Compare commits
1 Commits
deploy-202
...
constPropa
| Author | SHA1 | Date | |
|---|---|---|---|
| a5318a2c5c |
220
Pass_ID_List.md
220
Pass_ID_List.md
@@ -14,220 +14,10 @@
|
||||
|
||||
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:
|
||||
|
||||
在 `Mem2Reg` 优化遍中,当 `load` 和 `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` 的 `Use` 对象(通过 `User::operands` 列表)。
|
||||
* 一个 `Value` 持有所有使用它的 `User` 的 `Use` 对象(通过 `Value::uses` 列表)。
|
||||
|
||||
原始问题是:当一个 `LoadInst` 或 `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` 的 `uses` 列表中就会留下指向已删除 `LoadInst` / `StoreInst` 的 `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
|
||||
|
||||
在 `Mem2RegContext::renameVariables` 函数中,`load` 和 `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` 的 `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` 的 `uses` 列表中移除。
|
||||
|
||||
2. **处理 `StoreInst`:**
|
||||
当找到一个指向可提升 `AllocaInst` 的 `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` 或 `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` 和 `Value` 的两种 `Use` 关系:
|
||||
|
||||
1. **清理指令作为 `Value` 时的 `uses` 列表 (由 `replaceAllUsesWith` 完成):**
|
||||
在 `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` 的 `uses` 列表,并将这些 `Use` 对象从 `LoadInst` 的 `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` 的 `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` 或 `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` 从 `AllocaInst` 的 `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)时,其操作数**的 `use` 关系。
|
||||
|
||||
这就确保了当 `Mem2Reg` 遍历并删除 `load` 和 `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` 的 `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` 的 `load` 和 `store` 指令都被移除后,`AllocaInst` 的 `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
|
||||
|
||||
@@ -235,12 +25,4 @@ Branch 和 Return 指令: 这些是终结符指令,不产生一个可用于其
|
||||
|
||||
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
|
||||
需要保证符号表能正确维护域中的局部变量
|
||||
2732
doc/Doxyfile
Normal file
2732
doc/Doxyfile
Normal file
File diff suppressed because it is too large
Load Diff
BIN
doc/The Definitive ANTLR 4 Reference.pdf
Normal file
BIN
doc/The Definitive ANTLR 4 Reference.pdf
Normal file
Binary file not shown.
BIN
doc/arm/ARM_calling_convention.pdf
Executable file
BIN
doc/arm/ARM_calling_convention.pdf
Executable file
Binary file not shown.
BIN
doc/arm/DDI0403E_e_armv7m_arm.pdf
Executable file
BIN
doc/arm/DDI0403E_e_armv7m_arm.pdf
Executable file
Binary file not shown.
BIN
doc/arm/DDI0406C_d_armv7ar_arm.pdf
Executable file
BIN
doc/arm/DDI0406C_d_armv7ar_arm.pdf
Executable file
Binary file not shown.
27078
doc/arm/DEN0013D_cortex_a_series_PG.pdf
Executable file
27078
doc/arm/DEN0013D_cortex_a_series_PG.pdf
Executable file
File diff suppressed because one or more lines are too long
BIN
doc/arm/GNU_Assembler_Directives.pdf
Executable file
BIN
doc/arm/GNU_Assembler_Directives.pdf
Executable file
Binary file not shown.
BIN
doc/arm/ISA_AArch32_xml_A_profile-2023-03.pdf
Executable file
BIN
doc/arm/ISA_AArch32_xml_A_profile-2023-03.pdf
Executable file
Binary file not shown.
BIN
doc/arm/Raspberry.pdf
Executable file
BIN
doc/arm/Raspberry.pdf
Executable file
Binary file not shown.
BIN
doc/arm/cheatsheetv1.3-1920x1080.png
Executable file
BIN
doc/arm/cheatsheetv1.3-1920x1080.png
Executable file
Binary file not shown.
|
After Width: | Height: | Size: 346 KiB |
169
doc/markdowns/IR.md
Normal file
169
doc/markdowns/IR.md
Normal 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` 或 `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 *> ¶mTypes)`,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` 和 `fScalar`:分别存储整数和浮点数常量的值。
|
||||
- 静态方法:`get(int value)` 和 `get(float value)`,用于获取常量值。
|
||||
- `getInt()` 和 `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` 和 `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()` 和 `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` 和 `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
156
doc/markdowns/IRbuilder.md
Normal file
@@ -0,0 +1,156 @@
|
||||
`IRBuilder.h` 文件定义了一个 `IRBuilder` 类,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` 和 `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` 和 `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
121
doc/markdowns/IRcpp.md
Normal 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 *> ¶mTypes)`: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 *> ¶mTypes)`: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)` 和 `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()` 和 `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` 和 `FunctionType`,用于表示 IR 中的类型。
|
||||
- **值**:实现了 `Value`、`ConstantValue` 和 `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` 和 `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
20878
doc/n1124.pdf
Normal file
File diff suppressed because one or more lines are too long
BIN
doc/riscv/RISC-V-Reader-Chinese-v1.pdf
Executable file
BIN
doc/riscv/RISC-V-Reader-Chinese-v1.pdf
Executable file
Binary file not shown.
BIN
doc/riscv/riscv-spec-20191213.pdf
Executable file
BIN
doc/riscv/riscv-spec-20191213.pdf
Executable file
Binary file not shown.
BIN
doc/slides/exp01-antlr.pdf
Executable file
BIN
doc/slides/exp01-antlr.pdf
Executable file
Binary file not shown.
BIN
doc/slides/exp02-ir.pdf
Executable file
BIN
doc/slides/exp02-ir.pdf
Executable file
Binary file not shown.
BIN
doc/slides/lab03-code generation.pdf
Executable file
BIN
doc/slides/lab03-code generation.pdf
Executable file
Binary file not shown.
BIN
doc/slides/lab04-scalar optimization.pdf
Executable file
BIN
doc/slides/lab04-scalar optimization.pdf
Executable file
Binary file not shown.
BIN
doc/slides/lab05-register allocation.pdf
Normal file
BIN
doc/slides/lab05-register allocation.pdf
Normal file
Binary file not shown.
BIN
doc/sysy-2022-runtime.pdf
Normal file
BIN
doc/sysy-2022-runtime.pdf
Normal file
Binary file not shown.
BIN
doc/sysy-2022-spec.pdf
Normal file
BIN
doc/sysy-2022-spec.pdf
Normal file
Binary file not shown.
@@ -60,7 +60,11 @@ display_file_content() {
|
||||
# 清理临时文件的函数
|
||||
clean_tmp() {
|
||||
echo "正在清理临时目录: ${TMP_DIR}"
|
||||
rm -rf "${TMP_DIR}"/*
|
||||
rm -rf "${TMP_DIR}"/*.s \
|
||||
"${TMP_DIR}"/*_sysyc_riscv64 \
|
||||
"${TMP_DIR}"/*_sysyc_riscv64.actual_out \
|
||||
"${TMP_DIR}"/*_sysyc_riscv64.expected_stdout \
|
||||
"${TMP_DIR}"/*_sysyc_riscv64.o
|
||||
echo "清理完成。"
|
||||
}
|
||||
|
||||
|
||||
@@ -21,7 +21,6 @@ QEMU_RISCV64="qemu-riscv64"
|
||||
# --- 初始化变量 ---
|
||||
EXECUTE_MODE=false
|
||||
CLEAN_MODE=false
|
||||
OPTIMIZE_FLAG="" # 用于存储 -O1 标志
|
||||
SYSYC_TIMEOUT=10 # sysyc 编译超时 (秒)
|
||||
GCC_TIMEOUT=10 # gcc 编译超时 (秒)
|
||||
EXEC_TIMEOUT=5 # qemu 自动化执行超时 (秒)
|
||||
@@ -40,7 +39,6 @@ show_help() {
|
||||
echo "选项:"
|
||||
echo " -e, --executable 编译为可执行文件并运行测试 (必须)。"
|
||||
echo " -c, --clean 清理 tmp 临时目录下的所有文件。"
|
||||
echo " -O1 启用 sysyc 的 -O1 优化。"
|
||||
echo " -sct N 设置 sysyc 编译超时为 N 秒 (默认: 10)。"
|
||||
echo " -gct N 设置 gcc 交叉编译超时为 N 秒 (默认: 10)。"
|
||||
echo " -et N 设置 qemu 自动化执行超时为 N 秒 (默认: 5)。"
|
||||
@@ -70,7 +68,7 @@ display_file_content() {
|
||||
fi
|
||||
}
|
||||
|
||||
# --- 参数解析 ---
|
||||
# --- 本次修改点: 整个参数解析逻辑被重写 ---
|
||||
# 使用标准的 while 循环来健壮地处理任意顺序的参数
|
||||
while [[ "$#" -gt 0 ]]; do
|
||||
case "$1" in
|
||||
@@ -82,10 +80,6 @@ while [[ "$#" -gt 0 ]]; do
|
||||
CLEAN_MODE=true
|
||||
shift # 消耗选项
|
||||
;;
|
||||
-O1)
|
||||
OPTIMIZE_FLAG="-O1"
|
||||
shift # 消耗选项
|
||||
;;
|
||||
-sct)
|
||||
if [[ -n "$2" && "$2" =~ ^[0-9]+$ ]]; then SYSYC_TIMEOUT="$2"; shift 2; else echo "错误: -sct 需要一个正整数参数。" >&2; exit 1; fi
|
||||
;;
|
||||
@@ -150,7 +144,6 @@ mkdir -p "${TMP_DIR}"
|
||||
TOTAL_CASES=${#SY_FILES[@]}
|
||||
|
||||
echo "SysY 单例测试运行器启动..."
|
||||
if [ -n "$OPTIMIZE_FLAG" ]; then echo "优化等级: ${OPTIMIZE_FLAG}"; fi
|
||||
echo "超时设置: sysyc=${SYSYC_TIMEOUT}s, gcc=${GCC_TIMEOUT}s, qemu=${EXEC_TIMEOUT}s"
|
||||
echo "失败输出最大行数: ${MAX_OUTPUT_LINES}"
|
||||
echo ""
|
||||
@@ -171,21 +164,9 @@ for sy_file in "${SY_FILES[@]}"; do
|
||||
echo "======================================================================"
|
||||
echo "正在处理: ${sy_file}"
|
||||
|
||||
# --- 本次修改点: 拷贝源文件到 tmp 目录 ---
|
||||
echo " 拷贝源文件到 ${TMP_DIR}..."
|
||||
cp "${sy_file}" "${TMP_DIR}/$(basename "${sy_file}")"
|
||||
if [ -f "${input_file}" ]; then
|
||||
cp "${input_file}" "${TMP_DIR}/$(basename "${input_file}")"
|
||||
fi
|
||||
if [ -f "${output_reference_file}" ]; then
|
||||
cp "${output_reference_file}" "${TMP_DIR}/$(basename "${output_reference_file}")"
|
||||
fi
|
||||
|
||||
# 步骤 1: sysyc 编译
|
||||
echo " 使用 sysyc 编译 (超时 ${SYSYC_TIMEOUT}s)..."
|
||||
timeout -s KILL ${SYSYC_TIMEOUT} "${SYSYC}" -S "${sy_file}" ${OPTIMIZE_FLAG} -o "${assembly_file}"
|
||||
timeout -s KILL ${SYSYC_TIMEOUT} "${SYSYC}" -s ir "${sy_file}" ${OPTIMIZE_FLAG} > "${ir_file}"
|
||||
# timeout -s KILL ${SYSYC_TIMEOUT} "${SYSYC}" -s asmd "${sy_file}" > "${assembly_debug_file}" 2>&1
|
||||
timeout -s KILL ${SYSYC_TIMEOUT} "${SYSYC}" -s ir "${sy_file}" > "${ir_file}"
|
||||
SYSYC_STATUS=$?
