* Add reference output tests for JS operations
This commit starts adding a test suite which checks in, to the
repository, test assertions for both the JS and wasm file outputs of a
Rust crate compiled with `#[wasm_bindgen]`. These aren't intended to be
exhaustive or large scale tests, but rather micro-tests to help observe
the changes in `wasm-bindgen`'s output over time.
The motivation for this commit is basically overhauling how all the GC
passes work in `wasm-bindgen` today. The reorganization is also included
in this commit as well.
Previously `wasm-bindgen` would, in an ad-hoc fashion, run the GC passes
of `walrus` in a bunch of places to ensure that less "garbage" was seen
by future passes. This not only was a source of slowdown but it also was
pretty brittle since `wasm-bindgen` kept breaking if extra iteams leaked
through.
The strategy taken in this commit is to have one precise location for a
GC pass, and everything goes through there. This is achieved by:
* All internal exports are removed immediately when generating the
nonstandard wasm interface types section. Internal exports,
intrinsics, and runtime support are all referenced by the various
instructions and/or sections that use them. This means that we now
have precise tracking of what an adapter uses.
* This in turn enables us to implement the `add_gc_roots` function for
`walrus` custom sections, which in turn allows walrus GC passes to do
what `unexport_unused_intrinsics` did before. That function is now no
longer necessary, but effectively works the same way. All intrinsics
are unexported at the beginning and then they're selectively
re-imported and re-exported through the JS glue generation pass as
necessary and defined by the bindings.
* Passes like the `anyref` pass are now much more precise about the
intrinsics that they work with. The `anyref` pass also deletes any
internal intrinsics found and also does some rewriting of the adapters
aftewards now to hook up calls to the heap count import to the heap
count intrinsic in the wasm module.
* Fix handling of __wbindgen_realloc
The final user of the `require_internal_export` function was
`__wbindgen_realloc`. This usage has now been removed by updating how we
handle usage of the `realloc` function.
The wasm interface types standard doesn't have a `realloc` function
slot, nor do I think it ever will. This means that as a polyfill for
wasm interface types we'll always have to support the lack of `realloc`.
For direct Rust to JS, however, we can still optionally handle
`realloc`. This is all handled with a few internal changes.
* Custom `StringToMemory` instructions now exist. These have an extra
`realloc` slot to store an intrinsic, if found.
* Our custom instructions are lowered to the standard instructions when
generating an interface types section.
* The `realloc` function, if present, is passed as an argument like the
malloc function when passing strings to wasm. If it's not present we
use a slower fallback, but if it's present we use the faster
implementation.
This should mean that there's little-to-no impact on existing users of
`wasm-bindgen`, but this should continue to still work for wasm
interface types polyfills and such. Additionally the GC passes now work
in that they don't delete `__wbindgen_realloc` which we later try to
reference.
* Add an empty test for the anyref pass
* Precisely track I32FromOptionAnyref's dependencies
This depends on the anyref table and a function to allocate an index if
the anyref pass is running, so be sure to track that in the instruction
itself for GC rooting.
* Trim extraneous exports from nop anyref module
Or if you're otherwise not using anyref slices, don't force some
intrinsics to exist.
* Remove globals from reference tests
Looks like these values adjust in slight but insignificant ways over
time
* Update the anyref xform tests
* Catch all closures by walking all `Descriptor` values and looking for
either `Function` or `Closure`.
* Update the correct arguments for wasm by ensuring that the closure
modifications skip the first two arguments.
This commit fixes an issue where bindings for
`__wbindgen_object_drop_ref` are generated even if the function isn't
actually used by the final wasm file. This is currently due to the fact
that we run gc passes pretty late in wasm-bindgen and one of the
intrinsics that ended up getting gc'd referenced the
`__wbindgen_object_drop_ref` intrinsic function.
The fix here is somewhat naive by just updating the intrinsic to not
actually use `__wbindgen_object_drop_ref`. This may not be a long-term
solution but it should be good enough for now at least.
Most of the CLI crates were already in the 2018 edition, and it turns
out that one of the macro crates was already in the 2018 edition so we
may as well move everything to the 2018 edition!
Always nice to remove those `extern crate` statements nowadays!
This commit also does a `cargo fmt --all` to make sure we're conforming
with style again.
This commit adds experimental support to `wasm-bindgen` to emit and
leverage the `anyref` native wasm type. This native type is still in a
proposal status (the reference-types proposal). The intention of
`anyref` is to be able to directly hold JS values in wasm and pass the
to imported functions, namely to empower eventual host bindings (now
renamed WebIDL bindings) integration where we can skip JS shims
altogether for many imports.
This commit doesn't actually affect wasm-bindgen's behavior at all
as-is, but rather this support requires an opt-in env var to be
configured. Once the support is stable in browsers it's intended that
this will add a CLI switch for turning on this support, eventually
defaulting it to `true` in the far future.
The basic strategy here is to take the `stack` and `slab` globals in the
generated JS glue and move them into wasm using a table. This new table
in wasm is managed at the fringes via injected shims. At
`wasm-bindgen`-time the CLI will rewrite exports and imports with shims
that actually use `anyref` if needed, performing loads/stores inside the
wasm module instead of externally in the wasm module.
This should provide a boost over what we have today, but it's not a
fantastic strategy long term. We have a more grand vision for `anyref`
being a first-class type in the language, but that's on a much longer
horizon and this is currently thought to be the best we can do in terms
of integration in the near future.
The stack/heap JS tables are combined into one wasm table. The stack
starts at the end of the table and grows down with a stack pointer (also
injected). The heap starts at the end and grows up (state managed in
linear memory). The anyref transformation here will hook up various
intrinsics in wasm-bindgen to the runtime functionality if the anyref
supoprt is enabled.
The main tricky treatment here was applied to closures, where we need JS
to use a different function pointer than the one Rust gives it to use a
JS function pointer empowered with anyref. This works by switching up a
bit how descriptors work, embedding the shims to call inside descriptors
rather than communicated at runtime. This means that we're accessing
constant values in the generated JS and we can just update the constant
value accessed.
