Internals & Dispatch

RepliBuild compiles C/C++ with Clang, reads the DWARF debug metadata the compiler emits about its own output, and generates Julia bindings that are correct by construction — struct offsets, enum underlying types, vtable slots, base-subobject offsets, and calling conventions all come from the compiler's own record rather than being guessed. Where DWARF is incomplete (enums the optimizer folded away, macro definitions, function-pointer typedefs) the Clang.jl AST fills the gap; where Julia's computed struct alignment disagrees with the DWARF size, the struct is treated as packed and routed away from ccall.

This page is the technical reference for how that pipeline is assembled — the stages, the three dispatch tiers, and the modules behind each. It is aimed at contributors and advanced integrators, not everyday use.

Tier 2 — the MLIR/JLCS marshalling layer — has its own page: ABI Marshalling as Compiler IR covers the dialect's types and ops, the SysV lowering, the thunk calling contract, source-level debugging, and the failure classes the design exists to make loud. This page describes where it sits; that one describes what it is.

The pipeline

Source becomes a loadable Julia module in seven stages, each owned by one module:

1. DependencyResolver — resolve the [dependencies] table (git / local / system) and merge the external sources into the build graph.
2. Discovery — scan files and the #include graph; write or refresh replibuild.toml.
3. Compiler — Clang compiles each translation unit to LLVM IR, cached on source mtime plus a compile fingerprint (flags, defines, include dirs, LLVM version, target triple).
4. Linker — link, optimize, and assemble the IR into the .so (plus LTO bitcode when enable_lto is on).
5. DWARFParser — llvm-dwarfdump + nm produce ClassInfo / VtableInfo and the compilation_metadata.json layout facts.
6. Wrapper generator — DWARF + symbols drive the C or C++ generator to emit the Julia module.
7. JITManager — at module load, stand up a per-library MLIR thunk engine for any Tier-2 calls.

For C projects the link/optimize/assemble steps run in-process on Julia's resident libLLVM, version-matched to the JLL clang that emitted the IR (no external LLVM, no DWARF-dropping version skew); a failure is a hard error unless [link] fallback = true selects the external llvm-link/opt pipeline. C++ always uses the external pipeline plus the system MLIR dialect. Only the final .ll → .so codegen shells to clang in both buckets.

Three dispatch tiers

Each wrapped function is routed to exactly one calling tier, chosen from its DWARF signature:

The routing decision lives in Wrapper/DispatchLogic.jl (is_ccall_safe, is_c_lto_safe); the Tier selection logic section below lists the exact checks.

Wrapper

Source: src/Wrapper.jl, src/Wrapper/

The Wrapper package generates Julia FFI modules from DWARF metadata and binary symbol tables. It is structured as a two-track system: a C generator and a C++ generator, selected automatically via config.wrap.language.

Module layout

Language selection

wrap.language is an extensible dispatch key — "c" and "cpp" are the first two targets, with additional language generators planned:

[wrap]
language = "c"   # selects C generator + clang toolchain
language = "cpp" # selects C++ generator + clang++ toolchain (default)

discover() sets this automatically based on the scanned source files. Adding a new language means adding a generator under src/Wrapper/<Lang>/ and registering it in Wrapper/Generator.jl.

Tier selection logic

The function is_ccall_safe() in src/Wrapper/DispatchLogic.jl is the core dispatch decision. It inspects each function's DWARF metadata and returns false when the signature needs an MLIR thunk (Tier 2), true when it can be called directly (Tier 1 or Tier 3).

Choosing between the two direct tiers is a second, C-only step. is_c_lto_safe() marks a function Tier-1-eligible unless its return type is a packed struct or a by-value union — C has no templates, vtables, STL, or inheritance, so those are the only ABI hazards, and by-value parameters are fine either way. Eligibility is necessary but not sufficient: with [wrap.tier1] enable = true the function must also survive slicing, the hazard/size policy, and the dlsym pre-flight. Anything that does not, and everything when Tier 1 is off, emits ccall.

