smda

A minimalist recursive x86 / x64 / AArch64 disassembler library, optimized for accurate Control Flow Graph (CFG) recovery from PE, ELF, and Mach-O binaries and arbitrary memory dumps.

The output is a collection of functions, basic blocks, and instructions with their respective edges (block-to-block, function-to-function). Optionally, references to the Windows API can be inferred via the ApiScout method.

smda-rs is a Rust port of danielplohmann/smda (Python). It powers capa-rs, the Rust port of Mandiant's capability extractor.

Features

• Input formats: PE (32 / 64-bit), ELF (32 / 64-bit), Mach-O (Intel + ARM64, thin and fat).
• Architectures: x86, x86_64, AArch64 (0.6.0+).
• Function discovery: prologue scan (MSVC + GCC / clang endbr64 family + ARM64 stp x29, x30, [sp, #-N]!), call-target propagation, PE exception-handler (.pdata) seeding, PE export-table seeding.
• Per-function output: basic blocks, in / out references, API calls (ApiScout — embedded Win7 + WinXP DBs), stack-string refs, block-to-block edges, is_exported, PIC + opcode hashes, dominator tree + nesting depth.
• Report-level: oep, find_function_by_offset / find_block_by_offset lookups, per-disassembly timeout via parse_with_timeout.

Architecture-aware decoding (0.6.0)

The decoder lives behind a small Decoder trait with two backends:

• X86Decoder — wraps iced_x86. Variable-width, 32 / 64-bit modes. Same x86 path as 0.5.x; zero behavioural change.
• Aarch64Decoder — wraps disarm64. Fixed 4-byte instructions, 64-bit only. Validated at 98%+ clean memory-operand extraction on real Apple-silicon ARM64 binaries (Rust release builds, /bin/ls) before integration.

Smda decides which decoder to use. The caller passes &[u8]; smda inspects the header and routes:

• ELF e_machine == EM_AARCH64 (183) → AArch64.
• PE coff_header.machine == 0xAA64 → AArch64.
• Mach-O cputype == CPU_TYPE_ARM64 (0x100000C) → AArch64. For fat (universal) binaries, the slice preference is configurable via SmdaConfig::macho_arch_preference: default is HostNative (picks the slice matching the host machine — ARM64 on Apple-silicon, x86_64 on Intel/AMD Linux/Windows), with explicit Aarch64First / X86_64First / X86First overrides for analysts who want consistent slice selection regardless of host.
• Everything else falls through to the existing x86 32/64-bit detection.

DecodedInsn is an enum (X86(IcedInsn) / Aarch64(ArmInsn)); the typed accessors on function::Instruction (mnemonic_enum, op_kind, memory_base, flow_control, is_call, is_jmp, is_ret, format_mnemonic, format_operands, length, bytes_in, get_printable_len) keep their 0.5.x signatures and dispatch internally.

ARM64 function-discovery depth in 0.6.0 is minimum-viable — exports + entry point as candidate seeds, then the recursive call-target propagation does the rest. A typical ARM64 executable with no exports will surface 1 function (the entry point) and everything it calls; an ARM64 dylib with N exports surfaces N + transitively-reachable functions. The x86 prologue-scan analysers don't have ARM64 equivalents in this release — that, plus the deeper passes (jump-table walking, indirect-call register tracking, tail-call detection past b/bl, ARM64 PE .pdata packed unwind, typed AArch64 operand extraction for downstream offset: rules in capa-rs, AArch64 mnemonic IDF), is the 0.6.1 work. x86/x64 binaries are unaffected — same code, same output as 0.5.2.

Carried over from 0.5.x

• Zero-copy disassembly. BinaryInfo<'a> borrows the input bytes directly. No mapped-image allocation, no per-instruction byte clone, no DisassemblyReport.buffer.
• Modern Linux ELF coverage: added GCC / clang endbr64 (F3 0F 1E FA) plus the extended GCC AMD64 prologue family (48 89 5C 24 ??, 48 83 EC ??, 41 57 41 56). On CET-enabled binaries (most modern distros) function discovery improves dramatically — one test ELF went from 3280 → 10106 functions. MSVC samples unchanged.
• Linux exit-syscall recognition: mov eax, 60; syscall (and exit_group / int 0x80 equivalents) now end the containing function correctly.
• PE exports as candidate seeds: the export RVA list, previously only surfaced in the public report, now seeds the function-candidate scanner. Free coverage win on stripped DLLs.
• New report fields: report.oep (original entry point VA), function.is_exported (PE only), function.stringrefs (VAs of stack-string writes — wires up the existing Instruction::get_printable_len).
• New lookups: report.find_function_by_offset(addr) / find_block_by_offset(addr).
• Timeout support: Disassembler::parse_with_timeout(..., Duration) + new Error::AnalysisTimeout for batch processors of untrusted samples.
• Section-table abstraction. Byte access goes through binary_info.bytes_at(va, len) -> Result<&[u8]>, which looks up the VA in a small per-binary SectionMap table and returns a borrowed slice into the input. Replaces the old contiguous mapped image.
• Instruction slimmed down. The 0.3.x per-instruction mnemonic: String, operands: Option<String>, and bytes: String (hex) fields are gone. Use the typed iced accessors (mnemonic_enum(), op_kind(), flow_control(), …) for hot paths, or format_mnemonic() / format_operands() / bytes_in(&binary_info) for on-demand formatting.
• Decoders are pure-Rust — iced-x86 for x86 (no C/C++ build dep, ~2–3× faster than capstone) and disarm64 for AArch64 (table-generated from the ARM ISA JSON, MIT-licensed, also no C dep).
• Same security guards. All the checked-arithmetic, allocation caps, and bounds checks added in 0.3.0 are preserved — the pe::map_binary and elf::map_binary rewrites kept every defensive check, just changed the return type from Vec<u8> to Vec<SectionMap>.

