Skip to main content

byteflow/jit/
compiler.rs

1#![allow(unsafe_code)]
2
3use std::collections::{HashMap, HashSet, VecDeque};
4
5use cranelift_codegen::ir::types;
6use cranelift_codegen::ir::{AbiParam, Block, Function, InstBuilder, MachMemFlags, Signature, Value as ClifValue};
7use cranelift_frontend::{FunctionBuilder, FunctionBuilderContext};
8use cranelift_module::{Linkage, Module};
9
10use crate::{Chunk, Instruction, Opcode, Value};
11
12use super::error::CompileError;
13use super::exit::{JIT_BUDGET, JIT_CONTINUE, JIT_RETURN, JIT_TRAP};
14use super::frame::{
15    JitEntry, OFF_BUDGET, OFF_CALL_DEPTH, OFF_CALL_STACK, OFF_EXIT_KIND, OFF_FUNCTION,
16    OFF_PC, OFF_REGISTER_COUNT, OFF_RETURN_REG, MAX_JIT_CALL_DEPTH,
17};
18use super::trace::{CompiledTrace, TraceKey, TraceSpan, MAX_TRACE_LENGTH};
19
20/// Maps bytecode register indices to Cranelift SSA values for the current trace.
21pub struct RegisterMap {
22    values: Vec<Option<ClifValue>>,
23}
24
25impl RegisterMap {
26    pub fn new(registers: usize) -> Self {
27        Self {
28            values: vec![None; registers],
29        }
30    }
31
32    pub fn get(&self, reg: usize) -> Option<ClifValue> {
33        self.values.get(reg).copied().flatten()
34    }
35
36    pub fn set(&mut self, reg: usize, value: ClifValue) {
37        if let Some(slot) = self.values.get_mut(reg) {
38            *slot = Some(value);
39        }
40    }
41
42    pub fn clear(&mut self) {
43        for slot in &mut self.values {
44            *slot = None;
45        }
46    }
47}
48
49pub struct TraceCompiler<'a> {
50    module: &'a mut cranelift_jit::JITModule,
51}
52
53impl<'a> TraceCompiler<'a> {
54    pub fn new(module: &'a mut cranelift_jit::JITModule) -> Self {
55        Self { module }
56    }
57
58    pub fn compile_trace(
59        &mut self,
60        chunk: &Chunk,
61        key: TraceKey,
62    ) -> Result<CompiledTrace, CompileError> {
63        let def = chunk.function(key.function).ok_or(CompileError::EmptyTrace {
64            function: key.function,
65            pc: key.entry_pc,
66        })?;
67        let region = collect_trace_region(chunk, key)?;
68        if region.blocks.is_empty() {
69            return Err(CompileError::EmptyTrace {
70                function: key.function,
71                pc: key.entry_pc,
72            });
73        }
74
75        let end_pc = region_max_pc(&region, key)?;
76        let span = TraceSpan {
77            function: key.function,
78            range: key.entry_pc..end_pc.saturating_add(1),
79        };
80
81        let target = self.module.target_config();
82        let pointer_type = target.pointer_type();
83        let mut sig = Signature::new(target.default_call_conv);
84        sig.params.push(AbiParam::new(pointer_type));
85
86        let func_id = self
87            .module
88            .declare_function(
89                &format!("trace_{}_{}", key.function, key.entry_pc),
90                Linkage::Local,
91                &sig,
92            )
93            .map_err(|e| CompileError::Module(e.to_string()))?;
94
95        let mut ctx = cranelift_codegen::Context::new();
96        ctx.func = Function::with_name_signature(
97            cranelift_codegen::ir::UserFuncName::user(0, func_id.as_u32()),
98            sig,
99        );
100
101        let mut func_ctx = FunctionBuilderContext::new();
102        let mut builder = FunctionBuilder::new(&mut ctx.func, &mut func_ctx);
103        let entry = builder.create_block();
104        builder.append_block_params_for_function_params(entry);
105        builder.switch_to_block(entry);
106
107        let frame_ptr = builder.block_params(entry)[0];
108        let slots_ptr = builder.ins().load(pointer_type, MachMemFlags::new(), frame_ptr, 0);
109        let budget_ptr = field_ptr(&mut builder, frame_ptr, pointer_type, OFF_BUDGET);
110        let call_depth_ptr = field_ptr(&mut builder, frame_ptr, pointer_type, OFF_CALL_DEPTH);
111        let call_stack_ptr = {
112            let addr = field_ptr(&mut builder, frame_ptr, pointer_type, OFF_CALL_STACK);
113            builder
114                .ins()
115                .load(pointer_type, MachMemFlags::new(), addr, 0)
116        };
117
118        let supports_calls = region.blocks.iter().any(|block| {
119            block
120                .pcs
121                .iter()
122                .any(|pc| chunk.code[*pc as usize].op == Opcode::Call)
123        });
124
125        let trap_block = builder.create_block();
126        let return_block = builder.create_block();
127
128        let mut pc_to_block = HashMap::new();
129        for block in &region.blocks {
130            pc_to_block.insert(block.start_pc, builder.create_block());
131        }
132
133        let entry_clif = *pc_to_block
134            .get(&key.entry_pc)
135            .ok_or(CompileError::EmptyTrace {
136                function: key.function,
137                pc: key.entry_pc,
138            })?;
139        builder.ins().jump(entry_clif, &[]);
140
141        for block in &region.blocks {
142            let clif_block = pc_to_block[&block.start_pc];