Checks performed:

1. STL container types — Any STL type in parameters or return forces Tier 2
2. Return type safety:
   • Template returns (contains <) → Tier 2 (unpredictable ABI)
   • Struct return by value > 16 bytes → Tier 2 (too large for ccall sret)
   • Non-POD class return → Tier 2
   • Packed struct return (DWARF size != Julia aligned size) → Tier 2
3. Parameter type safety:
   • Union parameters → Tier 2
   • Packed struct parameters → Tier 2
4. Exception safety — Per-function is_noexcept flag from DWARF. If absent (function may throw) and the module's may_throw setting is on, the function routes through jlcs.try_call rather than ccall.

For struct-graph cases where pairwise heuristics miss transitive layout mismatches (a non-packed struct that contains a packed struct, for example), src/IRGen/DAGDiff.jl performs a structural type-graph diff to surface bad cases and produces a topo-sorted lowering order for multi-type thunks.

Functions routed to Tier 2 are further divided between JIT dispatch (JITManager.invoke()) and AOT thunks (ccall to _thunks.so), controlled by the aot_thunks config flag.

Idiomatic wrapper generation

Beyond raw ccall bindings, the wrapper generator clusters related C++ functions by class name to produce idiomatic Julia types:

1. Factory detection: Functions matching create_X, new_X, make_X, alloc_X, init_X, or returning X* are identified as constructors.
2. Destructor detection: Functions matching delete_X, destroy_X, free_X, dealloc_X, or X_destroy are identified as destructors.
3. Method clustering: Functions taking X* as their first parameter and associated with the same DWARF class are grouped as instance methods.

The result is a mutable struct ManagedX with a raw Ptr{Cvoid} handle, a registered finalizer calling the C++ destructor, and multiple-dispatch method proxies that pass the pointer via Base.unsafe_convert.

extract_symbols(binary_path::String, registry::TypeRegistry; demangle::Bool=true, method::Symbol=:nm)

is_enum_like(cpp_type::String)::Bool

is_function_pointer_like(cpp_type::String)::Bool

is_struct_like(cpp_type::String)::Bool

wrap_basic(config::RepliBuildConfig, library_path::String; generate_docs::Bool=true)

wrap_library(config::RepliBuildConfig, library_path::String;
             headers::Vector{String}=String[],
             generate_tests::Bool=false,
             generate_docs::Bool=true)

ParamInfo

SymbolInfo

TypeRegistry

TypeStrictness

Compiler

Source: src/Builder/Compiler.jl

The Compiler module handles the translation of C/C++ source code into LLVM IR and shared libraries. It oversees the entire build pipeline from dependency management down to IR optimization.

Build pipeline

1. Auto-discovery and dependency resolution: Scans the project directory, resolving file paths and external git/local dependencies to merge into the build graph.
2. Pre-processing (shims and templates): Dynamically generates C/C++ shim files for configured macros and explicitly instantiates templates based on replibuild.toml settings. This allows normally invisible constructs to manifest in the final binary and DWARF metadata. Macro shims are pinned to default symbol visibility, and a header-collision guard verifies each direct shim #include resolves inside the project/dependency tree — the shim TU lives under the build cache, so a bare include could otherwise fall through the -I path to a system-installed header at a different version and silently bake wrong macro values.
3. Compilation to LLVM IR: Translates source code into .ll text format — .c via the JLL clang, .cpp via system clang++. The per-file cache is keyed on source mtime plus a compile fingerprint (flags, defines, include dirs, LLVM version, target triple); config changes can never silently reuse stale IR.
4. IR transformation and sanitization: Strips attributes incompatible with Julia's internal LLVM JIT, removes va_start/va_end intrinsics from varargs function bodies (varargs are routed through true-variadic @ccall wrapper generation), and cleans mismatched debug metadata.
5. Link / optimize / assemble: For C, these steps run in-process on Julia's resident libLLVM — LLVM.link! for linking, the new pass manager (default<O…>) for optimization, and in-process bitcode assembly — version-matched to the JLL clang that emitted the IR. A failure is a hard error; [link] fallback = true selects the external llvm-link/opt pipeline instead. C++ always uses the external pipeline.
6. Static promotion (_promote_statics_libllvm, C in-process bucket, [link] promote_statics): see below.
7. Codegen: The final .ll → .so step shells to clang/clang++.

