Compare commits

..

1 Commits

Author SHA1 Message Date
ef4bdfc8eb [deploy]部署版本2 2025-07-24 13:30:21 +08:00
468 changed files with 103596 additions and 16389 deletions

5
.gitignore vendored
View File

@ -23,6 +23,7 @@
# Compiled Static libraries
*.lai
*.la
*.a
*.lib
# Executables
@ -36,7 +37,6 @@ doxygen
!/testdata/functional/*.out
!/testdata/h_functional/*.out
!/testdata/performance/*.out
build/
.antlr
.vscode/
@ -53,5 +53,4 @@ __init__.py
.DS_*
antlr/
.clang-format
antlr/

View File

@ -1,246 +0,0 @@
# 记录中端遍的开发进度
| 名称 | 优化级别 | 开发进度 |
| ------------ | ------------ | ---------- |
| CFG优化 | 函数级 | 已完成 |
| DCE | 函数级 | 待正确性测试 |
| Mem2Reg | 函数级 | 待正确性测试 |
| Reg2Mem | 函数级 | 待正确性测试 |
# 部分优化遍的说明
## Mem2Reg
Mem2Reg 遍的主要目标是将那些不必要的、只用于局部标量变量的内存分配 (alloca 指令) 消除,并将这些变量的值转换为 SSA 形式。这有助于减少内存访问,提高代码效率,并为后续的优化创造更好的条件。
通过Mem2Reg理解删除指令时对use关系的维护
`Mem2Reg` 优化遍中,当 `load``store` 指令被删除时,其 `use` 关系(即它们作为操作数与其他 `Value` 对象之间的连接)的正确消除是一个关键问题,尤其涉及到 `AllocaInst`
结合您提供的 `Mem2RegContext::renameVariables` 代码和我们之前讨论的 `usedelete` 逻辑,下面是 `use` 关系如何被正确消除的详细过程:
### 问题回顾:`Use` 关系的双向性
在您的 IR 设计中,`Use` 对象扮演着连接 `User`(使用者,如 `LoadInst`)和 `Value`(被使用者,如 `AllocaInst`)的双向角色:
* 一个 `User` 持有对其操作数 `Value``Use` 对象(通过 `User::operands` 列表)。
* 一个 `Value` 持有所有使用它的 `User``Use` 对象(通过 `Value::uses` 列表)。
原始问题是:当一个 `LoadInst``StoreInst` 被删除时,如果不对其作为操作数与 `AllocaInst` 之间的 `Use` 关系进行明确清理,`AllocaInst``uses` 列表中就会留下指向已删除 `LoadInst` / `StoreInst``Use` 对象,导致内部的 `User*` 指针悬空,在后续访问时引发 `segmentation fault`
### `Mem2Reg` 中 `load`/`store` 指令的删除行为
`Mem2RegContext::renameVariables` 函数中,`load``store` 指令被处理时,其行为如下:
1. **处理 `LoadInst`**
当找到一个指向可提升 `AllocaInst``LoadInst` 时,其用途会被 `replaceAllUsesWith(allocaToValueStackMap[alloca].top())` 替换。这意味着任何原本使用 `LoadInst` 本身计算结果的指令,现在都直接使用 SSA 值栈顶部的 `Value`
**重点:** 这一步处理的是 `LoadInst` 作为**被使用的值 (Value)** 时,其 `uses` 列表的清理。即,将 `LoadInst` 的所有使用者重定向到新的 SSA 值,并把这些 `Use` 对象从 `LoadInst``uses` 列表中移除。
2. **处理 `StoreInst`**
当找到一个指向可提升 `AllocaInst``StoreInst` 时,`StoreInst` 存储的值会被压入值栈。`StoreInst` 本身并不产生可被其他指令直接使用的值(其类型是 `void`),所以它没有 `uses` 列表需要替换。
**重点:** `StoreInst` 的主要作用是更新内存状态,在 SSA 形式下,它被移除后需要清理它作为**使用者 (User)** 时的操作数关系。
在这两种情况下,一旦 `load``store` 指令的 SSA 转换完成,它们都会通过 `instIter = SysYIROptUtils::usedelete(instIter)` 被显式删除。
### `SysYIROptUtils::usedelete` 如何正确消除 `Use` 关系
关键在于对 `SysYIROptUtils::usedelete` 函数的修改,使其在删除指令时,同时处理该指令作为 `User``Value` 的两种 `Use` 关系:
1. **清理指令作为 `Value` 时的 `uses` 列表 (由 `replaceAllUsesWith` 完成)**
`usedelete` 函数中,`inst->replaceAllUsesWith(UndefinedValue::get(inst->getType()))` 的调用至关重要。这确保了:
* 如果被删除的 `Instruction`(例如 `LoadInst`)产生了结果值并被其他指令使用,所有这些使用者都会被重定向到 `UndefinedValue`(或者 `Mem2Reg` 中具体的 SSA 值)。
* 这个过程会遍历 `LoadInst``uses` 列表,并将这些 `Use` 对象从 `LoadInst``uses` 列表中移除。这意味着 `LoadInst` 自己不再被任何其他指令使用。
2. **清理指令作为 `User` 时其操作数的 `uses` 列表 (由 `RemoveUserOperandUses` 完成)**
这是您提出的、并已集成到 `usedelete` 中的关键改进点。对于一个被删除的 `Instruction`(它同时也是 `User`),我们需要清理它**自己使用的操作数**所维护的 `use` 关系。
* 例如,`LoadInst %op1` 使用了 `%op1`(一个 `AllocaInst`)。当 `LoadInst` 被删除时,`AllocaInst``uses` 列表中有一个 `Use` 对象指向这个 `LoadInst`
* `RemoveUserOperandUses` 函数会遍历被删除 `User`(即 `LoadInst``StoreInst`)的 `operands` 列表。
* 对于 `operands` 列表中的每个 `std::shared_ptr<Use> use_ptr`,它会获取 `Use` 对象内部指向的 `Value`(例如 `AllocaInst*`),然后调用 `value->removeUse(use_ptr)`
* 这个 `removeUse` 调用会负责将 `use_ptr``AllocaInst``uses` 列表中删除。
### 总结
通过在 `SysYIROptUtils::usedelete` 中同时执行这两个步骤:
* `replaceAllUsesWith`:处理被删除指令**作为结果被使用**时的 `use` 关系。
* `RemoveUserOperandUses`:处理被删除指令**作为使用者User其操作数**的 `use` 关系。
这就确保了当 `Mem2Reg` 遍历并删除 `load``store` 指令时,无论是它们作为 `Value` 的使用者,还是它们作为 `User` 的操作数,所有相关的 `Use` 对象都能被正确地从 `Value``uses` 列表中移除,从而避免了悬空指针和后续的 `segmentation fault`
最后,当所有指向某个 `AllocaInst``load``store` 指令都被移除后,`AllocaInst``uses` 列表将变得干净(只包含 Phi 指令,如果它们在 SSA 转换中需要保留 Alloca 作为操作数),这时在 `Mem2RegContext::cleanup()` 阶段,`SysYIROptUtils::usedelete(alloca)` 就可以安全地删除 `AllocaInst` 本身了。
## Reg2Mem
我们的Reg2Mem 遍的主要目标是作为 Mem2Reg 的一种逆操作,但更具体是解决后端无法识别 PhiInst 指令的问题。主要的速录是将函数参数和 PhiInst 指令的结果从 SSA 形式转换回内存形式,通过插入 alloca、load 和 store 指令来实现。其他非 Phi 的指令结果将保持 SSA 形式。
## SCCP
SCCP稀疏条件常量传播是一种编译器优化技术它结合了常量传播和死代码消除。其核心思想是在程序执行过程中尝试识别并替换那些在编译时就能确定其值的变量常量同时移除那些永远不会被执行到的代码块不可达代码
以下是 SCCP 的实现思路:
1. 核心数据结构与工作列表:
Lattice 值Lattice Value: SCCP 使用三值格Three-Valued Lattice来表示变量的状态
Top (T): 初始状态,表示变量的值未知,但可能是一个常量。
Constant (C): 表示变量的值已经确定为一个具体的常量。
Bottom (⊥): 表示变量的值不确定或不是一个常量(例如,它可能在运行时有多个不同的值,或者从内存中加载)。一旦变量状态变为 Bottom它就不能再变回 Constant 或 Top。
SSAPValue: 封装了 Lattice 值和常量具体值(如果状态是 Constant
*valState (map<Value, SSAPValue>):** 存储程序中每个 Value变量、指令结果等的当前 SCCP Lattice 状态。
*ExecutableBlocks (set<BasicBlock>):** 存储在分析过程中被确定为可执行的基本块。
工作列表 (Worklists):
cfgWorkList (queue<pair<BasicBlock, BasicBlock>>):** 存储待处理的控制流图CFG边。当一个块被标记为可执行时它的后继边会被添加到这个列表。
*ssaWorkList (queue<Instruction>):** 存储待处理的 SSA (Static Single Assignment) 指令。当一个指令的任何操作数的状态发生变化时,该指令就会被添加到这个列表,需要重新评估。
2. 初始化:
所有 Value 的状态都被初始化为 Top。
所有基本块都被初始化为不可执行。
函数的入口基本块被标记为可执行,并且该块中的所有指令被添加到 ssaWorkList。
3. 迭代过程 (Fixed-Point Iteration)
SCCP 的核心是一个迭代过程,它交替处理 CFG 工作列表和 SSA 工作列表,直到达到一个不动点(即没有更多的状态变化)。
处理 cfgWorkList:
从 cfgWorkList 中取出一个边 (prev, next)。
如果 next 块之前是不可执行的,现在通过 prev 块可达,则将其标记为可执行 (markBlockExecutable)。
一旦 next 块变为可执行,其内部的所有指令(特别是 Phi 指令)都需要被重新评估,因此将它们添加到 ssaWorkList。
处理 ssaWorkList:
从 ssaWorkList 中取出一个指令 inst。
重要: 只有当 inst 所在的块是可执行的,才处理该指令。不可执行块中的指令不参与常量传播。
计算新的 Lattice 值 (computeLatticeValue): 根据指令类型和其操作数的当前 Lattice 状态,计算 inst 的新的 Lattice 状态。
常量折叠: 如果所有操作数都是常量,则可以直接执行运算并得到一个新的常量结果。
Bottom 传播: 如果任何操作数是 Bottom或者运算规则导致不确定例如除以零则结果为 Bottom。
Phi 指令的特殊处理: Phi 指令的值取决于其所有可执行的前驱块传入的值。
如果所有可执行前驱都提供了相同的常量 C则 Phi 结果为 C。
如果有任何可执行前驱提供了 Bottom或者不同的可执行前驱提供了不同的常量则 Phi 结果为 Bottom。
如果所有可执行前驱都提供了 Top则 Phi 结果仍为 Top。
更新状态: 如果 inst 的新计算出的 Lattice 值与它当前存储的值不同,则更新 valState[inst]。
传播变化: 如果 inst 的状态发生变化,那么所有使用 inst 作为操作数的指令都可能受到影响,需要重新评估。因此,将 inst 的所有使用者添加到 ssaWorkList。
处理终结符指令 (BranchInst, ReturnInst):
对于条件分支 BranchInst如果其条件操作数变为常量
如果条件为真,则只有真分支的目标块是可达的,将该边添加到 cfgWorkList。
如果条件为假,则只有假分支的目标块是可达的,将该边添加到 cfgWorkList。
如果条件不是常量Top 或 Bottom则两个分支都可能被执行将两边的边都添加到 cfgWorkList。
这会影响 CFG 的可达性分析,可能导致新的块被标记为可执行。
4. 应用优化 (Transformation)
当两个工作列表都为空,达到不动点后,程序代码开始进行实际的修改:
常量替换:
遍历所有指令。如果指令的 valState 为 Constant则用相应的 ConstantValue 替换该指令的所有用途 (replaceAllUsesWith)。
将该指令标记为待删除。
对于指令的操作数,如果其 valState 为 Constant则直接将操作数替换为对应的 ConstantValue常量折叠
删除死指令: 遍历所有标记为待删除的指令,并从其父基本块中删除它们。
删除不可达基本块: 遍历函数中的所有基本块。如果一个基本块没有被标记为可执行 (ExecutableBlocks 中不存在),则将其从函数中删除。但入口块不能删除。
简化分支指令:
遍历所有可执行的基本块的终结符指令。
对于条件分支 BranchInst如果其条件操作数在 valState 中是 Constant
如果条件为真,则将该条件分支替换为一个无条件跳转到真分支目标块的指令。
如果条件为假,则将该条件分支替换为一个无条件跳转到假分支目标块的指令。
更新 CFG移除不可达的分支边和其前驱信息。
computeLatticeValue 的具体逻辑:
这个函数是 SCCP 的核心逻辑,它定义了如何根据指令类型和操作数的当前 Lattice 状态来计算指令结果的 Lattice 状态。
二元运算 (Add, Sub, Mul, Div, Rem, ICmp, And, Or):
如果任何一个操作数是 Bottom结果就是 Bottom。
如果任何一个操作数是 Top结果就是 Top。
如果两个操作数都是 Constant执行实际的常量运算结果是一个新的 Constant。
一元运算 (Neg, Not):
如果操作数是 Bottom结果就是 Bottom。
如果操作数是 Top结果就是 Top。
如果操作数是 Constant执行实际的常量运算结果是一个新的 Constant。
Load 指令: 通常情况下Load 的结果会被标记为 Bottom因为内存内容通常在编译时无法确定。但如果加载的是已知的全局常量可能可以确定。在提供的代码中它通常返回 Bottom。
Store 指令: Store 不产生值,所以其 SSAPValue 保持 Top 或不关心。
Call 指令: 大多数 Call 指令(尤其是对外部或有副作用的函数)的结果都是 Bottom。对于纯函数如果所有参数都是常量理论上可以折叠但这需要额外的分析。
GetElementPtr (GEP) 指令: GEP 计算内存地址。如果所有索引都是常量,地址本身是常量。但 SCCP 关注的是数据值,因此这里通常返回 Bottom除非有特定的指针常量跟踪。
Phi 指令: 如上所述,基于所有可执行前驱的传入值进行聚合。
Alloc 指令: Alloc 分配内存,返回一个指针。其内容通常是 Bottom。
Branch 和 Return 指令: 这些是终结符指令,不产生一个可用于其他指令的值,通常 SSAPValue 保持 Top 或不关心。
类型转换 (ZExt, SExt, Trunc, FtoI, ItoF): 如果操作数是 Constant则执行相应的类型转换结果仍为 Constant。对于浮点数转换由于 SSAPValue 的 constantVal 为 int 类型,所以对浮点数的操作会保守地返回 Bottom。
未处理的指令: 默认情况下,任何未明确处理的指令都被保守地假定为产生 Bottom 值。
浮点数处理的注意事项:
在提供的代码中SSAPValue 的 constantVal 是 int 类型。这使得浮点数常量传播变得复杂。对于浮点数相关的指令kFAdd, kFMul, kFCmp, kFNeg, kFNot, kItoF, kFtoI 等),如果不能将浮点值准确地存储在 int 中,或者不能可靠地执行浮点运算,那么通常会保守地将结果设置为 Bottom。一个更完善的 SCCP 实现会使用 std::variant<int, float> 或独立的浮点常量存储来处理浮点数。
# 后续优化可能涉及的改动
## 1将所有的alloca集中到entryblock中
好处优化友好性方便mem2reg提升
目前没有实现这个机制,如果想要实现首先解决同一函数不同域的同名变量命名区分
需要保证符号表能正确维护域中的局部变量
# 关于中端优化提升编译器性能的TODO
## usedelete_withinstdelte方法
这个方法删除了use关系并移除了指令逻辑是根据Instruction* inst去find对应的迭代器并erase
有些情况下外部持有迭代器和inst,可以省略find过程

View File

@ -37,38 +37,4 @@ mysysy/ $ bash setup.sh
```
### 配套脚本
TODO: 需要完善)
### TODO_list:
除开注释中的TODO后续时间充足可以考虑的TODO:
- store load指令由于gep指令的引入, 维度信息的记录是非必须的, 考虑删除
- use def关系经过mem2reg和phi函数明确转换为ssa形式, 以及函数参数通过value数组明确定义, 使得基本块的args参数信息记录非必须, 考虑删除
---
## 编译器后端 TODO 列表
### 1. `CALL` 指令处理不完善 (高优先级)
* **问题描述**:当前 `RISCv64RegAlloc::getInstrUseDef()` 方法中,对 `CALL` 指令的 `use`/`def` 分析不完整。它正确识别了返回值为 `def` 和参数为 `use`,但**没有将所有调用者保存 (Caller-saved) 的物理寄存器(`T0-T6`, `A0-A7`)标记为隐式 `def` (即 `CALL` 会破坏它们)**。
* **潜在后果**
* **活跃性分析错误**:寄存器分配器可能会错误地认为某个跨函数调用活跃的虚拟寄存器是安全的,并将其分配给 `T` 或 `A` 寄存器。
* **值被破坏**:在 `CALL` 指令执行后,这些 `T` 或 `A` 寄存器中本应保留的值会被被调用的函数破坏,导致程序行为异常。
* **参考文件**`RISCv64RegAlloc.cpp` (在 `getInstrUseDef` 函数中对 `RVOpcodes::CALL` 的处理)。
### 2. `T6` 寄存器作为溢出寄存器的问题 (中等优先级)
* **问题描述**`RISCv64RegAlloc::rewriteFunction()` 方法中,所有未能成功着色并被溢出 (spilled) 的虚拟寄存器,都被统一替换为物理寄存器 `T6`。
* **问题 2.1`T6` 是调用者保存寄存器,但未被调用者保存**
* `T6` 属于调用者保存寄存器 (`T0-T6` 范围)。
* 标准 ABI 要求,如果一个调用者保存寄存器在函数调用前后都活跃(例如,它存储了一个被溢出的变量,而这个变量在 `CALL` 指令之后还需要用到),那么**调用者**有责任在 `CALL` 前保存该寄存器,并在 `CALL` 后恢复它。
* 目前的 `rewriteFunction` 没有为 `T6` 插入这种保存/恢复逻辑。
* **潜在后果**:如果一个溢出变量被分配到 `T6`,并且它跨函数调用活跃,那么 `putint` 或其他任何被调用的函数可能会随意使用 `T6`,从而破坏该溢出变量的值。
* **问题 2.2:所有溢出变量共用一个 `T6`**
* 将所有溢出变量映射到同一个物理寄存器 `T6` 是一种简化的溢出策略。
* **潜在后果**:这意味着,每当需要使用一个溢出变量时,其值必须从栈中加载到 `T6`;每当一个溢出变量被定义时,其值必须从 `T6` 存储回栈。这会引入大量的 `load`/`store` 指令,并导致 `T6` 本身成为一个高度冲突的寄存器,严重降低代码效率。
* **参考文件**`RISCv64RegAlloc.cpp` (在 `rewriteFunction` 函数中处理 `spilled_vregs` 的部分)。
TODO: 需要完善)

2732
doc/Doxyfile Normal file

File diff suppressed because it is too large Load Diff

Binary file not shown.

Binary file not shown.

BIN
doc/arm/DDI0403E_e_armv7m_arm.pdf Executable file

Binary file not shown.

Binary file not shown.

File diff suppressed because one or more lines are too long

Binary file not shown.

Binary file not shown.

BIN
doc/arm/Raspberry.pdf Executable file

Binary file not shown.

Binary file not shown.

After

Width:  |  Height:  |  Size: 346 KiB

169
doc/markdowns/IR.md Normal file
View File

@ -0,0 +1,169 @@
这个头文件定义了一个用于生成中间表示IR的数据结构主要用于编译器前端将抽象语法树AST转换为中间代码。以下是文件中定义的主要类和功能的整理和解释
---
### **1. 类型系统Type System**
#### **1.1 `Type` 类**
- **作用**:表示所有基本标量类型(如 `int``float``void` 等)以及指针类型和函数类型。
- **成员**
- `Kind` 枚举:表示类型的种类(如 `kInt``kFloat``kPointer` 等)。
- `kind`:当前类型的种类。
- 构造函数:`Type(Kind kind)`,用于初始化类型。
- 静态方法:如 `getIntType()``getFloatType()` 等,用于获取特定类型的单例对象。
- 类型检查方法:如 `isInt()``isFloat()` 等,用于检查当前类型是否为某种类型。
- `getSize()`:获取类型的大小。
- `as<T>()`:将当前类型动态转换为派生类(如 `PointerType``FunctionType`)。
#### **1.2 `PointerType` 类**
- **作用**:表示指针类型,派生自 `Type`
- **成员**
- `baseType`:指针指向的基础类型。
- 静态方法:`get(Type *baseType)`,用于获取指向 `baseType` 的指针类型。
- `getBaseType()`:获取指针指向的基础类型。
#### **1.3 `FunctionType` 类**
- **作用**:表示函数类型,派生自 `Type`
- **成员**
- `returnType`:函数的返回类型。
- `paramTypes`:函数的参数类型列表。
- 静态方法:`get(Type *returnType, const std::vector<Type *> &paramTypes)`,用于获取函数类型。
- `getReturnType()`:获取函数的返回类型。
- `getParamTypes()`:获取函数的参数类型列表。
---
### **2. 中间表示IR**
#### **2.1 `Value` 类**
- **作用**:表示 IR 中的所有值(如指令、常量、参数等)。
- **成员**
- `Kind` 枚举:表示值的种类(如 `kAdd``kSub``kConstant` 等)。
- `kind`:当前值的种类。
- `type`:值的类型。
- `name`:值的名称。
- `uses`:值的用途列表(表示哪些指令使用了该值)。
- 构造函数:`Value(Kind kind, Type *type, const std::string &name)`
- 类型检查方法:如 `isInt()``isFloat()` 等。
- `getUses()`:获取值的用途列表。
- `replaceAllUsesWith(Value *value)`:将该值的所有用途替换为另一个值。
- `print(std::ostream &os)`:打印值的表示。
#### **2.2 `ConstantValue` 类**
- **作用**:表示编译时常量(如整数常量、浮点数常量)。
- **成员**
- `iScalar``fScalar`:分别存储整数和浮点数常量的值。
- 静态方法:`get(int value)``get(float value)`,用于获取常量值。
- `getInt()``getFloat()`:获取常量的值。
#### **2.3 `Argument` 类**
- **作用**:表示函数或基本块的参数。
- **成员**
- `block`:参数所属的基本块。
- `index`:参数的索引。
- 构造函数:`Argument(Type *type, BasicBlock *block, int index, const std::string &name)`
- `getParent()`:获取参数所属的基本块。
- `getIndex()`:获取参数的索引。
#### **2.4 `BasicBlock` 类**
- **作用**:表示基本块,包含一系列指令。
- **成员**
- `parent`:基本块所属的函数。
- `instructions`:基本块中的指令列表。
- `arguments`:基本块的参数列表。
- `successors``predecessors`:基本块的后继和前驱列表。
- 构造函数:`BasicBlock(Function *parent, const std::string &name)`
- `getParent()`:获取基本块所属的函数。
- `getInstructions()`:获取基本块中的指令列表。
- `createArgument()`:为基本块创建一个参数。
#### **2.5 `Instruction` 类**
- **作用**:表示 IR 中的指令,派生自 `User`
- **成员**
- `Kind` 枚举:表示指令的种类(如 `kAdd``kSub``kLoad` 等)。
- `kind`:当前指令的种类。
- `parent`:指令所属的基本块。
- 构造函数:`Instruction(Kind kind, Type *type, BasicBlock *parent, const std::string &name)`
- `getParent()`:获取指令所属的基本块。
- `getFunction()`:获取指令所属的函数。
- 指令分类方法:如 `isBinary()``isUnary()``isMemory()` 等。
#### **2.6 `User` 类**
- **作用**:表示使用其他值的指令或全局值,派生自 `Value`
- **成员**
- `operands`:指令的操作数列表。
- 构造函数:`User(Kind kind, Type *type, const std::string &name)`
- `getOperand(int index)`:获取指定索引的操作数。
- `addOperand(Value *value)`:添加一个操作数。
- `replaceOperand(int index, Value *value)`:替换指定索引的操作数。
#### **2.7 具体指令类**
- **`CallInst`**:表示函数调用指令。
- **`UnaryInst`**:表示一元操作指令(如取反、类型转换)。
- **`BinaryInst`**:表示二元操作指令(如加法、减法)。
- **`ReturnInst`**:表示返回指令。
- **`UncondBrInst`**:表示无条件跳转指令。
- **`CondBrInst`**:表示条件跳转指令。
- **`AllocaInst`**:表示栈内存分配指令。
- **`LoadInst`**:表示从内存加载值的指令。
- **`StoreInst`**:表示将值存储到内存的指令。
---
### **3. 模块和函数**
#### **3.1 `Function` 类**
- **作用**:表示函数,包含多个基本块。
- **成员**
- `parent`:函数所属的模块。
- `blocks`:函数中的基本块列表。
- 构造函数:`Function(Module *parent, Type *type, const std::string &name)`
- `getReturnType()`:获取函数的返回类型。
- `getParamTypes()`:获取函数的参数类型列表。
- `addBasicBlock()`:为函数添加一个基本块。
#### **3.2 `GlobalValue` 类**
- **作用**:表示全局变量或常量。
- **成员**
- `parent`:全局值所属的模块。
- `hasInit`:是否有初始化值。
- `isConst`:是否是常量。
- 构造函数:`GlobalValue(Module *parent, Type *type, const std::string &name, const std::vector<Value *> &dims, Value *init)`
- `init()`:获取全局值的初始化值。
#### **3.3 `Module` 类**
- **作用**:表示整个编译单元(如一个源文件)。
- **成员**
- `children`:模块中的所有值(如函数、全局变量)。
- `functions`:模块中的函数列表。
- `globals`:模块中的全局变量列表。
- `createFunction()`:创建一个函数。
- `createGlobalValue()`:创建一个全局变量。
- `getFunction()`:获取指定名称的函数。
- `getGlobalValue()`:获取指定名称的全局变量。
---
### **4. 工具类**
#### **4.1 `Use` 类**
- **作用**:表示值与其使用者之间的关系。
- **成员**
- `index`:值在使用者操作数列表中的索引。
- `user`:使用者。
- `value`:被使用的值。
- 构造函数:`Use(int index, User *user, Value *value)`
- `getValue()`:获取被使用的值。
#### **4.2 `range` 类**
- **作用**:封装迭代器对 `[begin, end)`,用于遍历容器。
- **成员**
- `begin()``end()`:返回范围的起始和结束迭代器。
- `size()`:返回范围的大小。
- `empty()`:判断范围是否为空。
---
### **5. 总结**
- **类型系统**`Type``PointerType``FunctionType` 用于表示 IR 中的类型。
- **中间表示**`Value``ConstantValue``Instruction` 等用于表示 IR 中的值和指令。
- **模块和函数**`Module``Function``GlobalValue` 用于组织 IR 的结构。
- **工具类**`Use``range` 用于辅助实现 IR 的数据结构和遍历。
这个头文件定义了一个完整的 IR 数据结构,适用于编译器前端将 AST 转换为中间代码,并支持后续的优化和目标代码生成。

156
doc/markdowns/IRbuilder.md Normal file
View File

@ -0,0 +1,156 @@
`IRBuilder.h` 文件定义了一个 `IRBuilder`用于简化中间表示IR的构建过程。`IRBuilder` 提供了创建各种 IR 指令的便捷方法,并将这些指令插入到指定的基本块中。以下是对文件中主要内容的整理和解释:
---
### **1. `IRBuilder` 类的作用**
`IRBuilder` 是一个工具类用于在生成中间表示IR时简化指令的创建和插入操作。它的主要功能包括
- 提供创建各种 IR 指令的工厂方法。
- 将创建的指令插入到指定的基本块中。
- 支持在基本块的任意位置插入指令。
---
### **2. 主要成员**
#### **2.1 成员变量**
- **`block`**:当前操作的基本块。
- **`position`**:当前操作的插入位置(基本块中的迭代器)。
#### **2.2 构造函数**
- **默认构造函数**`IRBuilder()`
- **带参数的构造函数**
- `IRBuilder(BasicBlock *block)`:初始化 `IRBuilder`,并设置当前基本块和插入位置(默认在基本块末尾)。
- `IRBuilder(BasicBlock *block, BasicBlock::iterator position)`:初始化 `IRBuilder`,并设置当前基本块和插入位置。
#### **2.3 设置方法**
- **`setPosition(BasicBlock *block, BasicBlock::iterator position)`**:设置当前基本块和插入位置。
- **`setPosition(BasicBlock::iterator position)`**:设置当前插入位置。
#### **2.4 获取方法**
- **`getBasicBlock()`**:获取当前基本块。
- **`getPosition()`**:获取当前插入位置。
---
### **3. 指令创建方法**
`IRBuilder` 提供了多种工厂方法,用于创建不同类型的 IR 指令。这些方法会将创建的指令插入到当前基本块的指定位置。
#### **3.1 函数调用指令**
- **`createCallInst(Function *callee, const std::vector<Value *> &args, const std::string &name)`**
- 创建一个函数调用指令。
- 参数:
- `callee`:被调用的函数。
- `args`:函数参数列表。
- `name`:指令的名称(可选)。
- 返回:`CallInst*`
#### **3.2 一元操作指令**
- **`createUnaryInst(Instruction::Kind kind, Type *type, Value *operand, const std::string &name)`**
- 创建一个一元操作指令(如取反、类型转换)。
- 参数:
- `kind`:指令的类型(如 `kNeg``kFtoI` 等)。
- `type`:指令的结果类型。
- `operand`:操作数。
- `name`:指令的名称(可选)。
- 返回:`UnaryInst*`
- **具体一元操作指令**
- `createNegInst(Value *operand, const std::string &name)`:创建整数取反指令。
- `createNotInst(Value *operand, const std::string &name)`:创建逻辑取反指令。
- `createFtoIInst(Value *operand, const std::string &name)`:创建浮点数转整数指令。
- `createFNegInst(Value *operand, const std::string &name)`:创建浮点数取反指令。
- `createIToFInst(Value *operand, const std::string &name)`:创建整数转浮点数指令。
#### **3.3 二元操作指令**
- **`createBinaryInst(Instruction::Kind kind, Type *type, Value *lhs, Value *rhs, const std::string &name)`**
- 创建一个二元操作指令(如加法、减法)。
- 参数:
- `kind`:指令的类型(如 `kAdd``kSub` 等)。
- `type`:指令的结果类型。
- `lhs``rhs`:左操作数和右操作数。
- `name`:指令的名称(可选)。
- 返回:`BinaryInst*`
- **具体二元操作指令**
- 整数运算:
- `createAddInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数加法指令。
- `createSubInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数减法指令。
- `createMulInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数乘法指令。
- `createDivInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数除法指令。
- `createRemInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数取余指令。
- 整数比较:
- `createICmpEQInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数相等比较指令。
- `createICmpNEInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数不等比较指令。
- `createICmpLTInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数小于比较指令。
- `createICmpLEInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数小于等于比较指令。
- `createICmpGTInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数大于比较指令。
- `createICmpGEInst(Value *lhs, Value *rhs, const std::string &name)`:创建整数大于等于比较指令。
- 浮点数运算:
- `createFAddInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数加法指令。
- `createFSubInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数减法指令。
- `createFMulInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数乘法指令。
- `createFDivInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数除法指令。
- `createFRemInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数取余指令。
- 浮点数比较:
- `createFCmpEQInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数相等比较指令。
- `createFCmpNEInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数不等比较指令。
- `createFCmpLTInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数小于比较指令。
- `createFCmpLEInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数小于等于比较指令。
- `createFCmpGTInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数大于比较指令。
- `createFCmpGEInst(Value *lhs, Value *rhs, const std::string &name)`:创建浮点数大于等于比较指令。
#### **3.4 控制流指令**
- **`createReturnInst(Value *value)`**
- 创建返回指令。
- 参数:
- `value`:返回值(可选)。
- 返回:`ReturnInst*`
- **`createUncondBrInst(BasicBlock *block, std::vector<Value *> args)`**
- 创建无条件跳转指令。
- 参数:
- `block`:目标基本块。
- `args`:跳转参数(可选)。
- 返回:`UncondBrInst*`
- **`createCondBrInst(Value *condition, BasicBlock *thenBlock, BasicBlock *elseBlock, const std::vector<Value *> &thenArgs, const std::vector<Value *> &elseArgs)`**
- 创建条件跳转指令。
- 参数:
- `condition`:跳转条件。
- `thenBlock`:条件为真时的目标基本块。
- `elseBlock`:条件为假时的目标基本块。
- `thenArgs``elseArgs`:跳转参数(可选)。
- 返回:`CondBrInst*`
#### **3.5 内存操作指令**
- **`createAllocaInst(Type *type, const std::vector<Value *> &dims, const std::string &name)`**
- 创建栈内存分配指令。
- 参数:
- `type`:分配的类型。
- `dims`:数组维度(可选)。
- `name`:指令的名称(可选)。
- 返回:`AllocaInst*`
- **`createLoadInst(Value *pointer, const std::vector<Value *> &indices, const std::string &name)`**
- 创建加载指令。
- 参数:
- `pointer`:指针值。
- `indices`:数组索引(可选)。
- `name`:指令的名称(可选)。
- 返回:`LoadInst*`
- **`createStoreInst(Value *value, Value *pointer, const std::vector<Value *> &indices, const std::string &name)`**
- 创建存储指令。
- 参数:
- `value`:要存储的值。
- `pointer`:指针值。
- `indices`:数组索引(可选)。
- `name`:指令的名称(可选)。
- 返回:`StoreInst*`
---
### **4. 总结**
- `IRBuilder` 是一个用于简化 IR 构建的工具类,提供了创建各种 IR 指令的工厂方法。
- 通过 `IRBuilder`,可以方便地在指定基本块的任意位置插入指令。
- 该类的设计使得 IR 的生成更加模块化和易于维护。

121
doc/markdowns/IRcpp.md Normal file
View File

@ -0,0 +1,121 @@
这个 `IR.cpp` 文件实现了 `IR.h` 中定义的中间表示IR数据结构的功能。它包含了类型系统、值、指令、基本块、函数和模块的具体实现以及一些辅助函数用于打印 IR 的内容。以下是对文件中主要内容的整理和解释:
---
### **1. 辅助函数**
#### **1.1 `interleave` 函数**
- **作用**:用于在输出流中插入分隔符(如逗号)来打印容器中的元素。
- **示例**
```cpp
interleave(os, container, ", ");
```
#### **1.2 打印函数**
- **`printVarName`**:打印变量名,全局变量以 `@` 开头,局部变量以 `%` 开头。
- **`printBlockName`**:打印基本块名,以 `^` 开头。
- **`printFunctionName`**:打印函数名,以 `@` 开头。
- **`printOperand`**:打印操作数,如果是常量则直接打印值,否则打印变量名。
---
### **2. 类型系统**
#### **2.1 `Type` 类的实现**
- **静态方法**
- `getIntType()`、`getFloatType()`、`getVoidType()`、`getLabelType()`:返回对应类型的单例对象。
- `getPointerType(Type *baseType)`:返回指向 `baseType` 的指针类型。
- `getFunctionType(Type *returnType, const vector<Type *> &paramTypes)`:返回函数类型。
- **`getSize()`**:返回类型的大小(如 `int` 和 `float` 为 4 字节,指针为 8 字节)。
- **`print()`**:打印类型的表示。
#### **2.2 `PointerType` 类的实现**
- **静态方法**
- `get(Type *baseType)`:返回指向 `baseType` 的指针类型,使用 `std::map` 缓存已创建的指针类型。
- **`getBaseType()`**:返回指针指向的基础类型。
#### **2.3 `FunctionType` 类的实现**
- **静态方法**
- `get(Type *returnType, const vector<Type *> &paramTypes)`:返回函数类型,使用 `std::set` 缓存已创建的函数类型。
- **`getReturnType()`** 和 `getParamTypes()`:分别返回函数的返回类型和参数类型列表。
---
### **3. 值Value**
#### **3.1 `Value` 类的实现**
- **`replaceAllUsesWith(Value *value)`**:将该值的所有用途替换为另一个值。
- **`isConstant()`**:判断值是否为常量(包括常量值、全局值和函数)。
#### **3.2 `ConstantValue` 类的实现**
- **静态方法**
- `get(int value)``get(float value)`:返回整数或浮点数常量,使用 `std::map` 缓存已创建的常量。
- **`getInt()``getFloat()`**:返回常量的值。
- **`print()`**:打印常量的值。
#### **3.3 `Argument` 类的实现**
- **构造函数**:初始化参数的类型、所属基本块和索引。
- **`print()`**:打印参数的表示。
---
### **4. 基本块BasicBlock**
#### **4.1 `BasicBlock` 类的实现**
- **构造函数**:初始化基本块的名称和所属函数。
- **`print()`**:打印基本块的表示,包括参数和指令。
---
### **5. 指令Instruction**
#### **5.1 `Instruction` 类的实现**
- **构造函数**:初始化指令的类型、所属基本块和名称。
- **`print()`**:由具体指令类实现。
#### **5.2 具体指令类的实现**
- **`CallInst`**:表示函数调用指令。
- **`print()`**:打印函数调用的表示。
- **`UnaryInst`**:表示一元操作指令(如取反、类型转换)。
- **`print()`**:打印一元操作的表示。
- **`BinaryInst`**:表示二元操作指令(如加法、减法)。
- **`print()`**:打印二元操作的表示。
- **`ReturnInst`**:表示返回指令。
- **`print()`**:打印返回指令的表示。
- **`UncondBrInst`**:表示无条件跳转指令。
- **`print()`**:打印无条件跳转的表示。
- **`CondBrInst`**:表示条件跳转指令。
- **`print()`**:打印条件跳转的表示。
- **`AllocaInst`**:表示栈内存分配指令。
- **`print()`**:打印内存分配的表示。
- **`LoadInst`**:表示从内存加载值的指令。
- **`print()`**:打印加载指令的表示。
- **`StoreInst`**:表示将值存储到内存的指令。
- **`print()`**:打印存储指令的表示。
---
### **6. 函数Function**
#### **6.1 `Function` 类的实现**
- **构造函数**:初始化函数的名称、返回类型和参数类型。
- **`print()`**:打印函数的表示,包括基本块和指令。
---
### **7. 模块Module**
#### **7.1 `Module` 类的实现**
- **`print()`**:打印模块的表示,包括所有函数和全局变量。
---
### **8. 用户User**
#### **8.1 `User` 类的实现**
- **`setOperand(int index, Value *value)`**:设置指定索引的操作数。
- **`replaceOperand(int index, Value *value)`**:替换指定索引的操作数,并更新用途列表。
---
### **9. 总结**
- **类型系统**:实现了 `Type``PointerType``FunctionType`,用于表示 IR 中的类型。
- **值**:实现了 `Value``ConstantValue``Argument`,用于表示 IR 中的值和参数。
- **基本块**:实现了 `BasicBlock`,用于组织指令。
- **指令**:实现了多种具体指令类(如 `CallInst``BinaryInst` 等),用于表示 IR 中的操作。
- **函数和模块**:实现了 `Function``Module`,用于组织 IR 的结构。
- **打印功能**:通过 `print()` 方法,可以将 IR 的内容输出为可读的文本格式。
这个文件是编译器中间表示的核心实现能够将抽象语法树AST转换为中间代码并支持后续的优化和目标代码生成。

20878
doc/n1124.pdf Normal file

File diff suppressed because one or more lines are too long

Binary file not shown.

BIN
doc/riscv/riscv-spec-20191213.pdf Executable file

Binary file not shown.

BIN
doc/slides/exp01-antlr.pdf Executable file

Binary file not shown.

BIN
doc/slides/exp02-ir.pdf Executable file

Binary file not shown.

Binary file not shown.

Binary file not shown.

Binary file not shown.

BIN
doc/sysy-2022-runtime.pdf Normal file

Binary file not shown.

BIN
doc/sysy-2022-spec.pdf Normal file

Binary file not shown.

Binary file not shown.

Binary file not shown.