|
||||
if [ $SYSYC_STATUS -eq 124 ]; then
|
||||
echo -e "\e[31m错误: SysY 编译 ${sy_file} IR超时\e[0m"
|
||||
@@ -194,10 +175,12 @@ for sy_file in "${SY_FILES[@]}"; do
|
||||
echo -e "\e[31m错误: SysY 编译 ${sy_file} IR失败,退出码: ${SYSYC_STATUS}\e[0m"
|
||||
is_passed=0
|
||||
fi
|
||||
timeout -s KILL ${SYSYC_TIMEOUT} "${SYSYC}" -S "${sy_file}" -o "${assembly_file}"
|
||||
if [ $? -ne 0 ]; then
|
||||
echo -e "\e[31m错误: SysY 编译失败或超时。\e[0m"
|
||||
is_passed=0
|
||||
fi
|
||||
# timeout -s KILL ${SYSYC_TIMEOUT} "${SYSYC}" -s asmd "${sy_file}" > "${assembly_debug_file}" 2>&1
|
||||
|
||||
# 步骤 2: GCC 编译
|
||||
if [ "$is_passed" -eq 1 ]; then
|
||||
|
||||
@@ -16,8 +16,8 @@ SYSYC="${BUILD_BIN_DIR}/sysyc"
|
||||
GCC_RISCV64="riscv64-linux-gnu-gcc"
|
||||
QEMU_RISCV64="qemu-riscv64"
|
||||
|
||||
# --- 新增功能: 初始化变量 ---
|
||||
EXECUTE_MODE=false
|
||||
OPTIMIZE_FLAG="" # 用于存储 -O1 标志
|
||||
SYSYC_TIMEOUT=10 # sysyc 编译超时 (秒)
|
||||
GCC_TIMEOUT=10 # gcc 编译超时 (秒)
|
||||
EXEC_TIMEOUT=5 # qemu 执行超时 (秒)
|
||||
@@ -35,7 +35,6 @@ show_help() {
|
||||
echo "选项:"
|
||||
echo " -e, --executable 编译为可执行文件并运行测试。"
|
||||
echo " -c, --clean 清理 'tmp' 目录下的所有生成文件。"
|
||||
echo " -O1 启用 sysyc 的 -O1 优化。"
|
||||
echo " -set [f|h|p|all]... 指定要运行的测试集 (functional, h_functional, performance)。可多选,默认为 all。"
|
||||
echo " -sct N 设置 sysyc 编译超时为 N 秒 (默认: 10)。"
|
||||
echo " -gct N 设置 gcc 交叉编译超时为 N 秒 (默认: 10)。"
|
||||
@@ -86,12 +85,9 @@ while [[ "$#" -gt 0 ]]; do
|
||||
clean_tmp
|
||||
exit 0
|
||||
;;
|
||||
-O1)
|
||||
OPTIMIZE_FLAG="-O1"
|
||||
shift
|
||||
;;
|
||||
-set)
|
||||
shift # 移过 '-set'
|
||||
# 消耗所有后续参数直到遇到下一个选项
|
||||
while [[ "$#" -gt 0 && ! "$1" =~ ^- ]]; do
|
||||
TEST_SETS+=("$1")
|
||||
shift
|
||||
@@ -129,6 +125,7 @@ SET_MAP[p]="performance"
|
||||
|
||||
SEARCH_PATHS=()
|
||||
|
||||
# 如果未指定测试集,或指定了 'all',则搜索所有目录
|
||||
if [ ${#TEST_SETS[@]} -eq 0 ] || [[ " ${TEST_SETS[@]} " =~ " all " ]]; then
|
||||
SEARCH_PATHS+=("${TESTDATA_DIR}")
|
||||
else
|
||||
@@ -141,13 +138,13 @@ else
|
||||
done
|
||||
fi
|
||||
|
||||
# 如果没有有效的搜索路径,则退出
|
||||
if [ ${#SEARCH_PATHS[@]} -eq 0 ]; then
|
||||
echo -e "\e[31m错误: 没有找到有效的测试集目录,测试中止。\e[0m"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
echo "SysY 测试运行器启动..."
|
||||
if [ -n "$OPTIMIZE_FLAG" ]; then echo "优化等级: ${OPTIMIZE_FLAG}"; fi
|
||||
echo "输入目录: ${SEARCH_PATHS[@]}"
|
||||
echo "临时目录: ${TMP_DIR}"
|
||||
echo "执行模式: ${EXECUTE_MODE}"
|
||||
@@ -157,6 +154,7 @@ if ${EXECUTE_MODE}; then
|
||||
fi
|
||||
echo ""
|
||||
|
||||
# 使用构建好的路径查找 .sy 文件并排序
|
||||
sy_files=$(find "${SEARCH_PATHS[@]}" -name "*.sy" | sort -V)
|
||||
if [ -z "$sy_files" ]; then
|
||||
echo "在指定目录中未找到任何 .sy 文件。"
|
||||
@@ -164,6 +162,7 @@ if [ -z "$sy_files" ]; then
|
||||
fi
|
||||
TOTAL_CASES=$(echo "$sy_files" | wc -w)
|
||||
|
||||
# --- 修复: 使用 here-string (<<<) 代替管道 (|) 来避免子 shell 问题 ---
|
||||
while IFS= read -r sy_file; do
|
||||
is_passed=1 # 1 表示通过, 0 表示失败
|
||||
|
||||
@@ -177,8 +176,10 @@ while IFS= read -r sy_file; do
|
||||
output_actual_file="${TMP_DIR}/${output_base_name}_sysyc_riscv64.actual_out"
|
||||
|
||||
echo "正在处理: $(basename "$sy_file") (路径: ${relative_path_no_ext}.sy)"
|
||||
|
||||
# 步骤 1: 使用 sysyc 编译 .sy 到 .s
|
||||
echo " 使用 sysyc 编译 (超时 ${SYSYC_TIMEOUT}s)..."
|
||||
timeout -s KILL ${SYSYC_TIMEOUT} "${SYSYC}" -S "${sy_file}" -o "${assembly_file}" ${OPTIMIZE_FLAG}
|
||||
timeout -s KILL ${SYSYC_TIMEOUT} "${SYSYC}" -S "${sy_file}" -o "${assembly_file}"
|
||||
SYSYC_STATUS=$?
|
||||
if [ $SYSYC_STATUS -eq 124 ]; then
|
||||
echo -e "\e[31m错误: SysY 编译 ${sy_file} 超时\e[0m"
|
||||
@@ -188,7 +189,9 @@ while IFS= read -r sy_file; do
|
||||
is_passed=0
|
||||
fi
|
||||
|
||||
# 只有当 EXECUTE_MODE 为 true 且上一步成功时才继续
|
||||
if ${EXECUTE_MODE} && [ "$is_passed" -eq 1 ]; then
|
||||
# 步骤 2: 使用 riscv64-linux-gnu-gcc 编译 .s 到可执行文件
|
||||
echo " 使用 gcc 编译 (超时 ${GCC_TIMEOUT}s)..."
|
||||
timeout -s KILL ${GCC_TIMEOUT} "${GCC_RISCV64}" "${assembly_file}" -o "${executable_file}" -L"${LIB_DIR}" -lsysy_riscv -static
|
||||
GCC_STATUS=$?
|
||||
@@ -210,6 +213,7 @@ while IFS= read -r sy_file; do
|
||||
continue
|
||||
fi
|
||||
|
||||
# 步骤 3, 4, 5: 只有当编译都成功时才执行
|
||||
if [ "$is_passed" -eq 1 ]; then
|
||||
echo " 正在执行 (超时 ${EXEC_TIMEOUT}s)..."
|
||||
|
||||
|
||||
@@ -5,15 +5,12 @@ add_library(riscv64_backend_lib STATIC
|
||||
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
|
||||
|
||||
@@ -8,6 +8,11 @@ namespace sysy {
|
||||
|
||||
char CalleeSavedHandler::ID = 0;
|
||||
|
||||
// 辅助函数,用于判断一个物理寄存器是否为浮点寄存器
|
||||
static bool is_fp_reg(PhysicalReg reg) {
|
||||
return reg >= PhysicalReg::F0 && reg <= PhysicalReg::F31;
|
||||
}
|
||||
|
||||
bool CalleeSavedHandler::runOnFunction(Function *F, AnalysisManager& AM) {
|
||||
// This pass works on MachineFunction level, not IR level
|
||||
return false;
|
||||
@@ -15,37 +20,114 @@ bool CalleeSavedHandler::runOnFunction(Function *F, AnalysisManager& AM) {
|
||||
|
||||
void CalleeSavedHandler::runOnMachineFunction(MachineFunction* mfunc) {
|
||||
StackFrameInfo& frame_info = mfunc->getFrameInfo();
|
||||
const std::set<PhysicalReg>& used_callee_saved = frame_info.used_callee_saved_regs;
|
||||
|
||||
std::set<PhysicalReg> used_callee_saved;
|
||||
|
||||
// 1. 扫描所有指令,找出被使用的callee-saved寄存器
|
||||
// 这个Pass在RegAlloc之后运行,所以可以访问到物理寄存器
|
||||
for (auto& mbb : mfunc->getBlocks()) {
|
||||
for (auto& instr : mbb->getInstructions()) {
|
||||
for (auto& op : instr->getOperands()) {
|
||||
|
||||
auto check_and_insert_reg = [&](RegOperand* reg_op) {
|
||||
if (reg_op && !reg_op->isVirtual()) {
|
||||
PhysicalReg preg = reg_op->getPReg();
|
||||
|
||||
// 检查整数 s1-s11
|
||||
if (preg >= PhysicalReg::S1 && preg <= PhysicalReg::S11) {
|
||||
used_callee_saved.insert(preg);
|
||||
}
|
||||
// 检查浮点 fs0-fs11 (f8,f9,f18-f27)
|
||||
else if ((preg >= PhysicalReg::F8 && preg <= PhysicalReg::F9) || (preg >= PhysicalReg::F18 && preg <= PhysicalReg::F27)) {
|
||||
used_callee_saved.insert(preg);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
if (op->getKind() == MachineOperand::KIND_REG) {
|
||||
check_and_insert_reg(static_cast<RegOperand*>(op.get()));
|
||||
} else if (op->getKind() == MachineOperand::KIND_MEM) {
|
||||
check_and_insert_reg(static_cast<MemOperand*>(op.get())->getBase());
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
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);
|
||||
// 2. 计算并更新 frame_info
|
||||
frame_info.callee_saved_size = used_callee_saved.size() * 8;
|
||||
|
||||
// 为了布局确定性和恢复顺序一致,对寄存器排序
|
||||
std::vector<PhysicalReg> sorted_regs(used_callee_saved.begin(), used_callee_saved.end());
|
||||
std::sort(sorted_regs.begin(), sorted_regs.end());
|
||||
|
||||
// 3. 在函数序言中插入保存指令
|
||||
MachineBasicBlock* entry_block = mfunc->getBlocks().front().get();
|
||||
auto& entry_instrs = entry_block->getInstructions();
|
||||
// 插入点在函数入口标签之后,或者就是最开始
|
||||
auto insert_pos = entry_instrs.begin();
|
||||
if (!entry_instrs.empty() && entry_instrs.front()->getOpcode() == RVOpcodes::LABEL) {
|
||||
insert_pos = std::next(insert_pos);
|
||||
}
|
||||
size = regs_to_save.size() * 8; // 每个寄存器占8字节 (64-bit)
|
||||
frame_info.callee_saved_size = size;
|
||||
|
||||
std::vector<std::unique_ptr<MachineInstr>> save_instrs;
|
||||
// [关键] 从局部变量区域之后开始分配空间
|
||||
int current_offset = - (16 + frame_info.locals_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;
|
||||
for (PhysicalReg reg : sorted_regs) {
|
||||
current_offset -= 8;
|
||||
RVOpcodes save_op = is_fp_reg(reg) ? RVOpcodes::FSD : RVOpcodes::SD;
|
||||
|
||||
// 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;
|
||||
auto save_instr = std::make_unique<MachineInstr>(save_op);
|
||||
save_instr->addOperand(std::make_unique<RegOperand>(reg));
|
||||
save_instr->addOperand(std::make_unique<MemOperand>(
|
||||
std::make_unique<RegOperand>(PhysicalReg::S0), // 基址为帧指针 s0
|
||||
std::make_unique<ImmOperand>(current_offset)
|
||||
));
|
||||
save_instrs.push_back(std::move(save_instr));
|
||||
}
|
||||
|
||||
if (!save_instrs.empty()) {
|
||||
entry_instrs.insert(insert_pos,
|
||||
std::make_move_iterator(save_instrs.begin()),
|
||||
std::make_move_iterator(save_instrs.end()));
|
||||
}
|
||||
|
||||
// 4. 在函数结尾(ret之前)插入恢复指令
|
||||
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>> restore_instrs;
|
||||
// [关键] 使用与保存时完全相同的逻辑来计算偏移量
|
||||
current_offset = - (16 + frame_info.locals_size);
|
||||
|
||||
for (PhysicalReg reg : sorted_regs) {
|
||||
current_offset -= 8;
|
||||
RVOpcodes restore_op = is_fp_reg(reg) ? RVOpcodes::FLD : RVOpcodes::LD;
|
||||
|
||||
auto restore_instr = std::make_unique<MachineInstr>(restore_op);
|
||||
restore_instr->addOperand(std::make_unique<RegOperand>(reg));
|
||||
restore_instr->addOperand(std::make_unique<MemOperand>(
|
||||
std::make_unique<RegOperand>(PhysicalReg::S0),
|
||||
std::make_unique<ImmOperand>(current_offset)
|
||||
));
|
||||
restore_instrs.push_back(std::move(restore_instr));
|
||||
}
|
||||
|
||||
if (!restore_instrs.empty()) {
|
||||
mbb->getInstructions().insert(it,
|
||||
std::make_move_iterator(restore_instrs.begin()),
|
||||
std::make_move_iterator(restore_instrs.end()));
|
||||
}
|
||||
goto next_block_label;
|
||||
}
|
||||
}
|
||||
