Level)). Passes read from the table, never write if (OS == Darwin) branches. Adding a new platform = filling one row in describeTriple() + adding one case in each extractor's switch.

Supported Triples

Platform	Triple	Object Format
macOS arm64	arm64-apple-macos	Mach-O
macOS x86_64	x86_64-apple-macos	Mach-O
Linux x86_64	x86_64-linux-gnu	ELF
Linux arm64	aarch64-linux-gnu	ELF
Windows x86_64	x86_64-pc-windows-msvc	COFF
Windows arm64	aarch64-pc-windows-msvc	COFF
Android arm64	aarch64-linux-android	ELF
Android x86_64	x86_64-linux-android	ELF
iOS arm64	arm64-apple-ios	Mach-O

---

Platform-Specific Development Checklists

Windows x64 Checklist

• ABI: Win64 calling convention — RCX, RDX, R8, R9 for first 4 integer args; 32-byte shadow space on stack
• Codegen: createWinX86_64TargetCodeGenInfo() in ModuleEmitter.cpp; AVX level selection via Target.getABI()
• Object format: COFF — section names use .text, .data, .rdata; no segment prefixes
• Triple dispatch: Triple.getOS() == llvm::Triple::Win32 + Triple.getArch() == llvm::Triple::x86_64
• Stack probe: Windows requires __chkstk for stack allocations > 4KB; dyncode mode uses inline probe-stack dyncode to avoid CRT dependency
• TLS/PEB access: movq %gs:0x60, $0 for PEB pointer (user mode)
• DynCode imports: PEB walk (WinPEBImportPass) resolves kernel32.dll / ntdll.dll exports at runtime
• Kernel mode: KernelImportABI::WindowsKernelResolverShim; no PEB, no syscall stubs
• Build host: _popen / _pclose / _chdir / FindFirstFileA instead of POSIX equivalents
• CMake: detect MSVC via _NEVERC_HOST_MSVC; static CRT (/MT) via CMAKE_MSVC_RUNTIME_LIBRARY
• LTO disabled on Windows hosts: Full LTO miscompiles under Windows clang due to unspecified-evaluation-order UB; keep non-LTO until clean
• Linker: lld-link; dead-code elimination is on by default (no --gc-sections)
• Cross-compilation: bundled MSVC SDK in runtime/; no external SDK needed from macOS/Linux host
• Shell redirects: >nul 2>&1 instead of >/dev/null 2>&1
• Default shell: cmd.exe (not /bin/sh)
• Path separator: use llvm::sys::path::native() for backslash normalization
• Binary format: .exe / .dll output; NEVERC_STRIP_BINARY OFF on Windows

Windows arm64 Checklist

• ABI: Windows ARM64 uses its own calling convention (not pure AAPCS); X0–X7 for args, X18 reserved for TEB
• Codegen: createWindowsAArch64TargetCodeGenInfo() — separate from generic AArch64; AArch64ABIKind::AAPCS but with Win32 OS dispatch
• TLS/TEB access: ldr $0, [x18, #0x60] for PEB pointer (X18 = TEB on Windows ARM64)
• Stack probe: custom AArch64 probe-stack; different from x64 __chkstk
• BTI/PAC: Windows ARM64 supports branch target identification; check LangOpts.BranchTargetEnforcement module flags
• DynCode inject flags: TargetInjectFlags_Windows_AArch64.def — distinct from Unix AArch64 flags
• Kernel inject flags: TargetKernelFlags_Windows_AArch64.def
• Object format: COFF (same as Windows x64)
• Syscall ABI: SyscallABI::WindowsPEB — no direct syscall instruction; all imports via PEB walk
• Register convention: same ArgRegs as Unix arm64 (x0–x7, 8 regs) but calling convention semantics differ

Linux x64 Checklist

• ABI: System V AMD64 — RDI, RSI, RDX, R10, R8, R9 for syscall args; red zone present
• Codegen: createX86_64TargetCodeGenInfo() (non-Win32 default path)
• Object format: ELF — section .text; isOSBinFormatELF() gates ELF-specific module flags
• Syscall: syscall instruction; RAX = syscall number, RAX = return value
• DynCode imports: SyscallStubPass generates inline syscall stubs
• Kernel mode: KernelImportABI::LinuxKallsymsShim; resolve kernel symbols via kallsyms
• Stack probe: no __chkstk; Linux typically has guard pages; but dyncode mode may still need inline probing for large allocations
• Linker GC: --gc-sections for dead code elimination
• Build host: standard POSIX — popen, pclose, chdir, glob
• Exit code: WIFEXITED(Status) ? WEXITSTATUS(Status) : 1
• ELF module flags: getTriple().isOSBinFormatELF() → emit ELF-specific metadata