143            builder.switch_to_block(clif_block);
144            let mut registers = RegisterMap::new(def.num_registers as usize);
145
146            for pc in &block.pcs {
147                emit_budget_tick(
148                    &mut builder,
149                    budget_ptr,
150                    frame_ptr,
151                    pointer_type,
152                    *pc,
153                );
154                let instr = chunk.code[*pc as usize];
155                let mut emit_ctx = EmitContext {
156                    builder: &mut builder,
157                    chunk,
158                    registers: &mut registers,
159                    slots_ptr,
160                    frame_ptr,
161                    pointer_type,
162                    trap_block,
163                    return_block,
164                    pc_to_block: &pc_to_block,
165                    call_depth_ptr,
166                    call_stack_ptr,
167                    supports_calls,
168                };
169                emit_instruction(&mut emit_ctx, *pc, instr)?;
170            }
171        }
172
173        builder.switch_to_block(trap_block);
174        builder.seal_block(trap_block);
175        let trap_pc = region_max_pc(&region, key)?;
176        write_exit(
177            &mut builder,
178            frame_ptr,
179            pointer_type,
180            JIT_TRAP,
181            trap_pc,
182            0,
183        );
184
185        builder.switch_to_block(return_block);
186        builder.seal_block(return_block);
187        let (last_pc, return_reg) = match last_return_in_region(chunk, &region) {
188            Some(pair) => pair,
189            // Side-exit-only traces never reach this block; metadata is unused.
190            None => (key.entry_pc, 0),
191        };
192        write_exit(
193            &mut builder,
194            frame_ptr,
195            pointer_type,
196            JIT_RETURN,
197            last_pc,
198            return_reg,
199        );
200
201        builder.seal_all_blocks();
202        builder.finalize(self.module.target_config());
203
204        self.module
205            .define_function(func_id, &mut ctx)
206            .map_err(|e| CompileError::Module(e.to_string()))?;
207        self.module.clear_context(&mut ctx);
208        self.module
209            .finalize_definitions()
210            .map_err(|e| CompileError::Module(e.to_string()))?;
211
212        let code = self.module.get_finalized_function(func_id);
213        let entry: JitEntry = unsafe { std::mem::transmute(code) };
214
215        Ok(CompiledTrace { span, entry })
216    }
217}
218
219struct TraceBlock {
220    start_pc: u32,
221    pcs: Vec<u32>,
222}
223
224struct TraceRegion {
225    blocks: Vec<TraceBlock>,
226}
227
228fn region_max_pc(region: &TraceRegion, key: TraceKey) -> Result<u32, CompileError> {
229    region
230        .blocks
231        .iter()
232        .flat_map(|b| b.pcs.iter())
233        .copied()
234        .max()
235        .ok_or(CompileError::EmptyTrace {
236            function: key.function,
237            pc: key.entry_pc,
238        })
239}
240
241fn last_return_in_region(chunk: &Chunk, region: &TraceRegion) -> Option<(u32, u32)> {
242    region
243        .blocks
244        .iter()
245        .flat_map(|b| b.pcs.iter().map(|pc| (*pc, chunk.code[*pc as usize])))
246        .rev()
247        .find(|(_, instr)| instr.op == Opcode::Return)
248        .map(|(pc, instr)| (pc, u32::from(instr.a)))
249}
250
251fn jump_target(jpc: u32, imm: i32) -> u32 {
252    (i64::from(jpc) + 1 + i64::from(imm)) as u32
253}
254
255fn branch_target(bpc: u32, imm: i32) -> u32 {
256    jump_target(bpc, imm)
257}
258
259struct EmitContext<'a, 'b> {
260    builder: &'a mut FunctionBuilder<'b>,
261    chunk: &'a Chunk,
262    registers: &'a mut RegisterMap,
263    slots_ptr: ClifValue,
264    frame_ptr: ClifValue,
265    pointer_type: types::Type,
266    trap_block: Block,
267    return_block: Block,
268    pc_to_block: &'a HashMap<u32, Block>,
269    call_depth_ptr: ClifValue,
270    call_stack_ptr: ClifValue,
271    supports_calls: bool,
272}
273
274fn emit_instruction(
275    ctx: &mut EmitContext<'_, '_>,
276    pc: u32,
277    instr: Instruction,
278) -> Result<(), CompileError> {
279    let builder = &mut *ctx.builder;
280    let chunk = ctx.chunk;
281    let registers = &mut *ctx.registers;
282    let slots_ptr = ctx.slots_ptr;
283    let frame_ptr = ctx.frame_ptr;
284    let pointer_type = ctx.pointer_type;
285    let trap_block = ctx.trap_block;
286    let return_block = ctx.return_block;
287    let supports_calls = ctx.supports_calls;
288    let pc_to_block = ctx.pc_to_block;
289    let call_depth_ptr = ctx.call_depth_ptr;
290    let call_stack_ptr = ctx.call_stack_ptr;
291    match instr.op {
292        Opcode::LoadImm => {
293            let v = builder.ins().iconst(types::I64, i64::from(instr.imm));
294            registers.set(instr.a as usize, v);
295        }
296        Opcode::LoadConst => {
297            let idx = instr.imm as u32;
298            let konst = chunk.constant(idx).ok_or(CompileError::UnsupportedOpcode {