View File

@ -0,0 +1,171 @@
#include "AddressCalculationExpansion.h"
#include <iostream>
#include <vector>
#include "IR.h"
#include "IRBuilder.h"
extern int DEBUG;
namespace sysy {
bool AddressCalculationExpansion::run() {
bool changed = false;
for (auto& funcPair : pModule->getFunctions()) {
Function* func = funcPair.second.get();
for (auto& bb_ptr : func->getBasicBlocks()) {
BasicBlock* bb = bb_ptr.get();
for (auto it = bb->getInstructions().begin(); it != bb->getInstructions().end(); ) {
Instruction* inst = it->get();
Value* basePointer = nullptr;
Value* valueToStore = nullptr;
size_t firstIndexOperandIdx = 0;
size_t numBaseOperands = 0;
if (inst->isLoad()) {
numBaseOperands = 1;
basePointer = inst->getOperand(0);
firstIndexOperandIdx = 1;
} else if (inst->isStore()) {
numBaseOperands = 2;
valueToStore = inst->getOperand(0);
basePointer = inst->getOperand(1);
firstIndexOperandIdx = 2;
} else {
++it;
continue;
}
if (inst->getNumOperands() <= numBaseOperands) {
++it;
continue;
}
std::vector<int> dims;
if (AllocaInst* allocaInst = dynamic_cast<AllocaInst*>(basePointer)) {
for (const auto& use_ptr : allocaInst->getDims()) {
Value* dimValue = use_ptr->getValue();
if (ConstantValue* constVal = dynamic_cast<ConstantValue*>(dimValue)) {
dims.push_back(constVal->getInt());
} else {
std::cerr << "Warning: AllocaInst dimension is not a constant integer. Skipping GEP expansion for: ";
SysYPrinter::printValue(allocaInst);
std::cerr << "\n";
dims.clear();
break;
}
}
} else if (GlobalValue* globalValue = dynamic_cast<GlobalValue*>(basePointer)) {
// 遍历 GlobalValue 的所有维度操作数
for (const auto& use_ptr : globalValue->getDims()) {
Value* dimValue = use_ptr->getValue();
// 将维度值转换为常量整数
if (ConstantInteger* constVal = dynamic_cast<ConstantInteger*>(dimValue)) {
dims.push_back(constVal->getInt());
} else {
// 如果维度不是常量整数,则无法处理。
// 根据 IR.h 中 GlobalValue 的构造函数,这种情况不应发生,但作为安全检查是好的。
std::cerr << "Warning: GlobalValue dimension is not a constant integer. Skipping GEP expansion for: ";
SysYPrinter::printValue(globalValue);
std::cerr << "\n";
dims.clear(); // 清空已收集的部分维度信息
break;
}
}
} else {
std::cerr << "Warning: Base pointer is not AllocaInst/GlobalValue or its array dimensions cannot be determined for GEP expansion. Skipping GEP for: ";
SysYPrinter::printValue(basePointer);
std::cerr << " in instruction ";
SysYPrinter::printInst(inst);
std::cerr << "\n";
++it;
continue;
}
if (dims.empty() && (inst->getNumOperands() > numBaseOperands)) {
if (DEBUG) {
std::cerr << "ACE Warning: Could not get valid array dimensions for ";
SysYPrinter::printValue(basePointer);
std::cerr << " in instruction ";
SysYPrinter::printInst(inst);
std::cerr << " (expected dimensions for indices, but got none).\n";
}
++it;
continue;
}
std::vector<Value*> indexOperands;
for (size_t i = firstIndexOperandIdx; i < inst->getNumOperands(); ++i) {
indexOperands.push_back(inst->getOperand(i));
}
if (AllocaInst* allocaInst = dynamic_cast<AllocaInst*>(basePointer)) {
if (allocaInst->getNumDims() != indexOperands.size()) {
if (DEBUG) {
std::cerr << "ACE Warning: Index count (" << indexOperands.size() << ") does not match AllocaInst dimensions (" << allocaInst->getNumDims() << ") for instruction ";
SysYPrinter::printInst(inst);
std::cerr << "\n";
}
++it;
continue;
}
}
Value* totalOffset = ConstantInteger::get(0);
pBuilder->setPosition(bb, it);
for (size_t i = 0; i < indexOperands.size(); ++i) {
Value* index = indexOperands[i];
int stride = calculateStride(dims, i);
Value* strideConst = ConstantInteger::get(stride);
Type* intType = Type::getIntType();
BinaryInst* currentDimOffsetInst = pBuilder->createBinaryInst(Instruction::kMul, intType, index, strideConst);
BinaryInst* newTotalOffsetInst = pBuilder->createBinaryInst(Instruction::kAdd, intType, totalOffset, currentDimOffsetInst);
totalOffset = newTotalOffsetInst;
}
// 计算有效地址effective_address = basePointer + totalOffset
Value* effective_address = pBuilder->createBinaryInst(Instruction::kAdd, basePointer->getType(), basePointer, totalOffset);
// 创建新的 LoadInst 或 StoreInstindices 为空
Instruction* newInst = nullptr;
if (inst->isLoad()) {
newInst = pBuilder->createLoadInst(effective_address, {});
inst->replaceAllUsesWith(newInst);
} else { // StoreInst
newInst = pBuilder->createStoreInst(valueToStore, effective_address, {});
}
Instruction* oldInst = it->get();
++it;
for (size_t i = 0; i < oldInst->getNumOperands(); ++i) {
Value* operandValue = oldInst->getOperand(i);
if (operandValue) {
for (auto use_it = operandValue->getUses().begin(); use_it != operandValue->getUses().end(); ++use_it) {
if ((*use_it)->getUser() == oldInst && (*use_it)->getIndex() == i) {
operandValue->removeUse(*use_it);
break;
}
}
}
}
bb->getInstructions().erase(std::prev(it));
changed = true;
if (DEBUG) {
std::cerr << "ACE: Computed effective address:\n";
SysYPrinter::printInst(dynamic_cast<Instruction*>(effective_address));
std::cerr << "ACE: New Load/Store instruction:\n";
SysYPrinter::printInst(newInst);
std::cerr << "--------------------------------\n";
}
}
}
}
return changed;
}
} // namespace sysy

View File

@ -1,24 +1,52 @@
# src/CMakeLists.txt
# add_subdirectory 命令会负责遍历子目录并查找其内部的 CMakeLists.txt 文件
add_subdirectory(frontend)
add_subdirectory(midend)
add_subdirectory(backend/RISCv64)
# 移除 ANTLR 代码生成相关配置
# list(APPEND CMAKE_MODULE_PATH "${ANTLR_RUNTIME}/cmake")
# include(FindANTLR)
# antlr_target(SysYGen SysY.g4
# LEXER PARSER
# OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
# VISITOR
# )
# 构建 sysyc 可执行文件,链接各个模块的库
# 移除 SysYParser 库的构建(如果不需要独立库)
# add_library(SysYParser SHARED ${ANTLR_SysYGen_CXX_OUTPUTS})
# target_include_directories(SysYParser PUBLIC ${ANTLR_RUNTIME}/runtime/src)
# target_link_libraries(SysYParser PUBLIC antlr4_shared)
# 构建 sysyc 可执行文件,使用手动提供的 SysYLexer.cpp、SysYParser.cpp 等文件
add_executable(sysyc
sysyc.cpp
sysyc.cpp
SysYLexer.cpp # 手动提供的文件
SysYParser.cpp # 手动提供的文件
SysYVisitor.cpp # 手动提供的文件
IR.cpp
SysYIRGenerator.cpp
SysYIRPrinter.cpp
SysYIROptPre.cpp
SysYIRAnalyser.cpp
# DeadCodeElimination.cpp
AddressCalculationExpansion.cpp
# Mem2Reg.cpp
# Reg2Mem.cpp
RISCv64Backend.cpp
RISCv64ISel.cpp
RISCv64RegAlloc.cpp
RISCv64AsmPrinter.cpp
RISCv64Passes.cpp
)
# 链接各个模块的库
target_link_libraries(sysyc PRIVATE
frontend_lib
midend_lib
riscv64_backend_lib
antlr4_shared
)
# 设置 include 路径,包含项目顶层 include 目录
# 设置 include 路径,包含 ANTLR 运行时库和项目头文件
target_include_directories(sysyc PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/include # 项目头文件目录
${ANTLR_RUNTIME}/runtime/src # ANTLR运行时库头文件
)
${CMAKE_CURRENT_SOURCE_DIR}/include # 项目头文件目录
${ANTLR_RUNTIME}/runtime/src # ANTLR 运行时库头文件
)
# 保留 ANTLR 运行时库的链接
target_link_libraries(sysyc PRIVATE antlr4_shared)
# 保留其他编译选项
target_compile_options(sysyc PRIVATE -frtti)
# 可选:线程支持(如果需要,取消注释)
# set(THREADS_PREFER_PTHREAD_FLAG ON)
# find_package(Threads REQUIRED)
# target_link_libraries(sysyc PRIVATE Threads::Threads)

750
src/IR.cpp Normal file
View File

@ -0,0 +1,750 @@
#include "IR.h"
#include <algorithm>
#include <cassert>
#include <memory>
#include <queue>
#include <set>
#include <sstream>
#include <vector>
#include "IRBuilder.h"
/**
* @file IR.cpp
*
* @brief 定义IR相关类型与操作的源文件
*/
namespace sysy {
//===----------------------------------------------------------------------===//
// Types
//===----------------------------------------------------------------------===//
auto Type::getIntType() -> Type * {
static Type intType(kInt);
return &intType;
}
auto Type::getFloatType() -> Type * {
static Type floatType(kFloat);
return &floatType;
}
auto Type::getVoidType() -> Type * {
static Type voidType(kVoid);
return &voidType;
}
auto Type::getLabelType() -> Type * {
static Type labelType(kLabel);
return &labelType;
}
auto Type::getPointerType(Type *baseType) -> Type * {
// forward to PointerType
return PointerType::get(baseType);
}
auto Type::getFunctionType(Type *returnType, const std::vector<Type *> &paramTypes) -> Type * {
// forward to FunctionType
return FunctionType::get(returnType, paramTypes);
}
auto Type::getSize() const -> unsigned {
switch (kind) {
case kInt:
case kFloat:
return 4;
case kLabel:
case kPointer:
case kFunction:
return 8;
case kVoid:
return 0;
}
return 0;
}
PointerType* PointerType::get(Type *baseType) {
static std::map<Type *, std::unique_ptr<PointerType>> pointerTypes;
auto iter = pointerTypes.find(baseType);
if (iter != pointerTypes.end()) {
return iter->second.get();
}
auto type = new PointerType(baseType);
assert(type);
auto result = pointerTypes.emplace(baseType, type);
return result.first->second.get();
}
FunctionType*FunctionType::get(Type *returnType, const std::vector<Type *> &paramTypes) {
static std::set<std::unique_ptr<FunctionType>> functionTypes;
auto iter =
std::find_if(functionTypes.begin(), functionTypes.end(), [&](const std::unique_ptr<FunctionType> &type) -> bool {
if (returnType != type->getReturnType() ||
paramTypes.size() != static_cast<size_t>(type->getParamTypes().size())) {
return false;
}
return std::equal(paramTypes.begin(), paramTypes.end(), type->getParamTypes().begin());
});
if (iter != functionTypes.end()) {
return iter->get();
}
auto type = new FunctionType(returnType, paramTypes);
assert(type);
auto result = functionTypes.emplace(type);
return result.first->get();
}
void Value::replaceAllUsesWith(Value *value) {
for (auto &use : uses) {
use->getUser()->setOperand(use->getIndex(), value);
}
uses.clear();
}
// Implementations for static members
std::unordered_map<ConstantValueKey, ConstantValue*, ConstantValueHash, ConstantValueEqual> ConstantValue::mConstantPool;
std::unordered_map<Type*, UndefinedValue*> UndefinedValue::UndefValues;
ConstantValue* ConstantValue::get(Type* type, ConstantValVariant val) {
ConstantValueKey key = {type, val};
auto it = mConstantPool.find(key);
if (it != mConstantPool.end()) {
return it->second;
}
ConstantValue* newConstant = nullptr;
if (std::holds_alternative<int>(val)) {
newConstant = new ConstantInteger(type, std::get<int>(val));
} else if (std::holds_alternative<float>(val)) {
newConstant = new ConstantFloating(type, std::get<float>(val));
} else {
assert(false && "Unsupported ConstantValVariant type");
}
mConstantPool[key] = newConstant;
return newConstant;
}
ConstantInteger* ConstantInteger::get(Type* type, int val) {
return dynamic_cast<ConstantInteger*>(ConstantValue::get(type, val));
}
ConstantFloating* ConstantFloating::get(Type* type, float val) {
return dynamic_cast<ConstantFloating*>(ConstantValue::get(type, val));
}
UndefinedValue* UndefinedValue::get(Type* type) {
assert(!type->isVoid() && "Cannot get UndefinedValue of void type!");
auto it = UndefValues.find(type);
if (it != UndefValues.end()) {
return it->second;
}
UndefinedValue* newUndef = new UndefinedValue(type);
UndefValues[type] = newUndef;
return newUndef;
}
auto Function::getCalleesWithNoExternalAndSelf() -> std::set<Function *> {
std::set<Function *> result;
for (auto callee : callees) {
if (parent->getExternalFunctions().count(callee->getName()) == 0U && callee != this) {
result.insert(callee);
}
}
return result;
}
// 函数克隆,后续函数级优化(内联等)需要用到
Function * Function::clone(const std::string &suffix) const {
std::stringstream ss;
std::map<BasicBlock *, BasicBlock *> oldNewBlockMap;
IRBuilder builder;
auto newFunction = new Function(parent, type, name);
newFunction->getEntryBlock()->setName(blocks.front()->getName());
oldNewBlockMap.emplace(blocks.front().get(), newFunction->getEntryBlock());
auto oldBlockListIter = std::next(blocks.begin());
while (oldBlockListIter != blocks.end()) {
auto newBlock = newFunction->addBasicBlock(oldBlockListIter->get()->getName());
oldNewBlockMap.emplace(oldBlockListIter->get(), newBlock);
oldBlockListIter++;
}
for (const auto &oldNewBlockItem : oldNewBlockMap) {
auto oldBlock = oldNewBlockItem.first;
auto newBlock = oldNewBlockItem.second;
for (const auto &oldPred : oldBlock->getPredecessors()) {
newBlock->addPredecessor(oldNewBlockMap.at(oldPred));
}
for (const auto &oldSucc : oldBlock->getSuccessors()) {
newBlock->addSuccessor(oldNewBlockMap.at(oldSucc));
}
}
std::map<Value *, Value *> oldNewValueMap;
std::map<Value *, bool> isAddedToCreate;
std::map<Value *, bool> isCreated;
std::queue<Value *> toCreate;
for (const auto &oldBlock : blocks) {
for (const auto &inst : oldBlock->getInstructions()) {
isAddedToCreate.emplace(inst.get(), false);
isCreated.emplace(inst.get(), false);
}
}
for (const auto &oldBlock : blocks) {
for (const auto &inst : oldBlock->getInstructions()) {
for (const auto &valueUse : inst->getOperands()) {
auto value = valueUse->getValue();
if (oldNewValueMap.find(value) == oldNewValueMap.end()) {
auto oldAllocInst = dynamic_cast<AllocaInst *>(value);
if (oldAllocInst != nullptr) {
std::vector<Value *> dims;
for (const auto &dim : oldAllocInst->getDims()) {
dims.emplace_back(dim->getValue());
}
ss << oldAllocInst->getName() << suffix;
auto newAllocInst =
new AllocaInst(oldAllocInst->getType(), dims, oldNewBlockMap.at(oldAllocInst->getParent()), ss.str());
ss.str("");
oldNewValueMap.emplace(oldAllocInst, newAllocInst);
if (isAddedToCreate.find(oldAllocInst) == isAddedToCreate.end()) {
isAddedToCreate.emplace(oldAllocInst, true);
} else {
isAddedToCreate.at(oldAllocInst) = true;
}
if (isCreated.find(oldAllocInst) == isCreated.end()) {
isCreated.emplace(oldAllocInst, true);
} else {
isCreated.at(oldAllocInst) = true;
}
}
}
}
if (inst->getKind() == Instruction::kAlloca) {
if (oldNewValueMap.find(inst.get()) == oldNewValueMap.end()) {
auto oldAllocInst = dynamic_cast<AllocaInst *>(inst.get());
std::vector<Value *> dims;
for (const auto &dim : oldAllocInst->getDims()) {
dims.emplace_back(dim->getValue());
}
ss << oldAllocInst->getName() << suffix;
auto newAllocInst =
new AllocaInst(oldAllocInst->getType(), dims, oldNewBlockMap.at(oldAllocInst->getParent()), ss.str());
ss.str("");
oldNewValueMap.emplace(oldAllocInst, newAllocInst);
if (isAddedToCreate.find(oldAllocInst) == isAddedToCreate.end()) {
isAddedToCreate.emplace(oldAllocInst, true);
} else {
isAddedToCreate.at(oldAllocInst) = true;
}
if (isCreated.find(oldAllocInst) == isCreated.end()) {
isCreated.emplace(oldAllocInst, true);
} else {
isCreated.at(oldAllocInst) = true;
}
}
}
}
}
for (const auto &oldBlock : blocks) {
for (const auto &inst : oldBlock->getInstructions()) {
for (const auto &valueUse : inst->getOperands()) {
auto value = valueUse->getValue();
if (oldNewValueMap.find(value) == oldNewValueMap.end()) {
auto globalValue = dynamic_cast<GlobalValue *>(value);
auto constVariable = dynamic_cast<ConstantVariable *>(value);
auto constantValue = dynamic_cast<ConstantValue *>(value);
auto functionValue = dynamic_cast<Function *>(value);
if (globalValue != nullptr || constantValue != nullptr || constVariable != nullptr ||
functionValue != nullptr) {
if (functionValue == this) {
oldNewValueMap.emplace(value, newFunction);
} else {
oldNewValueMap.emplace(value, value);
}
isCreated.emplace(value, true);
isAddedToCreate.emplace(value, true);
}
}
}
}
}
for (const auto &oldBlock : blocks) {
for (const auto &inst : oldBlock->getInstructions()) {
if (inst->getKind() != Instruction::kAlloca) {
bool isReady = true;
for (const auto &use : inst->getOperands()) {
auto value = use->getValue();
if (dynamic_cast<BasicBlock *>(value) == nullptr && !isCreated.at(value)) {
isReady = false;
break;
}
}
if (isReady) {
toCreate.push(inst.get());
isAddedToCreate.at(inst.get()) = true;
}
}
}
}
while (!toCreate.empty()) {
auto inst = dynamic_cast<Instruction *>(toCreate.front());
toCreate.pop();
bool isReady = true;
for (const auto &valueUse : inst->getOperands()) {
auto value = dynamic_cast<Instruction *>(valueUse->getValue());
if (value != nullptr && !isCreated.at(value)) {
isReady = false;
break;
}
}
if (!isReady) {
toCreate.push(inst);
continue;
}
isCreated.at(inst) = true;
switch (inst->getKind()) {
case Instruction::kAdd:
case Instruction::kSub:
case Instruction::kMul:
case Instruction::kDiv:
case Instruction::kRem:
case Instruction::kICmpEQ:
case Instruction::kICmpNE:
case Instruction::kICmpLT:
case Instruction::kICmpGT:
case Instruction::kICmpLE:
case Instruction::kICmpGE:
case Instruction::kAnd:
case Instruction::kOr:
case Instruction::kFAdd:
case Instruction::kFSub:
case Instruction::kFMul:
case Instruction::kFDiv:
case Instruction::kFCmpEQ:
case Instruction::kFCmpNE:
case Instruction::kFCmpLT:
case Instruction::kFCmpGT:
case Instruction::kFCmpLE:
case Instruction::kFCmpGE: {
auto oldBinaryInst = dynamic_cast<BinaryInst *>(inst);
auto lhs = oldBinaryInst->getLhs();
auto rhs = oldBinaryInst->getRhs();
Value *newLhs;
Value *newRhs;
newLhs = oldNewValueMap[lhs];
newRhs = oldNewValueMap[rhs];
ss << oldBinaryInst->getName() << suffix;
auto newBinaryInst = new BinaryInst(oldBinaryInst->getKind(), oldBinaryInst->getType(), newLhs, newRhs,
oldNewBlockMap.at(oldBinaryInst->getParent()), ss.str());
ss.str("");
oldNewValueMap.emplace(oldBinaryInst, newBinaryInst);
break;
}
case Instruction::kNeg:
case Instruction::kNot:
case Instruction::kFNeg:
case Instruction::kFNot:
case Instruction::kItoF:
case Instruction::kFtoI: {
auto oldUnaryInst = dynamic_cast<UnaryInst *>(inst);
auto hs = oldUnaryInst->getOperand();
Value *newHs;
newHs = oldNewValueMap.at(hs);
ss << oldUnaryInst->getName() << suffix;
auto newUnaryInst = new UnaryInst(oldUnaryInst->getKind(), oldUnaryInst->getType(), newHs,
oldNewBlockMap.at(oldUnaryInst->getParent()), ss.str());
ss.str("");
oldNewValueMap.emplace(oldUnaryInst, newUnaryInst);
break;
}
case Instruction::kCall: {
auto oldCallInst = dynamic_cast<CallInst *>(inst);
std::vector<Value *> newArgumnts;
for (const auto &arg : oldCallInst->getArguments()) {
newArgumnts.emplace_back(oldNewValueMap.at(arg->getValue()));
}
ss << oldCallInst->getName() << suffix;
CallInst *newCallInst;
newCallInst =
new CallInst(oldCallInst->getCallee(), newArgumnts, oldNewBlockMap.at(oldCallInst->getParent()), ss.str());
ss.str("");
// if (oldCallInst->getCallee() != this) {
// newCallInst = new CallInst(oldCallInst->getCallee(), newArgumnts,
// oldNewBlockMap.at(oldCallInst->getParent()),
// oldCallInst->getName());
// } else {
// newCallInst = new CallInst(newFunction, newArgumnts, oldNewBlockMap.at(oldCallInst->getParent()),
// oldCallInst->getName());
// }
oldNewValueMap.emplace(oldCallInst, newCallInst);
break;
}
case Instruction::kCondBr: {
auto oldCondBrInst = dynamic_cast<CondBrInst *>(inst);
auto oldCond = oldCondBrInst->getCondition();
Value *newCond;
newCond = oldNewValueMap.at(oldCond);
auto newCondBrInst = new CondBrInst(newCond, oldNewBlockMap.at(oldCondBrInst->getThenBlock()),
oldNewBlockMap.at(oldCondBrInst->getElseBlock()), {}, {},
oldNewBlockMap.at(oldCondBrInst->getParent()));
oldNewValueMap.emplace(oldCondBrInst, newCondBrInst);
break;
}
case Instruction::kBr: {
auto oldBrInst = dynamic_cast<UncondBrInst *>(inst);
auto newBrInst =
new UncondBrInst(oldNewBlockMap.at(oldBrInst->getBlock()), {}, oldNewBlockMap.at(oldBrInst->getParent()));
oldNewValueMap.emplace(oldBrInst, newBrInst);
break;
}
case Instruction::kReturn: {
auto oldReturnInst = dynamic_cast<ReturnInst *>(inst);
auto oldRval = oldReturnInst->getReturnValue();
Value *newRval = nullptr;
if (oldRval != nullptr) {
newRval = oldNewValueMap.at(oldRval);
}
auto newReturnInst =
new ReturnInst(newRval, oldNewBlockMap.at(oldReturnInst->getParent()), oldReturnInst->getName());
oldNewValueMap.emplace(oldReturnInst, newReturnInst);
break;
}
case Instruction::kAlloca: {
assert(false);
}
case Instruction::kLoad: {
auto oldLoadInst = dynamic_cast<LoadInst *>(inst);
auto oldPointer = oldLoadInst->getPointer();
Value *newPointer;
newPointer = oldNewValueMap.at(oldPointer);
std::vector<Value *> newIndices;
for (const auto &index : oldLoadInst->getIndices()) {
newIndices.emplace_back(oldNewValueMap.at(index->getValue()));
}
ss << oldLoadInst->getName() << suffix;
auto newLoadInst = new LoadInst(newPointer, newIndices, oldNewBlockMap.at(oldLoadInst->getParent()), ss.str());
ss.str("");
oldNewValueMap.emplace(oldLoadInst, newLoadInst);
break;
}
case Instruction::kStore: {
auto oldStoreInst = dynamic_cast<StoreInst *>(inst);
auto oldPointer = oldStoreInst->getPointer();
auto oldValue = oldStoreInst->getValue();
Value *newPointer;
Value *newValue;
std::vector<Value *> newIndices;
newPointer = oldNewValueMap.at(oldPointer);
newValue = oldNewValueMap.at(oldValue);
for (const auto &index : oldStoreInst->getIndices()) {
newIndices.emplace_back(oldNewValueMap.at(index->getValue()));
}
auto newStoreInst = new StoreInst(newValue, newPointer, newIndices,
oldNewBlockMap.at(oldStoreInst->getParent()), oldStoreInst->getName());
oldNewValueMap.emplace(oldStoreInst, newStoreInst);
break;
}
case Instruction::kLa: {
auto oldLaInst = dynamic_cast<LaInst *>(inst);
auto oldPointer = oldLaInst->getPointer();
Value *newPointer;
std::vector<Value *> newIndices;
newPointer = oldNewValueMap.at(oldPointer);
for (const auto &index : oldLaInst->getIndices()) {
newIndices.emplace_back(oldNewValueMap.at(index->getValue()));
}
ss << oldLaInst->getName() << suffix;
auto newLaInst = new LaInst(newPointer, newIndices, oldNewBlockMap.at(oldLaInst->getParent()), ss.str());
ss.str("");
oldNewValueMap.emplace(oldLaInst, newLaInst);
break;
}
case Instruction::kGetSubArray: {
auto oldGetSubArrayInst = dynamic_cast<GetSubArrayInst *>(inst);
auto oldFather = oldGetSubArrayInst->getFatherArray();
auto oldChild = oldGetSubArrayInst->getChildArray();
Value *newFather;
Value *newChild;
std::vector<Value *> newIndices;
newFather = oldNewValueMap.at(oldFather);
newChild = oldNewValueMap.at(oldChild);
for (const auto &index : oldGetSubArrayInst->getIndices()) {
newIndices.emplace_back(oldNewValueMap.at(index->getValue()));
}
ss << oldGetSubArrayInst->getName() << suffix;
auto newGetSubArrayInst =
new GetSubArrayInst(dynamic_cast<LVal *>(newFather), dynamic_cast<LVal *>(newChild), newIndices,
oldNewBlockMap.at(oldGetSubArrayInst->getParent()), ss.str());
ss.str("");
oldNewValueMap.emplace(oldGetSubArrayInst, newGetSubArrayInst);
break;
}
case Instruction::kMemset: {
auto oldMemsetInst = dynamic_cast<MemsetInst *>(inst);
auto oldPointer = oldMemsetInst->getPointer();
auto oldValue = oldMemsetInst->getValue();
Value *newPointer;
Value *newValue;
newPointer = oldNewValueMap.at(oldPointer);
newValue = oldNewValueMap.at(oldValue);
auto newMemsetInst = new MemsetInst(newPointer, oldMemsetInst->getBegin(), oldMemsetInst->getSize(), newValue,
oldNewBlockMap.at(oldMemsetInst->getParent()), oldMemsetInst->getName());
oldNewValueMap.emplace(oldMemsetInst, newMemsetInst);
break;
}
case Instruction::kInvalid:
case Instruction::kPhi: {
break;
}
default:
assert(false);
}
for (const auto &userUse : inst->getUses()) {
auto user = userUse->getUser();
if (!isAddedToCreate.at(user)) {
toCreate.push(user);
isAddedToCreate.at(user) = true;
}
}
}
for (const auto &oldBlock : blocks) {
auto newBlock = oldNewBlockMap.at(oldBlock.get());
builder.setPosition(newBlock, newBlock->end());
for (const auto &inst : oldBlock->getInstructions()) {
builder.insertInst(dynamic_cast<Instruction *>(oldNewValueMap.at(inst.get())));
}
}
for (const auto &param : blocks.front()->getArguments()) {
newFunction->getEntryBlock()->insertArgument(dynamic_cast<AllocaInst *>(oldNewValueMap.at(param)));
}
return newFunction;
}
/**
* 设置操作数
*/
void User::setOperand(unsigned index, Value *value) {
assert(index < getNumOperands());
operands[index]->setValue(value);
value->addUse(operands[index]);
}
/**
* 替换操作数
*/
void User::replaceOperand(unsigned index, Value *value) {
assert(index < getNumOperands());
auto &use = operands[index];
use->getValue()->removeUse(use);
use->setValue(value);
value->addUse(use);
}
/**
* phi相关函数
*/
Value* PhiInst::getvalfromBlk(BasicBlock* blk){
refreshB2VMap();
if( blk2val.find(blk) != blk2val.end()) {
return blk2val.at(blk);
}
return nullptr;
}
BasicBlock* PhiInst::getBlkfromVal(Value* val){
// 返回第一个值对应的基本块
for(unsigned i = 0; i < vsize; i++) {
if(getValue(i) == val) {
return getBlock(i);
}
}
return nullptr;
}
void PhiInst::delValue(Value* val){
//根据value删除对应的基本块和值
unsigned i = 0;
BasicBlock* blk = getBlkfromVal(val);
for(i = 0; i < vsize; i++) {
if(getValue(i) == val) {
break;
}
}
removeOperand(2 * i + 1); // 删除blk
removeOperand(2 * i); // 删除val
vsize--;
blk2val.erase(blk); // 删除blk2val映射
}
void PhiInst::delBlk(BasicBlock* blk){
//根据Blk删除对应的基本块和值
unsigned i = 0;
Value* val = getvalfromBlk(blk);
for(i = 0; i < vsize; i++) {
if(getBlock(i) == blk) {
break;
}
}
removeOperand(2 * i + 1); // 删除blk
removeOperand(2 * i); // 删除val
vsize--;
blk2val.erase(blk); // 删除blk2val映射
}
void PhiInst::replaceBlk(BasicBlock* newBlk, unsigned k){
refreshB2VMap();
Value* val = blk2val.at(getBlock(k));
// 替换基本块
setOperand(2 * k + 1, newBlk);
// 替换blk2val映射
blk2val.erase(getBlock(k));
blk2val.emplace(newBlk, val);
}
void PhiInst::replaceold2new(BasicBlock* oldBlk, BasicBlock* newBlk){
refreshB2VMap();
Value* val = blk2val.at(oldBlk);
// 替换基本块
delBlk(oldBlk);
addIncoming(val, newBlk);
}
void PhiInst::refreshB2VMap(){
blk2val.clear();
for(unsigned i = 0; i < vsize; i++) {
blk2val.emplace(getBlock(i), getValue(i));
}
}
CallInst::CallInst(Function *callee, const std::vector<Value *> &args, BasicBlock *parent, const std::string &name)
: Instruction(kCall, callee->getReturnType(), parent, name) {
addOperand(callee);
for (auto arg : args) {
addOperand(arg);
}
}
/**
* 获取被调用函数的指针
*/
Function * CallInst::getCallee() const { return dynamic_cast<Function *>(getOperand(0)); }
/**
* 获取变量指针
*/
auto SymbolTable::getVariable(const std::string &name) const -> User * {
auto node = curNode;
while (node != nullptr) {
auto iter = node->varList.find(name);
if (iter != node->varList.end()) {
return iter->second;
}
node = node->pNode;
}
return nullptr;
}
/**
* 添加变量到符号表
*/
auto SymbolTable::addVariable(const std::string &name, User *variable) -> User * {
User *result = nullptr;
if (curNode != nullptr) {
std::stringstream ss;
auto iter = variableIndex.find(name);
if (iter != variableIndex.end()) {
ss << name << iter->second ;
iter->second += 1;
} else {
variableIndex.emplace(name, 1);
ss << name << 0 ;
}
variable->setName(ss.str());
curNode->varList.emplace(name, variable);
auto global = dynamic_cast<GlobalValue *>(variable);
auto constvar = dynamic_cast<ConstantVariable *>(variable);
if (global != nullptr) {
globals.emplace_back(global);
} else if (constvar != nullptr) {
consts.emplace_back(constvar);
}
result = variable;
}
return result;
}
/**
* 获取全局变量
*/
auto SymbolTable::getGlobals() -> std::vector<std::unique_ptr<GlobalValue>> & { return globals; }
/**
* 获取常量
*/
auto SymbolTable::getConsts() const -> const std::vector<std::unique_ptr<ConstantVariable>> & { return consts; }
/**
* 进入新的作用域
*/
void SymbolTable::enterNewScope() {
auto newNode = new SymbolTableNode;
nodeList.emplace_back(newNode);
if (curNode != nullptr) {
curNode->children.emplace_back(newNode);
}
newNode->pNode = curNode;
curNode = newNode;
}
/**
* 进入全局作用域
*/
void SymbolTable::enterGlobalScope() { curNode = nodeList.front().get(); }
/**
* 离开作用域
*/
void SymbolTable::leaveScope() { curNode = curNode->pNode; }
/**
* 是否位于全局作用域
*/
auto SymbolTable::isInGlobalScope() const -> bool { return curNode->pNode == nullptr; }
/**
*移动指令
*/
auto BasicBlock::moveInst(iterator sourcePos, iterator targetPos, BasicBlock *block) -> iterator {
auto inst = sourcePos->release();
inst->setParent(block);
block->instructions.emplace(targetPos, inst);
return instructions.erase(sourcePos);
}
} // namespace sysy

View File

@ -1,22 +1,15 @@
#include "RISCv64AsmPrinter.h"
#include "RISCv64ISel.h"
#include <stdexcept>
#include <sstream>
#include <iostream>
namespace sysy {
// 检查是否为内存加载/存储指令,以处理特殊的打印格式
bool isMemoryOp(RVOpcodes opcode) {
switch (opcode) {
// --- 整数加载/存储 (原有逻辑) ---
case RVOpcodes::LB: case RVOpcodes::LH: case RVOpcodes::LW: case RVOpcodes::LD:
case RVOpcodes::LBU: case RVOpcodes::LHU: case RVOpcodes::LWU:
case RVOpcodes::SB: case RVOpcodes::SH: case RVOpcodes::SW: case RVOpcodes::SD:
case RVOpcodes::FLW:
case RVOpcodes::FSW:
// 如果未来支持双精度也在这里添加FLD/FSD
// case RVOpcodes::FLD:
// case RVOpcodes::FSD:
return true;
default:
return false;
@ -29,12 +22,57 @@ void RISCv64AsmPrinter::run(std::ostream& os, bool debug) {
OS = &os;
*OS << ".globl " << MFunc->getName() << "\n";
*OS << MFunc->getName() << ":\n";
printPrologue();
for (auto& mbb : MFunc->getBlocks()) {
printBasicBlock(mbb.get(), debug);
}
}
void RISCv64AsmPrinter::printPrologue() {
StackFrameInfo& frame_info = MFunc->getFrameInfo();
// 序言需要为保存ra和s0预留16字节
int total_stack_size = frame_info.locals_size + frame_info.spill_size + 16;
int aligned_stack_size = (total_stack_size + 15) & ~15;
frame_info.total_size = aligned_stack_size;
if (aligned_stack_size > 0) {
*OS << " addi sp, sp, -" << aligned_stack_size << "\n";
*OS << " sd ra, " << (aligned_stack_size - 8) << "(sp)\n";
*OS << " sd s0, " << (aligned_stack_size - 16) << "(sp)\n";
*OS << " addi s0, sp, " << aligned_stack_size << "\n";
}
// 忠实还原保存函数入口参数的逻辑
Function* F = MFunc->getFunc();
if (F && F->getEntryBlock()) {
int arg_idx = 0;
RISCv64ISel* isel = MFunc->getISel();
for (AllocaInst* alloca_for_param : F->getEntryBlock()->getArguments()) {
if (arg_idx >= 8) break;
unsigned vreg = isel->getVReg(alloca_for_param);
if (frame_info.alloca_offsets.count(vreg)) {
int offset = frame_info.alloca_offsets.at(vreg);
auto arg_reg = static_cast<PhysicalReg>(static_cast<int>(PhysicalReg::A0) + arg_idx);
*OS << " sw " << regToString(arg_reg) << ", " << offset << "(s0)\n";
}
arg_idx++;
}
}
}
void RISCv64AsmPrinter::printEpilogue() {
int aligned_stack_size = MFunc->getFrameInfo().total_size;
if (aligned_stack_size > 0) {
*OS << " ld ra, " << (aligned_stack_size - 8) << "(sp)\n";
*OS << " ld s0, " << (aligned_stack_size - 16) << "(sp)\n";
*OS << " addi sp, sp, " << aligned_stack_size << "\n";
}
}
void RISCv64AsmPrinter::printBasicBlock(MachineBasicBlock* mbb, bool debug) {
if (!mbb->getName().empty()) {
*OS << mbb->getName() << ":\n";
@ -46,6 +84,9 @@ void RISCv64AsmPrinter::printBasicBlock(MachineBasicBlock* mbb, bool debug) {
void RISCv64AsmPrinter::printInstruction(MachineInstr* instr, bool debug) {
auto opcode = instr->getOpcode();
if (opcode == RVOpcodes::RET) {
printEpilogue();
}
if (opcode == RVOpcodes::LABEL) {
// 标签直接打印,不加缩进
@ -61,7 +102,7 @@ void RISCv64AsmPrinter::printInstruction(MachineInstr* instr, bool debug) {
case RVOpcodes::ADD: *OS << "add "; break; case RVOpcodes::ADDI: *OS << "addi "; break;
case RVOpcodes::ADDW: *OS << "addw "; break; case RVOpcodes::ADDIW: *OS << "addiw "; break;
case RVOpcodes::SUB: *OS << "sub "; break; case RVOpcodes::SUBW: *OS << "subw "; break;
case RVOpcodes::MUL: *OS << "mul "; break; case RVOpcodes::MULW: *OS << "mulw "; break; case RVOpcodes::MULH: *OS << "mulh "; break;
case RVOpcodes::MUL: *OS << "mul "; break; case RVOpcodes::MULW: *OS << "mulw "; break;
case RVOpcodes::DIV: *OS << "div "; break; case RVOpcodes::DIVW: *OS << "divw "; break;
case RVOpcodes::REM: *OS << "rem "; break; case RVOpcodes::REMW: *OS << "remw "; break;
case RVOpcodes::XOR: *OS << "xor "; break; case RVOpcodes::XORI: *OS << "xori "; break;
@ -80,9 +121,7 @@ void RISCv64AsmPrinter::printInstruction(MachineInstr* instr, bool debug) {
case RVOpcodes::LHU: *OS << "lhu "; break; case RVOpcodes::LBU: *OS << "lbu "; break;
case RVOpcodes::SW: *OS << "sw "; break; case RVOpcodes::SH: *OS << "sh "; break;
case RVOpcodes::SB: *OS << "sb "; break; case RVOpcodes::LD: *OS << "ld "; break;
case RVOpcodes::SD: *OS << "sd "; break; case RVOpcodes::FLW: *OS << "flw "; break;
case RVOpcodes::FSW: *OS << "fsw "; break; case RVOpcodes::FLD: *OS << "fld "; break;
case RVOpcodes::FSD: *OS << "fsd "; break;
case RVOpcodes::SD: *OS << "sd "; break;
case RVOpcodes::J: *OS << "j "; break; case RVOpcodes::JAL: *OS << "jal "; break;
case RVOpcodes::JALR: *OS << "jalr "; break; case RVOpcodes::RET: *OS << "ret"; break;
case RVOpcodes::BEQ: *OS << "beq "; break; case RVOpcodes::BNE: *OS << "bne "; break;
@ -91,64 +130,24 @@ void RISCv64AsmPrinter::printInstruction(MachineInstr* instr, bool debug) {
case RVOpcodes::LI: *OS << "li "; break; case RVOpcodes::LA: *OS << "la "; break;
case RVOpcodes::MV: *OS << "mv "; break; case RVOpcodes::NEG: *OS << "neg "; break;
case RVOpcodes::NEGW: *OS << "negw "; break; case RVOpcodes::SEQZ: *OS << "seqz "; break;
case RVOpcodes::SNEZ: *OS << "snez "; break;
case RVOpcodes::FADD_S: *OS << "fadd.s "; break;
case RVOpcodes::FSUB_S: *OS << "fsub.s "; break;
case RVOpcodes::FMUL_S: *OS << "fmul.s "; break;
case RVOpcodes::FDIV_S: *OS << "fdiv.s "; break;
case RVOpcodes::FNEG_S: *OS << "fneg.s "; break;
case RVOpcodes::FEQ_S: *OS << "feq.s "; break;
case RVOpcodes::FLT_S: *OS << "flt.s "; break;
case RVOpcodes::FLE_S: *OS << "fle.s "; break;
case RVOpcodes::FCVT_S_W: *OS << "fcvt.s.w "; break;
case RVOpcodes::FCVT_W_S: *OS << "fcvt.w.s "; break;
case RVOpcodes::FCVT_W_S_RTZ: *OS << "fcvt.w.s "; break;
case RVOpcodes::FMV_S: *OS << "fmv.s "; break;
case RVOpcodes::FMV_W_X: *OS << "fmv.w.x "; break;
case RVOpcodes::FMV_X_W: *OS << "fmv.x.w "; break;
case RVOpcodes::CALL: { // 为CALL指令添加特殊处理逻辑
*OS << "call ";
// 遍历所有操作数,只寻找并打印函数名标签
for (const auto& op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_LABEL) {
printOperand(op.get());
break; // 找到标签后即可退出
}
}
*OS << "\n";
return; // 处理完毕,直接返回,不再执行后续的通用操作数打印
}
case RVOpcodes::SNEZ: *OS << "snez "; break;
case RVOpcodes::CALL: *OS << "call "; break;
case RVOpcodes::LABEL:
// printOperand(instr->getOperands()[0].get());
// *OS << ":";
break;
case RVOpcodes::FRAME_LOAD_W:
case RVOpcodes::FRAME_LOAD:
// It should have been eliminated by RegAlloc
if (!debug) throw std::runtime_error("FRAME pseudo-instruction not eliminated before AsmPrinter");
*OS << "frame_load_w "; break;
case RVOpcodes::FRAME_LOAD_D:
*OS << "frame_load "; break;
case RVOpcodes::FRAME_STORE:
// It should have been eliminated by RegAlloc
if (!debug) throw std::runtime_error("FRAME pseudo-instruction not eliminated before AsmPrinter");
*OS << "frame_load_d "; break;
case RVOpcodes::FRAME_STORE_W:
// It should have been eliminated by RegAlloc
if (!debug) throw std::runtime_error("FRAME pseudo-instruction not eliminated before AsmPrinter");
*OS << "frame_store_w "; break;
case RVOpcodes::FRAME_STORE_D:
// It should have been eliminated by RegAlloc
if (!debug) throw std::runtime_error("FRAME pseudo-instruction not eliminated before AsmPrinter");
*OS << "frame_store_d "; break;
*OS << "frame_store "; break;
case RVOpcodes::FRAME_ADDR:
// It should have been eliminated by RegAlloc
if (!debug) throw std::runtime_error("FRAME pseudo-instruction not eliminated before AsmPrinter");
*OS << "frame_addr "; break;
case RVOpcodes::FRAME_LOAD_F:
if (!debug) throw std::runtime_error("FRAME_LOAD_F not eliminated before AsmPrinter");
*OS << "frame_load_f "; break;
case RVOpcodes::FRAME_STORE_F:
if (!debug) throw std::runtime_error("FRAME_STORE_F not eliminated before AsmPrinter");
*OS << "frame_store_f "; break;
case RVOpcodes::PSEUDO_KEEPALIVE:
if (!debug) throw std::runtime_error("PSEUDO_KEEPALIVE not eliminated before AsmPrinter");
*OS << "keepalive "; break;
default:
throw std::runtime_error("Unknown opcode in AsmPrinter");
}
@ -238,30 +237,4 @@ std::string RISCv64AsmPrinter::regToString(PhysicalReg reg) {
}
}
std::string RISCv64AsmPrinter::formatInstr(const MachineInstr* instr) {
if (!instr) return "(null instr)";
// 使用 stringstream 作为临时的输出目标
std::stringstream ss;
// 关键: 临时将类成员 'OS' 指向我们的 stringstream
std::ostream* old_os = this->OS;
this->OS = &ss;
// 修正: 调用正确的内部打印函数 printMachineInstr
printInstruction(const_cast<MachineInstr*>(instr), false);
// 恢复旧的 ostream 指针
this->OS = old_os;
// 获取stringstream的内容并做一些清理
std::string result = ss.str();
size_t endpos = result.find_last_not_of(" \t\n\r");
if (std::string::npos != endpos) {
result = result.substr(0, endpos + 1);
}
return result;
}
} // namespace sysy

92
src/RISCv64Backend.cpp Normal file
View File

@ -0,0 +1,92 @@
#include "RISCv64Backend.h"
#include "RISCv64ISel.h"
#include "RISCv64RegAlloc.h"
#include "RISCv64AsmPrinter.h"
#include "RISCv64Passes.h" // 包含优化Pass的头文件
#include <sstream>
namespace sysy {
// 顶层入口
std::string RISCv64CodeGen::code_gen() {
return module_gen();
}
// 模块级代码生成
std::string RISCv64CodeGen::module_gen() {
std::stringstream ss;
// 1. 处理全局变量 (.data段)
if (!module->getGlobals().empty()) {
ss << ".data\n";
for (const auto& global : module->getGlobals()) {
ss << ".globl " << global->getName() << "\n";
ss << global->getName() << ":\n";
const auto& init_values = global->getInitValues();
for (size_t i = 0; i < init_values.getValues().size(); ++i) {
auto val = init_values.getValues()[i];
auto count = init_values.getNumbers()[i];
if (auto constant = dynamic_cast<ConstantValue*>(val)) {
for (unsigned j = 0; j < count; ++j) {
if (constant->isInt()) {
ss << " .word " << constant->getInt() << "\n";
} else {
float f = constant->getFloat();
uint32_t float_bits = *(uint32_t*)&f;
ss << " .word " << float_bits << "\n";
}
}
}
}
}
}
// 2. 处理函数 (.text段)
if (!module->getFunctions().empty()) {
ss << ".text\n";
for (const auto& func_pair : module->getFunctions()) {
if (func_pair.second.get()) {
ss << function_gen(func_pair.second.get());
}
}
}
return ss.str();
}
// function_gen 现在是包含具体优化名称的、完整的处理流水线
std::string RISCv64CodeGen::function_gen(Function* func) {
// === 完整的后端处理流水线 ===
// 阶段 1: 指令选择 (sysy::IR -> LLIR with virtual registers)
RISCv64ISel isel;
std::unique_ptr<MachineFunction> mfunc = isel.runOnFunction(func);
std::stringstream ss1;
RISCv64AsmPrinter printer1(mfunc.get());
printer1.run(ss1, true);
// 阶段 2: 指令调度 (Instruction Scheduling)
PreRA_Scheduler scheduler;
scheduler.runOnMachineFunction(mfunc.get());
// 阶段 3: 物理寄存器分配 (Register Allocation)
RISCv64RegAlloc reg_alloc(mfunc.get());
reg_alloc.run();
// 阶段 4: 窥孔优化 (Peephole Optimization)
PeepholeOptimizer peephole;
peephole.runOnMachineFunction(mfunc.get());
// 阶段 5: 局部指令调度 (Local Scheduling)
PostRA_Scheduler local_scheduler;
local_scheduler.runOnMachineFunction(mfunc.get());
// 阶段 6: 代码发射 (Code Emission)
std::stringstream ss;
RISCv64AsmPrinter printer(mfunc.get());
printer.run(ss);
if (DEBUG) ss << ss1.str(); // 将指令选择阶段的结果也包含在最终输出中
return ss.str();
}
} // namespace sysy

989
src/RISCv64ISel.cpp Normal file
View File

@ -0,0 +1,989 @@
#include "RISCv64ISel.h"
#include <stdexcept>
#include <set>
#include <functional>
#include <cmath> // For std::fabs
#include <limits> // For std::numeric_limits
#include <iostream>
namespace sysy {
// DAG节点定义 (内部实现)
struct RISCv64ISel::DAGNode {
enum NodeKind { CONSTANT, LOAD, STORE, BINARY, CALL, RETURN, BRANCH, ALLOCA_ADDR, UNARY, MEMSET };
NodeKind kind;
Value* value = nullptr;
std::vector<DAGNode*> operands;
std::vector<DAGNode*> users;
DAGNode(NodeKind k) : kind(k) {}
};
RISCv64ISel::RISCv64ISel() : vreg_counter(0), local_label_counter(0) {}
// 为一个IR Value获取或分配一个新的虚拟寄存器
unsigned RISCv64ISel::getVReg(Value* val) {
if (!val) {
throw std::runtime_error("Cannot get vreg for a null Value.");
}
if (vreg_map.find(val) == vreg_map.end()) {
if (vreg_counter == 0) {
vreg_counter = 1; // vreg 0 保留
}
vreg_map[val] = vreg_counter++;
}
return vreg_map.at(val);
}
// 主入口函数
std::unique_ptr<MachineFunction> RISCv64ISel::runOnFunction(Function* func) {
F = func;
if (!F) return nullptr;
MFunc = std::make_unique<MachineFunction>(F, this);
vreg_map.clear();
bb_map.clear();
vreg_counter = 0;
local_label_counter = 0;
select();
return std::move(MFunc);
}
// 指令选择主流程
void RISCv64ISel::select() {
for (const auto& bb_ptr : F->getBasicBlocks()) {
auto mbb = std::make_unique<MachineBasicBlock>(bb_ptr->getName(), MFunc.get());
bb_map[bb_ptr.get()] = mbb.get();
MFunc->addBlock(std::move(mbb));
}
if (F->getEntryBlock()) {
for (auto* arg_alloca : F->getEntryBlock()->getArguments()) {
getVReg(arg_alloca);
}
}
for (const auto& bb_ptr : F->getBasicBlocks()) {
selectBasicBlock(bb_ptr.get());
}
for (const auto& bb_ptr : F->getBasicBlocks()) {
CurMBB = bb_map.at(bb_ptr.get());
for (auto succ : bb_ptr->getSuccessors()) {
CurMBB->successors.push_back(bb_map.at(succ));
}
for (auto pred : bb_ptr->getPredecessors()) {
CurMBB->predecessors.push_back(bb_map.at(pred));
}
}
}
// 处理单个基本块
void RISCv64ISel::selectBasicBlock(BasicBlock* bb) {
CurMBB = bb_map.at(bb);
auto dag = build_dag(bb);
if (DEBUG) { // 使用 DEBUG 宏或变量来控制是否打印
print_dag(dag, bb->getName());
}
std::map<Value*, DAGNode*> value_to_node;
for(const auto& node : dag) {
if (node->value) {
value_to_node[node->value] = node.get();
}
}
std::set<DAGNode*> selected_nodes;
std::function<void(DAGNode*)> select_recursive =
[&](DAGNode* node) {
if (DEEPDEBUG) {
std::cout << "[DEEPDEBUG] select_recursive: Visiting node with kind: " << node->kind
<< " (Value: " << (node->value ? node->value->getName() : "null") << ")" << std::endl;
}
if (!node || selected_nodes.count(node)) return;
for (auto operand : node->operands) {
select_recursive(operand);
}
selectNode(node);
selected_nodes.insert(node);
};
for (const auto& inst_ptr : bb->getInstructions()) {
DAGNode* node_to_select = nullptr;
if (value_to_node.count(inst_ptr.get())) {
node_to_select = value_to_node.at(inst_ptr.get());
} else {
for(const auto& node : dag) {
if(node->value == inst_ptr.get()) {
node_to_select = node.get();
break;
}
}
}
if(node_to_select) {
select_recursive(node_to_select);
}
}
}
// 核心函数为DAG节点选择并生成MachineInstr (已修复和增强的完整版本)
void RISCv64ISel::selectNode(DAGNode* node) {
// 调用者select_recursive已经保证了操作数节点会先于当前节点被选择。
// 因此,这里我们只处理当前节点。
switch (node->kind) {
// [V2优点] 采纳“延迟物化”Late Materialization思想。
// 这两个节点仅作为标记不直接生成指令。它们的目的是在DAG中保留类型信息。
// 加载其值的责任被转移给了使用它们的父节点如STORE, BINARY等
// 这修复了之前版本中“使用未初始化虚拟寄存器”的根本性bug。
case DAGNode::CONSTANT:
case DAGNode::ALLOCA_ADDR:
if (node->value) {
// 确保它有一个关联的虚拟寄存器即可,不生成代码。
getVReg(node->value);
}
break;
case DAGNode::LOAD: {
auto dest_vreg = getVReg(node->value);
Value* ptr_val = node->operands[0]->value;
// [V1设计保留] 对于从栈变量加载,继续使用伪指令 FRAME_LOAD。
// 这种设计将栈帧布局的具体计算推迟到后续的 `eliminateFrameIndices` 阶段,保持了模块化。
if (auto alloca = dynamic_cast<AllocaInst*>(ptr_val)) {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::FRAME_LOAD);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(getVReg(alloca)));
CurMBB->addInstruction(std::move(instr));
} else if (auto global = dynamic_cast<GlobalValue*>(ptr_val)) {
// 对于全局变量,先用 la 加载其地址,再用 lw 加载其值。
auto addr_vreg = getNewVReg();
auto la = std::make_unique<MachineInstr>(RVOpcodes::LA);
la->addOperand(std::make_unique<RegOperand>(addr_vreg));
la->addOperand(std::make_unique<LabelOperand>(global->getName()));
CurMBB->addInstruction(std::move(la));
auto lw = std::make_unique<MachineInstr>(RVOpcodes::LW);
lw->addOperand(std::make_unique<RegOperand>(dest_vreg));
lw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(addr_vreg),
std::make_unique<ImmOperand>(0)
));
CurMBB->addInstruction(std::move(lw));
} else {
// 对于已经在虚拟寄存器中的指针地址,直接通过该地址加载。
auto ptr_vreg = getVReg(ptr_val);
auto lw = std::make_unique<MachineInstr>(RVOpcodes::LW);
lw->addOperand(std::make_unique<RegOperand>(dest_vreg));
lw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(ptr_vreg),
std::make_unique<ImmOperand>(0)
));
CurMBB->addInstruction(std::move(lw));
}
break;
}
case DAGNode::STORE: {
Value* val_to_store = node->operands[0]->value;
Value* ptr_val = node->operands[1]->value;
// [V2优点] 在STORE节点内部负责加载作为源的常量。
// 如果要存储的值是一个常量,就在这里生成 `li` 指令加载它。
if (auto val_const = dynamic_cast<ConstantValue*>(val_to_store)) {
if (DEBUG) {
std::cout << "[DEBUG] selectNode-BINARY: Found constant operand with value " << val_const->getInt()
<< ". Generating LI instruction." << std::endl;
}
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(getVReg(val_const)));
li->addOperand(std::make_unique<ImmOperand>(val_const->getInt()));
CurMBB->addInstruction(std::move(li));
}
auto val_vreg = getVReg(val_to_store);
// [V1设计保留] 同样,对于向栈变量的存储,使用 FRAME_STORE 伪指令。
if (auto alloca = dynamic_cast<AllocaInst*>(ptr_val)) {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::FRAME_STORE);
instr->addOperand(std::make_unique<RegOperand>(val_vreg));
instr->addOperand(std::make_unique<RegOperand>(getVReg(alloca)));
CurMBB->addInstruction(std::move(instr));
} else if (auto global = dynamic_cast<GlobalValue*>(ptr_val)) {
// 向全局变量存储。
auto addr_vreg = getNewVReg();
auto la = std::make_unique<MachineInstr>(RVOpcodes::LA);
la->addOperand(std::make_unique<RegOperand>(addr_vreg));
la->addOperand(std::make_unique<LabelOperand>(global->getName()));
CurMBB->addInstruction(std::move(la));
auto sw = std::make_unique<MachineInstr>(RVOpcodes::SW);
sw->addOperand(std::make_unique<RegOperand>(val_vreg));
sw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(addr_vreg),
std::make_unique<ImmOperand>(0)
));
CurMBB->addInstruction(std::move(sw));
} else {
// 向一个指针(存储在虚拟寄存器中)指向的地址存储。
auto ptr_vreg = getVReg(ptr_val);
auto sw = std::make_unique<MachineInstr>(RVOpcodes::SW);
sw->addOperand(std::make_unique<RegOperand>(val_vreg));
sw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(ptr_vreg),
std::make_unique<ImmOperand>(0)
));
CurMBB->addInstruction(std::move(sw));
}
break;
}
case DAGNode::BINARY: {
auto bin = dynamic_cast<BinaryInst*>(node->value);
Value* lhs = bin->getLhs();
Value* rhs = bin->getRhs();
if (bin->getKind() == BinaryInst::kAdd) {
Value* base = nullptr;
Value* offset = nullptr;
// [修改] 扩展基地址的判断,使其可以识别 AllocaInst 或 GlobalValue
if (dynamic_cast<AllocaInst*>(lhs) || dynamic_cast<GlobalValue*>(lhs)) {
base = lhs;
offset = rhs;
} else if (dynamic_cast<AllocaInst*>(rhs) || dynamic_cast<GlobalValue*>(rhs)) {
base = rhs;
offset = lhs;
}
// 如果成功匹配到地址计算模式
if (base) {
// 1. 先为偏移量加载常量(如果它是常量的话)
if (auto const_offset = dynamic_cast<ConstantValue*>(offset)) {
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(getVReg(const_offset)));
li->addOperand(std::make_unique<ImmOperand>(const_offset->getInt()));
CurMBB->addInstruction(std::move(li));
}
// 2. [修改] 根据基地址的类型,生成不同的指令来获取基地址
auto base_addr_vreg = getNewVReg(); // 创建一个新的临时vreg来存放基地址
// 情况一:基地址是局部栈变量
if (auto alloca_base = dynamic_cast<AllocaInst*>(base)) {
auto frame_addr_instr = std::make_unique<MachineInstr>(RVOpcodes::FRAME_ADDR);
frame_addr_instr->addOperand(std::make_unique<RegOperand>(base_addr_vreg));
frame_addr_instr->addOperand(std::make_unique<RegOperand>(getVReg(alloca_base)));
CurMBB->addInstruction(std::move(frame_addr_instr));
}
// 情况二:基地址是全局变量
else if (auto global_base = dynamic_cast<GlobalValue*>(base)) {
auto la_instr = std::make_unique<MachineInstr>(RVOpcodes::LA);
la_instr->addOperand(std::make_unique<RegOperand>(base_addr_vreg));
la_instr->addOperand(std::make_unique<LabelOperand>(global_base->getName()));
CurMBB->addInstruction(std::move(la_instr));
}
// 3. 生成真正的add指令计算最终地址这部分逻辑保持不变
auto final_addr_vreg = getVReg(bin); // 这是整个二元运算的结果vreg
auto offset_vreg = getVReg(offset);
auto add_instr = std::make_unique<MachineInstr>(RVOpcodes::ADD); // 指针运算是64位
add_instr->addOperand(std::make_unique<RegOperand>(final_addr_vreg));
add_instr->addOperand(std::make_unique<RegOperand>(base_addr_vreg));
add_instr->addOperand(std::make_unique<RegOperand>(offset_vreg));
CurMBB->addInstruction(std::move(add_instr));
return; // 地址计算处理完毕,直接返回
}
}
// [V2优点] 在BINARY节点内部按需加载常量操作数。
auto load_val_if_const = [&](Value* val) {
if (auto c = dynamic_cast<ConstantValue*>(val)) {
if (DEBUG) {
std::cout << "[DEBUG] selectNode-BINARY: Found constant operand with value " << c->getInt()
<< ". Generating LI instruction." << std::endl;
}
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(getVReg(c)));
li->addOperand(std::make_unique<ImmOperand>(c->getInt()));
CurMBB->addInstruction(std::move(li));
}
};
// 检查是否能应用立即数优化。
bool rhs_is_imm_opt = false;
if (auto rhs_const = dynamic_cast<ConstantValue*>(rhs)) {
if (bin->getKind() == BinaryInst::kAdd && rhs_const->getInt() >= -2048 && rhs_const->getInt() < 2048) {
rhs_is_imm_opt = true;
}
}
// 仅在不能作为立即数操作数时才需要提前加载。
load_val_if_const(lhs);
if (!rhs_is_imm_opt) {
load_val_if_const(rhs);
}
auto dest_vreg = getVReg(bin);
auto lhs_vreg = getVReg(lhs);
// [V2优点] 融合 ADDIW 优化。
if (rhs_is_imm_opt) {
auto rhs_const = dynamic_cast<ConstantValue*>(rhs);
auto instr = std::make_unique<MachineInstr>(RVOpcodes::ADDIW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<ImmOperand>(rhs_const->getInt()));
CurMBB->addInstruction(std::move(instr));
return; // 指令已生成,直接返回。
}
auto rhs_vreg = getVReg(rhs);
switch (bin->getKind()) {
case BinaryInst::kAdd: {
// 区分指针运算64位和整数运算32位
RVOpcodes opcode = (lhs->getType()->isPointer() || rhs->getType()->isPointer()) ? RVOpcodes::ADD : RVOpcodes::ADDW;
auto instr = std::make_unique<MachineInstr>(opcode);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case BinaryInst::kSub: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SUBW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case BinaryInst::kMul: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::MULW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case Instruction::kDiv: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::DIVW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case Instruction::kRem: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::REMW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case BinaryInst::kICmpEQ: { // 等于 (a == b) -> (subw; seqz)
auto sub = std::make_unique<MachineInstr>(RVOpcodes::SUBW);
sub->addOperand(std::make_unique<RegOperand>(dest_vreg));
sub->addOperand(std::make_unique<RegOperand>(lhs_vreg));
sub->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(sub));
auto seqz = std::make_unique<MachineInstr>(RVOpcodes::SEQZ);
seqz->addOperand(std::make_unique<RegOperand>(dest_vreg));
seqz->addOperand(std::make_unique<RegOperand>(dest_vreg));
CurMBB->addInstruction(std::move(seqz));
break;
}
case BinaryInst::kICmpNE: { // 不等于 (a != b) -> (subw; snez)
auto sub = std::make_unique<MachineInstr>(RVOpcodes::SUBW);
sub->addOperand(std::make_unique<RegOperand>(dest_vreg));
sub->addOperand(std::make_unique<RegOperand>(lhs_vreg));
sub->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(sub));
auto snez = std::make_unique<MachineInstr>(RVOpcodes::SNEZ);
snez->addOperand(std::make_unique<RegOperand>(dest_vreg));
snez->addOperand(std::make_unique<RegOperand>(dest_vreg));
CurMBB->addInstruction(std::move(snez));
break;
}
case BinaryInst::kICmpLT: { // 小于 (a < b) -> slt
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SLT);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case BinaryInst::kICmpGT: { // 大于 (a > b) -> (b < a) -> slt
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SLT);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case BinaryInst::kICmpLE: { // 小于等于 (a <= b) -> !(b < a) -> (slt; xori)
auto slt = std::make_unique<MachineInstr>(RVOpcodes::SLT);
slt->addOperand(std::make_unique<RegOperand>(dest_vreg));
slt->addOperand(std::make_unique<RegOperand>(rhs_vreg));
slt->addOperand(std::make_unique<RegOperand>(lhs_vreg));
CurMBB->addInstruction(std::move(slt));