next_block_label:;
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace sysy
|
||||
} // namespace sysy
|
||||
@@ -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
|
||||
@@ -1,22 +1,17 @@
|
||||
#include "PrologueEpilogueInsertion.h"
|
||||
#include "RISCv64LLIR.h" // 假设包含了 PhysicalReg, RVOpcodes 等定义
|
||||
#include "RISCv64ISel.h"
|
||||
#include "RISCv64RegAlloc.h" // 需要访问RegAlloc的结果
|
||||
#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();
|
||||
|
||||
// 使用标准的 Erase-Remove Idiom 来删除满足条件的元素
|
||||
instrs.erase(
|
||||
std::remove_if(instrs.begin(), instrs.end(),
|
||||
[](const std::unique_ptr<MachineInstr>& instr) {
|
||||
@@ -27,59 +22,39 @@ void PrologueEpilogueInsertionPass::runOnMachineFunction(MachineFunction* mfunc)
|
||||
);
|
||||
}
|
||||
|
||||
// 2. 确定需要保存的被调用者保存寄存器 (callee-saved)
|
||||
StackFrameInfo& frame_info = mfunc->getFrameInfo();
|
||||
Function* F = mfunc->getFunc();
|
||||
RISCv64ISel* isel = mfunc->getISel();
|
||||
|
||||
// [关键] 获取寄存器分配的结果 (vreg -> preg 的映射)
|
||||
// RegAlloc Pass 必须已经运行过
|
||||
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
|
||||
// 完全遵循 AsmPrinter 中的计算逻辑
|
||||
int total_stack_size = frame_info.locals_size +
|
||||
frame_info.spill_size +
|
||||
frame_info.callee_saved_size +
|
||||
16;
|
||||
16; // 为 ra 和 s0 固定的16字节
|
||||
|
||||
// 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;
|
||||
|
||||
// 只有在需要分配栈空间时才生成指令
|
||||
if (aligned_stack_size > 0) {
|
||||
// --- 4. 插入完整的序言 ---
|
||||
// --- 1. 插入序言 ---
|
||||
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
|
||||
// 1. addi sp, sp, -aligned_stack_size
|
||||
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
|
||||
// 2. sd ra, (aligned_stack_size - 8)(sp)
|
||||
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>(
|
||||
@@ -87,6 +62,8 @@ void PrologueEpilogueInsertionPass::runOnMachineFunction(MachineFunction* mfunc)
|
||||
std::make_unique<ImmOperand>(aligned_stack_size - 8)
|
||||
));
|
||||
prologue_instrs.push_back(std::move(save_ra));
|
||||
|
||||
// 3. sd s0, (aligned_stack_size - 16)(sp)
|
||||
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>(
|
||||
@@ -95,55 +72,66 @@ void PrologueEpilogueInsertionPass::runOnMachineFunction(MachineFunction* mfunc)
|
||||
));
|
||||
prologue_instrs.push_back(std::move(save_fp));
|
||||
|
||||
// 4.3. 设置新的帧指针 s0
|
||||
// 4. addi s0, sp, aligned_stack_size
|
||||
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));
|
||||
|
||||
// --- 在s0设置完毕后,使用物理寄存器加载栈参数 ---
|
||||
if (F && isel) {
|
||||
int arg_idx = 0;
|
||||
for (Argument* arg : F->getArguments()) {
|
||||
if (arg_idx >= 8) {
|
||||
unsigned vreg = isel->getVReg(arg);
|
||||
|
||||
if (frame_info.alloca_offsets.count(vreg) && vreg_to_preg_map.count(vreg)) {
|
||||
int offset = frame_info.alloca_offsets.at(vreg);
|
||||
PhysicalReg dest_preg = vreg_to_preg_map.at(vreg);
|
||||
Type* arg_type = arg->getType();
|
||||
|
||||
if (arg_type->isFloat()) {
|
||||
auto load_arg = std::make_unique<MachineInstr>(RVOpcodes::FLW);
|
||||
load_arg->addOperand(std::make_unique<RegOperand>(dest_preg));
|
||||
load_arg->addOperand(std::make_unique<MemOperand>(
|
||||
std::make_unique<RegOperand>(PhysicalReg::S0),
|
||||
std::make_unique<ImmOperand>(offset)
|
||||
));
|
||||
prologue_instrs.push_back(std::move(load_arg));
|
||||
} else {
|
||||
RVOpcodes load_op = arg_type->isPointer() ? RVOpcodes::LD : RVOpcodes::LW;
|
||||
auto load_arg = std::make_unique<MachineInstr>(load_op);
|
||||
load_arg->addOperand(std::make_unique<RegOperand>(dest_preg));
|
||||
load_arg->addOperand(std::make_unique<MemOperand>(
|
||||
std::make_unique<RegOperand>(PhysicalReg::S0),
|
||||
std::make_unique<ImmOperand>(offset)
|
||||
));
|
||||
prologue_instrs.push_back(std::move(load_arg));
|
||||
}
|
||||
}
|
||||
}
|
||||
arg_idx++;
|
||||
}
|
||||
}
|
||||
|
||||
// 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
|
||||
// 确定插入点
|
||||
auto insert_pos = entry_instrs.begin();
|
||||
|
||||
// 一次性将所有序言指令插入
|
||||
if (!prologue_instrs.empty()) {
|
||||
entry_instrs.insert(insert_pos,
|
||||
std::make_move_iterator(prologue_instrs.begin()),
|
||||
std::make_move_iterator(prologue_instrs.end()));
|
||||
}
|
||||
|
||||
// 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 ---
|
||||
// --- 2. 插入尾声 (此部分逻辑保持不变) ---
|
||||
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
|
||||
// 1. ld ra
|
||||
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>(
|
||||
@@ -151,6 +139,8 @@ void PrologueEpilogueInsertionPass::runOnMachineFunction(MachineFunction* mfunc)
|
||||
std::make_unique<ImmOperand>(aligned_stack_size - 8)
|
||||
));
|
||||
epilogue_instrs.push_back(std::move(restore_ra));
|
||||
|
||||
// 2. ld s0
|
||||
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>(
|
||||
@@ -159,18 +149,18 @@ void PrologueEpilogueInsertionPass::runOnMachineFunction(MachineFunction* mfunc)
|
||||
));
|
||||
epilogue_instrs.push_back(std::move(restore_fp));
|
||||
|
||||
// 5.3. 释放栈帧
|
||||
// 3. addi sp, sp, aligned_stack_size
|
||||
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()));
|
||||
|
||||
if (!epilogue_instrs.empty()) {
|
||||
mbb->getInstructions().insert(it,
|
||||
std::make_move_iterator(epilogue_instrs.begin()),
|
||||
std::make_move_iterator(epilogue_instrs.end()));
|
||||
}
|
||||
goto next_block;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -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
|
||||
@@ -1,8 +1,7 @@
|
||||
#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
|
||||
@@ -61,7 +60,7 @@ 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;
|
||||
@@ -82,7 +81,7 @@ void RISCv64AsmPrinter::printInstruction(MachineInstr* instr, bool debug) {
|
||||
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::FSD: *OS << "fsd "; 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;
|
||||
@@ -102,11 +101,10 @@ void RISCv64AsmPrinter::printInstruction(MachineInstr* instr, bool debug) {
|
||||
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
|
||||
case RVOpcodes::CALL: { // [核心修改] 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()) {
|
||||
@@ -238,30 +236,4 @@ 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
|
||||
@@ -1,13 +1,10 @@
|
||||
#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
|
||||
@@ -15,39 +12,6 @@ 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];
|
||||
@@ -75,36 +39,18 @@ std::string RISCv64CodeGen::module_gen() {
|
||||
|
||||
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;
|
||||
// 判断是否为大型零初始化数组,以便放入.bss段
|
||||
bool is_large_zero_array = false;
|
||||
if (init_values.getValues().size() == 1) {
|
||||
if (auto const_val = dynamic_cast<ConstantValue*>(init_values.getValues()[0])) {
|
||||
if (const_val->isInt() && const_val->getInt() == 0 && init_values.getNumbers()[0] > 16) {
|
||||
is_large_zero_array = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 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 {
|
||||
@@ -112,12 +58,12 @@ std::string RISCv64CodeGen::module_gen() {
|
||||
}
|
||||
}
|
||||
|
||||
// --- 步骤2:生成 .bss 段的代码 ---
|
||||
// --- 步骤2:生成 .bss 段的代码 (这部分不变) ---
|
||||
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);
|
||||
unsigned count = global->getInitValues().getNumbers()[0];
|
||||
unsigned total_size = count * 4; // 假设元素都是4字节
|
||||
|
||||
ss << " .align 3\n";
|
||||
ss << ".globl " << global->getName() << "\n";
|
||||
@@ -128,67 +74,33 @@ std::string RISCv64CodeGen::module_gen() {
|
||||
}
|
||||
}
|
||||
|
||||
// --- 步骤3:生成 .data 段的代码 ---
|
||||
// --- [修改] 步骤3:生成 .data 段的代码 ---
|
||||
// 我们需要检查 data_globals 和 常量列表是否都为空
|
||||
if (!data_globals.empty() || !module->getConsts().empty()) {
|
||||
ss << ".data\n";
|
||||
|
||||
// a. 处理普通的全局变量 (GlobalValue)
|
||||
// a. 先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());
|
||||
}
|
||||
printInitializer(ss, global->getInitValues());
|
||||
}
|
||||
|
||||
// b. 处理全局常量 (ConstantVariable)
|
||||
// b. [新增] 再处理全局常量 (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 ---
|
||||
// --- 处理函数 (.text段) 的逻辑保持不变 ---
|
||||
if (!module->getFunctions().empty()) {
|
||||
ss << ".text\n";
|
||||
for (const auto& func_pair : module->getFunctions()) {
|
||||
if (func_pair.second.get() && !func_pair.second->getBasicBlocks().empty()) {
|
||||
if (func_pair.second.get()) {
|
||||
ss << function_gen(func_pair.second.get());
|
||||
if (DEBUG) std::cerr << "Function: " << func_pair.first << " generated.\n";
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -203,58 +115,25 @@ std::string RISCv64CodeGen::function_gen(Function* func) {
|
||||
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);
|
||||
std::stringstream ss1;
|
||||
RISCv64AsmPrinter printer1(mfunc.get());
|
||||
printer1.run(ss1, 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 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();
|
||||
|
||||
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 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;
|
||||
@@ -264,7 +143,7 @@ std::string RISCv64CodeGen::function_gen(Function* func) {
|
||||
PrologueEpilogueInsertionPass pei_pass;
|
||||
pei_pass.runOnMachineFunction(mfunc.get());
|
||||
|
||||