299                opcode: Opcode::LoadConst,
300                pc,
301            })?;
302            let v = match konst {
303                Value::Int(i) => builder.ins().iconst(types::I64, *i),
304                Value::Bool(b) => builder.ins().iconst(types::I64, i64::from(*b)),
305                _ => {
306                    return Err(CompileError::UnsupportedOpcode {
307                        opcode: Opcode::LoadConst,
308                        pc,
309                    });
310                }
311            };
312            registers.set(instr.a as usize, v);
313        }
314        Opcode::Move => {
315            let v = load_reg(
316                builder,
317                registers,
318                slots_ptr,
319                pointer_type,
320                instr.b,
321                pc,
322            )?;
323            registers.set(instr.a as usize, v);
324        }
325        Opcode::Add | Opcode::Sub | Opcode::Mul | Opcode::Eq | Opcode::Lt | Opcode::Le => {
326            let lhs = load_reg(
327                builder,
328                registers,
329                slots_ptr,
330                pointer_type,
331                instr.b,
332                pc,
333            )?;
334            let rhs = load_reg(
335                builder,
336                registers,
337                slots_ptr,
338                pointer_type,
339                instr.c,
340                pc,
341            )?;
342            let result = match instr.op {
343                Opcode::Add => builder.ins().iadd(lhs, rhs),
344                Opcode::Sub => builder.ins().isub(lhs, rhs),
345                Opcode::Mul => builder.ins().imul(lhs, rhs),
346                Opcode::Eq => {
347                    let cmp = builder.ins().icmp(
348                        cranelift_codegen::ir::condcodes::IntCC::Equal,
349                        lhs,
350                        rhs,
351                    );
352                    bool_as_i64(builder, cmp)
353                }
354                Opcode::Lt => {
355                    let cmp = builder.ins().icmp(
356                        cranelift_codegen::ir::condcodes::IntCC::SignedLessThan,
357                        lhs,
358                        rhs,
359                    );
360                    bool_as_i64(builder, cmp)
361                }
362                Opcode::Le => {
363                    let cmp = builder.ins().icmp(
364                        cranelift_codegen::ir::condcodes::IntCC::SignedLessThanOrEqual,
365                        lhs,
366                        rhs,
367                    );
368                    bool_as_i64(builder, cmp)
369                }
370                _ => unreachable!(),
371            };
372            registers.set(instr.a as usize, result);
373        }
374        Opcode::Div | Opcode::Mod => {
375            let lhs = load_reg(
376                builder,
377                registers,
378                slots_ptr,
379                pointer_type,
380                instr.b,
381                pc,
382            )?;
383            let rhs = load_reg(
384                builder,
385                registers,
386                slots_ptr,
387                pointer_type,
388                instr.c,
389                pc,
390            )?;
391            flush_registers(builder, registers, slots_ptr, pointer_type);
392            let zero = builder.ins().iconst(types::I64, 0);
393            let is_zero = builder.ins().icmp(
394                cranelift_codegen::ir::condcodes::IntCC::Equal,
395                rhs,
396                zero,
397            );
398            let ok = builder.create_block();
399            builder.ins().brif(is_zero, trap_block, &[], ok, &[]);
400            builder.switch_to_block(ok);
401            builder.seal_block(ok);
402            let result = if instr.op == Opcode::Div {
403                builder.ins().sdiv(lhs, rhs)
404            } else {
405                builder.ins().srem(lhs, rhs)
406            };
407            registers.set(instr.a as usize, result);
408        }
409        Opcode::Neg => {
410            let v = load_reg(
411                builder,
412                registers,
413                slots_ptr,
414                pointer_type,
415                instr.b,
416                pc,
417            )?;
418            let zero = builder.ins().iconst(types::I64, 0);
419            registers.set(instr.a as usize, builder.ins().isub(zero, v));
420        }
421        Opcode::Jump => {
422            flush_registers(builder, registers, slots_ptr, pointer_type);
423            let target_pc = jump_target(pc, instr.imm);
424            branch_to_pc(
425                builder,
426                frame_ptr,
427                pointer_type,
428                pc_to_block,
429                target_pc,
430            );
431        }
432        Opcode::Branch => {
433            let cond = load_reg(
434                builder,
435                registers,
436                slots_ptr,
437                pointer_type,