Quick start

Add to your Cargo.toml:

[dependencies]
smda = "0.6"

Then disassemble a file:

use smda::{Disassembler, SmdaConfig};

fn main() -> smda::Result<()> {
    // Load the file yourself — the report borrows from this buffer
    // for the lifetime `'a`, so it must outlive the report.
    let buf = std::fs::read("Sample.exe")?;

    // 0.5.0: positional bool args were replaced by SmdaConfig so new
    // analysis knobs land without further API breaks. Every field has
    // a sensible default; chain only what you need.
    let cfg = SmdaConfig::new()
        .path("Sample.exe")
        .high_accuracy(false)        // slower, finds more functions
        .resolve_tailcalls(false);   // promote tail-call targets to functions

    let report = Disassembler::parse(&buf, &cfg)?;

    println!("format       : {:?}", report.format);
    println!("architecture : {:?}", report.architecture);
    println!("bitness      : {}", report.bitness);
    println!("base addr    : 0x{:x}", report.base_addr);
    println!("functions    : {}", report.functions.len());

    for (addr, func) in report.get_functions()?.iter().take(5) {
        let blocks = func.get_blocks()?;
        let insns  = func.get_num_instructions()?;
        println!("  0x{:08x}  {} blocks, {} insns", addr, blocks.len(), insns);
    }
    Ok(())
}

For raw memory dumps (shellcode, unpacked modules — x86 / x64 only in 0.6.0; ARM64 shellcode needs file-format wrapping until 0.6.1 ships an arch arg here):

use smda::{Disassembler, SmdaConfig};
use std::time::Duration;

let shellcode: &[u8] = &[/* … */];
let cfg = SmdaConfig::new().timeout(Duration::from_secs(10));
let report = Disassembler::parse_buffer(
    shellcode,
    0x1000,     // virtual base address
    64,         // bitness (32 or 64)
    &cfg,
)?;

Typed iced accessors

Each Instruction carries the fully-decoded iced_x86::Instruction (16 bytes, Copy) and exposes typed accessors. New code should prefer these over the on-demand string formatters — no allocation, no string parsing.

use smda::function::Instruction;
use smda::BinaryInfo;
use iced_x86::{FlowControl, Mnemonic, OpKind};

fn classify(ins: &Instruction, bi: &BinaryInfo<'_>) {
    // On-demand formatting (allocates a fresh String per call —
    // cache locally if you read it more than once per instruction).
    println!(
        "{:08x}  {:7} {}",
        ins.offset,
        ins.format_mnemonic(),
        ins.format_operands().unwrap_or_default(),
    );

    // Raw instruction bytes, borrowed from the input file (zero-copy).
    if let Ok(bytes) = ins.bytes_in(bi) {
        println!("  bytes: {}", hex::encode(bytes));
    }

    // Typed accessors — no string parsing, no allocation.
    if ins.is_call() {
        println!("  -> call");
    }
    if ins.is_conditional_jmp() {
        println!("  -> Jcc to 0x{:x}", ins.near_branch_target());
    }
    if ins.mnemonic_enum() == Mnemonic::Xor
        && ins.op_count() == 2
        && ins.op_kind(0) == OpKind::Register
        && ins.op_kind(1) == OpKind::Register
        && ins.op_register(0) == ins.op_register(1)
    {
        println!("  -> register clear ({:?})", ins.op_register(0));
    }
    if ins.flow_control() == FlowControl::Return {
        println!("  -> return");
    }
}

Requirements

• Rust 1.95 or newer (2024 edition).
• No C/C++ toolchain required — pure Rust.

Why a Rust port?

smda-rs exists to give capa-rs and other Rust-side static-analysis tools a fast, dependency-light recursive disassembler without pulling in capstone, vivisect, or a Python runtime.

Used by

• capa-rs — static capability extractor for PE / ELF / shellcode / .NET binaries.

License

Licensed under the MIT License.

Acknowledgements

• danielplohmann/smda — original Python implementation by Daniel Plohmann and Steffen Enders.
• iced-x86 — the Rust decoder powering the disassembler backend.