* All of the MxVectors share an inheritance chain. MxVector4 inherits
from MxVector3 which inherits from MxVector2.
* They all operate on a shared `float*` data member which points to the
underlying storage.
* There are also MxVector3/4Data classes, which inherit from Vector3/4,
but add concrete storage for the Vector data rather than just an
abstract data pointer.
* The same is true for MxMatrix, with there being an abstract and a
concrete variant of it.
* Also improve reccmp.py register matching algorithm. It previously
could not recognize an effective match when a swap had to take place
between two registers used on the same line. It turns out this happens
a lot in floating point math code so I adjusted the implementation to
break the disassembly lines on spaces rather than just linebreaks
allowing the existing effective match code to handle that case too.
* Rather than using <OFFSET> as a replacement for all offsets in a
function, label the offsets as <OFFSET1>, <OFFSET2>, etc. Doing this
will avoid false-positive 100% matches resulting from the same
function being called in two times where a different on should have
been called or vice versa. And the same for globals. I already
encountered one case of this in the wild.
* When a 100% match initially fails, try to make the functions match by
swapping register allocations. This makes it possible to get a 100%
match where the generated machine code differs only in register
allocation.
* Only apply the above when it is possible to reach a 100% match in that
way. Otherwise show the developer the unadultrated diff to avoid
complicating decompilation.
* In the result listing, show the functions which are "effective
matches" in this way as "100%*" instead of "100%".
* reccmp: avoid repeated execution of winepath
Executing winepath many times is slow,
so try we like to avoid it as much as possible.
When the path start with a known prefix, replace it with
a cached prefix and do some string manipulation.
This change reduces execution time of reccmp.py from 90s to 2s.
Which is nice.
m
* reccmp: continue looking when source cannot be found
Most often, the reasons is mismatched sources.
* reccmp: add basic logging + optional debug
* Read the addresses in the exe headers as little endian
Now we can use our own compiled LEGO1.LIB rather than one generated from the original. Also implements a script that tests them to help ensure future commits don't break them.
* Stubbed a bunch of classes and annotated them for later use. Heavily wip and more of pseudocode right now.
* Converted pseudocode into real code!
* Created a bunch more classes and added more information to exisiting ones
Did not error check, this was pushed just for reference
* More classes and implementation details. Still not checked for any errors
* Fixed code and decided on a way to handle virtual table stubs
* Some additional fixes
* More smaller fixes
* Added classes to project and made it compile
* Fixed function adresses that caused the python script to fail
* More classes and virtual function resolves. Builds and compares fine.
* Again more classes and virtual function resolves. Builds and compares fine.
* No clue, I guess forced update for line endings
* Finished up some work, compiles fine. All functions are STUB annotated to not pollute reccmp.py output.
* line ending change
* rename GetClassName/IsClass
Mirroring recent changes from master
* further conform to current master
* update project
* cleanup
* project only updates when you close msdev
---------
Co-authored-by: Cydra <cydra95@gmail.com>
Co-authored-by: itsmattkc <34096995+itsmattkc@users.noreply.github.com>
* Improve reccmp.py
* Now only shows the info for a single function when a specific function
is specified via -v
* Now colors the output by default
* Percentages are shown as green/yellow/red depending on the percentage
completed.
* Diff +/- lines are shown as green/red.
* Includes standard --no-color argument in case we need no color for
some tooling which consumes the output.
* Feedback
Sorry to everyone's muscle memory, but I think this is better. The idea for the name was "recomp compare", but it's too easy to read it as "recomp with a typo". This should fix that, as well as be slightly easier to write since it's shorter.
* recomp.py: use argparse to parse arguments
* Address code revew comments
* reccomp.py: -h/--help for help -H/--htmp for html
* update CI to use new arg
* slight string updates
---------
Co-authored-by: itsmattkc <34096995+itsmattkc@users.noreply.github.com>
* add test to compare assembly between functions
* ci: use abs path of wget
* ci: fix shell disambiguity
* ci: ensure capstone is installed
* ci: ensure correct filenames
* use better source for lego island files
* give me an idea of what the dir structure looks like
* make wine path function
* improved script and project
* fixed script on windows
* print debug info because now it literally only doesn't work on fucking github actions
* better source path resolving
For some reason, nmake compiles produce different symbols. I wonder if this affects the accuracy of the decomp.