438                instr.a,
439                pc,
440            )?;
441            flush_registers(builder, registers, slots_ptr, pointer_type);
442            let zero = builder.ins().iconst(types::I64, 0);
443            let is_falsy = builder.ins().icmp(
444                cranelift_codegen::ir::condcodes::IntCC::Equal,
445                cond,
446                zero,
447            );
448            let fall_pc = pc + 1;
449            let falsy_pc = branch_target(pc, instr.imm);
450            match (pc_to_block.get(&fall_pc), pc_to_block.get(&falsy_pc)) {
451                (Some(fall), Some(falsy)) => {
452                    builder.ins().brif(is_falsy, *falsy, &[], *fall, &[]);
453                }
454                (Some(fall), None) => {
455                    let side = builder.create_block();
456                    builder.ins().brif(is_falsy, side, &[], *fall, &[]);
457                    builder.switch_to_block(side);
458                    builder.seal_block(side);
459                    write_exit(
460                        builder,
461                        frame_ptr,
462                        pointer_type,
463                        JIT_CONTINUE,
464                        falsy_pc,
465                        0,
466                    );
467                }
468                (None, Some(falsy)) => {
469                    let side = builder.create_block();
470                    builder.ins().brif(is_falsy, *falsy, &[], side, &[]);
471                    builder.switch_to_block(side);
472                    builder.seal_block(side);
473                    write_exit(
474                        builder,
475                        frame_ptr,
476                        pointer_type,
477                        JIT_CONTINUE,
478                        fall_pc,
479                        0,
480                    );
481                }
482                (None, None) => {
483                    write_exit(
484                        builder,
485                        frame_ptr,
486                        pointer_type,
487                        JIT_CONTINUE,
488                        falsy_pc,
489                        0,
490                    );
491                }
492            }
493        }
494        Opcode::Return => {
495            let v = load_reg(
496                builder,
497                registers,
498                slots_ptr,
499                pointer_type,
500                instr.a,
501                pc,
502            )?;
503            store_slot(builder, slots_ptr, pointer_type, instr.a, v);
504            if supports_calls {
505                emit_return(
506                    builder,
507                    frame_ptr,
508                    pointer_type,
509                    call_depth_ptr,
510                    call_stack_ptr,
511                    slots_ptr,
512                    registers,
513                    instr.a,
514                    pc,
515                );
516            } else {
517                write_u32_field(builder, frame_ptr, pointer_type, OFF_PC, pc);
518                builder.ins().jump(return_block, &[]);
519            }
520        }
521        Opcode::Nop => {}
522        Opcode::Call => {
523            emit_call(
524                builder,
525                chunk,
526                registers,
527                slots_ptr,
528                frame_ptr,
529                pointer_type,
530                pc_to_block,
531                call_depth_ptr,
532                call_stack_ptr,
533                pc,
534                instr,
535            )?;
536        }
537        other => {
538            return Err(CompileError::UnsupportedOpcode { opcode: other, pc });
539        }
540    }
541    Ok(())
542}
543
544fn branch_to_pc(
545    builder: &mut FunctionBuilder<'_>,
546    frame_ptr: ClifValue,
547    pointer_type: types::Type,
548    pc_to_block: &HashMap<u32, Block>,
549    target_pc: u32,
550) {
551    if let Some(block) = pc_to_block.get(&target_pc) {
552        builder.ins().jump(*block, &[]);
553    } else {
554        write_exit(
555            builder,
556            frame_ptr,
557            pointer_type,
558            JIT_CONTINUE,
559            target_pc,
560            0,
561        );
562    }
563}
564
565fn flush_registers(
566    builder: &mut FunctionBuilder<'_>,
567    registers: &RegisterMap,
568    slots_ptr: ClifValue,
569    pointer_type: types::Type,
570) {
571    for (reg, value) in registers.values.iter().enumerate() {
572        if let Some(v) = *value {
573            store_slot(builder, slots_ptr, pointer_type, reg as u8, v);
574        }
575    }
576}
577
578fn load_reg(
579    builder: &mut FunctionBuilder<'_>,
580    registers: &mut RegisterMap,
581    slots_ptr: ClifValue,
582    pointer_type: types::Type,
583    reg: u8,
584    _pc: u32,
585) -> Result<ClifValue, CompileError> {
586    if let Some(v) = registers.get(reg as usize) {
587        return Ok(v);
588    }
589    let offset = builder.ins().iconst(pointer_type, i64::from(u32::from(reg) * 8));
