botw/Contributing.md

Contributing

To contribute to the project, you will need:

• A disassembler or a decompiler such as Hex-Rays or Ghidra.
• Python 3 and pip for the diff script
• These Python modules: capstone colorama cxxfilt pyelftools (install them with pip install ...)
• The original 1.5.0 main NSO executable, converted to ELF format with nx2elf.
  • To dump it, follow the instructions on the wiki. If you only have 1.6.0 on your console, you should still dump it.
  • Decompress the main NSO with hactool.
  • If you have a 1.6.0 dump, use xdelta3 on the decompressed NSO to turn it into a 1.5.0 NSO. The patch is available here.
  • The uncompressed NSO has the following SHA256 hash: d9fa308d0ee7c0ab081c66d987523385e1afe06f66731bbfa32628438521c106
  • Copy it to data/main.elf -- it is used for the diff script and other tools.

Experience with reverse engineering optimized C++ code is very useful but not necessary if you already know how to decompile C code.

Using a decompiler is strongly recommended for efficiency reasons. If you have IDA 7.0+, ping @leoetlino to get a copy of the IDC which will make decompilation easier and help with understanding the code more generally.

Feel free to join the Zelda Decompilation Discord server if you have any questions.

How to decompile

0. Open the executable in the disassembler of your choice.

1. Pick a function that you want to decompile.

   • Prefer choosing a function that you understand or that is already named in your IDA/Ghidra database.
   • You do not need to fully understand the function, but you should at least have a rough idea of what it does.
   • If you are feeling more ambitious, pick an entire C++ class! This usually allows understanding the code better.

2. Decompile it using Hex-Rays or Ghidra.

   • Rename variables, add structures, do everything you can to make the output as clean as possible.
   • Again, understanding the function is very important.
   • C++ code tends to make heavy use of inline functions. For example, inlined string comparisons or copies are very common and tend to obscure what the function does. Focus on the outline of the function.

3. Implement the function in C++.

   • Stay close to the original code, but not too close: your code should mostly look like normal, clean C++ code. If it does not, chances are that you won't get a good match at all.
   • Keep in mind that decompilers can only produce C pseudocode. Some function calls may be member function calls.
   • Identify inlined functions and uninline them. For example, if you see a string copy, do not write the copy loop manually! Instead, call the inline function and let the compiler inline the function for you.
   • Identify duplicate pieces of code: those are usually a sign that functions have been inlined.

4. Build.

5. Get the mangled name of your function. For example, if you are decompiling BaseProcMgr::createInstance:

   $ tools/print_decomp_symbols.py -a | grep BaseProcMgr::createInstance
   UNLISTED  ksys::act::BaseProcMgr::createInstance(sead::Heap*) (_ZN4ksys3act11BaseProcMgr14createInstanceEPN4sead4HeapE)
   

6. Add the mangled function name to the list of decompiled functions.

   • To do so, open data/uking_functions.csv and search for the name or the address of function you have decompiled, and add the mangled function name to the last column.
   • Example: 0x00000071010c0d60,sub_71010C0D60,136,_ZN4ksys4util13TaskQueueBaseD1Ev

7. Compare the assembly with ./diff.py --source <mangled function name>

   • This will bring up a two-column diff. The code on the left is the original code; the code on the right is your version of the function.
   • You may ignore address differences (which often show up in adrp+ldr pairs or bl or b).

8. Tweak the code to get a perfectly matching function.

   • Clang is usually quite reasonable so it is very common for functions -- even complicated code -- to match on the first try.
   • Focus on large differences. If you have large differences (e.g. entire sections of code being at the wrong location), focus on getting rid of them first and ignore small differences like regalloc or trivial reorderings.
   • Regalloc: If you only have regalloc differences left in a function that looks semantically equivalent, double-check whether it is truly equivalent: such differences are typically caused by using the wrong variable. It is rare for LLVM to use a different set of registers if the code is equivalent.
   • This is usually the most difficult part of matching decomp. Please ask on Discord if you need help!

9. Update the list of decompiled functions.

   • If you have a function that matches perfectly, great!
   • If there are still minor differences left, add a // NON_MATCHING: comment to explain what is wrong and add a ? at the end of the mangled function name in the CSV.
   • For major differences (lots of entirely red/green/blue lines in the diff), add a ! at the end of the function name.

10. Before opening a PR, reformat the code with clang-format and run tools/check.py.

Non-inlined functions

When implementing non-inlined functions, please compare the assembly output against the original function and make it match the original code. At this scale, that is pretty much the only reliable way to ensure accuracy and functional equivalency.

However, given the large number of functions, certain kinds of small differences can be ignored when a function would otherwise be equivalent:

• Regalloc differences.

  • Warning: ensure that the return type of the function is correct. Differences that involve the X0-X7, W0-W7 or S0-S3 registers at the end of a function are suspicious.

