During instance initialization, we build two sorts of arrays eagerly:
- We create an "anyfunc" (a `VMCallerCheckedAnyfunc`) for every function
in an instance.
- We initialize every element of a funcref table with an initializer to
a pointer to one of these anyfuncs.
Most instances will not touch (via call_indirect or table.get) all
funcref table elements. And most anyfuncs will never be referenced,
because most functions are never placed in tables or used with
`ref.func`. Thus, both of these initialization tasks are quite wasteful.
Profiling shows that a significant fraction of the remaining
instance-initialization time after our other recent optimizations is
going into these two tasks.
This PR implements two basic ideas:
- The anyfunc array can be lazily initialized as long as we retain the
information needed to do so. For now, in this PR, we just recreate the
anyfunc whenever a pointer is taken to it, because doing so is fast
enough; in the future we could keep some state to know whether the
anyfunc has been written yet and skip this work if redundant.
This technique allows us to leave the anyfunc array as uninitialized
memory, which can be a significant savings. Filling it with
initialized anyfuncs is very expensive, but even zeroing it is
expensive: e.g. in a large module, it can be >500KB.
- A funcref table can be lazily initialized as long as we retain a link
to its corresponding instance and function index for each element. A
zero in a table element means "uninitialized", and a slowpath does the
initialization.
Funcref tables are a little tricky because funcrefs can be null. We need
to distinguish "element was initially non-null, but user stored explicit
null later" from "element never touched" (ie the lazy init should not
blow away an explicitly stored null). We solve this by stealing the LSB
from every funcref (anyfunc pointer): when the LSB is set, the funcref
is initialized and we don't hit the lazy-init slowpath. We insert the
bit on storing to the table and mask it off after loading.
We do have to set up a precomputed array of `FuncIndex`s for the table
in order for this to work. We do this as part of the module compilation.
This PR also refactors the way that the runtime crate gains access to
information computed during module compilation.
Performance effect measured with in-tree benches/instantiation.rs, using
SpiderMonkey built for WASI, and with memfd enabled:
```
BEFORE:
sequential/default/spidermonkey.wasm
time: [68.569 us 68.696 us 68.856 us]
sequential/pooling/spidermonkey.wasm
time: [69.406 us 69.435 us 69.465 us]
parallel/default/spidermonkey.wasm: with 1 background thread
time: [69.444 us 69.470 us 69.497 us]
parallel/default/spidermonkey.wasm: with 16 background threads
time: [183.72 us 184.31 us 184.89 us]
parallel/pooling/spidermonkey.wasm: with 1 background thread
time: [69.018 us 69.070 us 69.136 us]
parallel/pooling/spidermonkey.wasm: with 16 background threads
time: [326.81 us 337.32 us 347.01 us]
WITH THIS PR:
sequential/default/spidermonkey.wasm
time: [6.7821 us 6.8096 us 6.8397 us]
change: [-90.245% -90.193% -90.142%] (p = 0.00 < 0.05)
Performance has improved.
sequential/pooling/spidermonkey.wasm
time: [3.0410 us 3.0558 us 3.0724 us]
change: [-95.566% -95.552% -95.537%] (p = 0.00 < 0.05)
Performance has improved.
parallel/default/spidermonkey.wasm: with 1 background thread
time: [7.2643 us 7.2689 us 7.2735 us]
change: [-89.541% -89.533% -89.525%] (p = 0.00 < 0.05)
Performance has improved.
parallel/default/spidermonkey.wasm: with 16 background threads
time: [147.36 us 148.99 us 150.74 us]
change: [-18.997% -18.081% -17.285%] (p = 0.00 < 0.05)
Performance has improved.
parallel/pooling/spidermonkey.wasm: with 1 background thread
time: [3.1009 us 3.1021 us 3.1033 us]
change: [-95.517% -95.511% -95.506%] (p = 0.00 < 0.05)
Performance has improved.
parallel/pooling/spidermonkey.wasm: with 16 background threads
time: [49.449 us 50.475 us 51.540 us]
change: [-85.423% -84.964% -84.465%] (p = 0.00 < 0.05)
Performance has improved.
```
So an improvement of something like 80-95% for a very large module (7420
functions in its one funcref table, 31928 functions total).
This PR introduces a new way of performing cooperative timeslicing that
is intended to replace the "fuel" mechanism. The tradeoff is that this
mechanism interrupts with less precision: not at deterministic points
where fuel runs out, but rather when the Engine enters a new epoch. The
generated code instrumentation is substantially faster, however, because
it does not need to do as much work as when tracking fuel; it only loads
the global "epoch counter" and does a compare-and-branch at backedges
and function prologues.
This change has been measured as ~twice as fast as fuel-based
timeslicing for some workloads, especially control-flow-intensive
workloads such as the SpiderMonkey JS interpreter on Wasm/WASI.
The intended interface is that the embedder of the `Engine` performs an
`engine.increment_epoch()` call periodically, e.g. once per millisecond.
An async invocation of a Wasm guest on a `Store` can specify a number of
epoch-ticks that are allowed before an async yield back to the
executor's event loop. (The initial amount and automatic "refills" are
configured on the `Store`, just as for fuel.) This call does only
signal-safe work (it increments an `AtomicU64`) so could be invoked from
a periodic signal, or from a thread that wakes up once per period.
* Implement the memory64 proposal in Wasmtime
This commit implements the WebAssembly [memory64 proposal][proposal] in
both Wasmtime and Cranelift. In terms of work done Cranelift ended up
needing very little work here since most of it was already prepared for
64-bit memories at one point or another. Most of the work in Wasmtime is
largely refactoring, changing a bunch of `u32` values to something else.