590    let addr = builder.ins().iadd(slots_ptr, offset);
591    let v = builder
592        .ins()
593        .load(types::I64, MachMemFlags::new(), addr, 0);
594    registers.set(reg as usize, v);
595    Ok(v)
596}
597
598fn collect_trace_region(chunk: &Chunk, key: TraceKey) -> Result<TraceRegion, CompileError> {
599    let mut queue = VecDeque::from([key.entry_pc]);
600    let mut seen_starts = HashSet::new();
601    let mut blocks = Vec::new();
602    let mut total = 0usize;
603
604    while let Some(start) = queue.pop_front() {
605        if !seen_starts.insert(start) {
606            continue;
607        }
608        if total >= MAX_TRACE_LENGTH {
609            break;
610        }
611
612        let mut pcs = Vec::new();
613        let mut pc = start;
614        while let Some(instr) = chunk.code.get(pc as usize).copied() {
615            if is_effect_opcode(instr.op) {
616                break;
617            }
618            if instr.op == Opcode::LoadConst {
619                let idx = instr.imm as u32;
620                match chunk.constant(idx) {
621                    Some(Value::Int(_) | Value::Bool(_)) => {}
622                    _ => break,
623                }
624            }
625            pcs.push(pc);
626            total += 1;
627            if total >= MAX_TRACE_LENGTH {
628                break;
629            }
630
631            match instr.op {
632                Opcode::Return => break,
633                Opcode::Jump => {
634                    queue.push_back(jump_target(pc, instr.imm));
635                    break;
636                }
637                Opcode::Branch => {
638                    queue.push_back(pc + 1);
639                    queue.push_back(branch_target(pc, instr.imm));
640                    break;
641                }
642                Opcode::Call => {
643                    let callee = instr.imm as u32;
644                    if let Some(def) = chunk.function(callee) {
645                        queue.push_back(def.entry);
646                    }
647                    queue.push_back(pc + 1);
648                    break;
649                }
650                _ => pc += 1,
651            }
652        }
653
654        if !pcs.is_empty() {
655            blocks.push(TraceBlock { start_pc: start, pcs });
656        }
657    }
658
659    Ok(TraceRegion { blocks })
660}
661
662fn is_effect_opcode(op: Opcode) -> bool {
663    matches!(
664        op,
665        Opcode::CallNative
666            | Opcode::Spawn
667            | Opcode::Yield
668            | Opcode::Sleep
669            | Opcode::Exit
670            | Opcode::SelfPid
671            | Opcode::Send
672            | Opcode::Receive
673            | Opcode::ReceiveTimeout
674            | Opcode::ReceiveMatch
675            | Opcode::ReceiveMatchImm
676            | Opcode::Ask
677            | Opcode::AskTimeout
678            | Opcode::Monitor
679            | Opcode::Demonitor
680            | Opcode::Link
681            | Opcode::Unlink
682            | Opcode::Delegate
683            | Opcode::Trap
684            | Opcode::Halt
685    )
686}
687
688fn field_ptr(
689    builder: &mut FunctionBuilder<'_>,
690    base: ClifValue,
691    pointer_type: types::Type,
692    offset: i32,
693) -> ClifValue {
694    let off = builder.ins().iconst(pointer_type, i64::from(offset));
695    builder.ins().iadd(base, off)
696}
697
698fn store_slot(
699    builder: &mut FunctionBuilder<'_>,
700    slots: ClifValue,
701    pointer_type: types::Type,
702    reg: u8,
703    value: ClifValue,
704) {
705    let offset = builder.ins().iconst(pointer_type, i64::from(u32::from(reg) * 8));
706    let addr = builder.ins().iadd(slots, offset);
707    builder.ins().store(MachMemFlags::new(), value, addr, 0);
708}
709
710fn bool_as_i64(builder: &mut FunctionBuilder<'_>, cond: ClifValue) -> ClifValue {
711    let one = builder.ins().iconst(types::I64, 1);
712    let zero = builder.ins().iconst(types::I64, 0);
713    builder.ins().select(cond, one, zero)
714}
715
716fn emit_budget_tick(
717    builder: &mut FunctionBuilder<'_>,
718    budget_ptr: ClifValue,
719    frame_ptr: ClifValue,
720    pointer_type: types::Type,
721    pc: u32,
722) {
723    let budget = builder.ins().load(types::I32, MachMemFlags::new(), budget_ptr, 0);
724    let zero = builder.ins().iconst(types::I32, 0);
725    let exhausted = builder.ins().icmp(
726        cranelift_codegen::ir::condcodes::IntCC::Equal,
727        budget,
728        zero,
729    );
730    let continue_insn = builder.create_block();
731    let budget_exit = builder.create_block();
732    builder.ins().brif(exhausted, budget_exit, &[], continue_insn, &[]);
733    builder.switch_to_block(budget_exit);
734    builder.seal_block(budget_exit);
735    write_exit(builder, frame_ptr, pointer_type, JIT_BUDGET, pc, 0);
736    builder.switch_to_block(continue_insn);
737    builder.seal_block(continue_insn);
738    let one = builder.ins().iconst(types::I32, 1);
739    let new_budget = builder.ins().isub(budget, one);
740    builder.ins().store(MachMemFlags::new(), new_budget, budget_ptr, 0);
