decomp-toolkit

Yet another GameCube/Wii decompilation toolkit.

decomp-toolkit functions both as a command-line tool for developers, and as a replacement for various parts of a decompilation project's build system.

Goals

• Automate as much as possible, allowing developers to focus on matching code rather than months-long tedious setup.
• Provide highly accurate and performant analysis and tooling.
• Provide everything in a single portable binary. This simplifies project setup: a script can simply fetch the binary from GitHub.
• Replace common usages of msys2 and GNU assembler, eliminating the need to depend on devkitPro.
• Integrate well with other decompilation tooling like objdiff and decomp.me.

Background

The goal of a matching decompilation project is to write C/C++ code that compiles back to the exact same binary as the original game. This often requires using the same compiler as the original game. (For GameCube and Wii, Metrowerks CodeWarrior)

When compiling C/C++ code, the compiler (in our case, mwcceppc) generates an object file (.o) for every source file. This object file contains the compiled machine code, as well as information that the linker (mwldeppc) uses to generate the final executable.

One way to verify that our code is a match is by taking any code that has been decompiled, and linking it alongside portions of the original binary that have not been decompiled yet. First, we create relocatable objects from the original binary:

(Heavily simplified)

Then, each object can be replaced by a decompiled version as matching code is written. If the linker still generates a binary that is byte-for-byte identical to the original, then we know that the decompiled code is a match.

decomp-toolkit provides tooling for analyzing and splitting the original binary into relocatable objects, as well as generating the linker script and other files needed to link the decompiled code.

Other approaches

Manual assembly

With existing GameCube/Wii decompilation tooling, the setup process is very tedious and error-prone. The general process is:

• Begin by disassembling the original binary with a tool like doldisasm.py. This produces one giant assembly file per section.
• Manually comb through the assembly files and fix many issues, like incorrect or missing relocations, incorrect or missing symbols, and more.
• Manually find-and-replace the auto-generated symbol names based on other sources, like other decompilation projects or a map file. (If you're lucky enough to have one)
• Manually determine data types and sizes, and convert them accordingly. (For example, .4byte -> .float, strings, etc)
• Manually split the assembly files into individual objects. This is a very tedious process, as it requires identifying the boundaries of each function, determining whether adjacent functions are related, finding associated data from each data section, and cut-and-pasting all of this into a new file.

Other downsides of this approach:

• Manually editing the assembly means that the result is not reproducible. You can't run the script again to make any updates, because your changes will be overwritten. This also means that the assembly files must be stored in version control, which is not ideal.
• Incorrectly splitting objects is very easy to do, and can be difficult to detect. For example, a .ctors entry must be located in the same object as the function it references, otherwise the linker will not generate the correct .ctors entry. extab and extabindex entries must also be located in the same object as the function they reference, have a label and have the correct size, and have a direct relocation rather than a section-relative relocation. Otherwise, the linker will crash with a cryptic error message.
• Relying on assembly means that you need an assembler. For GameCube/Wii, this means devkitPro, which is a large dependency and an obstacle for new contributors. The assembler also has some quirks that don't interact well with mwldeppc, which means that the object files must be manually post-processed to fix these issues. (See the elf fixup command)

With decomp-toolkit:

• Many analysis steps are automated and highly accurate. Many DOL files can be analyzed and split into re-linkable objects with no configuration.
• Signature analysis automatically labels common functions and objects, and allows for more accurate relocation rebuilding.
• Any manual adjustments are stored in configuration files, which are stored in version control.
• Splitting is simplified by updating a configuration file. The analyzer will check for common issues, like incorrectly split .ctors/.dtors/extab/extabindex entries. If the user hasn't configured a split for these, the analyzer will automatically split them along with their associated functions to ensure that the linker will generate everything correctly. This means that matching code can be written without worrying about splitting all sections up front.
• The splitter generates object files directly, with no assembler required. This means that we can avoid the devkitPro requirement. (Although we can still generate assembly files for viewing, editing, and compatibility with other tools)

dadosod

dadosod is a newer replacement for doldisasm.py. It has more accurate function and relocation analysis than doldisasm.py, as well as support for renaming symbols based on a map file. However, since it operates as a one-shot assembly generator, it still suffers from many of the same issues described above.

ppcdis

ppcdis is one of the tools that inspired decomp-toolkit. It has more accurate analysis than doldisasm.py, and has similar goals to decomp-toolkit. It also has some features that decomp-toolkit does not yet, like support for REL files.

However, decomp-toolkit has a few advantages:

• Faster and more accurate analysis. (See Analyzer features)
• Emits object files directly, with no assembler required.
• More robust handling of features like common BSS, .ctors/.dtors/extab/extabindex, and more.
• Requires very little configuration to start.
• Automatically labels common functions and objects with signature analysis.

Honorable mentions

splat is a binary splitting tool for N64 and PSX. Some ideas from splat inspired decomp-toolkit, like the symbol configuration format.

Terminology

DOL

A DOL file is the executable format used by GameCube and Wii games. It's essentially a raw binary with a header that contains information about the code and data sections, as well as the entry point.

ELF

An ELF file is the executable format used by most Unix-like operating systems. There are two common types of ELF files: relocatable and executable.

A relocatable ELF (.o, also called "object file") contains machine code and relocation information, and is used as input to the linker. Each object file is compiled from a single source file (.c, .cpp).

An executable ELF (.elf) contains the final machine code that can be loaded and executed. It can include information about symbols, debug information (DWARF), and sometimes information about the original relocations, but it is often missing some or all of these (referred to as "stripped").

Symbol

A symbol is a name that is assigned to a memory address. Symbols can be functions, variables, or other data.

