perf: 三处手写压帧收敛单一入口并免实参堆 Vec 物化

This commit is contained in:
2026-08-15 10:27:14 +08:00
parent 2239eeac24
commit 8bdcb43aba
7 changed files with 120 additions and 186 deletions
+20 -77
View File
@@ -1,5 +1,5 @@
use crate::native::NativeFn; use crate::native::NativeFn;
use crate::vm::{native_fn_ptr_to_fn, CallFrame, FrameContinuation, Vm}; use crate::vm::{native_fn_ptr_to_fn, FrameArgs, FrameContinuation, Vm};
use crate::{vm_debug, vm_trace}; use crate::{vm_debug, vm_trace};
use oxide_builtins::iterator::make_iterator_for_value; use oxide_builtins::iterator::make_iterator_for_value;
use oxide_builtins::{builtins_debug, builtins_trace}; use oxide_builtins::{builtins_debug, builtins_trace};
@@ -7,7 +7,6 @@ use oxide_bytecode::opcode;
use oxide_runtime_api::{to_boolean, NativeResult}; use oxide_runtime_api::{to_boolean, NativeResult};
use oxide_types::object::{Cell, JsObject, PropAttributes}; use oxide_types::object::{Cell, JsObject, PropAttributes};
use oxide_types::value::JsValue; use oxide_types::value::JsValue;
use std::sync::Arc;
impl Vm { impl Vm {
#[inline(always)] #[inline(always)]
@@ -90,7 +89,7 @@ impl Vm {
self.push_bytecode_frame( self.push_bytecode_frame(
callee, callee,
this, this,
&args, FrameArgs::Slice(&args),
None, None,
None, None,
JsValue::undefined(), JsValue::undefined(),
@@ -486,77 +485,21 @@ impl Vm {
self.sub_modules.len() self.sub_modules.len()
)); ));
} }
if self.frames.len() >= self.kernel_core.config.max_call_depth { // 收敛到统一压帧入口:this = derived_this(super() 把实例交予父构造器),
return Err(self.error_message_text("RangeError", "Maximum call stack size exceeded")); // new.target 保持外层类;构造结果写回 regs[254]。
} self.push_bytecode_frame(
super_ctor,
let sub_bytecode = Arc::clone(&self.sub_modules[sub_idx].bytecode); derived_this,
let sub_n_args = self.sub_modules[sub_idx].n_args as usize; FrameArgs::RegRange {
let sub_n_registers = self.sub_modules[sub_idx].n_registers; first: first_arg_reg,
let sub_param_base = self.sub_modules[sub_idx].param_base as usize; count: arg_count,
let caller_active_reg_limit = self.active_reg_limit.max(1); },
let caller_reg_limit = self.call_window_limit(caller_active_reg_limit, call_window); Some(254),
let saved_reg_offset = self.save_stack.len() as u32; Some(derived_this),
self.save_stack.extend_from_slice(&self.regs[..caller_reg_limit as usize]); new_target,
let saved_this = self.regs[254]; FrameContinuation::None,
let saved_new_target = self.regs[255]; call_window,
)?;
// 完整实参写入 spill 栈实参区(在帧的 spill 区之前),供 CREATE_ARGUMENTS 使用。
let args_base = self.spill_stack.len() as u32;
for i in 0..arg_count {
let src_reg = first_arg_reg.wrapping_add(i as u8) as usize;
self.spill_stack.push(self.regs[src_reg]);
}
let args_count = arg_count.min(u16::MAX as usize) as u16;
for i in 0..sub_n_args {
let src_reg = first_arg_reg.wrapping_add(i as u8) as usize;
self.regs[sub_param_base + i] = self.regs[src_reg];
}
self.regs[254] = derived_this;
self.regs[255] = new_target;
self.saved_bytecode_stack.push(std::mem::take(&mut self.bytecode));
self.saved_immutables_stack.push(self.active_immutables);
let function_name = self.sub_modules[sub_idx]
.function_name
.as_deref()
.map(|name| self.kernel_core.perm_interner().intern(name).0)
.unwrap_or(0);
self.frames.push(CallFrame {
return_addr: self.pc,
function_name,
caller_reg_limit,
caller_active_reg_limit,
saved_reg_offset,
spill_offset: self.spill_stack.len() as u32,
arguments_base: args_base,
arguments_count: args_count,
saved_this,
saved_new_target,
callee: super_ctor,
construct_result_reg: Some(254),
