Instead of keeping bytecode as a set of disjoint basic blocks on the
malloc heap, bytecode is now a contiguous sequence of bytes(!)
The transformation happens at the end of Bytecode::Generator::generate()
and the only really hairy part is rerouting jump labels.
This required solving a few problems:
- The interpreter execution loop had to change quite a bit, since we
were storing BasicBlock pointers all over the place, and control
transfer was done by redirecting the interpreter's current block.
- Exception handlers & finalizers are now stored per-bytecode-range
in a side table in Executable.
- The interpreter now has a plain program counter instead of a stream
iterator. This actually makes error stack generation a bit nicer
since we just have to deal with a number instead of reaching into
the iterator.
This yields a 25% performance improvement on this microbenchmark:
for (let i = 0; i < 1_000_000; ++i) { }
But basically everything gets faster. :^)
Before this change, all JumpFoo instructions inherited from Jump, which
forced the unconditional Jump to have an unusued "false target" member.
Also, labels were unnecessarily wrapped in Optional<>.
By defining each jump instruction separately, they all shrink in size,
and all ambiguity is removed.
This does two things:
* Clear exceptions when transferring control out of a finalizer
Otherwise they would resurface at the end of the next finalizer
(see test the new test case), or at the end of a function
* Pop one scheduled jump when transferring control out of a finalizer
This removes one old FIXME
Before this change both ExecutionContext and CallFrame were created
before executing function/module/script with a couple exceptions:
- executable created for default function argument evaluation has to
run in function's execution context.
- `execute_ast_node()` where executable compiled for ASTNode has to be
executed in running execution context.
This change moves all members previously owned by CallFrame into
ExecutionContext, and makes two exceptions where an executable that does
not have a corresponding execution context saves and restores registers
before running.
Now, all execution state lives in a single entity, which makes it a bit
easier to reason about and opens opportunities for optimizations, such
as moving registers and local variables into a single array.
When a PutById / PutByValue bytecode operation results in accessing a
nullish object, we now include the name of the property and the object
being accessed in the exception message (if available). This should make
it easier to debug live websites.
For example, the following errors would all previously produce a generic
error message of "ToObject on null or undefined":
> foo = null
> foo.bar = 1
Uncaught exception:
[TypeError] Cannot access property "bar" on null object "foo"
at <unknown>
> foo = { bar: undefined }
> foo.bar.baz = 1
Uncaught exception:
[TypeError] Cannot access property "baz" on undefined object "foo.bar"
at <unknown>
Note we certainly don't capture all possible nullish property write
accesses here. This just covers cases I've seen most on live websites;
we can cover more cases as they arise.
When a GetById / GetByValue bytecode operation results in accessing a
nullish object, we now include the name of the property and the object
being accessed in the exception message (if available). This should make
it easier to debug live websites.
For example, the following errors would all previously produce a generic
error message of "ToObject on null or undefined":
> foo = null
> foo.bar
Uncaught exception:
[TypeError] Cannot access property "bar" on null object "foo"
at <unknown>
> foo = { bar: undefined }
> foo.bar.baz
Uncaught exception:
[TypeError] Cannot access property "baz" on undefined object "foo.bar"
at <unknown>
Note we certainly don't capture all possible nullish property read
accesses here. This just covers cases I've seen most on live websites;
we can cover more cases as they arise.
As long as the inputs are Int32, we can convert them to UInt32 in a
spec-compliant way with a simple static_cast<u32>.
This allows calculations like `-3 >>> 2` to take the fast path as well,
which is extremely valuable for stuff like crypto code.
While we're doing this, also remove the fast paths from the generic
shift functions in Value.cpp, since we only end up there if we *didn't*
take the same fast path in the interpreter.
This patch adds a new "Peephole" pass for performing small, local
optimizations to bytecode.
We also introduce the first such optimization, fusing a sequence of
some comparison instruction FooCompare followed by a JumpIf into a
new set of JumpFooCompare instructions.
This gives a ~50% speed-up on the following microbenchmark:
for (let i = 0; i < 10_000_000; ++i) {
}
But more traditional benchmarks see a pretty sizable speed-up as well,
for example 15% on Kraken/ai-astar.js and 16% on Kraken/audio-dft.js :^)
Instead of emitting a NewBigInt instruction to construct a primitive
bigint from a parsed literal, we now instantiate the BigInt on the heap
during codegen.
Instead of emitting a NewString instruction to construct a primitive
string from a parsed literal, we now instantiate the PrimitiveString on
the heap during codegen.
Instead of looking these up in the VM execution context stack whenever
we need them, we now just cache them in the interpreter when entering
a new call frame.
Comparing two Values has to call the generic same_value() helper,
and we can avoid this by simply using a stronger type for built-in
native function handlers.
Instead of having Call refer to a range of VM registers, it now has
a trailing list of argument operands as part of the instruction.
This means we no longer have to shuffle every argument value into
a register before making a call, making bytecode smaller & faster. :^)
By handling common cases like Int32 arithmetic directly in the
instruction handler, we can avoid the cost of calling the generic helper
functions in Value.cpp.
Instead of splitting the postfix variants into ToNumeric + Inc/Dec,
we now have dedicated PostfixIncrement and PostfixDecrement instructions
that handle both outputs in one go.