Static promotion

Post-optimization and pre-codegen, every function or global that a Tier-1 bitcode slice may bind by declare but that cannot reach the .so's dynamic symbol table is renamed to an exported __rb_<lib>_<name> with external linkage and default visibility. The test is "not exportable", not "internal linkage": default<O…> runs no internalize pass, so it covers both internal/private linkage (file-local statics) and external-linkage-but-hidden/protected symbols — Lua's LUAI_FUNC functions and LUAI_DDEF tables are exactly the latter shape. Only internal constants stay internal: no symbol exists for a slice to bind, and read-only data has no divergence class.

The rename happens on the one module (<name>_abi.ll) that becomes both the .so and the slice source, so the two are bit-identical by construction — which is what removes the whole-module path's duplicated-static failure class rather than papering over it. The old→new map lands in compilation_metadata.json under promoted_symbols, and extract_symbols_from_binary filters __rb_* so promoted statics never surface as wrappable API.

Metadata extraction

At build time the compiler also extracts DWARF metadata into compilation_metadata.json (functions, struct definitions, enums, globals). DIE parsing is depth-aware: readelf DIE headers carry the tree depth, and member/enumerator/inheritance/template DIEs at depth d attribute to the type last seen at depth d−1 — so nested type definitions interleaved between members (routine clang output) cannot steal subsequent members from the enclosing class. The recorded emitter version is the compiler that actually produced the IR, not whichever llvm-config happens to be on the PATH.

ingest_library(config) -> String

sanitize_ir_for_julia(ir_text::String) -> String

Configuration Manager

Source: src/Builder/ConfigurationManager.jl

The single source of truth for all build settings. Handles TOML parsing, validation, and merging into a typed RepliBuildConfig struct.

Discovery

Source: src/Builder/Discovery.jl

Scans the filesystem to identify C/C++ source files, headers, and dependencies. Auto-detects project language (:c vs :cpp) from the scanned source extensions and sets wrap.language accordingly in the generated replibuild.toml.

discover(target_dir::String=pwd(); force::Bool=false, unsafe::Bool=false, build::Bool=false, wrap::Bool=false) -> String

# Discover only
toml_path = RepliBuild.Discovery.discover()

# Discover and build
toml_path = RepliBuild.Discovery.discover(build=true)

# Full pipeline: discover → build → wrap
toml_path = RepliBuild.Discovery.discover(build=true, wrap=true)

# Then use the TOML path:
RepliBuild.build(toml_path)
RepliBuild.wrap(toml_path)

DWARFParser

Source: src/Builder/DWARFParser.jl

Parses llvm-dwarfdump output to extract structured type information from compiled binaries. This is the bridge between C++ debug metadata and Julia wrapper generation.

Data structures

Extraction targets

export_vtable_json(vtinfo::VtableInfo, output_path::String)

parse_dwarf_output(dwarf_text::String) -> Dict{String, ClassInfo}

parse_symbol_table(nm_output::String) -> Tuple{Dict{String, UInt64}, Dict{String, UInt64}}

parse_vtables(binary_path::String) -> VtableInfo

read_vtable_data(binary_path::String, vtable_addr::UInt64, num_entries::Int) -> Vector{UInt64}

JLCSIRGenerator

Source: src/IRGen/JLCSIRGenerator.jl, src/IRGen/ir_gen/

Transforms parsed DWARF metadata (VtableInfo) into MLIR source text in the JLCS dialect. The generated IR is then parsed and either JIT-compiled by MLIRNative (Tier 2 JIT) or written to disk and AOT-compiled by ThunkBuilder (Tier 2 AOT). Both paths share this module — there is no separate AOT IR generator.