// 阶段 3.3: 大立即数合法化
|
||||
// 阶段 3.3: 清理产生的大立即数
|
||||
LegalizeImmediatesPass legalizer;
|
||||
legalizer.runOnMachineFunction(mfunc.get());
|
||||
|
||||
@@ -272,7 +151,7 @@ std::string RISCv64CodeGen::function_gen(Function* func) {
|
||||
std::stringstream ss;
|
||||
RISCv64AsmPrinter printer(mfunc.get());
|
||||
printer.run(ss);
|
||||
|
||||
if (DEBUG) ss << "\n" << ss1.str(); // 将指令选择阶段的结果也包含在最终输出中
|
||||
return ss.str();
|
||||
}
|
||||
|
||||
|
||||
@@ -1,10 +1,9 @@
|
||||
#include "RISCv64ISel.h"
|
||||
#include "IR.h" // For GlobalValue
|
||||
#include <stdexcept>
|
||||
#include <set>
|
||||
#include <functional>
|
||||
#include <cmath>
|
||||
#include <limits>
|
||||
#include <cmath> // For std::fabs
|
||||
#include <limits> // For std::numeric_limits
|
||||
#include <iostream>
|
||||
|
||||
namespace sysy {
|
||||
@@ -168,6 +167,33 @@ void RISCv64ISel::selectBasicBlock(BasicBlock* bb) {
|
||||
select_recursive(node_to_select);
|
||||
}
|
||||
}
|
||||
|
||||
if (CurMBB == MFunc->getBlocks().front().get()) { // 只对入口块操作
|
||||
auto keepalive = std::make_unique<MachineInstr>(RVOpcodes::PSEUDO_KEEPALIVE);
|
||||
for (Argument* arg : F->getArguments()) {
|
||||
keepalive->addOperand(std::make_unique<RegOperand>(getVReg(arg)));
|
||||
}
|
||||
|
||||
auto& instrs = CurMBB->getInstructions();
|
||||
auto insert_pos = instrs.end();
|
||||
|
||||
// 关键:检查基本块是否以一个“终止指令”结尾
|
||||
if (!instrs.empty()) {
|
||||
RVOpcodes last_op = instrs.back()->getOpcode();
|
||||
// 扩充了判断条件,涵盖所有可能的终止指令
|
||||
if (last_op == RVOpcodes::J || last_op == RVOpcodes::RET ||
|
||||
last_op == RVOpcodes::BEQ || last_op == RVOpcodes::BNE ||
|
||||
last_op == RVOpcodes::BLT || last_op == RVOpcodes::BGE ||
|
||||
last_op == RVOpcodes::BLTU || last_op == RVOpcodes::BGEU)
|
||||
{
|
||||
// 如果是,插入点就在这个终止指令之前
|
||||
insert_pos = std::prev(instrs.end());
|
||||
}
|
||||
}
|
||||
|
||||
// 在计算出的正确位置插入伪指令
|
||||
instrs.insert(insert_pos, std::move(keepalive));
|
||||
}
|
||||
}
|
||||
|
||||
// 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)
|
||||
@@ -183,12 +209,8 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
case DAGNode::CONSTANT:
|
||||
case DAGNode::ALLOCA_ADDR:
|
||||
if (node->value) {
|
||||
// GlobalValue objects (global variables) should not get virtual registers
|
||||
// since they represent memory addresses, not register-allocated values
|
||||
if (dynamic_cast<GlobalValue*>(node->value) == nullptr) {
|
||||
// 确保它有一个关联的虚拟寄存器即可,不生成代码。
|
||||
getVReg(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;
|
||||
|
||||
@@ -380,7 +402,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
Value* base = nullptr;
|
||||
Value* offset = nullptr;
|
||||
|
||||
// 扩展基地址的判断,使其可以识别 AllocaInst 或 GlobalValue
|
||||
// [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;
|
||||
@@ -399,7 +421,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
CurMBB->addInstruction(std::move(li));
|
||||
}
|
||||
|
||||
// 2. 根据基地址的类型,生成不同的指令来获取基地址
|
||||
// 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(Type::getIntType()); // 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
|
||||
@@ -430,7 +452,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
}
|
||||
}
|
||||
|
||||
// 在BINARY节点内部按需加载常量操作数。
|
||||
// [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) {
|
||||
@@ -461,7 +483,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
auto dest_vreg = getVReg(bin);
|
||||
auto lhs_vreg = getVReg(lhs);
|
||||
|
||||
// 融合 ADDIW 优化。
|
||||
// [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);
|
||||
@@ -517,15 +539,6 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
CurMBB->addInstruction(std::move(instr));
|
||||
break;
|
||||
}
|
||||
case Instruction::kSRA: {
|
||||
auto rhs_const = dynamic_cast<ConstantInteger*>(rhs);
|
||||
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SRAIW);
|
||||
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));
|
||||
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));
|
||||
@@ -745,9 +758,8 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
CurMBB->addInstruction(std::move(instr));
|
||||
break;
|
||||
}
|
||||
case Instruction::kFtoI: { // 浮点 to 整数 (使用硬件指令进行向零截断)
|
||||
// 直接生成一条带有 rtz 舍入模式的转换指令
|
||||
auto instr = std::make_unique<MachineInstr>(RVOpcodes::FCVT_W_S_RTZ);
|
||||
case Instruction::kFtoI: { // 浮点 to 整数
|
||||
auto instr = std::make_unique<MachineInstr>(RVOpcodes::FCVT_W_S);
|
||||
instr->addOperand(std::make_unique<RegOperand>(dest_vreg)); // 目标是整数vreg
|
||||
instr->addOperand(std::make_unique<RegOperand>(src_vreg)); // 源是浮点vreg
|
||||
CurMBB->addInstruction(std::move(instr));
|
||||
@@ -931,7 +943,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
|
||||
// --- 步骤 3: 生成CALL指令 ---
|
||||
auto call_instr = std::make_unique<MachineInstr>(RVOpcodes::CALL);
|
||||
// 如果函数有返回值,将它的目标虚拟寄存器作为第一个操作数
|
||||
// [协议] 如果函数有返回值,将它的目标虚拟寄存器作为第一个操作数
|
||||
if (!call->getType()->isVoid()) {
|
||||
unsigned dest_vreg = getVReg(call);
|
||||
call_instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
|
||||
@@ -1008,7 +1020,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
} else {
|
||||
// --- 处理整数/指针返回值 ---
|
||||
// 返回值需要被放入 a0
|
||||
// 在RETURN节点内加载常量返回值
|
||||
// [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));
|
||||
@@ -1022,7 +1034,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
}
|
||||
}
|
||||
}
|
||||
// 函数尾声(epilogue)不由RETURN节点生成,
|
||||
// [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));
|
||||
@@ -1036,7 +1048,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
auto then_bb_name = cond_br->getThenBlock()->getName();
|
||||
auto else_bb_name = cond_br->getElseBlock()->getName();
|
||||
|
||||
// 检查分支条件是否为编译期常量
|
||||
// [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
|
||||
@@ -1051,7 +1063,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
}
|
||||
// 如果条件不是常量,则执行标准流程
|
||||
else {
|
||||
// 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)
|
||||
// [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);
|
||||
@@ -1085,7 +1097,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
}
|
||||
|
||||
case DAGNode::MEMSET: {
|
||||
// Memset的核心展开逻辑在虚拟寄存器层面是正确的,无需修改。
|
||||
// [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。
|
||||
|
||||
@@ -1131,11 +1143,10 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
auto r_value_byte = getVReg(memset->getValue());
|
||||
|
||||
// 为memset内部逻辑创建新的临时虚拟寄存器
|
||||
Type* ptr_type = Type::getPointerType(Type::getIntType());
|
||||
auto r_counter = getNewVReg(ptr_type);
|
||||
auto r_end_addr = getNewVReg(ptr_type);
|
||||
auto r_current_addr = getNewVReg(ptr_type);
|
||||
auto r_temp_val = getNewVReg(Type::getIntType());
|
||||
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) {
|
||||
@@ -1226,7 +1237,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
|
||||
// --- Step 1: 获取基地址 (此部分逻辑正确,保持不变) ---
|
||||
auto base_ptr_node = node->operands[0];
|
||||
auto current_addr_vreg = getNewVReg(gep->getType());
|
||||
auto current_addr_vreg = getNewVReg();
|
||||
|
||||
if (auto alloca_base = dynamic_cast<AllocaInst*>(base_ptr_node->value)) {
|
||||
auto frame_addr_instr = std::make_unique<MachineInstr>(RVOpcodes::FRAME_ADDR);
|
||||
@@ -1268,20 +1279,15 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
// 如果步长为0(例如对一个void类型或空结构体索引),则不产生任何偏移
|
||||
if (stride != 0) {
|
||||
// --- 为当前索引和步长生成偏移计算指令 ---
|
||||
auto offset_vreg = getNewVReg(Type::getIntType());
|
||||
|
||||
// 处理索引 - 区分常量与动态值
|
||||
unsigned index_vreg;
|
||||
auto offset_vreg = getNewVReg();
|
||||
auto index_vreg = getVReg(indexValue);
|
||||
|
||||
// 如果索引是常量,先用 LI 指令加载到虚拟寄存器
|
||||
if (auto const_index = dynamic_cast<ConstantValue*>(indexValue)) {
|
||||
// 对于常量索引,直接创建新的虚拟寄存器
|
||||
index_vreg = getNewVReg(Type::getIntType());
|
||||
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
|
||||
li->addOperand(std::make_unique<RegOperand>(index_vreg));
|
||||
li->addOperand(std::make_unique<ImmOperand>(const_index->getInt()));
|
||||
CurMBB->addInstruction(std::move(li));
|
||||
} else {
|
||||
// 对于动态索引,使用已存在的虚拟寄存器
|
||||
index_vreg = getVReg(indexValue);
|
||||
}
|
||||
|
||||
// 优化:如果步长是1,可以直接移动(MV)作为偏移量,无需乘法
|
||||
@@ -1292,7 +1298,7 @@ void RISCv64ISel::selectNode(DAGNode* node) {
|
||||
CurMBB->addInstruction(std::move(mv));
|
||||
} else {
|
||||
// 步长不为1,需要生成乘法指令
|
||||
auto size_vreg = getNewVReg(Type::getIntType());
|
||||
auto size_vreg = getNewVReg();
|
||||
auto li_size = std::make_unique<MachineInstr>(RVOpcodes::LI);
|
||||
li_size->addOperand(std::make_unique<RegOperand>(size_vreg));
|
||||
li_size->addOperand(std::make_unique<ImmOperand>(stride));
|
||||
@@ -1439,7 +1445,7 @@ std::vector<std::unique_ptr<RISCv64ISel::DAGNode>> RISCv64ISel::build_dag(BasicB
|
||||
|
||||
// 依次添加所有索引作为后续的操作数
|
||||
for (auto index : gep->getIndices()) {
|
||||
// 从 Use 对象中获取真正的 Value*
|
||||
// [修复] 从 Use 对象中获取真正的 Value*
|
||||
gep_node->operands.push_back(get_operand_node(index->getValue(), value_to_node, nodes_storage));
|
||||
}
|
||||
} else if (auto load = dynamic_cast<LoadInst*>(inst)) {
|
||||
@@ -1467,7 +1473,7 @@ std::vector<std::unique_ptr<RISCv64ISel::DAGNode>> RISCv64ISel::build_dag(BasicB
