wasmer_compiler_llvm/
config.rs

1use crate::compiler::LLVMCompiler;
2pub use inkwell::OptimizationLevel as LLVMOptLevel;
3use inkwell::targets::{
4    CodeModel, InitializationConfig, RelocMode, Target as InkwellTarget, TargetMachine,
5    TargetMachineOptions, TargetTriple,
6};
7use itertools::Itertools;
8use std::fs::File;
9use std::io::{self, Write};
10use std::path::PathBuf;
11use std::sync::Arc;
12use std::{fmt::Debug, num::NonZero};
13use target_lexicon::BinaryFormat;
14use wasmer_compiler::misc::{CompiledKind, function_kind_to_filename};
15use wasmer_compiler::{Compiler, CompilerConfig, Engine, EngineBuilder, ModuleMiddleware};
16use wasmer_types::{
17    Features,
18    target::{Architecture, OperatingSystem, Target, Triple},
19};
20
21/// The InkWell ModuleInfo type
22pub type InkwellModule<'ctx> = inkwell::module::Module<'ctx>;
23
24/// The InkWell MemoryBuffer type
25pub type InkwellMemoryBuffer = inkwell::memory_buffer::MemoryBuffer;
26
27/// Callbacks to the different LLVM compilation phases.
28#[derive(Debug, Clone)]
29pub struct LLVMCallbacks {
30    debug_dir: PathBuf,
31}
32
33impl LLVMCallbacks {
34    pub fn new(debug_dir: PathBuf) -> Result<Self, io::Error> {
35        // Create the debug dir in case it doesn't exist
36        std::fs::create_dir_all(&debug_dir)?;
37        Ok(Self { debug_dir })
38    }
39
40    fn base_path(&self, module_hash: &Option<String>) -> PathBuf {
41        let mut path = self.debug_dir.clone();
42        if let Some(hash) = module_hash {
43            path.push(hash);
44        }
45        std::fs::create_dir_all(&path)
46            .unwrap_or_else(|_| panic!("cannot create debug directory: {}", path.display()));
47        path
48    }
49
50    pub fn preopt_ir(
51        &self,
52        kind: &CompiledKind,
53        module_hash: &Option<String>,
54        module: &InkwellModule,
55    ) {
56        let mut path = self.base_path(module_hash);
57        path.push(function_kind_to_filename(kind, ".preopt.ll"));
58        module
59            .print_to_file(&path)
60            .expect("Error while dumping pre optimized LLVM IR");
61    }
62    pub fn postopt_ir(
63        &self,
64        kind: &CompiledKind,
65        module_hash: &Option<String>,
66        module: &InkwellModule,
67    ) {
68        let mut path = self.base_path(module_hash);
69        path.push(function_kind_to_filename(kind, ".postopt.ll"));
70        module
71            .print_to_file(&path)
72            .expect("Error while dumping post optimized LLVM IR");
73    }
74    pub fn obj_memory_buffer(
75        &self,
76        kind: &CompiledKind,
77        module_hash: &Option<String>,
78        memory_buffer: &InkwellMemoryBuffer,
79    ) {
80        let mut path = self.base_path(module_hash);
81        path.push(function_kind_to_filename(kind, ".o"));
82        let mem_buf_slice = memory_buffer.as_slice();
83        let mut file =
84            File::create(path).expect("Error while creating debug object file from LLVM IR");
85        file.write_all(mem_buf_slice).unwrap();
86    }
87
88    pub fn asm_memory_buffer(
89        &self,
90        kind: &CompiledKind,
91        module_hash: &Option<String>,
92        asm_memory_buffer: &InkwellMemoryBuffer,
93    ) {
94        let mut path = self.base_path(module_hash);
95        path.push(function_kind_to_filename(kind, ".s"));
96        let mem_buf_slice = asm_memory_buffer.as_slice();
97        let mut file =
98            File::create(path).expect("Error while creating debug assembly file from LLVM IR");
99        file.write_all(mem_buf_slice).unwrap();
100    }
101}
102
103#[derive(Debug, Clone)]
104pub struct LLVM {
105    pub(crate) enable_nan_canonicalization: bool,
106    pub(crate) enable_non_volatile_memops: bool,
107    pub(crate) enable_verifier: bool,
108    pub(crate) enable_perfmap: bool,
109    pub(crate) opt_level: LLVMOptLevel,
110    is_pic: bool,
111    pub(crate) callbacks: Option<LLVMCallbacks>,
112    /// The middleware chain.
113    pub(crate) middlewares: Vec<Arc<dyn ModuleMiddleware>>,
114    /// Number of threads to use when compiling a module.
115    pub(crate) num_threads: NonZero<usize>,
116    pub(crate) verbose_asm: bool,
117}
118
119impl LLVM {
120    /// Creates a new configuration object with the default configuration
121    /// specified.
