Skip to content

0x70 Roadmap

From foundations to full parity.

Eleven milestones, headless core first. Generated from docs/PLAN.md; later milestones are intentionally coarse and are planned in detail when they are reached. The interface (M7) and the GNU disassembler (M10) were started early.

  1. M0Complete7/7 tasks

    Foundations

    wave 1, complete 2026-11-12

    • Workspace, lints, gate, CI, licence, docs skeleton, test checklist
    • bk-core + bk-marshal: address spaces, addresses, varnode data, XML and packed encode/decode (features/core-marshal.md)
    • bk-pcode: opcodes, op behaviours, float formats, multiprecision (features/pcode.md)
    • bk-sleigh-syntax: SLEIGH preprocessor, lexer, parser; parses all 150 .slaspec files (features/sleigh-syntax.md)
    • bk-lang: .ldefs, .pspec, .opinion, language registry (features/lang.md)
    • bk-bin + bk-elf: byte providers, binary reader, ELF parser (features/bin-elf.md)
    • bk-testkit + reference tools: build Ghidra's sleigh and decomp, compile every .sla, oracle wrappers (features/reference-tools.md)
  2. M1Mostly complete4/5 tasks

    SLEIGH runtime

    Tracks and order: ADR-023. Undefined behaviour in Ghidra's engine: ADR-024.

    • sla — bk-sla: the compiled-language model, validating reader, byte-exact writers; all 150 reference files round-trip (features/sla.md)
    • sleigh-decode — bk-sleigh: loading, context database, instruction decoding, operand printing; I/E/T records equal to pcodedump for every language (features/sleigh-decode.md)
    • sleigh-pcode — bk-sleigh: p-code generation (templates, handles, delay slots, crossbuild, user ops), Translate, basilisk sleigh decode; every record equal to pcodedump, the 20-operand exception aside (features/sleigh-pcode.md)
    • Benchmarks: language load time, decode and lift throughput at or above the reference (PERF.md): load 20 ms against 32; single-threaded decode 1.11× and decode + p-code 1.00× the g++ reference on x86-64 Linux, 1.32× and 1.38× on the M1; 6.7–7.9× at 8 threads. A clear margin for decode + p-code on x86-64 is a queued follow-up (compiled templates)
    • sleigh-language — bk-sleigh, bk-language: registers as a model, flow and operand types, overrides, compile on demand (features/sleigh-language.md, a draft; later wave)
  3. M2Mostly complete3/4 tasks

    SLEIGH compiler

    • sleigh-compiler-front — bk-sleigh-compiler: semantic analysis over bk-sleigh-syntax (symbols, tokens, constructors, macros, with blocks), p-code template compilation; symbols, print pieces, context changes and templates equal to Ghidra's for all 150 specifications (features/sleigh-compiler-front.md)
    • sleigh-compiler-back — bk-sleigh-compiler: pattern resolution, decision-tree construction, consistency checks, basilisk sleigh compile; all 150 specifications compile to the same inflated bytes as Ghidra's C++ compiler; SleighCompileRegressionTest ported as that equality (features/sleigh-compiler-back.md)
    • sleigh-snippet — bk-sleigh-snippet: p-code snippets of .cspec/.pspec (137 elements in 49 files; 588 compilations over 116 languages) compile to Ghidra's template with no difference; reference tool snippetdump; PcodeParserTest ported (features/sleigh-snippet.md)
    • Compiler diagnostics equal to the C++ compiler's, one case per message (in sleigh-compiler-front/-back)
  4. M3Current1/9 tasks

    Decompiler core, standalone

    Runs on Ghidra's <decompilertest> data tests (bytes + language + script), with no program database. Crates: ADR-026. How parity is measured: ADR-027. What the port carries for later improvements: ADR-025, ADR-028. The 31 tracks, their order, sizes and acceptance are in features/decomp-plan.md; this list follows its waves.

    • Wave 1 (leaf tracks; all five tracks done: D01, D02, D03 on 2026-11-14, D00 and D04 on 2026-11-15 — features/decomp-oracle.md, features/decomp-cfg.md, ADR-042; one follow-up in progress, the state dump's @scope section, format version 2): D01 bk-decomp-range (117 testcirclerange tests; features/decomp-range.md) · D02 bk-decomp-types (3 of the 8 testtypes tests, the other five need the IR and move to D08; features/decomp-types.md) · D03 bk-cspec (21 testfuncproto, 5 testparamstore, nine Java prototype-model classes; features/decomp-cspec.md) · D00 oracle tooling (state dumps, debug console, recorded p-code; features/decomp-oracle.md) · D04 bk-decomp-cfg block graph and structuring (features/decomp-cfg.md)
    • Wave 2: D05 bk-decomp-ir (features/decomp-ir.md, revised 2026-11-15; ADR-046) · D06 bk-decomp-symbols and the host trait (features/decomp-symbols.md; ADR-043, 044, 045; in progress) · D16a printer spike ✔ (2026-11-16: the answer is yes, ADR-048)
    • Wave 3: D07 core (function state, units, profiles, drivers) · D08 typeop and casts · D16c emitter
    • Wave 4: D09 flow and the data-test runner · D10 heritage · D11 actions A · R1–R9 rule families · D19 corpus extension (32 rules apply in no corpus yet)
    • Wave 5: D12 actions B · D13 jump tables · D14 merge
    • Wave 6: D15 structuring on the IR · D16b output tree
    • Wave 7: D17 driver · D18 injection (needs bk-sleigh-snippet)
    • All 93 data tests pass; differential runs against reference decomp
    • basilisk decompile on a data-test file; per-function benchmark against Ghidra
  5. M4Planned0/7 tasks

