This commit removes the stack map registry and instead uses the existing
information from the store's module registry to lookup stack maps.
A trait is now used to pass the lookup context to the runtime, implemented by
`Store` to do the lookup.
With this change, module registration in `Store` is now entirely limited to
inserting the module into the module registry.
This commit moves the shared signature registry out of `Store` and into
`Engine`.
This helps eliminate work that was performed whenever a `Module` was
instantiated into a `Store`.
Now a `Module` is registered with the shared signature registry upon creation,
storing the mapping from the module's signature index space to the shared index
space.
This also refactors the "frame info" registry into a general purpose "module
registry" that is used to look up trap information, signature information, and
(soon) stack map information.
This commit is intended to be a perf improvement for instantiation of
modules with lots of functions. Previously the `lookup_shared_signature`
callback was showing up quite high in profiles as part of instantiation.
As some background, this callback is used to translate from a module's
`SignatureIndex` to a `VMSharedSignatureIndex` which the instance
stores. This callback is called for two reasons, one is to translate all
of the module's own types into `VMSharedSignatureIndex` for the purposes
of `call_indirect` (the translation of that loads from this table to
compare indices). The second reason is that a `VMCallerCheckedAnyfunc`
is prepared for all functions and this embeds a `VMSharedSignatureIndex`
inside of it.
The slow part today is that the lookup callback was called
once-per-function and each lookup involved hashing a full
`WasmFuncType`. Albeit our hash algorithm is still Rust's default
SipHash algorithm which is quite slow, but we also shouldn't need to
re-hash each signature if we see it multiple times anyway.
The fix applied in this commit is to change this lookup callback to an
`enum` where one variant is that there's a table to lookup from. This
table is a `PrimaryMap` which means that lookup is quite fast. The only
thing we need to do is to prepare the table ahead of time. Currently
this happens on the instantiation path because in my measurments the
creation of the table is quite fast compared to the rest of
instantiation. If this becomes an issue, though, we can look into
creating the table as part of `SigRegistry::register_module` and caching
it somewhere (I'm not entirely sure where but I'm sure we can figure it
out).
There's in generally not a ton of efficiency around the `SigRegistry`
type. I'm hoping though that this fixes the next-lowest-hanging-fruit in
terms of performance without complicating the implementation too much. I
tried a few variants and this change seemed like the best balance
between simplicity and still a nice performance gain.
Locally I measured an improvement in instantiation time for a large-ish
module by reducing the time from ~3ms to ~2.6ms per instance.
This commit updates the implementation of `VMOffsets` to frontload all
checked arithmetic on construction of the `VMOffsets` which allows
eliding all checked arithmetic when accessing the fields of `VMOffsets`.
For testing and such this adds a new constructor as well from a new
`VMOffsetsFields` structure which is a clone of the old definition.
This should help speed up some profile hot spots I've been seeing where
with all the checked arithmetic on field sizes this was slowing down the
various accessors during instantiation (which uses `VMOffsets` to
initialize various fields of the `VMContext`).
* Combine stack-based cleanups for faster wasm calls
This commit is an extension of #2757 where the goal is to optimize entry
into WebAssembly. Currently wasmtime has two stack-based cleanups when
entering wasm, one for the externref activation table and another for
stack limits getting reset. This commit fuses these two cleanups
together into one and moves some code around which enables less captures
for fewer closures and such to speed up calls in to wasm a bit more.
Overall this drops the execution time from 88ns to 80ns locally for me.
This also updates the atomic orderings when updating the stack limit
from `SeqCst` to `Relaxed`. While `SeqCst` is a reasonable starting
point the usage here should be safe to use `Relaxed` since we're not
using the atomics to actually protect any memory, it's simply receiving
signals from other threads.
* Determine whether a pc is wasm via a global map
The macOS implementation of traps recently changed to using mach ports
for handlers instead of signal handlers. This means that a previously
relied upon invariant, each thread fixes its own trap, was broken. The
macOS implementation worked around this by maintaining a global map from
thread id to thread local information, however, to solve the problem.
This global map is quite slow though. It involves taking a lock and
updating a hash map on all calls into WebAssembly. In my local testing
this accounts for >70% of the overhead of calling into WebAssembly on
macOS. Naturally it'd be great to remove this!
This commit fixes this issue and removes the global lock/map that is
updated on all calls into WebAssembly. The fix is to maintain a global
map of wasm modules and their trap addresses in the `wasmtime` crate.
Doing so is relatively simple since we're already tracking this
information at the `Store` level.
Once we've got a global map then the macOS implementation can use this
from a foreign thread and everything works out.
Locally this brings the overhead, on macOS specifically, of calling into
wasm from 80ns to ~20ns.
