resource: Implement decodeSZSNxAsm64_
#196
Conversation
|
Do we allow assembly in the project? I am referring to this to be exact https://botw.link/about#how-can-i-help
|
|
That's for BotW, not sead. We don't have a clear policy on this for sead Personally I think that it's a bit iffy but we're already doing the same thing for Wii U... |
|
Also I'm not sure if this should be inline assembly or a separate .s file. If we can't get regalloc to work the same I think that suggests it's not inline asm |
|
For Wii U, it's definitely inline assembly for several reasons, like the double |
|
I'm a bit suspicious of that "mov #1000", followed by "sub #1" - even IDA displayed that as "mov #FFF" straight-away, so I don't see how the compiler would miss this optimization opportunity if not explicitly listed in the assembly, but I'll do some testing later. |
|
Next to the argument already listed above, here's another one: |
While I agree that this definitely smells like handwritten assembly, that argument is not quite true! Compilers can make mistakes too and/or UB can cause compilers to generate dumb code. |
|
It's assembly |
|
I mean, yeah, it's literally in the function name. The debate is whether it's a separate text file or inline assembly (I'm certain of the latter). |
|
I tried writing it as nice as possible, using templates for the usages of |
|
Any more comments on this, or should this be merged? |
Based on the name, this function just contains a bunch of assembly - so I didn't even bother trying to re-implement it in C++, but just copied the output of the disassembler here straight-away.
The labels are defined using just
[number]:to avoid creating actual labels in the executable for them. This also means that references to them must be suffixed with eitherbfor backwards-search orffor forwards-search, which is why some of those jump targets look weird.Additionally, I tried properly defining inputs/outputs of the assembly block. The auto-allocation for
srcanddstworks fine here, buterrordoesn't want to pick the actual targetw2- either it picksw0and generates conflicts withdstinx0(=> mismatch), or when marking it as clobbering variable (=&r), it decides to pickw8instead, which is also not our target - so I manually assigned it tow2to match the code here.At label
6, two more instructions are auto-generated (mov w0, w2 ; ret), thanks to the setup of the assembly block.This change is