|
||||
}
|
||||
}
|
||||
}
|
||||
if (bin->isFPBinary()) { // 假设浮点指令枚举值更大
|
||||
if (bin->getKind() >= Instruction::kFAdd) { // 假设浮点指令枚举值更大
|
||||
auto fbin_node = create_node(DAGNode::FBINARY, bin, value_to_node, nodes_storage);
|
||||
fbin_node->operands.push_back(get_operand_node(bin->getLhs(), value_to_node, nodes_storage));
|
||||
fbin_node->operands.push_back(get_operand_node(bin->getRhs(), value_to_node, nodes_storage));
|
||||
@@ -1543,7 +1549,7 @@ unsigned RISCv64ISel::getTypeSizeInBytes(Type* type) {
|
||||
}
|
||||
}
|
||||
|
||||
// 打印DAG图以供调试的辅助函数
|
||||
// [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;
|
||||
@@ -1639,8 +1645,4 @@ void RISCv64ISel::print_dag(const std::vector<std::unique_ptr<DAGNode>>& dag, co
|
||||
std::cerr << "======================================\n\n";
|
||||
}
|
||||
|
||||
unsigned int RISCv64ISel::getVRegCounter() const {
|
||||
return vreg_counter;
|
||||
}
|
||||
|
||||
} // namespace sysy
|
||||
@@ -1,122 +1,6 @@
|
||||
#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";
|
||||
}
|
||||
|
||||
}
|
||||
}
|
||||
@@ -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
@@ -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
|
||||
@@ -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
|
||||
@@ -20,8 +20,6 @@ public:
|
||||
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 printBasicBlock(MachineBasicBlock* mbb, bool debug = false);
|
||||
|
||||
@@ -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;
|
||||
};
|
||||
|
||||
|
||||
@@ -3,12 +3,6 @@
|
||||
|
||||
#include "RISCv64LLIR.h"
|
||||
|
||||
// Forward declarations
|
||||
namespace sysy {
|
||||
class GlobalValue;
|
||||
class Value;
|
||||
}
|
||||
|
||||
extern int DEBUG;
|
||||
extern int DEEPDEBUG;
|
||||
|
||||
@@ -22,8 +16,8 @@ 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;
|
||||
unsigned getNewVReg() { return vreg_counter++; }
|
||||
unsigned getNewVReg(Type* type);
|
||||
// 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; }
|
||||
|
||||
@@ -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>
|
||||
@@ -39,16 +38,14 @@ enum class PhysicalReg {
|
||||
|
||||
// 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_START_ID = 100000,
|
||||
PHYS_REG_END_ID = PHYS_REG_START_ID + 320, // 预留足够的空间
|
||||
|
||||
INVALID, ///< 无效寄存器标记
|
||||
};
|
||||
|
||||
// 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
|
||||
@@ -88,7 +85,6 @@ enum class RVOpcodes {
|
||||
// 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)
|
||||
@@ -199,11 +195,6 @@ 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 +274,14 @@ 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
|
||||
std::map<unsigned, PhysicalReg> vreg_to_preg_map;
|
||||
std::vector<PhysicalReg> callee_saved_regs; // 用于存储需要保存的被调用者保存寄存器列表
|
||||
};
|
||||
|
||||
// 机器函数
|
||||
@@ -305,7 +295,7 @@ 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));
|
||||
}
|
||||
|
||||
@@ -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
|
||||
@@ -8,10 +8,7 @@
|
||||
#include "CalleeSavedHandler.h"
|
||||
#include "LegalizeImmediates.h"
|
||||
#include "PrologueEpilogueInsertion.h"
|
||||
#include "EliminateFrameIndices.h"
|
||||
#include "Pass.h"
|
||||
#include "DivStrengthReduction.h"
|
||||
|
||||
|
||||
namespace sysy {
|
||||
|
||||
|
||||
@@ -3,15 +3,9 @@
|
||||
|
||||
#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 {
|
||||
|
||||
@@ -23,98 +17,58 @@ public:
|
||||
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*>;
|
||||
using LiveSet = std::set<unsigned>; // 活跃虚拟寄存器集合
|
||||
using InterferenceGraph = std::map<unsigned, std::set<unsigned>>;
|
||||
|
||||
// --- 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 ---
|
||||
// 栈帧管理
|
||||
void eliminateFrameIndices();
|
||||
|
||||
// 活跃性分析
|
||||
void analyzeLiveness();
|
||||
|
||||
MachineFunction* MFunc;
|
||||
RISCv64ISel* ISel;
|
||||
// 构建干扰图
|
||||
void buildInterferenceGraph();
|
||||
|
||||
// --- 算法数据结构 ---
|
||||
// 寄存器池
|
||||
// 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);
|
||||
|
||||
// 辅助函数,处理调用约定
|
||||
void handleCallingConvention();
|
||||
|
||||
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;
|
||||
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;
|
||||
// 存储vreg到IR Value*的反向映射
|
||||
// 这个map将在run()函数开始时被填充,并在rewriteFunction()中使用。
|
||||
std::map<unsigned, Value*> vreg_to_value_map;
|
||||
std::map<PhysicalReg, unsigned> preg_to_vreg_id_map; // 物理寄存器到特殊vreg ID的映射
|
||||
|
||||
// 用于计算类型大小的辅助函数
|
||||
unsigned getTypeSizeInBytes(Type* type);
|
||||
|
||||
// 辅助函数,用于打印集合
|
||||
static void printLiveSet(const LiveSet& s, const std::string& name, std::ostream& os);
|
||||
|
||||
// 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
|
||||
|
||||
@@ -202,7 +202,6 @@ class Use {
|
||||
|
||||
public:
|
||||
unsigned getIndex() const { return index; } ///< 返回value在User操作数中的位置
|
||||
void setIndex(int newIndex) { index = newIndex; } ///< 设置value在User操作数中的位置
|
||||
User* getUser() const { return user; } ///< 返回使用者
|
||||
Value* getValue() const { return value; } ///< 返回被使用的值
|
||||
void setValue(Value *newValue) { value = newValue; } ///< 将被使用的值设置为newValue
|
||||
@@ -230,14 +229,7 @@ class Value {
|
||||
std::list<std::shared_ptr<Use>>& getUses() { return uses; } ///< 获取使用关系列表
|
||||
void addUse(const std::shared_ptr<Use> &use) { uses.push_back(use); } ///< 添加使用关系
|
||||
void replaceAllUsesWith(Value *value); ///< 将原来使用该value的使用者全变为使用给定参数value并修改相应use关系
|
||||
void removeUse(const std::shared_ptr<Use> &use) {
|
||||
assert(use != nullptr && "Use cannot be null");
|
||||
assert(use->getValue() == this && "Use being removed does NOT point to this Value!");
|
||||
auto it = std::find(uses.begin(), uses.end(), use);
|
||||
assert(it != uses.end() && "Use not found in Value's uses");
|
||||
uses.remove(use);
|
||||
} ///< 删除使用关系use
|
||||
void removeAllUses();
|
||||
void removeUse(const std::shared_ptr<Use> &use) { uses.remove(use); } ///< 删除使用关系use
|
||||
};
|
||||
|
||||
/**
|
||||
@@ -522,15 +514,12 @@ public:
|
||||
explicit BasicBlock(Function *parent, const std::string &name = "")
|
||||
: Value(Type::getLabelType(), name), parent(parent) {}
|
||||
~BasicBlock() override {
|
||||
// for (auto pre : predecessors) {
|
||||
// pre->removeSuccessor(this);
|
||||
// }
|
||||
// for (auto suc : successors) {
|
||||
// suc->removePredecessor(this);
|
||||
// }
|
||||
// 这些关系应该在 BasicBlock 被从 Function 中移除时,
|
||||
// 由负责 CFG 优化的 Pass (例如 SCCP 的 RemoveDeadBlock) 显式地清理。
|
||||
// 析构函数只负责清理 BasicBlock 自身拥有的资源(例如,指令列表)。
|
||||
for (auto pre : predecessors) {
|
||||
pre->removeSuccessor(this);
|
||||
}
|
||||
for (auto suc : successors) {
|
||||
suc->removePredecessor(this);
|
||||
}
|
||||
}
|
||||
|
||||
public:
|
||||
@@ -584,9 +573,7 @@ public:
|
||||
if (iter != predecessors.end()) {
|
||||
predecessors.erase(iter);
|
||||
} else {
|
||||
// 如果没有找到前驱块,可能是因为它已经被移除或不存在
|
||||
// 这可能是一个错误情况,或者是因为在CFG优化过程中已经处理
|
||||
// assert(false && "Predecessor block not found in BasicBlock");
|
||||
assert(false);
|
||||
}
|
||||
}
|
||||
void removeSuccessor(BasicBlock *block) {
|
||||
@@ -594,9 +581,7 @@ public:
|
||||
if (iter != successors.end()) {
|
||||
successors.erase(iter);
|
||||
} else {
|
||||
// 如果没有找到后继块,可能是因为它已经被移除或不存在
|
||||
// 这可能是一个错误情况,或者是因为在CFG优化过程中已经处理
|
||||
// assert(false && "Successor block not found in BasicBlock");
|
||||
assert(false);
|
||||
}
|
||||
}
|
||||
void replacePredecessor(BasicBlock *oldBlock, BasicBlock *newBlock) {
|
||||
@@ -614,7 +599,7 @@ public:
|
||||
prev->addSuccessor(next);
|
||||
next->addPredecessor(prev);
|
||||
}
|
||||
iterator removeInst(iterator pos) { return instructions.erase(pos); }
|
||||
void removeInst(iterator pos) { instructions.erase(pos); }
|
||||
void removeInst(Instruction *inst) {
|
||||
auto pos = std::find_if(instructions.begin(), instructions.end(),
|
||||
[inst](const std::unique_ptr<Instruction> &i) { return i.get() == inst; });
|
||||
@@ -650,7 +635,11 @@ class User : public Value {
|
||||
operands.emplace_back(std::make_shared<Use>(operands.size(), this, value));
|
||||
value->addUse(operands.back());
|
||||
} ///< 增加操作数
|
||||
void removeOperand(unsigned index);
|
||||
void removeOperand(unsigned index) {
|
||||
auto value = getOperand(index);
|
||||
value->removeUse(operands[index]);
|
||||
operands.erase(operands.begin() + index);
|
||||
} ///< 移除操作数
|
||||
template <typename ContainerT>
|
||||
void addOperands(const ContainerT &newoperands) {
|
||||
for (auto value : newoperands) {
|
||||
@@ -706,19 +695,19 @@ class Instruction : public User {
|
||||
kCondBr = 0x1UL << 30,
|
||||
kBr = 0x1UL << 31,
|
||||
kReturn = 0x1UL << 32,
|
||||
kUnreachable = 0x1UL << 33,
|
||||
// mem op
|
||||
kAlloca = 0x1UL << 34,
|
||||
kLoad = 0x1UL << 35,
|
||||
kStore = 0x1UL << 36,
|
||||
kGetElementPtr = 0x1UL << 37,
|
||||
kMemset = 0x1UL << 38,
|
||||
kAlloca = 0x1UL << 33,
|
||||
kLoad = 0x1UL << 34,
|
||||
kStore = 0x1UL << 35,
|
||||
kGetElementPtr = 0x1UL << 36,
|
||||
kMemset = 0x1UL << 37,
|
||||
// kGetSubArray = 0x1UL << 38,
|
||||
// Constant Kind removed as Constants are now Values, not Instructions.