constructed_this: Some(derived_this),
is_derived_constructor: super_obj.is_derived_constructor(),
continuation: FrameContinuation::None,
});
self.bytecode = sub_bytecode;
let subs = Arc::clone(&self.sub_modules);
self.activate_immutables(sub_idx, &subs[sub_idx].constants);
self.cell_stack.push(Vec::with_capacity(subs[sub_idx].cells_needed as usize));
for (name, reg) in &self.sub_modules[sub_idx].builtin_reg_map.clone() {
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);
}
}
self.active_reg_limit = sub_n_registers.max(1);
self.pc = 0;
return Ok(true); return Ok(true);
} else { } else {
self.raise_error_kind("TypeError", "super constructor is not callable")?; self.raise_error_kind("TypeError", "super constructor is not callable")?;
@@ -794,7 +737,7 @@ impl Vm {
self.push_bytecode_frame( self.push_bytecode_frame(
callee, callee,
this_value, this_value,
&args, FrameArgs::Slice(&args),
None, None,
None, None,
JsValue::undefined(), JsValue::undefined(),
@@ -901,7 +844,7 @@ impl Vm {
self.push_bytecode_frame( self.push_bytecode_frame(
constructor, constructor,
this_value, this_value,
&args, FrameArgs::Slice(&args),
Some(rd as u8), Some(rd as u8),
Some(new_obj_val), Some(new_obj_val),
constructor, constructor,
@@ -989,7 +932,7 @@ impl Vm {
self.push_bytecode_frame( self.push_bytecode_frame(
super_ctor, super_ctor,
derived_this, derived_this,
&args, FrameArgs::Slice(&args),
Some(254), Some(254),
Some(derived_this), Some(derived_this),
new_target, new_target,
+3 -3
View File
@@ -194,8 +194,8 @@ impl Vm {
/// 附 length / callee 属性。第一版为 unmapped(非严格)语义,索引与形参不同步。 /// 附 length / callee 属性。第一版为 unmapped(非严格)语义,索引与形参不同步。
/// ///
/// # 边界与前提 /// # 边界与前提
/// - 实参区由 `push_bytecode_frame` / `dispatch_super_call` / `dispatch_new_expression` /// - 实参区由统一压帧入口 `push_bytecode_frame` 在推帧时写入 spill 栈;
/// 在推帧时写入 spill 栈;frames 为空(inline 同步调用)时读 `inline_args_*`。 /// frames 为空(inline 同步调用)时读 `inline_args_*`。
/// - 索引属性用 shape 槽存储(普通对象),length/callee 为不可枚举数据属性。 /// - 索引属性用 shape 槽存储(普通对象),length/callee 为不可枚举数据属性。
pub(crate) fn dispatch_create_arguments(&mut self, rd: usize) -> Result<(), String> { pub(crate) fn dispatch_create_arguments(&mut self, rd: usize) -> Result<(), String> {
let (base, count) = match self.frames.last() { let (base, count) = match self.frames.last() {
@@ -244,7 +244,7 @@ impl Vm {
/// 2. 实参个数超出固定形参数的部分作为数组元素写入。 /// 2. 实参个数超出固定形参数的部分作为数组元素写入。
/// ///
/// # 边界与前提 /// # 边界与前提
/// - 实参来源与 CREATE_ARGUMENTS 相同(push_bytecode_frame / inline 路径写入 spill 栈)。 /// - 实参来源与 CREATE_ARGUMENTS 相同(统一压帧入口 `push_bytecode_frame` 写入 spill 栈)。
/// - 实参不足 `fixed_count` 时数组为空。 /// - 实参不足 `fixed_count` 时数组为空。
pub(crate) fn dispatch_create_rest_array(&mut self, rd: usize, fixed_count: usize) -> Result<(), String> { pub(crate) fn dispatch_create_rest_array(&mut self, rd: usize, fixed_count: usize) -> Result<(), String> {
let (base, count) = match self.frames.last() { let (base, count) = match self.frames.last() {
+46 -3
View File
@@ -163,6 +163,42 @@ pub enum FrameContinuation {
AccessorSet, AccessorSet,
} }
/// 压帧实参来源:已物化切片或寄存器连续区间。
///
/// `RegRange` 直接引用调用方寄存器文件(CALL/NEW/SUPER_CALL 的实参在寄存器中
/// 天然连续排列),免去临时堆 `Vec<JsValue>` 物化,每次调用省 1 次分配/释放。
#[derive(Clone, Copy)]
pub(crate) enum FrameArgs<'a> {
/// 已物化实参:spread 展开、TailCall、accessor 与初始执行路径。
Slice(&'a [JsValue]),
/// 寄存器连续区间 `regs[first .. first+count)`。
RegRange { first: u8, count: usize },
}
impl FrameArgs<'_> {
/// 实参总数。
pub(crate) fn len(&self) -> usize {
match self {
FrameArgs::Slice(s) => s.len(),