Linux arm64 Checklist

• ABI: AAPCS64 — X0–X7 for args (8 registers); no shadow space
• Codegen: createAArch64TargetCodeGenInfo() with AArch64ABIKind::AAPCS
• Syscall: svc #0; X8 = syscall number, X0 = return value
• PAC/BTI: module flags for sign-return-address, branch-target-enforcement emitted when Arch == llvm::Triple::aarch64
• DynCode inject flags: TargetInjectFlags_Unix_AArch64.def
• Kernel inject flags: TargetKernelFlags_Unix_AArch64.def
• Object format: ELF
• Indirect branches: AllBlrPass (optional -fdyncode-all-blr) rewrites indirect branches to BLR for arm64

macOS arm64 Checklist

• ABI: DarwinPCS — variant of AAPCS with differences in va_arg, struct passing, and alignment
• Codegen: createAArch64TargetCodeGenInfo() with AArch64ABIKind::DarwinPCS (gated by Target.getABI() == "darwinpcs")
• Object format: Mach-O — section __text (prefixed with __); isOSBinFormatMachO() for Mach-O-specific paths
• Syscall: svc #0x80; X16 = syscall number (with SyscallNumberMask = 0 for arm64, = 0x2000000 for x86_64 Darwin), X0 = return value
• Kernel mode: KernelImportABI::DarwinXNUKextShim for kext symbol resolution
• Build host: Homebrew LLVM preferred; CMake auto-detects /opt/homebrew/opt/llvm/bin
• Linker GC: -Wl,-dead_strip (macOS-specific flag, not --gc-sections)
• Code signing: Developer ID signed + Apple notarized; no quarantine workaround needed
• llvm-ar: auto-detected when host is non-Apple clang to avoid libtool incompatibility
• PAC/BTI: BranchProtectionPAuthLR, sign-return-address-all, sign-return-address-with-bkey module flags

---

ABI & Calling Convention Reference

x86_64 ABI Split

	System V (Linux/macOS/Android)	Win64 (Windows)
Integer args	RDI, RSI, RDX, RCX, R8, R9	RCX, RDX, R8, R9
Syscall args	RDI, RSI, RDX, R10, R8, R9	N/A (PEB walk)
Syscall # reg	RAX	N/A
Return reg	RAX	RAX
Shadow space	No (128-byte red zone)	Yes (32 bytes)
Callee-saved	RBX, RBP, R12–R15	RBX, RBP, RDI, RSI, R12–R15
Struct return	RAX:RDX or memory	RAX or memory

AArch64 ABI Split

	AAPCS (Linux/Android)	DarwinPCS (macOS/iOS)	Windows ARM64
Integer args	X0–X7	X0–X7	X0–X7
Syscall # reg	X8	X16	N/A (PEB)
Syscall insn	svc #0	svc #0x80	N/A
TEB/TLS	N/A	N/A	X18 (reserved)
PEB access	N/A	N/A	ldr $0, [x18, #0x60]
va_arg	stack-based	slightly different alignment	Win64 rules

---

Build System Cross-Platform Checklist

CMake Configuration (NeverC.cmake)

• LLVM targets: always "AArch64;X86" — only two backends
• MSVC detection: _NEVERC_HOST_MSVC flag; avoid -O2, -march=native on MSVC
• Static CRT on Windows: CMAKE_MSVC_RUNTIME_LIBRARY = "MultiThreaded$<$<CONFIG:Debug>:Debug>"
• LLD selection: auto-detect lld-link (Windows), ld64.lld (macOS), ld.lld (Linux)
• LTO: disabled on Windows hosts and MSVC; enabled only on macOS/Linux non-Debug non-cross-compile
• PGO: two-phase generate/use; -fprofile-instr-generate / -fprofile-instr-use
• Section GC flags: -ffunction-sections -fdata-sections + platform-specific linker flags
• Native arch tuning: -march=native when NEVERC_NATIVE_ARCH=ON and not cross-compiling
• ccache/sccache: auto-detected via find_program
• llvm-ar: auto-used on macOS when host compiler is non-Apple Clang (avoids libtool incompatibility)