auto xori = std::make_unique<MachineInstr>(RVOpcodes::XORI);
xori->addOperand(std::make_unique<RegOperand>(dest_vreg));
xori->addOperand(std::make_unique<RegOperand>(dest_vreg));
xori->addOperand(std::make_unique<ImmOperand>(1));
CurMBB->addInstruction(std::move(xori));
break;
}
case BinaryInst::kICmpGE: { // 大于等于 (a >= b) -> !(a < b) -> (slt; xori)
auto slt = std::make_unique<MachineInstr>(RVOpcodes::SLT);
slt->addOperand(std::make_unique<RegOperand>(dest_vreg));
slt->addOperand(std::make_unique<RegOperand>(lhs_vreg));
slt->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(slt));
auto xori = std::make_unique<MachineInstr>(RVOpcodes::XORI);
xori->addOperand(std::make_unique<RegOperand>(dest_vreg));
xori->addOperand(std::make_unique<RegOperand>(dest_vreg));
xori->addOperand(std::make_unique<ImmOperand>(1));
CurMBB->addInstruction(std::move(xori));
break;
}
default:
throw std::runtime_error("Unsupported binary instruction in ISel");
}
break;
}
case DAGNode::UNARY: {
auto unary = dynamic_cast<UnaryInst*>(node->value);
auto dest_vreg = getVReg(unary);
auto src_vreg = getVReg(unary->getOperand());
switch (unary->getKind()) {
case UnaryInst::kNeg: { // 取负: 0 - src
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SUBW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::ZERO));
instr->addOperand(std::make_unique<RegOperand>(src_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case UnaryInst::kNot: { // 逻辑非: src == 0 ? 1 : 0
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SEQZ);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(src_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
default:
throw std::runtime_error("Unsupported unary instruction in ISel");
}
break;
}
case DAGNode::CALL: {
auto call = dynamic_cast<CallInst*>(node->value);
// 处理函数参数放入a0-a7物理寄存器
size_t num_operands = node->operands.size();
size_t reg_arg_count = std::min(num_operands, (size_t)8);
for (size_t i = 0; i < reg_arg_count; ++i) {
DAGNode* arg_node = node->operands[i];
auto arg_preg = static_cast<PhysicalReg>(static_cast<int>(PhysicalReg::A0) + i);
if (arg_node->kind == DAGNode::CONSTANT) {
if (auto const_val = dynamic_cast<ConstantValue*>(arg_node->value)) {
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(arg_preg));
li->addOperand(std::make_unique<ImmOperand>(const_val->getInt()));
CurMBB->addInstruction(std::move(li));
}
} else {
auto src_vreg = getVReg(arg_node->value);
auto mv = std::make_unique<MachineInstr>(RVOpcodes::MV);
mv->addOperand(std::make_unique<RegOperand>(arg_preg));
mv->addOperand(std::make_unique<RegOperand>(src_vreg));
CurMBB->addInstruction(std::move(mv));
}
}
if (num_operands > 8) {
size_t stack_arg_count = num_operands - 8;
int stack_space = stack_arg_count * 8; // RV64中每个参数槽位8字节
// 2a. 在栈上分配空间
auto alloc_instr = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
alloc_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
alloc_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
alloc_instr->addOperand(std::make_unique<ImmOperand>(-stack_space));
CurMBB->addInstruction(std::move(alloc_instr));
// 2b. 存储每个栈参数
for (size_t i = 8; i < num_operands; ++i) {
DAGNode* arg_node = node->operands[i];
unsigned src_vreg;
// 准备源寄存器
if (arg_node->kind == DAGNode::CONSTANT) {
// 如果是常量,先加载到临时寄存器
src_vreg = getNewVReg();
auto const_val = dynamic_cast<ConstantValue*>(arg_node->value);
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(src_vreg));
li->addOperand(std::make_unique<ImmOperand>(const_val->getInt()));
CurMBB->addInstruction(std::move(li));
} else {
src_vreg = getVReg(arg_node->value);
}
// 计算在栈上的偏移量
int offset = (i - 8) * 8;
// 生成 sd 指令
auto sd_instr = std::make_unique<MachineInstr>(RVOpcodes::SD);
sd_instr->addOperand(std::make_unique<RegOperand>(src_vreg));
sd_instr->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::SP),
std::make_unique<ImmOperand>(offset)
));
CurMBB->addInstruction(std::move(sd_instr));
}
}
auto call_instr = std::make_unique<MachineInstr>(RVOpcodes::CALL);
call_instr->addOperand(std::make_unique<LabelOperand>(call->getCallee()->getName()));
CurMBB->addInstruction(std::move(call_instr));
if (num_operands > 8) {
size_t stack_arg_count = num_operands - 8;
int stack_space = stack_arg_count * 8;
auto dealloc_instr = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
dealloc_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
dealloc_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
dealloc_instr->addOperand(std::make_unique<ImmOperand>(stack_space));
CurMBB->addInstruction(std::move(dealloc_instr));
}
// 处理返回值从a0移动到目标虚拟寄存器
if (!call->getType()->isVoid()) {
auto mv_instr = std::make_unique<MachineInstr>(RVOpcodes::MV);
mv_instr->addOperand(std::make_unique<RegOperand>(getVReg(call)));
mv_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::A0));
CurMBB->addInstruction(std::move(mv_instr));
}
break;
}
case DAGNode::RETURN: {
auto ret_inst_ir = dynamic_cast<ReturnInst*>(node->value);
if (ret_inst_ir && ret_inst_ir->hasReturnValue()) {
Value* ret_val = ret_inst_ir->getReturnValue();
// [V2优点] 在RETURN节点内加载常量返回值
if (auto const_val = dynamic_cast<ConstantValue*>(ret_val)) {
auto li_instr = std::make_unique<MachineInstr>(RVOpcodes::LI);
li_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::A0));
li_instr->addOperand(std::make_unique<ImmOperand>(const_val->getInt()));
CurMBB->addInstruction(std::move(li_instr));
} else {
auto mv_instr = std::make_unique<MachineInstr>(RVOpcodes::MV);
mv_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::A0));
mv_instr->addOperand(std::make_unique<RegOperand>(getVReg(ret_val)));
CurMBB->addInstruction(std::move(mv_instr));
}
}
// [V1设计保留] 函数尾声epilogue不由RETURN节点生成
// 而是由后续的AsmPrinter或其它Pass统一处理这是一种常见且有效的模块化设计。
auto ret_mi = std::make_unique<MachineInstr>(RVOpcodes::RET);
CurMBB->addInstruction(std::move(ret_mi));
break;
}
case DAGNode::BRANCH: {
// 处理条件分支
if (auto cond_br = dynamic_cast<CondBrInst*>(node->value)) {
Value* condition = cond_br->getCondition();
auto then_bb_name = cond_br->getThenBlock()->getName();
auto else_bb_name = cond_br->getElseBlock()->getName();
// [优化] 检查分支条件是否为编译期常量
if (auto const_cond = dynamic_cast<ConstantValue*>(condition)) {
// 如果条件是常量直接生成一个无条件跳转J而不是BNE
if (const_cond->getInt() != 0) { // 条件为 true
auto j_instr = std::make_unique<MachineInstr>(RVOpcodes::J);
j_instr->addOperand(std::make_unique<LabelOperand>(then_bb_name));
CurMBB->addInstruction(std::move(j_instr));
} else { // 条件为 false
auto j_instr = std::make_unique<MachineInstr>(RVOpcodes::J);
j_instr->addOperand(std::make_unique<LabelOperand>(else_bb_name));
CurMBB->addInstruction(std::move(j_instr));
}
}
// 如果条件不是常量,则执行标准流程
else {
// [修复] 为条件变量生成加载指令(如果它是常量的话,尽管上面已经处理了)
// 这一步是为了逻辑完整,以防有其他类型的常量没有被捕获
if (auto const_val = dynamic_cast<ConstantValue*>(condition)) {
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(getVReg(const_val)));
li->addOperand(std::make_unique<ImmOperand>(const_val->getInt()));
CurMBB->addInstruction(std::move(li));
}
auto cond_vreg = getVReg(condition);
// 生成 bne cond, zero, then_label (如果cond不为0则跳转到then)
auto br_instr = std::make_unique<MachineInstr>(RVOpcodes::BNE);
br_instr->addOperand(std::make_unique<RegOperand>(cond_vreg));
br_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::ZERO));
br_instr->addOperand(std::make_unique<LabelOperand>(then_bb_name));
CurMBB->addInstruction(std::move(br_instr));
// 为else分支生成无条件跳转 (后续Pass可以优化掉不必要的跳转)
auto j_instr = std::make_unique<MachineInstr>(RVOpcodes::J);
j_instr->addOperand(std::make_unique<LabelOperand>(else_bb_name));
CurMBB->addInstruction(std::move(j_instr));
}
}
// 处理无条件分支
else if (auto uncond_br = dynamic_cast<UncondBrInst*>(node->value)) {
auto j_instr = std::make_unique<MachineInstr>(RVOpcodes::J);
j_instr->addOperand(std::make_unique<LabelOperand>(uncond_br->getBlock()->getName()));
CurMBB->addInstruction(std::move(j_instr));
}
break;
}
case DAGNode::MEMSET: {
// [V1设计保留] Memset的核心展开逻辑在虚拟寄存器层面是正确的无需修改。
// 之前的bug是由于其输入地址、值、大小的虚拟寄存器未被正确初始化。
// 在修复了CONSTANT/ALLOCA_ADDR的加载问题后此处的逻辑现在可以正常工作。
if (DEBUG) {
std::cout << "[DEBUG] selectNode-MEMSET: Processing MEMSET node." << std::endl;
}
auto memset = dynamic_cast<MemsetInst*>(node->value);
Value* val_to_set = memset->getValue();
Value* size_to_set = memset->getSize();
Value* ptr_val = memset->getPointer();
auto dest_addr_vreg = getVReg(ptr_val);
if (auto const_val = dynamic_cast<ConstantValue*>(val_to_set)) {
if (DEBUG) {
std::cout << "[DEBUG] selectNode-MEMSET: Found constant 'value' operand (" << const_val->getInt() << "). Generating LI." << std::endl;
}
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(getVReg(const_val)));
li->addOperand(std::make_unique<ImmOperand>(const_val->getInt()));
CurMBB->addInstruction(std::move(li));
}
if (auto const_size = dynamic_cast<ConstantValue*>(size_to_set)) {
if (DEBUG) {
std::cout << "[DEBUG] selectNode-MEMSET: Found constant 'size' operand (" << const_size->getInt() << "). Generating LI." << std::endl;
}
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(getVReg(const_size)));
li->addOperand(std::make_unique<ImmOperand>(const_size->getInt()));
CurMBB->addInstruction(std::move(li));
}
if (auto alloca = dynamic_cast<AllocaInst*>(ptr_val)) {
if (DEBUG) {
std::cout << "[DEBUG] selectNode-MEMSET: Found 'pointer' operand is an AllocaInst. Generating FRAME_ADDR." << std::endl;
}
// 生成新的伪指令来获取栈地址
auto instr = std::make_unique<MachineInstr>(RVOpcodes::FRAME_ADDR);
instr->addOperand(std::make_unique<RegOperand>(dest_addr_vreg)); // 目标虚拟寄存器
instr->addOperand(std::make_unique<RegOperand>(getVReg(alloca))); // 源AllocaInst
CurMBB->addInstruction(std::move(instr));
}
auto r_dest_addr = getVReg(memset->getPointer());
auto r_num_bytes = getVReg(memset->getSize());
auto r_value_byte = getVReg(memset->getValue());
// 为memset内部逻辑创建新的临时虚拟寄存器
auto r_counter = getNewVReg();
auto r_end_addr = getNewVReg();
auto r_current_addr = getNewVReg();
auto r_temp_val = getNewVReg();
// 定义一系列lambda表达式来简化指令创建
auto add_instr = [&](RVOpcodes op, unsigned rd, unsigned rs1, unsigned rs2) {
auto i = std::make_unique<MachineInstr>(op);
i->addOperand(std::make_unique<RegOperand>(rd));
i->addOperand(std::make_unique<RegOperand>(rs1));
i->addOperand(std::make_unique<RegOperand>(rs2));
CurMBB->addInstruction(std::move(i));
};
auto addi_instr = [&](RVOpcodes op, unsigned rd, unsigned rs1, int64_t imm) {
auto i = std::make_unique<MachineInstr>(op);
i->addOperand(std::make_unique<RegOperand>(rd));
i->addOperand(std::make_unique<RegOperand>(rs1));
i->addOperand(std::make_unique<ImmOperand>(imm));
CurMBB->addInstruction(std::move(i));
};
auto store_instr = [&](RVOpcodes op, unsigned src, unsigned base, int64_t off) {
auto i = std::make_unique<MachineInstr>(op);
i->addOperand(std::make_unique<RegOperand>(src));
i->addOperand(std::make_unique<MemOperand>(std::make_unique<RegOperand>(base), std::make_unique<ImmOperand>(off)));
CurMBB->addInstruction(std::move(i));
};
auto branch_instr = [&](RVOpcodes op, unsigned rs1, unsigned rs2, const std::string& label) {
auto i = std::make_unique<MachineInstr>(op);
i->addOperand(std::make_unique<RegOperand>(rs1));
i->addOperand(std::make_unique<RegOperand>(rs2));
i->addOperand(std::make_unique<LabelOperand>(label));
CurMBB->addInstruction(std::move(i));
};
auto jump_instr = [&](const std::string& label) {
auto i = std::make_unique<MachineInstr>(RVOpcodes::J);
i->addOperand(std::make_unique<LabelOperand>(label));
CurMBB->addInstruction(std::move(i));
};
auto label_instr = [&](const std::string& name) {
auto i = std::make_unique<MachineInstr>(RVOpcodes::LABEL);
i->addOperand(std::make_unique<LabelOperand>(name));
CurMBB->addInstruction(std::move(i));
};
// 生成唯一的循环标签
int unique_id = this->local_label_counter++;
std::string loop_start_label = MFunc->getName() + "_memset_loop_start_" + std::to_string(unique_id);
std::string loop_end_label = MFunc->getName() + "_memset_loop_end_" + std::to_string(unique_id);
std::string remainder_label = MFunc->getName() + "_memset_remainder_" + std::to_string(unique_id);
std::string done_label = MFunc->getName() + "_memset_done_" + std::to_string(unique_id);
// 构造64位的填充值
addi_instr(RVOpcodes::ANDI, r_temp_val, r_value_byte, 255);
addi_instr(RVOpcodes::SLLI, r_value_byte, r_temp_val, 8);
add_instr(RVOpcodes::OR, r_temp_val, r_temp_val, r_value_byte);
addi_instr(RVOpcodes::SLLI, r_value_byte, r_temp_val, 16);
add_instr(RVOpcodes::OR, r_temp_val, r_temp_val, r_value_byte);
addi_instr(RVOpcodes::SLLI, r_value_byte, r_temp_val, 32);
add_instr(RVOpcodes::OR, r_temp_val, r_temp_val, r_value_byte);
// 计算循环边界
add_instr(RVOpcodes::ADD, r_end_addr, r_dest_addr, r_num_bytes);
auto mv = std::make_unique<MachineInstr>(RVOpcodes::MV);
mv->addOperand(std::make_unique<RegOperand>(r_current_addr));
mv->addOperand(std::make_unique<RegOperand>(r_dest_addr));
CurMBB->addInstruction(std::move(mv));
addi_instr(RVOpcodes::ANDI, r_counter, r_num_bytes, -8);
add_instr(RVOpcodes::ADD, r_counter, r_dest_addr, r_counter);
// 8字节主循环
label_instr(loop_start_label);
branch_instr(RVOpcodes::BGEU, r_current_addr, r_counter, loop_end_label);
store_instr(RVOpcodes::SD, r_temp_val, r_current_addr, 0);
addi_instr(RVOpcodes::ADDI, r_current_addr, r_current_addr, 8);
jump_instr(loop_start_label);
// 1字节收尾循环
label_instr(loop_end_label);
label_instr(remainder_label);
branch_instr(RVOpcodes::BGEU, r_current_addr, r_end_addr, done_label);
store_instr(RVOpcodes::SB, r_temp_val, r_current_addr, 0);
addi_instr(RVOpcodes::ADDI, r_current_addr, r_current_addr, 1);
jump_instr(remainder_label);
label_instr(done_label);
break;
}
default:
throw std::runtime_error("Unsupported DAGNode kind in ISel");
}
}
// 以下是忠实移植的DAG构建函数
RISCv64ISel::DAGNode* RISCv64ISel::create_node(int kind_int, Value* val, std::map<Value*, DAGNode*>& value_to_node, std::vector<std::unique_ptr<DAGNode>>& nodes_storage) {
auto kind = static_cast<DAGNode::NodeKind>(kind_int);
if (val && value_to_node.count(val) && kind != DAGNode::STORE && kind != DAGNode::RETURN && kind != DAGNode::BRANCH && kind != DAGNode::MEMSET) {
return value_to_node[val];
}
auto node = std::make_unique<DAGNode>(kind);
node->value = val;
DAGNode* raw_node_ptr = node.get();
nodes_storage.push_back(std::move(node));
if (val && !val->getType()->isVoid() && (dynamic_cast<Instruction*>(val) || dynamic_cast<GlobalValue*>(val))) {
value_to_node[val] = raw_node_ptr;
}
return raw_node_ptr;
}
RISCv64ISel::DAGNode* RISCv64ISel::get_operand_node(Value* val_ir, std::map<Value*, DAGNode*>& value_to_node, std::vector<std::unique_ptr<DAGNode>>& nodes_storage) {
if (value_to_node.count(val_ir)) {
return value_to_node[val_ir];
} else if (dynamic_cast<ConstantValue*>(val_ir)) {
return create_node(DAGNode::CONSTANT, val_ir, value_to_node, nodes_storage);
} else if (dynamic_cast<GlobalValue*>(val_ir)) {
return create_node(DAGNode::CONSTANT, val_ir, value_to_node, nodes_storage);
} else if (dynamic_cast<AllocaInst*>(val_ir)) {
return create_node(DAGNode::ALLOCA_ADDR, val_ir, value_to_node, nodes_storage);
}
return create_node(DAGNode::LOAD, val_ir, value_to_node, nodes_storage);
}
std::vector<std::unique_ptr<RISCv64ISel::DAGNode>> RISCv64ISel::build_dag(BasicBlock* bb) {
std::vector<std::unique_ptr<DAGNode>> nodes_storage;
std::map<Value*, DAGNode*> value_to_node;
for (const auto& inst_ptr : bb->getInstructions()) {
Instruction* inst = inst_ptr.get();
if (auto alloca = dynamic_cast<AllocaInst*>(inst)) {
create_node(DAGNode::ALLOCA_ADDR, alloca, value_to_node, nodes_storage);
} else if (auto store = dynamic_cast<StoreInst*>(inst)) {
auto store_node = create_node(DAGNode::STORE, store, value_to_node, nodes_storage);
store_node->operands.push_back(get_operand_node(store->getValue(), value_to_node, nodes_storage));
store_node->operands.push_back(get_operand_node(store->getPointer(), value_to_node, nodes_storage));
} else if (auto memset = dynamic_cast<MemsetInst*>(inst)) {
auto memset_node = create_node(DAGNode::MEMSET, memset, value_to_node, nodes_storage);
memset_node->operands.push_back(get_operand_node(memset->getPointer(), value_to_node, nodes_storage));
memset_node->operands.push_back(get_operand_node(memset->getBegin(), value_to_node, nodes_storage));
memset_node->operands.push_back(get_operand_node(memset->getSize(), value_to_node, nodes_storage));
memset_node->operands.push_back(get_operand_node(memset->getValue(), value_to_node, nodes_storage));
if (DEBUG) {
std::cout << "[DEBUG] build_dag: Created MEMSET node for: " << memset->getName() << std::endl;
for (size_t i = 0; i < memset_node->operands.size(); ++i) {
std::cout << " -> Operand " << i << " has kind: " << memset_node->operands[i]->kind << std::endl;
}
}
} else if (auto load = dynamic_cast<LoadInst*>(inst)) {
auto load_node = create_node(DAGNode::LOAD, load, value_to_node, nodes_storage);
load_node->operands.push_back(get_operand_node(load->getPointer(), value_to_node, nodes_storage));
} else if (auto bin = dynamic_cast<BinaryInst*>(inst)) {
if(value_to_node.count(bin)) continue;
if (bin->getKind() == BinaryInst::kSub) {
if (auto const_lhs = dynamic_cast<ConstantValue*>(bin->getLhs())) {
if (const_lhs->getInt() == 0) {
auto unary_node = create_node(DAGNode::UNARY, bin, value_to_node, nodes_storage);
unary_node->operands.push_back(get_operand_node(bin->getRhs(), value_to_node, nodes_storage));
continue;
}
}
}
auto bin_node = create_node(DAGNode::BINARY, bin, value_to_node, nodes_storage);
bin_node->operands.push_back(get_operand_node(bin->getLhs(), value_to_node, nodes_storage));
bin_node->operands.push_back(get_operand_node(bin->getRhs(), value_to_node, nodes_storage));
} else if (auto un = dynamic_cast<UnaryInst*>(inst)) {
if(value_to_node.count(un)) continue;
auto unary_node = create_node(DAGNode::UNARY, un, value_to_node, nodes_storage);
unary_node->operands.push_back(get_operand_node(un->getOperand(), value_to_node, nodes_storage));
} else if (auto call = dynamic_cast<CallInst*>(inst)) {
if(value_to_node.count(call)) continue;
auto call_node = create_node(DAGNode::CALL, call, value_to_node, nodes_storage);
for (auto arg : call->getArguments()) {
call_node->operands.push_back(get_operand_node(arg->getValue(), value_to_node, nodes_storage));
}
} else if (auto ret = dynamic_cast<ReturnInst*>(inst)) {
auto ret_node = create_node(DAGNode::RETURN, ret, value_to_node, nodes_storage);
if (ret->hasReturnValue()) {
ret_node->operands.push_back(get_operand_node(ret->getReturnValue(), value_to_node, nodes_storage));
}
} else if (auto cond_br = dynamic_cast<CondBrInst*>(inst)) {
auto br_node = create_node(DAGNode::BRANCH, cond_br, value_to_node, nodes_storage);
br_node->operands.push_back(get_operand_node(cond_br->getCondition(), value_to_node, nodes_storage));
} else if (auto uncond_br = dynamic_cast<UncondBrInst*>(inst)) {
create_node(DAGNode::BRANCH, uncond_br, value_to_node, nodes_storage);
}
}
return nodes_storage;
}
// [新] 打印DAG图以供调试的辅助函数
void RISCv64ISel::print_dag(const std::vector<std::unique_ptr<DAGNode>>& dag, const std::string& bb_name) {
// 检查是否有DEBUG宏或者全局变量避免在非调试模式下打印
// if (!DEBUG) return;
std::cerr << "=== DAG for Basic Block: " << bb_name << " ===\n";
std::set<DAGNode*> visited;
// 为节点分配临时ID方便阅读
std::map<DAGNode*, int> node_to_id;
int current_id = 0;
for (const auto& node_ptr : dag) {
node_to_id[node_ptr.get()] = current_id++;
}
// 将NodeKind枚举转换为字符串的辅助函数
auto get_kind_string = [](DAGNode::NodeKind kind) {
switch (kind) {
case DAGNode::CONSTANT: return "CONSTANT";
case DAGNode::LOAD: return "LOAD";
case DAGNode::STORE: return "STORE";
case DAGNode::BINARY: return "BINARY";
case DAGNode::CALL: return "CALL";
case DAGNode::RETURN: return "RETURN";
case DAGNode::BRANCH: return "BRANCH";
case DAGNode::ALLOCA_ADDR: return "ALLOCA_ADDR";
case DAGNode::UNARY: return "UNARY";
case DAGNode::MEMSET: return "MEMSET";
default: return "UNKNOWN";
}
};
// 递归打印节点的lambda表达式
std::function<void(DAGNode*, int)> print_node =
[&](DAGNode* node, int indent) {
if (!node) return;
std::string current_indent(indent, ' ');
int node_id = node_to_id.count(node) ? node_to_id[node] : -1;
std::cerr << current_indent << "Node#" << node_id << ": " << get_kind_string(node->kind);
// 尝试打印关联的虚拟寄存器
if (node->value && vreg_map.count(node->value)) {
std::cerr << " (vreg: %vreg" << vreg_map.at(node->value) << ")";
}
// 打印关联的IR Value信息
if (node->value) {
std::cerr << " [";
if (auto inst = dynamic_cast<Instruction*>(node->value)) {
std::cerr << inst->getKindString();
if (!inst->getName().empty()) {
std::cerr << "(" << inst->getName() << ")";
}
} else if (auto constant = dynamic_cast<ConstantValue*>(node->value)) {
std::cerr << "Const(" << constant->getInt() << ")";
} else if (auto global = dynamic_cast<GlobalValue*>(node->value)) {
std::cerr << "Global(" << global->getName() << ")";
} else if (auto alloca = dynamic_cast<AllocaInst*>(node->value)) {
std::cerr << "Alloca(" << alloca->getName() << ")";
}
std::cerr << "]";
}
std::cerr << "\n";
if (visited.count(node)) {
std::cerr << current_indent << " (已打印过子节点)\n";
return;
}
visited.insert(node);
if (!node->operands.empty()) {
std::cerr << current_indent << " Operands:\n";
for (auto operand : node->operands) {
print_node(operand, indent + 4);
}
}
};
// 从根节点(没有用户的节点,或有副作用的节点)开始打印
for (const auto& node_ptr : dag) {
if (node_ptr->users.empty() ||
node_ptr->kind == DAGNode::STORE ||
node_ptr->kind == DAGNode::RETURN ||
node_ptr->kind == DAGNode::BRANCH ||
node_ptr->kind == DAGNode::MEMSET)
{
print_node(node_ptr.get(), 0);
}
}
std::cerr << "======================================\n\n";
}
} // namespace sysy

54
src/RISCv64Passes.cpp Normal file
View File

@ -0,0 +1,54 @@
#include "RISCv64Passes.h"
#include <iostream>
namespace sysy {
// --- 寄存器分配前优化 ---
void PreRA_Scheduler::runOnMachineFunction(MachineFunction* mfunc) {
// TODO: 在此实现寄存器分配前的指令调度。
// 遍历mfunc中的每一个MachineBasicBlock。
// 对每个基本块内的MachineInstr列表进行重排。
//
// 实现思路:
// 1. 分析每个基本块内指令的数据依赖关系,构建依赖图(DAG)。
// 2. 根据目标处理器的流水线特性(指令延迟等),使用列表调度等算法对指令进行重排。
// 3. 此时操作的是虚拟寄存器,只存在真依赖,调度自由度最大。
//
// std::cout << "Running Pre-RA Instruction Scheduler..." << std::endl;
}
// --- 寄存器分配后优化 ---
void PeepholeOptimizer::runOnMachineFunction(MachineFunction* mfunc) {
// TODO: 在此实现窥孔优化。
// 遍历mfunc中的每一个MachineBasicBlock。
// 对每个基本块内的MachineInstr列表进行扫描和替换。
//
// 实现思路:
// 1. 维护一个大小固定例如3-5条指令的滑动窗口。
// 2. 识别特定的冗余模式,例如:
// - `mv a0, a1` 后紧跟 `mv a1, a0` (可消除的交换)
// - `sw t0, 12(s0)` 后紧跟 `lw t1, 12(s0)` (冗余加载)
// - 强度削减: `mul x, x, 2` -> `slli x, x, 1`
// 3. 识别后直接修改MachineInstr列表删除、替换或插入指令
//
// std::cout << "Running Post-RA Peephole Optimizer..." << std::endl;
}
void PostRA_Scheduler::runOnMachineFunction(MachineFunction* mfunc) {
// TODO: 在此实现寄存器分配后的局部指令调度。
// 遍历mfunc中的每一个MachineBasicBlock。
// 重点关注由寄存器分配器插入的spill/fill代码。
//
// 实现思路:
// 1. 识别出用于spill/fill的lw/sw指令。
// 2. 在不违反数据依赖(包括物理寄存器引入的伪依赖)的前提下,
// 尝试将lw指令向上移动使其与使用它的指令之间有足够的距离以隐藏访存延迟。
// 3. 同样可以尝试将sw指令向下移动。
//
// std::cout << "Running Post-RA Local Scheduler..." << std::endl;
}
} // namespace sysy

335
src/RISCv64RegAlloc.cpp Normal file
View File

@ -0,0 +1,335 @@
#include "RISCv64RegAlloc.h"
#include "RISCv64ISel.h"
#include <algorithm>
#include <vector>
namespace sysy {
RISCv64RegAlloc::RISCv64RegAlloc(MachineFunction* mfunc) : MFunc(mfunc) {
allocable_int_regs = {
PhysicalReg::T0, PhysicalReg::T1, PhysicalReg::T2, PhysicalReg::T3,
PhysicalReg::T4, PhysicalReg::T5, PhysicalReg::T6,
PhysicalReg::A0, PhysicalReg::A1, PhysicalReg::A2, PhysicalReg::A3,
PhysicalReg::A4, PhysicalReg::A5, PhysicalReg::A6, PhysicalReg::A7,
PhysicalReg::S0, PhysicalReg::S1, PhysicalReg::S2, PhysicalReg::S3,
PhysicalReg::S4, PhysicalReg::S5, PhysicalReg::S6, PhysicalReg::S7,
PhysicalReg::S8, PhysicalReg::S9, PhysicalReg::S10, PhysicalReg::S11,
};
}
void RISCv64RegAlloc::run() {
eliminateFrameIndices();
analyzeLiveness();
buildInterferenceGraph();
colorGraph();
rewriteFunction();
}
void RISCv64RegAlloc::eliminateFrameIndices() {
StackFrameInfo& frame_info = MFunc->getFrameInfo();
int current_offset = 20; // 这里写20是为了在$s0和第一个变量之间留出20字节的安全区
// 以防止一些函数调用方面的恶性bug。
Function* F = MFunc->getFunc();
RISCv64ISel* isel = MFunc->getISel();
for (auto& bb : F->getBasicBlocks()) {
for (auto& inst : bb->getInstructions()) {
if (auto alloca = dynamic_cast<AllocaInst*>(inst.get())) {
int size = 4;
if (!alloca->getDims().empty()) {
int num_elements = 1;
for (const auto& dim_use : alloca->getDims()) {
if (auto const_dim = dynamic_cast<ConstantValue*>(dim_use->getValue())) {
num_elements *= const_dim->getInt();
}
}
size *= num_elements;
}
current_offset += size;
unsigned alloca_vreg = isel->getVReg(alloca);
frame_info.alloca_offsets[alloca_vreg] = -current_offset;
}
}
}
frame_info.locals_size = current_offset;
for (auto& mbb : MFunc->getBlocks()) {
std::vector<std::unique_ptr<MachineInstr>> new_instructions;
for (auto& instr_ptr : mbb->getInstructions()) {
if (instr_ptr->getOpcode() == RVOpcodes::FRAME_LOAD) {
auto& operands = instr_ptr->getOperands();
unsigned dest_vreg = static_cast<RegOperand*>(operands[0].get())->getVRegNum();
unsigned alloca_vreg = static_cast<RegOperand*>(operands[1].get())->getVRegNum();
int offset = frame_info.alloca_offsets.at(alloca_vreg);
auto addr_vreg = isel->getNewVReg();
auto addi = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
addi->addOperand(std::make_unique<RegOperand>(addr_vreg));
addi->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
addi->addOperand(std::make_unique<ImmOperand>(offset));
new_instructions.push_back(std::move(addi));
auto lw = std::make_unique<MachineInstr>(RVOpcodes::LW);
lw->addOperand(std::make_unique<RegOperand>(dest_vreg));
lw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(addr_vreg),
std::make_unique<ImmOperand>(0)));
new_instructions.push_back(std::move(lw));
} else if (instr_ptr->getOpcode() == RVOpcodes::FRAME_STORE) {
auto& operands = instr_ptr->getOperands();
unsigned src_vreg = static_cast<RegOperand*>(operands[0].get())->getVRegNum();
unsigned alloca_vreg = static_cast<RegOperand*>(operands[1].get())->getVRegNum();
int offset = frame_info.alloca_offsets.at(alloca_vreg);
auto addr_vreg = isel->getNewVReg();
auto addi = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
addi->addOperand(std::make_unique<RegOperand>(addr_vreg));
addi->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
addi->addOperand(std::make_unique<ImmOperand>(offset));
new_instructions.push_back(std::move(addi));
auto sw = std::make_unique<MachineInstr>(RVOpcodes::SW);
sw->addOperand(std::make_unique<RegOperand>(src_vreg));
sw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(addr_vreg),
std::make_unique<ImmOperand>(0)));
new_instructions.push_back(std::move(sw));
} else if (instr_ptr->getOpcode() == RVOpcodes::FRAME_ADDR) { // [新] 处理FRAME_ADDR
auto& operands = instr_ptr->getOperands();
unsigned dest_vreg = static_cast<RegOperand*>(operands[0].get())->getVRegNum();
unsigned alloca_vreg = static_cast<RegOperand*>(operands[1].get())->getVRegNum();
int offset = frame_info.alloca_offsets.at(alloca_vreg);
// 将 `frame_addr rd, rs` 展开为 `addi rd, s0, offset`
auto addi = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
addi->addOperand(std::make_unique<RegOperand>(dest_vreg));
addi->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0)); // 基地址是帧指针 s0
addi->addOperand(std::make_unique<ImmOperand>(offset));
new_instructions.push_back(std::move(addi));
} else {
new_instructions.push_back(std::move(instr_ptr));
}
}
mbb->getInstructions() = std::move(new_instructions);
}
}
void RISCv64RegAlloc::getInstrUseDef(MachineInstr* instr, LiveSet& use, LiveSet& def) {
bool is_def = true;
auto opcode = instr->getOpcode();
// 预定义def和use规则
if (opcode == RVOpcodes::SW || opcode == RVOpcodes::SD ||
opcode == RVOpcodes::BEQ || opcode == RVOpcodes::BNE ||
opcode == RVOpcodes::BLT || opcode == RVOpcodes::BGE ||
opcode == RVOpcodes::RET || opcode == RVOpcodes::J) {
is_def = false;
}
if (opcode == RVOpcodes::CALL) {
// CALL会杀死所有调用者保存寄存器这是一个简化处理
// 同时也使用了传入a0-a7的参数
}
for (const auto& op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<RegOperand*>(op.get());
if (reg_op->isVirtual()) {
if (is_def) {
def.insert(reg_op->getVRegNum());
is_def = false;
} else {
use.insert(reg_op->getVRegNum());
}
}
} else if (op->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand*>(op.get());
if (mem_op->getBase()->isVirtual()) {
use.insert(mem_op->getBase()->getVRegNum());
}
}
}
}
void RISCv64RegAlloc::analyzeLiveness() {
bool changed = true;
while (changed) {
changed = false;
for (auto it = MFunc->getBlocks().rbegin(); it != MFunc->getBlocks().rend(); ++it) {
auto& mbb = *it;
LiveSet live_out;
for (auto succ : mbb->successors) {
if (!succ->getInstructions().empty()) {
auto first_instr = succ->getInstructions().front().get();
if (live_in_map.count(first_instr)) {
live_out.insert(live_in_map.at(first_instr).begin(), live_in_map.at(first_instr).end());
}
}
}
for (auto instr_it = mbb->getInstructions().rbegin(); instr_it != mbb->getInstructions().rend(); ++instr_it) {
MachineInstr* instr = instr_it->get();
LiveSet old_live_in = live_in_map[instr];
live_out_map[instr] = live_out;
LiveSet use, def;
getInstrUseDef(instr, use, def);
LiveSet live_in = use;
LiveSet diff = live_out;
for (auto vreg : def) {
diff.erase(vreg);
}
live_in.insert(diff.begin(), diff.end());
live_in_map[instr] = live_in;
live_out = live_in;
if (live_in_map[instr] != old_live_in) {
changed = true;
}
}
}
}
}
void RISCv64RegAlloc::buildInterferenceGraph() {
std::set<unsigned> all_vregs;
for (auto& mbb : MFunc->getBlocks()) {
for(auto& instr : mbb->getInstructions()) {
LiveSet use, def;
getInstrUseDef(instr.get(), use, def);
for(auto u : use) all_vregs.insert(u);
for(auto d : def) all_vregs.insert(d);
}
}
for (auto vreg : all_vregs) { interference_graph[vreg] = {}; }
for (auto& mbb : MFunc->getBlocks()) {
for (auto& instr : mbb->getInstructions()) {
LiveSet def, use;
getInstrUseDef(instr.get(), use, def);
const LiveSet& live_out = live_out_map.at(instr.get());
for (unsigned d : def) {
for (unsigned l : live_out) {
if (d != l) {
interference_graph[d].insert(l);
interference_graph[l].insert(d);
}
}
}
}
}
}
void RISCv64RegAlloc::colorGraph() {
std::vector<unsigned> sorted_vregs;
for (auto const& [vreg, neighbors] : interference_graph) {
sorted_vregs.push_back(vreg);
}
std::sort(sorted_vregs.begin(), sorted_vregs.end(), [&](unsigned a, unsigned b) {
return interference_graph[a].size() > interference_graph[b].size();
});
for (unsigned vreg : sorted_vregs) {
std::set<PhysicalReg> used_colors;
for (unsigned neighbor : interference_graph.at(vreg)) {
if (color_map.count(neighbor)) {
used_colors.insert(color_map.at(neighbor));
}
}
bool colored = false;
for (PhysicalReg preg : allocable_int_regs) {
if (used_colors.find(preg) == used_colors.end()) {
color_map[vreg] = preg;
colored = true;
break;
}
}
if (!colored) {
spilled_vregs.insert(vreg);
}
}
}
void RISCv64RegAlloc::rewriteFunction() {
StackFrameInfo& frame_info = MFunc->getFrameInfo();
int current_offset = frame_info.locals_size;
for (unsigned vreg : spilled_vregs) {
current_offset += 4;
frame_info.spill_offsets[vreg] = -current_offset;
}
frame_info.spill_size = current_offset - frame_info.locals_size;
for (auto& mbb : MFunc->getBlocks()) {
std::vector<std::unique_ptr<MachineInstr>> new_instructions;
for (auto& instr_ptr : mbb->getInstructions()) {
LiveSet use, def;
getInstrUseDef(instr_ptr.get(), use, def);
for (unsigned vreg : use) {
if (spilled_vregs.count(vreg)) {
int offset = frame_info.spill_offsets.at(vreg);
auto load = std::make_unique<MachineInstr>(RVOpcodes::LW);
load->addOperand(std::make_unique<RegOperand>(vreg));
load->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0),
std::make_unique<ImmOperand>(offset)
));
new_instructions.push_back(std::move(load));
}
}
new_instructions.push_back(std::move(instr_ptr));
for (unsigned vreg : def) {
if (spilled_vregs.count(vreg)) {
int offset = frame_info.spill_offsets.at(vreg);
auto store = std::make_unique<MachineInstr>(RVOpcodes::SW);
store->addOperand(std::make_unique<RegOperand>(vreg));
store->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0),
std::make_unique<ImmOperand>(offset)
));
new_instructions.push_back(std::move(store));
}
}
}
mbb->getInstructions() = std::move(new_instructions);
}
for (auto& mbb : MFunc->getBlocks()) {
for (auto& instr_ptr : mbb->getInstructions()) {
for (auto& op_ptr : instr_ptr->getOperands()) {
if(op_ptr->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<RegOperand*>(op_ptr.get());
if (reg_op->isVirtual()) {
unsigned vreg = reg_op->getVRegNum();
if (color_map.count(vreg)) {
reg_op->setPReg(color_map.at(vreg));
} else if (spilled_vregs.count(vreg)) {
reg_op->setPReg(PhysicalReg::T6); // 溢出统一用t6
}
}
} else if (op_ptr->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand*>(op_ptr.get());
auto base_reg_op = mem_op->getBase();
if(base_reg_op->isVirtual()){
unsigned vreg = base_reg_op->getVRegNum();
if(color_map.count(vreg)) {
base_reg_op->setPReg(color_map.at(vreg));
} else if (spilled_vregs.count(vreg)) {
base_reg_op->setPReg(PhysicalReg::T6);
}
}
}
}
}
}
}
} // namespace sysy

532
src/SysYIRAnalyser.cpp Normal file
View File

@ -0,0 +1,532 @@
#include "SysYIRAnalyser.h"
#include <iostream>
namespace sysy {
void ControlFlowAnalysis::init() {
// 初始化分析器
auto &functions = pModule->getFunctions();
for (const auto &function : functions) {
auto func = function.second.get();
auto basicBlocks = func->getBasicBlocks();
for (auto &basicBlock : basicBlocks) {
blockAnalysisInfo[basicBlock.get()] = new BlockAnalysisInfo();
blockAnalysisInfo[basicBlock.get()]->clear();
}
functionAnalysisInfo[func] = new FunctionAnalysisInfo();
functionAnalysisInfo[func]->clear();
}
}
void ControlFlowAnalysis::runControlFlowAnalysis() {
// 运行控制流分析
clear(); // 清空之前的分析结果
init(); // 初始化分析器
computeDomNode();
computeDomTree();
computeDomFrontierAllBlk();
}
void ControlFlowAnalysis::intersectOP4Dom(std::unordered_set<BasicBlock *> &dom, const std::unordered_set<BasicBlock *> &other) {
// 计算交集
for (auto it = dom.begin(); it != dom.end();) {
if (other.find(*it) == other.end()) {
// 如果other中没有这个基本块则从dom中删除
it = dom.erase(it);
} else {
++it;
}
}
}
auto ControlFlowAnalysis::findCommonDominator(BasicBlock *a, BasicBlock *b) -> BasicBlock * {
// 查找两个基本块的共同支配结点
while (a != b) {
BlockAnalysisInfo* infoA = blockAnalysisInfo[a];
BlockAnalysisInfo* infoB = blockAnalysisInfo[b];
// 如果深度不同,则向上移动到直接支配结点
// TODO空间换时间倍增优化优先级较低
while (infoA->getDomDepth() > infoB->getDomDepth()) {
a = const_cast<BasicBlock*>(infoA->getIdom());
infoA = blockAnalysisInfo[a];
}
while (infoB->getDomDepth() > infoA->getDomDepth()) {
b = const_cast<BasicBlock*>(infoB->getIdom());
infoB = blockAnalysisInfo[b];
}
if (a == b) break;
a = const_cast<BasicBlock*>(infoA->getIdom());
b = const_cast<BasicBlock*>(infoB->getIdom());
}
return a;
}
void ControlFlowAnalysis::computeDomNode(){
auto &functions = pModule->getFunctions();
// 分析每个函数内的基本块
for (const auto &function : functions) {
auto func = function.second.get();
auto basicBlocks = func->getBasicBlocks();
std::unordered_set<BasicBlock *> domSetTmp;
// 一开始把domSetTmp置为所有block
auto entry_block = func->getEntryBlock();
entry_block->setName("Entry");
blockAnalysisInfo[entry_block]->addDominants(entry_block);
for (auto &basicBlock : basicBlocks) {
domSetTmp.emplace(basicBlock.get());
}
// 初始化
for (auto &basicBlock : basicBlocks) {
if (basicBlock.get() != entry_block) {
blockAnalysisInfo[basicBlock.get()]->setDominants(domSetTmp);
// 先把所有block的必经结点都设为N
}
}
// 支配节点计算公式
//DOM[B]={B} {⋂P∈pred(B) DOM[P]}
// 其中pred(B)是B的所有前驱结点
// 迭代计算支配结点,直到不再变化
// 这里使用迭代法,直到支配结点不再变化
// TODOLengauer-Tarjan 算法可以更高效地计算支配结点
// 或者按照CFG拓扑序遍历效率更高
bool changed = true;
while (changed) {
changed = false;
// 循环非start结点
for (auto &basicBlock : basicBlocks) {
if (basicBlock.get() != entry_block) {
auto olddom =
blockAnalysisInfo[basicBlock.get()]->getDominants();
std::unordered_set<BasicBlock *> dom =
blockAnalysisInfo[basicBlock->getPredecessors().front()]->getDominants();
// 对于每个基本块,计算其支配结点
// 取其前驱结点的支配结点的交集和自己
for (auto pred : basicBlock->getPredecessors()) {
intersectOP4Dom(dom, blockAnalysisInfo[pred]->getDominants());
}
dom.emplace(basicBlock.get());
blockAnalysisInfo[basicBlock.get()]->setDominants(dom);
if (dom != olddom) {
changed = true;
}
}
}
}
}
}
// TODO SEMI-NCA算法改进
void ControlFlowAnalysis::computeDomTree() {
// 构造支配树
auto &functions = pModule->getFunctions();
for (const auto &function : functions) {
auto func = function.second.get();
auto basicBlocks = func->getBasicBlocks();
auto entry_block = func->getEntryBlock();
blockAnalysisInfo[entry_block]->setIdom(entry_block);
blockAnalysisInfo[entry_block]->setDomDepth(0); // 入口块深度为0
bool changed = true;
while (changed) {
changed = false;
for (auto &basicBlock : basicBlocks) {
if (basicBlock.get() == entry_block) continue;
BasicBlock *new_idom = nullptr;
for (auto pred : basicBlock->getPredecessors()) {
// 跳过未处理的前驱
if (blockAnalysisInfo[pred]->getIdom() == nullptr) continue;
// new_idom = (new_idom == nullptr) ? pred : findCommonDominator(new_idom, pred);
if (new_idom == nullptr)
new_idom = pred;
else
new_idom = findCommonDominator(new_idom, pred);
}
// 更新直接支配节点
if (new_idom && new_idom != blockAnalysisInfo[basicBlock.get()]->getIdom()) {
// 移除旧的支配关系
if (blockAnalysisInfo[basicBlock.get()]->getIdom()) {
blockAnalysisInfo[const_cast<BasicBlock*>(blockAnalysisInfo[basicBlock.get()]->getIdom())]->removeSdoms(basicBlock.get());
}
// 设置新的支配关系
// std::cout << "Block: " << basicBlock->getName()
// << " New Idom: " << new_idom->getName() << std::endl;
blockAnalysisInfo[basicBlock.get()]->setIdom(new_idom);
blockAnalysisInfo[new_idom]->addSdoms(basicBlock.get());
// 更新深度 = 直接支配节点深度 + 1
blockAnalysisInfo[basicBlock.get()]->setDomDepth(
blockAnalysisInfo[new_idom]->getDomDepth() + 1);
changed = true;
}
}
}
}
// for (auto &basicBlock : basicBlocks) {
// if (basicBlock.get() != func->getEntryBlock()) {
// auto dominats =
// blockAnalysisInfo[basicBlock.get()]->getDominants();
// bool found = false;
// // 从前驱结点开始寻找直接支配结点
// std::queue<BasicBlock *> q;
// for (auto pred : basicBlock->getPredecessors()) {
// q.push(pred);
// }
// // BFS遍历前驱结点直到找到直接支配结点
// while (!found && !q.empty()) {
// auto curr = q.front();
// q.pop();
// if (curr == basicBlock.get())
// continue;
// if (dominats.count(curr) != 0U) {
// blockAnalysisInfo[basicBlock.get()]->setIdom(curr);
// blockAnalysisInfo[curr]->addSdoms(basicBlock.get());
// found = true;
// } else {
// for (auto pred : curr->getPredecessors()) {
// q.push(pred);
// }
// }
// }
// }
// }
}
// std::unordered_set<BasicBlock *> ControlFlowAnalysis::computeDomFrontier(BasicBlock *block) {
// std::unordered_set<BasicBlock *> ret_list;
// // 计算 localDF
// for (auto local_successor : block->getSuccessors()) {
// if (local_successor->getIdom() != block) {
// ret_list.emplace(local_successor);
// }
// }
// // 计算 upDF
// for (auto up_successor : block->getSdoms()) {
// auto childrenDF = computeDF(up_successor);
// for (auto w : childrenDF) {
// if (block != w->getIdom() || block == w) {
// ret_list.emplace(w);
// }
// }
// }
// return ret_list;
// }
void ControlFlowAnalysis::computeDomFrontierAllBlk() {
auto &functions = pModule->getFunctions();
for (const auto &function : functions) {
auto func = function.second.get();
auto basicBlocks = func->getBasicBlocks();
// 按支配树深度排序(从深到浅)
std::vector<BasicBlock *> orderedBlocks;
for (auto &bb : basicBlocks) {
orderedBlocks.push_back(bb.get());
}
std::sort(orderedBlocks.begin(), orderedBlocks.end(),
[this](BasicBlock *a, BasicBlock *b) {
return blockAnalysisInfo[a]->getDomDepth() > blockAnalysisInfo[b]->getDomDepth();
});
// 计算支配边界
for (auto block : orderedBlocks) {
std::unordered_set<BasicBlock *> df;
// Local DF: 直接后继中不被当前块支配的
for (auto succ : block->getSuccessors()) {
// 当前块不支配该后继(即不是其直接支配节点)
if (blockAnalysisInfo[succ]->getIdom() != block) {
df.insert(succ);
}
}
// Up DF: 从支配子树中继承
for (auto child : blockAnalysisInfo[block]->getSdoms()) {
for (auto w : blockAnalysisInfo[child]->getDomFrontiers()) {
// 如果w不被当前块支配
if (block != blockAnalysisInfo[w]->getIdom()) {
df.insert(w);
}
}
}
blockAnalysisInfo[block]->setDomFrontiers(df);
}
}
}
// ==========================
// dataflow analysis utils
// ==========================
// 先引用学长的代码
// TODO: Worklist 增加逆后序遍历机制
void DataFlowAnalysisUtils::forwardAnalyze(Module *pModule){
std::map<DataFlowAnalysis *, bool> workAnalysis;
for (auto &dataflow : forwardAnalysisList) {
dataflow->init(pModule);
}
for (const auto &function : pModule->getFunctions()) {
for (auto &dataflow : forwardAnalysisList) {
workAnalysis.emplace(dataflow, false);
}
while (!workAnalysis.empty()) {
for (const auto &block : function.second->getBasicBlocks()) {
for (auto &elem : workAnalysis) {
if (elem.first->analyze(pModule, block.get())) {
elem.second = true;
}
}
}
std::map<DataFlowAnalysis *, bool> tmp;
std::remove_copy_if(workAnalysis.begin(), workAnalysis.end(), std::inserter(tmp, tmp.end()),
[](const std::pair<DataFlowAnalysis *, bool> &elem) -> bool { return !elem.second; });
workAnalysis.swap(tmp);
for (auto &elem : workAnalysis) {
elem.second = false;
}
}
}
}
void DataFlowAnalysisUtils::backwardAnalyze(Module *pModule) {
std::map<DataFlowAnalysis *, bool> workAnalysis;
for (auto &dataflow : backwardAnalysisList) {
dataflow->init(pModule);
}
for (const auto &function : pModule->getFunctions()) {
for (auto &dataflow : backwardAnalysisList) {
workAnalysis.emplace(dataflow, false);
}
while (!workAnalysis.empty()) {
for (const auto &block : function.second->getBasicBlocks()) {
for (auto &elem : workAnalysis) {
if (elem.first->analyze(pModule, block.get())) {
elem.second = true;
}
}
}
std::map<DataFlowAnalysis *, bool> tmp;
std::remove_copy_if(workAnalysis.begin(), workAnalysis.end(), std::inserter(tmp, tmp.end()),
[](const std::pair<DataFlowAnalysis *, bool> &elem) -> bool { return !elem.second; });
workAnalysis.swap(tmp);
for (auto &elem : workAnalysis) {
elem.second = false;
}
}
}
}
std::set<User *> ActiveVarAnalysis::getUsedSet(Instruction *inst) {
using Kind = Instruction::Kind;
std::vector<User *> operands;
for (const auto &operand : inst->getOperands()) {
operands.emplace_back(dynamic_cast<User *>(operand->getValue()));
}
std::set<User *> result;
switch (inst->getKind()) {
// phi op
case Kind::kPhi:
case Kind::kCall:
result.insert(std::next(operands.begin()), operands.end());
break;
case Kind::kCondBr:
result.insert(operands[0]);
break;
case Kind::kBr:
case Kind::kAlloca:
break;
// mem op
case Kind::kStore:
// StoreInst 的第一个操作数是被存储的值,第二个操作数是存储的变量
// 后续的是可能的数组维度
result.insert(operands[0]);
result.insert(operands.begin() + 2, operands.end());
break;
case Kind::kLoad:
case Kind::kLa: {
auto variable = dynamic_cast<AllocaInst *>(operands[0]);
auto global = dynamic_cast<GlobalValue *>(operands[0]);
auto constArray = dynamic_cast<ConstantVariable *>(operands[0]);
if ((variable != nullptr && variable->getNumDims() == 0) || (global != nullptr && global->getNumDims() == 0) ||
(constArray != nullptr && constArray->getNumDims() == 0)) {
result.insert(operands[0]);
}
result.insert(std::next(operands.begin()), operands.end());
break;
}
case Kind::kGetSubArray: {
for (unsigned i = 2; i < operands.size(); i++) {
// 数组的维度信息
result.insert(operands[i]);
}
break;
}
case Kind::kMemset: {
result.insert(std::next(operands.begin()), operands.end());
break;
}
case Kind::kInvalid:
// Binary
case Kind::kAdd:
case Kind::kSub:
case Kind::kMul:
case Kind::kDiv:
case Kind::kRem:
case Kind::kICmpEQ:
case Kind::kICmpNE:
case Kind::kICmpLT:
case Kind::kICmpLE:
case Kind::kICmpGT:
case Kind::kICmpGE:
case Kind::kFAdd:
case Kind::kFSub:
case Kind::kFMul:
case Kind::kFDiv:
case Kind::kFCmpEQ:
case Kind::kFCmpNE:
case Kind::kFCmpLT:
case Kind::kFCmpLE:
case Kind::kFCmpGT:
case Kind::kFCmpGE:
case Kind::kAnd:
case Kind::kOr:
// Unary
case Kind::kNeg:
case Kind::kNot:
case Kind::kFNot:
case Kind::kFNeg:
case Kind::kFtoI:
case Kind::kItoF:
// terminator
case Kind::kReturn:
result.insert(operands.begin(), operands.end());
break;
default:
assert(false);
break;
}
result.erase(nullptr);
return result;
}
User * ActiveVarAnalysis::getDefine(Instruction *inst) {
User *result = nullptr;
if (inst->isStore()) {
StoreInst* store = dynamic_cast<StoreInst *>(inst);
auto operand = store->getPointer();
AllocaInst* variable = dynamic_cast<AllocaInst *>(operand);
GlobalValue* global = dynamic_cast<GlobalValue *>(operand);
if ((variable != nullptr && variable->getNumDims() != 0) || (global != nullptr && global->getNumDims() != 0)) {
// 如果是数组变量或者全局变量,则不返回定义
// TODO兼容数组变量
result = nullptr;
} else {
result = dynamic_cast<User *>(operand);
}
} else if (inst->isPhi()) {
result = dynamic_cast<User *>(inst->getOperand(0));
} else if (inst->isBinary() || inst->isUnary() || inst->isCall() ||
inst->isLoad() || inst->isLa()) {
result = dynamic_cast<User *>(inst);
}
return result;
}
void ActiveVarAnalysis::init(Module *pModule) {
for (const auto &function : pModule->getFunctions()) {
for (const auto &block : function.second->getBasicBlocks()) {
activeTable.emplace(block.get(), std::vector<std::set<User *>>{});
for (unsigned i = 0; i < block->getNumInstructions() + 1; i++)
activeTable.at(block.get()).emplace_back();
}
}
}
// 活跃变量分析公式 每个块内的分析动作供分析器调用
bool ActiveVarAnalysis::analyze(Module *pModule, BasicBlock *block) {
bool changed = false; // 标记数据流结果是否有变化
std::set<User *> activeSet{}; // 当前计算的活跃变量集合
// 步骤1: 计算基本块出口的活跃变量集 (OUT[B])
// 公式: OUT[B] = _{S ∈ succ(B)} IN[S]
for (const auto &succ : block->getSuccessors()) {
// 获取后继块入口的活跃变量集 (IN[S])
auto succActiveSet = activeTable.at(succ).front();
// 合并所有后继块的入口活跃变量
activeSet.insert(succActiveSet.begin(), succActiveSet.end());
}
// 步骤2: 处理基本块出口处的活跃变量集
const auto &instructions = block->getInstructions();
const auto numInstructions = instructions.size();
// 获取旧的出口活跃变量集 (block出口对应索引numInstructions)
const auto &oldEndActiveSet = activeTable.at(block)[numInstructions];
// 检查出口活跃变量集是否有变化
if (!std::equal(activeSet.begin(), activeSet.end(),
oldEndActiveSet.begin(), oldEndActiveSet.end()))
{
changed = true; // 标记变化
activeTable.at(block)[numInstructions] = activeSet; // 更新出口活跃变量集
}
// 步骤3: 逆序遍历基本块中的指令
// 从最后一条指令开始向前计算每个程序点的活跃变量
auto instructionIter = instructions.end();
instructionIter--; // 指向最后一条指令
// 从出口向入口遍历 (索引从numInstructions递减到1)
for (unsigned i = numInstructions; i > 0; i--) {
auto inst = instructionIter->get(); // 当前指令
auto used = getUsedSet(inst);
User *defined = getDefine(inst);
// 步骤3.3: 计算指令入口的活跃变量 (IN[i])
// 公式: IN[i] = use_i (OUT[i] - def_i)
activeSet.erase(defined); // 移除被定义的变量 (OUT[i] - def_i)
activeSet.insert(used.begin(), used.end()); // 添加使用的变量
// 获取旧的入口活跃变量集 (位置i-1对应当前指令的入口)
const auto &oldActiveSet = activeTable.at(block)[i - 1];
// 检查活跃变量集是否有变化
if (!std::equal(activeSet.begin(), activeSet.end(),
oldActiveSet.begin(), oldActiveSet.end()))
{
changed = true; // 标记变化
activeTable.at(block)[i - 1] = activeSet; // 更新入口活跃变量集
}
instructionIter--; // 移动到前一条指令
}
return changed; // 返回数据流结果是否变化
}
auto ActiveVarAnalysis::getActiveTable() const -> const std::map<BasicBlock *, std::vector<std::set<User *>>> & {
return activeTable;
}
} // namespace sysy

1278
src/SysYIRGenerator.cpp Normal file

File diff suppressed because it is too large Load Diff

484
src/SysYIROptPre.cpp Normal file
View File

@ -0,0 +1,484 @@
#include "SysYIROptPre.h"
#include <cassert>
#include <list>
#include <map>
#include <memory>
#include <string>
#include <iostream>
#include "IR.h"
#include "IRBuilder.h"
namespace sysy {
/**
* use删除operand,以免扰乱后续分析
* instr: 要删除的指令
*/
void SysYOptPre::usedelete(Instruction *instr) {
for (auto &use : instr->getOperands()) {
Value* val = use->getValue();
// std::cout << delete << val->getName() << std::endl;
val->removeUse(use);
}
}
// 删除br后的无用指令
void SysYOptPre::SysYDelInstAfterBr() {
auto &functions = pModule->getFunctions();
for (auto &function : functions) {
auto basicBlocks = function.second->getBasicBlocks();
for (auto &basicBlock : basicBlocks) {
bool Branch = false;
auto &instructions = basicBlock->getInstructions();
auto Branchiter = instructions.end();
for (auto iter = instructions.begin(); iter != instructions.end(); ++iter) {
if (Branch)
usedelete(iter->get());
else if ((*iter)->isTerminator()){
Branch = true;
Branchiter = iter;
}
}
if (Branchiter != instructions.end()) ++Branchiter;
while (Branchiter != instructions.end())
Branchiter = instructions.erase(Branchiter);
if (Branch) { // 更新前驱后继关系
auto thelastinstinst = basicBlock->getInstructions().end();
--thelastinstinst;
auto &Successors = basicBlock->getSuccessors();
for (auto iterSucc = Successors.begin(); iterSucc != Successors.end();) {
(*iterSucc)->removePredecessor(basicBlock.get());
basicBlock->removeSuccessor(*iterSucc);
}
if (thelastinstinst->get()->isUnconditional()) {
BasicBlock* branchBlock = dynamic_cast<BasicBlock *>(thelastinstinst->get()->getOperand(0));
basicBlock->addSuccessor(branchBlock);
branchBlock->addPredecessor(basicBlock.get());
} else if (thelastinstinst->get()->isConditional()) {
BasicBlock* thenBlock = dynamic_cast<BasicBlock *>(thelastinstinst->get()->getOperand(1));
BasicBlock* elseBlock = dynamic_cast<BasicBlock *>(thelastinstinst->get()->getOperand(2));
basicBlock->addSuccessor(thenBlock);
basicBlock->addSuccessor(elseBlock);
thenBlock->addPredecessor(basicBlock.get());
elseBlock->addPredecessor(basicBlock.get());
}
}
}
}
}
void SysYOptPre::SysYBlockMerge() {
auto &functions = pModule->getFunctions(); //std::map<std::string, std::unique_ptr<Function>>
for (auto &function : functions) {
// auto basicBlocks = function.second->getBasicBlocks();
auto &func = function.second;
for (auto blockiter = func->getBasicBlocks().begin();
blockiter != func->getBasicBlocks().end();) {
if (blockiter->get()->getNumSuccessors() == 1) {
// 如果当前块只有一个后继块
// 且后继块只有一个前驱块
// 则将当前块和后继块合并
if (((blockiter->get())->getSuccessors()[0])->getNumPredecessors() == 1) {
// std::cout << "merge block: " << blockiter->get()->getName() << std::endl;
BasicBlock* block = blockiter->get();
BasicBlock* nextBlock = blockiter->get()->getSuccessors()[0];
auto nextarguments = nextBlock->getArguments();
// 删除br指令
if (block->getNumInstructions() != 0) {
auto thelastinstinst = block->end();
(--thelastinstinst);
if (thelastinstinst->get()->isUnconditional()) {
usedelete(thelastinstinst->get());
block->getInstructions().erase(thelastinstinst);
} else if (thelastinstinst->get()->isConditional()) {
// 如果是条件分支,判断条件是否相同,主要优化相同布尔表达式
if (thelastinstinst->get()->getOperand(1)->getName() == thelastinstinst->get()->getOperand(1)->getName()) {
usedelete(thelastinstinst->get());
block->getInstructions().erase(thelastinstinst);
}
}
}
// 将后继块的指令移动到当前块
// 并将后继块的父指针改为当前块
for (auto institer = nextBlock->begin(); institer != nextBlock->end();) {
institer->get()->setParent(block);
block->getInstructions().emplace_back(institer->release());
institer = nextBlock->getInstructions().erase(institer);
}
// 合并参数
// TODO是否需要去重?