Local symbols are only visible within the object file they are defined in.
These are usually defined as static in C/C++ or are compiler-generated.

Global symbols are visible to all object files, and their names must be unique.

Weak symbols are similar to global symbols, but can be replaced by a global symbol with the same name.
For example: the SDK defines a weak OSReport function, which can be replaced by a game-specific implementation.
Weak symbols are also used for functions generated by the compiler or as a result of C++ features, since they can exist in multiple object files. The linker will deduplicate these functions, keeping only the first copy.

Relocation

A relocation is essentially a pointer to a symbol. At compile time, the final address of a symbol is not known yet, therefore a relocation is needed. At link time, each symbol is assigned a final address, and the linker will use the relocations to update the machine code with the final addresses of the symbol.

Before:

# Unrelocated, instructions point to address 0 (unknown)
lis r3, 0
ori r3, r3, 0

After:

# Relocated, instructions point to 0x80001234
lis r3, 0x8000
ori r3, r3, 0x1234

Once the linker performs the relocation with the final address, the relocation is no longer needed. Still, sometimes the final ELF will still contain the relocation information, but the conversion to DOL will always remove it.

When we analyze a file, we attempt to rebuild the relocations. This is useful for several reasons:

• It allows us to split the file into relocatable objects. Each object can then be replaced with a decompiled version, as matching code is written.
• It allows us to modify or add code and data to the game and have all machine code still to point to the correct symbols, which may now be in a different location.
• It allows us to view the machine code in a disassembler and show symbol names instead of raw addresses.

Analyzer features

Function boundary analysis
Discovers function boundaries with high accuracy. Uses various heuristics to disambiguate tail calls from inner-function control flow.

Signature analysis
Utilizes a built-in signature database to identify common Metrowerks and SDK functions and objects.
This also helps decomp-toolkit automatically generate required splits, like __init_cpp_exceptions.

Relocation analysis
Performs control-flow analysis and rebuilds relocations with high accuracy.
With some manual tweaking (mainly in data), this should generate fully-shiftable objects.

Section analysis
Automatically identifies DOL and REL sections based on information from signature and relocation analysis.

Object analysis
Attempts to identify the type and size of data objects by analyzing usage.
Also attempts to identify string literals, wide string literals, and string tables.

Splitting
Generates split object files in memory based on user configuration.
In order to support relinking with mwldeppc.exe, any unsplit .ctors, .dtors, extab and extabindex entries are analyzed and automatically split along with their associated functions. This ensures that the linker will properly generate these sections without any additional configuration.
A topological sort is performed to determine the final link order of the split objects.

Object file writing
Writes object files directly, with no assembler required. (Bye devkitPPC!)
If desired, optionally writes GNU assembler-compatible files alongside the object files.

Linker script generation
Generates ldscript.lcf for mwldeppc.exe.

Future work

• Support REL and RSO files
• Add more signatures
• Rework CodeWarrior map parsing

Commands

ar create

Create a static library (.a) from the input objects.

$ dtk ar create out.a input_1.o input_2.o
# or
$ echo input_1.o >> rspfile
$ echo input_2.o >> rspfile
$ dtk ar create out.a @rspfile

demangle

Demangles CodeWarrior C++ symbols. A thin wrapper for cwdemangle.

$ dtk demangle 'BuildLight__9CGuiLightCFv'
CGuiLight::BuildLight() const

dol info

Analyzes a DOL file and outputs information section and symbol information.

$ dtk dol info input.dol

dol split

[!NOTE]
This command is a work-in-progress.

Analyzes and splits a DOL file into relocatable objects based on user configuration.

$ dtk dol split config.yml target

TODO: document configuration file

dwarf dump

Dumps DWARF 1.1 information from an ELF file. (Does not support DWARF 2+)

$ dtk dwarf dump input.elf

elf disasm

Disassemble an unstripped CodeWarrior ELF file. Attempts to automatically split objects and rebuild relocations when possible.

$ dtk elf disasm input.elf out

elf fixup

Fixes issues with GNU assembler-built objects to ensure compatibility with mwldeppc.exe.

• Strips empty sections
• Generates section symbols for all allocatable sections
• Where possible, replaces section-relative relocations with direct relocations.
• Adds an (asm) suffix to the file symbol. (For matching progress calculation)

# input and output can be the same
$ dtk elf fixup file.o file.o

elf2dol

Creates a DOL file from the provided ELF file.

$ dtk elf2dol input.elf output.dol

map

[!WARNING]
This command is currently broken.

Processes CodeWarrior map files and provides information about symbols and TUs.

$ dtk map entries Game.MAP 'Unit.o'
# Outputs all symbols that are referenced by Unit.o
# This is useful for finding deduplicated weak functions,
# which only show on first use in the link map.

$ dtk map symbol Game.MAP 'Function__5ClassFv'
# Outputs reference information for Function__5ClassFv
# CodeWarrior link maps can get very deeply nested,
# so this is useful for emitting direct references
# in a readable format.

rel info

Prints basic information about a REL file.

$ dtk rel info input.rel

rel merge

Merges a DOL file and associated RELs into a single ELF file, suitable for analysis in your favorite reverse engineering software.

$ dtk rel info main.dol rels/*.rel -o merged.elf

rso info

[!WARNING]
This command is not yet functional.

Prints basic information about an RSO file.

$ dtk rso info input.rso

shasum

Calculate and verify SHA-1 hashes.

$ dtk shasum baserom.dol
949c5ed7368aef547e0b0db1c3678f466e2afbff  baserom.dol

$ dtk shasum -c baserom.sha1 
baserom.dol: OK