Decomp permuter

Automatically permutes C files to better match a target binary. The permuter has two modes of operation:

• Random: purely at random, introduce temporary variables for values, change types, put statements on the same line...
• Manual: test all combinations of user-specified variations, using macros like PERM_GENERAL(a = b ? c : d;, if (b) a = c; else a = d;) to try both specified alternatives.

The modes can also be combined, by using the PERM_RANDOMIZE macro.

This tool supports MIPS (compiled by IDO, possibly GCC), PowerPC, and ARM32 assembly.

Usage

./permuter.py directory/ runs the permuter; see below for the meaning of the directory. Pass -h to see possible flags. -j is suggested (enables multi-threaded mode).

You'll first need to install a couple of prerequisites: python3 -m pip install pycparser pynacl toml Levenshtein (also dataclasses if on Python 3.6 or below) pynacl is optional and only necessary for the "permuter@home" networking feature. Levenshtein is optional and only necessary for using the Levenshtein diff algorithm (difflib is used by default).

The permuter expects as input one or more directory containing:

• a .c file with a single function,
• a .o file to match,
• a .sh file that compiles the .c file,
• a .toml file specifying settings.

For projects with a properly configured makefile, you should be able to set these up by running

./import.py <path/to/file.c> <path/to/file.s>

where file.c contains the function to be permuted, and file.s is its assembly in a self-contained file. Otherwise, see USAGE.md for more details.

For projects using Ninja instead of Make, add a permuter_settings.toml in the root or tools/ directory of the project:

build_system = "ninja"

Then import.py should work as expected if build.ninja is at the root of the project.

All of the possible randomizations are assigned a weight value that affects the frequency with which the randomization is chosen. The default set of weights is specified in default_weights.toml and vary based on the targeted compiler. These weights can be overridden by modifying settings.toml in the input directory.

The .c file may be modified with any of the following macros which affect manual permutation:

• PERM_GENERAL(a, b, ...) expands to any of a, b, ...
• PERM_VAR(a, b) sets the meta-variable a to b, PERM_VAR(a) expands to the meta-variable a.
• PERM_RANDOMIZE(code) expands to code, but allows randomization within that region. Multiple regions may be specified. A PERM_RANDOMIZE block is automatically added when there are no PERM macros.
• PERM_FORCE_SAMELINE(code) expands to code, but joined to a single line after round-tripping through the C parser library (which normally puts statements on separate lines). Can be useful for IDO where same-lineness affects codegen.
• PERM_LINESWAP(lines) expands to a permutation of the ordered set of non-whitespace lines (split by \n). Each line must contain zero or more complete C statements. (For incomplete statements use PERM_LINESWAP_TEXT, which is slower because it has to repeatedly parse C code.)
• PERM_INT(lo, hi) expands to an integer between lo and hi (which must be constants).
• PERM_IGNORE(code) expands to code, without passing it through the C parser library (pycparser)/randomizer. This can be used to avoid parse errors for non-standard C, e.g. asm blocks.
• PERM_PRETEND(code) expands to code for the purpose of the C parser/randomizer, but gets removed afterwards. This can be used together with PERM_IGNORE to enable the permuter to deal with input it isn't designed for (e.g. inline functions, C++, non-code).
• PERM_ONCE([key,] code) expands to either code or to nothing, such that each unique key gets expanded exactly once. key defaults to code. For example, PERM_ONCE(a;) b; PERM_ONCE(a;) expands to either a; b; or b; a;.

Arguments are split by a commas, exluding commas inside parenthesis. (,) is a special escape sequence that resolves to ,.

Nested macros are allowed, so e.g.

PERM_VAR(delayed, )
PERM_GENERAL(stmt;, PERM_VAR(delayed, stmt;))
...
PERM_VAR(delayed)

is an alternative way of writing PERM_ONCE.

If any multi-choice PERM macros are provided, automatic randomization will be disabled; to enable it you need to surround the function (or the relevant parts of it) with PERM_RANDOMIZE.

permuter@home

The permuter supports a distributed mode, where people can donate processor power to your permuter runs to speed them up. To use this, pass -J when running permuter.py and follow the instructions. (This can be combined with regular -j flags.) You will need to be granted access by someone who is already connected to a permuter network.

permuter@home is only available for a limited number of compilers (see the list for the main permuter network), and currently does not work on native Windows (but WSL does work).

To allow others to use your computer for permuter runs, do the following:

• install Docker (used for sandboxing and to ensure a consistent environment)
• if on Linux, add yourself to the Docker group: sudo usermod -aG docker $USER or set up rootless Docker
• install required packages: python3 -m pip install docker
• open a terminal, and run ./pah.py run-server to start the worker server. There are a few required arguments (e.g. how many cores to use), see --help for more details.

Anyone who is granted access to permuter@home can run a worker.

To set up a new permuter network, see src/net/controller/README.md.

FAQ

What do the scores mean? The scores are computed by taking diffs of objdump'd .o files, and giving different penalties for lines that are the same/use the same instruction/are reordered/don't match at all. 0 means the function matches fully. Stack positions are ignored unless --stack-diffs is passed (but beware that the permuter is currently quite bad at resolving stack differences). For more details, see scorer.py. It's far from a perfect system, and should probably be tweaked to look at e.g. the register diff graph.

What sort of non-matchings are the permuter good at? It's generally best towards the end, when mostly regalloc changes remain. If there are reorderings or functional changes, it's often easy to resolve those by hand, and neither the scorer nor the randomizer tends to play well with them.

Should I use this instead of trying to match code by hand? No, but it can be a good complement. PERM macros can be used to quickly test lots of variations of a function at once, in cases where there are interactions between several parts of a function. The randomization mode often finds lots of nonsensical changes that improve regalloc "by accident"; it's up to you to pick out the ones that look sensible. If none do, it can still be useful to know which parts of the function need to be changed to get the code nearer to matching. Having made one of the improvements, and the function can then be permuted again, to find further possible improvements.

Helping out

There's tons of room for helping out with the permuter! Many more randomization passes could be added, the scoring function is far from optimal, the permuter could be made easier to use, etc. etc. The GitHub Issues list has some ideas.

Ideally, mypy permuter.py and ./run-tests.sh should succeed with no errors, and files formatted with black. To setup a pre-commit hook for black, run:

pip install pre-commit black
pre-commit install

PRs that skip this are still welcome, however.