vdbgln() was responsible for ~10% of samples on pv's flamegraph for
RequestServer (under request_did_finish) when loading github.com in
Browser and recording a whole-system profile. This makes that almost
completely disappear.
BigEndianInputBitStream is the Core::Stream API's bitwise input stream
for big endian input data. The functionality and bitwise read API is
almost unchanged from AK::BitStream, except that this bit stream only
supports big endian operations.
As the behavior for mixing big endian and little endian reads on
AK::BitStream is unknown (and untested), it was never done anyways. So
this was a good opportunity to split up big endian and little endian
reading.
Another API improvement from AK::BitStream is the ability to specify
the return type of the bit read function. Always needing to static_cast
the result of BitStream::read_bits_big_endian into the desired type is
adding a lot of avoidable noise to the users (primarily FlacLoader).
Regardless of the backing type that the number would otherwise parse to,
if it is zero and the sign was supposed to be negative then it needs to
be a floating point number to represent the correct value.
This implements an 8-bit front stencil buffer. Stencil operations are
SIMD optimized. LibGL changes include:
* New `glStencilMask` and `glStencilMaskSeparate` functions
* New context parameter `GL_STENCIL_CLEAR_VALUE`
In this particular case, auto is a footgun, because we are very
certain about the type that we want to return. const-correctness
could have been violated (as Vector did), because Iterator did not
enforce that the returned pointer was actually const if the Iterator
was an Iterator over a const container.
Methods marked as const should always only return Foo const&, never
Foo&. Previously, this only worked correctly with Foo, but not with
Foo&: If someone fetched a T const&, this would have been expanded
to Foo& const& which is just Foo&. Therefore, they were able to modify
the elements of the Vector, even though it was marked as const.
This commit fixes that by introducing VisibleType as an alias for
RemoveReference<T> and using it throughout Vector.
There are also other cases where we don't require a mutable reference
to the underlying type, only a const reference (for example in a find
operation). In these cases, we now also correctly expand the type
to Foo const&.
This is particularly useful in the Kernel, where the physical pages of
a VMObject are stored as a FixedArray but often passed around as a Span
from which a new FixedArray should be cloned.
This makes the following code behave as expected:
Variant<int, String> x { some_string() };
x.visit(
[](String const&) {}, // Expectation is for this to be called
[](auto&) {});
This is useful for writing new data at the end of a ByteBuffer. For
instance, with the Stream API:
auto pending_bytes = TRY(stream.pending_bytes());
auto receive_buffer = TRY(buffer.get_bytes_for_writing(
pending_bytes));
TRY(stream.read(receive_buffer));
Except for tangential accessors such as data(), there is no more feature
of FixedArray that is untested after this large expansion of its test
cases. These tests, with the help of the new NoAllocationGuard, also
test the allocation contract that was fixated in the last commit.
Hopefully this builds confidence in future Kernel uses of FixedArray
as well as its establishment in the real-time parts of the audio
subsystem. I'm excited :^)
FixedArray always *almost* had the following allocation guarantees:
There is (possibly) one allocation in the constructor and one (or more)
deallocation(s) in the destructor. No other operation allocates or
deallocates. With this removal of the public clear() method, which
nobody except the test used anyways, those guarantees are now completely
true and furthermore fixated with an explanatory comment.
Because we don't have our LibC on Lagom, the allocation guard global
flag doesn't exist and NoAllocationGuard fails to build on Lagom.
Whoops. For now, just disable NoAllocationGuard's functionality here.
StringView::for_each_split_view allows you to process the splits in a
StringView without needing to allocate a Vector<StringView> to store
each of the parts.
Since we migrated the implementation from the normal split_view path, we
can also re-implement split_view in terms of for_each_split_view.
This mechanism was unsafe to use in any multithreaded context, since
the hook function was invoked on a raw pointer *after* decrementing
the local ref count.
Since we don't use it for anything anymore, let's just get rid of it.
Currently, we define a CaseInsensitiveStringTraits structure for String.
Using this structure for StringView involves allocating a String from
that view, and a second string to convert that intermediate string to
lowercase.
This defines CaseInsensitiveStringViewTraits (and the underlying helper
case_insensitive_string_hash) to avoid allocations.
NoAllocationGuard is an RAII stack guard that prevents allocations
while it exists. This is done through a thread-local global flag which
causes malloc to crash on a VERIFY if it is false. The guard allows for
recursion.
The intended use case for this class is in real-time audio code. In such
code, allocations are really bad, and this is an easy way of dynamically
enforcing the no-allocations rule while giving the user good feedback if
it is violated. Before real-time audio code is executed, e.g. in LibDSP,
a NoAllocationGuard is instantiated. This is not done with this commit,
as currently some code in LibDSP may still incorrectly allocate in real-
time situations.
Other use cases for the Kernel have also been added, so this commit
builds on the previous to add the support both in Userland and in the
Kernel.