FrameArgs::RegRange { count, .. } => *count,
}
}
/// 取第 `i` 个实参;超出实参范围返回 undefined(与切片 `get(i)` 语义一致)。
pub(crate) fn get(&self, vm: &Vm, i: usize) -> JsValue {
match self {
FrameArgs::Slice(s) => s.get(i).copied().unwrap_or(JsValue::undefined()),
FrameArgs::RegRange { first, count } => {
if i < *count {
vm.regs[first.wrapping_add(i as u8) as usize]
} else {
JsValue::undefined()
}
}
}
}
}
/// 一次函数调用的调用帧:记录返回地址、调用方寄存器窗口与 `this`/`new.target`。 /// 一次函数调用的调用帧:记录返回地址、调用方寄存器窗口与 `this`/`new.target`。
/// ///
/// 调用方寄存器窗口在 `save_stack` 中按 `saved_reg_offset` 保存,返回时由 /// 调用方寄存器窗口在 `save_stack` 中按 `saved_reg_offset` 保存,返回时由
@@ -1161,7 +1197,7 @@ impl Vm {
#[expect(clippy::too_many_arguments)] #[expect(clippy::too_many_arguments)]
pub(crate) fn push_bytecode_frame( pub(crate) fn push_bytecode_frame(
&mut self, callee: JsValue, this_value: JsValue, args: &[JsValue], construct_result_reg: Option<u8>, &mut self, callee: JsValue, this_value: JsValue, args: FrameArgs, construct_result_reg: Option<u8>,
constructed_this: Option<JsValue>, new_target: JsValue, continuation: FrameContinuation, call_window: u8, constructed_this: Option<JsValue>, new_target: JsValue, continuation: FrameContinuation, call_window: u8,
) -> Result<(), String> { ) -> Result<(), String> {
vm_trace!( vm_trace!(
@@ -1205,7 +1241,7 @@ impl Vm {
let saved_new_target = self.regs[255]; let saved_new_target = self.regs[255];
for i in 0..sub_n_args { for i in 0..sub_n_args {
self.regs[sub_param_base + i] = args.get(i).copied().unwrap_or(JsValue::undefined()); self.regs[sub_param_base + i] = args.get(self, i);
} }
self.regs[254] = if sub_is_arrow { obj.captured_this() } else { this_value }; self.regs[254] = if sub_is_arrow { obj.captured_this() } else { this_value };
self.regs[255] = new_target; self.regs[255] = new_target;
@@ -1222,7 +1258,14 @@ impl Vm {
// 完整实参写入 spill 栈实参区(在帧的 spill 区之前):CREATE_ARGUMENTS 据此 // 完整实参写入 spill 栈实参区(在帧的 spill 区之前):CREATE_ARGUMENTS 据此
// 构建 arguments 对象,帧恢复时随 spill 区截断一起丢弃。 // 构建 arguments 对象,帧恢复时随 spill 区截断一起丢弃。
let args_base = self.spill_stack.len() as u32; let args_base = self.spill_stack.len() as u32;
self.spill_stack.extend_from_slice(args); match args {
FrameArgs::Slice(s) => self.spill_stack.extend_from_slice(s),
FrameArgs::RegRange { first, count } => {
for i in 0..count {
self.spill_stack.push(self.regs[first.wrapping_add(i as u8) as usize]);
}
}
}
let args_count = args.len().min(u16::MAX as usize) as u16; let args_count = args.len().min(u16::MAX as usize) as u16;
self.frames.push(CallFrame { self.frames.push(CallFrame {
+29 -25
View File
@@ -1,4 +1,4 @@
use crate::vm::{Completion, FrameContinuation, TryHandler, Vm}; use crate::vm::{Completion, FrameArgs, FrameContinuation, TryHandler, Vm};
use crate::vm_trace; use crate::vm_trace;
use oxide_bytecode::opcode; use oxide_bytecode::opcode;
use oxide_types::object::PropAttributes; use oxide_types::object::PropAttributes;
@@ -34,35 +34,39 @@ impl Vm {
return Ok(true); return Ok(true);
} else if obj.sub_module_index() > 0 { } else if obj.sub_module_index() > 0 {
let sub_idx = obj.sub_module_index() as usize; let sub_idx = obj.sub_module_index() as usize;
let args: Vec<JsValue> = (0..arg_count) let this_value = self.regs[this_reg as usize];
.map(|i| self.regs[first_arg_reg.wrapping_add(i as u8) as usize]) let is_generator = sub_idx < self.sub_modules.len() && self.sub_modules[sub_idx].is_generator;
.collect(); let is_async = sub_idx < self.sub_modules.len() && self.sub_modules[sub_idx].is_async;
// 异步生成器函数调用返回异步生成器迭代器对象。 // 生成器/异步路径需把实参物化存进状态盒;普通字节码调用直接
