Book a Demo!
CoCalc Logo Icon
StoreFeaturesDocsShareSupportNewsAboutPoliciesSign UpSign In
bytecodealliance
GitHub Repository: bytecodealliance/wasmtime
Path: blob/main/crates/environ/src/component/translate/adapt.rs
1692 views
1
//! Identification and creation of fused adapter modules in Wasmtime.
2
//!
3
//! A major piece of the component model is the ability for core wasm modules to
4
//! talk to each other through the use of lifted and lowered functions. For
5
//! example one core wasm module can export a function which is lifted. Another
6
//! component could import that lifted function, lower it, and pass it as the
7
//! import to another core wasm module. This is what Wasmtime calls "adapter
8
//! fusion" where two core wasm functions are coming together through the
9
//! component model.
10
//!
11
//! There are a few ingredients during adapter fusion:
12
//!
13
//! * A core wasm function which is "lifted".
14
//! * A "lift type" which is the type that the component model function had in
15
//! the original component
16
//! * A "lower type" which is the type that the component model function has
17
//! in the destination component (the one the uses `canon lower`)
18
//! * Configuration options for both the lift and the lower operations such as
19
//! memories, reallocs, etc.
20
//!
21
//! With these ingredients combined Wasmtime must produce a function which
22
//! connects the two components through the options specified. The fused adapter
23
//! performs tasks such as validation of passed values, copying data between
24
//! linear memories, etc.
25
//!
26
//! Wasmtime's current implementation of fused adapters is designed to reduce
27
//! complexity elsewhere as much as possible while also being suitable for being
28
//! used as a polyfill for the component model in JS environments as well. To
29
//! that end Wasmtime implements a fused adapter with another wasm module that
30
//! it itself generates on the fly. The usage of WebAssembly for fused adapters
31
//! has a number of advantages:
32
//!
33
//! * There is no need to create a raw Cranelift-based compiler. This is where
34
//! majority of "unsafety" lives in Wasmtime so reducing the need to lean on
35
//! this or audit another compiler is predicted to weed out a whole class of
36
//! bugs in the fused adapter compiler.
37
//!
38
//! * As mentioned above generation of WebAssembly modules means that this is
39
//! suitable for use in JS environments. For example a hypothetical tool which
40
//! polyfills a component onto the web today would need to do something for
41
//! adapter modules, and ideally the adapters themselves are speedy. While
42
//! this could all be written in JS the adapting process is quite nontrivial
43
//! so sharing code with Wasmtime would be ideal.
44
//!
45
//! * Using WebAssembly insulates the implementation to bugs to a certain
46
//! degree. While logic bugs are still possible it should be much more
47
//! difficult to have segfaults or things like that. With adapters exclusively
48
//! executing inside a WebAssembly sandbox like everything else the failure
49
//! modes to the host at least should be minimized.
50
//!
51
//! * Integration into the runtime is relatively simple, the adapter modules are
52
//! just another kind of wasm module to instantiate and wire up at runtime.
53
//! The goal is that the `GlobalInitializer` list that is processed at runtime
54
//! will have all of its `Adapter`-using variants erased by the time it makes
55
//! its way all the way up to Wasmtime. This means that the support in
56
//! Wasmtime prior to adapter modules is actually the same as the support
57
//! after adapter modules are added, keeping the runtime fiddly bits quite
58
//! minimal.
59
//!
60
//! This isn't to say that this approach isn't without its disadvantages of
61
//! course. For now though this seems to be a reasonable set of tradeoffs for
62
//! the development stage of the component model proposal.
63
//!
64
//! ## Creating adapter modules
65
//!
66
//! With WebAssembly itself being used to implement fused adapters, Wasmtime
67
//! still has the question of how to organize the adapter functions into actual
68
//! wasm modules.
69
//!
70
//! The first thing you might reach for is to put all the adapters into the same
71
//! wasm module. This cannot be done, however, because some adapters may depend
72
//! on other adapters (transitively) to be created. This means that if
73
//! everything were in the same module there would be no way to instantiate the
74
//! module. An example of this dependency is an adapter (A) used to create a
75
//! core wasm instance (M) whose exported memory is then referenced by another
76
//! adapter (B). In this situation the adapter B cannot be in the same module
77
//! as adapter A because B needs the memory of M but M is created with A which
78
//! would otherwise create a circular dependency.
79
//!
