wasmer_compiler/
compiler.rs

1//! This module mainly outputs the `Compiler` trait that custom
2//! compilers will need to implement.
3
4use std::cmp::Reverse;
5
6use crate::progress::ProgressContext;
7use crate::types::{module::CompileModuleInfo, symbols::SymbolRegistry};
8use crate::{
9    FunctionBodyData, ModuleTranslationState,
10    lib::std::{boxed::Box, sync::Arc},
11    translator::ModuleMiddleware,
12    types::function::Compilation,
13};
14use crossbeam_channel::unbounded;
15use enumset::EnumSet;
16use itertools::Itertools;
17use wasmer_types::{
18    CompilationProgressCallback, Features, LocalFunctionIndex,
19    entity::PrimaryMap,
20    error::CompileError,
21    target::{CpuFeature, Target, UserCompilerOptimizations},
22};
23#[cfg(feature = "translator")]
24use wasmparser::{Validator, WasmFeatures};
25
26/// The compiler configuration options.
27pub trait CompilerConfig {
28    /// Enable Position Independent Code (PIC).
29    ///
30    /// This is required for shared object generation (Native Engine),
31    /// but will make the JIT Engine to fail, since PIC is not yet
32    /// supported in the JIT linking phase.
33    fn enable_pic(&mut self) {
34        // By default we do nothing, each backend will need to customize this
35        // in case they do something special for emitting PIC code.
36    }
37
38    /// Enable compiler IR verification.
39    ///
40    /// For compilers capable of doing so, this enables internal consistency
41    /// checking.
42    fn enable_verifier(&mut self) {
43        // By default we do nothing, each backend will need to customize this
44        // in case they create an IR that they can verify.
45    }
46
47    /// Enable generation of perfmaps to sample the JIT compiled frames.
48    fn enable_perfmap(&mut self) {
49        // By default we do nothing, each backend will need to customize this
50        // in case they create an IR that they can verify.
51    }
52
53    /// For the LLVM compiler, we can use non-volatile memory operations which lead to a better performance
54    /// (but are not 100% SPEC compliant).
55    fn enable_non_volatile_memops(&mut self) {}
56
57    /// Enable NaN canonicalization.
58    ///
59    /// NaN canonicalization is useful when trying to run WebAssembly
60    /// deterministically across different architectures.
61    fn canonicalize_nans(&mut self, _enable: bool) {
62        // By default we do nothing, each backend will need to customize this
63        // in case they create an IR that they can verify.
64    }
65
66    /// Gets the custom compiler config
67    fn compiler(self: Box<Self>) -> Box<dyn Compiler>;
68
69    /// Gets the default features for this compiler in the given target
70    fn default_features_for_target(&self, target: &Target) -> Features {
71        self.supported_features_for_target(target)
72    }
73
74    /// Gets the supported features for this compiler in the given target
75    fn supported_features_for_target(&self, _target: &Target) -> Features {
76        Features::default()
77    }
78
79    /// Pushes a middleware onto the back of the middleware chain.
80    fn push_middleware(&mut self, middleware: Arc<dyn ModuleMiddleware>);
81}
82
83impl<T> From<T> for Box<dyn CompilerConfig + 'static>
84where
85    T: CompilerConfig + 'static,
86{
87    fn from(other: T) -> Self {
88        Box::new(other)
89    }
90}
91
92/// An implementation of a Compiler from parsed WebAssembly module to Compiled native code.
93pub trait Compiler: Send + std::fmt::Debug {
94    /// Returns a descriptive name for this compiler.
95    ///
96    /// Note that this is an API breaking change since 3.0
97    fn name(&self) -> &str;
98
99    /// Returns the deterministic id of this compiler. Same compilers with different
100    /// optimizations map to different deterministic IDs.
101    fn deterministic_id(&self) -> String;
102
103    /// Add suggested optimizations to this compiler.
104    ///
105    /// # Note
106    ///
107    /// Not every compiler supports every optimization. This function may fail (i.e. not set the
108    /// suggested optimizations) silently if the underlying compiler does not support one or
109    /// more optimizations.
110    fn with_opts(
111        &mut self,
112        suggested_compiler_opts: &UserCompilerOptimizations,
113    ) -> Result<(), CompileError> {
114        _ = suggested_compiler_opts;
115        Ok(())
116    }
117
118    /// Validates a module.
119    ///
120    /// It returns the a succesful Result in case is valid, `CompileError` in case is not.