Platform Abstraction Layer (Platform.cpp)

All #ifdef _WIN32 branches:

Function	POSIX	Windows
shellExecute	popen / pclose + WEXITSTATUS	_popen / _pclose + raw status
shellExecuteNoCapture	>/dev/null 2>&1	>nul 2>&1
globFiles	glob() / globfree()	FindFirstFileA / FindNextFileA / FindClose
getDefaultShell	/bin/sh	cmd.exe
changeCwd	chdir	_chdir

Use llvm::sys::fs::* and llvm::sys::path::* whenever possible — they are already cross-platform.

---

DynCode Pipeline Cross-Platform Checklist

TargetDesc Table Fields

Every platform must populate:

• OS — DynCodeOS enum
• Arch — DynCodeArch enum
• Format — ObjectFormat enum (MachO/ELF/COFF)
• Syscall — SyscallABI (which syscall mechanism)
• TextSectionName — .text (ELF/COFF) or __text (Mach-O)
• AsmTemplate — inline asm for syscall instruction
• SyscallNumberReg / SyscallRetReg — register names
• SyscallNumberMask — 0x2000000 for macOS x86_64, 0 elsewhere
• ArgRegs / NumArgRegs — argument register array
• TCBReadAsm / TCBReadConstraint — PEB/TEB access (Windows only)
• DriverInjectFlags — from .def tables per (OS, Arch) pair
• KernelImport — KernelImportABI enum for kernel mode
• KernelInjectFlags — from .def tables per (OS, Arch) pair

Inject Flags .def Tables

Each (OS, Arch, Level) combination has its own .def file:

TargetInjectFlags_Unix_X86_64.def
TargetInjectFlags_Unix_AArch64.def
TargetInjectFlags_Windows_X86_64.def
TargetInjectFlags_Windows_AArch64.def
TargetKernelFlags_Unix_X86_64.def
TargetKernelFlags_Unix_AArch64.def
TargetKernelFlags_Windows_X86_64.def
TargetKernelFlags_Windows_AArch64.def

Plus UserExtra_* variants for extension. When adding a new platform, create matching .def files.

Pipeline Execution Order (Cross-Platform)

cc1 frontend (C → IR) → PIC default
  ↓
PipelineStartEP:
  ① ZeroRelocPass (Prep)        — all platforms
  ② IndirectBrPass              — all platforms
  ③ MemIntrinPass               — mem*/str*/bzero inlining (all)
  ④ StringRuntimePass           — builtin string → stack arena (all)
  ⑤ CompilerRtPass              — __udivti3 / i128 div inlining (all)
  ⑥ SyscallStubPass             — User + non-Windows only
  ⑦ WinPEBImportPass            — User + Windows only
  ⑧ KernelImportPass            — Kernel, all OS
  ⑨ Data2TextPass phase 1       — all platforms
     (extensible hooks: RunBefore/AfterPrep, RunBeforeInlining)
  ↓
LLVM optimizer (AlwaysInliner, SROA, SLPVectorize, InstCombine)
  ↓
OptimizerLastEP:
  ⑩ Data2TextPass phase 2       — all platforms
  ⑪ ZeroRelocPass (Stackify)    — all platforms
  ⑫ AllBlrPass                  — optional (-fdyncode-all-blr), arm64 only
     (extensible hooks: RunAfterInlining, RunAfterStackify)
  ↓
MIR (TargetPassConfig.addMachinePasses):
  ⑬ DynCodeMIRPrepPass        — per-target rewrite patterns/opcodes
     (extensible hooks: RunBeforePreEmit, RunAfterPreEmit)
  ↓
Extractor: MachO / ELF / COFF dispatcher
  → patch intra-.text relocs, reject external relocs/data, output flat .bin

Object Extractor Differences

	Mach-O	ELF	COFF
Extractor	MachOExtractor	ELFExtractor	COFFExtractor
Text section	__text	.text	.text
Reloc patching	intra-segment	intra-section	intra-section
External reloc	rejected	rejected	rejected
Data sections	rejected	rejected	rejected

