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

wasmer_compiler/
compiler.rs