FixedArray now doesn't expose any infallible constructors anymore.
Rather, it exposes fallible methods. Therefore, it can be used for
OOM-safe code.
This commit also converts the rest of the system to use the new API.
However, as an example, VMObject can't take advantage of this yet,
as we would have to endow VMObject with a fallible static
construction method, which would require a very fundamental change
to VMObject's whole inheritance hierarchy.
The previous implementation had some pretty short cycles and two fixed
points (1711463637 and 2389024350). If two keys hashed to one of these
values insertions and lookups would loop forever.
This version is based on a standard xorshift PRNG with period 2**32-1.
The all-zero state is usually forbidden, so we insert it into the cycle
at an arbitrary location.
This is just to allow removing the 'clang-format off' directive. This
concept is only used within this header, so it doesn't need to be in the
global namespace.
We have whittled away at the usages of these AK::Vector APIs in the
Kernel. This change disables them from being visible when building
the Kernel to make sure no new usages get added.
If we didn't define our own `move` and `forward` and inject it into the
`std` namespace, then we would error just after trying to `using` those,
if no one has included it before. Now, we will include `utility` from
the STD if we aren't replacing the `std`.
This was a premature optimization from the early days of SerenityOS.
The eternal heap was a simple bump pointer allocator over a static
byte array. My original idea was to avoid heap fragmentation and improve
data locality, but both ideas were rooted in cargo culting, not data.
We would reserve 4 MiB at boot and only ended up using ~256 KiB, wasting
the rest.
This patch replaces all kmalloc_eternal() usage by regular kmalloc().
This is an interface to downcast(), which degrades errors into runtime
errors, and allows seemingly-correct-but-not-quite constructs like the
following to appear to compile, but fail at runtime:
Variant<NonnullRefPtr<T>, U> foo = ...;
Variant<RefPtr<T>, U> bar = foo;
The expectation here is that `foo` is converted to a RefPtr<T> if it
contains one, and remains a U otherwise, but in reality, the
NonnullRefPtr<T> variant is simply dropped on the floor, and the
resulting variant becomes invalid, failing the assertion in downcast().
This commit adds a Variant<Ts...>(Variant<NewTs...>) constructor that
ensures that no alternative can be left out at compiletime, for the
users that were using this interface for merely increasing the number of
alternatives (for instance, LibSQL's Value class).
This is a raffinement of 49cbd4dcca.
Previously, the container was scanned to compute the size in the unhappy
path. Now, using `all_of` happy and unhappy path should be fast.
In order to reduce our reliance on __builtin_{ffs, clz, ctz, popcount},
this commit removes all calls to these functions and replaces them with
the equivalent functions in AK/BuiltinWrappers.h.
While watching Andreas' most recent video, I noticed that this function
only worked with 32 bit values, but was a serious performance
bottleneck for the kernel. As such, I reworked it to use `size_t`, so
it now can switch to 64-bit sweeps on 64-bit platforms. This caused
test-js to go from 12.5 seconds hot to 11.5 seconds hot on my machine
when running on KVM x86_64.
The goal of this file is to enable C++ overloaded functions for
standard builtin functions that we use. It contains fallback
implementations for systems that do not have the builtins available.
Before, if we couldn't read enough data out of the buffer, we would re-
fill the buffer and recursively call read(), which in turn reads data
from the buffer into the resliced target span. This incurs very
intensive superflous memmove's when large chunks of data are read from
a buffered stream.
This commit changes the behavior so that when we exhaust the buffer, we
first read any necessary additional data directly into the target, then
fill up the buffer again. Effectively, this results in drastically
reduced overhead from Buffered when reading large contiguous chunks.
Of course, Buffered is designed to speed up data access patterns with
small frequent reads, but it's nice to be able to combine both access
patterns on one stream without penalties either way.
The final performance gain is about an additional 80% of abench decoding
speed.
This unbreaks the /var/run/utmp system which starts out as an empty
string, and is then turned into an object by the first update.
This isn't necessarily the best way for this to work, but it's how
it used to work, so this just fixes the regression for now.
This fixes at least half of our LibC includes in the kernel. The source
of truth for errno codes and their description strings now lives in
Kernel/API/POSIX/errno.h as an enumeration, which LibC includes.
This is the same strategy that LLVM's compiler-rt uses to make sure that
each UBSAN error is only reported once, when UBSAN is *not* deadly.
Otherwise, each time we head through a UB codepath, we will log the same
error over and over. That behavior just adds noise to the logs and makes
it nearly impossible to run binaires that have some common code path
with flagged UB in them.
compiler-rt goes the extra step to make sure the "clear" action is
atomic, but we don't really have that many multi-threaded apps gettting
tested with UBSAN yet, so we can add that later.
This will allow us to avoid some potentially expensive type conversion
during lookup, like form String to StringView, which would allocate
memory otherwise.