---

Module Flags & Platform-Specific Metadata

AArch64-Only Module Flags

Only emitted when Arch == llvm::Triple::aarch64:

"branch-target-enforcement"    // BTI
"branch-protection-pauth-lr"   // PAC for LR
"sign-return-address"          // PAC sign return addr
"sign-return-address-all"      // sign all functions
"sign-return-address-with-bkey"// use B-key

Windows-Specific Module Flags

"ms-kernel"    // kernel mode (-fms-kernel); affects AsmPrinter on both X86 and AArch64
"cfguard"      // Control Flow Guard (value 2 = full, 1 = no-checks)
"ehcontguard"  // Exception Handling Continuation Guard

ELF-Specific Paths

if (getTriple().isOSBinFormatELF()) { /* ELF metadata */ }

Mach-O-Specific Paths

const bool isMachO = getTriple().isOSBinFormatMachO();

---

Testing Across Platforms

Test Organization

• DynCodeTests.cpp — core dyncode pipeline tests
• DynCodeCrossTargetTests.cpp — cross-target compilation tests
• DynCodeStressTests.cpp — stress / edge-case tests
• BasicTests.cpp — basic compiler functionality
• DriverTests.cpp — driver / CLI tests
• BuildTests.cpp — build system (make) tests
• Platform-specific loaders: loader_windows.c, loader_linux.c, loader_arm64_macos.c

Cross-Platform Test Checklist

• Test all 5 target triples: macOS arm64, Linux x64, Linux arm64, Windows x64, Windows arm64
• Test both User and Kernel execution levels
• Verify dyncode extraction for each object format (MachO/ELF/COFF)
• Verify import resolution: PEB walk (Windows), syscall stubs (Linux/macOS), kernel shims
• Test string runtime in dyncode mode across all targets
• Verify stack probe behavior on Windows targets (both x64 and arm64)
• Test bad-byte auditing across different target encodings
• Cross-compile from macOS host to all other targets

---

Compiler Architecture Overview

Classic Three-Phase Design

Source Code → Frontend → Middle-End (Optimizer) → Backend → Machine Code
                ↓              ↓                      ↓
             AST/IR      LLVM IR Passes          Target Code

NeverC-Specific Extensions

Source (.c/.nc) → Frontend → DynCode IR Passes → MIR Passes → Backend → Extractor → .bin
                                    ↓                   ↓                      ↓
                             ZeroReloc, Import    MIRPrepPass         MachO/ELF/COFF extract
                             Syscall, String      Rewrite patterns

Frontend Development

Lexical Analysis

enum class TokenKind {
    Identifier, Number, String, Keyword,
    Operator, Punctuation, EndOfFile
};

struct Token {
    TokenKind kind;
    std::string value;
    SourceLocation location;
};

Parser Implementation

• Recursive Descent: Easy to implement, good error messages
• Operator Precedence Parsing: Efficient for expression parsing
• LALR/LR: Use tools like Bison for complex grammars

AST Design

class Expr {
public:
    virtual ~Expr() = default;
    virtual llvm::Value* codegen() = 0;
};

class BinaryExpr : public Expr {
    std::unique_ptr<Expr> LHS, RHS;
    char Op;
public:
    llvm::Value* codegen() override {
        llvm::Value* L = LHS->codegen();
        llvm::Value* R = RHS->codegen();
        switch (Op) {
            case '+': return Builder.CreateFAdd(L, R, "addtmp");
            case '-': return Builder.CreateFSub(L, R, "subtmp");
            case '*': return Builder.CreateFMul(L, R, "multmp");
            case '/': return Builder.CreateFDiv(L, R, "divtmp");
        }
    }
};

LLVM IR Generation

Module and Context Setup

#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/IRBuilder.h"

class CodeGen {
    std::unique_ptr<llvm::LLVMContext> Context;
    std::unique_ptr<llvm::Module> Module;
    std::unique_ptr<llvm::IRBuilder<>> Builder;

public:
    CodeGen() {
        Context = std::make_unique<llvm::LLVMContext>();
        Module = std::make_unique<llvm::Module>("my_module", *Context);
        Builder = std::make_unique<llvm::IRBuilder<>>(*Context);
    }
};

