FuzzARM

A random basic test ROM generator for the GBA. It tests both ARM and THUMB mode instructions (Data processing/ALU operations/direct adding/subtraction/loading and storing in various different ways), so basically the basic functions of the CPU.

I plan on making it optional to add certain tests (for example, adding only THUMB tests, or not allowing load/store tests to be generated)

Disclaimer I do not own FASMARM in any way shape or form, nor have I contributed to it. I merely use it in this project to compile the generated .asm ROMS.

Generating a ROM:

You can generate a ROM in 2 ways:

From the command line You can generate a ROM from the command line. This way gives you more options, you can select instruction types to be tested, and if you want only, some or no THUMB mode tests. py main.py -h for more information:

usage: main.py [-h] [-T {some,all,none}] [-nM] [-nD] [-nLS] [--S SEED] N

A fuzzer for the Nintendo GameBoy Advance

positional arguments:
  N                   Number of tests to generate

optional arguments:
  -h, --help          show this help message and exit
  -T {some,all,none}  Amount of THUMB mode tests to generate, choose from 'some', 'all' or 'none'
  -nM                 Disable multiply tests
  -nD                 Disable data processing tests
  -nLS                Disable load/store tests
  --S SEED            Seed to seed the generator with

Double clicking If you double click main.py (with python installed on your system, and a python executable in the PATH), you will be prompted with how many tests you want to generate. Enter a valid number, and this many tests of all types will be generated. To disable tests, use the command line option described above.

As for the number of tests: you can give any value greater than or equal to 1. Note that you will have to make sure that the ROM size will not get too big, so you should not do over about ~1600000 tests (I did about 1000 for testing this script, that should be enough really, but if you really want you can add way more)

After this, the script will finish running and make a main.gba file in the same folder, or overwrite the file with that name if it exists.

ROM generation requirements:

The only libraries I used are random, numpy and subprocess. I think these are all builtin in Python (though I'm not sure about numpy).

ROM running requirements:

If you are building an emulator, I wrote this so that it can be ran from the very early stages of emulation. I added a dump to eWRAM of the results, but other than that it uses rendering mode 4 to display the results, so you will need that. Other than that, I do not use any SWIs/DMAs/Timers/IRQs/Weird IO registers (Only DISPCNT and KEYINPUT). You should be fine having most the basic instructions implemented, I use conditional operations, branches, arithmetic operations, and PSR transfers, but that should be about it.

I think these ROMs should be very useful for debugging purposes, as they do not have any timing requirements, as mentioned above. If the ROM doesn't run at all on your emulator, you should be able to compare logs with established emulators like mGBA or Nanoboy Advance (VBA does not always pass apparently).

The eWRAM dump:

Below is an example for the eWRAM dump for a failed test (VBA does not pass signed multiplication tests):

It dumps the results in the following format (also described in /asm/run_tests.asm:

1 word:  ['AAAA' OR 'TTTT'] for ARM or THUMB state
2 words: [opcode + shift] OR [multiplication opcode] OR [store opcode/load opcode]
1 word:  [????]

1 word:  [initial r0]
1 word:  [initial r1]
1 word:  [initial r2]
1 word:  [initial CPSR]

1 word:  [gotten  r3]
1 word:  [gotten  r4]
1 word:  [0000 0000]
1 word:  [gotten  CPSR]

1 word:  [expected r3]
1 word:  [expected r4]
1 word:  [0000 0000]
1 word:  [expected CPSR]

So the opcode format is for example tst lsl , or smull , or strh/ldrsh , padded with spaces to make it of length 12.

The ROMS in the repo:

If you don't want to generate ROMs yourself, I generated a few for you!

ARM_DataProcessing.gba: 10000 ARM mode Data Processing type tests.
ARM_Any.gba: 10000 ARM mode tests of all types.
THUMB_DataProcessing.gba: 10000 THUMB mode Data Processing type tests.
THUMB_Any.gba: 10000 THUMB mode tests of any type.
FuzzARM.gba: 10000 tests of any mode and of any type.

Output:

If you pass all tests, the ROM will display "End of testing" right away. If you fail any test, the ROM will hang on that test, and display the following things:

what instruction you failed (e.g. tst r0, r1, asr r2)
the input parameters (e.g. r0 001A5342, etc.)
the values it got
the values it expected

Note: the format for the opcode for multiplication seems to be a bit off, I'll have to fix that in the future. You can look at the way the tests work below to see more accurately what happens.

You can press any button (except L or R) to continue testing until you hit the next error. Once there are no more tests failing, "End of testing" will be displayed as well (though if you have failed tests before, old text might still be displayed).

The data processing tests work as follows:

r1 is shifted by r2 with the proper shift type, and stored in r3
r0 and r3 are operated upon, and stored in r4

Note: This means that the tests are not exactly the way they are output, in the sense that:

add r4, r0, r1, lsl r2

is in fact

mov r3, r1, lsl r2
add r4, r0, r3

Note: I 'retransfer' the initial CPSR flags for `ADC`/`SBC`/`RSC`, because those need the old value of C to operate upon.

This means that inbetween the shift and the lsl, for these ops, there is another msr instruction to produce the correct result.

The multiplication instructions work as follows:

no register is shifted
r0, r1 are operated upon, and the result is stored in r4 (and r3 in case of a MULL/MLAL instruction)

The load/store instructions work as follows:

no register is shifted
r4 (and r3 for LDM/SWP instructions) are stored
r4 (and r3 for LDM/SWP instructions) are changed to the value of r12, which always holds 0800XXXX as far as I'm aware
r4 (and r3 for LDM/SWP instructions) are loaded from the same addresses

Note: no strange edge case behavior is tested for loading and storing: all loads/stores tested are word aligned!

The ROM uses a PSR transfer to initialize the CPSR to some arbitrary value to improve testing accuracy.

Screenshots:

Not sign extending a sign extended halfword load in 'THUMB' mode:

Passing all tests:

DenSinH/FuzzARM