121    #[cfg(feature = "translator")]
122    fn validate_module(&self, features: &Features, data: &[u8]) -> Result<(), CompileError> {
123        let mut wasm_features = WasmFeatures::empty();
124        wasm_features.set(WasmFeatures::BULK_MEMORY, features.bulk_memory);
125        wasm_features.set(WasmFeatures::THREADS, features.threads);
126        wasm_features.set(WasmFeatures::REFERENCE_TYPES, features.reference_types);
127        wasm_features.set(WasmFeatures::MULTI_VALUE, features.multi_value);
128        wasm_features.set(WasmFeatures::SIMD, features.simd);
129        wasm_features.set(WasmFeatures::TAIL_CALL, features.tail_call);
130        wasm_features.set(WasmFeatures::MULTI_MEMORY, features.multi_memory);
131        wasm_features.set(WasmFeatures::MEMORY64, features.memory64);
132        wasm_features.set(WasmFeatures::EXCEPTIONS, features.exceptions);
133        wasm_features.set(WasmFeatures::EXTENDED_CONST, features.extended_const);
134        wasm_features.set(WasmFeatures::RELAXED_SIMD, features.relaxed_simd);
135        wasm_features.set(WasmFeatures::WIDE_ARITHMETIC, features.wide_arithmetic);
136        wasm_features.set(WasmFeatures::MUTABLE_GLOBAL, true);
137        wasm_features.set(WasmFeatures::SATURATING_FLOAT_TO_INT, true);
138        wasm_features.set(WasmFeatures::FLOATS, true);
139        wasm_features.set(WasmFeatures::SIGN_EXTENSION, true);
140        wasm_features.set(WasmFeatures::GC_TYPES, true);
141
142        let mut validator = Validator::new_with_features(wasm_features);
143        validator
144            .validate_all(data)
145            .map_err(|e| CompileError::Validate(format!("{e}")))?;
146        Ok(())
147    }
148
149    /// Compiles a parsed module.
150    ///
151    /// It returns the [`Compilation`] or a [`CompileError`].
152    fn compile_module(
153        &self,
154        target: &Target,
155        module: &CompileModuleInfo,
156        module_translation: &ModuleTranslationState,
157        // The list of function bodies
158        function_body_inputs: PrimaryMap<LocalFunctionIndex, FunctionBodyData<'_>>,
159        progress_callback: Option<&CompilationProgressCallback>,
160    ) -> Result<Compilation, CompileError>;
161
162    /// Compiles a module into a native object file.
163    ///
164    /// It returns the bytes as a `&[u8]` or a [`CompileError`].
165    fn experimental_native_compile_module(
166        &self,
167        _target: &Target,
168        _module: &CompileModuleInfo,
169        _module_translation: &ModuleTranslationState,
170        // The list of function bodies
171        _function_body_inputs: &PrimaryMap<LocalFunctionIndex, FunctionBodyData<'_>>,
172        _symbol_registry: &dyn SymbolRegistry,
173        // The metadata to inject into the wasmer_metadata section of the object file.
174        _wasmer_metadata: &[u8],
175    ) -> Option<Result<Vec<u8>, CompileError>> {
176        None
177    }
178
179    /// Get the middlewares for this compiler
180    fn get_middlewares(&self) -> &[Arc<dyn ModuleMiddleware>];
181
182    /// Get the CpuFeatues used by the compiler
183    fn get_cpu_features_used(&self, cpu_features: &EnumSet<CpuFeature>) -> EnumSet<CpuFeature> {
184        *cpu_features
185    }
186
187    /// Get whether `perfmap` is enabled or not.
188    fn get_perfmap_enabled(&self) -> bool {
189        false
190    }
191}
192
193/// A bucket containing a group of functions and their total size, used to balance compilation units for parallel compilation.
194pub struct FunctionBucket<'a> {
195    functions: Vec<(LocalFunctionIndex, &'a FunctionBodyData<'a>)>,
196    /// IR size of the bucket (in bytes).
197    pub size: usize,
198}
199
200impl<'a> FunctionBucket<'a> {
201    /// Creates a new, empty `FunctionBucket`.
202    pub fn new() -> Self {
203        Self {
204            functions: Vec::new(),
205            size: 0,
206        }
207    }
208}
209
210/// Build buckets sized by function length to keep compilation units balanced for parallel compilation.