741}
742
743fn write_u32_field(
744    builder: &mut FunctionBuilder<'_>,
745    frame_ptr: ClifValue,
746    pointer_type: types::Type,
747    offset: i32,
748    value: u32,
749) {
750    let addr = field_ptr(builder, frame_ptr, pointer_type, offset);
751    let val = builder.ins().iconst(types::I32, i64::from(value));
752    builder.ins().store(MachMemFlags::new(), val, addr, 0);
753}
754
755fn write_exit(
756    builder: &mut FunctionBuilder<'_>,
757    frame_ptr: ClifValue,
758    pointer_type: types::Type,
759    kind: u32,
760    pc: u32,
761    return_reg: u32,
762) {
763    write_u32_field(builder, frame_ptr, pointer_type, OFF_EXIT_KIND, kind);
764    write_u32_field(builder, frame_ptr, pointer_type, OFF_PC, pc);
765    write_u32_field(builder, frame_ptr, pointer_type, OFF_RETURN_REG, return_reg);
766    builder.ins().return_(&[]);
767}
768
769#[allow(clippy::too_many_arguments)]
770fn emit_call(
771    builder: &mut FunctionBuilder<'_>,
772    chunk: &Chunk,
773    registers: &mut RegisterMap,
774    slots_ptr: ClifValue,
775    frame_ptr: ClifValue,
776    pointer_type: types::Type,
777    pc_to_block: &HashMap<u32, Block>,
778    call_depth_ptr: ClifValue,
779    call_stack_ptr: ClifValue,
780    pc: u32,
781    instr: Instruction,
782) -> Result<(), CompileError> {
783    let callee = instr.imm as u32;
784    let argc = instr.b;
785    let dst = instr.a;
786    let def = chunk.function(callee).ok_or(CompileError::UnsupportedOpcode {
787        opcode: Opcode::Call,
788        pc,
789    })?;
790    let callee_entry = def.entry;
791    let Some(callee_block) = pc_to_block.get(&callee_entry) else {
792        flush_registers(builder, registers, slots_ptr, pointer_type);
793        write_exit(
794            builder,
795            frame_ptr,
796            pointer_type,
797            JIT_CONTINUE,
798            pc,
799            0,
800        );
801        return Ok(());
802    };
803
804    flush_registers(builder, registers, slots_ptr, pointer_type);
805
806    let depth = builder
807        .ins()
808        .load(types::I32, MachMemFlags::new(), call_depth_ptr, 0);
809    let max_depth = builder
810        .ins()
811        .iconst(types::I32, i64::from(MAX_JIT_CALL_DEPTH as u32));
812    let too_deep = builder.ins().icmp(
813        cranelift_codegen::ir::condcodes::IntCC::SignedGreaterThanOrEqual,
814        depth,
815        max_depth,
816    );
817    let effect_exit = builder.create_block();
818    let call_body = builder.create_block();
819    builder
820        .ins()
821        .brif(too_deep, effect_exit, &[], call_body, &[]);
822    builder.switch_to_block(effect_exit);
823    builder.seal_block(effect_exit);
824    write_exit(
825        builder,
826        frame_ptr,
827        pointer_type,
828        super::exit::JIT_EFFECT,
829        pc,
830        0,
831    );
832    builder.switch_to_block(call_body);
833    builder.seal_block(call_body);
834
835    let record_bytes = builder.ins().iconst(
836        types::I32,
837        i64::from(std::mem::size_of::<super::frame::JitCallRecord>() as u32),
838    );
839    let record_off = builder.ins().imul(depth, record_bytes);
840    let record_off_ptr = builder.ins().uextend(pointer_type, record_off);
841    let record_addr = builder.ins().iadd(call_stack_ptr, record_off_ptr);
842
843    let return_pc = builder.ins().iconst(types::I32, i64::from(pc + 1));
844    builder
845        .ins()
846        .store(MachMemFlags::new(), return_pc, record_addr, 0);
847
848    let func_ptr = field_ptr(builder, frame_ptr, pointer_type, OFF_FUNCTION);
849    let caller_fn = builder
850        .ins()
851        .load(types::I32, MachMemFlags::new(), func_ptr, 0);
852    let caller_fn_addr = {
853        let off = builder.ins().iconst(pointer_type, 4);
854        builder.ins().iadd(record_addr, off)
855    };
856    builder
857        .ins()
858        .store(MachMemFlags::new(), caller_fn, caller_fn_addr, 0);
859
860    let dest = builder.ins().iconst(types::I32, i64::from(u32::from(dst)));
861    let dest_addr = {
862        let off = builder.ins().iconst(pointer_type, 8);
863        builder.ins().iadd(record_addr, off)
864    };
865    builder
866        .ins()
867        .store(MachMemFlags::new(), dest, dest_addr, 0);
868
869    let reg_ptr = field_ptr(builder, frame_ptr, pointer_type, OFF_REGISTER_COUNT);
870    let caller_regs = builder
871        .ins()
872        .load(types::I32, MachMemFlags::new(), reg_ptr, 0);
873    let caller_regs_addr = {
874        let off = builder.ins().iconst(pointer_type, 12);
875        builder.ins().iadd(record_addr, off)
876    };
877    builder
878        .ins()
879        .store(MachMemFlags::new(), caller_regs, caller_regs_addr, 0);
880
881    let one = builder.ins().iconst(types::I32, 1);
882    let new_depth = builder.ins().iadd(depth, one);
883    builder
884        .ins()
885        .store(MachMemFlags::new(), new_depth, call_depth_ptr, 0);
886
887    for i in 0..argc {
888        let src_reg = u32::from(dst) + u32::from(i);
889        let src_offset = builder.ins().iconst(pointer_type, i64::from(src_reg * 8));