Submodules

Generation flow

generate_jlcs_ir(vtinfo, metadata; needed_symbols) produces a complete MLIR module:

1. Struct aliases + registration: type aliases for all extracted structs (packed structs as !jlcs.c_struct, padded structs as !llvm.struct with packed members inlined as LLVM literals)
2. Type info operations: jlcs.type_info for each class with non-empty members, carrying the DWARF-resolved destructor and the base/virtual-base tables
3. Function thunks: func.func @mangled_thunk wrappers carrying llvm.emit_c_interface — filtered by needed_symbols (the wrapper's thunk manifest, i.e. dead-thunk elimination). Each body unpacks %args_ptr (ciface convention), emits jlcs.marshal_arg for packed-struct parameters, jlcs.scope copy-construct/destruct brackets for non-trivial by-value class parameters, jlcs.ffe_call / jlcs.try_call (per-function noexcept routing) or jlcs.vcall (virtual instance methods with scalar/pointer signatures), and jlcs.marshal_ret for packed-struct returns
4. STL container thunks: Accessor thunks for detected STL containers (size, data, push_back, etc.)
5. Array-view thunks: rank-1 strided accessors for fixed-size primitive array members

DAGDiff

Source: src/IRGen/DAGDiff.jl

Structural type-graph diff used by tier selection and IR generation when a struct may contain other structs whose layouts disagree between Julia and C++. The pairwise check in is_ccall_safe() catches direct packed-vs-aligned mismatches; DAGDiff catches the transitive cases — a non-packed struct that contains a packed struct as a field, a struct chain through a typedef alias, etc. It outputs a topo-sorted lowering order so that the MLIR thunks for dependent types are emitted in the right sequence.

Slicer

Source: src/IRGen/Slicer.jl

Per-function bitcode slicing for Tier 1, on Julia's resident libLLVM. slice_library(abi_ll; targets, cache_dir) parses the promoted module once and clones it per target (LLVMCloneModule), then strips the clone to declarations: LLVMFunctionDeleteBody for every reached function, LLVMSetInitializer2(gv, NULL) for reached mutable and external constant globals, internalize + globaldce for everything unreached. Internal constants are embedded rather than declared. Every slice is verified before it is returned, and results are cached content-addressed under <cache>/slices/.

The closure is one level deep by construction — a declared function contributes no edges of its own — so slice size tracks the target function, not the library. lua_gettop cuts 15.8 MB down to 2.8 KB; luaL_openlibs lands at 6 KB. This is why max_slice_kb is a tripwire rather than a tuning knob.

Anything the Slicer cannot slice correctly comes back as a refusal with a reason, never as silently-wrong IR: a variadic target, a blockaddress into a body being deleted, alias/ifunc, or an unpromoted module (the fail-loud guard against slicing _opt.ll by mistake). Softer shapes come back as hazard flags for the generator's gate — :setjmp_family, :varargs_callee, :noinline, :weak, :inline_asm, :module_asm.

Each SliceResult also records the symbols the slice declares, post-DCE and excluding intrinsics. _tier1_preflight! in the C generator dlopens the .so RTLD_GLOBAL and dlsyms each one — the exact lookup ORC will perform at first call — because an unresolved declare does not raise: ORC prints Symbols not found: [...] and then blocks forever. A miss demotes that function to ccall; a .so that will not dlopen disables Tier 1 for the whole wrap.

ThunkBuilder

Source: src/Builder/ThunkBuilder.jl

AOT compilation path for Tier 2 thunks. When aot_thunks = true in replibuild.toml, this module drives the same JLCSIRGenerator.generate_jlcs_ir() used by the JIT path, lowers the result with MLIRNative.lower_to_llvm(), emits an object file through MLIRNative.emit_object(), and links it against the user's compiled library (clang/clang++ -shared, rpath'd to the library directory) into a companion shared library named <libname>_thunks.so. With [link] enable_lto on it additionally emits and assembles the thunks' own LTO bitcode. An AOT failure is a warning, not a build failure — the JIT path remains available.