211pub fn build_function_buckets<'a>(
212    function_body_inputs: &'a PrimaryMap<LocalFunctionIndex, FunctionBodyData<'a>>,
213    bucket_threshold_size: u64,
214) -> Vec<FunctionBucket<'a>> {
215    let mut function_bodies = function_body_inputs
216        .iter()
217        .sorted_by_key(|(id, body)| Reverse((body.data.len(), id.as_u32())))
218        .collect_vec();
219
220    let mut buckets = Vec::new();
221
222    while !function_bodies.is_empty() {
223        let mut next_function_body = Vec::with_capacity(function_bodies.len());
224        let mut bucket = FunctionBucket::new();
225
226        for (fn_index, fn_body) in function_bodies.into_iter() {
227            if bucket.size + fn_body.data.len() <= bucket_threshold_size as usize
228                // Huge functions must fit into a bucket!
229                || bucket.size == 0
230            {
231                bucket.size += fn_body.data.len();
232                bucket.functions.push((fn_index, fn_body));
233            } else {
234                next_function_body.push((fn_index, fn_body));
235            }
236        }
237
238        function_bodies = next_function_body;
239        buckets.push(bucket);
240    }
241
242    buckets
243}
244
245/// Represents a function that has been compiled by the backend compiler.
246pub trait CompiledFunction {}
247
248/// Translates a function from its input representation to a compiled form.
249pub trait FuncTranslator {}
250
251/// Compile function buckets largest-first via the channel (instead of Rayon's par_iter).
252#[allow(clippy::too_many_arguments)]
253pub fn translate_function_buckets<'a, C, T, F, G>(
254    pool: &rayon::ThreadPool,
255    func_translator_builder: F,
256    translate_fn: G,
257    progress: Option<ProgressContext>,
258    buckets: &[FunctionBucket<'a>],
259) -> Result<Vec<C>, CompileError>
260where
261    T: FuncTranslator,
262    C: CompiledFunction + Send + Sync,
263    F: Fn() -> T + Send + Sync + Copy,
264    G: Fn(&mut T, &LocalFunctionIndex, &FunctionBodyData) -> Result<C, CompileError>
265        + Send
266        + Sync
267        + Copy,
268{
269    let progress = progress.as_ref();
270
271    let functions = pool.install(|| {
272        let (bucket_tx, bucket_rx) = unbounded::<&FunctionBucket<'a>>();
273        for bucket in buckets {
274            bucket_tx.send(bucket).map_err(|e| {
275                CompileError::Resource(format!("cannot allocate crossbeam channel item: {e}"))
276            })?;
277        }
278        drop(bucket_tx);
279
280        let (result_tx, result_rx) =
281            unbounded::<Result<Vec<(LocalFunctionIndex, C)>, CompileError>>();
282
283        pool.scope(|s| {
284            let worker_count = pool.current_num_threads().max(1);
285            for _ in 0..worker_count {
286                let bucket_rx = bucket_rx.clone();
287                let result_tx = result_tx.clone();
288                s.spawn(move |_| {
289                    let mut func_translator = func_translator_builder();
290
291                    while let Ok(bucket) = bucket_rx.recv() {
292                        let bucket_result = (|| {
293                            let mut translated_functions = Vec::new();
294                            for (i, input) in bucket.functions.iter() {
295                                let translated = translate_fn(&mut func_translator, i, input)?;
296                                if let Some(progress) = progress {
297                                    progress.notify_steps(input.data.len() as u64)?;
298                                }
299                                translated_functions.push((*i, translated));
300                            }
301                            Ok(translated_functions)
302                        })();
303
304                        if result_tx.send(bucket_result).is_err() {
305                            break;
306                        }
307                    }
308                });
309            }
310        });
311
312        drop(result_tx);
313        let mut functions = Vec::with_capacity(buckets.iter().map(|b| b.functions.len()).sum());
314        for _ in 0..buckets.len() {
315            match result_rx.recv().map_err(|e| {
316                CompileError::Resource(format!("cannot allocate crossbeam channel item: {e}"))
317            })? {
318                Ok(bucket_functions) => functions.extend(bucket_functions),
319                Err(err) => return Err(err),
320            }
321        }
322        Ok(functions)
323    })?;
324
325    Ok(functions
326        .into_iter()
327        .sorted_by_key(|x| x.0)
328        .map(|(_, body)| body)
329        .collect_vec())
330}
331
332/// Byte size threshold for a function that is considered large.
333pub const WASM_LARGE_FUNCTION_THRESHOLD: u64 = 100_000;
334
335/// Estimated byte size of a trampoline (used for progress bar reporting).
336pub const WASM_TRAMPOLINE_ESTIMATED_BODY_SIZE: u64 = 1_000;
wasmer_compiler/compiler.rs

wasmer_compiler/
compiler.rs
