This patch provides a work-around for an apparent V8 bug where
initializing multiple isolates concurrently leads to a crash on
Windows.
At the time of writing the cause of this crash is not exactly
understood, but it seems to be related to the V8 internal
function win64_unwindinfo::RegisterNonABICompliantCodeRange(),
which didn't exist in older versions of V8.
* In order to prevent ArrayBuffers from getting garbage collected by V8,
we used to store a v8::Persistent<ArrayBuffer> in a map. This patch
introduces a custom ArrayBuffer allocator which doesn't use Persistent
handles, but instead stores a pointer to the actual ArrayBuffer data
alongside with a reference count. Since creating Persistent handles
has quite a bit of overhead, this change significantly increases
performance. Various HTTP server benchmarks report about 5-10% more
requests per second than before.
* Previously the Persistent handle that prevented garbage collection had
to be released manually, and this wasn't always done, which was
causing memory leaks. This has been resolved by introducing a new
`PinnedBuf` type in both Rust and C++ that automatically re-enables
garbage collection when it goes out of scope.
* Zero-copy buffers are now correctly wrapped in an Option if there is a
possibility that they're not present. This clears up a correctness
issue where we were creating zero-length slices from a null pointer,
which is against the rules.
Op dispatch is now dynamically dispatched, so slightly less efficient.
The immeasurable perf hit is a reasonable trade for the API simplicity
that is gained here.
Additionally, instead of polling ops in a loop until none of them are
ready, the isolate will now yield to the task system after delivering
the first batch of completed ops to the javascript side.
Although this makes performance a bit worse (about 15% fewer
requests/second on the 'deno_core_http_bench' benchmark), we feel that
the advantages are worth it:
* It resolves the extremely high worst-case latency that we were seeing
on deno_core_http_bench, in particular when using the multi-threaded
Tokio runtime, which would sometimes exceed a full second.
* Before this patch, the implementation of Isolate::poll() had to loop
through all sub-futures and poll each one of them, which doesn't scale
well as the number of futures managed by the isolate goes up. This
could lead to poor performance when e.g. a server is servicing
thousands of connected clients.
This change is made in preparation for using FuturesUnordered to track
futures that are spawned by the isolate. FuturesUnordered sets up
notififications for every future that it finds to be not ready when
polled, which causes a crash if attempted outside of a task context.
* Moves how snapshots are supplied to the Isolate. Previously they were
given by Behavior::startup_data() but it was only called once at
startup. It makes more sense (and simplifies Behavior) to pass it to the
constructor of Isolate.
* Adds new libdeno type deno_snapshot instead of overloading
deno_buf.
* Adds new libdeno method to delete snapshot deno_snapshot_delete().
* Renames deno_get_snapshot() to deno_snapshot_new().
* Makes StartupData hold references to snapshots. This was implicit when
it previously held a deno_buf but is made explicit now. Note that
include_bytes!() returns a &'static [u8] and we want to avoid
copying that.
Fixes some sed errors introduced in c43cfe.
Unfortunately moving libdeno required splitting build.rs into two parts,
one for cli and one for core.
I've also removed the arm64 build - it's complicating things at this
re-org and we're not even testing it. I need to swing back to it and get
tools/test.py running for it.