80
//! The second possibility of organizing adapter modules would be to place each
81
//! fused adapter into its own module. Each `canon lower` would effectively
82
//! become a core wasm module instantiation at that point. While this works it's
83
//! currently believed to be a bit too fine-grained. For example it would mean
84
//! that importing a dozen lowered functions into a module could possibly result
85
//! in up to a dozen different adapter modules. While this possibility could
86
//! work it has been ruled out as "probably too expensive at runtime".
87
//!
88
//! Thus the purpose and existence of this module is now evident -- this module
89
//! exists to identify what exactly goes into which adapter module. This will
90
//! evaluate the `GlobalInitializer` lists coming out of the `inline` pass and
91
//! insert `InstantiateModule` entries for where adapter modules should be
92
//! created.
93
//!
94
//! ## Partitioning adapter modules
95
//!
96
//! Currently this module does not attempt to be really all that fancy about
97
//! grouping adapters into adapter modules. The main idea is that most items
98
//! within an adapter module are likely to be close together since they're
99
//! theoretically going to be used for an instantiation of a core wasm module
100
//! just after the fused adapter was declared. With that in mind the current
101
//! algorithm is a one-pass approach to partitioning everything into adapter
102
//! modules.
103
//!
104
//! Adapters were identified in-order as part of the inlining phase of
105
//! translation where we're guaranteed that once an adapter is identified
106
//! it can't depend on anything identified later. The pass implemented here is
107
//! to visit all transitive dependencies of an adapter. If one of the
108
//! dependencies of an adapter is an adapter in the current adapter module
109
//! being built then the current module is finished and a new adapter module is
110
//! started. This should quickly partition adapters into contiugous chunks of
111
//! their index space which can be in adapter modules together.
112
//!
113
//! There's probably more general algorithms for this but for now this should be
114
//! fast enough as it's "just" a linear pass. As we get more components over
115
//! time this may want to be revisited if too many adapter modules are being
116
//! created.
117
118
use crate::EntityType;
119
use crate::component::translate::*;
120
use crate::fact;
121
use std::collections::HashSet;
122
123
/// Metadata information about a fused adapter.
124
#[derive(Debug, Clone, Hash, Eq, PartialEq)]
125
pub struct Adapter {
126
/// The type used when the original core wasm function was lifted.
127
///
128
/// Note that this could be different than `lower_ty` (but still matches
129
/// according to subtyping rules).
130
pub lift_ty: TypeFuncIndex,
131
/// Canonical ABI options used when the function was lifted.
132
pub lift_options: AdapterOptions,
133
/// The type used when the function was lowered back into a core wasm
134
/// function.
135
///
136
/// Note that this could be different than `lift_ty` (but still matches
137
/// according to subtyping rules).
138
pub lower_ty: TypeFuncIndex,
139
/// Canonical ABI options used when the function was lowered.
140
pub lower_options: AdapterOptions,
141
/// The original core wasm function which was lifted.
142
pub func: dfg::CoreDef,
143
}
144
145
/// The data model for objects that are not unboxed in locals.
146
#[derive(Debug, Clone, Hash, Eq, PartialEq)]
147
pub enum DataModel {
148
/// Data is stored in GC objects.
149
Gc {},
150
151
/// Data is stored in a linear memory.
152
LinearMemory {
153
/// An optional memory definition supplied.
154
memory: Option<dfg::CoreExport<MemoryIndex>>,
155
/// If `memory` is specified, whether it's a 64-bit memory.
156
memory64: bool,
157
/// An optional definition of `realloc` to used.
158
realloc: Option<dfg::CoreDef>,
159
},
160
}
161
162
/// Configuration options which can be specified as part of the canonical ABI
163
/// in the component model.
164
#[derive(Debug, Clone, Hash, Eq, PartialEq)]
165
pub struct AdapterOptions {
166
/// The Wasmtime-assigned component instance index where the options were
167
/// originally specified.
168
pub instance: RuntimeComponentInstanceIndex,
169
/// How strings are encoded.
170
pub string_encoding: StringEncoding,
171
/// The async callback function used by these options, if specified.
172
pub callback: Option<dfg::CoreDef>,
173
/// An optional definition of a `post-return` to use.
174
pub post_return: Option<dfg::CoreDef>,
175
/// Whether to use the async ABI for lifting or lowering.
176
pub async_: bool,
177
/// The core function type that is being lifted from / lowered to.
178
pub core_type: ModuleInternedTypeIndex,
179
/// The data model used by this adapter: linear memory or GC objects.