* Fix compiles
* Review comments
This commit implements a few optimizations, mainly inlining, that should
improve the performance of calling a WebAssembly function. This code
path can be quite hot depending on the embedding case and we hadn't
really put much effort into optimizing the nitty gritty.
The predominant optimization here is adding `#[inline]` to trivial
functions so performance is improved without having to compile with LTO.
Another optimization is to call `lazy_per_thread_init` when traps are
initialized per-thread (when a `Store` is created) rather than each time
a function is called. The next optimization is to change the unwind
reason in the `CallThreadState` to `MaybeUninit` to avoid extra checks
in the default case about whether we need to drop its variants (since in
the happy path we never need to drop it). The final optimization is to
optimize out a few checks when `async` support is disabled for a small
speed boost.
In a small benchmark where wasmtime calls a simple wasm function my
macOS computer dropped from 110ns to 86ns overhead, a 20% decrease. The
macOS overhead is still largely dominated by the global lock acquisition
and hash table management for traps right now, but I suspect the Linux
overhead is much better (should be on the order of ~30 or so ns).
We still have a long way to go to compete with SpiderMonkey which, in
testing, seem to have ~6ns overhead in calling the same wasm function on
my computer.
This change makes the storage of `Table` more internally consistent.
Elements are stored as raw pointers for both static and dynamic table storage.
Explicitly storing elements as pointers removes assumptions being made by the
pooling allocator in terms of the size and default representation of the
elements.
However, care must be made to properly clone externrefs for table operations.
* Consume fuel during function execution
This commit adds codegen infrastructure necessary to instrument wasm
code to consume fuel as it executes. Currently nothing is really done
with the fuel, but that'll come in later commits.
The focus of this commit is to implement the codegen infrastructure
necessary to consume fuel and account for fuel consumed correctly.
* Periodically check remaining fuel in wasm JIT code
This commit enables wasm code to periodically check to see if fuel has
run out. When fuel runs out an intrinsic is called which can do what it
needs to do in the result of fuel running out. For now a trap is thrown
to have at least some semantics in synchronous stores, but another
planned use for this feature is for asynchronous stores to periodically
yield back to the host based on fuel running out.
Checks for remaining fuel happen in the same locations as interrupt
checks, which is to say the start of the function as well as loop
headers.
* Improve codegen by caching `*const VMInterrupts`
The location of the shared interrupt value and fuel value is through a
double-indirection on the vmctx (load through the vmctx and then load
through that pointer). The second pointer in this chain, however, never
changes, so we can alter codegen to account for this and remove some
extraneous load instructions and hopefully reduce some register
pressure even maybe.
* Add tests fuel can abort infinite loops
* More fuzzing with fuel
Use fuel to time out modules in addition to time, using fuzz input to
figure out which.
* Update docs on trapping instructions
* Fix doc links
* Fix a fuzz test
* Change setting fuel to adding fuel
* Fix a doc link
* Squelch some rustdoc warnings
This commit fixes an issue where when looking up the stack map for a pc
within a function we might end up reading the *previous* function's
stack maps. This then later caused asserts to trip because we started
interpreting random data as a `VMExternRef` when it wasn't. The fix was
to add `None` markers for "this range has no stack map" in the function
ranges map.
Closes#2386
This new fuzz target exercises sequences of `table.get`s, `table.set`s, and
GCs.
It already found a couple bugs:
* Some leaks due to ref count cycles between stores and host-defined functions
closing over those stores.
* If there are no live references for a PC, Cranelift can avoid emiting an
associated stack map. This was running afoul of a debug assertion.
These instructions have fast, inline JIT paths for the common cases, and only
call out to host VM functions for the slow paths. This required some changes to
`cranelift-wasm`'s `FuncEnvironment`: instead of taking a `FuncCursor` to insert
an instruction sequence within the current basic block,
`FuncEnvironment::translate_table_{get,set}` now take a `&mut FunctionBuilder`
so that they can create whole new basic blocks. This is necessary for
implementing GC read/write barriers that involve branching (e.g. checking for
null, or whether a store buffer is at capacity).
Furthermore, it required that the `load`, `load_complex`, and `store`
instructions handle loading and storing through an `r{32,64}` rather than just
`i{32,64}` addresses. This involved making `r{32,64}` types acceptable
instantiations of the `iAddr` type variable, plus a few new instruction
encodings.
Part of #929
Also add configuration to CI to fail doc generation if any links are
broken. Unfortunately we can't blanket deny all warnings in rustdoc
since some are unconditional warnings, but for now this is hopefully
good enough.
Closes#1947
`funcref`s are implemented as `NonNull<VMCallerCheckedAnyfunc>`.
This should be more efficient than using a `VMExternRef` that points at a
`VMCallerCheckedAnyfunc` because it gets rid of an indirection, dynamic
allocation, and some reference counting.
Note that the null function reference is *NOT* a null pointer; it is a
`VMCallerCheckedAnyfunc` that has a null `func_ptr` member.
