Adds a quick perf report feature to `goalc` that lets you compare how
much faster / slower it takes to compile the projects, with some simple
features like filtering the files, adjusting for how large of a margin
of error in the speeds you care about, and which test iteration you want
to compare against.
This is something I plan to use as I work more in `goalc` as an easy way
to track / show the results.
Relates to #1353
This adds no new functionality or overhead to the compiler, yet. This is
the preliminary work that has:
- added code to the compiler in several spots to flag when something is
used without being properly required/imported/whatever (disabled by
default)
- that was used to generate project wide file dependencies (some
circulars were manually fixed)
- then that graph underwent a transitive reduction and the result was
written to all `jak1` source files.
The next step will be making this actually produce and use a dependency
graph. Some of the reasons why I'm working on this:
- eliminates more `game.gp` boilerplate. This includes the `.gd` files
to some extent (`*-ag` files and `tpage` files will still need to be
handled) this is the point of the new `bundles` form. This should make
it even easier to add a new file into the source tree.
- a build order that is actually informed from something real and
compiler warnings that tell you when you are using something that won't
be available at build time.
- narrows the search space for doing LSP actions -- like searching for
references. Since it would be way too much work to store in the compiler
every location where every symbol/function/etc is used, I have to do
ad-hoc searches. By having a dependency graph i can significantly reduce
that search space.
- opens the doors for common shared code with a legitimate pattern.
Right now jak 2 shares code from the jak 1 folder. This is basically a
hack -- but by having an explicit require syntax, it would be possible
to reference arbitrary file paths, such as a `common` folder.
Some stats:
- Jak 1 has about 2500 edges between files, including transitives
- With transitives reduced at the source code level, each file seems to
have a modest amount of explicit requirements.
Known issues:
- Tracking the location for where `defmacro`s and virtual state
definitions were defined (and therefore the file) is still problematic.
Because those forms are in a macro environment, the reader does not
track them. I'm wondering if a workaround could be to search the
reader's text_db by not just the `goos::Object` but by the text
position. But for the purposes of finishing this work, I just statically
analyzed and searched the code with throwaway python code.
This adds hfrag, but with a few remaining issues:
- The textures aren't animated. Instead, it just uses one texture.
- The texture filtering isn't as good as at it could be.
I also cleaned up a few issues with the background renderers:
- Cleaned up some stuff that is common to hfrag, tie, tfrag, shrub
- Moved time-of-day color packing stuff to FR3 creation, rather than at
level load. This appears to reduce the frame time spikes when a level is
first drawn by about 5 or 6 ms in big levels.
- Cleaned up the x86 specific stuff used in time of day. Now there's
only one place where we have an `ifdef`, rather than spreading it all
over the rendering code.
This fixes issues with certain Jak 3 levels not rendering because there
is a mismatch between the DGO name, nickname and real level name (bsp
name).
FR3s use a different filename, so you can delete the ones you have after
this is merged.
This affects custom levels, but I don't have that toolchain set up so
someone else will have to test that.
Patching up the extractor while working on the launcher, fixes:
- makes it so you can compile successfully given a folder path
(currently assumes your project path contains `iso_data`)
- ignore `buildinfo.json` from validation code.
- fixes an edge-case that could recursively fill up your entire
hard-drive!
- allows overriding the decompilation configuration via flag
- adds a way to specify where the ISO should be extracted to
I noticed that jak 3's compilation was spending a lot of time accessing
the `unordered_map`s we use to store constants and symbol types.
I repurposed the `EnvironmentMap` originally made for GOOS for this. It
turns out that we were copying the entire constant map whenever we
encountered a `deftype`, and fixed that too.
This speeds up jak3 compiles from ~16 to 11 seconds for me.
The `test-play` macro is back, though it doesn't call `play` yet. We can
at least load all of `game.cgo`, which involves loading a lot of the
code we've decompiled, loading/linking objects files compiled by
OpenGOAL (like dir-tpages), and loading/linking Jak's art-groups (for
jak 3 they are stored v5 format that I added to the linker).
There were no major issues - just a few forgotten mips2c entries and
minor bugs/functions that needed stubs. Most of the work was updating
the linker. Hopefully I'll never have to touch that code again - I think
it supports everything we need for jak 3!
