Ghidrust provides a hyper fast, AI native reverse engineering platform built in Rust, inspired by Ghidra. Includes experimental GPU kernels to rapidly decompile large files, a decryption suite, and network monitor for packet to binary reverse engineering.
GPU strategy matrix — all Auto Analysis types + decompile
Each analyzer has a dedicated GPU method class (not one printable kernel labeled 20 ways).
CPU remains the correctness oracle. Timing always splits PCIe (upload+download) vs on-device (kernels).
Research write-up with measured tables: PARALLEL_RE_RESEARCH.md (lab fixture, 8 MiB pad, ~175 MB RTTI seed, decompile multipass).
| Analyzer | Strategy class | GPU method (novel kernel / multipass) |
|---|---|---|
| ASCII Strings | printable_run |
Parallel printable-byte mark + host run compact |
| Unicode Strings | cstr_multi |
Host UTF-16LE scan; GPU seed uses multi-needle cstr family |
| Aggressive Instruction Finder | code_density |
Non-int3 / code-like density windows over exec image |
| Call Convention ID | prologue_abi |
Scan push/mov/sub-rsp / ret patterns → ABI hint seeds |
| Call-Fixup Installer | cstr_multi |
Multi-needle cstr for security_cookie / known fixups |
| Create Address Tables | ptr_chain |
Parallel LE u64 “in-image-looking” density seeds |
| Decompiler Parameter ID | spill_scan |
Scan rcx/rdx/r8/r9 spill encodings (48 89 4c/54/…) |
| Decompiler Switch Analysis | ptr_chain |
Jump-table density (same family as address tables) |
| Demangler Microsoft | cstr_multi |
.?AV / ? mangled prefix multi-needle |
| Embedded Media | magic_media |
PNG/JPEG/GIF/RIFF magic parallel match |
| Function ID | hash_window |
Fixed-window rolling hash over exec prologues |
| Function Start Search | prologue_seed |
55 48 89 e5 and 48 83 ec start seeds |
| Non-Returning Functions - Discovered | cstr_multi |
ExitProcess / abort-style API name needles |
| PDB MSDIA | cstr_multi |
MSF / PDB signature strings |
| PDB Universal | cstr_multi |
Same portable MSF/signature family (distinct needle set id) |
| Shared Return Calls | ret_epilogue |
c3 / c2 epilogue sites in exec |
| Stack | stack_frame |
sub rsp / mov [rbp+…] frame markers |
| Variadic Function Signature Override | cstr_multi |
printf/scanf family needles |
| WindowsPE x86 PE RTTI Analyzer | rtti_scan |
.?AV / .?AU / _ZTS type-info name scan |
| Windows x86 Propagate External Parameters | cstr_multi |
Known WinAPI name needles |
| WindowsResourceReference | magic_res |
VS_VERSION_INFO / resource markers |
| GPU Decompile (not Auto Analysis) | decomp_multipass |
VRAM multipass decode→leaders→blocks→emit |
Timing model
pcie_ms = t_upload + t_download (host↔device for that run)
device_ms = t_kernels (on-chip / VRAM pipeline; SIMT workgroup_size=256)
wall_ms ≈ setup + pcie_ms + device_ms (setup reported separately when cold)
GPU decompile VRAM multipass reports the same split (pcie_upload_ms / device_ms / pcie_download_ms) via bench_vram_decompile_vs_cpu, which also records CPU multipass wall (cpu_ms) and oracle equality.
Equality model (honest)
Analyzers — equal is seed-stage on the same haystack (including --large tiled pad):
- CPU:
cpu_emulate_kernel(algorithm twin of WGSL) - GPU: SIMT kernel atomic hit total + offset multiset (when under cap)
Not compared for analyzer equal: filtered Auto Analysis oracle counts (analyzer_oracle).
Host merge applies GPU seeds into Program analysis fields (merged_primary).
VRAM decompile — equal requires:
mid_pipeline_host_reads == 0normalize_pseudo(GPU) == normalize_pseudo(CPU multipass)ir_countmatch
Large mode does not force equal=true.
CLI
# fixture baseline (non-large)
ghidrust analyzer-bench fixtures/analysis_lab.pe --out metrics_fixture.json
# large tiled workload (≥8 MiB)
ghidrust analyzer-bench fixtures/analysis_lab.pe --large --out metrics_large.json
ghidrust analyzer-bench-matrix # print strategy matrix