This document is intended to be an introductory guide to decomp (tailored specifically for Dinosaur Planet). If this is your first time with decomp then please read on! If you're already familiar with decomp then you may still find some useful information here, as this document is a little specific to Dinosaur Planet.
- 1. Introduction
- 2. Decompiling a Function
- 3. Symbols
- 4. Data Sections
- 5. Optimization Levels and Debug Symbols
- 6. Decompilation Tools
It's important to first understand exactly what this project is attempting to accomplish. The goal is to create a matching decompilation. This means that when all reversed code is recompiled, the re-created ROM is byte-for-byte identical to the original.
This process is typically done on a per-function basis. The raw machine code for a function is extracted/disassembled from the original ROM into readable MIPS assembly code. Then, that function is decompiled into C code that when recompiled results in the exact same machine code that it started from.
Before continuing, please consider reading the overview of the project to get an idea of where everything is located. Don't worry if not everything makes sense yet, this document will attempt to explain more in detail.
Let's go over the full process of decompiling a single function. This section will be using the already decompiled function vec3_normalize as an example (you can find the final source in src/vec3.c).
Before we can begin decompiling a function, we need to extract its assembly code from the ROM. This is done automatically using a tool called splat. Currently, splat is already configured to extract most functions. Splat's configuration for Dinosaur Planet can be found at splat.yaml in the repository root.
Splat works by mapping an address range of the ROM to individual files. For example, vec3_normalize is currently part of the vec3.c file. You can see this mapping in splat.yaml:
# From 0x17180 until the next entry is mapped to src/vec3.c
- [0x17180, c, vec3]With this configuration, splat will take the recompiled contents of vec3.c and stitch them into the recompiled ROM starting at that exact address.
The final step is to temporarily include the original assembly of each function that is part of the file. This is done using the special pragma GLOBAL_ASM:
#pragma GLOBAL_ASM("asm/nonmatchings/vec3/vec3_normalize.s")Splat extracts each function it finds into the asm/nonmatchings directory as individual .s files grouped by the name of the mapped file they are a part of. The GLOBAL_ASM pragma is used by another tool, asm-processor, which is responsible for actually including the referenced .s file at that position in the C source.
Great! Now, we can start to decompile this function.
First, let's create some space to start decompiling and reversing vec3_normalize. We'll use an #if directive to temporarily replace the GLOBAL_ASM pragma with an actual C function:
#if 0
#pragma GLOBAL_ASM("asm/nonmatchings/vec3/vec3_normalize.s")
#else
// We don't know anything about the function signature yet, so we'll
// just make it void for now.
void vec3_normalize() {
}
#endifNext, let's look at the full assembly code of the function:
glabel vec3_normalize
addiu $sp, $sp, -0x18
sw $ra, 0x14($sp)
lwc1 $f2, ($a0)
lwc1 $f14, 4($a0)
lwc1 $f0, 8($a0)
mul.s $f4, $f2, $f2
sw $a0, 0x18($sp)
mul.s $f6, $f14, $f14
add.s $f8, $f4, $f6
mul.s $f10, $f0, $f0
jal sqrtf
add.s $f12, $f10, $f8
mtc1 $zero, $f16
lw $a0, 0x18($sp)
mov.s $f12, $f0
c.eq.s $f0, $f16
lui $at, 0x3f80
bc1tl .L8001678C
mov.s $f0, $f12
mtc1 $at, $f18
lwc1 $f4, ($a0)
lwc1 $f10, 4($a0)
div.s $f2, $f18, $f0
lwc1 $f16, 8($a0)
mul.s $f6, $f4, $f2
nop
mul.s $f8, $f10, $f2
nop
mul.s $f18, $f16, $f2
swc1 $f6, ($a0)
swc1 $f8, 4($a0)
swc1 $f18, 8($a0)
mov.s $f0, $f12
.L8001678C:
lw $ra, 0x14($sp)
addiu $sp, $sp, 0x18
jr $ra
nop
At this point, it's very helpful to have a MIPS assembly manual on hand so we can look up the meaning of each instruction. Each line in the above disassembly is a single instruction in the format mnemonic [operand1,operand2,...]. For example, the first instruction is the mnemonic 'addiu' with the operands 'sp,sp,-0x18'.
