Skip to content

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.

3.3 KB
Raw Blame

GPU-resident full decompile process (discovered multi-pass SIMT method)

Status: implemented in ghidrust-decomp::gpu_decompile
Date: 2026-07

Problem this solves

Classical decompilation is branchy (CFG, structuring). Prior notes treated “decompile on GPU” as impossible.
Discovery: decompile can be re-factored into fixed-shape multi-pass stages that run entirely in VRAM:

  1. Host uploads the function code region once (single PCIe H→D).
  2. On-device kernels transform device buffers only (no mid-pipeline full-IR download).
  3. Host downloads only the final dump buffer once (D→H) and writes a scan-friendly file.

This avoids PCIe per-instruction / per-block round-trips that make naïve “GPU helper” decompilers lose.

Dataflow (residency)

HOST                         DEVICE (VRAM)
─────                        ─────────────
code_bytes ──upload──►  buf_code
                         │
              [K1 decode_walk]  → buf_ir[MAX_IR]
                         │
              [K2 mark_leaders] → buf_leaders[MAX_IR]
                         │
              [K3 build_blocks] → buf_blocks[MAX_BLOCKS], buf_edges[MAX_EDGES]
                         │
              [K4 emit_text]    → buf_emit[MAX_EMIT]  (ASCII pseudo-C dump)
                         │
host ◄──single map/read──┘
         write file .gdecomp

Invariant: mid_pipeline_host_reads == 0 (instrumented). Only the final emit buffer is read back.

Kernel roles (custom WGSL compute)

Kernel SIMT shape Input (device) Output (device)
decode_walk 1 active walker (or workgroup) walks x86 stream into fixed IR slots buf_code buf_ir
mark_leaders 1 thread per IR slot buf_ir buf_leaders
build_blocks compact leaders → block table + edges buf_ir, buf_leaders buf_blocks, buf_edges
emit_text walk blocks/insns into byte buffer all tables buf_emit

x86 decode uses the same opcode subset as Ghidrust’s hand-rolled CPU decoder for fixture/common prologue patterns (push/mov/xor/pop/ret/int3/jmp/jcc). Bounded multi-pass, not commercial Hex-Rays parity.

Dump format (.gdecomp — scan-friendly)

magic: "GDEC" + version u8=1
entry: u64 LE
name_len: u32 LE + utf8 name
n_ir / n_blocks / n_edges: u32 LE each
emit_len: u32 LE + utf8 pseudo-C

Text section is greppable; header is compact for tooling.

Correctness

  1. Multipass normal form: GPU emit_text and CPU multipass_emit_from_ir share the same template (void FUN, block_N, if (/* jcc */) { goto … }, // fall block_N, return;).
  2. Classic structural oracle: decompile_instructions mnemonics (until first ret) map to the same opcode sequence as multipass IR (structural_ops_match_classic).
  3. Leaders are race-free: mark_leaders only writes 1 into a zero-initialized buffer (never writes 0), so branch targets cannot lose leader status.

Relation to bulk scan

bulk_scan printable/pattern GPU kernels are orthogonal. This process is full decompile stages on GPU, not a rebranded string scan.

CLI

ghidrust gpu-decompile fixtures/tiny_x64.pe --out entry.gdecomp --metrics metrics.json