1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
use crate::compiler::LLVMCompiler;
use inkwell::targets::{
    CodeModel, InitializationConfig, RelocMode, Target as InkwellTarget, TargetMachine,
    TargetTriple,
};
pub use inkwell::OptimizationLevel as LLVMOptLevel;
use itertools::Itertools;
use std::fmt::Debug;
use std::sync::Arc;
use target_lexicon::Architecture;
use wasmer_compiler::{Compiler, CompilerConfig, Engine, EngineBuilder, ModuleMiddleware};
use wasmer_types::{FunctionType, LocalFunctionIndex, Target, Triple};

/// The InkWell ModuleInfo type
pub type InkwellModule<'ctx> = inkwell::module::Module<'ctx>;

/// The InkWell MemoryBuffer type
pub type InkwellMemoryBuffer = inkwell::memory_buffer::MemoryBuffer;

/// The compiled function kind, used for debugging in the `LLVMCallbacks`.
#[derive(Debug, Clone)]
pub enum CompiledKind {
    // A locally-defined function in the Wasm file.
    Local(LocalFunctionIndex),
    // A function call trampoline for a given signature.
    FunctionCallTrampoline(FunctionType),
    // A dynamic function trampoline for a given signature.
    DynamicFunctionTrampoline(FunctionType),
    // An entire Wasm module.
    Module,
}

/// Callbacks to the different LLVM compilation phases.
pub trait LLVMCallbacks: Debug + Send + Sync {
    fn preopt_ir(&self, function: &CompiledKind, module: &InkwellModule);
    fn postopt_ir(&self, function: &CompiledKind, module: &InkwellModule);
    fn obj_memory_buffer(&self, function: &CompiledKind, memory_buffer: &InkwellMemoryBuffer);
}

#[derive(Debug, Clone)]
pub struct LLVM {
    pub(crate) enable_nan_canonicalization: bool,
    pub(crate) enable_verifier: bool,
    pub(crate) opt_level: LLVMOptLevel,
    is_pic: bool,
    pub(crate) callbacks: Option<Arc<dyn LLVMCallbacks>>,
    /// The middleware chain.
    pub(crate) middlewares: Vec<Arc<dyn ModuleMiddleware>>,
}

impl LLVM {
    /// Creates a new configuration object with the default configuration
    /// specified.
    pub fn new() -> Self {
        Self {
            enable_nan_canonicalization: false,
            enable_verifier: false,
            opt_level: LLVMOptLevel::Aggressive,
            is_pic: false,
            callbacks: None,
            middlewares: vec![],
        }
    }

    /// The optimization levels when optimizing the IR.
    pub fn opt_level(&mut self, opt_level: LLVMOptLevel) -> &mut Self {
        self.opt_level = opt_level;
        self
    }

    /// Callbacks that will triggered in the different compilation
    /// phases in LLVM.
    pub fn callbacks(&mut self, callbacks: Option<Arc<dyn LLVMCallbacks>>) -> &mut Self {
        self.callbacks = callbacks;
        self
    }

    fn reloc_mode(&self) -> RelocMode {
        if self.is_pic {
            RelocMode::PIC
        } else {
            RelocMode::Static
        }
    }

    fn code_model(&self) -> CodeModel {
        // We normally use the large code model, but when targeting shared
        // objects, we are required to use PIC. If we use PIC anyways, we lose
        // any benefit from large code model and there's some cost on all
        // platforms, plus some platforms (MachO) don't support PIC + large
        // at all.
        if self.is_pic {
            CodeModel::Small
        } else {
            CodeModel::Large
        }
    }

    fn target_triple(&self, target: &Target) -> TargetTriple {
        // Hack: we're using is_pic to determine whether this is a native
        // build or not.
        let operating_system = if target.triple().operating_system
            == wasmer_types::OperatingSystem::Darwin
            && !self.is_pic
        {
            // LLVM detects static relocation + darwin + 64-bit and
            // force-enables PIC because MachO doesn't support that
            // combination. They don't check whether they're targeting
            // MachO, they check whether the OS is set to Darwin.
            //
            // Since both linux and darwin use SysV ABI, this should work.
            //  but not in the case of Aarch64, there the ABI is slightly different
            #[allow(clippy::match_single_binding)]
            match target.triple().architecture {
                _ => wasmer_types::OperatingSystem::Linux,
            }
        } else {
            target.triple().operating_system
        };
        let binary_format = if self.is_pic {
            target.triple().binary_format
        } else {
            target_lexicon::BinaryFormat::Elf
        };
        let triple = Triple {
            architecture: target.triple().architecture,
            vendor: target.triple().vendor.clone(),
            operating_system,
            environment: target.triple().environment,
            binary_format,
        };
        TargetTriple::create(&triple.to_string())
    }

    /// Generates the target machine for the current target
    pub fn target_machine(&self, target: &Target) -> TargetMachine {
        let triple = target.triple();
        let cpu_features = &target.cpu_features();

        match triple.architecture {
            Architecture::X86_64 | Architecture::X86_32(_) => {
                InkwellTarget::initialize_x86(&InitializationConfig {
                    asm_parser: true,
                    asm_printer: true,
                    base: true,
                    disassembler: true,
                    info: true,
                    machine_code: true,
                })
            }
            Architecture::Aarch64(_) => InkwellTarget::initialize_aarch64(&InitializationConfig {
                asm_parser: true,
                asm_printer: true,
                base: true,
                disassembler: true,
                info: true,
                machine_code: true,
            }),
            // Architecture::Arm(_) => InkwellTarget::initialize_arm(&InitializationConfig {
            //     asm_parser: true,
            //     asm_printer: true,
            //     base: true,
            //     disassembler: true,
            //     info: true,
            //     machine_code: true,
            // }),
            _ => unimplemented!("target {} not yet supported in Wasmer", triple),
        }

        // The CPU features formatted as LLVM strings
        // We can safely map to gcc-like features as the CPUFeatures
        // are compliant with the same string representations as gcc.
        let llvm_cpu_features = cpu_features
            .iter()
            .map(|feature| format!("+{}", feature.to_string()))
            .join(",");

        let target_triple = self.target_triple(target);
        let llvm_target = InkwellTarget::from_triple(&target_triple).unwrap();
        llvm_target
            .create_target_machine(
                &target_triple,
                "generic",
                &llvm_cpu_features,
                self.opt_level,
                self.reloc_mode(),
                self.code_model(),
            )
            .unwrap()
    }
}

impl CompilerConfig for LLVM {
    /// Emit code suitable for dlopen.
    fn enable_pic(&mut self) {
        // TODO: although we can emit PIC, the object file parser does not yet
        // support all the relocations.
        self.is_pic = true;
    }

    /// Whether to verify compiler IR.
    fn enable_verifier(&mut self) {
        self.enable_verifier = true;
    }

    fn enable_nan_canonicalization(&mut self) {
        self.enable_nan_canonicalization = true;
    }

    fn canonicalize_nans(&mut self, enable: bool) {
        self.enable_nan_canonicalization = enable;
    }

    /// Transform it into the compiler.
    fn compiler(self: Box<Self>) -> Box<dyn Compiler> {
        Box::new(LLVMCompiler::new(*self))
    }

    /// Pushes a middleware onto the back of the middleware chain.
    fn push_middleware(&mut self, middleware: Arc<dyn ModuleMiddleware>) {
        self.middlewares.push(middleware);
    }
}

impl Default for LLVM {
    fn default() -> LLVM {
        Self::new()
    }
}

impl From<LLVM> for Engine {
    fn from(config: LLVM) -> Self {
        EngineBuilder::new(config).engine()
    }
}