180
pub data_model: DataModel,
181
}
182
183
impl<'data> Translator<'_, 'data> {
184
/// This is the entrypoint of functionality within this module which
185
/// performs all the work of identifying adapter usages and organizing
186
/// everything into adapter modules.
187
///
188
/// This will mutate the provided `component` in-place and fill out the dfg
189
/// metadata for adapter modules.
190
pub(super) fn partition_adapter_modules(&mut self, component: &mut dfg::ComponentDfg) {
191
// Visit each adapter, in order of its original definition, during the
192
// partitioning. This allows for the guarantee that dependencies are
193
// visited in a topological fashion ideally.
194
let mut state = PartitionAdapterModules::default();
195
for (id, adapter) in component.adapters.iter() {
196
state.adapter(component, id, adapter);
197
}
198
state.finish_adapter_module();
199
200
// Now that all adapters have been partitioned into modules this loop
201
// generates a core wasm module for each adapter module, translates
202
// the module using standard core wasm translation, and then fills out
203
// the dfg metadata for each adapter.
204
for (module_id, adapter_module) in state.adapter_modules.iter() {
205
let mut module =
206
fact::Module::new(self.types.types(), self.tunables.debug_adapter_modules);
207
let mut names = Vec::with_capacity(adapter_module.adapters.len());
208
for adapter in adapter_module.adapters.iter() {
209
let name = format!("adapter{}", adapter.as_u32());
210
module.adapt(&name, &component.adapters[*adapter]);
211
names.push(name);
212
}
213
let wasm = module.encode();
214
let imports = module.imports().to_vec();
215
216
// Extend the lifetime of the owned `wasm: Vec<u8>` on the stack to
217
// a higher scope defined by our original caller. That allows to
218
// transform `wasm` into `&'data [u8]` which is much easier to work
219
// with here.
220
let wasm = &*self.scope_vec.push(wasm);
221
if log::log_enabled!(log::Level::Trace) {
222
match wasmprinter::print_bytes(wasm) {
223
Ok(s) => log::trace!("generated adapter module:\n{s}"),
224
Err(e) => log::trace!("failed to print adapter module: {e}"),
225
}
226
}
227
228
// With the wasm binary this is then pushed through general
229
// translation, validation, etc. Note that multi-memory is
230
// specifically enabled here since the adapter module is highly
231
// likely to use that if anything is actually indirected through
232
// memory.
233
self.validator.reset();
234
let static_module_index = self.static_modules.next_key();
235
let translation = ModuleEnvironment::new(
236
self.tunables,
237
&mut self.validator,
238
self.types.module_types_builder(),
239
static_module_index,
240
)
241
.translate(Parser::new(0), wasm)
242
.expect("invalid adapter module generated");
243
244
// Record, for each adapter in this adapter module, the module that
245
// the adapter was placed within as well as the function index of
246
// the adapter in the wasm module generated. Note that adapters are
247
// partitioned in-order so we're guaranteed to push the adapters
248
// in-order here as well. (with an assert to double-check)
249
for (adapter, name) in adapter_module.adapters.iter().zip(&names) {
250
let index = translation.module.exports[name];
251
let i = component.adapter_partitionings.push((module_id, index));
252
assert_eq!(i, *adapter);
253
}
254
255
// Finally the metadata necessary to instantiate this adapter
256
// module is also recorded in the dfg. This metadata will be used
257
// to generate `GlobalInitializer` entries during the linearization
258
// final phase.