for (auto &argm : nextarguments) {
argm->setParent(block);
block->insertArgument(argm);
}
// 更新前驱后继关系,类似树节点操作
block->removeSuccessor(nextBlock);
nextBlock->removePredecessor(block);
std::list<BasicBlock *> succshoulddel;
for (auto &succ : nextBlock->getSuccessors()) {
block->addSuccessor(succ);
succ->replacePredecessor(nextBlock, block);
succshoulddel.push_back(succ);
}
for (auto del : succshoulddel) {
nextBlock->removeSuccessor(del);
}
func->removeBasicBlock(nextBlock);
} else {
blockiter++;
}
} else {
blockiter++;
}
}
}
}
// 删除无前驱块兼容SSA后的处理
void SysYOptPre::SysYDelNoPreBLock() {
auto &functions = pModule->getFunctions(); // std::map<std::string, std::unique_ptr<sysy::Function>>
for (auto &function : functions) {
auto &func = function.second;
for (auto &block : func->getBasicBlocks()) {
block->setreachableFalse();
}
// 对函数基本块做一个拓扑排序,排查不可达基本块
auto entryBlock = func->getEntryBlock();
entryBlock->setreachableTrue();
std::queue<BasicBlock *> blockqueue;
blockqueue.push(entryBlock);
while (!blockqueue.empty()) {
auto block = blockqueue.front();
blockqueue.pop();
for (auto &succ : block->getSuccessors()) {
if (!succ->getreachable()) {
succ->setreachableTrue();
blockqueue.push(succ);
}
}
}
// 删除不可达基本块指令
for (auto blockIter = func->getBasicBlocks().begin(); blockIter != func->getBasicBlocks().end();blockIter++) {
if (!blockIter->get()->getreachable())
for (auto &iterInst : blockIter->get()->getInstructions())
usedelete(iterInst.get());
}
for (auto blockIter = func->getBasicBlocks().begin(); blockIter != func->getBasicBlocks().end();) {
if (!blockIter->get()->getreachable()) {
for (auto succblock : blockIter->get()->getSuccessors()) {
int indexphi = 1;
for (auto pred : succblock->getPredecessors()) {
if (pred == blockIter->get()) {
break;
}
indexphi++;
}
for (auto &phiinst : succblock->getInstructions()) {
if (phiinst->getKind() != Instruction::kPhi) {
break;
}
phiinst->removeOperand(indexphi);
}
}
// 删除不可达基本块,注意迭代器不可达问题
func->removeBasicBlock((blockIter++)->get());
} else {
blockIter++;
}
}
}
}
void SysYOptPre::SysYDelEmptyBlock() {
auto &functions = pModule->getFunctions();
for (auto &function : functions) {
// 收集不可达基本块
// 这里的不可达基本块是指没有实际指令的基本块
// 当一个基本块没有实际指令例如只有phi指令和一个uncondbr指令时也会被视作不可达
auto basicBlocks = function.second->getBasicBlocks();
std::map<sysy::BasicBlock *, BasicBlock *> EmptyBlocks;
// 空块儿和后继的基本块的映射
for (auto &basicBlock : basicBlocks) {
if (basicBlock->getNumInstructions() == 0) {
if (basicBlock->getNumSuccessors() == 1) {
EmptyBlocks[basicBlock.get()] = basicBlock->getSuccessors().front();
}
}
else{
// 如果只有phi指令和一个uncondbr。(phi)*(uncondbr)?
// 判断除了最后一个指令之外是不是只有phi指令
bool onlyPhi = true;
for (auto &inst : basicBlock->getInstructions()) {
if (!inst->isPhi() && !inst->isUnconditional()) {
onlyPhi = false;
break;
}
}
if(onlyPhi)
EmptyBlocks[basicBlock.get()] = basicBlock->getSuccessors().front();
}
}
// 更新基本块信息,增加必要指令
for (auto &basicBlock : basicBlocks) {
// 把空块转换成只有跳转指令的不可达块
if (distance(basicBlock->begin(), basicBlock->end()) == 0) {
if (basicBlock->getNumSuccessors() == 0) {
continue;
}
if (basicBlock->getNumSuccessors() > 1) {
assert("");
}
pBuilder->setPosition(basicBlock.get(), basicBlock->end());
pBuilder->createUncondBrInst(basicBlock->getSuccessors()[0], {});
continue;
}
auto thelastinst = basicBlock->getInstructions().end();
--thelastinst;
// 根据br指令传递的后继块信息跳过空块链
if (thelastinst->get()->isUnconditional()) {
BasicBlock* OldBrBlock = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0));
BasicBlock *thelastBlockOld = nullptr;
// 如果空块链表为多个块
while (EmptyBlocks.find(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))) !=
EmptyBlocks.end()) {
thelastBlockOld = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0));
thelastinst->get()->replaceOperand(0, EmptyBlocks[thelastBlockOld]);
}
basicBlock->removeSuccessor(OldBrBlock);
OldBrBlock->removePredecessor(basicBlock.get());
basicBlock->addSuccessor(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0)));
dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))->addPredecessor(basicBlock.get());
if (thelastBlockOld != nullptr) {
int indexphi = 0;
for (auto &pred : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))->getPredecessors()) {
if (pred == thelastBlockOld) {
break;
}
indexphi++;
}
// 更新phi指令的操作数
// 移除thelastBlockOld对应的phi操作数
for (auto &InstInNew : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))->getInstructions()) {
if (InstInNew->isPhi()) {
dynamic_cast<PhiInst *>(InstInNew.get())->removeOperand(indexphi + 1);
} else {
break;
}
}
}
} else if (thelastinst->get()->getKind() == Instruction::kCondBr) {
auto OldThenBlock = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1));
auto OldElseBlock = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2));
BasicBlock *thelastBlockOld = nullptr;
while (EmptyBlocks.find(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1))) !=
EmptyBlocks.end()) {
thelastBlockOld = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1));
thelastinst->get()->replaceOperand(
1, EmptyBlocks[dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1))]);
}
basicBlock->removeSuccessor(OldThenBlock);
OldThenBlock->removePredecessor(basicBlock.get());
// 处理 then 和 else 分支合并的情况
if (dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1)) ==
dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2))) {
auto thebrBlock = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1));
usedelete(thelastinst->get());
thelastinst = basicBlock->getInstructions().erase(thelastinst);
pBuilder->setPosition(basicBlock.get(), basicBlock->end());
pBuilder->createUncondBrInst(thebrBlock, {});
continue;
}
basicBlock->addSuccessor(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1)));
dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1))->addPredecessor(basicBlock.get());
// auto indexInNew = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))->getPredecessors().
if (thelastBlockOld != nullptr) {
int indexphi = 0;
for (auto &pred : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1))->getPredecessors()) {
if (pred == thelastBlockOld) {
break;
}
indexphi++;
}
for (auto &InstInNew : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1))->getInstructions()) {
if (InstInNew->isPhi()) {
dynamic_cast<PhiInst *>(InstInNew.get())->removeOperand(indexphi + 1);
} else {
break;
}
}
}
thelastBlockOld = nullptr;
while (EmptyBlocks.find(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2))) !=
EmptyBlocks.end()) {
thelastBlockOld = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2));
thelastinst->get()->replaceOperand(
2, EmptyBlocks[dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2))]);
}
basicBlock->removeSuccessor(OldElseBlock);
OldElseBlock->removePredecessor(basicBlock.get());
// 处理 then 和 else 分支合并的情况
if (dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1)) ==
dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2))) {
auto thebrBlock = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1));
usedelete(thelastinst->get());
thelastinst = basicBlock->getInstructions().erase(thelastinst);
pBuilder->setPosition(basicBlock.get(), basicBlock->end());
pBuilder->createUncondBrInst(thebrBlock, {});
continue;
}
basicBlock->addSuccessor(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2)));
dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2))->addPredecessor(basicBlock.get());
if (thelastBlockOld != nullptr) {
int indexphi = 0;
for (auto &pred : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2))->getPredecessors()) {
if (pred == thelastBlockOld) {
break;
}
indexphi++;
}
for (auto &InstInNew : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2))->getInstructions()) {
if (InstInNew->isPhi()) {
dynamic_cast<PhiInst *>(InstInNew.get())->removeOperand(indexphi + 1);
} else {
break;
}
}
}
} else {
if (basicBlock->getNumSuccessors() == 1) {
pBuilder->setPosition(basicBlock.get(), basicBlock->end());
pBuilder->createUncondBrInst(basicBlock->getSuccessors()[0], {});
auto thelastinst = basicBlock->getInstructions().end();
(--thelastinst);
auto OldBrBlock = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0));
sysy::BasicBlock *thelastBlockOld = nullptr;
while (EmptyBlocks.find(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))) !=
EmptyBlocks.end()) {
thelastBlockOld = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0));
thelastinst->get()->replaceOperand(
0, EmptyBlocks[dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))]);
}
basicBlock->removeSuccessor(OldBrBlock);
OldBrBlock->removePredecessor(basicBlock.get());
basicBlock->addSuccessor(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0)));
dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))->addPredecessor(basicBlock.get());
if (thelastBlockOld != nullptr) {
int indexphi = 0;
for (auto &pred : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))->getPredecessors()) {
if (pred == thelastBlockOld) {
break;
}
indexphi++;
}
for (auto &InstInNew : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))->getInstructions()) {
if (InstInNew->isPhi()) {
dynamic_cast<PhiInst *>(InstInNew.get())->removeOperand(indexphi + 1);
} else {
break;
}
}
}
}
}
}
for (auto iter = function.second->getBasicBlocks().begin(); iter != function.second->getBasicBlocks().end();) {
if (EmptyBlocks.find(iter->get()) != EmptyBlocks.end()) {
// EntryBlock跳过
if (iter->get() == function.second->getEntryBlock()) {
++iter;
continue;
}
for (auto &iterInst : iter->get()->getInstructions())
usedelete(iterInst.get());
// 删除不可达基本块的phi指令的操作数
for (auto &succ : iter->get()->getSuccessors()) {
int index = 0;
for (auto &pred : succ->getPredecessors()) {
if (pred == iter->get()) {
break;
}
index++;
}
for (auto &instinsucc : succ->getInstructions()) {
if (instinsucc->isPhi()) {
dynamic_cast<PhiInst *>(instinsucc.get())->removeOperand(index);
} else {
break;
}
}
}
function.second->removeBasicBlock((iter++)->get());
} else {
++iter;
}
}
}
}
// 如果函数没有返回指令,则添加一个默认返回指令(主要解决void函数没有返回指令的问题)
void SysYOptPre::SysYAddReturn() {
auto &functions = pModule->getFunctions();
for (auto &function : functions) {
auto &func = function.second;
auto basicBlocks = func->getBasicBlocks();
for (auto &block : basicBlocks) {
if (block->getNumSuccessors() == 0) {
// 如果基本块没有后继块,则添加一个返回指令
if (block->getNumInstructions() == 0) {
pBuilder->setPosition(block.get(), block->end());
pBuilder->createReturnInst();
}
auto thelastinst = block->getInstructions().end();
--thelastinst;
if (thelastinst->get()->getKind() != Instruction::kReturn) {
// std::cout << "Warning: Function " << func->getName() << " has no return instruction, adding default return." << std::endl;
pBuilder->setPosition(block.get(), block->end());
// TODO: 如果int float函数缺少返回值是否需要报错
if (func->getReturnType()->isInt()) {
pBuilder->createReturnInst(ConstantInteger::get(0));
} else if (func->getReturnType()->isFloat()) {
pBuilder->createReturnInst(ConstantFloating::get(0.0F));
} else {
pBuilder->createReturnInst();
}
}
}
}
}
}
} // namespace sysy

483
src/SysYIRPrinter.cpp Normal file
View File

@ -0,0 +1,483 @@
#include "SysYIRPrinter.h"
#include <cassert>
#include <fstream>
#include <iostream>
#include <string>
#include "IR.h"
namespace sysy {
void SysYPrinter::printIR() {
const auto &functions = pModule->getFunctions();
//TODO: Print target datalayout and triple (minimal required by LLVM)
printGlobalVariable();
for (const auto &iter : functions) {
if (iter.second->getName() == "main") {
printFunction(iter.second.get());
break;
}
}
for (const auto &iter : functions) {
if (iter.second->getName() != "main") {
printFunction(iter.second.get());
}
}
}
std::string SysYPrinter::getTypeString(Type *type) {
if (type->isVoid()) {
return "void";
} else if (type->isInt()) {
return "i32";
} else if (type->isFloat()) {
return "float";
} else if (auto ptrType = dynamic_cast<PointerType*>(type)) {
return getTypeString(ptrType->getBaseType()) + "*";
} else if (auto ptrType = dynamic_cast<FunctionType*>(type)) {
return getTypeString(ptrType->getReturnType());
}
assert(false && "Unsupported type");
return "";
}
std::string SysYPrinter::getValueName(Value *value) {
if (auto global = dynamic_cast<GlobalValue*>(value)) {
return "@" + global->getName();
} else if (auto inst = dynamic_cast<Instruction*>(value)) {
return "%" + inst->getName();
} else if (auto constVal = dynamic_cast<ConstantValue*>(value)) {
if (constVal->isFloat()) {
return std::to_string(constVal->getFloat());
}
return std::to_string(constVal->getInt());
} else if (auto constVar = dynamic_cast<ConstantVariable*>(value)) {
return constVar->getName();
}
assert(false && "Unknown value type");
return "";
}
void SysYPrinter::printType(Type *type) {
std::cout << getTypeString(type);
}
void SysYPrinter::printValue(Value *value) {
std::cout << getValueName(value);
}
void SysYPrinter::printGlobalVariable() {
auto &globals = pModule->getGlobals();
for (const auto &global : globals) {
std::cout << "@" << global->getName() << " = global ";
auto baseType = dynamic_cast<PointerType *>(global->getType())->getBaseType();
printType(baseType);
if (global->getNumDims() > 0) {
// Array type
std::cout << " [";
for (unsigned i = 0; i < global->getNumDims(); i++) {
if (i > 0) std::cout << " x ";
std::cout << getValueName(global->getDim(i));
}
std::cout << "]";
}
std::cout << " ";
if (global->getNumDims() > 0) {
// Array initializer
std::cout << "[";
auto values = global->getInitValues();
auto counterValues = values.getValues();
auto counterNumbers = values.getNumbers();
for (size_t i = 0; i < counterNumbers.size(); i++) {
if (i > 0) std::cout << ", ";
if (baseType->isFloat()) {
std::cout << "float " << dynamic_cast<ConstantValue*>(counterValues[i])->getFloat();
} else {
std::cout << "i32 " << dynamic_cast<ConstantValue*>(counterValues[i])->getInt();
}
}
std::cout << "]";
} else {
// Scalar initializer
if (baseType->isFloat()) {
std::cout << "float " << dynamic_cast<ConstantValue*>(global->getByIndex(0))->getFloat();
} else {
std::cout << "i32 " << dynamic_cast<ConstantValue*>(global->getByIndex(0))->getInt();
}
}
std::cout << ", align 4" << std::endl;
}
}
void SysYPrinter::printFunction(Function *function) {
// Function signature
std::cout << "define ";
printType(function->getReturnType());
std::cout << " @" << function->getName() << "(";
auto entryBlock = function->getEntryBlock();
const auto &args_types = function->getParamTypes();
auto &args = entryBlock->getArguments();
int i = 0;
for (const auto &args_type : args_types) {
if (i > 0) std::cout << ", ";
printType(args_type);
std::cout << " %" << args[i]->getName();
i++;
}
std::cout << ") {" << std::endl;
// Function body
for (const auto &blockIter : function->getBasicBlocks()) {
// Basic block label
BasicBlock* blockPtr = blockIter.get();
if (blockPtr == function->getEntryBlock()) {
std::cout << "entry:" << std::endl;
} else if (!blockPtr->getName().empty()) {
std::cout << blockPtr->getName() << ":" << std::endl;
}
// Instructions
for (const auto &instIter : blockIter->getInstructions()) {
auto inst = instIter.get();
std::cout << " ";
printInst(inst);
}
}
std::cout << "}" << std::endl << std::endl;
}
void SysYPrinter::printInst(Instruction *pInst) {
using Kind = Instruction::Kind;
switch (pInst->getKind()) {
case Kind::kAdd:
case Kind::kSub:
case Kind::kMul:
case Kind::kDiv:
case Kind::kRem:
case Kind::kFAdd:
case Kind::kFSub:
case Kind::kFMul:
case Kind::kFDiv:
case Kind::kICmpEQ:
case Kind::kICmpNE:
case Kind::kICmpLT:
case Kind::kICmpGT:
case Kind::kICmpLE:
case Kind::kICmpGE:
case Kind::kFCmpEQ:
case Kind::kFCmpNE:
case Kind::kFCmpLT:
case Kind::kFCmpGT:
case Kind::kFCmpLE:
case Kind::kFCmpGE:
case Kind::kAnd:
case Kind::kOr: {
auto binInst = dynamic_cast<BinaryInst *>(pInst);
// Print result variable if exists
if (!binInst->getName().empty()) {
std::cout << "%" << binInst->getName() << " = ";
}
// Operation name
switch (pInst->getKind()) {
case Kind::kAdd: std::cout << "add"; break;
case Kind::kSub: std::cout << "sub"; break;
case Kind::kMul: std::cout << "mul"; break;
case Kind::kDiv: std::cout << "sdiv"; break;
case Kind::kRem: std::cout << "srem"; break;
case Kind::kFAdd: std::cout << "fadd"; break;
case Kind::kFSub: std::cout << "fsub"; break;
case Kind::kFMul: std::cout << "fmul"; break;
case Kind::kFDiv: std::cout << "fdiv"; break;
case Kind::kICmpEQ: std::cout << "icmp eq"; break;
case Kind::kICmpNE: std::cout << "icmp ne"; break;
case Kind::kICmpLT: std::cout << "icmp slt"; break;
case Kind::kICmpGT: std::cout << "icmp sgt"; break;
case Kind::kICmpLE: std::cout << "icmp sle"; break;
case Kind::kICmpGE: std::cout << "icmp sge"; break;
case Kind::kFCmpEQ: std::cout << "fcmp oeq"; break;
case Kind::kFCmpNE: std::cout << "fcmp one"; break;
case Kind::kFCmpLT: std::cout << "fcmp olt"; break;
case Kind::kFCmpGT: std::cout << "fcmp ogt"; break;
case Kind::kFCmpLE: std::cout << "fcmp ole"; break;
case Kind::kFCmpGE: std::cout << "fcmp oge"; break;
case Kind::kAnd: std::cout << "and"; break;
case Kind::kOr: std::cout << "or"; break;
default: break;
}
// Types and operands
std::cout << " ";
printType(binInst->getType());
std::cout << " ";
printValue(binInst->getLhs());
std::cout << ", ";
printValue(binInst->getRhs());
std::cout << std::endl;
} break;
case Kind::kNeg:
case Kind::kNot:
case Kind::kFNeg:
case Kind::kFNot:
case Kind::kFtoI:
case Kind::kBitFtoI:
case Kind::kItoF:
case Kind::kBitItoF: {
auto unyInst = dynamic_cast<UnaryInst *>(pInst);
if (!unyInst->getName().empty()) {
std::cout << "%" << unyInst->getName() << " = ";
}
switch (pInst->getKind()) {
case Kind::kNeg: std::cout << "sub "; break;
case Kind::kNot: std::cout << "not "; break;
case Kind::kFNeg: std::cout << "fneg "; break;
case Kind::kFNot: std::cout << "fneg "; break; // FNot not standard, map to fneg
case Kind::kFtoI: std::cout << "fptosi "; break;
case Kind::kBitFtoI: std::cout << "bitcast "; break;
case Kind::kItoF: std::cout << "sitofp "; break;
case Kind::kBitItoF: std::cout << "bitcast "; break;
default: break;
}
printType(unyInst->getType());
std::cout << " ";
// Special handling for negation
if (pInst->getKind() == Kind::kNeg || pInst->getKind() == Kind::kNot) {
std::cout << "i32 0, ";
}
printValue(pInst->getOperand(0));
// For bitcast, need to specify destination type
if (pInst->getKind() == Kind::kBitFtoI || pInst->getKind() == Kind::kBitItoF) {
std::cout << " to ";
printType(unyInst->getType());
}
std::cout << std::endl;
} break;
case Kind::kCall: {
auto callInst = dynamic_cast<CallInst *>(pInst);
auto function = callInst->getCallee();
if (!callInst->getName().empty()) {
std::cout << "%" << callInst->getName() << " = ";
}
std::cout << "call ";
printType(callInst->getType());
std::cout << " @" << function->getName() << "(";
auto params = callInst->getArguments();
bool first = true;
for (auto &param : params) {
if (!first) std::cout << ", ";
first = false;
printType(param->getValue()->getType());
std::cout << " ";
printValue(param->getValue());
}
std::cout << ")" << std::endl;
} break;
case Kind::kCondBr: {
auto condBrInst = dynamic_cast<CondBrInst *>(pInst);
std::cout << "br i1 ";
printValue(condBrInst->getCondition());
std::cout << ", label %" << condBrInst->getThenBlock()->getName();
std::cout << ", label %" << condBrInst->getElseBlock()->getName();
std::cout << std::endl;
} break;
case Kind::kBr: {
auto brInst = dynamic_cast<UncondBrInst *>(pInst);
std::cout << "br label %" << brInst->getBlock()->getName();
std::cout << std::endl;
} break;
case Kind::kReturn: {
auto retInst = dynamic_cast<ReturnInst *>(pInst);
std::cout << "ret ";
if (retInst->getNumOperands() != 0) {
printType(retInst->getOperand(0)->getType());
std::cout << " ";
printValue(retInst->getOperand(0));
} else {
std::cout << "void";
}
std::cout << std::endl;
} break;
case Kind::kAlloca: {
auto allocaInst = dynamic_cast<AllocaInst *>(pInst);
std::cout << "%" << allocaInst->getName() << " = alloca ";
auto baseType = dynamic_cast<PointerType *>(allocaInst->getType())->getBaseType();
printType(baseType);
if (allocaInst->getNumDims() > 0) {
std::cout << ", ";
for (size_t i = 0; i < allocaInst->getNumDims(); i++) {
if (i > 0) std::cout << ", ";
printType(Type::getIntType());
std::cout << " ";
printValue(allocaInst->getDim(i));
}
}
std::cout << ", align 4" << std::endl;
} break;
case Kind::kLoad: {
auto loadInst = dynamic_cast<LoadInst *>(pInst);
std::cout << "%" << loadInst->getName() << " = load ";
printType(loadInst->getType());
std::cout << ", ";
printType(loadInst->getPointer()->getType());
std::cout << " ";
printValue(loadInst->getPointer());
if (loadInst->getNumIndices() > 0) {
std::cout << ", ";
for (size_t i = 0; i < loadInst->getNumIndices(); i++) {
if (i > 0) std::cout << ", ";
printType(Type::getIntType());
std::cout << " ";
printValue(loadInst->getIndex(i));
}
}
std::cout << ", align 4" << std::endl;
} break;
case Kind::kLa: {
auto laInst = dynamic_cast<LaInst *>(pInst);
std::cout << "%" << laInst->getName() << " = getelementptr inbounds ";
auto ptrType = dynamic_cast<PointerType*>(laInst->getPointer()->getType());
printType(ptrType->getBaseType());
std::cout << ", ";
printType(laInst->getPointer()->getType());
std::cout << " ";
printValue(laInst->getPointer());
std::cout << ", ";
for (size_t i = 0; i < laInst->getNumIndices(); i++) {
if (i > 0) std::cout << ", ";
printType(Type::getIntType());
std::cout << " ";
printValue(laInst->getIndex(i));
}
std::cout << std::endl;
} break;
case Kind::kStore: {
auto storeInst = dynamic_cast<StoreInst *>(pInst);
std::cout << "store ";
printType(storeInst->getValue()->getType());
std::cout << " ";
printValue(storeInst->getValue());
std::cout << ", ";
printType(storeInst->getPointer()->getType());
std::cout << " ";
printValue(storeInst->getPointer());
if (storeInst->getNumIndices() > 0) {
std::cout << ", ";
for (size_t i = 0; i < storeInst->getNumIndices(); i++) {
if (i > 0) std::cout << ", ";
printType(Type::getIntType());
std::cout << " ";
printValue(storeInst->getIndex(i));
}
}
std::cout << ", align 4" << std::endl;
} break;
case Kind::kMemset: {
auto memsetInst = dynamic_cast<MemsetInst *>(pInst);
std::cout << "call void @llvm.memset.p0.";
printType(memsetInst->getPointer()->getType());
std::cout << "(";
printType(memsetInst->getPointer()->getType());
std::cout << " ";
printValue(memsetInst->getPointer());
std::cout << ", i8 ";
printValue(memsetInst->getValue());
std::cout << ", i32 ";
printValue(memsetInst->getSize());
std::cout << ", i1 false)" << std::endl;
} break;
case Kind::kPhi: {
auto phiInst = dynamic_cast<PhiInst *>(pInst);
printValue(phiInst->getOperand(0));
std::cout << " = phi ";
printType(phiInst->getType());
for (unsigned i = 1; i < phiInst->getNumOperands(); i++) {
if (i > 0) std::cout << ", ";
std::cout << "[ ";
printValue(phiInst->getOperand(i));
std::cout << " ]";
}
std::cout << std::endl;
} break;
case Kind::kGetSubArray: {
auto getSubArrayInst = dynamic_cast<GetSubArrayInst *>(pInst);
std::cout << "%" << getSubArrayInst->getName() << " = getelementptr inbounds ";
auto ptrType = dynamic_cast<PointerType*>(getSubArrayInst->getFatherArray()->getType());
printType(ptrType->getBaseType());
std::cout << ", ";
printType(getSubArrayInst->getFatherArray()->getType());
std::cout << " ";
printValue(getSubArrayInst->getFatherArray());
std::cout << ", ";
bool firstIndex = true;
for (auto &index : getSubArrayInst->getIndices()) {
if (!firstIndex) std::cout << ", ";
firstIndex = false;
printType(Type::getIntType());
std::cout << " ";
printValue(index->getValue());
}
std::cout << std::endl;
} break;
default:
assert(false && "Unsupported instruction kind");
break;
}
}
} // namespace sysy

View File

@ -1,26 +0,0 @@
# src/backend/RISCv64/CMakeLists.txt
add_library(riscv64_backend_lib STATIC
RISCv64AsmPrinter.cpp
RISCv64Backend.cpp
RISCv64ISel.cpp
RISCv64LLIR.cpp
RISCv64RegAlloc.cpp
RISCv64LinearScan.cpp
Handler/CalleeSavedHandler.cpp
Handler/LegalizeImmediates.cpp
Handler/PrologueEpilogueInsertion.cpp
Handler/EliminateFrameIndices.cpp
Optimize/Peephole.cpp
Optimize/PostRA_Scheduler.cpp
Optimize/PreRA_Scheduler.cpp
Optimize/DivStrengthReduction.cpp
)
# 包含后端模块所需的头文件路径
target_include_directories(riscv64_backend_lib PUBLIC
${CMAKE_CURRENT_SOURCE_DIR}/../../include/backend/RISCv64 # 后端顶层头文件
${CMAKE_CURRENT_SOURCE_DIR}/../../include/backend/RISCv64/Handler # 增加 Handler 头文件路径
${CMAKE_CURRENT_SOURCE_DIR}/../../include/backend/RISCv64/Optimize # 增加 Optimize 头文件路径
${CMAKE_CURRENT_SOURCE_DIR}/../../include/midend # 增加 midend 头文件路径 (已存在)
${CMAKE_CURRENT_SOURCE_DIR}/../../include/midend/Pass # 增加 midend 头文件路径 (已存在)
)

View File

@ -1,51 +0,0 @@
#include "CalleeSavedHandler.h"
#include <set>
#include <vector>
#include <algorithm>
#include <iterator>
namespace sysy {
char CalleeSavedHandler::ID = 0;
bool CalleeSavedHandler::runOnFunction(Function *F, AnalysisManager& AM) {
// This pass works on MachineFunction level, not IR level
return false;
}
void CalleeSavedHandler::runOnMachineFunction(MachineFunction* mfunc) {
StackFrameInfo& frame_info = mfunc->getFrameInfo();
const std::set<PhysicalReg>& used_callee_saved = frame_info.used_callee_saved_regs;
if (used_callee_saved.empty()) {
frame_info.callee_saved_size = 0;
frame_info.callee_saved_regs_to_store.clear();
return;
}
// 1. 计算被调用者保存寄存器所需的总空间大小
// s0 总是由 PEI Pass 单独处理,这里不计入大小,但要确保它在列表中
int size = 0;
std::set<PhysicalReg> regs_to_save = used_callee_saved;
if (regs_to_save.count(PhysicalReg::S0)) {
regs_to_save.erase(PhysicalReg::S0);
}
size = regs_to_save.size() * 8; // 每个寄存器占8字节 (64-bit)
frame_info.callee_saved_size = size;
// 2. 创建一个有序的、需要保存的寄存器列表,以便后续 Pass 确定地生成代码
// s0 不应包含在此列表中,因为它由 PEI Pass 特殊处理
std::vector<PhysicalReg> sorted_regs(regs_to_save.begin(), regs_to_save.end());
std::sort(sorted_regs.begin(), sorted_regs.end(), [](PhysicalReg a, PhysicalReg b){
return static_cast<int>(a) < static_cast<int>(b);
});
frame_info.callee_saved_regs_to_store = sorted_regs;
// 3. 更新栈帧总大小。
// 这是初步计算PEI Pass 会进行最终的对齐。
frame_info.total_size = frame_info.locals_size +
frame_info.spill_size +
frame_info.callee_saved_size;
}
} // namespace sysy

View File

@ -1,235 +0,0 @@
#include "EliminateFrameIndices.h"
#include "RISCv64ISel.h"
#include <cassert>
#include <vector>
namespace sysy {
// getTypeSizeInBytes 是一个通用辅助函数,保持不变
unsigned EliminateFrameIndicesPass::getTypeSizeInBytes(Type* type) {
if (!type) {
assert(false && "Cannot get size of a null type.");
return 0;
}
switch (type->getKind()) {
case Type::kInt:
case Type::kFloat:
return 4;
case Type::kPointer:
return 8;
case Type::kArray: {
auto arrayType = type->as<ArrayType>();
return arrayType->getNumElements() * getTypeSizeInBytes(arrayType->getElementType());
}
default:
assert(false && "Unsupported type for size calculation.");
return 0;
}
}
void EliminateFrameIndicesPass::runOnMachineFunction(MachineFunction* mfunc) {
StackFrameInfo& frame_info = mfunc->getFrameInfo();
Function* F = mfunc->getFunc();
RISCv64ISel* isel = mfunc->getISel();
// 在这里处理栈传递的参数以便在寄存器分配前就将数据流显式化修复溢出逻辑的BUG。
// 2. 只为局部变量(AllocaInst)分配栈空间和计算偏移量
// 局部变量从 s0 下方(负偏移量)开始分配,紧接着为 ra 和 s0 预留的16字节之后
int local_var_offset = 16;
if(F) { // 确保函数指针有效
for (auto& bb : F->getBasicBlocks()) {
for (auto& inst : bb->getInstructions()) {
if (auto alloca = dynamic_cast<AllocaInst*>(inst.get())) {
Type* allocated_type = alloca->getType()->as<PointerType>()->getBaseType();
int size = getTypeSizeInBytes(allocated_type);
// 优化栈帧大小对于大数组使用4字节对齐小对象使用8字节对齐
if (size >= 256) { // 大数组优化
size = (size + 3) & ~3; // 4字节对齐
} else {
size = (size + 7) & ~7; // 8字节对齐
}
if (size == 0) size = 4; // 最小4字节
local_var_offset += size;
unsigned alloca_vreg = isel->getVReg(alloca);
// 局部变量使用相对于s0的负向偏移
frame_info.alloca_offsets[alloca_vreg] = -local_var_offset;
}
}
}
}
// 记录仅由AllocaInst分配的局部变量的总大小
frame_info.locals_size = local_var_offset - 16;
// 记录局部变量区域分配结束的最终偏移量
frame_info.locals_end_offset = -local_var_offset;
// 在函数入口为所有栈传递的参数插入load指令
// 这个步骤至关重要它在寄存器分配之前为这些参数的vreg创建了明确的“定义(def)”指令。
// 这解决了在高寄存器压力下当这些vreg被溢出时`rewriteProgram`找不到其定义点而崩溃的问题。
if (F && isel && !mfunc->getBlocks().empty()) {
MachineBasicBlock* entry_block = mfunc->getBlocks().front().get();
std::vector<std::unique_ptr<MachineInstr>> arg_load_instrs;
// 步骤 3.1: 生成所有加载栈参数的指令,暂存起来
int arg_idx = 0;
for (Argument* arg : F->getArguments()) {
// 根据ABI前8个整型/指针参数通过寄存器传递,这里只处理超出部分。
if (arg_idx >= 8) {
// 计算参数在调用者栈帧中的位置该位置相对于被调用者的帧指针s0是正向偏移。
// 第9个参数(arg_idx=8)位于 0(s0)第10个(arg_idx=9)位于 8(s0),以此类推。
int offset = (arg_idx - 8) * 8;
unsigned arg_vreg = isel->getVReg(arg);
Type* arg_type = arg->getType();
// 根据参数类型选择正确的加载指令
RVOpcodes load_op;
if (arg_type->isFloat()) {
load_op = RVOpcodes::FLW; // 单精度浮点
} else if (arg_type->isPointer()) {
load_op = RVOpcodes::LD; // 64位指针
} else {
load_op = RVOpcodes::LW; // 32位整数
}
// 创建加载指令: lw/ld/flw vreg, offset(s0)
auto load_instr = std::make_unique<MachineInstr>(load_op);
load_instr->addOperand(std::make_unique<RegOperand>(arg_vreg));
load_instr->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0), // 基址为帧指针
std::make_unique<ImmOperand>(offset)
));
arg_load_instrs.push_back(std::move(load_instr));
}
arg_idx++;
}
//仅当有需要加载的栈参数时,才执行插入逻辑
if (!arg_load_instrs.empty()) {
auto& entry_instrs = entry_block->getInstructions();
auto insertion_point = entry_instrs.begin(); // 默认插入点为块的开头
auto last_arg_save_it = entry_instrs.end();
// 步骤 3.2: 寻找一个安全的插入点。
// 遍历入口块的指令,找到最后一条保存“寄存器传递参数”的伪指令。
// 这样可以确保我们在所有 a0-a7 参数被保存之后,才执行可能覆盖它们的加载指令。
for (auto it = entry_instrs.begin(); it != entry_instrs.end(); ++it) {
MachineInstr* instr = it->get();
// 寻找代表保存参数到栈的伪指令
if (instr->getOpcode() == RVOpcodes::FRAME_STORE_W ||
instr->getOpcode() == RVOpcodes::FRAME_STORE_D ||
instr->getOpcode() == RVOpcodes::FRAME_STORE_F) {
// 检查被保存的值是否是寄存器参数 (arg_no < 8)
auto& operands = instr->getOperands();
if (operands.empty() || operands[0]->getKind() != MachineOperand::KIND_REG) continue;
unsigned src_vreg = static_cast<RegOperand*>(operands[0].get())->getVRegNum();
Value* ir_value = isel->getVRegValueMap().count(src_vreg) ? isel->getVRegValueMap().at(src_vreg) : nullptr;
if (auto ir_arg = dynamic_cast<Argument*>(ir_value)) {
if (ir_arg->getIndex() < 8) {
last_arg_save_it = it; // 找到了一个保存寄存器参数的指令,更新位置
}
}
}
}
// 如果找到了这样的保存指令,我们的插入点就在它之后
if (last_arg_save_it != entry_instrs.end()) {
insertion_point = std::next(last_arg_save_it);
}
// 步骤 3.3: 在计算出的安全位置,一次性插入所有新创建的参数加载指令
entry_instrs.insert(insertion_point,
std::make_move_iterator(arg_load_instrs.begin()),
std::make_move_iterator(arg_load_instrs.end()));
}
}
// 4. 遍历所有机器指令,将访问局部变量的伪指令展开为真实指令
for (auto& mbb : mfunc->getBlocks()) {
std::vector<std::unique_ptr<MachineInstr>> new_instructions;
for (auto& instr_ptr : mbb->getInstructions()) {
RVOpcodes opcode = instr_ptr->getOpcode();
if (opcode == RVOpcodes::FRAME_LOAD_W || opcode == RVOpcodes::FRAME_LOAD_D || opcode == RVOpcodes::FRAME_LOAD_F) {
RVOpcodes real_load_op;
if (opcode == RVOpcodes::FRAME_LOAD_W) real_load_op = RVOpcodes::LW;
else if (opcode == RVOpcodes::FRAME_LOAD_D) real_load_op = RVOpcodes::LD;
else real_load_op = RVOpcodes::FLW;
auto& operands = instr_ptr->getOperands();
unsigned dest_vreg = static_cast<RegOperand*>(operands[0].get())->getVRegNum();
unsigned alloca_vreg = static_cast<RegOperand*>(operands[1].get())->getVRegNum();
int offset = frame_info.alloca_offsets.at(alloca_vreg);
auto addr_vreg = isel->getNewVReg(Type::getPointerType(Type::getIntType()));
// 展开为: addi addr_vreg, s0, offset
auto addi = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
addi->addOperand(std::make_unique<RegOperand>(addr_vreg));
addi->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
addi->addOperand(std::make_unique<ImmOperand>(offset));
new_instructions.push_back(std::move(addi));
// 展开为: lw/ld/flw dest_vreg, 0(addr_vreg)
auto load_instr = std::make_unique<MachineInstr>(real_load_op);
load_instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
load_instr->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(addr_vreg),
std::make_unique<ImmOperand>(0)));
new_instructions.push_back(std::move(load_instr));
} else if (opcode == RVOpcodes::FRAME_STORE_W || opcode == RVOpcodes::FRAME_STORE_D || opcode == RVOpcodes::FRAME_STORE_F) {
RVOpcodes real_store_op;
if (opcode == RVOpcodes::FRAME_STORE_W) real_store_op = RVOpcodes::SW;
else if (opcode == RVOpcodes::FRAME_STORE_D) real_store_op = RVOpcodes::SD;
else real_store_op = RVOpcodes::FSW;
auto& operands = instr_ptr->getOperands();
unsigned src_vreg = static_cast<RegOperand*>(operands[0].get())->getVRegNum();
unsigned alloca_vreg = static_cast<RegOperand*>(operands[1].get())->getVRegNum();
int offset = frame_info.alloca_offsets.at(alloca_vreg);
auto addr_vreg = isel->getNewVReg(Type::getPointerType(Type::getIntType()));
// 展开为: addi addr_vreg, s0, offset
auto addi = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
addi->addOperand(std::make_unique<RegOperand>(addr_vreg));
addi->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
addi->addOperand(std::make_unique<ImmOperand>(offset));
new_instructions.push_back(std::move(addi));
// 展开为: sw/sd/fsw src_vreg, 0(addr_vreg)
auto store_instr = std::make_unique<MachineInstr>(real_store_op);
store_instr->addOperand(std::make_unique<RegOperand>(src_vreg));
store_instr->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(addr_vreg),
std::make_unique<ImmOperand>(0)));
new_instructions.push_back(std::move(store_instr));
} else if (instr_ptr->getOpcode() == RVOpcodes::FRAME_ADDR) {
auto& operands = instr_ptr->getOperands();
unsigned dest_vreg = static_cast<RegOperand*>(operands[0].get())->getVRegNum();
unsigned alloca_vreg = static_cast<RegOperand*>(operands[1].get())->getVRegNum();
int offset = frame_info.alloca_offsets.at(alloca_vreg);
// 将 `frame_addr rd, rs` 展开为 `addi rd, s0, offset`
auto addi = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
addi->addOperand(std::make_unique<RegOperand>(dest_vreg));
addi->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
addi->addOperand(std::make_unique<ImmOperand>(offset));
new_instructions.push_back(std::move(addi));
} else {
new_instructions.push_back(std::move(instr_ptr));
}
}
mbb->getInstructions() = std::move(new_instructions);
}
}
} // namespace sysy

View File

@ -1,171 +0,0 @@
#include "LegalizeImmediates.h"
#include "RISCv64ISel.h" // 需要包含它以调用 getNewVReg()
#include "RISCv64AsmPrinter.h"
#include <vector>
#include <iostream>
// 声明外部调试控制变量
extern int DEBUG;
extern int DEEPDEBUG;
namespace sysy {
char LegalizeImmediatesPass::ID = 0;
// 辅助函数检查一个立即数是否在RISC-V的12位有符号范围内
static bool isLegalImmediate(int64_t imm) {
return imm >= -2048 && imm <= 2047;
}
void LegalizeImmediatesPass::runOnMachineFunction(MachineFunction* mfunc) {
if (DEBUG) {
std::cerr << "===== Running Legalize Immediates Pass on function: " << mfunc->getName() << " =====\n";
}
// 定义我们保留的、用于暂存的物理寄存器
const PhysicalReg TEMP_REG = PhysicalReg::T5;
// 创建一个临时的AsmPrinter用于打印指令方便调试
RISCv64AsmPrinter temp_printer(mfunc);
if (DEEPDEBUG) {
temp_printer.setStream(std::cerr);
}
for (auto& mbb : mfunc->getBlocks()) {
if (DEEPDEBUG) {
std::cerr << "--- Processing Basic Block: " << mbb->getName() << " ---\n";
}
// 创建一个新的指令列表,用于存放合法化后的指令
std::vector<std::unique_ptr<MachineInstr>> new_instructions;
for (auto& instr_ptr : mbb->getInstructions()) {
if (DEEPDEBUG) {
std::cerr << " Checking: ";
// 打印指令时末尾会带换行符,所以这里不用 std::endl
temp_printer.printInstruction(instr_ptr.get(), true);
}
bool legalized = false; // 标记当前指令是否已被展开处理
switch (instr_ptr->getOpcode()) {
case RVOpcodes::ADDI:
case RVOpcodes::ADDIW: {
auto& operands = instr_ptr->getOperands();
// 确保操作数足够多,以防万一
if (operands.size() < 3) break;
auto imm_op = static_cast<ImmOperand*>(operands.back().get());
if (!isLegalImmediate(imm_op->getValue())) {
if (DEEPDEBUG) {
std::cerr << " >> ILLEGAL immediate (" << imm_op->getValue() << "). Expanding...\n";
}
// 立即数超出范围,需要展开
auto rd_op = std::make_unique<RegOperand>(*static_cast<RegOperand*>(operands[0].get()));
auto rs1_op = std::make_unique<RegOperand>(*static_cast<RegOperand*>(operands[1].get()));
// 1. li t5, immediate
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(TEMP_REG));
li->addOperand(std::make_unique<ImmOperand>(imm_op->getValue()));
// 2. add/addw rd, rs1, t5
auto new_op = (instr_ptr->getOpcode() == RVOpcodes::ADDI) ? RVOpcodes::ADD : RVOpcodes::ADDW;
auto add = std::make_unique<MachineInstr>(new_op);
add->addOperand(std::move(rd_op));
add->addOperand(std::move(rs1_op));
add->addOperand(std::make_unique<RegOperand>(TEMP_REG));
if (DEEPDEBUG) {
std::cerr << " New sequence:\n ";
temp_printer.printInstruction(li.get(), true);
std::cerr << " ";
temp_printer.printInstruction(add.get(), true);
}
new_instructions.push_back(std::move(li));
new_instructions.push_back(std::move(add));
legalized = true;
}
break;
}
// 处理所有内存加载/存储指令
case RVOpcodes::LB: case RVOpcodes::LH: case RVOpcodes::LW: case RVOpcodes::LD:
case RVOpcodes::LBU: case RVOpcodes::LHU: case RVOpcodes::LWU:
case RVOpcodes::SB: case RVOpcodes::SH: case RVOpcodes::SW: case RVOpcodes::SD:
case RVOpcodes::FLW: case RVOpcodes::FSW: {
auto& operands = instr_ptr->getOperands();
auto mem_op = static_cast<MemOperand*>(operands.back().get());
auto offset_op = mem_op->getOffset();
if (!isLegalImmediate(offset_op->getValue())) {
if (DEEPDEBUG) {
std::cerr << " >> ILLEGAL immediate offset (" << offset_op->getValue() << "). Expanding...\n";
}
// 偏移量超出范围,需要展开
auto data_reg_op = std::make_unique<RegOperand>(*static_cast<RegOperand*>(operands[0].get()));
auto base_reg_op = std::make_unique<RegOperand>(*mem_op->getBase());
// 1. li t5, offset
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(TEMP_REG));
li->addOperand(std::make_unique<ImmOperand>(offset_op->getValue()));
// 2. add t5, base_reg, t5 (计算最终地址结果也放在t5)
auto add = std::make_unique<MachineInstr>(RVOpcodes::ADD);
add->addOperand(std::make_unique<RegOperand>(TEMP_REG));
add->addOperand(std::move(base_reg_op));
add->addOperand(std::make_unique<RegOperand>(TEMP_REG));
// 3. lw/sw data_reg, 0(t5)
auto mem_instr = std::make_unique<MachineInstr>(instr_ptr->getOpcode());
mem_instr->addOperand(std::move(data_reg_op));
mem_instr->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(TEMP_REG),
std::make_unique<ImmOperand>(0)
));
if (DEEPDEBUG) {
std::cerr << " New sequence:\n ";
temp_printer.printInstruction(li.get(), true);
std::cerr << " ";
temp_printer.printInstruction(add.get(), true);
std::cerr << " ";
temp_printer.printInstruction(mem_instr.get(), true);
}
new_instructions.push_back(std::move(li));
new_instructions.push_back(std::move(add));
new_instructions.push_back(std::move(mem_instr));
legalized = true;
}
break;
}
default:
// 其他指令不需要处理
break;
}
if (!legalized) {
if (DEEPDEBUG) {
std::cerr << " -- Immediate is legal. Skipping.\n";
}
// 如果当前指令不需要合法化,直接将其移动到新列表中
new_instructions.push_back(std::move(instr_ptr));
}
}
// 用新的、已合法化的指令列表替换旧的列表
mbb->getInstructions() = std::move(new_instructions);
}
if (DEBUG) {
std::cerr << "===== Finished Legalize Immediates Pass =====\n\n";
}
}
} // namespace sysy

View File

@ -1,182 +0,0 @@
#include "PrologueEpilogueInsertion.h"
#include "RISCv64LLIR.h" // 假设包含了 PhysicalReg, RVOpcodes 等定义
#include "RISCv64ISel.h"
#include <algorithm>
#include <vector>
#include <set>
namespace sysy {
char PrologueEpilogueInsertionPass::ID = 0;
void PrologueEpilogueInsertionPass::runOnMachineFunction(MachineFunction* mfunc) {
StackFrameInfo& frame_info = mfunc->getFrameInfo();
Function* F = mfunc->getFunc();
RISCv64ISel* isel = mfunc->getISel();
// 1. 清理 KEEPALIVE 伪指令
for (auto& mbb : mfunc->getBlocks()) {
auto& instrs = mbb->getInstructions();
instrs.erase(
std::remove_if(instrs.begin(), instrs.end(),
[](const std::unique_ptr<MachineInstr>& instr) {
return instr->getOpcode() == RVOpcodes::PSEUDO_KEEPALIVE;
}
),
instrs.end()
);
}
// 2. 确定需要保存的被调用者保存寄存器 (callee-saved)
auto& vreg_to_preg_map = frame_info.vreg_to_preg_map;
std::set<PhysicalReg> used_callee_saved_regs_set;
const auto& callee_saved_int = getCalleeSavedIntRegs();
const auto& callee_saved_fp = getCalleeSavedFpRegs();
for (const auto& pair : vreg_to_preg_map) {
PhysicalReg preg = pair.second;
bool is_int_cs = std::find(callee_saved_int.begin(), callee_saved_int.end(), preg) != callee_saved_int.end();
bool is_fp_cs = std::find(callee_saved_fp.begin(), callee_saved_fp.end(), preg) != callee_saved_fp.end();
if ((is_int_cs && preg != PhysicalReg::S0) || is_fp_cs) {
used_callee_saved_regs_set.insert(preg);
}
}
frame_info.callee_saved_regs_to_store.assign(
used_callee_saved_regs_set.begin(), used_callee_saved_regs_set.end()
);
std::sort(frame_info.callee_saved_regs_to_store.begin(), frame_info.callee_saved_regs_to_store.end());
frame_info.callee_saved_size = frame_info.callee_saved_regs_to_store.size() * 8;
// 3. 计算最终的栈帧总大小,包含栈溢出保护
int total_stack_size = frame_info.locals_size +
frame_info.spill_size +
frame_info.callee_saved_size +
16;
// 栈溢出保护:增加最大栈帧大小以容纳大型数组
const int MAX_STACK_FRAME_SIZE = 8192; // 8KB to handle large arrays like 256*4*2 = 2048 bytes
if (total_stack_size > MAX_STACK_FRAME_SIZE) {
// 如果仍然超过限制,尝试优化对齐方式
std::cerr << "Warning: Stack frame size " << total_stack_size
<< " exceeds recommended limit " << MAX_STACK_FRAME_SIZE << " for function "
<< mfunc->getName() << std::endl;
}
// 优化减少对齐开销使用16字节对齐而非更大的对齐
int aligned_stack_size = (total_stack_size + 15) & ~15;
frame_info.total_size = aligned_stack_size;
if (aligned_stack_size > 0) {