if sub_idx < self.sub_modules.len() // 引用寄存器连续区间,免临时堆 Vec(每次调用省 1 次分配)。
&& self.sub_modules[sub_idx].is_generator if is_generator || is_async {
&& self.sub_modules[sub_idx].is_async let args: Vec<JsValue> = (0..arg_count)
{ .map(|i| self.regs[first_arg_reg.wrapping_add(i as u8) as usize])
let gen = .collect();
self.create_async_generator_object(callee, self.regs[this_reg as usize], &args)?; // 异步生成器函数调用返回异步生成器迭代器对象。
self.regs[0] = gen; if is_generator && is_async {
return Ok(false); let gen = self.create_async_generator_object(callee, this_value, &args)?;
} self.regs[0] = gen;
// 生成器函数调用返回迭代器对象,不执行函数体。 return Ok(false);
if sub_idx < self.sub_modules.len() && self.sub_modules[sub_idx].is_generator { }
let gen = self.create_generator_object(callee, self.regs[this_reg as usize], &args)?; // 生成器函数调用返回迭代器对象,不执行函数体。
self.regs[0] = gen; if is_generator {
return Ok(false); let gen = self.create_generator_object(callee, this_value, &args)?;
} self.regs[0] = gen;
// 异步函数调用返回 capability promise,立即同步执行 body 到首个 await。 return Ok(false);
if sub_idx < self.sub_modules.len() && self.sub_modules[sub_idx].is_async { }
let promise = self.create_async_object(callee, self.regs[this_reg as usize], &args)?; // 异步函数调用返回 capability promise,立即同步执行 body 到首个 await。
let promise = self.create_async_object(callee, this_value, &args)?;
self.regs[0] = promise; self.regs[0] = promise;
return Ok(false); return Ok(false);
} }
self.push_bytecode_frame( self.push_bytecode_frame(
callee, callee,
self.regs[this_reg as usize], this_value,
&args, FrameArgs::RegRange {
first: first_arg_reg,
count: arg_count,
},
None, None,
None, None,
JsValue::undefined(), JsValue::undefined(),
+17 -73
View File
@@ -1,7 +1,5 @@
use std::sync::Arc;
use crate::native::NativeFn; use crate::native::NativeFn;
use crate::vm::{native_fn_ptr_to_fn, CallFrame, ForInIter, FrameContinuation, Vm, MAX_PROTO_CHAIN_DEPTH}; use crate::vm::{native_fn_ptr_to_fn, ForInIter, FrameArgs, FrameContinuation, Vm, MAX_PROTO_CHAIN_DEPTH};
use crate::vm_trace; use crate::vm_trace;
use oxide_runtime_api::{to_boolean, to_string_full, NativeResult, VmHost}; use oxide_runtime_api::{to_boolean, to_string_full, NativeResult, VmHost};
use oxide_types::object::{JsObject, PropAttributes}; use oxide_types::object::{JsObject, PropAttributes};
@@ -106,81 +104,27 @@ impl Vm {
return self.raise_type_error("g is not a constructor").map(|_| true); return self.raise_type_error("g is not a constructor").map(|_| true);
} }
if self.frames.len() >= self.kernel_core.config.max_call_depth {
return Err(self.error_message_text("RangeError", "Maximum call stack size exceeded"));
}
let new_obj_val = JsValue::object(new_obj as *mut u8); let new_obj_val = JsValue::object(new_obj as *mut u8);
let sub_bytecode = Arc::clone(&self.sub_modules[sub_idx].bytecode); // 收敛到统一压帧入口:derived 构造器在 super() 前 this 为 undefined,
let sub_n_args = self.sub_modules[sub_idx].n_args as usize; // 基类构造器 this = 新对象;new.target = 构造器本身。
let sub_n_registers = self.sub_modules[sub_idx].n_registers; let this_value = if ctor_obj.is_derived_constructor() {
let sub_param_base = self.sub_modules[sub_idx].param_base as usize;
let caller_active_reg_limit = self.active_reg_limit.max(1);
let caller_reg_limit = self.call_window_limit(caller_active_reg_limit, call_window);
let saved_reg_offset = self.save_stack.len() as u32;
self.save_stack.extend_from_slice(&self.regs[..caller_reg_limit as usize]);
let saved_this = self.regs[254];
let saved_new_target = self.regs[255];
