Resulting in a massive rename across almost everywhere! Alongside the
namespace change, we now have the following names:
* JS::NonnullGCPtr -> GC::Ref
* JS::GCPtr -> GC::Ptr
* JS::HeapFunction -> GC::Function
* JS::CellImpl -> GC::Cell
* JS::Handle -> GC::Root
Instead, smuggle it in as a `void*` private data and let Javascript
aware code cast out that pointer to a VM&.
In order to make this split, rename JS::Cell to JS::CellImpl. Once we
have a LibGC, this will become GC::Cell. CellImpl then has no specific
knowledge of the VM& and Realm&. That knowledge is instead put into
JS::Cell, which inherits from CellImpl. JS::Cell is responsible for
JavaScript's realm initialization, as well as converting of the void*
private data to what it knows should be the VM&.
NanBoxedValue is intended to be a GC-allocatable type which is not
specific to javascript, towards the effort of factoring out the GC
implementation from LibJS.
Now that the heap has no knowledge about a JavaScript realm and is
purely for managing the memory of the heap, it does not make sense
to name this function to say that it is a non-realm variant.
The main motivation behind this is to remove JS specifics of the Realm
from the implementation of the Heap.
As a side effect of this change, this is a bit nicer to read than the
previous approach, and in my opinion, also makes it a little more clear
that this method is specific to a JavaScript Realm.
While this does mean that we keep one copy of the stack info in the VM,
and another in the Heap; keeping a separate instance removes one more
instance of coupling between the heap and LibJS specific details.
There is definitely a possibility I am misunderstanding the reason
behind it - but this does not appear neccessary. The VM owns both the
string cache and Heap. On destruction, the VM should clear out both
the heap and its string cache.
While this is used in the implementation of Runtime objects itself, Heap
seems like a more appropriate home. This will also help in factoring out
the GC implementation into it's own library as the heap explicitly has
knowledge of WeakContainer.
We were miscalculating the length of the buffer in pointer-sized chunks,
which is what the conservative root scan cares about.
This could cause some values to be prematurely garbage-collected.
This is really just a type alias for NonnullGCPtr<T>, but it provides
a way to have non-owning non-visited NonnullGCPtr<T> without getting
yelled at by the Clang plugin for catching GC errors.
We were previously dumping the address of the cell pointer instead of
the address of the cell itself. This was causing mysterious orphans
in GC dumps, and it took me way too long to figure this out.
...and don't let them leak out of their evaluation contexts.
Also keep the exceptions separate from the actual values.
This greatly reduces the number of assertions hit while entering random
data into a sheet.
This changes the remaining uses of the following functions across LibJS:
- String::format() => String::formatted()
- dbg() => dbgln()
- printf() => out(), outln()
- fprintf() => warnln()
I also removed the relevant 'LogStream& operator<<' overloads as they're
not needed anymore.
Previously we would iterate over all the live HeapBlocks in order to
learn if an arbitrary pointer-sized value was a pointer into a live
HeapBlock. This was quite time-consuming.
Instead of that, just put all the live HeapBlock*'s in a HashTable
and identify pointers by doing a bit-masked lookup into the table.
Instead of keeping all the HeapBlocks in one big list, we now split it
into two levels:
- Heap has a set of Allocators, each with a specific cell size.
- Allocators have two lists of blocks, "full" and "usable".
Allocating a new cell no longer has to scan the entire set of blocks,
but instead just needs to find the right allocator and then pop a cell
from its freelist. If all the blocks in the allocator are full, a new
block will be created.
Blocks are moved from the "full" to "usable" list after sweeping has
determined that they are not completely empty and not completely full.
There are certainly many ways we can improve on this. This patch is
mostly about getting the new allocator architecture in place. :^)
While initialization common runtime objects like functions, prototypes,
etc, we don't really care about tracking transitions for each and every
property added to them.
This patch puts objects into a "disable transitions" mode while we call
initialize() on them. After that, adding more properties will cause new
transitions to be generated and added to the chain.
This gives a ~10% speed-up on test-js. :^)
When scanning for potential heap pointers during conservative GC,
we look for any value that is an address somewhere inside a heap cell.
However, we were failing to account for the slack at the end of a
block (which occurs whenever the block storage size isn't an exact
multiple of the cell size.) Pointers inside the trailing slack were
misidentified as pointers into "last_cell+1".
Instead of skipping over them, we would treat this garbage data as a
live cell and try to mark it. I believe this is the test-js crash that
has been terrorizing Travis for a while. :^)
Taking a big step towards a world of multiple global object, this patch
adds a new JS::VM object that houses the JS::Heap.
This means that the Heap moves out of Interpreter, and the same Heap
can now be used by multiple Interpreters, and can also outlive them.
The VM keeps a stack of Interpreter pointers. We push/pop on this
stack when entering/exiting execution with a given Interpreter.
This allows us to make this change without disturbing too much of
the existing code.
There is still a 1-to-1 relationship between Interpreter and the
global object. This will change in the future.
Ultimately, the goal here is to make Interpreter a transient object
that only needs to exist while you execute some code. Getting there
will take a lot more work though. :^)
Note that in LibWeb, the global JS::VM is called main_thread_vm(),
to distinguish it from future worker VM's.
The fact that a `MarkedValueList` had to be created was just annoying,
so here's an alternative.
This patchset also removes some (now) unneeded MarkedValueList.h includes.
You can now pass print_report=true to Heap::collect_garbage() and it
will print out a little summary of the time spent, and counts of
live vs freed cells and blocks.
The SI prefixes "k", "M", "G" mean "10^3", "10^6", "10^9".
The IEC prefixes "Ki", "Mi", "Gi" mean "2^10", "2^20", "2^30".
Let's use the correct name, at least in code.
Only changes the name of the constants, no other behavior change.
To make sure that everything is set up correctly in objects before we
start adding properties to them, we split cell allocation into 3 steps:
1. Allocate a cell of appropriate size from the Heap
2. Call the C++ constructor on the cell
3. Call initialize() on the constructed object
The job of initialize() is to define all the initial properties.
Doing it in a second pass guarantees that the Object has a valid Shape
and can find its own GlobalObject.
In C++, it's invalid to cast a block of memory to a complex type without
invoking its constructor. It's even more invalid to simply cast a pointer to a
block of memory to a pointer to *an abstract type*.
To fix this, make sure FreelistEntry is a concrete type, and call its
constructor whenever appropriate.
Lagom now builds under macOS. Only two minor adjustments were required:
* LibCore TCP/UDP code can't use `SOCK_{NONBLOCK,CLOEXEC}` on macOS,
use ioctl() and fcntl() instead
* LibJS `Heap` code pthread usage ported to MacOS
A MarkedValueList is basically a Vector<JS::Value> that registers with
the Heap and makes sure that the stored values don't get GC'd.
Before this change, we were unsafely keeping Vector<JS::Value> in some
places, which is out-of-reach for the live reference finding logic
since Vector puts its elements on the heap by default.
We now pass all the JavaScript tests even when running with "js -g",
which does a GC on every heap allocation.