|
||||
// kConstant = 0x1UL << 37, // Conflicts with kMemset if kept as is
|
||||
// phi
|
||||
kPhi = 0x1UL << 39,
|
||||
kBitItoF = 0x1UL << 40,
|
||||
kBitFtoI = 0x1UL << 41,
|
||||
kSRA = 0x1UL << 42,
|
||||
kMulh = 0x1UL << 43
|
||||
};
|
||||
|
||||
protected:
|
||||
@@ -815,12 +804,6 @@ public:
|
||||
return "Memset";
|
||||
case kPhi:
|
||||
return "Phi";
|
||||
case kBitItoF:
|
||||
return "BitItoF";
|
||||
case kBitFtoI:
|
||||
return "BitFtoI";
|
||||
case kSRA:
|
||||
return "SRA";
|
||||
default:
|
||||
return "Unknown";
|
||||
}
|
||||
@@ -832,15 +815,11 @@ public:
|
||||
|
||||
bool isBinary() const {
|
||||
static constexpr uint64_t BinaryOpMask =
|
||||
(kAdd | kSub | kMul | kDiv | kRem | kAnd | kOr | kSRA | kMulh) |
|
||||
(kICmpEQ | kICmpNE | kICmpLT | kICmpGT | kICmpLE | kICmpGE);
|
||||
return kind & BinaryOpMask;
|
||||
}
|
||||
bool isFPBinary() const {
|
||||
static constexpr uint64_t FPBinaryOpMask =
|
||||
(kAdd | kSub | kMul | kDiv | kRem | kAnd | kOr) |
|
||||
(kICmpEQ | kICmpNE | kICmpLT | kICmpGT | kICmpLE | kICmpGE) |
|
||||
(kFAdd | kFSub | kFMul | kFDiv) |
|
||||
(kFCmpEQ | kFCmpNE | kFCmpLT | kFCmpGT | kFCmpLE | kFCmpGE);
|
||||
return kind & FPBinaryOpMask;
|
||||
return kind & BinaryOpMask;
|
||||
}
|
||||
bool isUnary() const {
|
||||
static constexpr uint64_t UnaryOpMask =
|
||||
@@ -853,7 +832,7 @@ public:
|
||||
return kind & MemoryOpMask;
|
||||
}
|
||||
bool isTerminator() const {
|
||||
static constexpr uint64_t TerminatorOpMask = kCondBr | kBr | kReturn | kUnreachable;
|
||||
static constexpr uint64_t TerminatorOpMask = kCondBr | kBr | kReturn;
|
||||
return kind & TerminatorOpMask;
|
||||
}
|
||||
bool isCmp() const {
|
||||
@@ -873,7 +852,6 @@ public:
|
||||
}
|
||||
bool isUnconditional() const { return kind == kBr; }
|
||||
bool isConditional() const { return kind == kCondBr; }
|
||||
bool isCondBr() const { return kind == kCondBr; }
|
||||
bool isPhi() const { return kind == kPhi; }
|
||||
bool isAlloca() const { return kind == kAlloca; }
|
||||
bool isLoad() const { return kind == kLoad; }
|
||||
@@ -882,7 +860,6 @@ public:
|
||||
bool isMemset() const { return kind == kMemset; }
|
||||
bool isCall() const { return kind == kCall; }
|
||||
bool isReturn() const { return kind == kReturn; }
|
||||
bool isUnreachable() const { return kind == kUnreachable; }
|
||||
bool isDefine() const {
|
||||
static constexpr uint64_t DefineOpMask = kAlloca | kStore | kPhi;
|
||||
return (kind & DefineOpMask) != 0U;
|
||||
@@ -908,54 +885,37 @@ class PhiInst : public Instruction {
|
||||
const std::string &name = "")
|
||||
: Instruction(Kind::kPhi, type, parent, name), vsize(rhs.size()) {
|
||||
assert(rhs.size() == Blocks.size() && "PhiInst: rhs and Blocks must have the same size");
|
||||
for(size_t i = 0; i < vsize; ++i) {
|
||||
for(size_t i = 0; i < rhs.size(); ++i) {
|
||||
addOperand(rhs[i]);
|
||||
addOperand(Blocks[i]);
|
||||
blk2val[Blocks[i]] = rhs[i];
|
||||
}
|
||||
}
|
||||
|
||||
public:
|
||||
Value* getValue(unsigned k) const {return getOperand(2 * k);} ///< 获取位置为k的值
|
||||
BasicBlock* getBlock(unsigned k) const {return dynamic_cast<BasicBlock*>(getOperand(2 * k + 1));}
|
||||
|
||||
auto& getincomings() const {return blk2val;} ///< 获取所有的基本块和对应的值
|
||||
|
||||
Value* getvalfromBlk(BasicBlock* blk);
|
||||
BasicBlock* getBlkfromVal(Value* val);
|
||||
|
||||
unsigned getNumIncomingValues() const { return vsize; } ///< 获取传入值的数量
|
||||
Value *getIncomingValue(unsigned Idx) const { return getOperand(Idx * 2); } ///< 获取指定位置的传入值
|
||||
BasicBlock *getIncomingBlock(unsigned Idx) const {return dynamic_cast<BasicBlock *>(getOperand(Idx * 2 + 1)); } ///< 获取指定位置的传入基本块
|
||||
|
||||
Value* getValfromBlk(BasicBlock* block);
|
||||
BasicBlock* getBlkfromVal(Value* value);
|
||||
|
||||
void addIncoming(Value *value, BasicBlock *block) {
|
||||
assert(value && block && "PhiInst: value and block cannot be null");
|
||||
assert(value && block && "PhiInst: value and block must not be null");
|
||||
addOperand(value);
|
||||
addOperand(block);
|
||||
blk2val[block] = value;
|
||||
vsize++;
|
||||
} ///< 添加传入值和对应的基本块
|
||||
void removeIncoming(unsigned Idx) {
|
||||
assert(Idx < vsize && "PhiInst: Index out of bounds");
|
||||
auto blk = getIncomingBlock(Idx);
|
||||
removeOperand(Idx * 2 + 1); // Remove block
|
||||
removeOperand(Idx * 2); // Remove value
|
||||
blk2val.erase(blk);
|
||||
vsize--;
|
||||
} ///< 移除指定位置的传入值和对应的基本块
|
||||
// 移除指定的传入值或基本块
|
||||
void removeIncomingValue(Value *value);
|
||||
void removeIncomingBlock(BasicBlock *block);
|
||||
// 设置指定位置的传入值或基本块
|
||||
void setIncomingValue(unsigned Idx, Value *value);
|
||||
void setIncomingBlock(unsigned Idx, BasicBlock *block);
|
||||
// 替换指定位置的传入值或基本块(原理是删除再添加)保留旧块或者旧值
|
||||
void replaceIncomingValue(Value *oldValue, Value *newValue);
|
||||
void replaceIncomingBlock(BasicBlock *oldBlock, BasicBlock *newBlock);
|
||||
// 替换指定位置的传入值或基本块(原理是删除再添加)
|
||||
void replaceIncomingValue(Value *oldValue, Value *newValue, BasicBlock *newBlock);
|
||||
void replaceIncomingBlock(BasicBlock *oldBlock, BasicBlock *newBlock, Value *newValue);
|
||||
void refreshMap() {
|
||||
blk2val.clear();
|
||||
for (unsigned i = 0; i < vsize; ++i) {
|
||||
blk2val[getIncomingBlock(i)] = getIncomingValue(i);
|
||||
}
|
||||
} ///< 刷新块到值的映射关系
|
||||
|
||||
void delValue(Value* val);
|
||||
void delBlk(BasicBlock* blk);
|
||||
|
||||
void replaceBlk(BasicBlock* newBlk, unsigned k);
|
||||
void replaceold2new(BasicBlock* oldBlk, BasicBlock* newBlk);
|
||||
void refreshB2VMap();
|
||||
|
||||
auto getValues() { return make_range(std::next(operand_begin()), operand_end()); }
|
||||
};
|
||||
|
||||
@@ -1099,10 +1059,12 @@ class UncondBrInst : public Instruction {
|
||||
friend class Function;
|
||||
|
||||
protected:
|
||||
UncondBrInst(BasicBlock *block,
|
||||
UncondBrInst(BasicBlock *block, std::vector<Value *> args,
|
||||
BasicBlock *parent = nullptr)
|
||||
: Instruction(kBr, Type::getVoidType(), parent, "") {
|
||||
// assert(block->getNumArguments() == args.size());
|
||||
addOperand(block);
|
||||
addOperands(args);
|
||||
}
|
||||
|
||||
public:
|
||||
@@ -1110,16 +1072,6 @@ public:
|
||||
auto getArguments() const {
|
||||
return make_range(std::next(operand_begin()), operand_end());
|
||||
}
|
||||
std::vector<BasicBlock *> getSuccessors() const {
|
||||
std::vector<BasicBlock *> succs;
|
||||
// 假设无条件分支的目标块是它的第一个操作数
|
||||
if (getNumOperands() > 0) {
|
||||
if (auto target_bb = dynamic_cast<BasicBlock *>(getOperand(0))) {
|
||||
succs.push_back(target_bb);
|
||||
}
|
||||
}
|
||||
return succs;
|
||||
}
|
||||
|
||||
}; // class UncondBrInst
|
||||
|
||||
@@ -1131,12 +1083,17 @@ class CondBrInst : public Instruction {
|
||||
friend class Function;
|
||||
|
||||
protected:
|
||||
CondBrInst(Value *condition, BasicBlock *thenBlock, BasicBlock *elseBlock,
|
||||
BasicBlock *parent = nullptr)
|
||||
CondBrInst(Value *condition, BasicBlock *thenBlock, BasicBlock *elseBlock,
|
||||
const std::vector<Value *> &thenArgs,
|
||||
const std::vector<Value *> &elseArgs, BasicBlock *parent = nullptr)
|
||||
: Instruction(kCondBr, Type::getVoidType(), parent, "") {
|
||||
// assert(thenBlock->getNumArguments() == thenArgs.size() and
|
||||
// elseBlock->getNumArguments() == elseArgs.size());
|
||||
addOperand(condition);
|
||||
addOperand(thenBlock);
|
||||
addOperand(elseBlock);
|
||||
addOperands(thenArgs);
|
||||
addOperands(elseArgs);
|
||||
}
|
||||
public:
|
||||
Value* getCondition() const { return getOperand(0); }
|
||||
@@ -1146,39 +1103,29 @@ public:
|
||||
BasicBlock* getElseBlock() const {
|
||||
return dynamic_cast<BasicBlock *>(getOperand(2));
|
||||
}
|
||||
std::vector<BasicBlock *> getSuccessors() const {
|
||||
std::vector<BasicBlock *> succs;
|
||||
// 假设条件分支的真实块是第二个操作数,假块是第三个操作数
|
||||
// 操作数通常是:[0] 条件值, [1] TrueTargetBlock, [2] FalseTargetBlock
|
||||
if (getNumOperands() > 2) {
|
||||
if (auto true_bb = getThenBlock()) {
|
||||
succs.push_back(true_bb);
|
||||
}
|
||||
if (auto false_bb = getElseBlock()) {
|
||||
succs.push_back(false_bb);
|
||||
}
|
||||
}
|
||||
return succs;
|
||||
}
|
||||
// auto getThenArguments() const {
|
||||
// auto begin = std::next(operand_begin(), 3);
|
||||
// // auto end = std::next(begin, getThenBlock()->getNumArguments());
|
||||
// return make_range(begin, end);
|
||||
// }
|
||||
// auto getElseArguments() const {
|
||||
// auto begin =
|
||||
// std::next(operand_begin(), 3 + getThenBlock()->getNumArguments());
|
||||
// auto end = operand_end();
|
||||
// return make_range(begin, end);
|
||||
// }
|
||||
|
||||
}; // class CondBrInst
|
||||
|
||||
class UnreachableInst : public Instruction {
|
||||
public:
|
||||
// 构造函数:设置指令类型为 kUnreachable
|
||||
explicit UnreachableInst(const std::string& name, BasicBlock *parent = nullptr)
|
||||
: Instruction(kUnreachable, Type::getVoidType(), parent, "") {}
|
||||
|
||||
};
|
||||
|
||||
//! Allocate memory for stack variables, used for non-global variable declartion
|
||||
class AllocaInst : public Instruction {
|
||||
friend class IRBuilder;
|
||||
friend class Function;
|
||||
protected:
|
||||
AllocaInst(Type *type,
|
||||
AllocaInst(Type *type, const std::vector<Value *> &dims = {},
|
||||
BasicBlock *parent = nullptr, const std::string &name = "")
|
||||
: Instruction(kAlloca, type, parent, name) {
|
||||
addOperands(dims);
|
||||
}
|
||||
|
||||
public:
|
||||
@@ -1186,6 +1133,9 @@ public:
|
||||
Type* getAllocatedType() const {
|
||||
return getType()->as<PointerType>()->getBaseType();
|
||||
} ///< 获取分配的类型
|
||||
int getNumDims() const { return getNumOperands(); }
|
||||
auto getDims() const { return getOperands(); }
|
||||
Value* getDim(int index) { return getOperand(index); }
|
||||
|
||||
}; // class AllocaInst
|
||||
|
||||
@@ -1232,15 +1182,21 @@ class LoadInst : public Instruction {
|
||||
friend class Function;
|
||||
|
||||
protected:
|
||||
LoadInst(Value *pointer,
|
||||
LoadInst(Value *pointer, const std::vector<Value *> &indices = {},
|
||||
BasicBlock *parent = nullptr, const std::string &name = "")
|
||||
: Instruction(kLoad, pointer->getType()->as<PointerType>()->getBaseType(),
|
||||
parent, name) {
|
||||
addOperand(pointer);
|
||||
addOperands(indices);
|
||||
}
|
||||
|
||||
public:
|
||||
int getNumIndices() const { return getNumOperands() - 1; }
|
||||
Value* getPointer() const { return getOperand(0); }
|
||||
auto getIndices() const {
|
||||
return make_range(std::next(operand_begin()), operand_end());
|
||||
}
|
||||
Value* getIndex(int index) const { return getOperand(index + 1); }
|
||||
|
||||
}; // class LoadInst
|
||||
|
||||
@@ -1251,15 +1207,22 @@ class StoreInst : public Instruction {
|
||||
|
||||