// 完整实参写入 spill 栈实参区(在帧的 spill 区之前),供 CREATE_ARGUMENTS 使用。
let args_base = self.spill_stack.len() as u32;
for i in 0..arg_count {
let src_reg = first_arg_reg.wrapping_add(i as u8) as usize;
self.spill_stack.push(self.regs[src_reg]);
}
let args_count = arg_count.min(u16::MAX as usize) as u16;
for i in 0..sub_n_args {
let src_reg = first_arg_reg.wrapping_add(i as u8) as usize;
self.regs[sub_param_base + i] = self.regs[src_reg];
}
self.regs[254] = if ctor_obj.is_derived_constructor() {
JsValue::undefined() JsValue::undefined()
} else { } else {
new_obj_val new_obj_val
}; };
self.regs[255] = constructor; self.push_bytecode_frame(
constructor,
self.saved_bytecode_stack.push(std::mem::take(&mut self.bytecode)); this_value,
self.saved_immutables_stack.push(self.active_immutables); FrameArgs::RegRange {
first: first_arg_reg,
self.frames.push(CallFrame { count: arg_count,
return_addr: self.pc, },
function_name: self.sub_modules[sub_idx] Some(rd as u8),
.function_name Some(new_obj_val),
.as_deref() constructor,
.map(|name| self.kernel_core.perm_interner().intern(name).0) FrameContinuation::None,
.unwrap_or(0), call_window,
caller_reg_limit, )?;
caller_active_reg_limit,
saved_reg_offset,
spill_offset: self.spill_stack.len() as u32,
arguments_base: args_base,
arguments_count: args_count,
saved_this,
saved_new_target,
callee: constructor,
construct_result_reg: Some(rd as u8),
constructed_this: Some(new_obj_val),
is_derived_constructor: ctor_obj.is_derived_constructor(),
continuation: FrameContinuation::None,
});
self.bytecode = sub_bytecode;
let subs = Arc::clone(&self.sub_modules);
self.activate_immutables(sub_idx, &subs[sub_idx].constants);
self.cell_stack.push(Vec::with_capacity(subs[sub_idx].cells_needed as usize));
for (name, reg) in &self.sub_modules[sub_idx].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);
}
}
self.active_reg_limit = sub_n_registers.max(1);
self.pc = 0;
Ok(true) Ok(true)
} else { } else {
let error = let error =
+3 -3
View File
@@ -1,4 +1,4 @@
use crate::vm::{FrameContinuation, Vm, MAX_PROTO_CHAIN_DEPTH}; use crate::vm::{FrameArgs, FrameContinuation, Vm, MAX_PROTO_CHAIN_DEPTH};
use crate::{ic_trace, vm_trace}; use crate::{ic_trace, vm_trace};
use oxide_kernel::prop_forge::PropTemplate; use oxide_kernel::prop_forge::PropTemplate;
use oxide_runtime_api as coercion; use oxide_runtime_api as coercion;
@@ -126,7 +126,7 @@ impl Vm {
self.push_bytecode_frame( self.push_bytecode_frame(
getter, getter,
receiver, receiver,
&[], FrameArgs::Slice(&[]),
None, None,
None, None,
JsValue::undefined(), JsValue::undefined(),
@@ -326,7 +326,7 @@ impl Vm {
self.push_bytecode_frame( self.push_bytecode_frame(
setter, setter,
receiver, receiver,
&[val], FrameArgs::Slice(&[val]),
None, None,
None, None,
JsValue::undefined(), JsValue::undefined(),
+2 -2
View File
@@ -2,7 +2,7 @@ use std::sync::{Arc, OnceLock};
use oxide_bytecode::module::CompiledModule; use oxide_bytecode::module::CompiledModule;
use crate::vm::{CallFrame, FrameContinuation, InlineSyncState, Vm}; use crate::vm::{CallFrame, FrameArgs, FrameContinuation, InlineSyncState, Vm};
use crate::{vm_debug, vm_info, vm_trace, vm_warn}; use crate::{vm_debug, vm_info, vm_trace, vm_warn};
use oxide_types::object::JsObject; use oxide_types::object::JsObject;
use oxide_types::value::JsValue; use oxide_types::value::JsValue;
@@ -205,7 +205,7 @@ impl Vm {
self.push_bytecode_frame( self.push_bytecode_frame(
callee, callee,
this_value, this_value,
args, FrameArgs::Slice(args),
None, None,
None, None,
JsValue::undefined(), JsValue::undefined(),