122    pub fn new() -> Self {
123        Self {
124            enable_nan_canonicalization: false,
125            enable_non_volatile_memops: false,
126            enable_verifier: false,
127            enable_perfmap: false,
128            opt_level: LLVMOptLevel::Aggressive,
129            is_pic: false,
130            callbacks: None,
131            middlewares: vec![],
132            verbose_asm: false,
133            num_threads: std::thread::available_parallelism().unwrap_or(NonZero::new(1).unwrap()),
134        }
135    }
136
137    /// The optimization levels when optimizing the IR.
138    pub fn opt_level(&mut self, opt_level: LLVMOptLevel) -> &mut Self {
139        self.opt_level = opt_level;
140        self
141    }
142
143    pub fn num_threads(&mut self, num_threads: NonZero<usize>) -> &mut Self {
144        self.num_threads = num_threads;
145        self
146    }
147
148    pub fn verbose_asm(&mut self, verbose_asm: bool) -> &mut Self {
149        self.verbose_asm = verbose_asm;
150        self
151    }
152
153    /// Callbacks that will triggered in the different compilation
154    /// phases in LLVM.
155    pub fn callbacks(&mut self, callbacks: Option<LLVMCallbacks>) -> &mut Self {
156        self.callbacks = callbacks;
157        self
158    }
159
160    /// For the LLVM compiler, we can use non-volatile memory operations which lead to a better performance
161    /// (but are not 100% SPEC compliant).
162    pub fn non_volatile_memops(&mut self, enable_non_volatile_memops: bool) -> &mut Self {
163        self.enable_non_volatile_memops = enable_non_volatile_memops;
164        self
165    }
166
167    fn reloc_mode(&self, binary_format: BinaryFormat) -> RelocMode {
168        if matches!(binary_format, BinaryFormat::Macho) {
169            return RelocMode::Static;
170        }
171
172        if self.is_pic {
173            RelocMode::PIC
174        } else {
175            RelocMode::Static
176        }
177    }
178
179    fn code_model(&self, binary_format: BinaryFormat) -> CodeModel {
180        // We normally use the large code model, but when targeting shared
181        // objects, we are required to use PIC. If we use PIC anyways, we lose
182        // any benefit from large code model and there's some cost on all
183        // platforms, plus some platforms (MachO) don't support PIC + large
184        // at all.
185        if matches!(binary_format, BinaryFormat::Macho) {
186            return CodeModel::Default;
187        }
188
189        if self.is_pic {
190            CodeModel::Small
191        } else {
192            CodeModel::Large
193        }
194    }
195
196    pub(crate) fn target_operating_system(&self, target: &Target) -> OperatingSystem {
197        match target.triple().operating_system {
198            OperatingSystem::Darwin(deployment) if !self.is_pic => {
199                // LLVM detects static relocation + darwin + 64-bit and
200                // force-enables PIC because MachO doesn't support that
201                // combination. They don't check whether they're targeting
202                // MachO, they check whether the OS is set to Darwin.
203                //
204                // Since both linux and darwin use SysV ABI, this should work.
205                //  but not in the case of Aarch64, there the ABI is slightly different
206                #[allow(clippy::match_single_binding)]
207                match target.triple().architecture {
208                    Architecture::Aarch64(_) => OperatingSystem::Darwin(deployment),
209                    _ => OperatingSystem::Linux,
210                }
211            }
212            other => other,
213        }
214    }
215
216    pub(crate) fn target_binary_format(&self, target: &Target) -> target_lexicon::BinaryFormat {
217        if self.is_pic {
218            target.triple().binary_format
219        } else {
220            match self.target_operating_system(target) {
221                OperatingSystem::Darwin(_) => target_lexicon::BinaryFormat::Macho,
222                _ => target_lexicon::BinaryFormat::Elf,
223            }
224        }
225    }
226
227    fn target_triple(&self, target: &Target) -> TargetTriple {
228        let architecture = if target.triple().architecture
229            == Architecture::Riscv64(target_lexicon::Riscv64Architecture::Riscv64gc)
230        {
231            target_lexicon::Architecture::Riscv64(target_lexicon::Riscv64Architecture::Riscv64)
232        } else {
233            target.triple().architecture
234        };
235        // Hack: we're using is_pic to determine whether this is a native
236        // build or not.