890        let src_addr = builder.ins().iadd(slots_ptr, src_offset);
891        let value = builder
892            .ins()
893            .load(types::I64, MachMemFlags::new(), src_addr, 0);
894        let dst_offset = builder.ins().iconst(pointer_type, i64::from(u32::from(i) * 8));
895        let dst_addr = builder.ins().iadd(slots_ptr, dst_offset);
896        builder
897            .ins()
898            .store(MachMemFlags::new(), value, dst_addr, 0);
899    }
900
901    let callee_fn = builder.ins().iconst(types::I32, i64::from(callee));
902    builder
903        .ins()
904        .store(MachMemFlags::new(), callee_fn, func_ptr, 0);
905    let callee_regs = builder.ins().iconst(types::I32, i64::from(def.num_registers));
906    builder
907        .ins()
908        .store(MachMemFlags::new(), callee_regs, reg_ptr, 0);
909
910    registers.clear();
911    builder.ins().jump(*callee_block, &[]);
912    Ok(())
913}
914
915#[allow(clippy::too_many_arguments)]
916fn emit_return(
917    builder: &mut FunctionBuilder<'_>,
918    frame_ptr: ClifValue,
919    pointer_type: types::Type,
920    call_depth_ptr: ClifValue,
921    call_stack_ptr: ClifValue,
922    slots_ptr: ClifValue,
923    registers: &mut RegisterMap,
924    return_reg: u8,
925    pc: u32,
926) {
927    let depth = builder
928        .ins()
929        .load(types::I32, MachMemFlags::new(), call_depth_ptr, 0);
930    let zero = builder.ins().iconst(types::I32, 0);
931    let is_outer = builder.ins().icmp(
932        cranelift_codegen::ir::condcodes::IntCC::Equal,
933        depth,
934        zero,
935    );
936    let outer = builder.create_block();
937    let inner = builder.create_block();
938    builder
939        .ins()
940        .brif(is_outer, outer, &[], inner, &[]);
941    builder.switch_to_block(outer);
942    builder.seal_block(outer);
943    write_exit(
944        builder,
945        frame_ptr,
946        pointer_type,
947        JIT_RETURN,
948        pc,
949        u32::from(return_reg),
950    );
951    builder.switch_to_block(inner);
952    builder.seal_block(inner);
953
954    let one = builder.ins().iconst(types::I32, 1);
955    let new_depth = builder.ins().isub(depth, one);
956    builder
957        .ins()
958        .store(MachMemFlags::new(), new_depth, call_depth_ptr, 0);
959
960    let record_bytes = builder.ins().iconst(
961        types::I32,
962        i64::from(std::mem::size_of::<super::frame::JitCallRecord>() as u32),
963    );
964    let record_off = builder.ins().imul(new_depth, record_bytes);
965    let record_off_ptr = builder.ins().uextend(pointer_type, record_off);
966    let record_addr = builder.ins().iadd(call_stack_ptr, record_off_ptr);
967
968    let return_pc = builder
969        .ins()
970        .load(types::I32, MachMemFlags::new(), record_addr, 0);
971    let return_fn_addr = {
972        let off = builder.ins().iconst(pointer_type, 4);
973        builder.ins().iadd(record_addr, off)
974    };
975    let return_fn = builder
976        .ins()
977        .load(types::I32, MachMemFlags::new(), return_fn_addr, 0);
978    let dest_addr = {
979        let off = builder.ins().iconst(pointer_type, 8);
980        builder.ins().iadd(record_addr, off)
981    };
982    let dest_reg = builder
983        .ins()
984        .load(types::I32, MachMemFlags::new(), dest_addr, 0);
985    let caller_regs_addr = {
986        let off = builder.ins().iconst(pointer_type, 12);
987        builder.ins().iadd(record_addr, off)
988    };
989    let caller_regs = builder
990        .ins()
991        .load(types::I32, MachMemFlags::new(), caller_regs_addr, 0);
992
993    let ret_offset = builder.ins().iconst(pointer_type, i64::from(u32::from(return_reg) * 8));
994    let ret_addr = builder.ins().iadd(slots_ptr, ret_offset);
995    let ret_val = builder
996        .ins()
997        .load(types::I64, MachMemFlags::new(), ret_addr, 0);
998    let eight = builder.ins().iconst(types::I32, 8);
999    let dest_offset = builder.ins().imul(dest_reg, eight);
1000    let dest_offset_ptr = builder.ins().uextend(pointer_type, dest_offset);
1001    let dest_slot = builder.ins().iadd(slots_ptr, dest_offset_ptr);
1002    builder
1003        .ins()
1004        .store(MachMemFlags::new(), ret_val, dest_slot, 0);
1005
1006    let func_ptr = field_ptr(builder, frame_ptr, pointer_type, OFF_FUNCTION);
1007    builder
1008        .ins()
1009        .store(MachMemFlags::new(), return_fn, func_ptr, 0);
1010    let reg_ptr = field_ptr(builder, frame_ptr, pointer_type, OFF_REGISTER_COUNT);
1011    builder
1012        .ins()
1013        .store(MachMemFlags::new(), caller_regs, reg_ptr, 0);
1014
1015    registers.clear();
1016    write_dynamic_exit(
1017        builder,
1018        frame_ptr,
1019        pointer_type,
1020        JIT_CONTINUE,
1021        return_pc,
1022        0,
1023    );
1024    let _ = pc;
1025}
1026
1027fn write_dynamic_exit(
1028    builder: &mut FunctionBuilder<'_>,
1029    frame_ptr: ClifValue,
1030    pointer_type: types::Type,
1031    kind: u32,
1032    pc: ClifValue,
1033    return_reg: u32,
1034) {