The Julia wrapper then ccalls into the AOT thunks rather than calling JITManager.invoke. There is no MLIR JIT at runtime — libJLCS.so is only needed at build time for the lowering step. After AOT compilation, the user can ship the wrapped library + thunks .so without bundling LLVM/MLIR runtime libraries.

MLIRNative

Source: src/IRGen/MLIRNative.jl

Low-level ccall bindings to libJLCS.so, the compiled JLCS MLIR dialect shared library. Provides context management, module parsing, JIT engine creation, LLVM lowering, symbol lookup, and the object/IR emission used by the AOT path. Building the dialect (cd src/mlir && ./build.sh) is required only for the C++/Tier-2 bucket.

Two behaviours here are load-bearing beyond plain FFI plumbing. parse_module names the parse buffer after a content-hashed file it writes under the library's .debug/mlir/ — MLIR's parser stamps that name onto every op as a FileLineColLoc, and the lowering turns it into the emitted DWARF's DIFile, which is what makes a JIT'd thunk steppable in gdb. And jit_source_path falls back to a temp directory when .debug is unwritable (a read-only install), because losing co-location costs nothing while losing the source view costs the whole capability.

The dialect itself — its two types, fourteen ops, lowering pass, and calling contract — is documented in ABI Marshalling as Compiler IR.

JITManager

Source: src/IRGen/JITManager.jl

Runtime for Tier 2 dispatch: one MLIR execution engine per wrapped binary (LibraryEngine), held in a process-wide GLOBAL_JIT behind a shared, lock-free thunk cache.

Key design points

• Per-library engines. initialize_global_jit(binary_path) is called from each generated module's __init__ and creates (or reuses) the engine for its binary. Multiple wrappers coexist in one session — previously the first wrapper won and the second library's entire Tier 2 silently died, found while composing box2d with pugixml. A per-library initialization failure degrades only that library, and a missing-symbol error names every engine that was searched.
• Manifest-driven initialization: initialize_global_jit() reads thunk_manifest.json — the thunks the wrapper actually dispatches to — so dead thunks are never generated. Any initialization failure (including the pre-flight rejection of untranslatable IR types in libJLCS) degrades the module to "Tier 2 disabled" with ccall wrappers intact, never a process crash.
• Symbol registration before lowering: the engine is given the library and libJLCS.so as shared libraries, and the C++ runtime EH symbols (__gxx_personality_v0, __cxa_begin_catch, __cxa_end_catch) plus the jlcs_* exception helpers are registered explicitly, since JIT'd landing pads reference them by name.
• Lock-free hot path: _lookup_cached() reads from an @atomic snapshot of the symbol dictionary with no locking. The cache is published copy-on-write — a fresh dict is built with the new entry and atomically swapped in. Readers always see a stable, immutable snapshot.
• Arity specialization: invoke is @generated, emitting arity-specialized code for any argument count — stack-allocated Refs and a fixed-size Ptr{Cvoid}[], allocation-free at every arity. A thunk slot holds a pointer to the argument's storage, so Ref(x) is the right shape for an isbits x; the two kinds that are already an indirection — an AbstractString, and a Base.Ref that is not a Ptr (what a caller passes for a C++ T const& parameter) — are flattened to a raw pointer first and GC-preserved across the call. Both invoke methods share one _arg_marshal_plan, because two copies is how a fix to one silently misses the other.
• @generated return dispatch: _invoke_call resolves at compile time whether the return type is a primitive (direct ccall return) or a struct (sret buffer allocation). An unresolved Any return fails loudly with the actual cause instead of corrupting memory.
• Exception propagation: After every Tier 2 call, _check_pending_exception() polls the thread-local exception buffer set by jlcs.try_call lowering. If a C++ exception was caught during the call, a CxxException is thrown with the original what() message.

Calling convention

All Tier 2 functions use a unified ciface calling convention:

Debug

Source: src/Debug/Debug.jl

Static inspection of what the Tier-2 pipeline actually emitted, for a package this process never built. thunks(pkg) lists the thunk symbols; mlir_body(pkg, symbol) prints the generated dialect; disassemble(pkg; symbol=…) shells to objdump -dS so dialect ops and machine code interleave; dwarf(pkg; section=…) shows the address → MLIR-line table; walk(pkg, symbol) does the common combination in one call.

