This repository holds benchmarking code for various x86/x86-64 disassembler libraries.
Kubuntu 21.04, gcc 10.3.0, Rust 1.52.1, i5-6600K.
Test file: xul.dll from Firefox 86.0.1.7739 (77MB code section).
| Library | MB/s | % |
|---|---|---|
| iced (Rust) | 256.69 MB/s | 100.00% |
| yaxpeax (Rust) | 182.08 MB/s | 70.93% |
| diStorm (C) | 92.17 MB/s | 35.91% |
| udis86 (C) | 80.72 MB/s | 31.44% |
| Zydis (min) (C) | 61.33 MB/s | 23.89% |
| XED (C) | 55.39 MB/s | 21.58% |
| bddisasm (C) | 35.48 MB/s | 13.82% |
| Zydis (C) | 34.19 MB/s | 13.32% |
BeaEngine (C), Capstone (C) don't support decode only.
| Library | MB/s | % |
|---|---|---|
| iced (Rust) | 149.21 MB/s | 100.00% |
| diStorm (C) | 58.85 MB/s | 39.44% |
| yaxpeax (Rust) | 40.05 MB/s | 26.84% |
| Zydis (C) | 22.24 MB/s | 14.91% |
| bddisasm (C) | 16.00 MB/s | 10.72% |
| XED (C) | 15.67 MB/s | 10.50% |
| Capstone (C) | 14.02 MB/s | 9.39% |
| BeaEngine (C) | 13.42 MB/s | 8.99% |
| udis86 (C) | 12.02 MB/s | 8.05% |
Windows:
REM Start "x64 Native Tools Command Prompt for VS 2019"
REM Start git bash:
"C:\Program Files\Git\bin\bash.exe"Windows/Linux/macOS:
git clone --recursive 'https://github.com/icedland/disas-bench.git'
cd disas-bench
./make-all.sh
# Windows: python
python3 -mvenv venv
# Windows: source venv/Scripts/activate
source venv/bin/activate
pip install -r requirements.txt
# Optional args: <code-offset> <code-len> <filename> [loop-count]
python bench.pyThe optional bench.py arguments are:
<code-offset>= offset of the code section (in decimal or 0x hex)<code-len>= length of the code section (in decimal or 0x hex)<filename>= 64-bit x86 binary file to decode and format[loop-count]= optional loop count. Total number of bytes decoded and formatted is<code-len> * [loop-count]
You can use dumpbin.exe (Windows) or objdump to get the offset and size of the code section.
Find .text section and use <code-len> = virtual size (4CFA6B6 below) and <code-offset> = file pointer to raw data (400 below). All values are in hex so add a 0x prefix when passing the values to bench.py.
C:\path> dumpbin -headers filename.dll
...
SECTION HEADER #1
.text name
4CFA6B6 virtual size
1000 virtual address (0000000180001000 to 0000000184CFB6B5)
4CFA800 size of raw data
400 file pointer to raw data (00000400 to 04CFABFF)
...
Find .text section and set <code-len> to the first 32-bit value (04cfa6b6 below) and <code-offset> to the last 32-bit value (00000400 below). All values are in hex so add a 0x prefix when passing the values to bench.py.
$ objdump -h filename
...
Sections:
Idx Name Size VMA LMA File off Algn
0 .text 04cfa6b6 0000000180001000 0000000180001000 00000400 2**4
CONTENTS, ALLOC, LOAD, READONLY, CODE
...
If you feel like the benchmark for a lib doesn't drive it to its full potential or treats it unfairly, I'd be happy to accept PRs with improvements!