• Instruction reorderings when it is obvious the function is still semantically equivalent (e.g. two add/mov instructions that operate on entirely different registers being reordered)

Header utilities or inlined functions

For header-only utilities (like container classes), use pilot/debug builds, assertion messages and common sense to try to undo function inlining. For example, if you see the same assertion appear in many functions and the file name is a header file, or if you see identical snippets of code in many different places, chances are that you are dealing with an inlined function. In that case, you should refactor the inlined code into its own function.

Also note that introducing inlined functions is sometimes necessary to get the desired codegen.

If a function is inlined, you should try as hard as possible to make it match perfectly. For inlined functions, it is better to use weird code or small hacks to force a match as differences would otherwise appear in every single function that inlines the non-matching code, which drastically complicates matching other functions. If a hack is used, wrap it inside a #ifdef MATCHING_HACK_{PLATFORM} (see below for a list of defines).

Matching hacks

This project sometimes uses small hacks to force particular code to be generated by the compiler. Those have no semantic effects but can help with matching assembly code especially when the hacks are used for functions that are inlined.

• MATCHING_HACK_NX_CLANG: Hacks for Switch, when compiling with Clang.

For people who are familiar with C or other decomp projects

Given the impossibility of automatically splitting the assembly and generating a matching binary, the sheer size of the main executable and the usage of many software libraries, this project takes a different and somewhat experimental approach to matching decompilation.

Instead of trying to match the entire executable, each function is matched individually and source code is organized in whichever way makes the most sense. Libraries are not treated as being part of the game code, but as external dependencies. The result is that the codebase looks a lot more like a regular software project than a decompilation codebase. Since C++ code makes heavy use of inline functions and zero-cost abstractions that disappear in compiled code, contributors have a lot more leeway when it comes to organizing files and adding abstractions.

How easy is it to get matching code?

Compared to other decomp projects for older compilers: extremely easy. So outrageously easy that it is almost unfair to their contributors.

Clang is an extremely reasonable compiler with much fewer memes than older compilers such as IDO:

• Stack reordering issues are extremely rare, given that AArch64 uses its registers a lot more efficiently. And even when the stack is used, things Just Work™ in the vast majority of cases.
• Pure register allocation (regalloc) issues are almost non-existent. If you see something that looks like a regalloc problem, it usually means your code is not semantically equivalent.
• No if (1) shenanigans.
• No same line memes (codegen being different if two statements are put on the same line).
• Whitespace doesn't matter.

In general, two equivalent constructs that should clearly produce the same code actually produce the exact same code. There are exceptions, of course, but many things simply do not matter at all for matching. Inline functions do sometimes affect codegen, though.

Getting perfect matches on the first try happens pretty routinely, even for medium-sized and large functions (>1kB).

Most functions tend to call several other inline functions, notably utility functions from sead; as many core sead modules have already been reversed, decompiling a function sometimes only requires recognizing those function calls: decompilation at a higher level of abstraction!

Writing proper C++

Unlike most other decompilation projects, this one targets a large modern game that is written in C++. While C and C++ have similar syntax, C++ is somewhat more complex than C and has many more language features. To avoid getting lost in C++ code, please familiarize yourself with the following, preferably before decompiling:

• namespaces
  • Instead of using prefixes such as z_ to avoid name conflicts, C++ code instead relies on namespaces.
• classes, including inheritance, polymorphism, virtual functions
  • C++ classes/structs are basically C structs on steroids. Notably, C++ classes can contain member functions.
  • Member functions get an implicit this argument, which is passed as if it were the first argument.
  • Virtual member functions are member functions that can be overridden in derived classes.
  • Virtual member functions are usually implemented with a virtual function table or "vtable", which is a table of function pointers.
• const correctness for member functions
• iterators
• range-based for loops
• using nullptr instead of NULL
• C++11 / C++14 / C++17 features more generally

Project tools

• Check all decompiled functions for issues: tools/check.py
• To compare assembly: ./diff.py <mangled function name>
  • The function must be listed in data/uking_functions.csv first.
    • To do so, search for the name or the address of function you have decompiled, and add the mangled function name to the last column.
  • Pass the --source flag to show source code interleaved with assembly code.
  • Add the --inlines flag to show inline function calls. This is not enabled by default because it usually produces too much output to be useful.
  • For more options, see asm-differ.
• To print progress: tools/progress.py
  • Note that progress is only approximate because of inline functions, templating and compiler-generated functions.
• To print AI class decompilation status: tools/ai_progress.py
  • Use this to figure out which AI classes have not been decompiled yet.
• To dump symbols: tools/print_decomp_symbols.py

  • -u for undefined symbols (default)

  • -a for all symbols

  • Useful for getting the mangled name of a function. For example:

    $ tools/print_decomp_symbols.py -a | grep BaseProcMgr::createInstance
    UNLISTED  ksys::act::BaseProcMgr::createInstance(sead::Heap*) (_ZN4ksys3act11BaseProcMgr14createInstanceEPN4sead4HeapE)