protected:
|
||||
StoreInst(Value *value, Value *pointer,
|
||||
const std::vector<Value *> &indices = {},
|
||||
BasicBlock *parent = nullptr, const std::string &name = "")
|
||||
: Instruction(kStore, Type::getVoidType(), parent, name) {
|
||||
addOperand(value);
|
||||
addOperand(pointer);
|
||||
addOperands(indices);
|
||||
}
|
||||
|
||||
public:
|
||||
int getNumIndices() const { return getNumOperands() - 2; }
|
||||
Value* getValue() const { return getOperand(0); }
|
||||
Value* getPointer() const { return getOperand(1); }
|
||||
auto getIndices() const {
|
||||
return make_range(std::next(operand_begin(), 2), operand_end());
|
||||
}
|
||||
Value* getIndex(int index) const { return getOperand(index + 2); }
|
||||
|
||||
}; // class StoreInst
|
||||
|
||||
@@ -1398,18 +1361,20 @@ protected:
|
||||
|
||||
protected:
|
||||
GlobalValue(Module *parent, Type *type, const std::string &name,
|
||||
const std::vector<Value *> &dims = {},
|
||||
ValueCounter init = {})
|
||||
: Value(type, name), parent(parent) {
|
||||
assert(type->isPointer());
|
||||
// addOperands(dims);
|
||||
// 维度信息已经被记录到Type中,dim只是为了方便初始化
|
||||
numDims = 0;
|
||||
numDims = dims.size();
|
||||
if (init.size() == 0) {
|
||||
unsigned num = 1;
|
||||
auto arrayType = type->as<ArrayType>();
|
||||
while (arrayType) {
|
||||
numDims++;
|
||||
num *= arrayType->getNumElements();
|
||||
arrayType = arrayType->getElementType()->as<ArrayType>();
|
||||
for (unsigned i = 0; i < numDims; i++) {
|
||||
// Assume dims elements are ConstantInteger and cast appropriately
|
||||
auto dim_val = dynamic_cast<ConstantInteger*>(dims[i]);
|
||||
assert(dim_val && "GlobalValue dims must be constant integers");
|
||||
num *= dim_val->getInt();
|
||||
}
|
||||
if (dynamic_cast<PointerType *>(type)->getBaseType() == Type::getFloatType()) {
|
||||
init.push_back(ConstantFloating::get(0.0F), num); // Use new constant factory
|
||||
@@ -1421,6 +1386,9 @@ protected:
|
||||
}
|
||||
|
||||
public:
|
||||
// unsigned getNumDims() const { return numDims; } ///< 获取维度数量
|
||||
// Value* getDim(unsigned index) const { return getOperand(index); } ///< 获取位置为index的维度
|
||||
// auto getDims() const { return getOperands(); } ///< 获取维度列表
|
||||
unsigned getNumIndices() const {
|
||||
return numDims;
|
||||
} ///< 获取维度数量
|
||||
@@ -1462,19 +1430,13 @@ class ConstantVariable : public Value {
|
||||
ValueCounter initValues; ///< 值
|
||||
|
||||
protected:
|
||||
ConstantVariable(Module *parent, Type *type, const std::string &name, const ValueCounter &init)
|
||||
ConstantVariable(Module *parent, Type *type, const std::string &name, const ValueCounter &init,
|
||||
const std::vector<Value *> &dims = {})
|
||||
: Value(type, name), parent(parent) {
|
||||
assert(type->isPointer());
|
||||
// numDims = dims.size();
|
||||
numDims = 0;
|
||||
if(type->as<PointerType>()->getBaseType()->isArray()) {
|
||||
auto arrayType = type->as<ArrayType>();
|
||||
while (arrayType) {
|
||||
numDims++;
|
||||
arrayType = arrayType->getElementType()->as<ArrayType>();
|
||||
}
|
||||
}
|
||||
numDims = dims.size();
|
||||
initValues = init;
|
||||
// addOperands(dims); 同GlobalValue,维度信息已经被记录到Type中,dim只是为了方便初始化
|
||||
}
|
||||
|
||||
public:
|
||||
@@ -1506,6 +1468,9 @@ class ConstantVariable : public Value {
|
||||
|
||||
return getByIndex(index);
|
||||
} ///< 通过多维索引indices获取初始值
|
||||
// unsigned getNumDims() const { return numDims; } ///< 获取维度数量
|
||||
// Value* getDim(unsigned index) const { return getOperand(index); } ///< 获取位置为index的维度
|
||||
// auto getDims() const { return getOperands(); } ///< 获取维度列表
|
||||
const ValueCounter& getInitValues() const { return initValues; } ///< 获取初始值
|
||||
};
|
||||
|
||||
@@ -1564,12 +1529,13 @@ class Module {
|
||||
return result.first->second.get();
|
||||
} ///< 创建外部函数
|
||||
///< 变量创建伴随着符号表的更新
|
||||
GlobalValue* createGlobalValue(const std::string &name, Type *type, const ValueCounter &init = {}) {
|
||||
GlobalValue* createGlobalValue(const std::string &name, Type *type, const std::vector<Value *> &dims = {},
|
||||
const ValueCounter &init = {}) {
|
||||
bool isFinished = variableTable.isCurNodeNull();
|
||||
if (isFinished) {
|
||||
variableTable.enterGlobalScope();
|
||||
}
|
||||
auto result = variableTable.addVariable(name, new GlobalValue(this, type, name, init));
|
||||
auto result = variableTable.addVariable(name, new GlobalValue(this, type, name, dims, init));
|
||||
if (isFinished) {
|
||||
variableTable.leaveScope();
|
||||
}
|
||||
@@ -1578,8 +1544,9 @@ class Module {
|
||||
}
|
||||
return dynamic_cast<GlobalValue *>(result);
|
||||
} ///< 创建全局变量
|
||||
ConstantVariable* createConstVar(const std::string &name, Type *type, const ValueCounter &init) {
|
||||
auto result = variableTable.addVariable(name, new ConstantVariable(this, type, name, init));
|
||||
ConstantVariable* createConstVar(const std::string &name, Type *type, const ValueCounter &init,
|
||||
const std::vector<Value *> &dims = {}) {
|
||||
auto result = variableTable.addVariable(name, new ConstantVariable(this, type, name, init, dims));
|
||||
if (result == nullptr) {
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
@@ -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()) {
|
||||
@@ -280,7 +275,7 @@ 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;
|
||||
@@ -291,8 +286,9 @@ class IRBuilder {
|
||||
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;
|
||||
@@ -312,6 +308,24 @@ class IRBuilder {
|
||||
block->getInstructions().emplace(block->begin(), inst);
|
||||
return inst;
|
||||
} ///< 创建Phi指令
|
||||
// GetElementPtrInst* createGetElementPtrInst(Value *basePointer,
|
||||
// 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 GetElementPtrInst(basePointer, indices, block, newName);
|
||||
// assert(inst);
|
||||
// block->getInstructions().emplace(position, inst);
|
||||
// return inst;
|
||||
// }
|
||||
/**
|
||||
* @brief 根据 LLVM 设计模式创建 GEP 指令。
|
||||
* 它会自动推断返回类型,无需手动指定。
|
||||
|
||||
@@ -6,82 +6,30 @@
|
||||
#include <set>
|
||||
#include <vector>
|
||||
#include <algorithm>
|
||||
#include <functional>
|
||||
|
||||
namespace sysy {
|
||||
|
||||
// 支配树分析结果类
|
||||
// 支配树分析结果类 (保持不变)
|
||||
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
|
||||
BasicBlock* getImmediateDominator(BasicBlock* BB) const;
|
||||
const std::set<BasicBlock*>* getDominanceFrontier(BasicBlock* BB) const;
|
||||
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 computeIDoms(Function* F);
|
||||
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);
|
||||
std::map<BasicBlock*, std::set<BasicBlock*>> Dominators;
|
||||
std::map<BasicBlock*, BasicBlock*> IDoms;
|
||||
std::map<BasicBlock*, std::set<BasicBlock*>> DominanceFrontiers;
|
||||
std::map<BasicBlock*, std::set<BasicBlock*>> DominatorTreeChildren;
|
||||
};
|
||||
|
||||
|
||||
|
||||
@@ -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
|
||||
14
src/include/midend/Pass/Optimize/ConstPropagation.h
Normal file
14
src/include/midend/Pass/Optimize/ConstPropagation.h
Normal file
@@ -0,0 +1,14 @@
|
||||
#pragma once
|
||||
|
||||
#include "Pass.h"
|
||||
|
||||
namespace sysy {
|
||||
|
||||
class ConstPropagation : public OptimizationPass {
|
||||
public:
|
||||
ConstPropagation() : OptimizationPass("ConstPropagation", Granularity::Function) {}
|
||||
bool runOnFunction(Function *F, AnalysisManager& AM) override;
|
||||
static char ID;
|
||||
};
|
||||
|
||||
} // namespace sysy
|
||||
@@ -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
|
||||
@@ -75,7 +75,11 @@ private:
|
||||
// --------------------------------------------------------------------
|
||||
|
||||
// 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);
|
||||
// alloca: 当前正在处理的 AllocaInst
|
||||
// currentBB: 当前正在遍历的基本块
|
||||
// dt: 支配树分析结果
|
||||
// valueStack: 存储当前 AllocaInst 在当前路径上可见的 SSA 值栈
|
||||
void renameVariables(AllocaInst* alloca, BasicBlock* currentBB);
|
||||
|
||||
// --------------------------------------------------------------------
|
||||
// 阶段4: 清理
|
||||
|
||||
@@ -1,139 +1,196 @@
|
||||
#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)
|
||||
// 稀疏条件常量传播类
|
||||
// Sparse Conditional Constant Propagation
|
||||
/*
|
||||
伪代码
|
||||
function SCCP_Optimization(Module):
|
||||
for each Function in Module:
|
||||
changed = true
|
||||
while changed:
|
||||
changed = false
|
||||
// 阶段1: 常量传播与折叠
|
||||
changed |= PropagateConstants(Function)
|
||||
// 阶段2: 控制流简化
|
||||
changed |= SimplifyControlFlow(Function)
|
||||
end while
|
||||
end for
|
||||
|
||||
function PropagateConstants(Function):
|
||||
// 初始化
|
||||
executableBlocks = {entryBlock}
|
||||
valueState = map<Value, State> // 值->状态映射
|
||||
instWorkList = Queue()
|
||||
edgeWorkList = Queue()
|
||||
|
||||
// 初始化工作列表
|
||||
for each inst in entryBlock:
|
||||
instWorkList.push(inst)
|
||||
|
||||
// 迭代处理
|
||||
while !instWorkList.empty() || !edgeWorkList.empty():
|
||||
// 处理指令工作列表
|
||||
while !instWorkList.empty():
|
||||
inst = instWorkList.pop()
|
||||
// 如果指令是可执行基本块中的
|
||||
if executableBlocks.contains(inst.parent):
|
||||
ProcessInstruction(inst)
|
||||
|
||||
// 处理边工作列表
|
||||
while !edgeWorkList.empty():
|
||||
edge = edgeWorkList.pop()
|
||||
ProcessEdge(edge)
|
||||
|
||||
// 应用常量替换
|
||||
for each inst in Function:
|
||||
if valueState[inst] == CONSTANT:
|
||||
ReplaceWithConstant(inst, valueState[inst].constant)
|
||||
changed = true
|
||||
|
||||
return changed
|
||||
|
||||
function ProcessInstruction(Instruction inst):
|
||||
switch inst.type:
|
||||
//二元操作
|
||||
case BINARY_OP:
|
||||
lhs = GetValueState(inst.operands[0])
|
||||
rhs = GetValueState(inst.operands[1])
|
||||
if lhs == CONSTANT && rhs == CONSTANT:
|
||||
newState = ComputeConstant(inst.op, lhs.value, rhs.value)
|
||||
UpdateState(inst, newState)
|
||||
else if lhs == BOTTOM || rhs == BOTTOM:
|
||||
UpdateState(inst, BOTTOM)
|
||||
//phi
|
||||
case PHI:
|
||||
mergedState = ⊤
|
||||
for each incoming in inst.incomings:
|
||||
// 检查每个输入的状态
|
||||
if executableBlocks.contains(incoming.block):
|
||||
incomingState = GetValueState(incoming.value)
|
||||
mergedState = Meet(mergedState, incomingState)
|
||||
UpdateState(inst, mergedState)
|
||||
// 条件分支
|
||||
case COND_BRANCH:
|
||||
cond = GetValueState(inst.condition)
|
||||
if cond == CONSTANT:
|
||||
// 判断条件分支
|
||||
if cond.value == true:
|
||||
AddEdgeToWorkList(inst.parent, inst.trueTarget)
|
||||
else:
|
||||
AddEdgeToWorkList(inst.parent, inst.falseTarget)
|
||||
else if cond == BOTTOM:
|
||||
AddEdgeToWorkList(inst.parent, inst.trueTarget)
|
||||
AddEdgeToWorkList(inst.parent, inst.falseTarget)
|
||||
|
||||
case UNCOND_BRANCH:
|
||||
AddEdgeToWorkList(inst.parent, inst.target)
|
||||
|
||||
// 其他指令处理...