// --- 4. 插入完整的序言 ---
MachineBasicBlock* entry_block = mfunc->getBlocks().front().get();
auto& entry_instrs = entry_block->getInstructions();
std::vector<std::unique_ptr<MachineInstr>> prologue_instrs;
// 4.1. 分配栈帧
auto alloc_stack = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
alloc_stack->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
alloc_stack->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
alloc_stack->addOperand(std::make_unique<ImmOperand>(-aligned_stack_size));
prologue_instrs.push_back(std::move(alloc_stack));
// 4.2. 保存 ra 和 s0
auto save_ra = std::make_unique<MachineInstr>(RVOpcodes::SD);
save_ra->addOperand(std::make_unique<RegOperand>(PhysicalReg::RA));
save_ra->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::SP),
std::make_unique<ImmOperand>(aligned_stack_size - 8)
));
prologue_instrs.push_back(std::move(save_ra));
auto save_fp = std::make_unique<MachineInstr>(RVOpcodes::SD);
save_fp->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
save_fp->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::SP),
std::make_unique<ImmOperand>(aligned_stack_size - 16)
));
prologue_instrs.push_back(std::move(save_fp));
// 4.3. 设置新的帧指针 s0
auto set_fp = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
set_fp->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
set_fp->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
set_fp->addOperand(std::make_unique<ImmOperand>(aligned_stack_size));
prologue_instrs.push_back(std::move(set_fp));
// 4.4. 保存所有使用到的被调用者保存寄存器
int next_available_offset = -(16 + frame_info.locals_size + frame_info.spill_size);
for (const auto& reg : frame_info.callee_saved_regs_to_store) {
// 改为“先更新,后使用”逻辑
next_available_offset -= 8; // 先为当前寄存器分配下一个可用槽位
RVOpcodes store_op = isFPR(reg) ? RVOpcodes::FSD : RVOpcodes::SD;
auto save_cs_reg = std::make_unique<MachineInstr>(store_op);
save_cs_reg->addOperand(std::make_unique<RegOperand>(reg));
save_cs_reg->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0),
std::make_unique<ImmOperand>(next_available_offset) // 使用新计算出的正确偏移
));
prologue_instrs.push_back(std::move(save_cs_reg));
// 不再需要在循环末尾递减
}
// 4.5. 将所有生成的序言指令一次性插入到函数入口
entry_instrs.insert(entry_instrs.begin(),
std::make_move_iterator(prologue_instrs.begin()),
std::make_move_iterator(prologue_instrs.end()));
// --- 5. 插入完整的尾声 ---
for (auto& mbb : mfunc->getBlocks()) {
for (auto it = mbb->getInstructions().begin(); it != mbb->getInstructions().end(); ++it) {
if ((*it)->getOpcode() == RVOpcodes::RET) {
std::vector<std::unique_ptr<MachineInstr>> epilogue_instrs;
// 5.1. 恢复被调用者保存寄存器
int next_available_offset_restore = -(16 + frame_info.locals_size + frame_info.spill_size);
for (const auto& reg : frame_info.callee_saved_regs_to_store) {
next_available_offset_restore -= 8; // 为下一个寄存器准备偏移
RVOpcodes load_op = isFPR(reg) ? RVOpcodes::FLD : RVOpcodes::LD;
auto restore_cs_reg = std::make_unique<MachineInstr>(load_op);
restore_cs_reg->addOperand(std::make_unique<RegOperand>(reg));
restore_cs_reg->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0),
std::make_unique<ImmOperand>(next_available_offset_restore) // 使用当前偏移
));
epilogue_instrs.push_back(std::move(restore_cs_reg));
}
// 5.2. 恢复 ra 和 s0
auto restore_ra = std::make_unique<MachineInstr>(RVOpcodes::LD);
restore_ra->addOperand(std::make_unique<RegOperand>(PhysicalReg::RA));
restore_ra->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::SP),
std::make_unique<ImmOperand>(aligned_stack_size - 8)
));
epilogue_instrs.push_back(std::move(restore_ra));
auto restore_fp = std::make_unique<MachineInstr>(RVOpcodes::LD);
restore_fp->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
restore_fp->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::SP),
std::make_unique<ImmOperand>(aligned_stack_size - 16)
));
epilogue_instrs.push_back(std::move(restore_fp));
// 5.3. 释放栈帧
auto dealloc_stack = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
dealloc_stack->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
dealloc_stack->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
dealloc_stack->addOperand(std::make_unique<ImmOperand>(aligned_stack_size));
epilogue_instrs.push_back(std::move(dealloc_stack));
// 将尾声指令插入到 RET 指令之前
mbb->getInstructions().insert(it,
std::make_move_iterator(epilogue_instrs.begin()),
std::make_move_iterator(epilogue_instrs.end()));
goto next_block;
}
}
next_block:;
}
}
}
} // namespace sysy

View File

@ -1,282 +0,0 @@
#include "DivStrengthReduction.h"
#include <cmath>
#include <cstdint>
namespace sysy {
char DivStrengthReduction::ID = 0;
bool DivStrengthReduction::runOnFunction(Function *F, AnalysisManager& AM) {
// This pass works on MachineFunction level, not IR level
return false;
}
void DivStrengthReduction::runOnMachineFunction(MachineFunction *mfunc) {
if (!mfunc)
return;
bool debug = false; // Set to true for debugging
if (debug)
std::cout << "Running DivStrengthReduction optimization..." << std::endl;
int next_temp_reg = 1000;
auto createTempReg = [&]() -> int {
return next_temp_reg++;
};
struct MagicInfo {
int64_t magic;
int shift;
};
auto computeMagic = [](int64_t d, bool is_32bit) -> MagicInfo {
int word_size = is_32bit ? 32 : 64;
uint64_t ad = std::abs(d);
if (ad == 0) return {0, 0};
int l = std::floor(std::log2(ad));
if ((ad & (ad - 1)) == 0) { // power of 2
l = 0; // special case for power of 2, shift will be calculated differently
}
__int128_t one = 1;
__int128_t num;
int total_shift;
if (is_32bit) {
total_shift = 31 + l;
num = one << total_shift;
} else {
total_shift = 63 + l;
num = one << total_shift;
}
__int128_t den = ad;
int64_t magic = (num / den) + 1;
return {magic, total_shift};
};
auto isPowerOfTwo = [](int64_t n) -> bool {
return n > 0 && (n & (n - 1)) == 0;
};
auto getPowerOfTwoExponent = [](int64_t n) -> int {
if (n <= 0 || (n & (n - 1)) != 0) return -1;
int shift = 0;
while (n > 1) {
n >>= 1;
shift++;
}
return shift;
};
struct InstructionReplacement {
size_t index;
size_t count_to_erase;
std::vector<std::unique_ptr<MachineInstr>> newInstrs;
};
for (auto &mbb_uptr : mfunc->getBlocks()) {
auto &mbb = *mbb_uptr;
auto &instrs = mbb.getInstructions();
std::vector<InstructionReplacement> replacements;
for (size_t i = 0; i < instrs.size(); ++i) {
auto *instr = instrs[i].get();
bool is_32bit = (instr->getOpcode() == RVOpcodes::DIVW);
if (instr->getOpcode() != RVOpcodes::DIV && !is_32bit) {
continue;
}
if (instr->getOperands().size() != 3) {
continue;
}
auto *dst_op = instr->getOperands()[0].get();
auto *src1_op = instr->getOperands()[1].get();
auto *src2_op = instr->getOperands()[2].get();
int64_t divisor = 0;
bool const_divisor_found = false;
size_t instructions_to_replace = 1;
if (src2_op->getKind() == MachineOperand::KIND_IMM) {
divisor = static_cast<ImmOperand *>(src2_op)->getValue();
const_divisor_found = true;
} else if (src2_op->getKind() == MachineOperand::KIND_REG) {
if (i > 0) {
auto *prev_instr = instrs[i - 1].get();
if (prev_instr->getOpcode() == RVOpcodes::LI && prev_instr->getOperands().size() == 2) {
auto *li_dst_op = prev_instr->getOperands()[0].get();
auto *li_imm_op = prev_instr->getOperands()[1].get();
if (li_dst_op->getKind() == MachineOperand::KIND_REG && li_imm_op->getKind() == MachineOperand::KIND_IMM) {
auto *div_reg_op = static_cast<RegOperand *>(src2_op);
auto *li_dst_reg_op = static_cast<RegOperand *>(li_dst_op);
if (div_reg_op->isVirtual() && li_dst_reg_op->isVirtual() &&
div_reg_op->getVRegNum() == li_dst_reg_op->getVRegNum()) {
divisor = static_cast<ImmOperand *>(li_imm_op)->getValue();
const_divisor_found = true;
instructions_to_replace = 2;
}
}
}
}
}
if (!const_divisor_found) {
continue;
}
auto *dst_reg = static_cast<RegOperand *>(dst_op);
auto *src1_reg = static_cast<RegOperand *>(src1_op);
if (divisor == 0) continue;
std::vector<std::unique_ptr<MachineInstr>> newInstrs;
if (divisor == 1) {
auto moveInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::ADDW : RVOpcodes::ADD);
moveInstr->addOperand(std::make_unique<RegOperand>(*dst_reg));
moveInstr->addOperand(std::make_unique<RegOperand>(*src1_reg));
moveInstr->addOperand(std::make_unique<RegOperand>(PhysicalReg::ZERO));
newInstrs.push_back(std::move(moveInstr));
}
else if (divisor == -1) {
auto negInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::SUBW : RVOpcodes::SUB);
negInstr->addOperand(std::make_unique<RegOperand>(*dst_reg));
negInstr->addOperand(std::make_unique<RegOperand>(PhysicalReg::ZERO));
negInstr->addOperand(std::make_unique<RegOperand>(*src1_reg));
newInstrs.push_back(std::move(negInstr));
}
else if (isPowerOfTwo(std::abs(divisor))) {
int shift = getPowerOfTwoExponent(std::abs(divisor));
int temp_reg = createTempReg();
auto sraSignInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::SRAIW : RVOpcodes::SRAI);
sraSignInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
sraSignInstr->addOperand(std::make_unique<RegOperand>(*src1_reg));
sraSignInstr->addOperand(std::make_unique<ImmOperand>(is_32bit ? 31 : 63));
newInstrs.push_back(std::move(sraSignInstr));
auto srlInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::SRLIW : RVOpcodes::SRLI);
srlInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
srlInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
srlInstr->addOperand(std::make_unique<ImmOperand>((is_32bit ? 32 : 64) - shift));
newInstrs.push_back(std::move(srlInstr));
auto addInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::ADDW : RVOpcodes::ADD);
addInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
addInstr->addOperand(std::make_unique<RegOperand>(*src1_reg));
addInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
newInstrs.push_back(std::move(addInstr));
auto sraInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::SRAIW : RVOpcodes::SRAI);
sraInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
sraInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
sraInstr->addOperand(std::make_unique<ImmOperand>(shift));
newInstrs.push_back(std::move(sraInstr));
if (divisor < 0) {
auto negInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::SUBW : RVOpcodes::SUB);
negInstr->addOperand(std::make_unique<RegOperand>(*dst_reg));
negInstr->addOperand(std::make_unique<RegOperand>(PhysicalReg::ZERO));
negInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
newInstrs.push_back(std::move(negInstr));
} else {
auto moveInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::ADDW : RVOpcodes::ADD);
moveInstr->addOperand(std::make_unique<RegOperand>(*dst_reg));
moveInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
moveInstr->addOperand(std::make_unique<RegOperand>(PhysicalReg::ZERO));
newInstrs.push_back(std::move(moveInstr));
}
}
else {
auto magic_info = computeMagic(divisor, is_32bit);
int magic_reg = createTempReg();
int temp_reg = createTempReg();
auto loadInstr = std::make_unique<MachineInstr>(RVOpcodes::LI);
loadInstr->addOperand(std::make_unique<RegOperand>(magic_reg));
loadInstr->addOperand(std::make_unique<ImmOperand>(magic_info.magic));
newInstrs.push_back(std::move(loadInstr));
if (is_32bit) {
auto mulInstr = std::make_unique<MachineInstr>(RVOpcodes::MUL);
mulInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
mulInstr->addOperand(std::make_unique<RegOperand>(*src1_reg));
mulInstr->addOperand(std::make_unique<RegOperand>(magic_reg));
newInstrs.push_back(std::move(mulInstr));
auto sraInstr = std::make_unique<MachineInstr>(RVOpcodes::SRAI);
sraInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
sraInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
sraInstr->addOperand(std::make_unique<ImmOperand>(magic_info.shift));
newInstrs.push_back(std::move(sraInstr));
} else {
auto mulhInstr = std::make_unique<MachineInstr>(RVOpcodes::MULH);
mulhInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
mulhInstr->addOperand(std::make_unique<RegOperand>(*src1_reg));
mulhInstr->addOperand(std::make_unique<RegOperand>(magic_reg));
newInstrs.push_back(std::move(mulhInstr));
int post_shift = magic_info.shift - 63;
if (post_shift > 0) {
auto sraInstr = std::make_unique<MachineInstr>(RVOpcodes::SRAI);
sraInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
sraInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
sraInstr->addOperand(std::make_unique<ImmOperand>(post_shift));
newInstrs.push_back(std::move(sraInstr));
}
}
int sign_reg = createTempReg();
auto sraSignInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::SRAIW : RVOpcodes::SRAI);
sraSignInstr->addOperand(std::make_unique<RegOperand>(sign_reg));
sraSignInstr->addOperand(std::make_unique<RegOperand>(*src1_reg));
sraSignInstr->addOperand(std::make_unique<ImmOperand>(is_32bit ? 31 : 63));
newInstrs.push_back(std::move(sraSignInstr));
auto subInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::SUBW : RVOpcodes::SUB);
subInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
subInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
subInstr->addOperand(std::make_unique<RegOperand>(sign_reg));
newInstrs.push_back(std::move(subInstr));
if (divisor < 0) {
auto negInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::SUBW : RVOpcodes::SUB);
negInstr->addOperand(std::make_unique<RegOperand>(*dst_reg));
negInstr->addOperand(std::make_unique<RegOperand>(PhysicalReg::ZERO));
negInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
newInstrs.push_back(std::move(negInstr));
} else {
auto moveInstr = std::make_unique<MachineInstr>(is_32bit ? RVOpcodes::ADDW : RVOpcodes::ADD);
moveInstr->addOperand(std::make_unique<RegOperand>(*dst_reg));
moveInstr->addOperand(std::make_unique<RegOperand>(temp_reg));
moveInstr->addOperand(std::make_unique<RegOperand>(PhysicalReg::ZERO));
newInstrs.push_back(std::move(moveInstr));
}
}
if (!newInstrs.empty()) {
size_t start_index = i;
if (instructions_to_replace == 2) {
start_index = i - 1;
}
replacements.push_back({start_index, instructions_to_replace, std::move(newInstrs)});
}
}
for (auto it = replacements.rbegin(); it != replacements.rend(); ++it) {
instrs.erase(instrs.begin() + it->index, instrs.begin() + it->index + it->count_to_erase);
instrs.insert(instrs.begin() + it->index,
std::make_move_iterator(it->newInstrs.begin()),
std::make_move_iterator(it->newInstrs.end()));
}
}
}
} // namespace sysy

View File

@ -1,652 +0,0 @@
#include "Peephole.h"
#include <functional>
namespace sysy {
char PeepholeOptimizer::ID = 0;
bool PeepholeOptimizer::runOnFunction(Function *F, AnalysisManager& AM) {
// This pass works on MachineFunction level, not IR level
return false;
}
void PeepholeOptimizer::runOnMachineFunction(MachineFunction *mfunc) {
if (!mfunc)
return;
using namespace sysy;
// areRegsEqual: 检查两个寄存器操作数是否相等(考虑虚拟和物理寄存器)。
auto areRegsEqual = [](RegOperand *r1, RegOperand *r2) {
if (!r1 || !r2 || r1->isVirtual() != r2->isVirtual()) {
return false;
}
if (r1->isVirtual()) {
return r1->getVRegNum() == r2->getVRegNum();
} else {
return r1->getPReg() == r2->getPReg();
}
};
// 改进的 isRegUsedLater 函数 - 更完整和准确的实现
auto isRegUsedLater =
[&](const std::vector<std::unique_ptr<MachineInstr>> &instrs,
RegOperand *reg, size_t start_idx) -> bool {
for (size_t j = start_idx; j < instrs.size(); ++j) {
auto *instr = instrs[j].get();
auto opcode = instr->getOpcode();
// 检查所有操作数
for (size_t k = 0; k < instr->getOperands().size(); ++k) {
bool isDefOperand = false;
// 更完整的定义操作数判断逻辑
if (k == 0) { // 第一个操作数通常是目标寄存器
switch (opcode) {
// 算术和逻辑指令 - 第一个操作数是定义
case RVOpcodes::MV:
case RVOpcodes::ADDI:
case RVOpcodes::SLLI:
case RVOpcodes::SRLI:
case RVOpcodes::SRAI:
case RVOpcodes::SLTI:
case RVOpcodes::SLTIU:
case RVOpcodes::XORI:
case RVOpcodes::ORI:
case RVOpcodes::ANDI:
case RVOpcodes::ADD:
case RVOpcodes::SUB:
case RVOpcodes::SLL:
case RVOpcodes::SLT:
case RVOpcodes::SLTU:
case RVOpcodes::XOR:
case RVOpcodes::SRL:
case RVOpcodes::SRA:
case RVOpcodes::OR:
case RVOpcodes::AND:
case RVOpcodes::MUL:
case RVOpcodes::DIV:
case RVOpcodes::REM:
case RVOpcodes::LW:
case RVOpcodes::LH:
case RVOpcodes::LB:
case RVOpcodes::LHU:
case RVOpcodes::LBU:
// 存储指令 - 第一个操作数是使用(要存储的值)
case RVOpcodes::SW:
case RVOpcodes::SH:
case RVOpcodes::SB:
// 分支指令 - 第一个操作数是使用
case RVOpcodes::BEQ:
case RVOpcodes::BNE:
case RVOpcodes::BLT:
case RVOpcodes::BGE:
case RVOpcodes::BLTU:
case RVOpcodes::BGEU:
// 跳转指令 - 可能使用寄存器
case RVOpcodes::JALR:
isDefOperand = false;
break;
default:
// 对于未知指令,保守地假设第一个操作数可能是使用
isDefOperand = false;
break;
}
}
// 如果不是定义操作数,检查是否使用了目标寄存器
if (!isDefOperand) {
if (instr->getOperands()[k]->getKind() == MachineOperand::KIND_REG) {
auto *use_reg =
static_cast<RegOperand *>(instr->getOperands()[k].get());
if (areRegsEqual(reg, use_reg))
return true;
}
// 检查内存操作数中的基址寄存器
if (instr->getOperands()[k]->getKind() == MachineOperand::KIND_MEM) {
auto *mem =
static_cast<MemOperand *>(instr->getOperands()[k].get());
if (areRegsEqual(reg, mem->getBase()))
return true;
}
}
}
}
return false;
};
// 检查寄存器是否在指令中被重新定义(用于更精确的分析)
auto isRegRedefinedAt =
[](MachineInstr *instr, RegOperand *reg,
const std::function<bool(RegOperand *, RegOperand *)> &areRegsEqual)
-> bool {
if (instr->getOperands().empty())
return false;
auto opcode = instr->getOpcode();
// 只有当第一个操作数是定义操作数时才检查
switch (opcode) {
case RVOpcodes::MV:
case RVOpcodes::ADDI:
case RVOpcodes::ADD:
case RVOpcodes::SUB:
case RVOpcodes::MUL:
case RVOpcodes::LW:
// ... 其他定义指令
if (instr->getOperands()[0]->getKind() == MachineOperand::KIND_REG) {
auto *def_reg =
static_cast<RegOperand *>(instr->getOperands()[0].get());
return areRegsEqual(reg, def_reg);
}
break;
default:
break;
}
return false;
};
// 检查是否为存储-加载模式,支持不同大小的访问
auto isStoreLoadPattern = [](MachineInstr *store_instr,
MachineInstr *load_instr) -> bool {
auto store_op = store_instr->getOpcode();
auto load_op = load_instr->getOpcode();
// 检查存储-加载对应关系
return (store_op == RVOpcodes::SW && load_op == RVOpcodes::LW) || // 32位
(store_op == RVOpcodes::SH &&
load_op == RVOpcodes::LH) || // 16位有符号
(store_op == RVOpcodes::SH &&
load_op == RVOpcodes::LHU) || // 16位无符号
(store_op == RVOpcodes::SB &&
load_op == RVOpcodes::LB) || // 8位有符号
(store_op == RVOpcodes::SB &&
load_op == RVOpcodes::LBU) || // 8位无符号
(store_op == RVOpcodes::SD && load_op == RVOpcodes::LD); // 64位
};
// 检查两个内存访问是否访问相同的内存位置
auto areMemoryAccessesEqual =
[&areRegsEqual](MachineInstr *store_instr, MemOperand *store_mem,
MachineInstr *load_instr, MemOperand *load_mem) -> bool {
// 基址寄存器必须相同
if (!areRegsEqual(store_mem->getBase(), load_mem->getBase())) {
return false;
}
// 偏移量必须相同
if (store_mem->getOffset()->getValue() !=
load_mem->getOffset()->getValue()) {
return false;
}
// 检查访问大小是否兼容
auto store_op = store_instr->getOpcode();
auto load_op = load_instr->getOpcode();
// 获取访问大小(字节数)
auto getAccessSize = [](RVOpcodes opcode) -> int {
switch (opcode) {
case RVOpcodes::LB:
case RVOpcodes::LBU:
case RVOpcodes::SB:
return 1; // 8位
case RVOpcodes::LH:
case RVOpcodes::LHU:
case RVOpcodes::SH:
return 2; // 16位
case RVOpcodes::LW:
case RVOpcodes::SW:
return 4; // 32位
case RVOpcodes::LD:
case RVOpcodes::SD:
return 8; // 64位
default:
return -1; // 未知
}
};
int store_size = getAccessSize(store_op);
int load_size = getAccessSize(load_op);
// 只有访问大小完全匹配时才能进行优化
// 这避免了部分重叠访问的复杂情况
return store_size > 0 && store_size == load_size;
};
// 简单的内存别名分析:检查两个内存访问之间是否可能有冲突的内存操作
auto isMemoryAccessSafe =
[&](const std::vector<std::unique_ptr<MachineInstr>> &instrs,
size_t store_idx, size_t load_idx, MemOperand *mem) -> bool {
// 检查存储和加载之间是否有可能影响内存的指令
for (size_t j = store_idx + 1; j < load_idx; ++j) {
auto *between_instr = instrs[j].get();
auto between_op = between_instr->getOpcode();
// 检查是否有其他内存写入操作
switch (between_op) {
case RVOpcodes::SW:
case RVOpcodes::SH:
case RVOpcodes::SB:
case RVOpcodes::SD: {
// 如果有其他存储操作,需要检查是否可能访问相同的内存
if (between_instr->getOperands().size() >= 2 &&
between_instr->getOperands()[1]->getKind() ==
MachineOperand::KIND_MEM) {
auto *other_mem =
static_cast<MemOperand *>(between_instr->getOperands()[1].get());
// 保守的别名分析:如果使用不同的基址寄存器,假设可能别名
if (!areRegsEqual(mem->getBase(), other_mem->getBase())) {
return false; // 可能的别名,不安全
}
// 如果基址相同但偏移量不同,检查是否重叠
int64_t offset1 = mem->getOffset()->getValue();
int64_t offset2 = other_mem->getOffset()->getValue();
// 获取访问大小来检查重叠
auto getAccessSize = [](RVOpcodes opcode) -> int {
switch (opcode) {
case RVOpcodes::SB:
return 1;
case RVOpcodes::SH:
return 2;
case RVOpcodes::SW:
return 4;
case RVOpcodes::SD:
return 8;
default:
return 4; // 默认假设4字节
}
};
int size1 = getAccessSize(RVOpcodes::SW); // 从原存储指令推断
int size2 = getAccessSize(between_op);
// 检查内存区域是否重叠
bool overlaps =
!(offset1 + size1 <= offset2 || offset2 + size2 <= offset1);
if (overlaps) {
return false; // 内存重叠,不安全
}
}
break;
}
// 函数调用可能有副作用
case RVOpcodes::JAL:
case RVOpcodes::JALR:
return false; // 函数调用可能修改内存,不安全
// 原子操作或其他可能修改内存的指令
// 根据具体的RISC-V扩展添加更多指令
default:
// 对于未知指令,采用保守策略
// 可以根据具体需求调整
break;
}
}
return true; // 没有发现潜在的内存冲突
};
// isPowerOfTwo: 检查数值是否为2的幂次并返回其指数。
auto isPowerOfTwo = [](int64_t n) -> int {
if (n <= 0 || (n & (n - 1)) != 0)
return -1;
int shift = 0;
while (n > 1) {
n >>= 1;
shift++;
}
return shift;
};
for (auto &mbb_uptr : mfunc->getBlocks()) {
auto &mbb = *mbb_uptr;
auto &instrs = mbb.getInstructions();
if (instrs.size() < 2)
continue; // 基本块至少需要两条指令进行窥孔
// 遍历指令序列进行窥孔优化
for (size_t i = 0; i + 1 < instrs.size();) {
auto *mi1 = instrs[i].get();
auto *mi2 = instrs[i + 1].get();
bool changed = false;
// 1. 消除冗余交换移动: mv a, b; mv b, a -> mv a, b
if (mi1->getOpcode() == RVOpcodes::MV &&
mi2->getOpcode() == RVOpcodes::MV) {
if (mi1->getOperands().size() == 2 && mi2->getOperands().size() == 2) {
if (mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_REG &&
mi2->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi2->getOperands()[1]->getKind() == MachineOperand::KIND_REG) {
auto *dst1 = static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *src1 = static_cast<RegOperand *>(mi1->getOperands()[1].get());
auto *dst2 = static_cast<RegOperand *>(mi2->getOperands()[0].get());
auto *src2 = static_cast<RegOperand *>(mi2->getOperands()[1].get());
if (areRegsEqual(dst1, src2) && areRegsEqual(src1, dst2)) {
instrs.erase(instrs.begin() + i + 1); // 移除第二条指令
changed = true;
}
}
}
}
// 2. 冗余加载消除: sw t0, offset(base); lw t1, offset(base) -> 替换或消除
// lw 添加ld sd支持
else if (isStoreLoadPattern(mi1, mi2)) {
if (mi1->getOperands().size() == 2 && mi2->getOperands().size() == 2) {
if (mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_MEM &&
mi2->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi2->getOperands()[1]->getKind() == MachineOperand::KIND_MEM) {
auto *store_val =
static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *store_mem =
static_cast<MemOperand *>(mi1->getOperands()[1].get());
auto *load_val =
static_cast<RegOperand *>(mi2->getOperands()[0].get());
auto *load_mem =
static_cast<MemOperand *>(mi2->getOperands()[1].get());
// 检查内存访问是否匹配(基址、偏移量和访问大小)
if (areMemoryAccessesEqual(mi1, store_mem, mi2, load_mem)) {
// 进行简单的内存别名分析
if (isMemoryAccessSafe(instrs, i, i + 1, store_mem)) {
if (areRegsEqual(store_val, load_val)) {
// sw r1, mem; lw r1, mem -> 消除冗余的lw
instrs.erase(instrs.begin() + i + 1);
changed = true;
} else {
// sw r1, mem; lw r2, mem -> 替换lw为mv r2, r1
auto newInstr = std::make_unique<MachineInstr>(RVOpcodes::MV);
newInstr->addOperand(std::make_unique<RegOperand>(*load_val));
newInstr->addOperand(
std::make_unique<RegOperand>(*store_val));
instrs[i + 1] = std::move(newInstr);
changed = true;
}
}
}
}
}
}
// 3. 强度削减: mul y, x, 2^n -> slli y, x, n
else if (mi1->getOpcode() == RVOpcodes::MUL &&
mi1->getOperands().size() == 3) {
auto *dst_op = mi1->getOperands()[0].get();
auto *src1_op = mi1->getOperands()[1].get();
auto *src2_op = mi1->getOperands()[2].get();
if (dst_op->getKind() == MachineOperand::KIND_REG) {
auto *dst_reg = static_cast<RegOperand *>(dst_op);
RegOperand *src_reg = nullptr;
int shift = -1;
if (src1_op->getKind() == MachineOperand::KIND_REG &&
src2_op->getKind() == MachineOperand::KIND_IMM) {
shift =
isPowerOfTwo(static_cast<ImmOperand *>(src2_op)->getValue());
if (shift >= 0)
src_reg = static_cast<RegOperand *>(src1_op);
} else if (src1_op->getKind() == MachineOperand::KIND_IMM &&
src2_op->getKind() == MachineOperand::KIND_REG) {
shift =
isPowerOfTwo(static_cast<ImmOperand *>(src1_op)->getValue());
if (shift >= 0)
src_reg = static_cast<RegOperand *>(src2_op);
}
if (src_reg && shift >= 0 &&
shift <= 31) { // RISC-V 移位量限制 (0-31)
auto newInstr = std::make_unique<MachineInstr>(RVOpcodes::SLLI);
newInstr->addOperand(std::make_unique<RegOperand>(*dst_reg));
newInstr->addOperand(std::make_unique<RegOperand>(*src_reg));
newInstr->addOperand(std::make_unique<ImmOperand>(shift));
instrs[i] = std::move(newInstr);
changed = true;
}
}
}
// 4. 地址计算优化: addi dst, base, imm1; lw/sw val, imm2(dst) -> lw/sw
// val, (imm1+imm2)(base)
else if (mi1->getOpcode() == RVOpcodes::ADDI &&
mi1->getOperands().size() == 3) {
auto opcode2 = mi2->getOpcode();
if (opcode2 == RVOpcodes::LW || opcode2 == RVOpcodes::SW) {
if (mi2->getOperands().size() == 2 &&
mi2->getOperands()[1]->getKind() == MachineOperand::KIND_MEM &&
mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[2]->getKind() == MachineOperand::KIND_IMM) {
auto *addi_dst =
static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *addi_base =
static_cast<RegOperand *>(mi1->getOperands()[1].get());
auto *addi_imm =
static_cast<ImmOperand *>(mi1->getOperands()[2].get());
auto *mem_op =
static_cast<MemOperand *>(mi2->getOperands()[1].get());
auto *mem_base = mem_op->getBase();
auto *mem_imm = mem_op->getOffset();
// 检查 ADDI 的目标寄存器是否是内存操作的基址
if (areRegsEqual(addi_dst, mem_base)) {
// 改进的使用检查:考虑寄存器可能在后续被重新定义的情况
bool canOptimize = true;
// 检查从 i+2 开始的指令
for (size_t j = i + 2; j < instrs.size(); ++j) {
auto *later_instr = instrs[j].get();
// 如果寄存器被重新定义,那么它后面的使用就不相关了
if (isRegRedefinedAt(later_instr, addi_dst, areRegsEqual)) {
break; // 寄存器被重新定义,可以安全优化
}
// 如果寄存器被使用,则不能优化
if (isRegUsedLater(instrs, addi_dst, j)) {
canOptimize = false;
break;
}
}
if (canOptimize) {
int64_t new_offset = addi_imm->getValue() + mem_imm->getValue();
// 检查新偏移量是否符合 RISC-V 12位有符号立即数范围
if (new_offset >= -2048 && new_offset <= 2047) {
auto new_mem_op = std::make_unique<MemOperand>(
std::make_unique<RegOperand>(*addi_base),
std::make_unique<ImmOperand>(new_offset));
mi2->getOperands()[1] = std::move(new_mem_op);
instrs.erase(instrs.begin() + i);
changed = true;
}
}
}
}
}
}
// 5. 冗余移动指令消除: mv x, y; op z, x, ... -> op z, y, ... (如果 x
// 之后不再使用)
else if (mi1->getOpcode() == RVOpcodes::MV &&
mi1->getOperands().size() == 2) {
if (mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_REG) {
auto *mv_dst = static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *mv_src = static_cast<RegOperand *>(mi1->getOperands()[1].get());
// 检查第二条指令是否使用了 mv 的目标寄存器
std::vector<size_t> use_positions;
for (size_t k = 1; k < mi2->getOperands().size(); ++k) {
if (mi2->getOperands()[k]->getKind() == MachineOperand::KIND_REG) {
auto *use_reg =
static_cast<RegOperand *>(mi2->getOperands()[k].get());
if (areRegsEqual(mv_dst, use_reg)) {
use_positions.push_back(k);
}
}
// 也检查内存操作数中的基址寄存器
else if (mi2->getOperands()[k]->getKind() ==
MachineOperand::KIND_MEM) {
auto *mem =
static_cast<MemOperand *>(mi2->getOperands()[k].get());
if (areRegsEqual(mv_dst, mem->getBase())) {
// 对于内存操作数我们需要创建新的MemOperand
auto new_mem = std::make_unique<MemOperand>(
std::make_unique<RegOperand>(*mv_src),
std::make_unique<ImmOperand>(mem->getOffset()->getValue()));
mi2->getOperands()[k] = std::move(new_mem);
use_positions.push_back(k); // 标记已处理
}
}
}
if (!use_positions.empty()) {
// 改进的后续使用检查
bool canOptimize = true;
for (size_t j = i + 2; j < instrs.size(); ++j) {
auto *later_instr = instrs[j].get();
// 如果寄存器被重新定义,后续使用就不相关了
if (isRegRedefinedAt(later_instr, mv_dst, areRegsEqual)) {
break;
}
// 检查是否还有其他使用
if (isRegUsedLater(instrs, mv_dst, j)) {
canOptimize = false;
break;
}
}
if (canOptimize) {
// 替换所有寄存器使用(内存操作数已在上面处理)
for (size_t pos : use_positions) {
if (mi2->getOperands()[pos]->getKind() ==
MachineOperand::KIND_REG) {
mi2->getOperands()[pos] =
std::make_unique<RegOperand>(*mv_src);
}
}
instrs.erase(instrs.begin() + i);
changed = true;
}
}
}
}
// 6. 连续加法指令合并: addi t1, t0, imm1; addi t2, t1, imm2 -> addi t2,
// t0, (imm1+imm2)
else if (mi1->getOpcode() == RVOpcodes::ADDI &&
mi2->getOpcode() == RVOpcodes::ADDI) {
if (mi1->getOperands().size() == 3 && mi2->getOperands().size() == 3) {
if (mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[2]->getKind() == MachineOperand::KIND_IMM &&
mi2->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi2->getOperands()[1]->getKind() == MachineOperand::KIND_REG &&
mi2->getOperands()[2]->getKind() == MachineOperand::KIND_IMM) {
auto *addi1_dst =
static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *addi1_src =
static_cast<RegOperand *>(mi1->getOperands()[1].get());
auto *addi1_imm =
static_cast<ImmOperand *>(mi1->getOperands()[2].get());
auto *addi2_dst =
static_cast<RegOperand *>(mi2->getOperands()[0].get());
auto *addi2_src =
static_cast<RegOperand *>(mi2->getOperands()[1].get());
auto *addi2_imm =
static_cast<ImmOperand *>(mi2->getOperands()[2].get());
// 检查第一个ADDI的目标是否是第二个ADDI的源
if (areRegsEqual(addi1_dst, addi2_src)) {
// 改进的中间寄存器使用检查
bool canOptimize = true;
for (size_t j = i + 2; j < instrs.size(); ++j) {
auto *later_instr = instrs[j].get();
// 如果中间寄存器被重新定义,后续使用不相关
if (isRegRedefinedAt(later_instr, addi1_dst, areRegsEqual)) {
break;
}
// 检查是否有其他使用
if (isRegUsedLater(instrs, addi1_dst, j)) {
canOptimize = false;
break;
}
}
if (canOptimize) {
int64_t new_imm = addi1_imm->getValue() + addi2_imm->getValue();
// 检查新立即数范围
if (new_imm >= -2048 && new_imm <= 2047) {
auto newInstr =
std::make_unique<MachineInstr>(RVOpcodes::ADDI);
newInstr->addOperand(
std::make_unique<RegOperand>(*addi2_dst));
newInstr->addOperand(
std::make_unique<RegOperand>(*addi1_src));
newInstr->addOperand(std::make_unique<ImmOperand>(new_imm));
instrs[i + 1] = std::move(newInstr);
instrs.erase(instrs.begin() + i);
changed = true;
}
}
}
}
}
}
// 7. ADD with zero optimization: add r1, r2, zero -> mv r1, r2
else if (mi1->getOpcode() == RVOpcodes::ADD &&
mi1->getOperands().size() == 3) {
if (mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[2]->getKind() == MachineOperand::KIND_REG) {
auto *add_dst =
static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *add_src1 =
static_cast<RegOperand *>(mi1->getOperands()[1].get());
auto *add_src2 =
static_cast<RegOperand *>(mi1->getOperands()[2].get());
// 检查第二个源操作数是否为ZERO寄存器
if (!add_src2->isVirtual() &&
add_src2->getPReg() == PhysicalReg::ZERO) {
// 创建新的 MV 指令
auto newInstr = std::make_unique<MachineInstr>(RVOpcodes::MV);
newInstr->addOperand(std::make_unique<RegOperand>(*add_dst));
newInstr->addOperand(std::make_unique<RegOperand>(*add_src1));
instrs[i] = std::move(newInstr);
changed = true;
}
}
}
// 根据是否发生变化调整遍历索引
if (!changed) {
++i; // 没有优化,继续检查下一对指令
} else {
// 发生变化,适当回退以捕获新的优化机会。
// 这是一种安全的回退策略,可以触发连锁优化,且不会导致无限循环。
if (i > 0) {
--i;
}
}
}
}
}
} // namespace sysy

View File

@ -1,416 +0,0 @@
#include "PostRA_Scheduler.h"
#include <algorithm>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#define MAX_SCHEDULING_BLOCK_SIZE 10000 // 限制调度块大小,避免过大导致性能问题
namespace sysy {
char PostRA_Scheduler::ID = 0;
// 检查指令是否是加载指令 (LW, LD)
bool isLoadInstr(MachineInstr *instr) {
RVOpcodes opcode = instr->getOpcode();
return opcode == RVOpcodes::LW || opcode == RVOpcodes::LD ||
opcode == RVOpcodes::LH || opcode == RVOpcodes::LB ||
opcode == RVOpcodes::LHU || opcode == RVOpcodes::LBU ||
opcode == RVOpcodes::LWU;
}
// 检查指令是否是存储指令 (SW, SD)
bool isStoreInstr(MachineInstr *instr) {
RVOpcodes opcode = instr->getOpcode();
return opcode == RVOpcodes::SW || opcode == RVOpcodes::SD ||
opcode == RVOpcodes::SH || opcode == RVOpcodes::SB;
}
// 检查指令是否为控制流指令
bool isControlFlowInstr(MachineInstr *instr) {
RVOpcodes opcode = instr->getOpcode();
return opcode == RVOpcodes::RET || opcode == RVOpcodes::J ||
opcode == RVOpcodes::BEQ || opcode == RVOpcodes::BNE ||
opcode == RVOpcodes::BLT || opcode == RVOpcodes::BGE ||
opcode == RVOpcodes::BLTU || opcode == RVOpcodes::BGEU ||
opcode == RVOpcodes::CALL;
}
// 预计算指令信息的缓存
static std::unordered_map<MachineInstr *, InstrRegInfo> instr_info_cache;
// 获取指令定义的寄存器 - 优化版本
std::unordered_set<PhysicalReg> getDefinedRegisters(MachineInstr *instr) {
std::unordered_set<PhysicalReg> defined_regs;
RVOpcodes opcode = instr->getOpcode();
// 特殊处理CALL指令
if (opcode == RVOpcodes::CALL) {
// CALL指令可能定义返回值寄存器
if (!instr->getOperands().empty() &&
instr->getOperands().front()->getKind() == MachineOperand::KIND_REG) {
auto reg_op =
static_cast<RegOperand *>(instr->getOperands().front().get());
if (!reg_op->isVirtual()) {
defined_regs.insert(reg_op->getPReg());
}
}
return defined_regs;
}
// 存储指令不定义寄存器
if (isStoreInstr(instr)) {
return defined_regs;
}
// 分支指令不定义寄存器
if (opcode == RVOpcodes::BEQ || opcode == RVOpcodes::BNE ||
opcode == RVOpcodes::BLT || opcode == RVOpcodes::BGE ||
opcode == RVOpcodes::BLTU || opcode == RVOpcodes::BGEU ||
opcode == RVOpcodes::J || opcode == RVOpcodes::RET) {
return defined_regs;
}
// 对于其他指令,第一个寄存器操作数通常是定义的
if (!instr->getOperands().empty() &&
instr->getOperands().front()->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<RegOperand *>(instr->getOperands().front().get());
if (!reg_op->isVirtual()) {
defined_regs.insert(reg_op->getPReg());
}
}
return defined_regs;
}
// 获取指令使用的寄存器 - 优化版本
std::unordered_set<PhysicalReg> getUsedRegisters(MachineInstr *instr) {
std::unordered_set<PhysicalReg> used_regs;
RVOpcodes opcode = instr->getOpcode();
// 特殊处理CALL指令
if (opcode == RVOpcodes::CALL) {
bool first_reg_skipped = false;
for (const auto &op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_REG) {
if (!first_reg_skipped) {
first_reg_skipped = true;
continue; // 跳过返回值寄存器
}
auto reg_op = static_cast<RegOperand *>(op.get());
if (!reg_op->isVirtual()) {
used_regs.insert(reg_op->getPReg());
}
}
}
return used_regs;
}
// 对于存储指令,所有寄存器操作数都是使用的
if (isStoreInstr(instr)) {
for (const auto &op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<RegOperand *>(op.get());
if (!reg_op->isVirtual()) {
used_regs.insert(reg_op->getPReg());
}
} else if (op->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand *>(op.get());
if (!mem_op->getBase()->isVirtual()) {
used_regs.insert(mem_op->getBase()->getPReg());
}
}
}
return used_regs;
}
// 对于分支指令,所有寄存器操作数都是使用的
if (opcode == RVOpcodes::BEQ || opcode == RVOpcodes::BNE ||
opcode == RVOpcodes::BLT || opcode == RVOpcodes::BGE ||
opcode == RVOpcodes::BLTU || opcode == RVOpcodes::BGEU) {
for (const auto &op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<RegOperand *>(op.get());
if (!reg_op->isVirtual()) {
used_regs.insert(reg_op->getPReg());
}
}
}
return used_regs;
}
// 对于其他指令,除了第一个寄存器操作数(通常是定义),其余都是使用的
bool first_reg = true;
for (const auto &op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_REG) {
if (first_reg) {
first_reg = false;
continue; // 跳过第一个寄存器(定义)
}
auto reg_op = static_cast<RegOperand *>(op.get());
if (!reg_op->isVirtual()) {
used_regs.insert(reg_op->getPReg());
}
} else if (op->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand *>(op.get());
if (!mem_op->getBase()->isVirtual()) {
used_regs.insert(mem_op->getBase()->getPReg());
}
}
}
return used_regs;
}
// 获取内存访问的基址和偏移
MemoryAccess getMemoryAccess(MachineInstr *instr) {
if (!isLoadInstr(instr) && !isStoreInstr(instr)) {
return MemoryAccess();
}
// 查找内存操作数
for (const auto &op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand *>(op.get());
if (!mem_op->getBase()->isVirtual()) {
return MemoryAccess(mem_op->getBase()->getPReg(),
mem_op->getOffset()->getValue());
}
}
}
return MemoryAccess();
}
// 预计算指令信息
InstrRegInfo &getInstrInfo(MachineInstr *instr) {
auto it = instr_info_cache.find(instr);
if (it != instr_info_cache.end()) {
return it->second;
}
InstrRegInfo &info = instr_info_cache[instr];
info.defined_regs = getDefinedRegisters(instr);
info.used_regs = getUsedRegisters(instr);
info.is_load = isLoadInstr(instr);
info.is_store = isStoreInstr(instr);
info.is_control_flow = isControlFlowInstr(instr);
info.mem_access = getMemoryAccess(instr);
return info;
}
// 检查内存依赖 - 优化版本
bool hasMemoryDependency(const InstrRegInfo &info1, const InstrRegInfo &info2) {
// 如果都不是内存指令,没有内存依赖
if (!info1.is_load && !info1.is_store && !info2.is_load && !info2.is_store) {
return false;
}
const MemoryAccess &mem1 = info1.mem_access;
const MemoryAccess &mem2 = info2.mem_access;
if (!mem1.valid || !mem2.valid) {
// 如果无法确定内存访问模式,保守地认为存在依赖
return true;
}
// 如果访问相同的内存位置
if (mem1.base_reg == mem2.base_reg && mem1.offset == mem2.offset) {
// Store->Load: RAW依赖
// Load->Store: WAR依赖
// Store->Store: WAW依赖
return info1.is_store || info2.is_store;
}
// 不同内存位置通常没有依赖,但为了安全起见,
// 如果涉及store指令我们需要更保守
if (info1.is_store && info2.is_load) {
// 保守处理不同store和load之间可能有别名
return false; // 这里可以根据需要调整策略
}
return false;
}
// 检查两个指令之间是否存在依赖关系 - 优化版本
bool hasDependency(MachineInstr *instr1, MachineInstr *instr2) {
const InstrRegInfo &info1 = getInstrInfo(instr1);
const InstrRegInfo &info2 = getInstrInfo(instr2);
// 检查RAW依赖instr1定义的寄存器是否被instr2使用
for (const auto &reg : info1.defined_regs) {
if (info2.used_regs.find(reg) != info2.used_regs.end()) {
return true; // RAW依赖 - instr2读取instr1写入的值
}
}
// 检查WAR依赖instr1使用的寄存器是否被instr2定义
for (const auto &reg : info1.used_regs) {
if (info2.defined_regs.find(reg) != info2.defined_regs.end()) {
return true; // WAR依赖 - instr2覆盖instr1需要的值
}
}
// 检查WAW依赖两个指令定义相同寄存器
for (const auto &reg : info1.defined_regs) {
if (info2.defined_regs.find(reg) != info2.defined_regs.end()) {
return true; // WAW依赖 - 两条指令写入同一寄存器
}
}
// 检查内存依赖
if (hasMemoryDependency(info1, info2)) {
return true;
}
return false;
}
// 检查是否可以安全地将instr1和instr2交换位置 - 优化版本
bool canSwapInstructions(MachineInstr *instr1, MachineInstr *instr2) {
const InstrRegInfo &info1 = getInstrInfo(instr1);
const InstrRegInfo &info2 = getInstrInfo(instr2);
// 不能移动控制流指令
if (info1.is_control_flow || info2.is_control_flow) {
return false;
}
// 检查双向依赖关系
return !hasDependency(instr1, instr2) && !hasDependency(instr2, instr1);
}
// 新增:验证调度结果的正确性 - 优化版本
void validateSchedule(const std::vector<MachineInstr *> &instr_list) {
for (int i = 0; i < (int)instr_list.size(); i++) {
for (int j = i + 1; j < (int)instr_list.size(); j++) {
MachineInstr *earlier = instr_list[i];
MachineInstr *later = instr_list[j];
const InstrRegInfo &info_earlier = getInstrInfo(earlier);
const InstrRegInfo &info_later = getInstrInfo(later);
// 检查是否存在被违反的依赖关系
// 检查RAW依赖
for (const auto &reg : info_earlier.defined_regs) {
if (info_later.used_regs.find(reg) != info_later.used_regs.end()) {
// 这是正常的依赖关系earlier应该在later之前
continue;
}
}
// 检查内存依赖
if (hasMemoryDependency(info_earlier, info_later)) {
const MemoryAccess &mem1 = info_earlier.mem_access;
const MemoryAccess &mem2 = info_later.mem_access;
if (mem1.valid && mem2.valid && mem1.base_reg == mem2.base_reg &&
mem1.offset == mem2.offset) {
if (info_earlier.is_store && info_later.is_load) {
// Store->Load依赖顺序正确
continue;
}
}
}
}
}
}
// 在基本块内对指令进行调度优化 - 优化版本
void scheduleBlock(MachineBasicBlock *mbb) {
auto &instructions = mbb->getInstructions();
if (instructions.size() <= 1)
return;
if (instructions.size() > MAX_SCHEDULING_BLOCK_SIZE) {
return; // 跳过超大块,防止卡住
}
// 清理缓存,避免无效指针
instr_info_cache.clear();
std::vector<MachineInstr *> instr_list;
instr_list.reserve(instructions.size()); // 预分配容量
for (auto &instr : instructions) {
instr_list.push_back(instr.get());
}
// 预计算所有指令的信息
for (auto *instr : instr_list) {
getInstrInfo(instr);
}
// 使用更严格的调度策略,避免破坏依赖关系
bool changed = true;
int max_iterations = 10; // 限制迭代次数避免死循环
int iteration = 0;
while (changed && iteration < max_iterations) {
changed = false;
iteration++;
for (int i = 0; i < (int)instr_list.size() - 1; i++) {
MachineInstr *instr1 = instr_list[i];
MachineInstr *instr2 = instr_list[i + 1];
const InstrRegInfo &info1 = getInstrInfo(instr1);
const InstrRegInfo &info2 = getInstrInfo(instr2);
// 只进行非常保守的优化
bool should_swap = false;
// 策略1: 将load指令提前减少load-use延迟
if (info2.is_load && !info1.is_load && !info1.is_store) {
should_swap = canSwapInstructions(instr1, instr2);
}
// 策略2: 将非关键store指令延后为其他指令让路
else if (info1.is_store && !info2.is_load && !info2.is_store) {
should_swap = canSwapInstructions(instr1, instr2);
}
if (should_swap) {
std::swap(instr_list[i], instr_list[i + 1]);
changed = true;
// 调试输出
// std::cout << "Swapped instructions at positions " << i << " and " <<
// (i+1) << std::endl;
}
}
}
// 验证调度结果的正确性
validateSchedule(instr_list);
// 将调度后的指令顺序写回
std::unordered_map<MachineInstr *, std::unique_ptr<MachineInstr>> instr_map;
instr_map.reserve(instructions.size()); // 预分配容量
for (auto &instr : instructions) {
instr_map[instr.get()] = std::move(instr);
}
instructions.clear();
instructions.reserve(instr_list.size()); // 预分配容量
for (auto instr : instr_list) {
instructions.push_back(std::move(instr_map[instr]));
}
}
bool PostRA_Scheduler::runOnFunction(Function *F, AnalysisManager &AM) {
// 这个函数在IR级别运行但我们需要在机器指令级别运行
// 所以我们返回false表示没有对IR进行修改
return false;
}
void PostRA_Scheduler::runOnMachineFunction(MachineFunction *mfunc) {
// std::cout << "Running Post-RA Local Scheduler... " << std::endl;
// 遍历每个机器基本块
for (auto &mbb : mfunc->getBlocks()) {
scheduleBlock(mbb.get());
}
// 清理全局缓存
instr_info_cache.clear();
}
} // namespace sysy

View File

@ -1,466 +0,0 @@
#include "PreRA_Scheduler.h"