For the first step of decompilation, we'll convert each individual instruction into the equivalent C code:
void vec3_normalize(f32 *a0) {
// addiu $sp, $sp, -0x18 ; We'll skip the first two instructions,
// sw $ra, 0x14($sp) ; which are just part of the stack setup
// ; First is some float value loads from $a0,
// lwc1 $f2, ($a0) ; the first parameter of the function
f32 f2 = *a0;
// lwc1 $f14, 4($a0)
f32 f14 = *(a0+4)
// lwc1 $f0, 8($a0)
f32 f0 = *(a0+8)
// mul.s $f4, $f2, $f2 ; Next, is some arithmetic
f32 f4 = f2 * f2;
// sw $a0, 0x18($sp) ; There's an upcoming function call that
// ; doesn't use the $a0 register, so the code
// ; must save it to the stack first
*(sp+0x18) = a0;
// mul.s $f6, $f14, $f14
f32 f6 = f14 * f14;
// add.s $f8, $f4, $f6
f32 f8 = f4 + f6;
// mul.s $f10, $f0, $f0
f32 f10 = f0 * f0;
// jal sqrtf ; Here we have a function call with a
// add.s $f12, $f10, $f8 ; branch delay slot. The delay slot is
// ; executed first in this case.
f32 f12 = f10 + f8;
// ; Note: Normally, function calls will use
// ; the a0-a3 registers but sqrtf is
// ; implemented in assembly. Looking at its
// ; code shows it takes $f12 as its input
// ; and outputs to $f0
f32 f0 = sqrtf(f12);
// mtc1 $zero, $f16
f32 f16 = 0;
// lw $a0, 0x18($sp) ; $a0 is restored from the stack
a0 = *(sp+0x18);
// mov.s $f12, $f0
f32 f12 = f0;
// c.eq.s $f0, $f16 ; Here we have a conditional branch with
// lui $at, 0x3f80 ; a 'likely' delay slot. The delay slot
// bc1tl .L8001678C ; in this case is only executed if the
// mov.s $f0, $f12 ; condition is true
int at = 0x3f800000;
if (f0 == f16) {
f0 = f12;
goto L8001678C;
}
// mtc1 $at, $f18 ; More arithmetic
f32 f18 = at;
// lwc1 $f4, ($a0)
f32 f4 = *a0;
// lwc1 $f10, 4($a0)
f32 f10 = *(a0+4);
// div.s $f2, $f18, $f0
f32 f2 = f18 / f0;
// lwc1 $f16, 8($a0)
f32 f16 = *(a0+8);
// mul.s $f6, $f4, $f2
f32 f6 = f4 * f2;
// nop
// mul.s $f8, $f10, $f2
f32 f8 = f10 * f2;
// nop
// mul.s $f18, $f16, $f2
f32 f18 = f16 * f2;
// swc1 $f6, ($a0)
*a0 = f6;
// swc1 $f8, 4($a0)
*(a0+4) = f8;
// swc1 $f18, 8($a0)
*(a0+8) = f18;
// mov.s $f0, $f12
f32 f0 = f12;
L8001678C:
// lw $ra, 0x14($sp) ; Finally, the stack is restored from
// addiu $sp, $sp, 0x18 ; the previous frame
// jr $ra ; Note: Returning a float uses the $f0
// ; register instead of $v0 like normal.
// ; We can infer that a float is returned
// ; since $f0 is set at the end of the
// ; function and not used.
return f0;
// nop
}Whew! That was a bit of code to go through and annotate. It's worth mentioning at this point that auto-decompilers can be used instead of manually translating each instruction, such as mips2c and Ghidra.
It's unlikely that a tool will decompile assembly into a perfectly matching function, but it can still save a lot of time.
Now that we have each instruction translated, we can start analyzing the function and start reducing code into readable C (currently the code won't even compile).
The first interesting code we can see is the use of $a0:
f32 f2 = *a0;
f32 f14 = *(a0+4);
f32 f0 = *(a0+8);
// ...