    Program model, storage, loaders

    • ref-headless: Ghidra itself built and run headless on the build host as the oracle for everything that is Java in Ghidra — per imported file an export of memory blocks, symbols, relocations, functions, references and data types in a comparable form; readers in bk-testkit (the M4 counterpart of D00; without it loaders and the program model could only be checked by ported unit tests)
    • Storage engine spike and decision (ADR-003), then bk-store
    • bk-program: memory blocks, address map, code units, symbols and namespaces, references, functions and variables, equates, comments, bookmarks, properties, relocations, register values
    • bk-datatypes: built-in types, composites, archives, data organisation from .cspec
    • bk-loader + ELF, PE, Mach-O, raw binary, Intel hex; relocation handlers per processor
    • basilisk import
    • From the roadmap check against Ghidra's source (research/plan-programs/roadmap-check.md), also in M4: the corpus of test binaries (ADR-057); a program builder for tests (449 Ghidra test files use ProgramBuilder); the domain-object layer — transactions, undo, events, options — which Ghidra keeps in Framework/Project; external locations, program trees, the source map, function tags; the read-only reader of Ghidra's packed databases if the owner allows it (the 16 shipped .gdt archives and the .fidb files; the analyzer applying the archives is on by default); the file-open part of the file-system layer; the binary, hex and XML exporters (moved from M6); the editing commands with undo and redo (moved from M7, which binds them to actions)
  6. M5Planned0/6 tasks

    Disassembly and auto-analysis

    • Flow-following disassembler, instruction prototypes, context propagation; the command basilisk disasm on a loaded program (ADR-047)
    • Auto-analysis manager and scheduling
    • Analyzers in Ghidra's priority order: entry points, function starts (pattern files), references, switch tables, stack, data and strings, no-return, call-fixups, demanglers, format-specific analyzers
    • Decompiler wired to the program (parameter ID, switch analysis, decompiler-dependent analyzers)
    • Block models, symbolic propagation, the byte-search engine. Ghidra's default analysis runs two decompiler analyzers and several analyzers filed under M6: M5's whole-binary oracle runs with those switched off until they land (analysis profiles, ADR-056)
    • basilisk analyze (headless: import, analyse, export); whole-binary differential tests against Ghidra headless
  7. M6Planned0/5 tasks

    Headless core complete

    • Exporters (C, XML/SARIF, binary, hex), import/export round trips
    • DWARF, PDB, GNU and Microsoft demanglers (the GNU one ported from libiberty, ADR-018), Function ID
    • Emulator (bk-emu) and the processor emulator tests (126 classes); needs sleigh-language, its tests need M4 and built binaries; Ghidra's JIT is 45,439 of the module's 71,882 lines and needs a decision (owner)
    • RTTI class recovery (28,042 lines of Ghidra scripts)
    • Assembler (bk-asm)
  8. M7Planned0/3 tasks

    egui UI

    • Shell: docking, pop-out windows, actions and key bindings, task and progress, options
    • Listing, decompiler view with cross-highlighting, byte viewer, symbol tree and table, data type manager, function graph, references, search, bookmarks, memory map
    • Editing commands with undo/redo; program diff
  9. M8Planned0/1 tasks

    Projects and collaboration

    • Projects, versioned programs, multi-user server, merge
  10. M9Planned0/5 tasks

    Plugin API

    ADR-014, ADR-035, ADR-036, ADR-037

    Python hosts Basilisk through an extension module (basilisk-py); the tracks and their acceptance are in features/plugin-api.md. The work is spread over the milestones its parts depend on rather than left to the end.

    • P0 reference fixtures · P1 host and packaging (python -m basilisk, wheels on three systems, the binary's hand-over) · P3 the flat scripting layer — can start now
    • P2 program-model bindings · P4a loaders — after M4
    • P4b analyzers and exporters — after M5/M6 · P5 decompiler bindings and proposers — after the decompiler is wired to a program
    • P6 interface hooks (bk-plugin-ui) — after bk-ui's action registry and view kinds accept run-time entries
    • P7 ports of Ghidra's bundled scripts (28 % port once the program model is bound, 53 % with decompiler, data-type and analysis bindings); importable module in PyGhidra's role
  11. M10Planned0/2 tasks

    Remaining modules

    • Debugger and trace model, Version Tracking, BSim, file-system browser and container formats, remaining loaders and processors' Java extensions, everything left in PARITY.md
    • Ghidra's remaining GPL modules: DMG; the GNU disassembler over binutils built from source (bk-binutils-sys, bk-gnu-disasm, ADR-019)