A number of internal and public interfaces are changing as a result of
this commit, for example:
* Acessors on `wasmtime::Memory` that work with pages now all return
`u64` unconditionally rather than `u32`. This makes it possible to
accommodate 64-bit memories with this API, but we may also want to
consider `usize` here at some point since the host can't grow past
`usize`-limited pages anyway.
* The `wasmtime::Limits` structure is removed in favor of
minimum/maximum methods on table/memory types.
* Many libcall intrinsics called by jit code now unconditionally take
`u64` arguments instead of `u32`. Return values are `usize`, however,
since the return value, if successful, is always bounded by host
memory while arguments can come from any guest.
* The `heap_addr` clif instruction now takes a 64-bit offset argument
instead of a 32-bit one. It turns out that the legalization of
`heap_addr` already worked with 64-bit offsets, so this change was
fairly trivial to make.
* The runtime implementation of mmap-based linear memories has changed
to largely work in `usize` quantities in its API and in bytes instead
of pages. This simplifies various aspects and reflects that
mmap-memories are always bound by `usize` since that's what the host
is using to address things, and additionally most calculations care
about bytes rather than pages except for the very edge where we're
going to/from wasm.
Overall I've tried to minimize the amount of `as` casts as possible,
using checked `try_from` and checked arithemtic with either error
handling or explicit `unwrap()` calls to tell us about bugs in the
future. Most locations have relatively obvious things to do with various
implications on various hosts, and I think they should all be roughly of
the right shape but time will tell. I mostly relied on the compiler
complaining that various types weren't aligned to figure out
type-casting, and I manually audited some of the more obvious locations.
I suspect we have a number of hidden locations that will panic on 32-bit
hosts if 64-bit modules try to run there, but otherwise I think we
should be generally ok (famous last words). In any case I wouldn't want
to enable this by default naturally until we've fuzzed it for some time.
In terms of the actual underlying implementation, no one should expect
memory64 to be all that fast. Right now it's implemented with
"dynamic" heaps which have a few consequences:
* All memory accesses are bounds-checked. I'm not sure how aggressively
Cranelift tries to optimize out bounds checks, but I suspect not a ton
since we haven't stressed this much historically.
* Heaps are always precisely sized. This means that every call to
`memory.grow` will incur a `memcpy` of memory from the old heap to the
new. We probably want to at least look into `mremap` on Linux and
otherwise try to implement schemes where dynamic heaps have some
reserved pages to grow into to help amortize the cost of
`memory.grow`.
The memory64 spec test suite is scheduled to now run on CI, but as with
all the other spec test suites it's really not all that comprehensive.
I've tried adding more tests for basic things as I've had to implement
guards for them, but I wouldn't really consider the testing adequate
from just this PR itself. I did try to take care in one test to actually
allocate a 4gb+ heap and then avoid running that in the pooling
allocator or in emulation because otherwise that may fail or take
excessively long.
[proposal]: https://github.com/WebAssembly/memory64/blob/master/proposals/memory64/Overview.md
* Fix some tests
* More test fixes
* Fix wasmtime tests
* Fix doctests
* Revert to 32-bit immediate offsets in `heap_addr`
This commit updates the generation of addresses in wasm code to always
use 32-bit offsets for `heap_addr`, and if the calculated offset is
bigger than 32-bits we emit a manual add with an overflow check.
* Disable memory64 for spectest fuzzing
* Fix wrong offset being added to heap addr
* More comments!
* Clarify bytes/pages
* Consolidate address calculations for atomics
This commit consolidates all calcuations of guest addresses into one
`prepare_addr` function. This notably remove the atomics-specifics paths
as well as the `prepare_load` function (now renamed to `prepare_addr`
and folded into `get_heap_addr`).
The goal of this commit is to simplify how addresses are managed in the
code generator for atomics to use all the shared infrastrucutre of other
loads/stores as well. This additionally fixes#3132 via the use of
`heap_addr` in clif for all operations.
I also added a number of tests for loads/stores with varying alignments.
Originally I was going to allow loads/stores to not be aligned since
that's what the current formal specification says, but the overview of
the threads proposal disagrees with the formal specification, so I
figured I'd leave it as-is but adding tests probably doesn't hurt.
Closes#3132
* Fix old backend
* Guarantee misalignment checks happen before out-of-bounds
This commit slims down the list of builtin intrinsics. It removes the
duplicated intrinsics for imported and locally defined items, instead
always using one intrinsic for both. This was previously inconsistently
applied where some intrinsics got two copies (one for imported one for
local) and other intrinsics got only one copy. This does add an extra
branch in intrinsics since they need to determine whether something is
local or not, but that's generally much lower cost than the intrinsics
themselves.
This also removes the `memory32_size` intrinsic, instead inlining the
codegen directly into the clif IR. This matches what the `table.size`
instruction does and removes the need for a few functions on a
`wasmtime_runtime::Instance`.
* 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 adds initial (gated) support for the multi-memory wasm
proposal. This was actually quite easy since almost all of wasmtime
already expected multi-memory to be implemented one day. The only real
substantive change is the `memory.copy` intrinsic changes, which now
accounts for the source/destination memories possibly being different.
This commit extracts the two implementations of `Compiler` into two
separate crates, `wasmtime-cranelfit` and `wasmtime-lightbeam`. The
`wasmtime-jit` crate then depends on these two and instantiates them
appropriately. The goal here is to start reducing the weight of the
`wasmtime-environ` crate, which currently serves as a common set of
types between all `wasmtime-*` crates. Long-term I'd like to remove the
dependency on Cranelift from `wasmtime-environ`, but that's going to
take a lot more work.
In the meantime I figure it's a good way to get started by separating
out the lightbeam/cranelift function compilers from the
`wasmtime-environ` crate. We can continue to iterate on moving things
out in the future, too.