*a0 = f6;
*(a0+4) = f8;
*(a0+8) = f18;This looks a lot like the access of struct fields. From this, we could assume that $a0 holds a pointer to a structure with three four-byte wide fields. Since this function is known to be working with three-dimensional vectors, let's use the following definition:
typedef struct {
f32 x, y, z;
} Vec3f;Additionally, we were able to infer that this function returns a single float. Now that the return value and parameters of the function are known, let's update the signature:
f32 vec3_normalize(Vec3f *a0)Next up, let's start reducing and cleaning up the function implementation. Many of the local variables derived from $f registers aren't individual variables but are part of larger expressions.
f32 vec3_normalize(Vec3f *a0) {
f32 f18, f2;
f32 f12 = sqrtf((a0->z * a0->z) + ((a0->x * a0->x) + (a0->y * a0->y)));
if (f12 == 0.0f) {
goto L8001678C;
}
f18 = 0x3f800000;
f2 = f18 / f12;
a0->x = a0->x * f2;
a0->y = a0->y * f2;
a0->z = a0->z * f2;
L8001678C:
return f12;
}This looks significantly better already (and compiles)! Let's check our progress so far by diffing our implementation against the original ROM. We can do this by building the ROM (./dino.py build) and then running diff on the function's symbol (./dino.py diff vec3_normalize). Looks like everything matches except for one instruction: lui at,0x3f80, which from our implementation is currently lui at,0x4e7e. If we trace this instruction back to the C code, we can see it's from the line f18 = 0x3f800000; Let's fix it.
Tip: When running
diffconsider passing the flags-m -o -w(or just-mow). This will tell diff to run build first, compare object files (to display symbol names), and watch the source file for changes and auto re-build, respectively.
Looking back at the assembly, we can see these two instructions:
lui at,0x3f80
mtc1 at,$f18
These two instructions are actually 'bitwise' casting the integer 0x3f800000 into a float, rather than simply assigning the float 1,065,353,216.0 to f18. Let's use a floating point converter to see what 0x3f800000 is bit-for-bit as a float. Turns out it's actually 1.0! If we replace the statement with f18 = 1.0f;, re-build, and re-run the diff, we can now see that the function matches!
Before we call this function done however, let's clean it up by getting rid of the goto and using better variable names based on what we know about vector normalization:
Note: Most functions will not use
gotoat all and usually won't match with them still in place. Instead, branches should be translated intoifand sometimes simpleswitchstatements.
f32 vec3_normalize(Vec3f *v) {
f32 length, lengthInv;
length = sqrtf((v->z * v->z) + ((v->x * v->x) + (v->y * v->y)));
if (length != 0.0f) {
lengthInv = 1.0f / length;
v->x = v->x * lengthInv;
v->y = v->y * lengthInv;
v->z = v->z * lengthInv;
}
return length;
}Congratulations 🎉, you've gone through the full decompilation process of a function!
There is still so much more to reversing assembly, but this is the general process of breaking it down one step at a time. For much more information on decompiling assembly from IDO, please see the ever-evolving -O2 decompilation (for IDO 5.3 and 7.1).
An important part of decomp is giving names to memory addresses, known as symbols. This includes global/local static variables and functions.
Part of splat's job is to automatically generate symbols for addresses that it sees referenced. The full list of these generated symbols can be found (after running ./dino.py extract) in the undefined_funcs_auto.txt and undefined_syms_auto.txt files in the repository root. Initially, splat will also generate a name for each symbol, such as func_<address> and D_<address> for functions and variables respectively. These names can be overwritten using the symbol_addrs.txt file (also in the repository root).
The format of each line in symbol_addrs.txt is as follows:
<symbol name> = <VRAM address in hex>; // type:<data or func> size:<byte size in hex>
For example:
gActorCount = 0x800b2934; // type:data size:0x2
texture_load = 0x8003cda8; // type:func
The size attribute is optional.
Sometimes, there will be an address that should have a symbol that splat was unable to detect. This usually happens when the address is not explicitly referenced in the final assembly code. In this case, the symbol cannot be added to symbol_addrs.txt and instead must be added to undefined_funcs.txt or undefined_syms.txt for functions and variables respectively (found in the repository root). These files are fed directly into the linker and as such should not contain the type or size attributes (just the format <name> = <VRAM address in hex>;).
Note: When modifying
symbol_addrs.txt, you will need to re-run./dino.py extractto re-generate the linker the script with the new symbols. If you just modifyundefined_funcs.txtorundefined_syms.txt, then this is not required.