1035    write_u32_field(builder, frame_ptr, pointer_type, OFF_EXIT_KIND, kind);
1036    let pc_ptr = field_ptr(builder, frame_ptr, pointer_type, OFF_PC);
1037    builder.ins().store(MachMemFlags::new(), pc, pc_ptr, 0);
1038    write_u32_field(builder, frame_ptr, pointer_type, OFF_RETURN_REG, return_reg);
1039    builder.ins().return_(&[]);
1040}
1041
1042#[cfg(test)]
1043mod tests {
1044    use std::sync::Arc;
1045
1046    use crate::{bytecode::builder::ChunkBuilder, NativeTable, Opcode, Vm};
1047
1048    use super::*;
1049    use crate::jit::dispatch::force_compile;
1050    use crate::jit::trace::{JitContext, TraceKey};
1051
1052    type TestResult = Result<(), Box<dyn std::error::Error>>;
1053
1054    #[test]
1055    fn compiles_load_imm_add_return() -> TestResult {
1056        let mut b = ChunkBuilder::new("jit");
1057        b.begin_function("main", 0, 3);
1058        b.emit_load_imm(0, 41);
1059        b.emit_load_imm(1, 1);
1060        b.emit_binop(Opcode::Add, 2, 0, 1);
1061        b.emit_return(2);
1062        let chunk = Arc::new(b.finish());
1063
1064        let mut ctx = JitContext::new(chunk.clone())?;
1065        let key = TraceKey {
1066            function: 0,
1067            entry_pc: 0,
1068        };
1069        let mut compiler = TraceCompiler::new(ctx.module_mut());
1070        let trace = compiler.compile_trace(&chunk, key)?;
1071        assert!(trace.span.range.contains(&0));
1072        Ok(())
1073    }
1074
1075    #[test]
1076    fn compiles_branch_falsy_taken() -> TestResult {
1077        let mut b = ChunkBuilder::new("jit-branch");
1078        b.begin_function("main", 0, 1);
1079        let done = b.new_label();
1080        let ret = b.new_label();
1081        b.emit_load_imm(0, 0);
1082        b.emit_branch(0, done);
1083        b.emit_load_imm(0, 99);
1084        b.emit_jump(ret);
1085        b.bind_label(done);
1086        b.emit_load_imm(0, 7);
1087        b.bind_label(ret);
1088        b.emit_return(0);
1089        let chunk = Arc::new(b.finish());
1090
1091        let mut ctx = JitContext::new(chunk.clone())?;
1092        let key = TraceKey {
1093            function: 0,
1094            entry_pc: 0,
1095        };
1096        let mut compiler = TraceCompiler::new(ctx.module_mut());
1097        let trace = compiler.compile_trace(&chunk, key)?;
1098        let mut slots = vec![0i64; 1];
1099        let ret = super::super::dispatch::run_compiled_trace_ref(
1100            &trace,
1101            &mut slots,
1102            0,
1103            10_000,
1104            0,
1105            1,
1106        );
1107        assert_eq!(slots[0], 7);
1108        assert!(matches!(
1109            ret.into_reason(),
1110            super::super::exit::ExitReason::Return { .. }
1111        ));
1112        Ok(())
1113    }
1114
1115    #[test]
1116    fn compiles_unconditional_jump() -> TestResult {
1117        let mut b = ChunkBuilder::new("jit-jump");
1118        b.begin_function("main", 0, 1);
1119        let skip = b.new_label();
1120        let ret = b.new_label();
1121        b.emit_load_imm(0, 1);
1122        b.emit_jump(skip);
1123        b.emit_load_imm(0, 99);
1124        b.bind_label(skip);
1125        b.emit_load_imm(0, 41);
1126        b.bind_label(ret);
1127        b.emit_return(0);
1128        let chunk = Arc::new(b.finish());
1129
1130        let mut ctx = JitContext::new(chunk.clone())?;
1131        let key = TraceKey {
1132            function: 0,
1133            entry_pc: 0,
1134        };
1135        let mut compiler = TraceCompiler::new(ctx.module_mut());
1136        let trace = compiler.compile_trace(&chunk, key)?;
1137        let mut slots = vec![0i64; 1];
1138        let ret = super::super::dispatch::run_compiled_trace_ref(
1139            &trace,
1140            &mut slots,
1141            0,
1142            10_000,
1143            0,
1144            1,
1145        );
1146        assert!(matches!(
1147            ret.into_reason(),
1148            super::super::exit::ExitReason::Return { return_reg: 0 }
1149        ));
1150        assert_eq!(slots[0], 41);
1151        Ok(())
1152    }
1153
1154    #[test]
1155    fn compiles_intra_chunk_call() -> TestResult {
1156        let mut b = ChunkBuilder::new("jit-call");
1157        b.begin_function("double", 0, 1);
1158        b.emit_load_imm(0, 21);
1159        b.emit_return(0);
1160        b.begin_function("main", 1, 1);
1161        b.emit_load_imm(0, 0);
1162        b.emit_call(0, 0, 1);
1163        b.emit_return(0);
1164        let chunk = Arc::new(b.finish());
1165
1166        let mut ctx = JitContext::new(chunk.clone())?;
1167        let key = TraceKey {
1168            function: 1,
1169            entry_pc: chunk.functions[1].entry,
1170        };
1171        force_compile(&mut ctx, &chunk, key)?;
1172        let mut vm = Vm::new(chunk, NativeTable::empty(), key.function, &[])?;
1173        let result = super::super::dispatch::run_vm_with_jit(&mut vm, 10_000, &mut ctx);
1174        assert!(matches!(
1175            result,
1176            crate::VmResult::Complete(crate::Value::Int(21))
1177        ));
1178        Ok(())
1179    }
1180}