259
assert_eq!(imports.len(), translation.module.imports().len());
260
let args = imports
261
.iter()
262
.zip(translation.module.imports())
263
.map(|(arg, (_, _, ty))| fact_import_to_core_def(component, arg, ty))
264
.collect::<Vec<_>>();
265
let static_module_index2 = self.static_modules.push(translation);
266
assert_eq!(static_module_index, static_module_index2);
267
let id = component.adapter_modules.push((static_module_index, args));
268
assert_eq!(id, module_id);
269
}
270
}
271
}
272
273
fn fact_import_to_core_def(
274
dfg: &mut dfg::ComponentDfg,
275
import: &fact::Import,
276
ty: EntityType,
277
) -> dfg::CoreDef {
278
fn unwrap_memory(def: &dfg::CoreDef) -> dfg::CoreExport<MemoryIndex> {
279
match def {
280
dfg::CoreDef::Export(e) => e.clone().map_index(|i| match i {
281
EntityIndex::Memory(i) => i,
282
_ => unreachable!(),
283
}),
284
_ => unreachable!(),
285
}
286
}
287
288
let mut simple_intrinsic = |trampoline: dfg::Trampoline| {
289
let signature = ty.unwrap_func();
290
let index = dfg
291
.trampolines
292
.push((signature.unwrap_module_type_index(), trampoline));
293
dfg::CoreDef::Trampoline(index)
294
};
295
match import {
296
fact::Import::CoreDef(def) => def.clone(),
297
fact::Import::Transcode {
298
op,
299
from,
300
from64,
301
to,
302
to64,
303
} => {
304
let from = dfg.memories.push(unwrap_memory(from));
305
let to = dfg.memories.push(unwrap_memory(to));
306
let signature = ty.unwrap_func();
307
let index = dfg.trampolines.push((
308
signature.unwrap_module_type_index(),
309
dfg::Trampoline::Transcoder {
310
op: *op,
311
from,
312
from64: *from64,
313
to,
314
to64: *to64,
315
},
316
));
317
dfg::CoreDef::Trampoline(index)
318
}
319
fact::Import::ResourceTransferOwn => simple_intrinsic(dfg::Trampoline::ResourceTransferOwn),
320
fact::Import::ResourceTransferBorrow => {
321
simple_intrinsic(dfg::Trampoline::ResourceTransferBorrow)
322
}
323
fact::Import::ResourceEnterCall => simple_intrinsic(dfg::Trampoline::ResourceEnterCall),
324
fact::Import::ResourceExitCall => simple_intrinsic(dfg::Trampoline::ResourceExitCall),
325
fact::Import::PrepareCall { memory } => simple_intrinsic(dfg::Trampoline::PrepareCall {
326
memory: memory.as_ref().map(|v| dfg.memories.push(unwrap_memory(v))),
327
}),
328
fact::Import::SyncStartCall { callback } => {
329
simple_intrinsic(dfg::Trampoline::SyncStartCall {
330
callback: callback.clone().map(|v| dfg.callbacks.push(v)),
331
})
332
}
333
fact::Import::AsyncStartCall {
334
callback,
335
post_return,
336
} => simple_intrinsic(dfg::Trampoline::AsyncStartCall {
337
callback: callback.clone().map(|v| dfg.callbacks.push(v)),
338
post_return: post_return.clone().map(|v| dfg.post_returns.push(v)),
339
}),
340
fact::Import::FutureTransfer => simple_intrinsic(dfg::Trampoline::FutureTransfer),
341
fact::Import::StreamTransfer => simple_intrinsic(dfg::Trampoline::StreamTransfer),
342
fact::Import::ErrorContextTransfer => {
343
simple_intrinsic(dfg::Trampoline::ErrorContextTransfer)
344
}
345
}
346
}
347
348
#[derive(Default)]
349
struct PartitionAdapterModules {
350
/// The next adapter module that's being created. This may be empty.
351
next_module: AdapterModuleInProgress,
352
353
/// The set of items which are known to be defined which the adapter module
354
/// in progress is allowed to depend on.
355
defined_items: HashSet<Def>,
356
357
/// Finished adapter modules that won't be added to.
358
///
359
/// In theory items could be added to preexisting modules here but to keep
360
/// this pass linear this is never modified after insertion.
361
adapter_modules: PrimaryMap<dfg::AdapterModuleId, AdapterModuleInProgress>,
362
}
363
364
#[derive(Default)]
365
struct AdapterModuleInProgress {
366
/// The adapters which have been placed into this module.
367
adapters: Vec<dfg::AdapterId>,
368
}
369
370
/// Items that adapters can depend on.
371
///
372
/// Note that this is somewhat of a flat list and is intended to mostly model
373
/// core wasm instances which are side-effectful unlike other host items like
374
/// lowerings or always-trapping functions.
375
#[derive(Copy, Clone, Hash, Eq, PartialEq)]
376
enum Def {
377
Adapter(dfg::AdapterId),
378
Instance(dfg::InstanceId),
379
}
380
381
impl PartitionAdapterModules {
382
fn adapter(&mut self, dfg: &dfg::ComponentDfg, id: dfg::AdapterId, adapter: &Adapter) {
383
// Visit all dependencies of this adapter and if anything depends on
384
// the current adapter module in progress then a new adapter module is
385
// started.