Each compilation unit (.c files, .s, etc.) contributes code and data to various sections of its resulting object file (and ultimately the final linked ELF file). The common sections are:
.text- Executable machine code, each compiled function ends up here..data- Initialized global/local static variables..rodata- Initialized global/local static constants (also includes things like jump tables and string literals)..bss- Uninitialized global/local static variables and constants.
The .data, .rodata, and .bss sections are known as data sections (or data segments). Over the course of decomp, these sections need to be defined in the splat.yaml configuration for each file. Data sections are defined very similarly to code in splat, for example:
# From 0x92370 until the next entry is the .data section (in the ROM)
# for the resulting object file of src/crash.c
- [0x92370, .data, crash]The preceding example will tell splat to generate the linker script such that the .data section of crash.c's object file (build/src/crash.o) will be inserted at the address 0x92370 in the final ROM.
Note: Modifying anything in
splat.yamlwill require you to re-run./dino.py extractto re-extract code/data and, more importantly, to re-generate the linker script.
As an example, let's take a look at the function func_8005D3A4, which happens to be the first function in video.c with a switch statement that compiles into a jump table.
If we look through the first few lines of assembly, we can see that the function references a jump table at the VRAM address 0x8009ACC0. In order to get this function to match, the .rodata section for the file (video.c) needs to marked in splat.yaml such that the switch statement, when compiled, places its jump table at 0x8009ACC0. To do this, we'll first need to convert the VRAM address of the jump table to its ROM address.
To convert the address, we need to first look at the current splat config. The configured segment for code contains:
start: 0x1000 # ROM offset
vram: 0x80000400 # VRAM offsetWith this, we can convert the VRAM address to the ROM address as follows:
(0x8009ACC0 - 0x80000400) + 0x1000 = 0x9B8C0
Finally, we tell splat about the .rodata section for video.c:
- [0x9B8C0, .rodata, video]Note: This works in
video.c's case because this jump table is the first piece of data contributed to.rodatafor the file. If there was, for example, another jump table further up in the file, then splat would need the ROM address of that jump table.
Just about all game code in Dinosaur Planet was originally compiled with the same compiler flags. However, the game statically links other libraries that were compiled differently. The two IDO cc flags that primarily influence assembly generation are:
-O[0-3]- Defines the optimization level to compile with (-O0is off and-O3is all optimizations).-g[0-3]- Defines the symbol tables to be generated for debugging (-g0is off and-g3is for full debugging).
The flag combinations currently known for the decomp are:
-O2 -g3- The default. Used for almost everything.-O2 -g0- Used for compiling modified libultra code.-O1 -g0- Used for compiling unmodified libultra code.
Sometimes, to get a function to match you will need to change the optimization and debugging levels that the containing file is compiled with. This can be done currently by placing the file under a directory named O1 (for -O1 -g0) or g0 (for -O2 -g0).
Note: The direct parent directory must be
O1org0for the flags to be changed. For example,src/O1/stuff/file.cwill not work, butsrc/stuff/O1/file.cwill.
There are many tools out there to assist with decomp. You can find a full list of tools and other resources in the contribution guide.
Here are some recommended tools:
decomp.me is an online collaborative space for decomp. Here, you can create a 'scratch' for a function and get a live diff of your current C implementation and the expected assembly code. You can also share a link to your function for others to play around with. The site has a preset for the Dinosaur Planet decomp that will automatically set up the correct compiler version and default flags.
Note: To use decomp.me, you'll need a 'context' file for the function you're working on. This is essentially just a C header file that contains all of the declarations of types, functions, and macros that are used by your function. You can generate this file by running
./dino.py context <path to your file.c>(you may need to remove function implementations at the bottom of the generated file).
mips2c can be used to automatically decompile a function from assembly into C. There is an online and offline version of the tool. This tool can save a lot of time by producing an initial C implementation to work with that you can then tweak until the function matches.
Ghidra is a full reverse engineering tool suite. It can be used for many things such as exploring the full assembly of the ROM, getting instant C decompilations of each function, tracing references between addresses, and testing function and structure definitions. Ghidra can be a bit overwhelming at first, but once learned it can be extremely valuable for analyzing and understanding functions and data structures.
Note: To use Ghidra with the Dinosaur Planet ROM, you will need the custom loader plugin N64LoaderWV-DinoPlanet. This plugin allows Ghidra to read the N64 style ROM and analyze Dinosaur Planet DLL code.