- `fma-sphere`
- `prim-beam-h`
- `cam-start`
- `ragdoll`
- `light-trails-h`
- `light-trails`
- `menu`
- `water`
- `water-flow`
- `hud`
- `hud-classes`
- `progress`
- `progress-draw`
---
The `get-texture` macro replaces calls to `lookup-texture-by-id` and
`lookup-texture-by-id-fast`. The `defpart` macro detection was modified
to print a pair like `(texture-name tpage-name)` for the texture field
that gets turned into a `texture-id` constant. Only used in Jak 3 at the
moment, I'll probably go through the other games at a later point.
Reverse engineer the skinning matrix calculation and port to GOAL. This
is about 3x faster than the MIPS2c version.
As usual, there is a `*use-new-bones*` flag to go back to the old
version.
Fix for a bug in the compiler's `.div.vf` implementation (only happens
if src/dst are the same), and fix for a typo in the register allocator
that would sometimes cause it not to consider xmm8-xmm15.
This updates `fmt` to the latest version and moves to just being a copy
of their repo to make updating easier (no editing their cmake / figuring
out which files to minimally include).
The motivation for this is now that we switched to C++ 20, there were a
ton of deprecated function usages that is going away in future compiler
versions. This gets rid of all those warnings.
Adds two new lump types: `cell-info` and `buzzer-info`.
`cell-info` takes a `game-task` enum value and defines an `eco-info`
lump for power cells.
`buzzer-info` takes a `game-task` enum value and an integer and
automatically calculates the required value for the buzzer task with the
formula `task + buzzer * (1 << 16)`.
When giving entities custom actor IDs in your level JSON, it is possible
to break entity lookups by actor ID if the actors are not sorted by ID
because `entity-by-aid` expects them to be in order.
This sorts the actor list by ID before generating the level file and
also checks for any duplicates.
This sets up the C Kernel for Jak 3, and makes it possible to build and
load code built with `goalc --jak3`.
There's not too much interesting here, other than they switched to a
system where symbol IDs (unique numbers less than 2^14) are generated at
compile time, and those get included in the object file itself.
This is kind of annoying, since it means all tools that produce a GOAL
object file need to work together to assign unique symbol IDs. And since
the symbol IDs can't conflict, and are only a number between 0 and 2^14,
you can't just hash and hope for no collisions.
We work around this by ignoring the IDs and re-assigning our own. I
think this is very similar to what the C Kernel did on early builds of
Jak 3 which supported loading old format level files, which didn't have
the IDs included.
As far as I can tell, this shouldn't cause any problems. It defeats all
of their fancy tricks to save memory by not storing the symbol string,
but we don't care.
This sets out the bones of a Jak 3 build, many things are stubbed out,
guessed, or copied from Jak 2 but it should at least be good enough to:
run `task set-game-jak3`
launch the repl
run builds from the repl
build outputs themselves are untested but the build itself runs without
errors
---------
Co-authored-by: Tyler Wilding <xtvaser@gmail.com>
Updates Zydis to it's latest commit, this should fix building the
project on intel macs with a more recent version of macOS. This likely
needs some sanity checks that the debugger stuff still works as
expected.
This adds support for using enums in lumps using the new lump types
`enum-int32` and `enum-uint32`. Also adds these other new lump types:
- `water-height` (takes 3 meter floats, an enum and another optional
meter float)
- `eco-info` (takes an enum and an int)
- `vector3m` (3 meter floats + `w` set to 1.0)
The logger used in `goalc` tries to print an already-formatted string
`message` using `fmt::print(message);` Usually this doesn't cause
problems, but if you try to print, for example, an exception that has
special characters (notably `{`) it will try to do another round of
formatting/replacements, despite not having any args to replace with,
which ends up throwing another exception. This is why errors when
parsing custom level JSON cause the REPL to exit.
I've hopefully identified all the various instances of this across the
codebase
Major change to how `deftype` shows up in our code:
- the decompiler will no longer emit the `offset-assert`,
`method-count-assert`, `size-assert` and `flag-assert` parameters. There
are extremely few cases where having this in the decompiled code is
helpful, as the types there come from `all-types` which already has
those parameters. This also doesn't break type consistency because:
- the asserts aren't compared.
- the first step of the test uses `all-types`, which has the asserts,
which will throw an error if they're bad.
- the decompiler won't emit the `heap-base` parameter unless necessary
now.
- the decompiler will try its hardest to turn a fixed-offset field into
an `overlay-at` field. It falls back to the old offset if all else
fails.