386
self.adapter_options(dfg, &adapter.lift_options);
387
self.adapter_options(dfg, &adapter.lower_options);
388
self.core_def(dfg, &adapter.func);
389
390
// With all dependencies visited this adapter is added to the next
391
// module.
392
//
393
// This will either get added the preexisting module if this adapter
394
// didn't depend on anything in that module itself or it will be added
395
// to a fresh module if this adapter depended on something that the
396
// current adapter module created.
397
log::debug!("adding {id:?} to adapter module");
398
self.next_module.adapters.push(id);
399
}
400
401
fn adapter_options(&mut self, dfg: &dfg::ComponentDfg, options: &AdapterOptions) {
402
if let Some(def) = &options.callback {
403
self.core_def(dfg, def);
404
}
405
if let Some(def) = &options.post_return {
406
self.core_def(dfg, def);
407
}
408
match &options.data_model {
409
DataModel::Gc {} => {
410
// Nothing to do here yet.
411
}
412
DataModel::LinearMemory {
413
memory,
414
memory64: _,
415
realloc,
416
} => {
417
if let Some(memory) = memory {
418
self.core_export(dfg, memory);
419
}
420
if let Some(def) = realloc {
421
self.core_def(dfg, def);
422
}
423
}
424
}
425
}
426
427
fn core_def(&mut self, dfg: &dfg::ComponentDfg, def: &dfg::CoreDef) {
428
match def {
429
dfg::CoreDef::Export(e) => self.core_export(dfg, e),
430
dfg::CoreDef::Adapter(id) => {
431
// If this adapter is already defined then we can safely depend
432
// on it with no consequences.
433
if self.defined_items.contains(&Def::Adapter(*id)) {
434
log::debug!("using existing adapter {id:?} ");
435
return;
436
}
437
438
log::debug!("splitting module needing {id:?} ");
439
440
// .. otherwise we found a case of an adapter depending on an
441
// adapter-module-in-progress meaning that the current adapter
442
// module must be completed and then a new one is started.
443
self.finish_adapter_module();
444
assert!(self.defined_items.contains(&Def::Adapter(*id)));
445
}
446
447
// These items can't transitively depend on an adapter
448
dfg::CoreDef::Trampoline(_) | dfg::CoreDef::InstanceFlags(_) => {}
449
}
450
}
451
452
fn core_export<T>(&mut self, dfg: &dfg::ComponentDfg, export: &dfg::CoreExport<T>) {
453
// When an adapter depends on an exported item it actually depends on
454
// the instance of that exported item. The caveat here is that the
455
// adapter not only depends on that particular instance, but also all
456
// prior instances to that instance as well because instance
457
// instantiation order is fixed and cannot change.
458
//
459
// To model this the instance index space is looped over here and while
460
// an instance hasn't been visited it's visited. Note that if an
461
// instance has already been visited then all prior instances have
462
// already been visited so there's no need to continue.
463
let mut instance = export.instance;
464
while self.defined_items.insert(Def::Instance(instance)) {
465
self.instance(dfg, instance);
466
if instance.as_u32() == 0 {
467
break;
468
}
469
instance = dfg::InstanceId::from_u32(instance.as_u32() - 1);
470
}
471
}
472
473
fn instance(&mut self, dfg: &dfg::ComponentDfg, instance: dfg::InstanceId) {
474
log::debug!("visiting instance {instance:?}");
475
476
// ... otherwise if this is the first timet he instance has been seen
477
// then the instances own arguments are recursively visited to find
478
// transitive dependencies on adapters.
479
match &dfg.instances[instance] {
480
dfg::Instance::Static(_, args) => {
481
for arg in args.iter() {
482
self.core_def(dfg, arg);
483
}
484
}
485
dfg::Instance::Import(_, args) => {
486
for (_, values) in args {
487
for (_, def) in values {
488
self.core_def(dfg, def);
489
}
490
}
491
}
492
}
493
}
494
495
fn finish_adapter_module(&mut self) {
496
if self.next_module.adapters.is_empty() {
497
return;
498
}
499
500
// Reset the state of the current module-in-progress and then flag all
501
// pending adapters as now defined since the current module is being
502
// committed.
503
let module = mem::take(&mut self.next_module);
504
for adapter in module.adapters.iter() {
505
let inserted = self.defined_items.insert(Def::Adapter(*adapter));
506
assert!(inserted);
507
}
508
let idx = self.adapter_modules.push(module);
509
log::debug!("finishing adapter module {idx:?}");
510
}
511
}
512
513