#include "RISCv64LLIR.h"
#include <algorithm>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#define MAX_SCHEDULING_BLOCK_SIZE 1000 // 严格限制调度块大小
namespace sysy {
char PreRA_Scheduler::ID = 0;
// 检查指令是否是加载指令 (LW, LD)
static bool isLoadInstr(MachineInstr *instr) {
RVOpcodes opcode = instr->getOpcode();
return opcode == RVOpcodes::LW || opcode == RVOpcodes::LD ||
opcode == RVOpcodes::LH || opcode == RVOpcodes::LB ||
opcode == RVOpcodes::LHU || opcode == RVOpcodes::LBU ||
opcode == RVOpcodes::LWU;
}
// 检查指令是否是存储指令 (SW, SD)
static bool isStoreInstr(MachineInstr *instr) {
RVOpcodes opcode = instr->getOpcode();
return opcode == RVOpcodes::SW || opcode == RVOpcodes::SD ||
opcode == RVOpcodes::SH || opcode == RVOpcodes::SB;
}
// 检查指令是否为分支指令
static bool isBranchInstr(MachineInstr *instr) {
RVOpcodes opcode = instr->getOpcode();
return opcode == RVOpcodes::BEQ || opcode == RVOpcodes::BNE ||
opcode == RVOpcodes::BLT || opcode == RVOpcodes::BGE ||
opcode == RVOpcodes::BLTU || opcode == RVOpcodes::BGEU;
}
// 检查指令是否为跳转指令
static bool isJumpInstr(MachineInstr *instr) {
RVOpcodes opcode = instr->getOpcode();
return opcode == RVOpcodes::J;
}
// 检查指令是否为返回指令
static bool isReturnInstr(MachineInstr *instr) {
return instr->getOpcode() == RVOpcodes::RET;
}
// 检查指令是否为调用指令
static bool isCallInstr(MachineInstr *instr) {
return instr->getOpcode() == RVOpcodes::CALL;
}
// 检查指令是否为块终结指令(必须保持在块尾)
static bool isTerminatorInstr(MachineInstr *instr) {
return isBranchInstr(instr) || isJumpInstr(instr) || isReturnInstr(instr);
}
// 检查指令是否有副作用(需要谨慎处理)
static bool hasSideEffect(MachineInstr *instr) {
return isStoreInstr(instr) || isCallInstr(instr) || isTerminatorInstr(instr);
}
// 检查指令是否涉及内存操作
static bool hasMemoryAccess(MachineInstr *instr) {
return isLoadInstr(instr) || isStoreInstr(instr);
}
// 获取内存访问位置信息
struct MemoryLocation {
unsigned base_reg;
int64_t offset;
bool is_valid;
MemoryLocation() : base_reg(0), offset(0), is_valid(false) {}
MemoryLocation(unsigned base, int64_t off)
: base_reg(base), offset(off), is_valid(true) {}
bool operator==(const MemoryLocation &other) const {
return is_valid && other.is_valid && base_reg == other.base_reg &&
offset == other.offset;
}
};
// 缓存指令分析信息
struct InstrInfo {
std::unordered_set<unsigned> defined_regs;
std::unordered_set<unsigned> used_regs;
MemoryLocation mem_location;
bool is_load;
bool is_store;
bool is_terminator;
bool is_call;
bool has_side_effect;
bool has_memory_access;
InstrInfo() : is_load(false), is_store(false), is_terminator(false),
is_call(false), has_side_effect(false), has_memory_access(false) {}
};
// 指令信息缓存
static std::unordered_map<MachineInstr*, InstrInfo> instr_info_cache;
// 获取指令定义的虚拟寄存器 - 优化版本
static std::unordered_set<unsigned> getDefinedVirtualRegisters(MachineInstr *instr) {
std::unordered_set<unsigned> defined_regs;
RVOpcodes opcode = instr->getOpcode();
// CALL指令可能定义返回值寄存器
if (opcode == RVOpcodes::CALL) {
if (!instr->getOperands().empty() &&
instr->getOperands().front()->getKind() == MachineOperand::KIND_REG) {
auto reg_op =
static_cast<RegOperand *>(instr->getOperands().front().get());
if (reg_op->isVirtual()) {
defined_regs.insert(reg_op->getVRegNum());
}
}
return defined_regs;
}
// 存储指令和终结指令不定义寄存器
if (isStoreInstr(instr) || isTerminatorInstr(instr)) {
return defined_regs;
}
// 其他指令的第一个操作数通常是目标寄存器
if (!instr->getOperands().empty() &&
instr->getOperands().front()->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<RegOperand *>(instr->getOperands().front().get());
if (reg_op->isVirtual()) {
defined_regs.insert(reg_op->getVRegNum());
}
}
return defined_regs;
}
// 获取指令使用的虚拟寄存器 - 优化版本
static std::unordered_set<unsigned> getUsedVirtualRegisters(MachineInstr *instr) {
std::unordered_set<unsigned> used_regs;
RVOpcodes opcode = instr->getOpcode();
// CALL指令跳过第一个操作数返回值其余为参数
if (opcode == RVOpcodes::CALL) {
bool first_reg_skipped = false;
for (const auto &op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_REG) {
if (!first_reg_skipped) {
first_reg_skipped = true;
continue;
}
auto reg_op = static_cast<RegOperand *>(op.get());
if (reg_op->isVirtual()) {
used_regs.insert(reg_op->getVRegNum());
}
}
}
return used_regs;
}
// 存储指令和终结指令:所有操作数都是使用的
if (isStoreInstr(instr) || isTerminatorInstr(instr)) {
for (const auto &op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<RegOperand *>(op.get());
if (reg_op->isVirtual()) {
used_regs.insert(reg_op->getVRegNum());
}
} else if (op->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand *>(op.get());
if (mem_op->getBase()->isVirtual()) {
used_regs.insert(mem_op->getBase()->getVRegNum());
}
}
}
return used_regs;
}
// 其他指令:跳过第一个操作数(目标寄存器),其余为源操作数
bool first_reg = true;
for (const auto &op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_REG) {
if (first_reg) {
first_reg = false;
continue;
}
auto reg_op = static_cast<RegOperand *>(op.get());
if (reg_op->isVirtual()) {
used_regs.insert(reg_op->getVRegNum());
}
} else if (op->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand *>(op.get());
if (mem_op->getBase()->isVirtual()) {
used_regs.insert(mem_op->getBase()->getVRegNum());
}
}
}
return used_regs;
}
// 获取内存访问位置
static MemoryLocation getMemoryLocation(MachineInstr *instr) {
if (!isLoadInstr(instr) && !isStoreInstr(instr)) {
return MemoryLocation();
}
for (const auto &op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand *>(op.get());
if (mem_op->getBase()->isVirtual()) {
return MemoryLocation(mem_op->getBase()->getVRegNum(),
mem_op->getOffset()->getValue());
}
}
}
return MemoryLocation();
}
// 预计算并缓存指令信息
static const InstrInfo& getInstrInfo(MachineInstr *instr) {
auto it = instr_info_cache.find(instr);
if (it != instr_info_cache.end()) {
return it->second;
}
InstrInfo& info = instr_info_cache[instr];
info.defined_regs = getDefinedVirtualRegisters(instr);
info.used_regs = getUsedVirtualRegisters(instr);
info.mem_location = getMemoryLocation(instr);
info.is_load = isLoadInstr(instr);
info.is_store = isStoreInstr(instr);
info.is_terminator = isTerminatorInstr(instr);
info.is_call = isCallInstr(instr);
info.has_side_effect = hasSideEffect(instr);
info.has_memory_access = hasMemoryAccess(instr);
return info;
}
// 检查两个内存位置是否可能别名
static bool mayAlias(const MemoryLocation &loc1, const MemoryLocation &loc2) {
if (!loc1.is_valid || !loc2.is_valid) {
return true; // 保守处理:未知位置可能别名
}
// 不同基址寄存器,保守假设可能别名
if (loc1.base_reg != loc2.base_reg) {
return true;
}
// 相同基址寄存器,检查偏移
return loc1.offset == loc2.offset;
}
// 检查两个指令之间是否存在数据依赖 - 优化版本
static bool hasDataDependency(MachineInstr *first, MachineInstr *second) {
const InstrInfo& info_first = getInstrInfo(first);
const InstrInfo& info_second = getInstrInfo(second);
// RAW依赖: second读取first写入的寄存器
for (const auto &reg : info_first.defined_regs) {
if (info_second.used_regs.find(reg) != info_second.used_regs.end()) {
return true;
}
}
// WAR依赖: second写入first读取的寄存器
for (const auto &reg : info_first.used_regs) {
if (info_second.defined_regs.find(reg) != info_second.defined_regs.end()) {
return true;
}
}
// WAW依赖: 两个指令写入同一寄存器
for (const auto &reg : info_first.defined_regs) {
if (info_second.defined_regs.find(reg) != info_second.defined_regs.end()) {
return true;
}
}
return false;
}
// 检查两个指令之间是否存在内存依赖 - 优化版本
static bool hasMemoryDependency(MachineInstr *first, MachineInstr *second) {
const InstrInfo& info_first = getInstrInfo(first);
const InstrInfo& info_second = getInstrInfo(second);
if (!info_first.has_memory_access || !info_second.has_memory_access) {
return false;
}
// 如果至少有一个是存储指令,需要检查别名
if (info_first.is_store || info_second.is_store) {
return mayAlias(info_first.mem_location, info_second.mem_location);
}
return false; // 两个加载指令之间没有依赖
}
// 检查两个指令之间是否存在控制依赖 - 优化版本
static bool hasControlDependency(MachineInstr *first, MachineInstr *second) {
const InstrInfo& info_first = getInstrInfo(first);
const InstrInfo& info_second = getInstrInfo(second);
// 终结指令与任何其他指令都有控制依赖
if (info_first.is_terminator) {
return true; // first是终结指令second不能移动到first之前
}
if (info_second.is_terminator) {
return false; // second是终结指令可以保持在后面
}
// CALL指令具有控制副作用但可以参与有限的调度
if (info_first.is_call || info_second.is_call) {
// CALL指令之间保持顺序
if (info_first.is_call && info_second.is_call) {
return true;
}
// 其他情况允许调度(通过数据依赖控制)
}
return false;
}
// 综合检查两个指令是否可以交换 - 优化版本
static bool canSwapInstructions(MachineInstr *first, MachineInstr *second) {
// 检查所有类型的依赖
if (hasDataDependency(first, second) || hasDataDependency(second, first)) {
return false;
}
if (hasMemoryDependency(first, second)) {
return false;
}
if (hasControlDependency(first, second) ||
hasControlDependency(second, first)) {
return false;
}
return true;
}
// 找到基本块中的调度边界 - 优化版本
static std::vector<size_t>
findSchedulingBoundaries(const std::vector<MachineInstr *> &instrs) {
std::vector<size_t> boundaries;
boundaries.reserve(instrs.size() / 10); // 预估边界数量
boundaries.push_back(0); // 起始边界
for (size_t i = 0; i < instrs.size(); i++) {
const InstrInfo& info = getInstrInfo(instrs[i]);
// 终结指令前后都是边界
if (info.is_terminator) {
if (i > 0)
boundaries.push_back(i);
if (i + 1 < instrs.size())
boundaries.push_back(i + 1);
}
// 跳转目标标签也可能是边界(这里简化处理)
}
boundaries.push_back(instrs.size()); // 结束边界
// 去重并排序
std::sort(boundaries.begin(), boundaries.end());
boundaries.erase(std::unique(boundaries.begin(), boundaries.end()),
boundaries.end());
return boundaries;
}
// 在单个调度区域内进行指令调度 - 优化版本
static void scheduleRegion(std::vector<MachineInstr *> &instrs, size_t start,
size_t end) {
if (end - start <= 1) {
return; // 区域太小,无需调度
}
// 保守的调度策略:
// 1. 只对小规模区域进行调度
// 2. 优先将加载指令向前调度,以隐藏内存延迟
// 3. 确保不破坏数据依赖和内存依赖
// 简单的调度算法:只尝试将加载指令尽可能前移
for (size_t i = start + 1; i < end; i++) {
const InstrInfo& info = getInstrInfo(instrs[i]);
if (info.is_load) {
// 尝试将加载指令向前移动
for (size_t j = i; j > start; j--) {
// 检查是否可以与前一条指令交换
if (canSwapInstructions(instrs[j - 1], instrs[j])) {
std::swap(instrs[j - 1], instrs[j]);
} else {
// 一旦遇到依赖关系就停止移动
break;
}
}
}
}
}
static void scheduleBlock(MachineBasicBlock *mbb) {
auto &instructions = mbb->getInstructions();
if (instructions.size() <= 1 ||
instructions.size() > MAX_SCHEDULING_BLOCK_SIZE) {
return;
}
// 清理缓存,避免无效指针
instr_info_cache.clear();
// 构建指令列表
std::vector<MachineInstr *> instr_list;
instr_list.reserve(instructions.size()); // 预分配容量
for (auto &instr : instructions) {
instr_list.push_back(instr.get());
}
// 预计算所有指令信息
for (auto* instr : instr_list) {
getInstrInfo(instr);
}
// 找到调度边界
std::vector<size_t> boundaries = findSchedulingBoundaries(instr_list);
// 在每个调度区域内进行局部调度
for (size_t i = 0; i < boundaries.size() - 1; i++) {
size_t region_start = boundaries[i];
size_t region_end = boundaries[i + 1];
scheduleRegion(instr_list, region_start, region_end);
}
// 重建指令序列
std::unordered_map<MachineInstr *, std::unique_ptr<MachineInstr>> instr_map;
instr_map.reserve(instructions.size()); // 预分配容量
for (auto &instr : instructions) {
instr_map[instr.get()] = std::move(instr);
}
instructions.clear();
instructions.reserve(instr_list.size()); // 预分配容量
for (auto *instr : instr_list) {
instructions.push_back(std::move(instr_map[instr]));
}
}
bool PreRA_Scheduler::runOnFunction(Function *F, AnalysisManager &AM) {
return false;
}
void PreRA_Scheduler::runOnMachineFunction(MachineFunction *mfunc) {
for (auto &mbb : mfunc->getBlocks()) {
scheduleBlock(mbb.get());
}
// 清理全局缓存
instr_info_cache.clear();
}
} // namespace sysy

View File

@ -1,279 +0,0 @@
#include "RISCv64Backend.h"
#include "RISCv64ISel.h"
#include "RISCv64RegAlloc.h"
#include "RISCv64LinearScan.h" // <--- 新增此行
#include "RISCv64AsmPrinter.h"
#include "RISCv64Passes.h"
#include <sstream>
#include <future> // <--- 新增此行
#include <chrono> // <--- 新增此行
#include <iostream> // <--- 新增此行,用于打印超时警告
namespace sysy {
// 顶层入口
std::string RISCv64CodeGen::code_gen() {
return module_gen();
}
unsigned RISCv64CodeGen::getTypeSizeInBytes(Type* type) {
if (!type) {
assert(false && "Cannot get size of a null type.");
return 0;
}
switch (type->getKind()) {
// 对于SysY语言基本类型int和float都占用4字节
case Type::kInt:
case Type::kFloat:
return 4;
// 指针类型在RISC-V 64位架构下占用8字节
// 虽然SysY没有'int*'语法但数组变量在IR层面本身就是指针类型
case Type::kPointer:
return 8;
// 数组类型的总大小 = 元素数量 * 单个元素的大小
case Type::kArray: {
auto arrayType = type->as<ArrayType>();
// 递归调用以计算元素大小
return arrayType->getNumElements() * getTypeSizeInBytes(arrayType->getElementType());
}
// 其他类型如Void, Label等不占用栈空间或者不应该出现在这里
default:
// 如果遇到未处理的类型,触发断言,方便调试
// assert(false && "Unsupported type for size calculation.");
return 0; // 对于像Label或Void这样的类型返回0是合理的
}
}
void printInitializer(std::stringstream& ss, const ValueCounter& init_values) {
for (size_t i = 0; i < init_values.getValues().size(); ++i) {
auto val = init_values.getValues()[i];
auto count = init_values.getNumbers()[i];
if (auto constant = dynamic_cast<ConstantValue*>(val)) {
for (unsigned j = 0; j < count; ++j) {
if (constant->isInt()) {
ss << " .word " << constant->getInt() << "\n";
} else {
float f = constant->getFloat();
uint32_t float_bits = *(uint32_t*)&f;
ss << " .word " << float_bits << "\n";
}
}
}
}
}
std::string RISCv64CodeGen::module_gen() {
std::stringstream ss;
// --- 步骤1将全局变量(GlobalValue)分为.data和.bss两组 ---
std::vector<GlobalValue*> data_globals;
std::vector<GlobalValue*> bss_globals;
for (const auto& global_ptr : module->getGlobals()) {
GlobalValue* global = global_ptr.get();
// 使用更健壮的逻辑来判断是否为大型零初始化数组
bool is_all_zeros = true;
const auto& init_values = global->getInitValues();
// 检查初始化值是否全部为0
if (init_values.getValues().empty()) {
// 如果 ValueCounter 为空GlobalValue 的构造函数会确保它是零初始化的
is_all_zeros = true;
} else {
for (auto val : init_values.getValues()) {
if (auto const_val = dynamic_cast<ConstantValue*>(val)) {
if (!const_val->isZero()) {
is_all_zeros = false;
break;
}
} else {
// 如果初始值包含非常量(例如,另一个全局变量的地址),则不认为是纯零初始化
is_all_zeros = false;
break;
}
}
}
// 使用 getTypeSizeInBytes 检查总大小是否超过阈值 (16个整数 = 64字节)
Type* allocated_type = global->getType()->as<PointerType>()->getBaseType();
unsigned total_size = getTypeSizeInBytes(allocated_type);
bool is_large_zero_array = is_all_zeros && (total_size > 64);
if (is_large_zero_array) {
bss_globals.push_back(global);
} else {
data_globals.push_back(global);
}
}
// --- 步骤2生成 .bss 段的代码 ---
if (!bss_globals.empty()) {
ss << ".bss\n";
for (GlobalValue* global : bss_globals) {
Type* allocated_type = global->getType()->as<PointerType>()->getBaseType();
unsigned total_size = getTypeSizeInBytes(allocated_type);
ss << " .align 3\n";
ss << ".globl " << global->getName() << "\n";
ss << ".type " << global->getName() << ", @object\n";
ss << ".size " << global->getName() << ", " << total_size << "\n";
ss << global->getName() << ":\n";
ss << " .space " << total_size << "\n";
}
}
// --- 步骤3生成 .data 段的代码 ---
if (!data_globals.empty() || !module->getConsts().empty()) {
ss << ".data\n";
// a. 处理普通的全局变量 (GlobalValue)
for (GlobalValue* global : data_globals) {
Type* allocated_type = global->getType()->as<PointerType>()->getBaseType();
unsigned total_size = getTypeSizeInBytes(allocated_type);
ss << " .align 3\n";
ss << ".globl " << global->getName() << "\n";
ss << ".type " << global->getName() << ", @object\n";
ss << ".size " << global->getName() << ", " << total_size << "\n";
ss << global->getName() << ":\n";
bool is_all_zeros = true;
const auto& init_values = global->getInitValues();
if (init_values.getValues().empty()) {
is_all_zeros = true;
} else {
for (auto val : init_values.getValues()) {
if (auto const_val = dynamic_cast<ConstantValue*>(val)) {
if (!const_val->isZero()) {
is_all_zeros = false;
break;
}
} else {
is_all_zeros = false;
break;
}
}
}
if (is_all_zeros) {
ss << " .zero " << total_size << "\n";
} else {
// 对于有非零初始值的变量,保持原有的打印逻辑。
printInitializer(ss, global->getInitValues());
}
}
// b. 处理全局常量 (ConstantVariable)
for (const auto& const_ptr : module->getConsts()) {
ConstantVariable* cnst = const_ptr.get();
Type* allocated_type = cnst->getType()->as<PointerType>()->getBaseType();
unsigned total_size = getTypeSizeInBytes(allocated_type);
ss << " .align 3\n";
ss << ".globl " << cnst->getName() << "\n";
ss << ".type " << cnst->getName() << ", @object\n";
ss << ".size " << cnst->getName() << ", " << total_size << "\n";
ss << cnst->getName() << ":\n";
printInitializer(ss, cnst->getInitValues());
}
}
// --- 步骤4处理函数 (.text段) 的逻辑 ---
if (!module->getFunctions().empty()) {
ss << ".text\n";
for (const auto& func_pair : module->getFunctions()) {
if (func_pair.second.get() && !func_pair.second->getBasicBlocks().empty()) {
ss << function_gen(func_pair.second.get());
if (DEBUG) std::cerr << "Function: " << func_pair.first << " generated.\n";
}
}
}
return ss.str();
}
std::string RISCv64CodeGen::function_gen(Function* func) {
// === 完整的后端处理流水线 ===
// 阶段 1: 指令选择 (sysy::IR -> LLIR with virtual registers)
RISCv64ISel isel;
std::unique_ptr<MachineFunction> mfunc = isel.runOnFunction(func);
// 第一次调试打印输出
std::stringstream ss_after_isel;
RISCv64AsmPrinter printer_isel(mfunc.get());
printer_isel.run(ss_after_isel, true);
if (DEBUG) {
std::cerr << "====== Intermediate Representation after Instruction Selection ======\n"
<< ss_after_isel.str();
}
// 阶段 2: 消除帧索引 (展开伪指令,计算局部变量偏移)
// 这个Pass必须在寄存器分配之前运行
EliminateFrameIndicesPass efi_pass;
efi_pass.runOnMachineFunction(mfunc.get());
if (DEBUG) {
std::cerr << "====== stack info after eliminate frame indices ======\n";
mfunc->dumpStackFrameInfo(std::cerr);
std::stringstream ss_after_eli;
printer_isel.run(ss_after_eli, true);
std::cerr << "====== LLIR after eliminate frame indices ======\n"
<< ss_after_eli.str();
}
// // 阶段 2: 除法强度削弱优化 (Division Strength Reduction)
// DivStrengthReduction div_strength_reduction;
// div_strength_reduction.runOnMachineFunction(mfunc.get());
// // 阶段 2.1: 指令调度 (Instruction Scheduling)
// PreRA_Scheduler scheduler;
// scheduler.runOnMachineFunction(mfunc.get());
// 阶段 3: 物理寄存器分配 (Register Allocation)
RISCv64RegAlloc reg_alloc(mfunc.get());
reg_alloc.run();
if (DEBUG) {
std::cerr << "====== stack info after reg alloc ======\n";
mfunc->dumpStackFrameInfo(std::cerr);
}
// 阶段 3.1: 处理被调用者保存寄存器
CalleeSavedHandler callee_handler;
callee_handler.runOnMachineFunction(mfunc.get());
if (DEBUG) {
std::cerr << "====== stack info after callee handler ======\n";
mfunc->dumpStackFrameInfo(std::cerr);
}
// // 阶段 4: 窥孔优化 (Peephole Optimization)
// PeepholeOptimizer peephole;
// peephole.runOnMachineFunction(mfunc.get());
// 阶段 5: 局部指令调度 (Local Scheduling)
PostRA_Scheduler local_scheduler;
local_scheduler.runOnMachineFunction(mfunc.get());
// 阶段 3.2: 插入序言和尾声
PrologueEpilogueInsertionPass pei_pass;
pei_pass.runOnMachineFunction(mfunc.get());
// 阶段 3.3: 大立即数合法化
LegalizeImmediatesPass legalizer;
legalizer.runOnMachineFunction(mfunc.get());
// 阶段 6: 代码发射 (Code Emission)
std::stringstream ss;
RISCv64AsmPrinter printer(mfunc.get());
printer.run(ss);
return ss.str();
}
} // namespace sysy

File diff suppressed because it is too large Load Diff

View File

@ -1,122 +0,0 @@
#include "RISCv64LLIR.h"
#include <vector>
#include <iostream> // 用于 std::ostream 和 std::cerr
#include <string> // 用于 std::string
namespace sysy {
// 辅助函数:将 PhysicalReg 枚举转换为可读的字符串
std::string regToString(PhysicalReg reg) {
switch (reg) {
case PhysicalReg::ZERO: return "x0"; case PhysicalReg::RA: return "ra";
case PhysicalReg::SP: return "sp"; case PhysicalReg::GP: return "gp";
case PhysicalReg::TP: return "tp"; case PhysicalReg::T0: return "t0";
case PhysicalReg::T1: return "t1"; case PhysicalReg::T2: return "t2";
case PhysicalReg::S0: return "s0"; case PhysicalReg::S1: return "s1";
case PhysicalReg::A0: return "a0"; case PhysicalReg::A1: return "a1";
case PhysicalReg::A2: return "a2"; case PhysicalReg::A3: return "a3";
case PhysicalReg::A4: return "a4"; case PhysicalReg::A5: return "a5";
case PhysicalReg::A6: return "a6"; case PhysicalReg::A7: return "a7";
case PhysicalReg::S2: return "s2"; case PhysicalReg::S3: return "s3";
case PhysicalReg::S4: return "s4"; case PhysicalReg::S5: return "s5";
case PhysicalReg::S6: return "s6"; case PhysicalReg::S7: return "s7";
case PhysicalReg::S8: return "s8"; case PhysicalReg::S9: return "s9";
case PhysicalReg::S10: return "s10"; case PhysicalReg::S11: return "s11";
case PhysicalReg::T3: return "t3"; case PhysicalReg::T4: return "t4";
case PhysicalReg::T5: return "t5"; case PhysicalReg::T6: return "t6";
case PhysicalReg::F0: return "f0"; case PhysicalReg::F1: return "f1";
case PhysicalReg::F2: return "f2"; case PhysicalReg::F3: return "f3";
case PhysicalReg::F4: return "f4"; case PhysicalReg::F5: return "f5";
case PhysicalReg::F6: return "f6"; case PhysicalReg::F7: return "f7";
case PhysicalReg::F8: return "f8"; case PhysicalReg::F9: return "f9";
case PhysicalReg::F10: return "f10"; case PhysicalReg::F11: return "f11";
case PhysicalReg::F12: return "f12"; case PhysicalReg::F13: return "f13";
case PhysicalReg::F14: return "f14"; case PhysicalReg::F15: return "f15";
case PhysicalReg::F16: return "f16"; case PhysicalReg::F17: return "f17";
case PhysicalReg::F18: return "f18"; case PhysicalReg::F19: return "f19";
case PhysicalReg::F20: return "f20"; case PhysicalReg::F21: return "f21";
case PhysicalReg::F22: return "f22"; case PhysicalReg::F23: return "f23";
case PhysicalReg::F24: return "f24"; case PhysicalReg::F25: return "f25";
case PhysicalReg::F26: return "f26"; case PhysicalReg::F27: return "f27";
case PhysicalReg::F28: return "f28"; case PhysicalReg::F29: return "f29";
case PhysicalReg::F30: return "f30"; case PhysicalReg::F31: return "f31";
default: return "UNKNOWN_REG";
}
}
// 打印栈帧信息的完整实现
void MachineFunction::dumpStackFrameInfo(std::ostream& os) const {
const StackFrameInfo& info = frame_info;
os << "--- Stack Frame Info for function '" << getName() << "' ---\n";
// 打印尺寸信息
os << " Sizes:\n";
os << " Total Size: " << info.total_size << " bytes\n";
os << " Locals Size: " << info.locals_size << " bytes\n";
os << " Spill Size: " << info.spill_size << " bytes\n";
os << " Callee-Saved Size: " << info.callee_saved_size << " bytes\n";
os << "\n";
// 打印 Alloca 变量的偏移量
os << " Alloca Offsets (vreg -> offset from FP):\n";
if (info.alloca_offsets.empty()) {
os << " (None)\n";
} else {
for (const auto& pair : info.alloca_offsets) {
os << " %vreg" << pair.first << " -> " << pair.second << "\n";
}
}
os << "\n";
// 打印溢出变量的偏移量
os << " Spill Offsets (vreg -> offset from FP):\n";
if (info.spill_offsets.empty()) {
os << " (None)\n";
} else {
for (const auto& pair : info.spill_offsets) {
os << " %vreg" << pair.first << " -> " << pair.second << "\n";
}
}
os << "\n";
// 打印使用的被调用者保存寄存器
os << " Used Callee-Saved Registers:\n";
if (info.used_callee_saved_regs.empty()) {
os << " (None)\n";
} else {
os << " { ";
for (const auto& reg : info.used_callee_saved_regs) {
os << regToString(reg) << " ";
}
os << "}\n";
}
os << "\n";
// 打印需要保存/恢复的被调用者保存寄存器 (有序)
os << " Callee-Saved Registers to Store/Restore:\n";
if (info.callee_saved_regs_to_store.empty()) {
os << " (None)\n";
} else {
os << " [ ";
for (const auto& reg : info.callee_saved_regs_to_store) {
os << regToString(reg) << " ";
}
os << "]\n";
}
os << "\n";
// 打印最终的寄存器分配结果
os << " Final Register Allocation Map (vreg -> preg):\n";
if (info.vreg_to_preg_map.empty()) {
os << " (None)\n";
} else {
for (const auto& pair : info.vreg_to_preg_map) {
os << " %vreg" << pair.first << " -> " << regToString(pair.second) << "\n";
}
}
os << "---------------------------------------------------\n";
}
}

View File

@ -1,517 +0,0 @@
#include "RISCv64LinearScan.h"
#include "RISCv64LLIR.h"
#include "RISCv64ISel.h"
#include <iostream>
#include <set>
extern int DEBUG;
namespace sysy {
RISCv64LinearScan::RISCv64LinearScan(MachineFunction* mfunc)
: MFunc(mfunc),
ISel(mfunc->getISel()),
vreg_type_map(ISel->getVRegTypeMap()) {
// 初始化可用的物理寄存器池,与图着色版本保持一致
// 整数寄存器
allocable_int_regs = {
PhysicalReg::T0, PhysicalReg::T1, PhysicalReg::T2, PhysicalReg::T3, PhysicalReg::T4, /*T5保留作为大立即数加载寄存器*/ PhysicalReg::T6,
PhysicalReg::A0, PhysicalReg::A1, PhysicalReg::A2, PhysicalReg::A3, PhysicalReg::A4, PhysicalReg::A5, PhysicalReg::A6, PhysicalReg::A7,
PhysicalReg::S1, PhysicalReg::S2, PhysicalReg::S3, PhysicalReg::S4, PhysicalReg::S5, PhysicalReg::S6, PhysicalReg::S7,
PhysicalReg::S8, PhysicalReg::S9, PhysicalReg::S10, PhysicalReg::S11,
};
// 浮点寄存器
allocable_fp_regs = {
PhysicalReg::F0, PhysicalReg::F1, PhysicalReg::F2, PhysicalReg::F3, PhysicalReg::F4, PhysicalReg::F5, PhysicalReg::F6, PhysicalReg::F7,
PhysicalReg::F10, PhysicalReg::F11, PhysicalReg::F12, PhysicalReg::F13, PhysicalReg::F14, PhysicalReg::F15, PhysicalReg::F16, PhysicalReg::F17,
PhysicalReg::F8, PhysicalReg::F9, PhysicalReg::F18, PhysicalReg::F19, PhysicalReg::F20, PhysicalReg::F21, PhysicalReg::F22,
PhysicalReg::F23, PhysicalReg::F24, PhysicalReg::F25, PhysicalReg::F26, PhysicalReg::F27,
PhysicalReg::F28, PhysicalReg::F29, PhysicalReg::F30, PhysicalReg::F31,
};
// 新增识别所有通过寄存器传递的参数并建立vreg到物理寄存器(preg)的映射
// 这等同于图着色算法中的“预着色”步骤。
if (MFunc->getFunc()) {
int int_arg_idx = 0;
int fp_arg_idx = 0;
for (Argument* arg : MFunc->getFunc()->getArguments()) {
unsigned arg_vreg = ISel->getVReg(arg);
if (arg->getType()->isFloat()) {
if (fp_arg_idx < 8) { // fa0-fa7
auto preg = static_cast<PhysicalReg>(static_cast<int>(PhysicalReg::F10) + fp_arg_idx);
abi_vreg_map[arg_vreg] = preg;
fp_arg_idx++;
}
} else { // 整数或指针
if (int_arg_idx < 8) { // a0-a7
auto preg = static_cast<PhysicalReg>(static_cast<int>(PhysicalReg::A0) + int_arg_idx);
abi_vreg_map[arg_vreg] = preg;
int_arg_idx++;
}
}
}
}
}
void RISCv64LinearScan::run() {
if (DEBUG) std::cerr << "===== Running Linear Scan Register Allocation for function: " << MFunc->getName() << " =====\n";
bool changed = true;
while(changed) {
// 1. 准备阶段
linearizeBlocks();
computeLiveIntervals();
// 2. 执行线性扫描
changed = linearScan();
// 3. 如果有溢出,重写代码,然后下一轮重新开始
if (changed) {
rewriteProgram();
if (DEBUG) std::cerr << "--- Spilling detected, re-running linear scan ---\n";
}
}
// 4. 将最终分配结果应用到机器指令
applyAllocation();
// 5. 收集用到的被调用者保存寄存器
MFunc->getFrameInfo().vreg_to_preg_map = this->vreg_to_preg_map;
collectUsedCalleeSavedRegs();
if (DEBUG) std::cerr << "===== Finished Linear Scan Register Allocation =====\n\n";
}
// 步骤 1.1: 对基本块进行线性化,这里我们简单地按现有顺序排列
void RISCv64LinearScan::linearizeBlocks() {
linear_order_blocks.clear();
for (auto& mbb : MFunc->getBlocks()) {
linear_order_blocks.push_back(mbb.get());
}
}
// RISCv64LinearScan.cpp
void RISCv64LinearScan::computeLiveIntervals() {
instr_numbering.clear();
live_intervals.clear();
unhandled.clear();
// a. 对所有指令进行线性编号并记录CALL指令的位置
int num = 0;
std::set<int> call_locations;
for (auto* mbb : linear_order_blocks) {
for (auto& instr : mbb->getInstructions()) {
instr_numbering[instr.get()] = num;
if (instr->getOpcode() == RVOpcodes::CALL) {
call_locations.insert(num);
}
num += 2; // 指令编号间隔为2方便在溢出重写时插入指令
}
}
// b. 遍历所有指令记录每个vreg首次和末次出现的位置
std::map<unsigned, std::pair<int, int>> vreg_ranges; // vreg -> {first_instr_num, last_instr_num}
for (auto* mbb : linear_order_blocks) {
for (auto& instr_ptr : mbb->getInstructions()) {
const MachineInstr* instr = instr_ptr.get();
int instr_num = instr_numbering.at(instr);
std::set<unsigned> use, def;
getInstrUseDef(instr, use, def);
auto all_vregs = use;
all_vregs.insert(def.begin(), def.end());
for (unsigned vreg : all_vregs) {
if (vreg_ranges.find(vreg) == vreg_ranges.end()) {
vreg_ranges[vreg] = {instr_num, instr_num};
} else {
vreg_ranges[vreg].second = std::max(vreg_ranges[vreg].second, instr_num);
}
}
}
}
// c. 根据记录的边界创建LiveInterval对象并检查是否跨越CALL
for (auto const& [vreg, range] : vreg_ranges) {
live_intervals.emplace(vreg, LiveInterval(vreg));
auto& interval = live_intervals.at(vreg);
interval.start = range.first;
interval.end = range.second;
// 检查此区间是否跨越了任何CALL指令
auto it = call_locations.lower_bound(interval.start);
if (it != call_locations.end() && *it < interval.end) {
interval.crosses_call = true;
}
}
// d. 将所有计算出的活跃区间放入 unhandled 列表
for (auto& pair : live_intervals) {
unhandled.push_back(&pair.second);
}
std::sort(unhandled.begin(), unhandled.end(), [](const LiveInterval* a, const LiveInterval* b){
return a->start < b->start;
});
}
// RISCv64LinearScan.cpp
// 在类的定义中添加一个辅助函数来判断寄存器类型
bool isCalleeSaved(PhysicalReg preg) {
if (preg >= PhysicalReg::S1 && preg <= PhysicalReg::S11) return true;
if (preg == PhysicalReg::S0) return true; // s0 通常也作为被调用者保存
// 浮点寄存器
if (preg >= PhysicalReg::F8 && preg <= PhysicalReg::F9) return true;
if (preg >= PhysicalReg::F18 && preg <= PhysicalReg::F27) return true;
return false;
}
// 线性扫描主算法
bool RISCv64LinearScan::linearScan() {
active.clear();
spilled_vregs.clear();
vreg_to_preg_map.clear();
// 将寄存器池分为调用者保存和被调用者保存两类
std::set<PhysicalReg> free_caller_int_regs, free_callee_int_regs;
std::set<PhysicalReg> free_caller_fp_regs, free_callee_fp_regs;
for (auto preg : allocable_int_regs) {
if (isCalleeSaved(preg)) free_callee_int_regs.insert(preg);
else free_caller_int_regs.insert(preg);
}
for (auto preg : allocable_fp_regs) {
if (isCalleeSaved(preg)) free_callee_fp_regs.insert(preg);
else free_caller_fp_regs.insert(preg);
}
// 预处理ABI参数寄存器
vreg_to_preg_map.insert(abi_vreg_map.begin(), abi_vreg_map.end());
std::vector<LiveInterval*> normal_unhandled;
for(LiveInterval* interval : unhandled) {
if(abi_vreg_map.count(interval->vreg)) {
active.push_back(interval);
PhysicalReg preg = abi_vreg_map.at(interval->vreg);
if (isFPVReg(interval->vreg)) {
if(isCalleeSaved(preg)) free_callee_fp_regs.erase(preg); else free_caller_fp_regs.erase(preg);
} else {
if(isCalleeSaved(preg)) free_callee_int_regs.erase(preg); else free_caller_int_regs.erase(preg);
}
} else {
normal_unhandled.push_back(interval);
}
}
unhandled = normal_unhandled;
std::sort(active.begin(), active.end(), [](const LiveInterval* a, const LiveInterval* b){ return a->end < b->end; });
// 主循环
for (LiveInterval* current : unhandled) {
// a. 释放active列表中已结束的区间
std::vector<LiveInterval*> new_active;
for (LiveInterval* active_interval : active) {
if (active_interval->end < current->start) {
PhysicalReg preg = vreg_to_preg_map.at(active_interval->vreg);
if (isFPVReg(active_interval->vreg)) {
if(isCalleeSaved(preg)) free_callee_fp_regs.insert(preg); else free_caller_fp_regs.insert(preg);
} else {
if(isCalleeSaved(preg)) free_callee_int_regs.insert(preg); else free_caller_int_regs.insert(preg);
}
} else {
new_active.push_back(active_interval);
}
}
active = new_active;
// b. 约束化地为当前区间分配寄存器
bool is_fp = isFPVReg(current->vreg);
auto& free_caller = is_fp ? free_caller_fp_regs : free_caller_int_regs;
auto& free_callee = is_fp ? free_callee_fp_regs : free_callee_int_regs;
PhysicalReg allocated_preg = PhysicalReg::INVALID;
if (current->crosses_call) {
// 跨调用区间:必须使用被调用者保存寄存器
if (!free_callee.empty()) {
allocated_preg = *free_callee.begin();
free_callee.erase(allocated_preg);
}
} else {
// 非跨调用区间:优先使用调用者保存寄存器
if (!free_caller.empty()) {
allocated_preg = *free_caller.begin();
free_caller.erase(allocated_preg);
} else if (!free_callee.empty()) {
allocated_preg = *free_callee.begin();
free_callee.erase(allocated_preg);
}
}
if (allocated_preg != PhysicalReg::INVALID) {
vreg_to_preg_map[current->vreg] = allocated_preg;
active.push_back(current);
std::sort(active.begin(), active.end(), [](const LiveInterval* a, const LiveInterval* b){ return a->end < b->end; });
} else {
// c. 没有可用寄存器,需要溢出
spillAtInterval(current);
}
}
return !spilled_vregs.empty();
}
void RISCv64LinearScan::chooseRegForInterval(LiveInterval* current) {
bool is_fp = isFPVReg(current->vreg);
auto& free_regs = is_fp ? free_fp_regs : free_int_regs;
if (!free_regs.empty()) {
// 有可用寄存器
PhysicalReg preg = *free_regs.begin();
free_regs.erase(free_regs.begin());
vreg_to_preg_map[current->vreg] = preg;
active.push_back(current);
// 保持 active 列表按结束点排序
std::sort(active.begin(), active.end(), [](const LiveInterval* a, const LiveInterval* b){
return a->end < b->end;
});
} else {
// 没有可用寄存器,需要溢出
spillAtInterval(current);
}
}
void RISCv64LinearScan::spillAtInterval(LiveInterval* current) {
LiveInterval* spill_candidate = nullptr;
// 启发式溢出:
// 如果current需要callee-saved则从active中找一个占用callee-saved且结束最晚的区间比较
// 否则找active中结束最晚的区间
// 这里简化处理总是找active中结束最晚的区间
auto last_active = active.back();
if (last_active->end > current->end) {
// 溢出active中的区间
spill_candidate = last_active;
PhysicalReg preg = vreg_to_preg_map.at(spill_candidate->vreg);
vreg_to_preg_map[current->vreg] = preg; // 把换出的寄存器给current
// 更新active列表
active.pop_back();
active.push_back(current);
std::sort(active.begin(), active.end(), [](const LiveInterval* a, const LiveInterval* b){ return a->end < b->end; });
spilled_vregs.insert(spill_candidate->vreg);
} else {
// 溢出当前区间
spilled_vregs.insert(current->vreg);
}
}
// 步骤 3: 重写程序,插入溢出代码
void RISCv64LinearScan::rewriteProgram() {
StackFrameInfo& frame_info = MFunc->getFrameInfo();
int spill_offset = frame_info.locals_size; // 溢出区域接在局部变量之后
for (unsigned vreg : spilled_vregs) {
if (frame_info.spill_offsets.count(vreg)) continue; // 避免重复分配
int size = isFPVReg(vreg) ? 4 : (vreg_type_map.at(vreg)->isPointer() ? 8 : 4);
spill_offset += size;
spill_offset = (spill_offset + 7) & ~7; // 8字节对齐
frame_info.spill_offsets[vreg] = -(16 + spill_offset);
}
frame_info.spill_size = spill_offset - frame_info.locals_size;
for (auto& mbb : MFunc->getBlocks()) {
auto& instrs = mbb->getInstructions();
std::vector<std::unique_ptr<MachineInstr>> new_instrs;
for (auto it = instrs.begin(); it != instrs.end(); ++it) {
auto& instr = *it;
std::set<unsigned> use_vregs, def_vregs;
getInstrUseDef(instr.get(), use_vregs, def_vregs);
// 建立溢出vreg到新临时vreg的映射
std::map<unsigned, unsigned> use_remap;
std::map<unsigned, unsigned> def_remap;
// 1. 为所有溢出的USE创建LOAD指令和映射
for (unsigned old_vreg : use_vregs) {
if (spilled_vregs.count(old_vreg) && use_remap.find(old_vreg) == use_remap.end()) {
Type* type = vreg_type_map.at(old_vreg);
unsigned new_temp_vreg = ISel->getNewVReg(type);
use_remap[old_vreg] = new_temp_vreg;
RVOpcodes load_op = isFPVReg(old_vreg) ? RVOpcodes::FLW : (type->isPointer() ? RVOpcodes::LD : RVOpcodes::LW);
auto load = std::make_unique<MachineInstr>(load_op);
load->addOperand(std::make_unique<RegOperand>(new_temp_vreg));
load->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0),
std::make_unique<ImmOperand>(frame_info.spill_offsets.at(old_vreg))
));
new_instrs.push_back(std::move(load));
}
}
// 2. 为所有溢出的DEF创建映射
for (unsigned old_vreg : def_vregs) {
if (spilled_vregs.count(old_vreg) && def_remap.find(old_vreg) == def_remap.end()) {
Type* type = vreg_type_map.at(old_vreg);
unsigned new_temp_vreg = ISel->getNewVReg(type);
def_remap[old_vreg] = new_temp_vreg;
}
}
// 3. 基于角色精确地替换原指令中的操作数
auto opcode = instr->getOpcode();
auto& operands = instr->getOperands();
auto replace_reg_op = [](RegOperand* reg_op, const std::map<unsigned, unsigned>& remap) {
if (reg_op->isVirtual() && remap.count(reg_op->getVRegNum())) {
reg_op->setVRegNum(remap.at(reg_op->getVRegNum()));
}
};
if (op_info.count(opcode)) {
const auto& info = op_info.at(opcode);
// 替换 Defs
for (int idx : info.first) {
if (idx < operands.size() && operands[idx]->getKind() == MachineOperand::KIND_REG) {
replace_reg_op(static_cast<RegOperand*>(operands[idx].get()), def_remap);
}
}
// 替换 Uses
for (int idx : info.second) {
if (idx < operands.size()) {
if (operands[idx]->getKind() == MachineOperand::KIND_REG) {
replace_reg_op(static_cast<RegOperand*>(operands[idx].get()), use_remap);
} else if (operands[idx]->getKind() == MachineOperand::KIND_MEM) {
replace_reg_op(static_cast<MemOperand*>(operands[idx].get())->getBase(), use_remap);
}
}
}
} else if (opcode == RVOpcodes::CALL) {
// 特殊处理 CALL 指令
if (!operands.empty() && operands[0]->getKind() == MachineOperand::KIND_REG) {
replace_reg_op(static_cast<RegOperand*>(operands[0].get()), def_remap);
}
for (size_t i = 1; i < operands.size(); ++i) {
if (operands[i]->getKind() == MachineOperand::KIND_REG) {
replace_reg_op(static_cast<RegOperand*>(operands[i].get()), use_remap);
}
}
}
// 4. 将修改后的指令放入新列表
new_instrs.push_back(std::move(instr));
// 5. 为所有溢出的DEF创建STORE指令
for(const auto& pair : def_remap) {
unsigned old_vreg = pair.first;
unsigned new_temp_vreg = pair.second;
Type* type = vreg_type_map.at(old_vreg);
RVOpcodes store_op = isFPVReg(old_vreg) ? RVOpcodes::FSW : (type->isPointer() ? RVOpcodes::SD : RVOpcodes::SW);
auto store = std::make_unique<MachineInstr>(store_op);
store->addOperand(std::make_unique<RegOperand>(new_temp_vreg));
store->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0),
std::make_unique<ImmOperand>(frame_info.spill_offsets.at(old_vreg))
));
new_instrs.push_back(std::move(store));
}
}
instrs = std::move(new_instrs);
}
}
// 步骤 4: 应用最终分配结果
void RISCv64LinearScan::applyAllocation() {
for (auto& mbb : MFunc->getBlocks()) {
for (auto& instr_ptr : mbb->getInstructions()) {
for (auto& op_ptr : instr_ptr->getOperands()) {
if (op_ptr->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<RegOperand*>(op_ptr.get());
if (reg_op->isVirtual()) {
unsigned vreg = reg_op->getVRegNum();
if (vreg_to_preg_map.count(vreg)) {
reg_op->setPReg(vreg_to_preg_map.at(vreg));
} else {
// 如果一个vreg最终没有颜色这通常意味着它是一个短生命周期的临时变量
// 在溢出重写中产生,但在下一轮分配前就被优化掉了。
// 或者是一个从未被使用的定义。
// 给他一个临时寄存器以防万一。
reg_op->setPReg(PhysicalReg::T5);
}
}
} else if (op_ptr->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand*>(op_ptr.get());
auto reg_op = mem_op->getBase();
if (reg_op->isVirtual()) {
unsigned vreg = reg_op->getVRegNum();
if (vreg_to_preg_map.count(vreg)) {
reg_op->setPReg(vreg_to_preg_map.at(vreg));
} else {
reg_op->setPReg(PhysicalReg::T5);
}
}
}
}
}
}
}
void RISCv64LinearScan::getInstrUseDef(const MachineInstr* instr, std::set<unsigned>& use, std::set<unsigned>& def) {
// 这个函数与图着色版本中的 getInstrUseDef 逻辑完全相同,此处直接复用
auto opcode = instr->getOpcode();
const auto& operands = instr->getOperands();
// op_info 的定义已被移到函数外部的命名空间中
auto get_vreg_id_if_virtual = [&](const MachineOperand* op, std::set<unsigned>& s) {
if (op->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<const RegOperand*>(op);
if (reg_op->isVirtual()) s.insert(reg_op->getVRegNum());
} else if (op->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<const MemOperand*>(op);
auto reg_op = mem_op->getBase();
if (reg_op->isVirtual()) s.insert(reg_op->getVRegNum());
}
};
if (op_info.count(opcode)) {
const auto& info = op_info.at(opcode);
for (int idx : info.first) if (idx < operands.size()) get_vreg_id_if_virtual(operands[idx].get(), def);
for (int idx : info.second) if (idx < operands.size()) get_vreg_id_if_virtual(operands[idx].get(), use);
// MemOperand 的基址寄存器总是一个 use
for (const auto& op : operands) if (op->getKind() == MachineOperand::KIND_MEM) get_vreg_id_if_virtual(op.get(), use);
} else if (opcode == RVOpcodes::CALL) {
// CALL指令的特殊处理
// 第一个操作数如果有是def返回值
if (!operands.empty() && operands[0]->getKind() == MachineOperand::KIND_REG) get_vreg_id_if_virtual(operands[0].get(), def);
// 后续的寄存器操作数是use参数
for (size_t i = 1; i < operands.size(); ++i) if (operands[i]->getKind() == MachineOperand::KIND_REG) get_vreg_id_if_virtual(operands[i].get(), use);
}
}
// 辅助函数: 判断是否为浮点vreg
bool RISCv64LinearScan::isFPVReg(unsigned vreg) const {
return vreg_type_map.count(vreg) && vreg_type_map.at(vreg)->isFloat();
}
// 辅助函数: 收集被使用的被调用者保存寄存器
void RISCv64LinearScan::collectUsedCalleeSavedRegs() {
StackFrameInfo& frame_info = MFunc->getFrameInfo();
frame_info.used_callee_saved_regs.clear();
const auto& callee_saved_int = getCalleeSavedIntRegs();
const auto& callee_saved_fp = getCalleeSavedFpRegs();
std::set<PhysicalReg> callee_saved_set(callee_saved_int.begin(), callee_saved_int.end());
callee_saved_set.insert(callee_saved_fp.begin(), callee_saved_fp.end());
callee_saved_set.insert(PhysicalReg::S0); // s0总是被用作帧指针
for(const auto& pair : vreg_to_preg_map) {
PhysicalReg preg = pair.second;
if(callee_saved_set.count(preg)) {
frame_info.used_callee_saved_regs.insert(preg);
}
}
}
} // namespace sysy

File diff suppressed because it is too large Load Diff

View File

@ -1,17 +0,0 @@
# src/frontend/CMakeLists.txt
add_library(frontend_lib STATIC
SysYBaseVisitor.cpp
SysY.g4
SysYLexer.cpp
SysYParser.cpp
SysYVisitor.cpp
)
# 包含前端模块所需的头文件路径
target_include_directories(frontend_lib PUBLIC
${CMAKE_CURRENT_SOURCE_DIR}/../include/frontend # 前端头文件
${ANTLR_RUNTIME}/runtime/src # ANTLR 运行时头文件
)
# 链接 ANTLR 运行时库
target_link_libraries(frontend_lib PRIVATE antlr4_shared)

View File

@ -0,0 +1,59 @@
#pragma once
#include "IR.h" // 假设IR.h包含了Module, Function, BasicBlock, Instruction, Value, IRBuilder, Type等定义
#include "IRBuilder.h" // 需要IRBuilder来创建新指令
#include "SysYIRPrinter.h" // 新增: 用于调试输出
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>
#include <list> // 用于迭代和修改指令列表
#include <algorithm> // for std::reverse (if needed, although not used in final version)
#include <iostream> // MODIFICATION: 用于警告输出
namespace sysy {
/**
* @brief AddressCalculationExpansion Pass
*
* 这是一个IR优化Pass用于将LoadInst和StoreInst中包含的多维数组索引
* 显式地转换为IR中的BinaryInst乘法和加法序列并生成带有线性偏移量的
* LoadInst/StoreInst。
*
* 目的确保在寄存器分配之前所有中间地址计算的结果都有明确的IR指令和对应的虚拟寄存器
* 从而避免在后端DAG构建时临时创建值而导致寄存器分配缺失的问题。
*
* SysY语言特性
* - 无指针类型所有数组访问的基地址是alloca或global的AllocaType/ArrayType
* - 数据类型只有int和float且都占用4字节。
* - LoadInst和StoreInst直接接受多个索引作为额外操作数。
*/
class AddressCalculationExpansion {
private:
Module* pModule;
IRBuilder* pBuilder; // 用于在IR中插入新指令
// 数组元素的固定大小根据SysY特性int和float都是4字节
static const int ELEMENT_SIZE = 4;
// 辅助函数:根据数组的维度信息和当前索引的维度,计算该索引的步长(字节数)
// dims: 包含所有维度大小的vector例如 {2, 3, 4}
// currentDimIndex: 当前正在处理的索引在 dims 中的位置 (0, 1, 2...)
