use std::sync::{Arc, OnceLock};

use oxide_bytecode::module::CompiledModule;

use crate::vm::{CallFrame, InlineSyncState, Vm};
use crate::{vm_debug, vm_info, vm_trace, vm_warn};
use oxide_types::object::JsObject;
use oxide_types::value::JsValue;

/// 按 `flat_id` 下标收集整棵子模块树为平表，供 `run()` 装载。
fn collect_flat_modules(module: &CompiledModule) -> Vec<CompiledModule> {
    let mut out: Vec<Option<CompiledModule>> = Vec::new();
    place_flat(module, &mut out);
    out.into_iter().map(|m| m.expect("flat_id slot must be filled")).collect()
}

fn place_flat(module: &CompiledModule, out: &mut Vec<Option<CompiledModule>>) {
    let idx = module.flat_id as usize;
    if idx >= out.len() {
        out.resize(idx + 1, None);
    }
    out[idx] = Some(module.clone());
    for sub in &module.sub_modules {
        place_flat(sub, out);
    }
}

impl Vm {
    /// 保存当前完整 VM 执行状态到堆上（内联同步调用与生成器恢复共用）。
    pub(crate) fn save_inline_state(&mut self) -> Box<InlineSyncState> {
        vm_trace!("save_inline_state: pc={} depth={}", self.pc, self.frames.len());
        Box::new(InlineSyncState {
            regs: Box::new(self.regs),
            pc: self.pc,
            bytecode: std::mem::take(&mut self.bytecode),
            active_immutables: self.active_immutables,
            active_reg_limit: self.active_reg_limit,
            root_reg_limit: self.root_reg_limit,
            try_stack: std::mem::take(&mut self.try_stack),
            frames: std::mem::take(&mut self.frames),
            exception_value: self.exception_value.take(),
            pending_exception: self.pending_exception.take(),
            pending_error_kind: self.pending_error_kind.take(),
            pending_completion: self.pending_completion,
            for_in_iters: std::mem::take(&mut self.iters.for_in_iters),
            for_of_iters: std::mem::take(&mut self.iters.for_of_iters),
            last_for_of_result: self.iters.last_for_of_result,
            saved_bytecode_stack: std::mem::take(&mut self.saved_bytecode_stack),
            saved_immutables_stack: std::mem::take(&mut self.saved_immutables_stack),
            save_stack: std::mem::take(&mut self.save_stack),
            spill_stack: std::mem::take(&mut self.spill_stack),
            cell_stack: std::mem::take(&mut self.cell_stack),
            inline_callee: self.inline_callee,
            inline_args_base: self.inline_args_base,
            inline_args_count: self.inline_args_count,
        })
    }

    /// 把 [`save_inline_state`] 保存的状态完整恢复回 VM。
    pub(crate) fn restore_inline_state(&mut self, saved: Box<InlineSyncState>) {
        vm_trace!("restore_inline_state: pc={}", saved.pc);
        self.regs = *saved.regs;
        self.pc = saved.pc;
        self.bytecode = saved.bytecode;
        self.active_immutables = saved.active_immutables;
        self.active_reg_limit = saved.active_reg_limit;
        self.root_reg_limit = saved.root_reg_limit;
        self.try_stack = saved.try_stack;
        self.frames = saved.frames;
        self.exception_value = saved.exception_value;
        self.pending_exception = saved.pending_exception;
        self.pending_error_kind = saved.pending_error_kind;
        self.pending_completion = saved.pending_completion;
        self.iters.for_in_iters = saved.for_in_iters;
        self.iters.for_of_iters = saved.for_of_iters;
        self.iters.last_for_of_result = saved.last_for_of_result;
        self.saved_bytecode_stack = saved.saved_bytecode_stack;
        self.saved_immutables_stack = saved.saved_immutables_stack;
        self.save_stack = saved.save_stack;
        self.spill_stack = saved.spill_stack;
        self.cell_stack = saved.cell_stack;
        self.inline_callee = saved.inline_callee;
        self.inline_args_base = saved.inline_args_base;
        self.inline_args_count = saved.inline_args_count;
    }