237
238        let operating_system = self.target_operating_system(target);
239        let binary_format = self.target_binary_format(target);
240
241        let triple = Triple {
242            architecture,
243            vendor: target.triple().vendor.clone(),
244            operating_system,
245            environment: target.triple().environment,
246            binary_format,
247        };
248        TargetTriple::create(&triple.to_string())
249    }
250
251    /// Generates the target machine for the current target
252    pub fn target_machine(&self, target: &Target) -> TargetMachine {
253        self.target_machine_with_opt(target, true)
254    }
255
256    pub(crate) fn target_machine_with_opt(
257        &self,
258        target: &Target,
259        enable_optimization: bool,
260    ) -> TargetMachine {
261        let triple = target.triple();
262        let cpu_features = &target.cpu_features();
263
264        match triple.architecture {
265            Architecture::X86_64 | Architecture::X86_32(_) => {
266                InkwellTarget::initialize_x86(&InitializationConfig {
267                    asm_parser: true,
268                    asm_printer: true,
269                    base: true,
270                    disassembler: true,
271                    info: true,
272                    machine_code: true,
273                })
274            }
275            Architecture::Aarch64(_) => InkwellTarget::initialize_aarch64(&InitializationConfig {
276                asm_parser: true,
277                asm_printer: true,
278                base: true,
279                disassembler: true,
280                info: true,
281                machine_code: true,
282            }),
283            Architecture::Riscv64(_) | Architecture::Riscv32(_) => {
284                InkwellTarget::initialize_riscv(&InitializationConfig {
285                    asm_parser: true,
286                    asm_printer: true,
287                    base: true,
288                    disassembler: true,
289                    info: true,
290                    machine_code: true,
291                })
292            }
293            Architecture::LoongArch64 => {
294                InkwellTarget::initialize_loongarch(&InitializationConfig {
295                    asm_parser: true,
296                    asm_printer: true,
297                    base: true,
298                    disassembler: true,
299                    info: true,
300                    machine_code: true,
301                })
302            }
303            _ => unimplemented!("target {} not yet supported in Wasmer", triple),
304        }
305
306        // The CPU features formatted as LLVM strings
307        // We can safely map to gcc-like features as the CPUFeatures
308        // are compliant with the same string representations as gcc.
309        let llvm_cpu_features = cpu_features
310            .iter()
311            .map(|feature| format!("+{feature}"))
312            .join(",");
313
314        let target_triple = self.target_triple(target);
315        let llvm_target = InkwellTarget::from_triple(&target_triple).unwrap();
316        let mut llvm_target_machine_options = TargetMachineOptions::new()
317            .set_cpu(match triple.architecture {
318                Architecture::Riscv64(_) => "generic-rv64",
319                Architecture::Riscv32(_) => "generic-rv32",
320                Architecture::LoongArch64 => "generic-la64",
321                _ => "generic",
322            })
323            .set_features(match triple.architecture {
324                Architecture::Riscv64(_) => "+m,+a,+c,+d,+f",
325                Architecture::Riscv32(_) => "+m,+a,+c,+d,+f",
326                Architecture::LoongArch64 => "+f,+d",
327                _ => &llvm_cpu_features,
328            })
329            .set_level(if enable_optimization {
330                self.opt_level
331            } else {
332                LLVMOptLevel::None
333            })
334            .set_reloc_mode(self.reloc_mode(self.target_binary_format(target)))
335            .set_code_model(match triple.architecture {
336                Architecture::LoongArch64 | Architecture::Riscv64(_) | Architecture::Riscv32(_) => {
337                    CodeModel::Medium
338                }
339                _ => self.code_model(self.target_binary_format(target)),
340            });
341        if let Architecture::Riscv64(_) = triple.architecture {
342            llvm_target_machine_options = llvm_target_machine_options.set_abi("lp64d");
343        }
344        let target_machine = llvm_target
345            .create_target_machine_from_options(&target_triple, llvm_target_machine_options)
346            .unwrap();
347        target_machine.set_asm_verbosity(self.verbose_asm);
348        target_machine
349    }
350}
351
352impl CompilerConfig for LLVM {
353    /// Emit code suitable for dlopen.
354    fn enable_pic(&mut self) {
355        // TODO: although we can emit PIC, the object file parser does not yet
356        // support all the relocations.
357        self.is_pic = true;
358    }
359
360    fn enable_perfmap(&mut self) {
361        self.enable_perfmap = true
362    }
363
364    /// Whether to verify compiler IR.
365    fn enable_verifier(&mut self) {
366        self.enable_verifier = true;
367    }
368
369    /// For the LLVM compiler, we can use non-volatile memory operations which lead to a better performance
370    /// (but are not 100% SPEC compliant).
371    fn enable_non_volatile_memops(&mut self) {
372        self.enable_non_volatile_memops = true;
373    }
374
375    fn canonicalize_nans(&mut self, enable: bool) {
376        self.enable_nan_canonicalization = enable;
377    }
378
379    /// Transform it into the compiler.
380    fn compiler(self: Box<Self>) -> Box<dyn Compiler> {
381        Box::new(LLVMCompiler::new(*self))
382    }
383
384    /// Pushes a middleware onto the back of the middleware chain.
385    fn push_middleware(&mut self, middleware: Arc<dyn ModuleMiddleware>) {
386        self.middlewares.push(middleware);
387    }
388
389    fn supported_features_for_target(&self, _target: &Target) -> wasmer_types::Features {
390        let mut feats = Features::default();
391        feats.exceptions(true);
392        feats.relaxed_simd(true);
393        feats.wide_arithmetic(true);
394        feats
395    }
396}
397
398impl Default for LLVM {
399    fn default() -> LLVM {
400        Self::new()
401    }
402}
403
404impl From<LLVM> for Engine {
405    fn from(config: LLVM) -> Self {
406        EngineBuilder::new(config).engine()
407    }
408}
wasmer_compiler_llvm/config.rs

wasmer_compiler_llvm/
config.rs