Target-Aware Codegen Dispatch

std::unique_ptr<TargetCodeGenInfo> createTargetCodeGenInfo(ModuleEmitter &ME) {
    const llvm::Triple &Triple = ME.getTarget().getTriple();

    switch (Triple.getArch()) {
    case llvm::Triple::aarch64:
        switch (Triple.getOS()) {
        case llvm::Triple::Win32:
            return createWindowsAArch64TargetCodeGenInfo(ME, Kind);
        default:
            return createAArch64TargetCodeGenInfo(ME, Kind);
        }
    case llvm::Triple::x86_64:
        switch (Triple.getOS()) {
        case llvm::Triple::Win32:
            return createWinX86_64TargetCodeGenInfo(ME, AVXLevel);
        default:
            return createX86_64TargetCodeGenInfo(ME, AVXLevel);
        }
    }
}

Optimization Pass Pipeline

New Pass Manager

#include "llvm/Passes/PassBuilder.h"

void optimizeModule(llvm::Module& M) {
    llvm::PassBuilder PB;
    llvm::LoopAnalysisManager LAM;
    llvm::FunctionAnalysisManager FAM;
    llvm::CGSCCAnalysisManager CGAM;
    llvm::ModuleAnalysisManager MAM;

    PB.registerModuleAnalyses(MAM);
    PB.registerCGSCCAnalyses(CGAM);
    PB.registerFunctionAnalyses(FAM);
    PB.registerLoopAnalyses(LAM);
    PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);

    llvm::ModulePassManager MPM = PB.buildPerModuleDefaultPipeline(
        llvm::OptimizationLevel::O2);
    MPM.run(M, MAM);
}

JIT Compilation

LLVM ORC JIT

#include "llvm/ExecutionEngine/Orc/LLJIT.h"

auto JIT = llvm::orc::LLJITBuilder().create();
if (!JIT) handleError(JIT.takeError());

(*JIT)->addIRModule(llvm::orc::ThreadSafeModule(
    std::move(Module), std::move(Context)));

auto Sym = (*JIT)->lookup("main");
auto* MainFn = (int(*)())Sym->getAddress();
int result = MainFn();

Language Implementation Patterns

Memory-Safe Languages

• Use LLVM's memory sanitizer hooks
• Implement bounds checking with GEP introspection
• Reference counting or garbage collection integration

Type Systems

• Implement type inference during AST construction
• Generate appropriate LLVM types (i32, float, struct, ptr)
• Handle generic types via monomorphization or boxing

Error Handling

• Generate exception handling via LLVM's landingpad/invoke
• Implement Result/Option types as tagged unions
• Use LLVM's personality functions for unwinding

Development Workflow

1. Start Simple: Begin with Kaleidoscope tutorial
2. Incremental Features: Add one language feature at a time
3. Test Extensively: Unit tests for each compiler phase
4. Use LLVM Tools: opt, llc, llvm-dis for debugging IR
5. Profile and Optimize: Focus on common code patterns
6. Cross-Platform First: Ensure every new feature works across all 5 target triples

Cross-Platform Development Golden Rules

1. Never hardcode OS checks in passes — use TargetDesc table lookup
2. Always use llvm::sys::path / llvm::sys::fs — not raw POSIX or Win32 APIs
3. Gate #ifdef _WIN32 to Platform.cpp — keep it out of compiler logic
4. Test Windows ARM64 separately — its ABI is a unique hybrid (AAPCS regs + Win64 semantics)
5. Remember SyscallNumberMask — macOS x86_64 uses 0x2000000 offset, everything else is 0
6. Mach-O section names have __ prefix — __text not .text
7. Windows has no direct syscall in dyncode — always PEB walk, even on ARM64
8. LTO is not safe on Windows hosts — known UB surfaces under LTO; build non-LTO on Windows CI

Resources

Official Tutorials

• LLVM Kaleidoscope: Building a language from scratch
• Clang internals: Frontend implementation patterns
• Writing an LLVM Backend: Target code generation

Community Projects

See DIY Compiler section in README.md for 100+ example implementations across different language paradigms.

Getting Detailed Information

When you need detailed and up-to-date resource links, tool lists, or project references, fetch the latest data from:

https://raw.githubusercontent.com/gmh5225/awesome-llvm-security/refs/heads/main/README.md

This README contains comprehensive curated lists of:

• 100+ DIY compiler implementations (DIY Compiler section)
• Toolchain configurations and IDE setup
• Compiler development tutorials and books