Targeted C rewrites for matching decompilation. Give it a function's C. It returns variant sources, each changing one thing the compiler is sensitive to: a literal, a local's lifetime, which register carries an argument, how a branch is laid out, the order of stores. You compile each one with the game's original compiler and keep it only when the bytes match.
It only proposes. It never compiles, and it never decides whether something matches: your compiler and your judge (objdiff, a linked ROM, decomp.me) do that. That keeps it safe to run beside any matching workflow.
It complements decomp-permuter. The permuter makes thousands of random tree changes. These are a few hundred deliberate rewrites, each tied to a pattern old compilers are known for.
It started in decompdecomp, which uses it in its repair pipeline. This repository has kept its history.
MIPS first: most families and all the code readers understand MIPS (N64, PS1, PS2, PSP). Families that only read C work for any processor. The interface may change before 1.0.
uv tool install git+https://github.com/RetroReversing/decomp-repair
decomprepair src/game.c --function func_80012345 --arch mips --out proposals/
decomprepair src/game.c --function func_80012345 --timing # how long each family took
decomprepair src/game.c --function func_80012345 --json # proposals and timings as JSONFrom Python:
from decomprepair.propose import propose
report = propose(source, 'func_80012345', arch='mips', header=header, asm=target_asm)
for proposal in report.proposals:
print(proposal.family, proposal.module) # compile proposal.source and compare the bytes
for timing in report.timings:
print(timing.module, timing.ms, timing.skipped or timing.error)Some families need more than the C: the header (header), the processor (arch), the target's assembly (asm)
or the compiler flags (flags). Leave an input out and those families are skipped; the report says which input
each one needed.
The package is flat. Each *_repairs.py module is one family, with a variants(source, name, ...) function that
returns sources or (label, source) pairs. The other modules are the readers the families share: csource for C
text, mips and mips_* for MIPS code, symbols for symbol tables, m2cfix for m2c output. propose.py runs
them all.
The mips_* readers also give evidence about a byte difference that a repair can act on: argument and literal
differences, frame sizes, return paths, control-flow regions, runtime helpers, small leaf templates. They are
hints, never match verdicts. mips_count, mips_compare, mips_status, mips_fcsr and
mips_status_return_spacing propose fixed CP0/FPU leaves that C can't express; mips_debug_break.header() is the
header a compiler puts before each file so they compile, and their candidate(asm, name, registered) proposes only
when the caller says its compiler includes it.
The permuter's rewrites are here too. permuter makes simple random rewrites (statement and declaration order,
casts, register), permuter_local a few cheap ordering hypotheses, and permuter_initialization rejects variants
that read a local before it's set. permuter_gen.py is a program, not a module: run it by path with
decomp-permuter's Python and folder, and it writes that
randomizer's variants, guarded by its helpers beside it (permuter_capture, permuter_dependencies,
permuter_formatting). Running the candidates, compiling and scoring them is the caller's job.
Also: switch_tables (case groups in the target's jump-table order), switch_constant_dispatch,
initializer_clusters, loop_storage_shapes, callee_contracts (prototype mismatches), layout_frontiers
(reapply an edit along a saved layout of the source), expression_recipes (templates from an edit that matched;
storing them is the caller's), repair_tokens (lossless C tokens), and ps1_gte, the fixed PS1 GTE macros a
compiler can put before each file.
From decompdecomp 1.0.10-alpha.2: permuter_gen reads MSVC's words (__int64, calling conventions,
__declspec) and a C++ function's own C (msvc_as_c, cxx_as_c, Restorer), so PC and Xbox near misses reach
decomp-permuter; csource reads conditional compilation (surely_compiled, hidden_by, at_file_scope); and
mips says what's alignment padding (is_padding, ends_in_delay_branch).
indexed_record_repairs proposes bounded compositions for repeated indexed-record copies.
It uses existing record layouts, pure subscripts and raw MIPS base evidence to try subarray
anchors, retained word loads, explicit conversion lifetimes, byte-store order, byte decodes
and narrow index carriers. Unrelated ordinary fields remain eligible beside a volatile word;
existing volatile access order is preserved. Candidate-owned word/counter qualification is
an explicit target-supported hypothesis, never a shared-header change.
At most 64 complete recipe candidates are generated. There is no 512-instruction body cutoff for this lane; source size, record counts and component choices remain bounded. Unknown layouts, effectful indices, qualified byte stores, ambiguous roots, observable line macros and unsupported conversion macros decline. The library never compiles or accepts a result.
uv sync
uv run pytest -q
uv run ruff check src testsNothing in src/decomprepair may import decompdecomp (decompnet); tests/test_library.py checks this.
AGPL-3.0-only (LICENSE), like decompdecomp, which this code comes from. Some readers began in decomp.wtf
(LICENSE-decomp.wtf).