    pub(crate) fn call_bytecode_function_inline(
        &mut self, callee: JsValue, callee_obj: &JsObject, receiver: JsValue, args: &[JsValue],
    ) -> Result<JsValue, String> {
        if callee_obj.sub_module_index() == 0 {
            return Err(self.error_message_text("TypeError", "accessor is not callable"));
        }
        let sub_idx = callee_obj.sub_module_index() as usize;
        vm_debug!(
            "call_bytecode_function_inline: sub_idx={}, args={}, depth={}",
            sub_idx,
            args.len(),
            self.frames.len()
        );
        if sub_idx >= self.sub_modules.len() {
            return Err(format!(
                "accessor sub_module_index {} out of bounds (max {})",
                sub_idx,
                self.sub_modules.len()
            ));
        }
        if self.frames.len() >= self.kernel_core.config.max_call_depth {
            self.raise_error_kind("RangeError", "Maximum call stack size exceeded")?;
            return Ok(JsValue::undefined());
        }
        if self.native_call_depth >= self.kernel_core.config.max_call_depth {
            self.raise_error_kind("RangeError", "Maximum call stack size exceeded")?;
            return Ok(JsValue::undefined());
        }
        // 异步生成器函数调用返回异步生成器迭代器对象，next/return/throw 返回 Promise。
        if self.sub_modules[sub_idx].is_generator && self.sub_modules[sub_idx].is_async {
            return self.create_async_generator_object(callee, receiver, args);
        }
        // 生成器函数调用返回迭代器对象，不执行函数体。
        if self.sub_modules[sub_idx].is_generator {
            return self.create_generator_object(callee, receiver, args);
        }
        // 异步函数调用返回 capability promise，立即同步执行 body 到首个 await。
        if self.sub_modules[sub_idx].is_async {
            return self.create_async_object(callee, receiver, args);
        }
        self.native_call_depth += 1;

        let subs = Arc::clone(&self.sub_modules);
        let sub = &subs[sub_idx];

        vm_trace!("call_bytecode: saving state pc={} depth={}", self.pc, self.frames.len());
        let saved = self.save_inline_state();

        self.regs = [JsValue::undefined(); 256];
        self.pc = 0;
        self.bytecode = sub.bytecode.clone();
        self.activate_immutables(sub_idx, &sub.constants);
        self.active_reg_limit = sub.n_registers.max(1);
        self.root_reg_limit = self.active_reg_limit;
        self.cell_stack.push(Vec::with_capacity(sub.cells_needed as usize));
        self.inline_callee = Some(callee);
        // inline 同步调用无 CallFrame：完整实参写入 spill 栈实参区，供 CREATE_ARGUMENTS 读取。
        self.inline_args_base = self.spill_stack.len() as u32;
        self.spill_stack.extend_from_slice(args);
        self.inline_args_count = args.len().min(u16::MAX as usize) as u16;
        for i in 0..sub.n_args as usize {
            self.regs[sub.param_base as usize + i] = args.get(i).copied().unwrap_or(JsValue::undefined());
        }
        self.regs[254] = if sub.is_arrow { callee_obj.captured_this() } else { receiver };
        self.regs[255] = JsValue::undefined();
        for (name, reg) in &sub.builtin_reg_map {
            let si = self.kernel_core.perm_interner().intern(name.as_str()).0;
            let global = self.session.global_object();
            if let Some(pos) = self.kernel_core.shape_forge().lookup_position(global.shape_id(), si) {
                self.regs[*reg as usize] = global.get_prop_at(pos);
            }
        }
        let _ = callee;

        vm_trace!(
            "call_bytecode: dispatching sub_idx={} n_regs={} n_args={} arrow={}",
            sub_idx,
            sub.n_registers,
            sub.n_args,
            sub.is_arrow
        );
        let result = self.dispatch();
        self.native_call_depth -= 1;

        vm_trace!("call_bytecode: restoring state pc={} result={:?}", saved.pc, result.as_ref().ok());
        self.restore_inline_state(saved);
        result
    }

    pub(crate) fn restore_frame(&mut self, frame: CallFrame) {
        vm_trace!(
            "restore_frame: return_addr={} caller_reg_limit={}",
            frame.return_addr,
            frame.caller_reg_limit
        );
        if let Some(saved_bc) = self.saved_bytecode_stack.pop() {
            self.bytecode = saved_bc;
        }
        vm_debug!(
            "restore_frame: return_addr={} bc_len={} saved_stack={} fn={:?}",
            frame.return_addr,
            self.bytecode.len(),
            self.saved_bytecode_stack.len(),
            self.kernel_core
                .perm_interner()
                .lookup(frame.function_name)
                .map(|s| s.to_string())
        );
        if let Some(saved_imm) = self.saved_immutables_stack.pop() {
            self.active_immutables = saved_imm;
        }
        let offset = frame.saved_reg_offset as usize;
        let len = frame.caller_reg_limit as usize;
        self.regs[..len].copy_from_slice(&self.save_stack[offset..offset + len]);
        self.save_stack.truncate(offset);
        // spill 帧边界截断：子函数写入的 spill 数据在帧恢复后丢弃。
        self.spill_stack.truncate(frame.spill_offset as usize);
        self.regs[254] = frame.saved_this;
        self.regs[255] = frame.saved_new_target;
        self.active_reg_limit = frame.caller_reg_limit;
        self.pc = frame.return_addr;
    }