The object file it disassembles only exists when the JIT's object cache was enabled, and MLIR requires that at engine-creation time — so it is read from REPLIBUILD_JIT_OBJDUMP before the wrapper loads, not passed as an argument. Nothing here links gdb or LLVM: it shells to objdump and llvm-dwarfdump. See Debugging a thunk for the live-process counterpart.

BuildBridge

Source: src/Builder/BuildBridge.jl

Low-level compiler driver that shells out to clang, clang++, llvm-link, opt, llvm-as, and nm. All subprocess invocations go through this module, providing a single point of control for toolchain interaction. It serves the C++ pipeline and the C bucket's [link] fallback = true escape hatch; the default C path links and optimizes in-process on Julia's libLLVM (see Compiler).

LLVMEnvironment

Source: src/Builder/LLVMEnvironment.jl

Detects the system LLVM/Clang toolchain by searching standard paths and version-suffixed binaries. Falls back to LLVM_full_jll when no system toolchain is found. Caches results in ~/.replibuild/toolchain.toml with a 24-hour TTL.

EnvironmentDoctor

Source: src/Builder/EnvironmentDoctor.jl

check_environment() validates both toolchain buckets: the C bucket (JLL clang + Julia's resident libLLVM — no external install required) and the C++/Tier 2 bucket (system LLVM/MLIR 21+, Clang, mlir-tblgen, CMake 3.20+, and libJLCS.so). Returns a ToolchainStatus struct indicating which tiers are available, with OS-specific install instructions for missing components.

DependencyResolver

Source: src/Builder/DependencyResolver.jl

Processes the [dependencies] table from replibuild.toml. Supports three dependency types:

The exclude list is applied after scanning. Resolved source files merge into the compilation graph before the compile step.

PackageRegistry

Source: src/Builder/PackageRegistry.jl

Local package registry at ~/.replibuild/registry/. Provides:

• register() — Store a project's build configuration
• use() — Build + wrap + load, with artifact caching in ~/.replibuild/builds/<hash>/; on a local miss, fetches the package config from the RepliBuild-Hub community registry
• search() — Query the Hub index by name, description, tags, or language
• list_registry() — Print all registered packages with hash, source, and build status
• unregister() — Remove a package and clean cached builds

The build-cache key (hash_config) covers the TOML, sources, headers, and project git HEAD plus the generator fingerprint — RepliBuild's own version and git revision — so upgrading RepliBuild invalidates wrappers produced by older codegen. Cached wrappers resolve their .so sibling-first via @__DIR__, with the baked absolute path as fallback.

The REPLIBUILD_HOME environment variable overrides the default registry location; REPLIBUILD_HUB_URL points Hub operations at a private mirror.

STLWrappers

Source: src/Wrapper/Cpp/STLWrappers.jl

Detects STL container types (std::vector, std::string, std::map, etc.) in DWARF metadata and generates accessor functions. These are used by the MLIR IR generator (src/IRGen/ir_gen/STLContainerGen.jl) to produce JIT thunks for STL container methods.

ASTWalker

Source: src/Builder/ASTWalker.jl

Clang.jl-based AST walker for enum extraction. Handles enum class, hex values, namespaces, and other constructs that are difficult to extract reliably from DWARF alone. Replaces the earlier regex-based approach.

ClangJLBridge

Source: src/Builder/ClangJLBridge.jl

Integration module for Clang.jl header parsing. Used by the wrapper generator when use_clang_jl = true to supplement DWARF metadata with AST-level information.

Scaffold

Source: src/Builder/PackageRegistry.jl (scaffold_package function)

Generates a distributable Julia package from a registered RepliBuild project. The scaffolded package includes the compiled shared library, generated wrapper module, and a standard Julia Project.toml — ready for Pkg.add().