Part of #929
This allows us to detect when stack walking has failed to walk the whole stack,
and we are potentially missing on-stack roots, and therefore it would be unsafe
to do a GC because we could free objects too early, leading to use-after-free.
When we detect this scenario, we skip the GC.
For host VM code, we use plain reference counting, where cloning increments
the reference count, and dropping decrements it. We can avoid many of the
on-stack increment/decrement operations that typically plague the
performance of reference counting via Rust's ownership and borrowing system.
Moving a `VMExternRef` avoids mutating its reference count, and borrowing it
either avoids the reference count increment or delays it until if/when the
`VMExternRef` is cloned.
When passing a `VMExternRef` into compiled Wasm code, we don't want to do
reference count mutations for every compiled `local.{get,set}`, nor for
every function call. Therefore, we use a variation of **deferred reference
counting**, where we only mutate reference counts when storing
`VMExternRef`s somewhere that outlives the activation: into a global or
table. Simultaneously, we over-approximate the set of `VMExternRef`s that
are inside Wasm function activations. Periodically, we walk the stack at GC
safe points, and use stack map information to precisely identify the set of
`VMExternRef`s inside Wasm activations. Then we take the difference between
this precise set and our over-approximation, and decrement the reference
count for each of the `VMExternRef`s that are in our over-approximation but
not in the precise set. Finally, the over-approximation is replaced with the
precise set.
The `VMExternRefActivationsTable` implements the over-approximized set of
`VMExternRef`s referenced by Wasm activations. Calling a Wasm function and
passing it a `VMExternRef` moves the `VMExternRef` into the table, and the
compiled Wasm function logically "borrows" the `VMExternRef` from the
table. Similarly, `global.get` and `table.get` operations clone the gotten
`VMExternRef` into the `VMExternRefActivationsTable` and then "borrow" the
reference out of the table.
When a `VMExternRef` is returned to host code from a Wasm function, the host
increments the reference count (because the reference is logically
"borrowed" from the `VMExternRefActivationsTable` and the reference count
from the table will be dropped at the next GC).
For more general information on deferred reference counting, see *An
Examination of Deferred Reference Counting and Cycle Detection* by Quinane:
https://openresearch-repository.anu.edu.au/bitstream/1885/42030/2/hon-thesis.pdf
cc #929Fixes#1804
If you aren't expecting `VMExternRef`'s pointer-equality semantics, then these
trait implementations can be foot guns. Instead of implementing the trait, make
free functions in the `VMExternRef` namespace. This way, callers have to be a
little more explicit.
This is enough to get an `externref -> externref` identity function
passing.
However, `externref`s that are dropped by compiled Wasm code are (safely)
leaked. Follow up work will leverage cranelift's stack maps to resolve this
issue.
`VMExternRef` is a reference-counted box for any kind of data that is
external and opaque to running Wasm. Sometimes it might hold a Wasmtime
thing, other times it might hold something from a Wasmtime embedder and is
opaque even to us. It is morally equivalent to `Rc<dyn Any>` in Rust, but
additionally always fits in a pointer-sized word. `VMExternRef` is
non-nullable, but `Option<VMExternRef>` is a null pointer.
The one part of `VMExternRef` that can't ever be opaque to us is the
reference count. Even when we don't know what's inside an `VMExternRef`, we
need to be able to manipulate its reference count as we add and remove
references to it. And we need to do this from compiled Wasm code, so it must
be `repr(C)`!
`VMExternRef` itself is just a pointer to an `VMExternData`, which holds the
opaque, boxed value, its reference count, and its vtable pointer.
The `VMExternData` struct is *preceded* by the dynamically-sized value boxed
up and referenced by one or more `VMExternRef`s:
```ignore
,-------------------------------------------------------.
| |
V |
+----------------------------+-----------+-----------+ |
| dynamically-sized value... | ref_count | value_ptr |---'
+----------------------------+-----------+-----------+
| VMExternData |
+-----------------------+
^
+-------------+ |
| VMExternRef |-------------------+
+-------------+ |
|
+-------------+ |
| VMExternRef |-------------------+
+-------------+ |
|
... ===
|
+-------------+ |
| VMExternRef |-------------------'
+-------------+
```
The `value_ptr` member always points backwards to the start of the
dynamically-sized value (which is also the start of the heap allocation for
this value-and-`VMExternData` pair). Because it is a `dyn` pointer, it is
fat, and also points to the value's `Any` vtable.
The boxed value and the `VMExternRef` footer are held a single heap
allocation. The layout described above is used to make satisfying the
value's alignment easy: we just need to ensure that the heap allocation used
to hold everything satisfies its alignment. It also ensures that we don't
need a ton of excess padding between the `VMExternData` and the value for
values with large alignment.