This patch moves us away from the accumulator-based bytecode format to
one with explicit source and destination registers.
The new format has multiple benefits:
- ~25% faster on the Kraken and Octane benchmarks :^)
- Fewer instructions to accomplish the same thing
- Much easier for humans to read(!)
Because this change requires a fundamental shift in how bytecode is
generated, it is quite comprehensive.
Main implementation mechanism: generate_bytecode() virtual function now
takes an optional "preferred dst" operand, which allows callers to
communicate when they have an operand that would be optimal for the
result to go into. It also returns an optional "actual dst" operand,
which is where the completion value (if any) of the AST node is stored
after the node has "executed".
One thing of note that's new: because instructions can now take locals
as operands, this means we got rid of the GetLocal instruction.
A side-effect of that is we have to think about the temporal deadzone
(TDZ) a bit differently for locals (GetLocal would previously check
for empty values and interpret that as a TDZ access and throw).
We now insert special ThrowIfTDZ instructions in places where a local
access may be in the TDZ, to maintain the correct behavior.
There are a number of progressions and regressions from this test:
A number of async generator tests have been accidentally fixed while
converting the implementation to the new bytecode format. It didn't
seem useful to preserve bugs in the original code when converting it.
Some "does eval() return the correct completion value" tests have
regressed, in particular ones related to propagating the appropriate
completion after control flow statements like continue and break.
These are all fairly obscure issues, and I believe we can continue
working on them separately.
The net test262 result is a progression though. :^)
The JIT compiler was an interesting experiment, but ultimately the
security & complexity cost of doing arbitrary code generation at runtime
is far too high.
In subsequent commits, the bytecode format will change drastically, and
instead of rewriting the JIT to fit the new bytecode, this patch simply
removes the JIT instead.
Other engines, JavaScriptCore in particular, have already proven that
it's possible to handle the vast majority of contemporary web content
with an interpreter. They are currently ~5x faster than us on benchmarks
when running without a JIT. We need to catch up to them before
considering performance techniques with a heavy security cost.
perform_call() wants a ReadonlySpan<Value>, so just grab a slice of the
current register window instead of making a MarkedVector.
10% speed-up on this function call microbenchmark:
function callee(a, b, c) { }
function caller(callee) {
for (let i = 0; i < 10_000_000; ++i)
callee(1, 2, 3)
}
caller(callee)
This commit un-deprecates DeprecatedString, and repurposes it as a byte
string.
As the null state has already been removed, there are no other
particularly hairy blockers in repurposing this type as a byte string
(what it _really_ is).
This commit is auto-generated:
$ xs=$(ack -l \bDeprecatedString\b\|deprecated_string AK Userland \
Meta Ports Ladybird Tests Kernel)
$ perl -pie 's/\bDeprecatedString\b/ByteString/g;
s/deprecated_string/byte_string/g' $xs
$ clang-format --style=file -i \
$(git diff --name-only | grep \.cpp\|\.h)
$ gn format $(git ls-files '*.gn' '*.gni')
When iterating over an iterable, we get back a JS object with the fields
"value" and "done".
Before this change, we've had two dedicated instructions for retrieving
the two fields: IteratorResultValue and IteratorResultDone. These had no
fast path whatsoever and just did a generic [[Get]] access to fetch the
corresponding property values.
By replacing the instructions with GetById("value") and GetById("done"),
they instantly get caching and JIT fast paths for free, making iterating
over iterables much faster. :^)
26% speed-up on this microbenchmark:
function go(a) {
for (const p of a) {
}
}
const a = [];
a.length = 1_000_000;
go(a);
This patch makes IteratorRecord an Object. Although it's not exposed to
author code, this does allow us to store it in a VM register.
Now that we can store it in a VM register, we don't need to convert it
back and forth between IteratorRecord and Object when accessing it from
bytecode.
The big win here is avoiding 3 [[Get]] accesses on every iteration step
of for..of loops. There are also a bunch of smaller efficiencies gained.
20% speed-up on this microbenchmark:
function go(a) {
for (const p of a) {
}
}
const a = [];
a.length = 1_000_000;
go(a);
Allows the bytecode interpreter to call the builtins c++
implementation directly without making a javascript call
just as the JIT.
Kraken test speedups: imaging-gaussian-blur.js (1.5x) and
audio-oscillator.js (1.2x)
The number of registers in a call frame never changes, so we can
allocate it at the end of the CallFrame object and save ourselves the
cost of allocating separate Vector storage for every call frame.
Instead of allocating these in a mixture of ways, we now always put
them on the malloc heap, and keep an intrusive linked list of them
that we can iterate for GC marking purposes.
This will not meaningfully affect short array literals, but it does
give us a bit of extra perf when evaluating huge array expressions like
in Kraken/imaging-darkroom.js
Until now, the unwind context stack has not been maintained by jitted
code, which meant we were unable to support the `with` statement.
This is a first step towards supporting that by making jitted code
call out to C++ to update the unwind context stack when entering/leaving
unwind contexts.
We also introduce a new "Catch" bytecode instruction that moves the
current exception into the accumulator. It's always emitted at the start
of a "catch" block.