int calculateStride(const std::vector<int>& dims, size_t currentDimIndex) {
int stride = ELEMENT_SIZE; // 最内层元素大小 (4字节)
// 乘以当前维度之后的所有维度的大小
for (size_t i = currentDimIndex + 1; i < dims.size(); ++i) {
stride *= dims[i];
}
return stride;
}
public:
AddressCalculationExpansion(Module* module, IRBuilder* builder)
: pModule(module), pBuilder(builder) {}
// 运行此Pass
bool run();
};
} // namespace sysy

View File

@ -0,0 +1,39 @@
#pragma once
#include "IR.h"
#include "SysYIRAnalyser.h"
#include "SysYIRPrinter.h"
namespace sysy {
class DeadCodeElimination {
private:
Module *pModule;
ControlFlowAnalysis *pCFA; // 控制流分析指针
ActiveVarAnalysis *pAVA; // 活跃变量分析指针
DataFlowAnalysisUtils dataFlowAnalysisUtils; // 数据流分析工具类
public:
explicit DeadCodeElimination(Module *pMoudle,
ControlFlowAnalysis *pCFA = nullptr,
ActiveVarAnalysis *pAVA = nullptr)
: pModule(pMoudle), pCFA(pCFA), pAVA(pAVA), dataFlowAnalysisUtils() {} // 构造函数
// TODO根据参数传入的passes来运行不同的死代码删除流程
// void runDCEPipeline(const std::vector<std::string>& passes = {
// "dead-store", "redundant-load-store", "dead-load", "dead-alloca", "dead-global"
// });
void runDCEPipeline(); // 运行死代码删除
void eliminateDeadStores(Function* func, bool& changed); // 消除无用存储
void eliminateDeadLoads(Function* func, bool& changed); // 消除无用加载
void eliminateDeadAllocas(Function* func, bool& changed); // 消除无用内存分配
void eliminateDeadGlobals(bool& changed); // 消除无用全局变量
void eliminateDeadIndirectiveAllocas(Function* func, bool& changed); // 消除无用间接内存分配(phi节点)
void eliminateDeadRedundantLoadStore(Function* func, bool& changed); // 消除冗余加载和存储
bool isGlobal(Value *val);
bool isArr(Value *val);
void usedelete(Instruction *instr);
};
} // namespace sysy

File diff suppressed because it is too large Load Diff

View File

@ -126,7 +126,7 @@ class IRBuilder {
UnaryInst * createFNotInst(Value *operand, const std::string &name = "") {
return createUnaryInst(Instruction::kFNot, Type::getIntType(), operand, name);
} ///< 创建浮点取非指令
UnaryInst * createItoFInst(Value *operand, const std::string &name = "") {
UnaryInst * createIToFInst(Value *operand, const std::string &name = "") {
return createUnaryInst(Instruction::kItoF, Type::getFloatType(), operand, name);
} ///< 创建整型转浮点指令
UnaryInst * createBitItoFInst(Value *operand, const std::string &name = "") {
@ -217,12 +217,6 @@ class IRBuilder {
BinaryInst * createOrInst(Value *lhs, Value *rhs, const std::string &name = "") {
return createBinaryInst(Instruction::kOr, Type::getIntType(), lhs, rhs, name);
} ///< 创建按位或指令
BinaryInst * createSRAInst(Value *lhs, Value *rhs, const std::string &name = "") {
return createBinaryInst(Instruction::kSRA, Type::getIntType(), lhs, rhs, name);
} ///< 创建算术右移指令
BinaryInst * createMulhInst(Value *lhs, Value *rhs, const std::string &name = "") {
return createBinaryInst(Instruction::kMulh, Type::getIntType(), lhs, rhs, name);
} ///< 创建高位乘法指令
CallInst * createCallInst(Function *callee, const std::vector<Value *> &args, const std::string &name = "") {
std::string newName;
if (name.empty() && callee->getReturnType() != Type::getVoidType()) {
@ -245,30 +239,31 @@ class IRBuilder {
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建return指令
UncondBrInst * createUncondBrInst(BasicBlock *thenBlock) {
auto inst = new UncondBrInst(thenBlock, block);
UncondBrInst * createUncondBrInst(BasicBlock *thenBlock, const std::vector<Value *> &args) {
auto inst = new UncondBrInst(thenBlock, args, block);
assert(inst);
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建无条件指令
CondBrInst * createCondBrInst(Value *condition, BasicBlock *thenBlock, BasicBlock *elseBlock) {
auto inst = new CondBrInst(condition, thenBlock, elseBlock, block);
CondBrInst * createCondBrInst(Value *condition, BasicBlock *thenBlock, BasicBlock *elseBlock,
const std::vector<Value *> &thenArgs, const std::vector<Value *> &elseArgs) {
auto inst = new CondBrInst(condition, thenBlock, elseBlock, thenArgs, elseArgs, block);
assert(inst);
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建条件跳转指令
UnreachableInst * createUnreachableInst(const std::string &name = "") {
auto inst = new UnreachableInst(name, block);
assert(inst);
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建不可达指令
AllocaInst * createAllocaInst(Type *type, const std::string &name = "") {
auto inst = new AllocaInst(type, block, name);
AllocaInst * createAllocaInst(Type *type, const std::vector<Value *> &dims = {}, const std::string &name = "") {
auto inst = new AllocaInst(type, dims, block, name);
assert(inst);
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建分配指令
AllocaInst * createAllocaInstWithoutInsert(Type *type, const std::vector<Value *> &dims = {}, BasicBlock *parent = nullptr,
const std::string &name = "") {
auto inst = new AllocaInst(type, dims, parent, name);
assert(inst);
return inst;
} ///< 创建不插入指令列表的分配指令[仅用于phi指令]
LoadInst * createLoadInst(Value *pointer, const std::vector<Value *> &indices = {}, const std::string &name = "") {
std::string newName;
if (name.empty()) {
@ -280,111 +275,71 @@ class IRBuilder {
newName = name;
}
auto inst = new LoadInst(pointer, block, newName);
auto inst = new LoadInst(pointer, indices, block, newName);
assert(inst);
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建load指令
LaInst * createLaInst(Value *pointer, const std::vector<Value *> &indices = {}, const std::string &name = "") {
std::string newName;
if (name.empty()) {
std::stringstream ss;
ss << tmpIndex;
newName = ss.str();
tmpIndex++;
} else {
newName = name;
}
auto inst = new LaInst(pointer, indices, block, newName);
assert(inst);
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建la指令
GetSubArrayInst * createGetSubArray(LVal *fatherArray, const std::vector<Value *> &indices, const std::string &name = "") {
assert(fatherArray->getLValNumDims() > indices.size());
std::vector<Value *> subDims;
auto dims = fatherArray->getLValDims();
auto iter = std::next(dims.begin(), indices.size());
while (iter != dims.end()) {
subDims.emplace_back(*iter);
iter++;
}
std::string childArrayName;
std::stringstream ss;
ss << "A"
<< "%" << tmpIndex;
childArrayName = ss.str();
tmpIndex++;
auto fatherArrayValue = dynamic_cast<Value *>(fatherArray);
auto childArray = new AllocaInst(fatherArrayValue->getType(), subDims, block, childArrayName);
auto inst = new GetSubArrayInst(fatherArray, childArray, indices, block, childArrayName);
assert(inst);
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建获取部分数组指令
MemsetInst * createMemsetInst(Value *pointer, Value *begin, Value *size, Value *value, const std::string &name = "") {
auto inst = new MemsetInst(pointer, begin, size, value, block, name);
assert(inst);
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建memset指令
StoreInst * createStoreInst(Value *value, Value *pointer, const std::string &name = "") {
auto inst = new StoreInst(value, pointer, block, name);
StoreInst * createStoreInst(Value *value, Value *pointer, const std::vector<Value *> &indices = {},
const std::string &name = "") {
auto inst = new StoreInst(value, pointer, indices, block, name);
assert(inst);
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建store指令
PhiInst * createPhiInst(Type *type, const std::vector<Value*> &vals = {}, const std::vector<BasicBlock*> &blks = {}, const std::string &name = "") {
std::string newName;
if (name.empty()) {
std::stringstream ss;
ss << tmpIndex;
newName = ss.str();
tmpIndex++;
} else {
newName = name;
}
auto inst = new PhiInst(type, vals, blks, block, newName);
auto predNum = block->getNumPredecessors();
auto inst = new PhiInst(type, vals, blks, block, name);
assert(inst);
block->getInstructions().emplace(block->begin(), inst);
return inst;
} ///< 创建Phi指令
/**
* @brief LLVM GEP
*
*/
GetElementPtrInst *createGetElementPtrInst(Value *basePointer, const std::vector<Value *> &indices,
const std::string &name = "") {
Type *ResultElementType = getIndexedType(basePointer->getType(), indices);
if (!ResultElementType) {
assert(false && "Invalid GEP indexing!");
return nullptr;
}
Type *ResultType = PointerType::get(ResultElementType);
std::string newName;
if (name.empty()) {
std::stringstream ss;
ss << tmpIndex;
newName = ss.str();
tmpIndex++;
} else {
newName = name;
}
auto inst = new GetElementPtrInst(ResultType, basePointer, indices, block, newName);
assert(inst);
block->getInstructions().emplace(position, inst);
return inst;
}
static Type *getIndexedType(Type *pointerType, const std::vector<Value *> &indices) {
assert(pointerType->isPointer() && "base must be a pointer type!");
// GEP 的类型推断从基指针所指向的类型开始。
// 例如:
// - 如果 pointerType 是 `[20 x [10 x i32]]*``currentWalkType` 初始为 `[20 x [10 x i32]]`。
// - 如果 pointerType 是 `i32*``currentWalkType` 初始为 `i32`。
// - 如果 pointerType 是 `i32**``currentWalkType` 初始为 `i32*`。
Type *currentWalkType = pointerType->as<PointerType>()->getBaseType();
// 遍历所有索引来深入类型层次结构。
// `indices` 向量包含了所有 GEP 索引,包括由 `visitLValue` 等函数添加的初始 `0` 索引。
for (int i = 0; i < indices.size(); ++i) {
if (currentWalkType->isArray()) {
// 情况一:当前遍历类型是 `ArrayType`。
// 索引用于选择数组元素,`currentWalkType` 更新为数组的元素类型。
currentWalkType = currentWalkType->as<ArrayType>()->getElementType();
} else if (currentWalkType->isPointer()) {
// 情况二:当前遍历类型是 `PointerType`。
// 这意味着我们正在通过一个指针来访问其指向的内存。
// 索引用于选择该指针所指向的“数组”的元素。
// `currentWalkType` 更新为该指针所指向的基础类型。
// 例如:如果 `currentWalkType` 是 `i32*`,它将变为 `i32`。
// 如果 `currentWalkType` 是 `[10 x i32]*`,它将变为 `[10 x i32]`。
currentWalkType = currentWalkType->as<PointerType>()->getBaseType();
} else {
// 情况三:当前遍历类型是标量类型 (例如 `i32`, `float` 等非聚合、非指针类型)。
//
// 如果 `currentWalkType` 是标量,并且当前索引 `i` **不是** `indices` 向量中的最后一个索引,
// 这意味着尝试对一个标量类型进行进一步的结构性索引,这是**无效的**。
// 例如:`int x; x[0];` 对应的 GEP 链中,`x` 的类型是 `i32`,再加 `[0]` 索引就是错误。
//
// 如果 `currentWalkType` 是标量,且这是**最后一个索引** (`i == indices.size() - 1`)
// 那么 GEP 是合法的,它只是计算一个偏移地址,最终的类型就是这个标量类型。
// 此时 `currentWalkType` 保持不变,循环结束。
if (i < indices.size() - 1) {
assert(false && "Invalid GEP indexing: attempting to index into a non-aggregate/non-pointer type with further indices.");
return nullptr; // 返回空指针表示类型推断失败
}
// 如果是最后一个索引,且当前类型是标量,则类型保持不变,这是合法的。
// 循环会自然结束,返回正确的 `currentWalkType`。
}
}
// 所有索引处理完毕后,`currentWalkType` 就是 GEP 指令最终计算出的地址所指向的元素的类型。
return currentWalkType;
}
};
} // namespace sysy

59
src/include/Mem2Reg.h Normal file
View File

@ -0,0 +1,59 @@
#pragma once
#include <list>
#include <memory>
#include <stack>
#include <unordered_map>
#include <unordered_set>
#include "IR.h"
#include "IRBuilder.h"
#include "SysYIRAnalyser.h"
namespace sysy {
/**
* 实现静态单变量赋值核心类 mem2reg
*/
class Mem2Reg {
private:
Module *pModule;
IRBuilder *pBuilder;
ControlFlowAnalysis *controlFlowAnalysis; // 控制流分析
ActiveVarAnalysis *activeVarAnalysis; // 活跃变量分析
DataFlowAnalysisUtils dataFlowAnalysisUtils;
public:
Mem2Reg(Module *pMoudle, IRBuilder *pBuilder,
ControlFlowAnalysis *pCFA = nullptr, ActiveVarAnalysis *pAVA = nullptr) :
pModule(pMoudle), pBuilder(pBuilder), controlFlowAnalysis(pCFA), activeVarAnalysis(pAVA), dataFlowAnalysisUtils()
{} // 初始化函数
void mem2regPipeline(); ///< mem2reg
private:
// phi节点的插入需要计算IDF
std::unordered_set<BasicBlock *> computeIterDf(const std::unordered_set<BasicBlock *> &blocks); ///< 计算定义块集合的迭代支配边界
auto computeValue2Blocks() -> void; ///< 计算value2block的映射(不包括数组和global)
auto preOptimize1() -> void; ///< llvm memtoreg预优化1: 删除不含load的alloc和store
auto preOptimize2() -> void; ///< llvm memtoreg预优化2: 针对某个变量的Defblocks只有一个块的情况
auto preOptimize3() -> void; ///< llvm memtoreg预优化3: 针对某个变量的所有读写都在同一个块中的情况
auto insertPhi() -> void; ///< 为所有变量的迭代支配边界插入phi结点
auto rename(BasicBlock *block, std::unordered_map<Value *, int> &count,
std::unordered_map<Value *, std::stack<Instruction *>> &stacks) -> void; ///< 单个块的重命名
auto renameAll() -> void; ///< 重命名所有块
// private helper function.
private:
auto getPredIndex(BasicBlock *n, BasicBlock *s) -> int; ///< 获取前驱索引
auto cascade(Instruction *instr, bool &changed, Function *func, BasicBlock *block,
std::list<std::unique_ptr<Instruction>> &instrs) -> void; ///< 消除级联关系
auto isGlobal(Value *val) -> bool; ///< 判断是否是全局变量
auto isArr(Value *val) -> bool; ///< 判断是否是数组
auto usedelete(Instruction *instr) -> void; ///< 删除指令相关的value-use-user关系
};
} // namespace sysy

View File

@ -12,25 +12,22 @@ namespace sysy {
class RISCv64AsmPrinter {
public:
RISCv64AsmPrinter(MachineFunction* mfunc);
// 主入口
void run(std::ostream& os, bool debug = false);
void printInstruction(MachineInstr* instr, bool debug = false);
// 辅助函数
void setStream(std::ostream& os) { OS = &os; }
// 辅助函数
std::string regToString(PhysicalReg reg);
std::string formatInstr(const MachineInstr *instr);
private:
// 打印各个部分
void printPrologue();
void printEpilogue();
void printBasicBlock(MachineBasicBlock* mbb, bool debug = false);
void printInstruction(MachineInstr* instr, bool debug = false);
// 辅助函数
std::string regToString(PhysicalReg reg);
void printOperand(MachineOperand* op);
MachineFunction* MFunc;
std::ostream* OS = nullptr;
std::ostream* OS;
};
} // namespace sysy

View File

@ -22,9 +22,6 @@ private:
// 函数级代码生成 (实现新的流水线)
std::string function_gen(Function* func);
// 私有辅助函数,用于根据类型计算其占用的字节数。
unsigned getTypeSizeInBytes(Type* type);
Module* module;
};

View File

@ -3,12 +3,6 @@
#include "RISCv64LLIR.h"
// Forward declarations
namespace sysy {
class GlobalValue;
class Value;
}
extern int DEBUG;
extern int DEEPDEBUG;
@ -22,12 +16,7 @@ public:
// 公开接口以便后续模块如RegAlloc可以查询或创建vreg
unsigned getVReg(Value* val);
unsigned getNewVReg(Type* type);
unsigned getVRegCounter() const;
// 获取 vreg_map 的公共接口
const std::map<Value*, unsigned>& getVRegMap() const { return vreg_map; }
const std::map<unsigned, Value*>& getVRegValueMap() const { return vreg_to_value_map; }
const std::map<unsigned, Type*>& getVRegTypeMap() const { return vreg_type_map; }
unsigned getNewVReg() { return vreg_counter++; }
private:
// DAG节点定义作为ISel的内部实现细节
@ -44,10 +33,7 @@ private:
std::vector<std::unique_ptr<DAGNode>> build_dag(BasicBlock* bb);
DAGNode* get_operand_node(Value* val_ir, std::map<Value*, DAGNode*>&, std::vector<std::unique_ptr<DAGNode>>&);
DAGNode* create_node(int kind, Value* val, std::map<Value*, DAGNode*>&, std::vector<std::unique_ptr<DAGNode>>&);
// 用于计算类型大小的辅助函数
unsigned getTypeSizeInBytes(Type* type);
// 打印DAG图以供调试
void print_dag(const std::vector<std::unique_ptr<DAGNode>>& dag, const std::string& bb_name);
// 状态
@ -57,8 +43,6 @@ private:
// 映射关系
std::map<Value*, unsigned> vreg_map;
std::map<unsigned, Value*> vreg_to_value_map;
std::map<unsigned, Type*> vreg_type_map;
std::map<const BasicBlock*, MachineBasicBlock*> bb_map;
unsigned vreg_counter;

View File

@ -3,7 +3,6 @@
#include "IR.h" // 确保包含了您自己的IR头文件
#include <string>
#include <iostream>
#include <vector>
#include <memory>
#include <cstdint>
@ -19,36 +18,14 @@ namespace sysy {
// 物理寄存器定义
enum class PhysicalReg {
// --- 特殊功能寄存器 ---
ZERO, RA, SP, GP, TP,
// --- 整数寄存器 (按调用约定分组) ---
// 临时寄存器 (调用者保存)
T0, T1, T2, T3, T4, T5, T6,
// 保存寄存器 (被调用者保存)
S0, S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11,
// 参数/返回值寄存器 (调用者保存)
A0, A1, A2, A3, A4, A5, A6, A7,
// --- 浮点寄存器 ---
F0, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11,
F12, F13, F14, F15, F16, F17, F18, F19, F20, F21,
F22, F23, F24, F25, F26, F27, F28, F29, F30, F31,
// 用于内部表示物理寄存器在干扰图中的节点ID一个简单的特殊ID确保不与vreg_counter冲突
// 假设 vreg_counter 不会达到这么大的值
PHYS_REG_START_ID = 1000000,
PHYS_REG_END_ID = PHYS_REG_START_ID + 320, // 预留足够的空间
INVALID, ///< 无效寄存器标记
ZERO, RA, SP, GP, TP, T0, T1, T2, S0, S1, A0, A1, A2, A3, A4, A5, A6, A7, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, T3, T4, T5, T6,
F0, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12, F13, F14, F15,F16, F17, F18, F19, F20, F21, F22, F23, F24, F25, F26, F27, F28, F29, F30, F31
};
// RISC-V 指令操作码枚举
enum class RVOpcodes {
// 算术指令
ADD, ADDI, ADDW, ADDIW, SUB, SUBW, MUL, MULW, MULH, DIV, DIVW, REM, REMW,
ADD, ADDI, ADDW, ADDIW, SUB, SUBW, MUL, MULW, DIV, DIVW, REM, REMW,
// 逻辑指令
XOR, XORI, OR, ORI, AND, ANDI,
// 移位指令
@ -66,100 +43,12 @@ enum class RVOpcodes {
CALL,
// 特殊标记,非指令
LABEL,
// 浮点指令 (RISC-V 'F' 扩展)
// 浮点加载与存储
FLW, // flw rd, offset(rs1)
FSW, // fsw rs2, offset(rs1)
FLD, // fld rd, offset(rs1)
FSD, // fsd rs2, offset(rs1)
// 浮点算术运算 (单精度)
FADD_S, // fadd.s rd, rs1, rs2
FSUB_S, // fsub.s rd, rs1, rs2
FMUL_S, // fmul.s rd, rs1, rs2
FDIV_S, // fdiv.s rd, rs1, rs2
// 浮点比较 (单精度)
FEQ_S, // feq.s rd, rs1, rs2 (结果写入整数寄存器rd)
FLT_S, // flt.s rd, rs1, rs2 (less than)
FLE_S, // fle.s rd, rs1, rs2 (less than or equal)
// 浮点转换
FCVT_S_W, // fcvt.s.w rd, rs1 (有符号整数 -> 单精度浮点)
FCVT_W_S, // fcvt.w.s rd, rs1 (单精度浮点 -> 有符号整数)
FCVT_W_S_RTZ, // fcvt.w.s rd, rs1, rtz (使用向零截断模式)
// 浮点传送/移动
FMV_S, // fmv.s rd, rs1 (浮点寄存器之间)
FMV_W_X, // fmv.w.x rd, rs1 (整数寄存器位模式 -> 浮点寄存器)
FMV_X_W, // fmv.x.w rd, rs1 (浮点寄存器位模式 -> 整数寄存器)
FNEG_S, // fneg.s rd, rs (浮点取负)
// 伪指令
FRAME_LOAD_W, // 从栈帧加载 32位 Word (对应 lw)
FRAME_LOAD_D, // 从栈帧加载 64位 Doubleword (对应 ld)
FRAME_STORE_W, // 保存 32位 Word 到栈帧 (对应 sw)
FRAME_STORE_D, // 保存 64位 Doubleword 到栈帧 (对应 sd)
FRAME_LOAD_F, // 从栈帧加载单精度浮点数
FRAME_STORE_F, // 将单精度浮点数存入栈帧
FRAME_ADDR, // 获取栈帧变量的地址
PSEUDO_KEEPALIVE, // 保持寄存器活跃,防止优化器删除
// 新增伪指令,用于解耦栈帧处理
FRAME_LOAD, // 从栈帧加载 (AllocaInst)
FRAME_STORE, // 保存到栈帧 (AllocaInst)
FRAME_ADDR, // [新] 获取栈帧变量的地址
};
inline bool isGPR(PhysicalReg reg) {
return reg >= PhysicalReg::ZERO && reg <= PhysicalReg::T6;
}
// 判断一个物理寄存器是否是浮点寄存器 (FPR)
inline bool isFPR(PhysicalReg reg) {
return reg >= PhysicalReg::F0 && reg <= PhysicalReg::F31;
}
// 获取所有调用者保存的整数寄存器 (t0-t6, a0-a7)
inline const std::vector<PhysicalReg>& getCallerSavedIntRegs() {
static const std::vector<PhysicalReg> regs = {
PhysicalReg::T0, PhysicalReg::T1, PhysicalReg::T2, PhysicalReg::T3,
PhysicalReg::T4, PhysicalReg::T5, PhysicalReg::T6,
PhysicalReg::A0, PhysicalReg::A1, PhysicalReg::A2, PhysicalReg::A3,
PhysicalReg::A4, PhysicalReg::A5, PhysicalReg::A6, PhysicalReg::A7
};
return regs;
}
// 获取所有被调用者保存的整数寄存器 (s0-s11)
inline const std::vector<PhysicalReg>& getCalleeSavedIntRegs() {
static const std::vector<PhysicalReg> regs = {
PhysicalReg::S0, PhysicalReg::S1, PhysicalReg::S2, PhysicalReg::S3,
PhysicalReg::S4, PhysicalReg::S5, PhysicalReg::S6, PhysicalReg::S7,
PhysicalReg::S8, PhysicalReg::S9, PhysicalReg::S10, PhysicalReg::S11
};
return regs;
}
// 获取所有调用者保存的浮点寄存器 (ft0-ft11, fa0-fa7)
inline const std::vector<PhysicalReg>& getCallerSavedFpRegs() {
static const std::vector<PhysicalReg> regs = {
PhysicalReg::F0, PhysicalReg::F1, PhysicalReg::F2, PhysicalReg::F3,
PhysicalReg::F4, PhysicalReg::F5, PhysicalReg::F6, PhysicalReg::F7,
PhysicalReg::F8, PhysicalReg::F9, PhysicalReg::F10, PhysicalReg::F11, // ft0-ft11 和 fa0-fa7 在标准ABI中重叠
PhysicalReg::F12, PhysicalReg::F13, PhysicalReg::F14, PhysicalReg::F15,
PhysicalReg::F16, PhysicalReg::F17
};
return regs;
}
// 获取所有被调用者保存的浮点寄存器 (fs0-fs11)
inline const std::vector<PhysicalReg>& getCalleeSavedFpRegs() {
static const std::vector<PhysicalReg> regs = {
PhysicalReg::F18, PhysicalReg::F19, PhysicalReg::F20, PhysicalReg::F21,
PhysicalReg::F22, PhysicalReg::F23, PhysicalReg::F24, PhysicalReg::F25,
PhysicalReg::F26, PhysicalReg::F27, PhysicalReg::F28, PhysicalReg::F29,
PhysicalReg::F30, PhysicalReg::F31
};
return regs;
}
class MachineOperand;
class RegOperand;
class ImmOperand;
@ -199,11 +88,6 @@ public:
preg = new_preg;
is_virtual = false;
}
void setVRegNum(unsigned new_vreg_num) {
vreg_num = new_vreg_num;
is_virtual = true; // 确保设置vreg时操作数状态正确
}
private:
unsigned vreg_num = 0;
PhysicalReg preg = PhysicalReg::ZERO;
@ -283,15 +167,10 @@ private:
// 栈帧信息
struct StackFrameInfo {
int locals_size = 0; // 仅为AllocaInst分配的大小
int locals_end_offset = 0; // 记录局部变量分配结束后的偏移量(相对于s0为负)
int spill_size = 0; // 仅为溢出分配的大小
int total_size = 0; // 总大小
int callee_saved_size = 0; // 保存寄存器的大小
std::map<unsigned, int> alloca_offsets; // <AllocaInst的vreg, 栈偏移>
std::map<unsigned, int> spill_offsets; // <溢出vreg, 栈偏移>
std::set<PhysicalReg> used_callee_saved_regs; // 使用的保存寄存器
std::map<unsigned, PhysicalReg> vreg_to_preg_map; // RegAlloc最终的分配结果
std::vector<PhysicalReg> callee_saved_regs_to_store; // 已排序的、需要存取的被调用者保存寄存器
};
// 机器函数
@ -305,7 +184,7 @@ public:
StackFrameInfo& getFrameInfo() { return frame_info; }
const std::vector<std::unique_ptr<MachineBasicBlock>>& getBlocks() const { return blocks; }
std::vector<std::unique_ptr<MachineBasicBlock>>& getBlocks() { return blocks; }
void dumpStackFrameInfo(std::ostream& os = std::cerr) const;
void addBlock(std::unique_ptr<MachineBasicBlock> block) {
blocks.push_back(std::move(block));
}

View File

@ -0,0 +1,61 @@
#ifndef RISCV64_PASSES_H
#define RISCV64_PASSES_H
#include "RISCv64LLIR.h"
namespace sysy {
/**
* @class Pass
* @brief 所有优化Pass的抽象基类 (可选,但推荐)
* * 定义一个通用的接口,所有优化都应该实现它。
*/
class Pass {
public:
virtual ~Pass() = default;
virtual void runOnMachineFunction(MachineFunction* mfunc) = 0;
};
// --- 寄存器分配前优化 ---
/**
* @class PreRA_Scheduler
* @brief 寄存器分配前的指令调度器
* * 在虚拟寄存器上进行操作,此时调度自由度最大,
* 主要目标是隐藏指令延迟,提高流水线效率。
*/
class PreRA_Scheduler : public Pass {
public:
void runOnMachineFunction(MachineFunction* mfunc) override;
};
// --- 寄存器分配后优化 ---
/**
* @class PeepholeOptimizer
* @brief 窥孔优化器
* * 在已分配物理寄存器的指令流上,通过一个小的滑动窗口来查找
* 并替换掉一些冗余或低效的指令模式。
*/
class PeepholeOptimizer : public Pass {
public:
void runOnMachineFunction(MachineFunction* mfunc) override;
};
/**
* @class PostRA_Scheduler
* @brief 寄存器分配后的局部指令调度器
* * 主要目标是优化寄存器分配器插入的spill/fill代码(lw/sw)
* 尝试将加载指令提前,以隐藏其访存延迟。
*/
class PostRA_Scheduler : public Pass {
public:
void runOnMachineFunction(MachineFunction* mfunc) override;
};
} // namespace sysy
#endif // RISCV64_PASSES_H

View File

@ -0,0 +1,56 @@
#ifndef RISCV64_REGALLOC_H
#define RISCV64_REGALLOC_H
#include "RISCv64LLIR.h"
namespace sysy {
class RISCv64RegAlloc {
public:
RISCv64RegAlloc(MachineFunction* mfunc);
// 模块主入口
void run();
private:
using LiveSet = std::set<unsigned>; // 活跃虚拟寄存器集合
using InterferenceGraph = std::map<unsigned, std::set<unsigned>>;
// 栈帧管理
void eliminateFrameIndices();
// 活跃性分析
void analyzeLiveness();
// 构建干扰图
void buildInterferenceGraph();
// 图着色分配寄存器
void colorGraph();
// 重写函数替换vreg并插入溢出代码
void rewriteFunction();
// 辅助函数获取指令的Use/Def集合
void getInstrUseDef(MachineInstr* instr, LiveSet& use, LiveSet& def);
MachineFunction* MFunc;
// 活跃性分析结果
std::map<const MachineInstr*, LiveSet> live_in_map;
std::map<const MachineInstr*, LiveSet> live_out_map;
// 干扰图
InterferenceGraph interference_graph;
// 图着色结果
std::map<unsigned, PhysicalReg> color_map; // vreg -> preg
std::set<unsigned> spilled_vregs; // 被溢出的vreg集合
// 可用的物理寄存器池
std::vector<PhysicalReg> allocable_int_regs;
};
} // namespace sysy
#endif // RISCV64_REGALLOC_H

23
src/include/Reg2Mem.h Normal file
View File

@ -0,0 +1,23 @@
#pragma once
#include "IR.h"
#include "IRBuilder.h"
namespace sysy {
/**
* Reg2Mem(后端未做phi指令翻译)
*/
class Reg2Mem {
private:
Module *pModule;
IRBuilder *pBuilder;
public:
Reg2Mem(Module *pMoudle, IRBuilder *pBuilder) : pModule(pMoudle), pBuilder(pBuilder) {}
void DeletePhiInst();
// 删除UD关系, 因为删除了phi指令会修改ud关系
void usedelete(Instruction *instr);
};
} // namespace sysy

340
src/include/SysYFormatter.h Normal file
View File

@ -0,0 +1,340 @@
#pragma once
#include "SysYBaseVisitor.h"
#include "SysYParser.h"
#include <ostream>
namespace sysy {
class SysYFormatter : public SysYBaseVisitor {
protected:
std::ostream &os;
int indent = 0;
public:
SysYFormatter(std::ostream &os) : os(os), indent(0) {}
protected:
struct Indentor {
static constexpr int TabSize = 2;
int &indent;
Indentor(int &indent) : indent(indent) { indent += TabSize; }
~Indentor() { indent -= TabSize; }
};
std::ostream &space() { return os << std::string(indent, ' '); }
template <typename T>
std::ostream &interleave(const T &container, const std::string sep = ", ") {
auto b = container.begin(), e = container.end();
(*b)->accept(this);
for (b = std::next(b); b != e; b = std::next(b)) {
os << sep;
(*b)->accept(this);
}
return os;
}
public:
// virtual std::any visitModule(SysYParser::ModuleContext *ctx) override {
// return visitChildren(ctx);
// }
virtual std::any visitBtype(SysYParser::BtypeContext *ctx) override {
os << ctx->getText();
return 0;
}
virtual std::any visitDecl(SysYParser::DeclContext *ctx) override {
space();
if (ctx->CONST())
os << ctx->CONST()->getText() << ' ';
ctx->btype()->accept(this);
os << ' ';
interleave(ctx->varDef(), ", ") << ';' << '\n';
return 0;
}
virtual std::any visitVarDef(SysYParser::VarDefContext *ctx) override {
ctx->lValue()->accept(this);
if (ctx->initValue()) {
os << ' ' << '=' << ' ';
ctx->initValue()->accept(this);
}
return 0;
}
virtual std::any visitInitValue(SysYParser::InitValueContext *ctx) override {
if (not ctx->exp()) {
os << '{';
auto values = ctx->initValue();
if (values.size())
interleave(values, ", ");
os << '}';
}
return 0;
}
virtual std::any visitFunc(SysYParser::FuncContext *ctx) override {
ctx->funcType()->accept(this);
os << ' ' << ctx->ID()->getText() << '(';
if (ctx->funcFParams())
ctx->funcFParams()->accept(this);
os << ')' << ' ';
ctx->blockStmt()->accept(this);
os << '\n';
return 0;
}
virtual std::any visitFuncType(SysYParser::FuncTypeContext *ctx) override {
os << ctx->getText();
return 0;
}
virtual std::any
visitFuncFParams(SysYParser::FuncFParamsContext *ctx) override {
interleave(ctx->funcFParam(), ", ");
return 0;
}
virtual std::any
visitFuncFParam(SysYParser::FuncFParamContext *ctx) override {
ctx->btype()->accept(this);
os << ' ' << ctx->ID()->getText();
if (not ctx->LBRACKET().empty()) {
os << '[';
auto exp = ctx->exp();
if (not exp.empty()) {
os << '[';
interleave(exp, "][") << ']';
}
}
return 0;
}
virtual std::any visitBlockStmt(SysYParser::BlockStmtContext *ctx) override {
os << '{' << '\n';
{
Indentor indentor(indent);
auto items = ctx->blockItem();
if (not items.empty())
interleave(items, "");
}
space() << ctx->RBRACE()->getText() << '\n';
return 0;
}
// virtual std::any visitBlockItem(SysYParser::BlockItemContext *ctx)
// override {
// return visitChildren(ctx);
// }
// virtual std::any visitStmt(SysYParser::StmtContext *ctx) override {
// return visitChildren(ctx);
// }
virtual std::any
visitAssignStmt(SysYParser::AssignStmtContext *ctx) override {
space();
ctx->lValue()->accept(this);
os << " = ";
ctx->exp()->accept(this);
os << ';' << '\n';
return 0;
}
virtual std::any visitExpStmt(SysYParser::ExpStmtContext *ctx) override {
space();
ctx->exp()->accept(this);
os << ';' << '\n';
return 0;
}
void wrapBlock(SysYParser::StmtContext *stmt) {
bool isBlock = stmt->blockStmt();
if (isBlock) {
stmt->accept(this);
} else {
os << "{\n";
{
Indentor indentor(indent);
stmt->accept(this);
}
space() << "}\n";
}
};
virtual std::any visitIfStmt(SysYParser::IfStmtContext *ctx) override {
space();
os << ctx->IF()->getText() << " (";
ctx->exp()->accept(this);
os << ") ";
auto stmt = ctx->stmt();
auto ifStmt = stmt[0];
wrapBlock(ifStmt);
if (stmt.size() == 2) {
auto elseStmt = stmt[1];
wrapBlock(elseStmt);
}
return 0;
}
virtual std::any visitWhileStmt(SysYParser::WhileStmtContext *ctx) override {
space();
os << ctx->WHILE()->getText() << " (";
ctx->exp()->accept(this);
os << ") ";
wrapBlock(ctx->stmt());
return 0;
}
virtual std::any visitBreakStmt(SysYParser::BreakStmtContext *ctx) override {
space() << ctx->BREAK()->getText() << ';' << '\n';
return 0;
}
virtual std::any
visitContinueStmt(SysYParser::ContinueStmtContext *ctx) override {
space() << ctx->CONTINUE()->getText() << ';' << '\n';
return 0;
}
virtual std::any
visitReturnStmt(SysYParser::ReturnStmtContext *ctx) override {
space() << ctx->RETURN()->getText();
if (ctx->exp()) {
os << ' ';
ctx->exp()->accept(this);
}
os << ';' << '\n';
return 0;
}
// virtual std::any visitEmptyStmt(SysYParser::EmptyStmtContext *ctx)
// override {
// return visitChildren(ctx);
// }
virtual std::any
visitRelationExp(SysYParser::RelationExpContext *ctx) override {
auto lhs = ctx->exp(0);
auto rhs = ctx->exp(1);
std::string op =
ctx->LT() ? "<" : (ctx->LE() ? "<=" : (ctx->GT() ? ">" : ">="));
lhs->accept(this);
os << ' ' << op << ' ';
rhs->accept(this);
return 0;
}
virtual std::any
visitMultiplicativeExp(SysYParser::MultiplicativeExpContext *ctx) override {
auto lhs = ctx->exp(0);
auto rhs = ctx->exp(1);
std::string op = ctx->MUL() ? "*" : (ctx->DIV() ? "/" : "%");
lhs->accept(this);
os << ' ' << op << ' ';
rhs->accept(this);
return 0;
}
// virtual std::any visitLValueExp(SysYParser::LValueExpContext *ctx)
// override {
// return visitChildren(ctx);
// }
// virtual std::any visitNumberExp(SysYParser::NumberExpContext *ctx)
// override {
// return visitChildren(ctx);
// }
virtual std::any visitAndExp(SysYParser::AndExpContext *ctx) override {
ctx->exp(0)->accept(this);
os << " && ";
ctx->exp(1)->accept(this);
return 0;
}
virtual std::any visitUnaryExp(SysYParser::UnaryExpContext *ctx) override {
std::string op = ctx->ADD() ? "+" : (ctx->SUB() ? "-" : "!");
os << op;
ctx->exp()->accept(this);
return 0;
}
virtual std::any visitParenExp(SysYParser::ParenExpContext *ctx) override {
os << '(';
ctx->exp()->accept(this);
os << ')';
return 0;
}
virtual std::any visitStringExp(SysYParser::StringExpContext *ctx) override {
return visitChildren(ctx);
}
virtual std::any visitOrExp(SysYParser::OrExpContext *ctx) override {
ctx->exp(0)->accept(this);
os << " || ";
ctx->exp(1)->accept(this);
return 0;
}
// virtual std::any visitCallExp(SysYParser::CallExpContext *ctx) override {
// return visitChildren(ctx);
// }
virtual std::any
visitAdditiveExp(SysYParser::AdditiveExpContext *ctx) override {
auto lhs = ctx->exp(0);
auto rhs = ctx->exp(1);
std::string op = ctx->ADD() ? "+" : "-";
lhs->accept(this);
os << ' ' << op << ' ';
rhs->accept(this);
return 0;
}
virtual std::any visitEqualExp(SysYParser::EqualExpContext *ctx) override {
auto lhs = ctx->exp(0);
auto rhs = ctx->exp(1);
std::string op = ctx->EQ() ? "==" : "!=";
lhs->accept(this);
os << ' ' << op << ' ';
rhs->accept(this);
return 0;
}
virtual std::any visitCall(SysYParser::CallContext *ctx) override {
os << ctx->ID()->getText() << '(';
if (ctx->funcRParams())
ctx->funcRParams()->accept(this);
os << ')';
return 0;
}
virtual std::any visitLValue(SysYParser::LValueContext *ctx) override {
os << ctx->ID()->getText();
auto exp = ctx->exp();
if (not exp.empty()) {
os << '[';
interleave(exp, "][") << ']';
}
return 0;
}
virtual std::any visitNumber(SysYParser::NumberContext *ctx) override {
os << ctx->getText();
return 0;
}
virtual std::any visitString(SysYParser::StringContext *ctx) override {
os << ctx->getText();
return 0;
}
virtual std::any
visitFuncRParams(SysYParser::FuncRParamsContext *ctx) override {
interleave(ctx->exp(), ", ");
return 0;
}
};
} // namespace sysy

View File

@ -0,0 +1,465 @@
#pragma once
#include "IR.h"
namespace sysy {
// 前向声明
class Loop;
// 基本块分析信息类
class BlockAnalysisInfo {
public:
using block_list = std::vector<BasicBlock*>;
using block_set = std::unordered_set<BasicBlock*>;
protected:
// 支配树相关
int domdepth = 0; ///< 支配节点所在深度
BasicBlock* idom = nullptr; ///< 直接支配结点
block_list sdoms; ///< 支配树后继
block_set dominants; ///< 必经结点集合
block_set dominant_frontiers; ///< 支配边界
// 后续添加循环分析相关
// Loop* loopbelong = nullptr; ///< 所属循环
// int loopdepth = 0; ///< 循环深度
public:
// getterface
const int getDomDepth() const { return domdepth; }
const BasicBlock* getIdom() const { return idom; }
const block_list& getSdoms() const { return sdoms; }
const block_set& getDominants() const { return dominants; }
const block_set& getDomFrontiers() const { return dominant_frontiers; }
// 支配树操作
void setDomDepth(int depth) { domdepth = depth; }
void setIdom(BasicBlock* block) { idom = block; }
void addSdoms(BasicBlock* block) { sdoms.push_back(block); }
void clearSdoms() { sdoms.clear(); }
void removeSdoms(BasicBlock* block) {
sdoms.erase(std::remove(sdoms.begin(), sdoms.end(), block), sdoms.end());
}
void addDominants(BasicBlock* block) { dominants.emplace(block); }
void addDominants(const block_set& blocks) { dominants.insert(blocks.begin(), blocks.end()); }
void setDominants(BasicBlock* block) {
dominants.clear();
addDominants(block);
}
void setDominants(const block_set& doms) {
dominants = doms;
}
void setDomFrontiers(const block_set& df) {
dominant_frontiers = df;
}
// TODO循环分析操作方法
// 清空所有分析信息
void clear() {
domdepth = -1;
idom = nullptr;
sdoms.clear();
dominants.clear();
dominant_frontiers.clear();
// loopbelong = nullptr;
// loopdepth = 0;
}
};
// 函数分析信息类
class FunctionAnalysisInfo {
public:
// 函数属性
enum FunctionAttribute : uint64_t {
PlaceHolder = 0x0UL,
Pure = 0x1UL << 0,
SelfRecursive = 0x1UL << 1,
SideEffect = 0x1UL << 2,
NoPureCauseMemRead = 0x1UL << 3
};
// 数据结构
using Loop_list = std::list<std::unique_ptr<Loop>>;
using block_loop_map = std::unordered_map<BasicBlock*, Loop*>;
using value_block_map = std::unordered_map<Value*, BasicBlock*>;
using value_block_count_map = std::unordered_map<Value*, std::unordered_map<BasicBlock*, int>>;
// 分析数据
FunctionAttribute attribute = PlaceHolder; ///< 函数属性
std::set<Function*> callees; ///< 函数调用集合
Loop_list loops; ///< 所有循环
Loop_list topLoops; ///< 顶层循环
// block_loop_map basicblock2Loop; ///< 基本块到循环映射
std::list<std::unique_ptr<AllocaInst>> indirectAllocas; ///< 间接分配内存
// 值定义/使用信息
value_block_map value2AllocBlocks; ///< 值分配位置映射
value_block_count_map value2DefBlocks; ///< 值定义位置映射
value_block_count_map value2UseBlocks; ///< 值使用位置映射
// 函数属性操作
FunctionAttribute getAttribute() const { return attribute; }
void setAttribute(FunctionAttribute attr) { attribute = static_cast<FunctionAttribute>(attribute | attr); }
void clearAttribute() { attribute = PlaceHolder; }
// 调用关系操作
void addCallee(Function* callee) { callees.insert(callee); }
void removeCallee(Function* callee) { callees.erase(callee); }
void clearCallees() { callees.clear(); }
// 值-块映射操作
BasicBlock* getAllocBlockByValue(Value* value) {
auto it = value2AllocBlocks.find(value);
return it != value2AllocBlocks.end() ? it->second : nullptr;
}
std::unordered_set<BasicBlock *> getDefBlocksByValue(Value *value) {
std::unordered_set<BasicBlock *> blocks;
if (value2DefBlocks.count(value) > 0) {
for (const auto &pair : value2DefBlocks[value]) {
blocks.insert(pair.first);
}
}
return blocks;
}
std::unordered_set<BasicBlock *> getUseBlocksByValue(Value *value) {
std::unordered_set<BasicBlock *> blocks;
if (value2UseBlocks.count(value) > 0) {
for (const auto &pair : value2UseBlocks[value]) {
blocks.insert(pair.first);
}
}
return blocks;
}
// 值定义/使用操作
void addValue2AllocBlocks(Value* value, BasicBlock* block) { value2AllocBlocks[value] = block; }
void addValue2DefBlocks(Value* value, BasicBlock* block) { ++value2DefBlocks[value][block]; }
void addValue2UseBlocks(Value* value, BasicBlock* block) { ++value2UseBlocks[value][block]; }
// 获取值定义/使用信息
std::unordered_map<Value *, BasicBlock *>& getValue2AllocBlocks() {
return value2AllocBlocks;
}
std::unordered_map<Value *, std::unordered_map<BasicBlock *, int>>& getValue2DefBlocks() {
return value2DefBlocks;
}
std::unordered_map<Value *, std::unordered_map<BasicBlock *, int>>& getValue2UseBlocks() {
return value2UseBlocks;
}
std::unordered_set<Value *> getValuesOfDefBlock() {
std::unordered_set<Value *> values;
for (const auto &pair : value2DefBlocks) {
values.insert(pair.first);
}
return values;
}
// 删除信息操作
void removeValue2AllocBlock(Value *value) { value2AllocBlocks.erase(value); }
bool removeValue2DefBlock(Value *value, BasicBlock *block) {
bool changed = false;
if (--value2DefBlocks[value][block] == 0) {
value2DefBlocks[value].erase(block);
if (value2DefBlocks[value].empty()) {
value2DefBlocks.erase(value);
changed = true;
}
}
return changed;
}
bool removeValue2UseBlock(Value *value, BasicBlock *block) {
bool changed = false;
if (--value2UseBlocks[value][block] == 0) {
value2UseBlocks[value].erase(block);
if (value2UseBlocks[value].empty()) {
value2UseBlocks.erase(value);
changed = true;
}
}
return changed;
}
// 间接分配操作
void addIndirectAlloca(AllocaInst* alloca) { indirectAllocas.emplace_back(alloca); }
std::list<std::unique_ptr<AllocaInst>>& getIndirectAllocas() { return indirectAllocas; }
// TODO循环分析操作
// 清空所有分析信息
void clear() {
attribute = PlaceHolder;
callees.clear();
loops.clear();
topLoops.clear();
// basicblock2Loop.clear();
indirectAllocas.clear();
value2AllocBlocks.clear();
value2DefBlocks.clear();
value2UseBlocks.clear();
}
};
// 循环类 - 未实现优化
class Loop {
public:
using block_list = std::vector<BasicBlock *>;
using block_set = std::unordered_set<BasicBlock *>;
using Loop_list = std::vector<Loop *>;
protected:
Function *parent; // 所属函数
block_list blocksInLoop; // 循环内的基本块
BasicBlock *preheaderBlock = nullptr; // 前驱块
BasicBlock *headerBlock = nullptr; // 循环头
block_list latchBlock; // 回边块
block_set exitingBlocks; // 退出块
block_set exitBlocks; // 退出目标块
Loop *parentloop = nullptr; // 父循环
Loop_list subLoops; // 子循环
size_t loopID; // 循环ID
unsigned loopDepth; // 循环深度
Instruction *indCondVar = nullptr; // 循环条件变量
Instruction::Kind IcmpKind; // 比较类型
Value *indEnd = nullptr; // 循环结束值
AllocaInst *IndPhi = nullptr; // 循环变量
ConstantValue *indBegin = nullptr; // 循环起始值
ConstantValue *indStep = nullptr; // 循环步长
std::set<GlobalValue *> GlobalValuechange; // 循环内改变的全局变量
int StepType = 0; // 循环步长类型
bool parallelable = false; // 是否可并行
public:
explicit Loop(BasicBlock *header, const std::string &name = "")
: headerBlock(header) {
blocksInLoop.push_back(header);
}
void setloopID() {
static unsigned loopCount = 0;
loopCount = loopCount + 1;
loopID = loopCount;
}
ConstantValue* getindBegin() { return indBegin; }
ConstantValue* getindStep() { return indStep; }
void setindBegin(ConstantValue *indBegin2set) { indBegin = indBegin2set; }
void setindStep(ConstantValue *indStep2set) { indStep = indStep2set; }
void setStepType(int StepType2Set) { StepType = StepType2Set; }
int getStepType() { return StepType; }
size_t getLoopID() { return loopID; }
BasicBlock* getHeader() const { return headerBlock; }
BasicBlock* getPreheaderBlock() const { return preheaderBlock; }
block_list& getLatchBlocks() { return latchBlock; }
block_set& getExitingBlocks() { return exitingBlocks; }
block_set& getExitBlocks() { return exitBlocks; }
Loop* getParentLoop() const { return parentloop; }
void setParentLoop(Loop *parent) { parentloop = parent; }
void addBasicBlock(BasicBlock *bb) { blocksInLoop.push_back(bb); }
void addSubLoop(Loop *loop) { subLoops.push_back(loop); }
void setLoopDepth(unsigned depth) { loopDepth = depth; }
block_list& getBasicBlocks() { return blocksInLoop; }
Loop_list& getSubLoops() { return subLoops; }
unsigned getLoopDepth() const { return loopDepth; }
bool isLoopContainsBasicBlock(BasicBlock *bb) const {
return std::find(blocksInLoop.begin(), blocksInLoop.end(), bb) != blocksInLoop.end();
}
void addExitingBlock(BasicBlock *bb) { exitingBlocks.insert(bb); }
void addExitBlock(BasicBlock *bb) { exitBlocks.insert(bb); }
void addLatchBlock(BasicBlock *bb) { latchBlock.push_back(bb); }
void setPreheaderBlock(BasicBlock *bb) { preheaderBlock = bb; }
void setIndexCondInstr(Instruction *instr) { indCondVar = instr; }
void setIcmpKind(Instruction::Kind kind) { IcmpKind = kind; }
Instruction::Kind getIcmpKind() const { return IcmpKind; }
bool isSimpleLoopInvariant(Value *value) ;
void setIndEnd(Value *value) { indEnd = value; }
void setIndPhi(AllocaInst *phi) { IndPhi = phi; }
Value* getIndEnd() const { return indEnd; }
AllocaInst* getIndPhi() const { return IndPhi; }
Instruction* getIndCondVar() const { return indCondVar; }
void addGlobalValuechange(GlobalValue *globalvaluechange2add) {
GlobalValuechange.insert(globalvaluechange2add);
}
std::set<GlobalValue *>& getGlobalValuechange() {
return GlobalValuechange;
}
void setParallelable(bool flag) { parallelable = flag; }
bool isParallelable() const { return parallelable; }
};
// 控制流分析类
class ControlFlowAnalysis {
private:
Module *pModule; ///< 模块
std::unordered_map<BasicBlock*, BlockAnalysisInfo*> blockAnalysisInfo; // 基本块分析信息表
std::unordered_map<Function*, FunctionAnalysisInfo*> functionAnalysisInfo; // 函数分析信息
public:
explicit ControlFlowAnalysis(Module *pMoudle) : pModule(pMoudle) {}
// 获取基本块分析信息
BlockAnalysisInfo* getBlockAnalysisInfo(BasicBlock *block) {
auto it = blockAnalysisInfo.find(block);
if (it != blockAnalysisInfo.end()) {
return it->second;
}
return nullptr; // 如果未找到返回nullptr