- `overlay-at` now supports field "dereferencing" to specify the offset
that's within a field that's a structure, e.g.:
```lisp
(deftype foobar (structure)
((vec vector :inline)
(flags int32 :overlay-at (-> vec w))
)
)
```
in this structure, the offset of `flags` will be 12 because that is the
final offset of `vec`'s `w` field within this structure.
- **removed ID from all method declarations.** IDs are only ever
automatically assigned now. Fixes#3068.
- added an `:overlay` parameter to method declarations, in order to
declare a new method that goes on top of a previously-defined method.
Syntax is `:overlay <method-name>`. Please do not ever use this.
- added `state-methods` list parameter. This lets you quickly specify a
list of states to be put in the method table. Same syntax as the
`states` list parameter. The decompiler will try to put as many states
in this as it can without messing with the method ID order.
Also changes `defmethod` to make the first type definition (before the
arguments) optional. The type can now be inferred from the first
argument. Fixes#3093.
---------
Co-authored-by: Hat Kid <6624576+Hat-Kid@users.noreply.github.com>
Started at 349,880,038 allocations and 42s
- Switched to making `Symbol` in GOOS be a "fixed type", just a wrapper
around a `const char*` pointing to the string in the symbol table. This
is a step toward making a lot of things better, but by itself not a huge
improvement. Some things may be worse due to more temp `std::string`
allocations, but one day all these can be removed. On linux it saved
allocations (347,685,429), and saved a second or two (41 s).
- cache `#t` and `#f` in interpreter, better lookup for special
forms/builtins (hashtable of pointers instead of strings, vector for the
small special form list). Dropped time to 38s.
- special-case in quasiquote when splicing is the last thing in a list.
Allocation dropped to 340,603,082
- custom hash table for environment lookups (lexical vars). Dropped to
36s and 314,637,194
- less allocation in `read_list` 311,613,616. Time about the same.
- `let` and `let*` in Interpreter.cpp 191,988,083, time down to 28s.
This change adds a few new features:
- Decompiler automatically knows the type of `find-parent-method` use in
jak 1 and jak2 when used in a method or virtual state handler.
- Decompiler inserts a call to `call-parent-method` or
`find-parent-state`
- Removed most casts related to these functions
There are still a few minor issues around this:
- There are still some casts needed when using `post` methods, as `post`
is just a `function`, and needs a cast to `(function none)` or similar.
It didn't seem easy to change the type of `post`, so I'm not going to
worry about it for this PR. It only shows up in like 3 places in jak 2.
(and 0 in jak 1)
- If "call the handler if it's not #f" logic should probably be another
macro.
Fixes#805
Previously, `object` and `none` were both top-level types. This made
decompilation rather messy as they have no LCA and resulted in a lot of
variables coming out as type `none` which is very very wrong and
additionally there were plenty of casts to `object`. This changes it so
`none` becomes a child of `object` (it is still represented by
`NullType` which remains unusable in compilation).
This change makes `object` the sole top-level type, and the type that
can represent *any* GOAL object. I believe this matches the original
GOAL built-in type structure. A function that has a return type of
`object` can now return an integer or a `none` at the same time.
However, keep in mind that the return value of `(none)` is still
undefined, just as before. This also makes a cast to `object`
meaningless in 90% of the situations it showed up in (as every single
thing is already an `object`) and the decompiler will no longer emit
them. Casts to `none` are also reduced. Yay!
Additionally, state handlers also don't get the final `(none)` printed
out anymore. The return type of a state handler is completely
meaningless outside the event handler (which is return type `object`
anyway) so there are no limitations on what the last form needs to be. I
did this instead of making them return `object` to trick the decompiler
into not trying to output a variable to be used as a return value
(internally, in the decompiler they still have return type `none`, but
they have `object` elsewhere).
Fixes#1703Fixes#830Fixes#928
This will create a folder like `decompiler_out/jak1/entities` and save a
JSON file per level with all the actors.
Also, it should hopefully make custom level building a little faster.
- fix speaker names and time frames being uneditable in Jak 1
- added toggle to auto-selected a newly created scene as current
- changed the subtitle summary format slightly.
- current scene's name now appears in the UI
Some general improvements for the texture animator:
- Clouds are special cased, saving about 1 ms per frame
- Adjusting the amount of clouds now actually works.
- Fixed an issue with the brightness of clouds, and the way that they
fade out around the edges.
