Merging registers, constants and locals into single vector means:
- Better data locality
- No need to check type in Interpreter::get() and Interpreter::set()
which are very hot functions
Performance improvement is visible in almost all Octane and Kraken
tests.
When comparing two numbers, we can avoid a lot of implicit type
conversion nonsense and go straight to comparison, saving time in the
most common case.
These were out-of-line because we had some ideas about marking
instruction streams PROT_READ only, but that seems pretty arbitrary and
there's a lot of performance to be gained by putting these inline.
Instead, generate bytecode to execute their AST nodes and save the
resulting operands inside the NewClass instruction.
Moving property expression evaluation to happen before NewClass
execution also moves along creation of new private environment and
its population with private members (private members should be visible
during property evaluation).
Before:
- NewClass
After:
- CreatePrivateEnvironment
- AddPrivateName
- ...
- AddPrivateName
- NewClass
- LeavePrivateEnvironment
This patch stops emitting the BlockDeclarationInstantiation instruction
when there are no locals, and no function declarations in the scope.
We were spending 20% of CPU time on https://ventrella.com/Clusters/ just
creating empty environments for no reason.
If the minimal amount of required bindings is known in advance, it could
be used to ensure capacity to avoid resizing the internal vector that
holds bindings.
By doing that all instructions required for instantiation are emitted
once in compilation and then reused for subsequent calls, instead of
running generic instantiation process for each call.
If the callee is already a temporary register, we don't need to copy it
to *another* temporary before evaluating arguments. None of the
arguments will clobber the existing temporary anyway.
We only need to make copies of locals here, in case the locals are
modified by something like increment/decrement expressions.
Registers and constants can slip right through, without being Mov'ed
into a temporary first.
We know that `undefined` in the global scope is always the proper
undefined value. This commit takes advantage of that by simply emitting
a constant undefined value instead.
Unfortunately we can't be so sure in other scopes.
This helps some of the Cloudflare Turnstile stuff run faster, since they
are deliberately screwing with JS engines by asking us to do a bunch of
bitwise operations on e.g 65535.56
By rounding such values in bytecode generation, the interpreter can stay
on the happy path while executing, and finish quite a bit faster.
We now fuse sequences like [LessThan, JumpIf] to JumpLessThan.
This is only allowed for temporaries (i.e VM registers) with no other
references to them.
Before this change, switch codegen would interleave bytecode like this:
(test for case 1)
(code for case 1)
(test for case 2)
(code for case 2)
This meant that we often had to make many large jumps while looking for
the matching case, since code for each case can be huge.
It now looks like this instead:
(test for case 1)
(test for case 2)
(code for case 1)
(code for case 2)
This way, we can just fall through the tests until we hit one that fits,
without having to make any large jumps.
This removes a layer of indirection in the bytecode where we had to make
sure all the initializer elements were laid out in sequential registers.
Array expressions no longer clobber registers permanently, and they can
be reused immediately afterwards.
This patch adds a register freelist to Bytecode::Generator and switches
all operands inside the generator to a new ScopedOperand type that is
ref-counted and automatically frees the register when nothing uses it.
This dramatically reduces the size of bytecode executable register
windows, which were often in the several thousands of registers for
large functions. Most functions now use less than 100 registers.
The first block in every executable will always execute first, so if it
ends up doing a ResolveThisBinding, it's fine for all other blocks
within the same executable to use the same `this` value.
Once executed, this instruction will always produce the same result
in subsequent executions, so it's okay to cache it.
Unfortunately it may throw, so we can't just hoist it to the top of
every executable, since that would break observable execution order.
If we don't have enough stack space, throw an exception while we still
can, and give the caller a chance to recover.
This particular problem will go away once we make calls non-recursive.
This means inlining all the things. This yields a 40% speedup on the for
loop microbenchmark, and everything else gets faster as well. :^)
This makes compilation take foreeeever with GCC, so I'm only enabling it
for Clang in this commit. We should figure out how to make GCC compile
this without timing out CI, since the speedup is amazing.
This commit converts the main loop in Bytecode::Interpreter to use a
label table and computed goto for fast instruction dispatch.
This yields roughly 35% speedup on the for loop microbenchmark,
and makes everything else faster as well. :^)
Now that the interpreter is unrolled, we can advance the program counter
manually based on the current instruction type.
This makes most instructions a bit smaller. :^)
This commit adds a HANDLE_INSTRUCTION macro that expands to everything
needed to handle a single instruction (invoking the handler function,
checking for exceptions, and advancing the program counter).
This gives a ~15% speed-up on a for loop microbenchmark, and makes
basically everything faster.