    /// 重新执行当前已加载的 bytecode：清空执行状态并重置 IC 缓存后再次 dispatch。
    pub fn rerun(&mut self) -> Result<JsValue, String> {
        vm_info!("rerun: clearing IC caches");
        self.clear_execution_state();
        self.active_reg_limit = self.root_reg_limit;
        crate::ic_helper::clear_ic_caches(&mut self.bytecode);
        self.dispatch()
    }

    /// 加载并执行一个已编译模块，返回模块顶层执行结果或未捕获异常消息。
    ///
    /// 内部初始化寄存器/bytecode/immutables 与 builtin 寄存器预绑定，然后进入
    /// dispatch 主循环；执行完成或异常展开后返回。
    pub fn run(&mut self, module: &CompiledModule) -> Result<JsValue, String> {
        vm_debug!("run: starting bytecode execution, {} instructions", module.bytecode.len());
        self.clear_execution_state();
        self.cell_stack.clear();
        self.cell_stack.push(Vec::new());
        self.sub_modules = Arc::new(collect_flat_modules(module));
        self.immutables_cache = (0..self.sub_modules.len()).map(|_| OnceLock::new()).collect();
        self.bytecode = module.bytecode.clone();
        self.activate_immutables(0, &module.constants);
        self.root_reg_limit = module.n_registers.max(1);
        self.active_reg_limit = self.root_reg_limit;

        for (name, reg) in &module.builtin_reg_map {
            let si = self.kernel_core.perm_interner().intern(name.as_str()).0;
            let global = self.session.global_object();
            if let Some(pos) = self.kernel_core.shape_forge().lookup_position(global.shape_id(), si) {
                self.regs[*reg as usize] = global.get_prop_at(pos);
            }
        }

        // 全局代码的 this 绑定为全局对象（ECMA-262 全局执行上下文）。
        let global = self.session.global_object();
        self.regs[254] = JsValue::from_js_object(global.as_ptr() as *mut JsObject);

        let result = self.dispatch();
        // 顶层执行结束后 drain 微任务队列：Promise reactions 与 thenable 委托在此执行。
        self.drain_job_queue();
        result
    }

    pub(crate) fn unwind(&mut self) -> Result<(), String> {
        vm_debug!("unwind: {} try handlers on stack", self.try_stack.len());
        while let Some(mut handler) = self.try_stack.pop() {
            if handler.frame_depth > self.frames.len() {
                // 残留 handler 所属帧已返回（return 逃出泄漏的兜底）：丢弃防跳回
                // 死函数的 catch/finally——否则 catch 再抛会反复弹出同一 handler
                // 形成死循环。其 for_of_depth 快照属于已死帧，也不可据此关闭
                // 当前帧的迭代器，故整体跳过。
                continue;
            }
            while self.frames.len() > handler.frame_depth {
                if let Some(frame) = self.frames.pop() {
                    self.cell_stack.pop();
                    self.restore_frame(frame);
                }
            }
            // 对该处理器作用域内被中断的 for-of 循环执行 IteratorClose。
            self.close_for_of_above(handler.for_of_depth);
            if let Some(finally_pc) = handler.finally_pc {
                if handler.finally_active {
                    // 新异常在 finally 体内抛出：覆盖在途异常与完成，继续向外展开，
                    // 不再重入本 finally（否则已执行的 finally 会重复运行）。
                    self.pending_exception = None;
                    self.pending_error_kind = None;
                    self.pending_completion = None;
                    continue;
                }
                vm_trace!("unwind: entering finally at pc={}", finally_pc);
                self.pending_exception = Some(self.exception_value.take().unwrap_or(JsValue::undefined()));
                handler.finally_active = true;
                self.try_stack.push(handler);
                self.pc = finally_pc;
                return Ok(());
            }
            if let Some(catch_pc) = handler.catch_pc {
                vm_trace!("unwind: caught at pc={}", catch_pc);
                let exc = self.exception_value.take().unwrap_or(JsValue::undefined());
                self.regs[0] = exc;
                self.pc = catch_pc;
                return Ok(());
            }
        }
        self.close_for_of_above(0);
        while let Some(frame) = self.frames.pop() {
            self.cell_stack.pop();
            self.restore_frame(frame);
        }
        let exc = self.exception_value.take().unwrap_or(JsValue::undefined());
        // 保留逃逸的 JsValue，使 for-of 的 next() 抛出时可重新抛出原值而非展平后的
        // 字符串（由 dispatch_for_of_done/next 消费）。
        self.last_uncaught_value = Some(exc);
        let kind_str = self.pending_error_kind.take().unwrap_or("Error");
        let exc_text = self.error_text(exc);
        let msg = if exc_text.starts_with(kind_str) {
            format!("uncaught {exc_text}")
        } else {
            format!("uncaught {kind_str}: {exc_text}")
        };
        vm_warn!("unwind: uncaught {}: {}", kind_str, exc_text);
        Err(msg)
    }
}