}
FunctionAnalysisInfo* getFunctionAnalysisInfo(Function *func) {
auto it = functionAnalysisInfo.find(func);
if (it != functionAnalysisInfo.end()) {
return it->second;
}
return nullptr; // 如果未找到返回nullptr
}
void init(); // 初始化分析器
void computeDomNode(); // 计算必经结点
void computeDomTree(); // 构造支配树
// std::unordered_set<BasicBlock *> computeDomFrontier(BasicBlock *block) ; // 计算单个块的支配边界(弃用)
void computeDomFrontierAllBlk(); // 计算所有块的支配边界
void runControlFlowAnalysis(); // 运行控制流分析(主要是支配树和支配边界)
void clear(){
for (auto &pair : blockAnalysisInfo) {
delete pair.second; // 清理基本块分析信息
}
blockAnalysisInfo.clear();
for (auto &pair : functionAnalysisInfo) {
delete pair.second; // 清理函数分析信息
}
functionAnalysisInfo.clear();
} // 清空分析结果
~ControlFlowAnalysis() {
clear(); // 析构时清理所有分析信息
}
private:
void intersectOP4Dom(std::unordered_set<BasicBlock *> &dom, const std::unordered_set<BasicBlock *> &other); // 交集运算,
BasicBlock* findCommonDominator(BasicBlock *a, BasicBlock *b); // 查找两个基本块的共同支配结点
};
// 数据流分析类
// 该类为抽象类,具体的数据流分析器需要继承此类
// 因为每个数据流分析器的分析动作都不一样所以需要继承并实现analyze方法
class DataFlowAnalysis {
public:
virtual ~DataFlowAnalysis() = default;
public:
virtual void init(Module *pModule) {} ///< 分析器初始化
virtual auto analyze(Module *pModule, BasicBlock *block) -> bool { return true; } ///< 分析动作若完成则返回true;
virtual void clear() {} ///< 清空
};
// 数据流分析工具类
// 该类用于管理多个数据流分析器,提供统一的前向与后向分析接口
class DataFlowAnalysisUtils {
private:
std::vector<DataFlowAnalysis *> forwardAnalysisList; ///< 前向分析器列表
std::vector<DataFlowAnalysis *> backwardAnalysisList; ///< 后向分析器列表
public:
DataFlowAnalysisUtils() = default;
~DataFlowAnalysisUtils() {
clear(); // 析构时清理所有分析器
}
// 统一添加接口
void addAnalyzers(
std::vector<DataFlowAnalysis *> forwardList,
std::vector<DataFlowAnalysis *> backwardList = {})
{
forwardAnalysisList.insert(
forwardAnalysisList.end(),
forwardList.begin(),
forwardList.end());
backwardAnalysisList.insert(
backwardAnalysisList.end(),
backwardList.begin(),
backwardList.end());
}
// 单独添加接口
void addForwardAnalyzer(DataFlowAnalysis *analyzer) {
forwardAnalysisList.push_back(analyzer);
}
void addBackwardAnalyzer(DataFlowAnalysis *analyzer) {
backwardAnalysisList.push_back(analyzer);
}
// 设置分析器列表
void setAnalyzers(
std::vector<DataFlowAnalysis *> forwardList,
std::vector<DataFlowAnalysis *> backwardList)
{
forwardAnalysisList = std::move(forwardList);
backwardAnalysisList = std::move(backwardList);
}
// 清空列表
void clear() {
forwardAnalysisList.clear();
backwardAnalysisList.clear();
}
// 访问器
const auto& getForwardAnalyzers() const { return forwardAnalysisList; }
const auto& getBackwardAnalyzers() const { return backwardAnalysisList; }
public:
void forwardAnalyze(Module *pModule); ///< 执行前向分析
void backwardAnalyze(Module *pModule); ///< 执行后向分析
};
// 活跃变量分析类
// 提供def - use分析
// 未兼容数组变量但是考虑了维度的use信息
class ActiveVarAnalysis : public DataFlowAnalysis {
private:
std::map<BasicBlock *, std::vector<std::set<User *>>> activeTable; ///< 活跃信息表,存储每个基本块内的的活跃变量信息
public:
ActiveVarAnalysis() = default;
~ActiveVarAnalysis() override = default;
public:
static std::set<User*> getUsedSet(Instruction *inst);
static User* getDefine(Instruction *inst);
public:
void init(Module *pModule) override;
bool analyze(Module *pModule, BasicBlock *block) override;
// 外部活跃信息表访问器
const std::map<BasicBlock *, std::vector<std::set<User *>>> &getActiveTable() const;
void clear() override {
activeTable.clear(); // 清空活跃信息表
}
};
// 分析管理器 后续实现
// class AnalysisManager {
// };
} // namespace sysy

View File

@ -59,96 +59,15 @@ private:
std::unique_ptr<Module> module;
IRBuilder builder;
using ValueOrOperator = std::variant<Value*, int>;
std::vector<ValueOrOperator> BinaryExpStack; ///< 用于存储二元表达式的中缀表达式
std::vector<int> BinaryExpLenStack; ///< 用于存储该层次的二元表达式的长度
// 下面是用于后缀表达式的计算的数据结构
std::vector<ValueOrOperator> BinaryRPNStack; ///< 用于存储二元表达式的后缀表达式
std::vector<int> BinaryOpStack; ///< 用于存储二元表达式中缀表达式转换到后缀表达式的操作符栈
std::vector<Value *> BinaryValueStack; ///< 用于存储后缀表达式计算的操作数栈
// 约定操作符:
// 1: 'ADD', 2: 'SUB', 3: 'MUL', 4: 'DIV', 5: '%', 6: 'PLUS', 7: 'NEG', 8: 'NOT', 9: 'LPAREN', 10: 'RPAREN'
// 这里的操作符是为了方便后缀表达式的计算而设计
// 其中,'ADD', 'SUB', 'MUL', 'DIV', '%'
// 分别对应加法、减法、乘法、除法和取模
// 'PLUS' 和 'NEG' 分别对应一元加法和一元减法
// 'NOT' 对应逻辑非
// 'LPAREN' 和 'RPAREN' 分别对应左括号和右括号
enum BinaryOp {
ADD = 1, SUB = 2, MUL = 3, DIV = 4, MOD = 5, PLUS = 6, NEG = 7, NOT = 8, LPAREN = 9, RPAREN = 10,
};
int getOperatorPrecedence(int op) {
switch (op) {
case MUL: case DIV: case MOD: return 2;
case ADD: case SUB: return 1;
case PLUS: case NEG: case NOT: return 3;
case LPAREN: case RPAREN: return 0; // Parentheses have lowest precedence for stack logic
default: return -1; // Unknown operator
}
};
struct ExpKey {
BinaryOp op; ///< 操作符
Value *left; ///< 左操作数
Value *right; ///< 右操作数
ExpKey(BinaryOp op, Value *left, Value *right) : op(op), left(left), right(right) {}
bool operator<(const ExpKey &other) const {
if (op != other.op)
return op < other.op; ///< 比较操作符
if (left != other.left)
return left < other.left; ///< 比较左操作
return right < other.right; ///< 比较右操作数
} ///< 重载小于运算符用于比较ExpKey
};
struct UnExpKey {
BinaryOp op; ///< 一元操作符
Value *operand; ///< 操作数
UnExpKey(BinaryOp op, Value *operand) : op(op), operand(operand) {}
bool operator<(const UnExpKey &other) const {
if (op != other.op)
return op < other.op; ///< 比较操作符
return operand < other.operand; ///< 比较操作数
} ///< 重载小于运算符用于比较UnExpKey
};
struct GEPKey {
Value *basePointer;
std::vector<Value *> indices;
// 为 std::map 定义比较运算符,使得 GEPKey 可以作为键
bool operator<(const GEPKey &other) const {
if (basePointer != other.basePointer) {
return basePointer < other.basePointer;
}
// 逐个比较索引,确保顺序一致
if (indices.size() != other.indices.size()) {
return indices.size() < other.indices.size();
}
for (size_t i = 0; i < indices.size(); ++i) {
if (indices[i] != other.indices[i]) {
return indices[i] < other.indices[i];
}
}
return false; // 如果 basePointer 和所有索引都相同,则认为相等
}
};
std::map<GEPKey, Value*> availableGEPs; ///< 用于存储 GEP 的缓存
std::map<ExpKey, Value*> availableBinaryExpressions;
std::map<UnExpKey, Value*> availableUnaryExpressions;
std::map<Value*, Value*> availableLoads;
public:
SysYIRGenerator() = default;
bool HasReturnInst;
public:
Module *get() const { return module.get(); }
IRBuilder *getBuilder(){ return &builder; }
public:
std::any visitCompUnit(SysYParser::CompUnitContext *ctx) override;
std::any visitGlobalConstDecl(SysYParser::GlobalConstDeclContext *ctx) override;
@ -179,7 +98,7 @@ public:
std::any visitBlockStmt(SysYParser::BlockStmtContext* ctx) override;
// std::any visitStmt(SysYParser::StmtContext *ctx) override;
std::any visitAssignStmt(SysYParser::AssignStmtContext *ctx) override;
std::any visitExpStmt(SysYParser::ExpStmtContext *ctx) override;
// std::any visitExpStmt(SysYParser::ExpStmtContext *ctx) override;
// std::any visitBlkStmt(SysYParser::BlkStmtContext *ctx) override;
std::any visitIfStmt(SysYParser::IfStmtContext *ctx) override;
std::any visitWhileStmt(SysYParser::WhileStmtContext *ctx) override;
@ -213,29 +132,7 @@ public:
std::any visitLAndExp(SysYParser::LAndExpContext *ctx) override;
std::any visitLOrExp(SysYParser::LOrExpContext *ctx) override;
std::any visitConstExp(SysYParser::ConstExpContext *ctx) override;
bool isRightAssociative(int op);
Value* promoteType(Value* value, Type* targetType);
Value* computeExp(SysYParser::ExpContext *ctx, Type* targetType = nullptr);
Value* computeAddExp(SysYParser::AddExpContext *ctx, Type* targetType = nullptr);
void compute();
// 参数是发生 store 操作的目标地址/变量的 Value*
void invalidateExpressionsOnStore(Value* storedAddress);
// 清除因函数调用而失效的表达式缓存(保守策略)
void invalidateExpressionsOnCall();
// 在进入新的基本块时清空所有表达式缓存
void enterNewBasicBlock();
public:
// 获取GEP指令的地址
Value* getGEPAddressInst(Value* basePointer, const std::vector<Value*>& indices);
// 构建数组类型
Type* buildArrayType(Type* baseType, const std::vector<Value*>& dims);
unsigned countArrayDimensions(Type* type);
// std::any visitConstExp(SysYParser::ConstExpContext *ctx) override;
}; // class SysYIRGenerator

View File

@ -0,0 +1,37 @@
#pragma once
#include "IR.h"
#include "IRBuilder.h"
namespace sysy {
// 优化前对SysY IR的预处理也可以视作部分CFG优化
// 主要包括删除无用指令、合并基本块、删除空块等
// 这些操作可以在SysY IR生成时就完成但为了简化IR生成过程
// 这里将其放在SysY IR生成后进行预处理
// 同时兼容phi节点的处理可以再mem2reg后再次调用优化
class SysYOptPre {
private:
Module *pModule;
IRBuilder *pBuilder;
public:
SysYOptPre(Module *pMoudle, IRBuilder *pBuilder) : pModule(pMoudle), pBuilder(pBuilder) {}
void SysYOptimizateAfterIR(){
SysYDelInstAfterBr();
SysYBlockMerge();
SysYDelNoPreBLock();
SysYDelEmptyBlock();
SysYAddReturn();
}
void SysYDelInstAfterBr(); // 删除br后面的指令
void SysYDelEmptyBlock(); // 空块删除
void SysYDelNoPreBLock(); // 删除无前驱块
void SysYBlockMerge(); // 合并基本块(主要针对嵌套if while的exit块
// 也可以修改IR生成实现回填机制
void SysYAddReturn(); // 添加return指令(主要针对Void函数)
void usedelete(Instruction *instr); // use删除
};
} // namespace sysy

View File

@ -15,7 +15,6 @@ public:
public:
void printIR();
void printGlobalVariable();
void printGlobalConstant();
public:
@ -23,8 +22,6 @@ public:
static void printInst(Instruction *pInst);
static void printType(Type *type);
static void printValue(Value *value);
static void printBlock(BasicBlock *block);
static std::string getBlockName(BasicBlock *block);
static std::string getOperandName(Value *operand);
static std::string getTypeString(Type *type);
static std::string getValueName(Value *value);

View File

@ -1,33 +0,0 @@
#ifndef CALLEE_SAVED_HANDLER_H
#define CALLEE_SAVED_HANDLER_H
#include "RISCv64LLIR.h"
#include "Pass.h"
namespace sysy {
/**
* @class CalleeSavedHandler
* @brief 处理被调用者保存寄存器(Callee-Saved Registers)的Pass。
* * 这个Pass在寄存器分配之后运行。它的主要职责是
* 1. 扫描整个函数,找出所有被使用的 `s` 系列寄存器。
* 2. 在函数序言中插入 `sd` 指令来保存这些寄存器。
* 3. 在函数结尾ret指令前插入 `ld` 指令来恢复这些寄存器。
* 4. 正确计算因保存这些寄存器而需要的额外栈空间并更新StackFrameInfo。
*/
class CalleeSavedHandler : public Pass {
public:
static char ID;
CalleeSavedHandler() : Pass("callee-saved-handler", Granularity::Function, PassKind::Optimization) {}
void *getPassID() const override { return &ID; }
bool runOnFunction(Function *F, AnalysisManager& AM) override;
void runOnMachineFunction(MachineFunction* mfunc);
};
} // namespace sysy
#endif // CALLEE_SAVED_HANDLER_H

View File

@ -1,20 +0,0 @@
#ifndef ELIMINATE_FRAME_INDICES_H
#define ELIMINATE_FRAME_INDICES_H
#include "RISCv64LLIR.h"
namespace sysy {
class EliminateFrameIndicesPass {
public:
// Pass 的主入口函数
void runOnMachineFunction(MachineFunction* mfunc);
private:
// 帮助计算类型大小的辅助函数从原RegAlloc中移出
unsigned getTypeSizeInBytes(Type* type);
};
} // namespace sysy
#endif // ELIMINATE_FRAME_INDICES_H

View File

@ -1,36 +0,0 @@
#ifndef SYSY_LEGALIZE_IMMEDIATES_H
#define SYSY_LEGALIZE_IMMEDIATES_H
#include "RISCv64LLIR.h"
#include "Pass.h"
namespace sysy {
// MachineFunction 的前向声明在这里是可选的,因为 RISCv64LLIR.h 已经定义了它
// class MachineFunction;
/**
* @class LegalizeImmediatesPass
* @brief 一个用于“合法化”机器指令的Pass。
*
* 这个Pass的主要职责是遍历所有机器指令查找那些包含了超出
* 目标架构RISC-V编码范围的大立即数immediate的指令
* 并将它们展开成一个等价的、只包含合法立即数的指令序列。
*
* 它在指令选择之后、寄存器分配之前运行,确保进入后续阶段的
* 所有指令都符合硬件约束。
*/
class LegalizeImmediatesPass : public Pass {
public:
static char ID;
LegalizeImmediatesPass() : Pass("legalize-immediates", Granularity::Function, PassKind::Optimization) {}
void *getPassID() const override { return &ID; }
void runOnMachineFunction(MachineFunction* mfunc);
};
} // namespace sysy
#endif // SYSY_LEGALIZE_IMMEDIATES_H

View File

@ -1,35 +0,0 @@
#ifndef SYSY_PROLOGUE_EPILOGUE_INSERTION_H
#define SYSY_PROLOGUE_EPILOGUE_INSERTION_H
#include "RISCv64LLIR.h"
#include "Pass.h"
namespace sysy {
class MachineFunction;
/**
* @class PrologueEpilogueInsertionPass
* @brief 在函数中插入序言和尾声的机器指令。
*
* 这个Pass在所有栈帧大小计算完毕后包括局部变量、溢出槽、被调用者保存寄存器
* 在寄存器分配之后运行。它的职责是:
* 1. 根据 StackFrameInfo 中的最终栈大小,生成用于分配和释放栈帧的指令 (addi sp, sp, +/-size)。
* 2. 生成用于保存和恢复返回地址(ra)和旧帧指针(s0)的指令。
* 3. 将这些指令作为 MachineInstr 对象插入到 MachineFunction 的入口块和所有返回块中。
* 4. 这个Pass可能会生成带有大立即数的指令需要后续的 LegalizeImmediatesPass 来处理。
*/
class PrologueEpilogueInsertionPass : public Pass {
public:
static char ID;
PrologueEpilogueInsertionPass() : Pass("prologue-epilogue-insertion", Granularity::Function, PassKind::Optimization) {}
void *getPassID() const override { return &ID; }
void runOnMachineFunction(MachineFunction* mfunc);
};
} // namespace sysy
#endif // SYSY_PROLOGUE_EPILOGUE_INSERTION_H

View File

@ -1,30 +0,0 @@
#ifndef RISCV64_DIV_STRENGTH_REDUCTION_H
#define RISCV64_DIV_STRENGTH_REDUCTION_H
#include "RISCv64LLIR.h"
#include "Pass.h"
namespace sysy {
/**
* @class DivStrengthReduction
* @brief 除法强度削弱优化器
* * 将除法运算转换为乘法运算使用magic number算法
* 适用于除数为常数的情况,可以显著提高性能
*/
class DivStrengthReduction : public Pass {
public:
static char ID;
DivStrengthReduction() : Pass("div-strength-reduction", Granularity::Function, PassKind::Optimization) {}
void *getPassID() const override { return &ID; }
bool runOnFunction(Function *F, AnalysisManager& AM) override;
void runOnMachineFunction(MachineFunction* mfunc);
};
} // namespace sysy
#endif // RISCV64_DIV_STRENGTH_REDUCTION_H

View File

@ -1,30 +0,0 @@
#ifndef RISCV64_PEEPHOLE_H
#define RISCV64_PEEPHOLE_H
#include "RISCv64LLIR.h"
#include "Pass.h"
namespace sysy {
/**
* @class PeepholeOptimizer
* @brief 窥孔优化器
* * 在已分配物理寄存器的指令流上,通过一个小的滑动窗口来查找
* 并替换掉一些冗余或低效的指令模式。
*/
class PeepholeOptimizer : public Pass {
public:
static char ID;
PeepholeOptimizer() : Pass("peephole-optimizer", Granularity::Function, PassKind::Optimization) {}
void *getPassID() const override { return &ID; }
bool runOnFunction(Function *F, AnalysisManager& AM) override;
void runOnMachineFunction(MachineFunction* mfunc);
};
} // namespace sysy
#endif // RISCV64_PEEPHOLE_H

View File

@ -1,50 +0,0 @@
#ifndef POST_RA_SCHEDULER_H
#define POST_RA_SCHEDULER_H
#include "RISCv64LLIR.h"
#include "Pass.h"
namespace sysy {
/**
* @class PostRA_Scheduler
* @brief 寄存器分配后的局部指令调度器
* * 主要目标是优化寄存器分配器插入的spill/fill代码(lw/sw)
* 尝试将加载指令提前,以隐藏其访存延迟。
*/
struct MemoryAccess {
PhysicalReg base_reg;
int64_t offset;
bool valid;
MemoryAccess() : valid(false) {}
MemoryAccess(PhysicalReg base, int64_t off) : base_reg(base), offset(off), valid(true) {}
};
struct InstrRegInfo {
std::unordered_set<PhysicalReg> defined_regs;
std::unordered_set<PhysicalReg> used_regs;
bool is_load;
bool is_store;
bool is_control_flow;
MemoryAccess mem_access;
InstrRegInfo() : is_load(false), is_store(false), is_control_flow(false) {}
};
class PostRA_Scheduler : public Pass {
public:
static char ID;
PostRA_Scheduler() : Pass("post-ra-scheduler", Granularity::Function, PassKind::Optimization) {}
void *getPassID() const override { return &ID; }
bool runOnFunction(Function *F, AnalysisManager& AM) override;
void runOnMachineFunction(MachineFunction* mfunc);
};
} // namespace sysy
#endif // POST_RA_SCHEDULER_H

View File

@ -1,30 +0,0 @@
#ifndef PRE_RA_SCHEDULER_H
#define PRE_RA_SCHEDULER_H
#include "RISCv64LLIR.h"
#include "Pass.h"
namespace sysy {
/**
* @class PreRA_Scheduler
* @brief 寄存器分配前的指令调度器
* * 在虚拟寄存器上进行操作,此时调度自由度最大,
* 主要目标是隐藏指令延迟,提高流水线效率。
*/
class PreRA_Scheduler : public Pass {
public:
static char ID;
PreRA_Scheduler() : Pass("pre-ra-scheduler", Granularity::Function, PassKind::Optimization) {}
void *getPassID() const override { return &ID; }
bool runOnFunction(Function *F, AnalysisManager& AM) override;
void runOnMachineFunction(MachineFunction* mfunc);
};
} // namespace sysy
#endif // PRE_RA_SCHEDULER_H

View File

@ -1,104 +0,0 @@
#ifndef RISCV64_LINEARSCAN_H
#define RISCV64_LINEARSCAN_H
#include "RISCv64LLIR.h"
#include "RISCv64ISel.h"
#include <vector>
#include <map>
#include <set>
#include <algorithm>
namespace sysy {
// 前向声明
class MachineBasicBlock;
class MachineFunction;
class RISCv64ISel;
/**
* @brief 表示一个虚拟寄存器的活跃区间。
* 包含起始和结束指令编号。为了简化,我们不处理有“洞”的区间。
*/
struct LiveInterval {
unsigned vreg = 0;
int start = -1;
int end = -1;
bool crosses_call = false;
LiveInterval(unsigned vreg) : vreg(vreg) {}
// 用于排序,按起始点从小到大
bool operator<(const LiveInterval& other) const {
return start < other.start;
}
};
class RISCv64LinearScan {
public:
RISCv64LinearScan(MachineFunction* mfunc);
void run();
private:
// --- 核心算法流程 ---
void linearizeBlocks();
void computeLiveIntervals();
bool linearScan();
void rewriteProgram();
void applyAllocation();
void chooseRegForInterval(LiveInterval* current);
void spillAtInterval(LiveInterval* current);
// --- 辅助函数 ---
void getInstrUseDef(const MachineInstr* instr, std::set<unsigned>& use, std::set<unsigned>& def);
bool isFPVReg(unsigned vreg) const;
void collectUsedCalleeSavedRegs();
MachineFunction* MFunc;
RISCv64ISel* ISel;
// --- 线性扫描数据结构 ---
std::vector<MachineBasicBlock*> linear_order_blocks;
std::map<const MachineInstr*, int> instr_numbering;
std::map<unsigned, LiveInterval> live_intervals;
std::vector<LiveInterval*> unhandled;
std::vector<LiveInterval*> active; // 活跃且已分配物理寄存器的区间
std::set<unsigned> spilled_vregs; // 记录在本轮被决定溢出的vreg
// --- 寄存器池和分配结果 ---
std::vector<PhysicalReg> allocable_int_regs;
std::vector<PhysicalReg> allocable_fp_regs;
std::set<PhysicalReg> free_int_regs;
std::set<PhysicalReg> free_fp_regs;
std::map<unsigned, PhysicalReg> vreg_to_preg_map;
std::map<unsigned, PhysicalReg> abi_vreg_map;
const std::map<unsigned, Type*>& vreg_type_map;
};
static const std::map<RVOpcodes, std::pair<std::vector<int>, std::vector<int>>> op_info = {
{RVOpcodes::ADD, {{0}, {1, 2}}}, {RVOpcodes::SUB, {{0}, {1, 2}}}, {RVOpcodes::MUL, {{0}, {1, 2}}},
{RVOpcodes::DIV, {{0}, {1, 2}}}, {RVOpcodes::REM, {{0}, {1, 2}}}, {RVOpcodes::ADDW, {{0}, {1, 2}}},
{RVOpcodes::SUBW, {{0}, {1, 2}}}, {RVOpcodes::MULW, {{0}, {1, 2}}}, {RVOpcodes::DIVW, {{0}, {1, 2}}},
{RVOpcodes::REMW, {{0}, {1, 2}}}, {RVOpcodes::SLT, {{0}, {1, 2}}}, {RVOpcodes::SLTU, {{0}, {1, 2}}},
{RVOpcodes::ADDI, {{0}, {1}}}, {RVOpcodes::ADDIW, {{0}, {1}}}, {RVOpcodes::XORI, {{0}, {1}}},
{RVOpcodes::SLTI, {{0}, {1}}}, {RVOpcodes::SLTIU, {{0}, {1}}}, {RVOpcodes::LB, {{0}, {}}},
{RVOpcodes::LH, {{0}, {}}}, {RVOpcodes::LW, {{0}, {}}}, {RVOpcodes::LD, {{0}, {}}},
{RVOpcodes::LBU, {{0}, {}}}, {RVOpcodes::LHU, {{0}, {}}}, {RVOpcodes::LWU, {{0}, {}}},
{RVOpcodes::FLW, {{0}, {}}}, {RVOpcodes::FLD, {{0}, {}}}, {RVOpcodes::SB, {{}, {0, 1}}},
{RVOpcodes::SH, {{}, {0, 1}}}, {RVOpcodes::SW, {{}, {0, 1}}}, {RVOpcodes::SD, {{}, {0, 1}}},
{RVOpcodes::FSW, {{}, {0, 1}}}, {RVOpcodes::FSD, {{}, {0, 1}}}, {RVOpcodes::BEQ, {{}, {0, 1}}},
{RVOpcodes::BNE, {{}, {0, 1}}}, {RVOpcodes::BLT, {{}, {0, 1}}}, {RVOpcodes::BGE, {{}, {0, 1}}},
{RVOpcodes::JALR, {{0}, {1}}}, {RVOpcodes::LI, {{0}, {}}}, {RVOpcodes::LA, {{0}, {}}},
{RVOpcodes::MV, {{0}, {1}}}, {RVOpcodes::SEQZ, {{0}, {1}}}, {RVOpcodes::SNEZ, {{0}, {1}}},
{RVOpcodes::RET, {{}, {}}}, {RVOpcodes::FADD_S, {{0}, {1, 2}}}, {RVOpcodes::FSUB_S, {{0}, {1, 2}}},
{RVOpcodes::FMUL_S, {{0}, {1, 2}}}, {RVOpcodes::FDIV_S, {{0}, {1, 2}}}, {RVOpcodes::FEQ_S, {{0}, {1, 2}}},
{RVOpcodes::FLT_S, {{0}, {1, 2}}}, {RVOpcodes::FLE_S, {{0}, {1, 2}}}, {RVOpcodes::FCVT_S_W, {{0}, {1}}},
{RVOpcodes::FCVT_W_S, {{0}, {1}}}, {RVOpcodes::FMV_S, {{0}, {1}}}, {RVOpcodes::FMV_W_X, {{0}, {1}}},
{RVOpcodes::FMV_X_W, {{0}, {1}}}, {RVOpcodes::FNEG_S, {{0}, {1}}}
};
} // namespace sysy
#endif // RISCV64_LINEARSCAN_H

View File

@ -1,20 +0,0 @@
#ifndef RISCV64_PASSES_H
#define RISCV64_PASSES_H
#include "RISCv64LLIR.h"
#include "Peephole.h"
#include "PreRA_Scheduler.h"
#include "PostRA_Scheduler.h"
#include "CalleeSavedHandler.h"
#include "LegalizeImmediates.h"
#include "PrologueEpilogueInsertion.h"
#include "EliminateFrameIndices.h"
#include "Pass.h"
#include "DivStrengthReduction.h"
namespace sysy {
} // namespace sysy
#endif // RISCV64_PASSES_H

View File

@ -1,122 +0,0 @@
#ifndef RISCV64_REGALLOC_H
#define RISCV64_REGALLOC_H
#include "RISCv64LLIR.h"
#include "RISCv64ISel.h" // 包含 RISCv64ISel.h 以访问 ISel 和 Value 类型
#include <set>
#include <vector>
#include <map>
#include <stack>
extern int DEBUG;
extern int DEEPDEBUG;
extern int DEBUGLENGTH; // 用于限制调试输出的长度
extern int DEEPERDEBUG; // 用于更深层次的调试输出
namespace sysy {
class RISCv64RegAlloc {
public:
RISCv64RegAlloc(MachineFunction* mfunc);
// 模块主入口
void run();
private:
// 类型定义与Python版本对应
using VRegSet = std::set<unsigned>;
using InterferenceGraph = std::map<unsigned, VRegSet>;
using VRegStack = std::vector<unsigned>; // 使用vector模拟栈方便遍历
using MoveList = std::map<unsigned, std::set<const MachineInstr*>>;
using AliasMap = std::map<unsigned, unsigned>;
using ColorMap = std::map<unsigned, PhysicalReg>;
using VRegMoveSet = std::set<const MachineInstr*>;
// --- 核心算法流程 ---
void initialize();
void build();
void makeWorklist();
void simplify();
void coalesce();
void freeze();
void selectSpill();
void assignColors();
void rewriteProgram();
bool doAllocation();
void applyColoring();
void dumpState(const std::string &stage);
void precolorByCallingConvention();
// --- 辅助函数 ---
void getInstrUseDef(const MachineInstr* instr, VRegSet& use, VRegSet& def);
void getInstrUseDef_Liveness(const MachineInstr *instr, VRegSet &use, VRegSet &def);
void addEdge(unsigned u, unsigned v);
VRegSet adjacent(unsigned n);
VRegMoveSet nodeMoves(unsigned n);
bool moveRelated(unsigned n);
void decrementDegree(unsigned m);
void enableMoves(const VRegSet& nodes);
unsigned getAlias(unsigned n);
void addWorklist(unsigned u);
bool briggsHeuristic(unsigned u, unsigned v);
bool georgeHeuristic(unsigned u, unsigned v);
void combine(unsigned u, unsigned v);
void freezeMoves(unsigned u);
void collectUsedCalleeSavedRegs();
bool isFPVReg(unsigned vreg) const;
std::string regToString(PhysicalReg reg);
std::string regIdToString(unsigned id);
// --- 活跃性分析 ---
void analyzeLiveness();
MachineFunction* MFunc;
RISCv64ISel* ISel;
// --- 算法数据结构 ---
// 寄存器池
std::vector<PhysicalReg> allocable_int_regs;
std::vector<PhysicalReg> allocable_fp_regs;
int K_int; // 整数寄存器数量
int K_fp; // 浮点寄存器数量
// 节点集合
VRegSet precolored; // 预着色的节点 (物理寄存器)
VRegSet initial; // 初始的、所有待处理的虚拟寄存器节点
VRegSet simplifyWorklist;
VRegSet freezeWorklist;
VRegSet spillWorklist;
VRegSet spilledNodes;
VRegSet coalescedNodes;
VRegSet coloredNodes;
VRegStack selectStack;
// Move指令相关
std::set<const MachineInstr*> coalescedMoves;
std::set<const MachineInstr*> constrainedMoves;
std::set<const MachineInstr*> frozenMoves;
std::set<const MachineInstr*> worklistMoves;
std::set<const MachineInstr*> activeMoves;
// 数据结构
InterferenceGraph adjSet;
std::map<unsigned, VRegSet> adjList; // 邻接表
std::map<unsigned, int> degree;
MoveList moveList;
AliasMap alias;
ColorMap color_map;
// 活跃性分析结果
std::map<const MachineInstr*, VRegSet> live_in_map;
std::map<const MachineInstr*, VRegSet> live_out_map;
// VReg -> Value* 和 VReg -> Type* 的映射
const std::map<unsigned, Value*>& vreg_to_value_map;
const std::map<unsigned, Type*>& vreg_type_map;
};
} // namespace sysy
#endif // RISCV64_REGALLOC_H

View File

@ -1,107 +0,0 @@
#pragma once
#include "Pass.h" // 包含 Pass 框架
#include "IR.h" // 包含 IR 定义
#include <map>
#include <set>
#include <vector>
#include <algorithm>
#include <functional>
namespace sysy {
// 支配树分析结果类
class DominatorTree : public AnalysisResultBase {
public:
DominatorTree(Function* F);
// 获取指定基本块的所有支配者
const std::set<BasicBlock*>* getDominators(BasicBlock* BB) const;
// 获取指定基本块的即时支配者 (Immediate Dominator)
BasicBlock* getImmediateDominator(BasicBlock* BB) const;
// 获取指定基本块的支配边界 (Dominance Frontier)
const std::set<BasicBlock*>* getDominanceFrontier(BasicBlock* BB) const;
// 获取指定基本块在支配树中的子节点
const std::set<BasicBlock*>* getDominatorTreeChildren(BasicBlock* BB) const;
// 额外的 Getter获取所有支配者、即时支配者和支配边界的完整映射可选主要用于调试或特定场景
const std::map<BasicBlock*, std::set<BasicBlock*>>& getDominatorsMap() const { return Dominators; }
const std::map<BasicBlock*, BasicBlock*>& getIDomsMap() const { return IDoms; }
const std::map<BasicBlock*, std::set<BasicBlock*>>& getDominanceFrontiersMap() const { return DominanceFrontiers; }
// 计算所有基本块的支配者集合
void computeDominators(Function* F);
// 计算所有基本块的即时支配者(内部使用 Lengauer-Tarjan 算法)
void computeIDoms(Function* F);
// 计算所有基本块的支配边界
void computeDominanceFrontiers(Function* F);
// 计算支配树的结构(即每个节点的直接子节点)
void computeDominatorTreeChildren(Function* F);
private:
// 与该支配树关联的函数
Function* AssociatedFunction;
std::map<BasicBlock*, std::set<BasicBlock*>> Dominators; // 每个基本块的支配者集合
std::map<BasicBlock*, BasicBlock*> IDoms; // 每个基本块的即时支配者
std::map<BasicBlock*, std::set<BasicBlock*>> DominanceFrontiers; // 每个基本块的支配边界
std::map<BasicBlock*, std::set<BasicBlock*>> DominatorTreeChildren; // 支配树中每个基本块的子节点
// ==========================================================
// Lengauer-Tarjan 算法内部所需的数据结构和辅助函数
// 这些成员是私有的,以封装 LT 算法的复杂性并避免命名空间污染
// ==========================================================
// DFS 遍历相关:
std::map<BasicBlock*, int> dfnum_map; // 存储每个基本块的 DFS 编号
std::vector<BasicBlock*> vertex_vec; // 通过 DFS 编号反向查找对应的基本块指针
std::map<BasicBlock*, BasicBlock*> parent_map; // 存储 DFS 树中每个基本块的父节点
int df_counter; // DFS 计数器,也代表 DFS 遍历的总节点数 (N)
// 半支配者 (Semi-dominator) 相关:
std::map<BasicBlock*, BasicBlock*> sdom_map; // 存储每个基本块的半支配者
std::map<BasicBlock*, BasicBlock*> idom_map; // 存储每个基本块的即时支配者 (IDom)
std::map<BasicBlock*, std::vector<BasicBlock*>> bucket_map; // 桶结构,用于存储具有相同半支配者的节点,以延迟 IDom 计算
// 并查集 (Union-Find) 相关(用于 evalAndCompress 函数):
std::map<BasicBlock*, BasicBlock*> ancestor_map; // 并查集中的父节点(用于路径压缩)
std::map<BasicBlock*, BasicBlock*> label_map; // 并查集中,每个集合的代表节点(或其路径上 sdom 最小的节点)
// ==========================================================
// 辅助计算函数 (私有)
// ==========================================================
// 计算基本块的逆后序遍历 (Reverse Post Order, RPO) 顺序
// RPO 用于优化支配者计算和 LT 算法的效率
std::vector<BasicBlock*> computeReversePostOrder(Function* F);
// Lengauer-Tarjan 算法特定的辅助 DFS 函数
// 用于初始化 dfnum_map, vertex_vec, parent_map
void dfs_lt_helper(BasicBlock* u);
// 结合了并查集的 Find 操作和 LT 算法的 Eval 操作
// 用于在路径压缩时更新 label找到路径上 sdom 最小的节点
BasicBlock* evalAndCompress_lt_helper(BasicBlock* i);
// 并查集的 Link 操作
// 将 v_child 挂载到 u_parent 的并查集树下
void link_lt_helper(BasicBlock* u_parent, BasicBlock* v_child);
};
// 支配树分析遍
class DominatorTreeAnalysisPass : public AnalysisPass {
public:
// 唯一的 Pass ID
static void *ID;
DominatorTreeAnalysisPass() : AnalysisPass("DominatorTreeAnalysis", Pass::Granularity::Function) {}
// 实现 getPassID
void* getPassID() const override { return &ID; }
bool runOnFunction(Function* F, AnalysisManager &AM) override;
std::unique_ptr<AnalysisResultBase> getResult() override;
private:
std::unique_ptr<DominatorTree> CurrentDominatorTree;
};
} // namespace sysy

View File

@ -1,72 +0,0 @@
#pragma once
#include "IR.h" // 包含 IR 定义
#include "Pass.h" // 包含 Pass 框架
#include <algorithm> // for std::set_union, std::set_difference
#include <map>
#include <set>
#include <vector>
namespace sysy {
// 前向声明
class Function;
class BasicBlock;
class Value;
class Instruction;
// 活跃变量分析结果类
// 它将包含 LiveIn 和 LiveOut 集合
class LivenessAnalysisResult : public AnalysisResultBase {
public:
LivenessAnalysisResult(Function *F) : AssociatedFunction(F) {}
// 获取给定基本块的 LiveIn 集合
const std::set<Value *> *getLiveIn(BasicBlock *BB) const;
// 获取给定基本块的 LiveOut 集合
const std::set<Value *> *getLiveOut(BasicBlock *BB) const;
// 暴露内部数据结构,如果需要更直接的访问
const std::map<BasicBlock *, std::set<Value *>> &getLiveInSets() const { return liveInSets; }
const std::map<BasicBlock *, std::set<Value *>> &getLiveOutSets() const { return liveOutSets; }
// 核心计算方法,由 LivenessAnalysisPass 调用
void computeLiveness(Function *F);
private:
Function *AssociatedFunction; // 这个活跃变量分析是为哪个函数计算的
std::map<BasicBlock *, std::set<Value *>> liveInSets;
std::map<BasicBlock *, std::set<Value *>> liveOutSets;
// 辅助函数:计算基本块的 Def 和 Use 集合
// Def: 块内定义,且定义在所有使用之前的值
// Use: 块内使用,且使用在所有定义之前的值
void computeDefUse(BasicBlock *BB, std::set<Value *> &def, std::set<Value *> &use);
};
// 活跃变量分析遍
class LivenessAnalysisPass : public AnalysisPass {
public:
// 唯一的 Pass ID
static void *ID; // LLVM 风格的唯一 ID
LivenessAnalysisPass() : AnalysisPass("LivenessAnalysis", Pass::Granularity::Function) {}
// 实现 getPassID
void *getPassID() const override { return &ID; }
// 运行分析并返回结果。现在接受 AnalysisManager& AM 参数
bool runOnFunction(Function *F, AnalysisManager &AM) override;
// 获取分析结果的指针。
// 注意AnalysisManager 将会调用此方法来获取结果并进行缓存。
std::unique_ptr<AnalysisResultBase> getResult() override;
private:
// 存储当前分析计算出的 LivenessAnalysisResult 实例
// runOnFunction 每次调用都会创建新的 LivenessAnalysisResult 对象
std::unique_ptr<LivenessAnalysisResult> CurrentLivenessResult;
};
} // namespace sysy

View File

@ -1,20 +0,0 @@
#pragma once
#include "IR.h"
#include "Pass.h"
#include <queue>
#include <set>
namespace sysy {
class BuildCFG : public OptimizationPass {
public:
static void *ID;
BuildCFG() : OptimizationPass("BuildCFG", Granularity::Function) {}
bool runOnFunction(Function *F, AnalysisManager &AM) override;
void getAnalysisUsage(std::set<void *> &analysisDependencies, std::set<void *> &analysisInvalidations) const override;
void *getPassID() const override { return &ID; }
};
} // namespace sysy

View File

@ -1,63 +0,0 @@
#pragma once
#include "Pass.h"
#include "IR.h"
#include "SysYIROptUtils.h"
#include "Dom.h"
#include <unordered_set>
#include <queue>
namespace sysy {
// 前向声明分析结果类,确保在需要时可以引用
// class DominatorTreeAnalysisResult; // Pass.h 中已包含,这里不再需要
class SideEffectInfoAnalysisResult; // 假设有副作用分析结果类
// DCEContext 类用于封装DCE的内部逻辑和状态
// 这样可以避免静态变量在多线程或多次运行时的冲突,并保持代码的模块化
class DCEContext {
public:
// 运行DCE的主要方法
// func: 当前要优化的函数
// tp: 分析管理器,用于获取其他分析结果(如果需要)
void run(Function* func, AnalysisManager* AM, bool &changed);
private:
// 存储活跃指令的集合
std::unordered_set<Instruction*> alive_insts;
// 判断指令是否是“天然活跃”的(即总是保留的)
// inst: 要检查的指令
// 返回值: 如果指令是天然活跃的则为true否则为false
bool isAlive(Instruction* inst);
// 递归地将活跃指令及其依赖加入到 alive_insts 集合中
// inst: 要标记为活跃的指令
void addAlive(Instruction* inst);
};
// DCE 优化遍类,继承自 OptimizationPass
class DCE : public OptimizationPass {
public:
// 构造函数
DCE() : OptimizationPass("DCE", Granularity::Function) {}
// 静态成员作为该遍的唯一ID
static void *ID;
// 运行在函数上的优化逻辑
// F: 当前要优化的函数
// AM: 分析管理器,用于获取或使分析结果失效
// 返回值: 如果IR被修改则为true否则为false
bool runOnFunction(Function *F, AnalysisManager& AM) override;
// 声明该遍的分析依赖和失效信息
// analysisDependencies: 该遍运行前需要哪些分析结果
// analysisInvalidations: 该遍运行后会使哪些分析结果失效
void getAnalysisUsage(std::set<void *> &analysisDependencies, std::set<void *> &analysisInvalidations) const override;
// Pass 基类中的纯虚函数,必须实现
void *getPassID() const override { return &ID; }
};
} // namespace sysy

View File

@ -1,24 +0,0 @@
#pragma once
#include "../Pass.h"
namespace sysy {
class LargeArrayToGlobalPass : public OptimizationPass {
public:
static void *ID;
LargeArrayToGlobalPass() : OptimizationPass("LargeArrayToGlobal", Granularity::Module) {}
bool runOnModule(Module *M, AnalysisManager &AM) override;
void *getPassID() const override {
return &ID;
}
private:
unsigned calculateTypeSize(Type *type);
void convertAllocaToGlobal(AllocaInst *alloca, Function *F, Module *M);
std::string generateUniqueGlobalName(AllocaInst *alloca, Function *F);
};
} // namespace sysy

View File

@ -1,114 +0,0 @@
#pragma once
#include "Pass.h" // 包含Pass的基类定义
#include "IR.h" // 包含IR相关的定义如Instruction, Function, BasicBlock, AllocaInst, LoadInst, StoreInst, PhiInst等
#include "Dom.h" // 假设支配树分析的头文件,提供 DominatorTreeAnalysisResult
#include <vector>
#include <unordered_map>
#include <unordered_set>
#include <queue>
#include <stack> // 用于变量重命名阶段的SSA值栈
namespace sysy {
// 前向声明分析结果类,确保在需要时可以引用
class DominatorTree;
// Mem2RegContext 类,封装 mem2reg 遍的核心逻辑和状态
// 这样可以避免静态变量在多线程或多次运行时的冲突,并保持代码的模块化
class Mem2RegContext {
public:
Mem2RegContext(IRBuilder *builder) : builder(builder) {}
// 运行 mem2reg 优化的主要方法
// func: 当前要优化的函数
// tp: 分析管理器,用于获取支配树等分析结果
void run(Function* func, AnalysisManager* tp);
private:
IRBuilder *builder; // IR 构建器,用于插入指令
// 存储所有需要被提升的 AllocaInst
std::vector<AllocaInst*> promotableAllocas;
// 存储每个 AllocaInst 对应的 Phi 指令列表
// 键是 AllocaInst值是该 AllocaInst 在各个基本块中插入的 Phi 指令的列表
// (实际上,一个 AllocaInst 在一个基本块中只会有一个 Phi)
std::unordered_map<AllocaInst*, std::unordered_map<BasicBlock*, PhiInst*>> allocaToPhiMap;
// 存储每个 AllocaInst 对应的当前活跃 SSA 值栈
// 用于在变量重命名阶段追踪每个 AllocaInst 在不同控制流路径上的最新值
std::unordered_map<AllocaInst*, std::stack<Value*>> allocaToValueStackMap;
// 辅助映射,存储每个 AllocaInst 的所有 store 指令
std::unordered_map<AllocaInst*, std::unordered_set<StoreInst*>> allocaToStoresMap;
// 辅助映射,存储每个 AllocaInst 对应的定义基本块(包含 store 指令的块)
std::unordered_map<AllocaInst*, std::unordered_set<BasicBlock*>> allocaToDefBlocksMap;
// 支配树分析结果,用于 Phi 插入和变量重命名
DominatorTree* dt;
// --------------------------------------------------------------------
// 阶段1: 识别可提升的 AllocaInst
// --------------------------------------------------------------------
// 判断一个 AllocaInst 是否可以被提升到寄存器
// alloca: 要检查的 AllocaInst
// 返回值: 如果可以提升,则为 true否则为 false
bool isPromotableAlloca(AllocaInst* alloca);
// 收集所有对给定 AllocaInst 进行存储的 StoreInst
// alloca: 目标 AllocaInst
void collectStores(AllocaInst* alloca);
// --------------------------------------------------------------------
// 阶段2: 插入 Phi 指令 (Phi Insertion)
// --------------------------------------------------------------------
// 为给定的 AllocaInst 插入必要的 Phi 指令
// alloca: 目标 AllocaInst
// defBlocks: 包含对该 AllocaInst 进行 store 操作的基本块集合
void insertPhis(AllocaInst* alloca, const std::unordered_set<BasicBlock*>& defBlocks);
// --------------------------------------------------------------------
// 阶段3: 变量重命名 (Variable Renaming)
// --------------------------------------------------------------------
// 对支配树进行深度优先遍历,重命名变量并替换 load/store 指令
void renameVariables(BasicBlock* currentBB);
// --------------------------------------------------------------------
// 阶段4: 清理
// --------------------------------------------------------------------
// 删除所有原始的 AllocaInst、LoadInst 和 StoreInst
void cleanup();
};
// Mem2Reg 优化遍类,继承自 OptimizationPass
// 粒度为 Function表示它在每个函数上独立运行
class Mem2Reg : public OptimizationPass {
private:
IRBuilder *builder;
public:
// 构造函数
Mem2Reg(IRBuilder *builder) : OptimizationPass("Mem2Reg", Granularity::Function), builder(builder) {}
// 静态成员作为该遍的唯一ID
static void *ID;
// 运行在函数上的优化逻辑
// F: 当前要优化的函数
// AM: 分析管理器,用于获取支配树等分析结果,或使分析结果失效
// 返回值: 如果IR被修改则为true否则为false
bool runOnFunction(Function *F, AnalysisManager& AM) override;
// 声明该遍的分析依赖和失效信息
// analysisDependencies: 该遍运行前需要哪些分析结果
// analysisInvalidations: 该遍运行后会使哪些分析结果失效
void getAnalysisUsage(std::set<void *> &analysisDependencies, std::set<void *> &analysisInvalidations) const override;
void *getPassID() const override { return &ID; }
};
} // namespace sysy

View File

@ -1,59 +0,0 @@
#pragma once
#include "IR.h"
#include "IRBuilder.h" // 你的 IR Builder
#include "Liveness.h"
#include "Dom.h"
#include "Pass.h" // 你的 Pass 框架基类
#include <iostream> // 调试用
#include <map> // 用于 Value 到 AllocaInst 的映射
#include <set> // 可能用于其他辅助集合
#include <string>
#include <vector>
namespace sysy {
class Reg2MemContext {
public:
Reg2MemContext(IRBuilder *b) : builder(b) {}
// 运行 Reg2Mem 优化
void run(Function *func);
private:
IRBuilder *builder; // IR 构建器
// 存储 SSA Value 到对应的 AllocaInst 的映射
// 只有那些需要被"溢出"到内存的 SSA 值才会被记录在这里
std::map<Value *, AllocaInst *> valueToAllocaMap;
// 辅助函数:
// 1. 识别并为 SSA Value 分配 AllocaInst
void allocateMemoryForSSAValues(Function *func);
// 2. 将 SSA 值的使用替换为 Load/Store
void insertLoadsAndStores(Function *func);
// 3. 处理 Phi 指令,将其转换为 Load/Store
void rewritePhis(Function *func);
// 4. 清理 (例如,可能删除不再需要的 Phi 指令)
void cleanup(Function *func);
// 判断一个 Value 是否是 AllocaInst 可以为其分配内存的目标
// 通常指非指针类型的Instruction结果和Argument
bool isPromotableToMemory(Value *val);
};
class Reg2Mem : public OptimizationPass {
private:
IRBuilder *builder; ///< IR构建器用于插入指令
public:
static void *ID; ///< Pass的唯一标识符
Reg2Mem(IRBuilder* builder) : OptimizationPass("Reg2Mem", Pass::Granularity::Function), builder(builder) {}
bool runOnFunction(Function *F, AnalysisManager &AM) override;
void getAnalysisUsage(std::set<void *> &analysisDependencies, std::set<void *> &analysisInvalidations) const override;
void *getPassID() const override { return &ID; } ///< 获取 Pass ID
};
} // namespace sysy

View File

@ -1,139 +0,0 @@
#pragma once
#include "IR.h"
#include "Pass.h"
#include "SysYIROptUtils.h"
#include <cassert>
#include <iostream>
#include <map>
#include <queue>
#include <set>
#include <unordered_set>
#include <vector>
#include <variant>
#include <functional>
namespace sysy {
// 定义三值格 (Three-valued Lattice) 的状态
enum class LatticeVal {
Top, // (未知 / 未初始化)
Constant, // c (常量)
Bottom // ⊥ (不确定 / 变化 / 未定义)
};
// 新增枚举来区分常量的实际类型
enum class ValueType {
Integer,
Float,
Unknown // 用于 Top 和 Bottom 状态
};
// 用于表示 SSA 值的具体状态(包含格值和常量值)
struct SSAPValue {
LatticeVal state;
std::variant<int, float> constantVal; // 使用 std::variant 存储 int 或 float
ValueType constant_type; // 记录常量是整数还是浮点数
// 默认构造函数,初始化为 Top
SSAPValue() : state(LatticeVal::Top), constantVal(0), constant_type(ValueType::Unknown) {}
// 构造函数,用于创建 Bottom 状态
SSAPValue(LatticeVal s) : state(s), constantVal(0), constant_type(ValueType::Unknown) {
assert((s == LatticeVal::Top || s == LatticeVal::Bottom) && "SSAPValue(LatticeVal) only for Top/Bottom");
}
// 构造函数,用于创建 int Constant 状态
SSAPValue(int c) : state(LatticeVal::Constant), constantVal(c), constant_type(ValueType::Integer) {}
// 构造函数,用于创建 float Constant 状态
SSAPValue(float c) : state(LatticeVal::Constant), constantVal(c), constant_type(ValueType::Float) {}
// 比较操作符,用于判断状态是否改变
bool operator==(const SSAPValue &other) const {
if (state != other.state)
return false;
if (state == LatticeVal::Constant) {
if (constant_type != other.constant_type) return false; // 类型必须匹配
return constantVal == other.constantVal; // std::variant 会比较内部值
}
return true; // Top == Top, Bottom == Bottom
}
bool operator!=(const SSAPValue &other) const { return !(*this == other); }
};
// SCCP 上下文类,持有每个函数运行时的状态
class SCCPContext {
private:
IRBuilder *builder; // IR 构建器,用于插入指令和创建常量
// 工作列表
// 存储需要重新评估的指令
std::queue<Instruction *> instWorkList;
// 存储需要重新评估的控制流边 (pair: from_block, to_block)
std::queue<std::pair<BasicBlock *, BasicBlock *>> edgeWorkList;
// 格值映射SSA Value 到其当前状态
std::map<Value *, SSAPValue> valueState;
// 可执行基本块集合
std::unordered_set<BasicBlock *> executableBlocks;
// 追踪已访问的CFG边防止重复添加使用 SysYIROptUtils::PairHash
std::unordered_set<std::pair<BasicBlock*, BasicBlock*>, SysYIROptUtils::PairHash> visitedCFGEdges;
// 辅助函数:格操作 Meet
SSAPValue Meet(const SSAPValue &a, const SSAPValue &b);
// 辅助函数:获取值的当前状态,如果不存在则默认为 Top
SSAPValue GetValueState(Value *v);
// 辅助函数:更新值的状态,如果状态改变,将所有用户加入指令工作列表
void UpdateState(Value *v, SSAPValue newState);
// 辅助函数:将边加入边工作列表,并更新可执行块
void AddEdgeToWorkList(BasicBlock *fromBB, BasicBlock *toBB);
// 辅助函数:标记一个块为可执行
void MarkBlockExecutable(BasicBlock* block);
// 辅助函数:对二元操作进行常量折叠
SSAPValue ComputeConstant(BinaryInst *binaryinst, SSAPValue lhsVal, SSAPValue rhsVal);
// 辅助函数:对一元操作进行常量折叠
SSAPValue ComputeConstant(UnaryInst *unaryInst, SSAPValue operandVal);
// 主要优化阶段
// 阶段1: 常量传播与折叠
bool PropagateConstants(Function *func);
// 阶段2: 控制流简化
bool SimplifyControlFlow(Function *func);
// 辅助函数:处理单条指令
void ProcessInstruction(Instruction *inst);
// 辅助函数:处理单条控制流边
void ProcessEdge(const std::pair<BasicBlock *, BasicBlock *> &edge);
// 控制流简化辅助函数
// 查找所有可达的基本块 (基于常量条件)
std::unordered_set<BasicBlock *> FindReachableBlocks(Function *func);
// 移除死块
void RemoveDeadBlock(BasicBlock *bb, Function *func);
// 简化分支(将条件分支替换为无条件分支)
void SimplifyBranch(CondBrInst*brInst, bool condVal); // 保持 BranchInst
// 更新前驱块的终结指令(当一个后继块被移除时)
void UpdateTerminator(BasicBlock *predBB, BasicBlock *removedSucc);
// 移除 Phi 节点的入边(当其前驱块被移除时)
void RemovePhiIncoming(BasicBlock *phiParentBB, BasicBlock *removedPred);
public:
SCCPContext(IRBuilder *builder) : builder(builder) {}
// 运行 SCCP 优化
void run(Function *func, AnalysisManager &AM);
};
// SCCP 优化遍类,继承自 OptimizationPass
class SCCP : public OptimizationPass {
private:
IRBuilder *builder; // IR 构建器,作为 Pass 的成员,传入 Context
public:
SCCP(IRBuilder *builder) : OptimizationPass("SCCP", Granularity::Function), builder(builder) {}
static void *ID;
bool runOnFunction(Function *F, AnalysisManager &AM) override;
void getAnalysisUsage(std::set<void *> &analysisDependencies, std::set<void *> &analysisInvalidations) const override;
void *getPassID() const override { return &ID; }
};
} // namespace sysy

Some files were not shown because too many files have changed in this diff Show More