|
||||
|
||||
function ProcessEdge(Edge edge):
|
||||
fromBB, toBB = edge
|
||||
if !executableBlocks.contains(toBB):
|
||||
executableBlocks.add(toBB)
|
||||
for each inst in toBB:
|
||||
if inst is PHI:
|
||||
instWorkList.push(inst)
|
||||
else:
|
||||
instWorkList.push(inst) // 非PHI指令
|
||||
|
||||
// 更新PHI节点的输入
|
||||
for each phi in toBB.phis:
|
||||
instWorkList.push(phi)
|
||||
|
||||
function SimplifyControlFlow(Function):
|
||||
changed = false
|
||||
// 标记可达基本块
|
||||
ReachableBBs = FindReachableBlocks(Function.entry)
|
||||
|
||||
// 删除不可达块
|
||||
for each bb in Function.blocks:
|
||||
if !ReachableBBs.contains(bb):
|
||||
RemoveDeadBlock(bb)
|
||||
changed = true
|
||||
|
||||
// 简化条件分支
|
||||
for each bb in Function.blocks:
|
||||
terminator = bb.terminator
|
||||
if terminator is COND_BRANCH:
|
||||
cond = GetValueState(terminator.condition)
|
||||
if cond == CONSTANT:
|
||||
SimplifyBranch(terminator, cond.value)
|
||||
changed = true
|
||||
|
||||
return changed
|
||||
|
||||
function RemoveDeadBlock(BasicBlock bb):
|
||||
// 1. 更新前驱块的分支指令
|
||||
for each pred in bb.predecessors:
|
||||
UpdateTerminator(pred, bb)
|
||||
|
||||
// 2. 更新后继块的PHI节点
|
||||
for each succ in bb.successors:
|
||||
RemovePhiIncoming(succ, bb)
|
||||
|
||||
// 3. 删除块内所有指令
|
||||
for each inst in bb.instructions:
|
||||
inst.remove()
|
||||
|
||||
// 4. 从函数中移除基本块
|
||||
Function.removeBlock(bb)
|
||||
|
||||
function Meet(State a, State b):
|
||||
if a == ⊤: return b
|
||||
if b == ⊤: return a
|
||||
if a == ⊥ || b == ⊥: return ⊥
|
||||
if a.value == b.value: return a
|
||||
return ⊥
|
||||
|
||||
function UpdateState(Value v, State newState):
|
||||
oldState = valueState.get(v, ⊤)
|
||||
if newState != oldState:
|
||||
valueState[v] = newState
|
||||
for each user in v.users:
|
||||
if user is Instruction:
|
||||
instWorkList.push(user)
|
||||
|
||||
*/
|
||||
|
||||
enum class LatticeValue {
|
||||
Top, // ⊤ (Unknown)
|
||||
Constant, // c (Constant)
|
||||
Bottom // ⊥ (Undefined / Varying)
|
||||
};
|
||||
// LatticeValue: 用于表示值的状态,Top表示未知,Constant表示常量,Bottom表示未定义或变化的值。
|
||||
// 这里的LatticeValue用于跟踪每个SSA值(变量、指令结果)的状态,
|
||||
// 以便在SCCP过程中进行常量传播和控制流简化。
|
||||
|
||||
// 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
|
||||
};
|
||||
//TODO: 下列数据结构考虑集成到类中,避免重命名问题
|
||||
static std::set<Instruction *> Worklist;
|
||||
static std::unordered_set<BasicBlock*> Executable_Blocks;
|
||||
static std::queue<std::pair<BasicBlock *, BasicBlock *> > Executable_Edges;
|
||||
static std::map<Value*, LatticeValue> valueState;
|
||||
|
||||
// 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 {
|
||||
class SCCP {
|
||||
private:
|
||||
IRBuilder *builder; // IR 构建器,用于插入指令和创建常量
|
||||
Module *pModule;
|
||||
|
||||
// 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;
|
||||
public:
|
||||
SCCP(Module *pMoudle) : pModule(pMoudle) {}
|
||||
|
||||
// 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 run();
|
||||
bool PropagateConstants(Function *function);
|
||||
bool SimplifyControlFlow(Function *function);
|
||||
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);
|
||||
void RemoveDeadBlock(BasicBlock *bb);
|
||||
void UpdateState(Value *v, LatticeValue newState);
|
||||
LatticeValue Meet(LatticeValue a, LatticeValue b);
|
||||
LatticeValue GetValueState(Value *v);
|
||||
};
|
||||
|
||||
// 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
|
||||
} // namespace sysy
|
||||
|
||||
@@ -2,7 +2,6 @@
|
||||
|
||||
#include "IR.h"
|
||||
|
||||
extern int DEBUG;
|
||||
namespace sysy {
|
||||
|
||||
// 优化工具类,包含一些通用的优化方法
|
||||
@@ -11,80 +10,12 @@ namespace sysy {
|
||||
class SysYIROptUtils{
|
||||
|
||||
public:
|
||||
struct PairHash {
|
||||
template <class T1, class T2>
|
||||
std::size_t operator () (const std::pair<T1, T2>& p) const {
|
||||
auto h1 = std::hash<T1>{}(p.first);
|
||||
auto h2 = std::hash<T2>{}(p.second);
|
||||
|
||||
// 简单的组合哈希值,可以更复杂以减少冲突
|
||||
// 使用 boost::hash_combine 的简化版本
|
||||
return h1 ^ (h2 << 1);
|
||||
}
|
||||
};
|
||||
|
||||
static void RemoveUserOperandUses(User *user) {
|
||||
if (!user) {
|
||||
return;
|
||||
// 仅仅删除use关系
|
||||
static void usedelete(Instruction *instr) {
|
||||
for (auto &use : instr->getOperands()) {
|
||||
Value* val = use->getValue();
|
||||
val->removeUse(use);
|
||||
}
|
||||
|
||||
// 遍历 User 的 operands 列表。
|
||||
// 由于 operands 是 protected 成员,我们需要一个临时方法来访问它,
|
||||
// 或者在 User 类中添加一个 friend 声明。
|
||||
// 假设 User 内部有一个像 getOperands() 这样的公共方法返回 operands 的引用,
|
||||
// 或者将 SysYIROptUtils 声明为 User 的 friend。
|
||||
// 为了示例,我将假设可以直接访问 user->operands 或通过一个getter。
|
||||
// 如果无法直接访问,请在 IR.h 的 User 类中添加:
|
||||
// public: const std::vector<std::shared_ptr<Use>>& getOperands() const { return operands; }
|
||||
|
||||
// 迭代 copies of shared_ptr to avoid issues if removeUse modifies the list
|
||||
// (though remove should handle it, iterating a copy is safer or reverse iteration).
|
||||
// Since we'll clear the vector at the end, iterating forward is fine.
|
||||
for (const auto& use_ptr : user->getOperands()) { // 假设 getOperands() 可用
|
||||
if (use_ptr) {
|
||||
Value *val = use_ptr->getValue(); // 获取 Use 指向的 Value (如 AllocaInst)
|
||||
if (val) {
|
||||
val->removeUse(use_ptr); // 通知 Value 从其 uses 列表中移除此 Use 关系
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
static void usedelete(Instruction *inst) {
|
||||
assert(inst && "Instruction to delete cannot be null.");
|
||||
BasicBlock *parentBlock = inst->getParent();
|
||||
assert(parentBlock && "Instruction must have a parent BasicBlock to be deleted.");
|
||||
|
||||
// 步骤1: 处理所有使用者,将他们从使用 inst 变为使用 UndefinedValue
|
||||
// 这将清理 inst 作为 Value 时的 uses 列表
|
||||
if (!inst->getUses().empty()) {
|
||||
inst->replaceAllUsesWith(UndefinedValue::get(inst->getType()));
|
||||
}
|
||||
|
||||
// 步骤2: 清理 inst 作为 User 时的操作数关系
|
||||
// 通知 inst 所使用的所有 Value (如 AllocaInst),移除对应的 Use 关系。
|
||||
// 这里的 inst 实际上是一个 User*,所以可以安全地向下转型。
|
||||
RemoveUserOperandUses(static_cast<User*>(inst));
|
||||
|
||||
// 步骤3: 物理删除指令
|
||||
// 这会导致 Instruction 对象的 unique_ptr 销毁,从而调用其析构函数链。
|
||||
parentBlock->removeInst(inst);
|
||||
}
|
||||
|
||||
static BasicBlock::iterator usedelete(BasicBlock::iterator inst_it) {
|
||||
Instruction *inst_to_delete = inst_it->get();
|
||||
BasicBlock *parentBlock = inst_to_delete->getParent();
|
||||
assert(parentBlock && "Instruction must have a parent BasicBlock for iterator deletion.");
|
||||
|
||||
// 步骤1: 处理所有使用者
|
||||
if (!inst_to_delete->getUses().empty()) {
|
||||
inst_to_delete->replaceAllUsesWith(UndefinedValue::get(inst_to_delete->getType()));
|
||||
}
|
||||
|
||||
// 步骤2: 清理操作数关系
|
||||
RemoveUserOperandUses(static_cast<User*>(inst_to_delete));
|
||||
|
||||
// 步骤3: 物理删除指令并返回下一个迭代器
|
||||
return parentBlock->removeInst(inst_it);
|
||||
}
|
||||
|
||||
// 判断是否是全局变量
|
||||
@@ -95,17 +26,7 @@ public:
|
||||
// 判断是否是数组
|
||||
static bool isArr(Value *val) {
|
||||
auto aval = dynamic_cast<AllocaInst *>(val);
|
||||
// 如果是 AllocaInst 且通过Type::isArray()判断为数组类型
|
||||
return aval && aval->getType()->as<PointerType>()->getBaseType()->isArray();
|
||||
}
|
||||
// 判断是否是指向数组的指针
|
||||
static bool isArrPointer(Value *val) {
|
||||
auto aval = dynamic_cast<AllocaInst *>(val);
|
||||
// 如果是 AllocaInst 且通过Type::isPointer()判断为指针;
|
||||
auto baseType = aval->getType()->as<PointerType>()->getBaseType();
|
||||
// 在sysy中,函数的数组参数会退化成指针
|
||||
// 所以当AllocaInst的basetype是PointerType时(一维数组)或者是指向ArrayType的PointerType(多位数组)时,返回true
|
||||
return aval && (baseType->isPointer() || baseType->as<PointerType>()->getBaseType()->isArray());
|
||||
return aval != nullptr && aval->getNumDims() != 0;
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
@@ -279,7 +279,7 @@ private:
|
||||
IRBuilder *pBuilder;
|
||||
|
||||
public:
|
||||
PassManager() = delete;
|
||||
PassManager() = default;
|
||||
~PassManager() = default;
|
||||
|
||||
PassManager(Module *module, IRBuilder *builder) : pmodule(module) ,pBuilder(builder), analysisManager(module) {}
|
||||
|
||||
@@ -86,60 +86,7 @@ private:
|
||||
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;
|
||||
@@ -220,15 +167,6 @@ public:
|
||||
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);
|
||||
|
||||
@@ -10,9 +10,7 @@ add_library(midend_lib STATIC
|
||||
Pass/Optimize/Mem2Reg.cpp
|
||||
Pass/Optimize/Reg2Mem.cpp
|
||||
Pass/Optimize/SysYIRCFGOpt.cpp
|
||||
Pass/Optimize/SCCP.cpp
|
||||
Pass/Optimize/BuildCFG.cpp
|
||||
Pass/Optimize/LargeArrayToGlobal.cpp
|
||||
Pass/Optimize/ConstPropagation.cpp
|
||||
)
|
||||
|
||||
# 包含中端模块所需的头文件路径
|
||||
|
||||
@@ -118,46 +118,14 @@ ArrayType *ArrayType::get(Type *elementType, unsigned numElements) {
|
||||
return result.first->get();
|
||||
}
|
||||
|
||||
// void Value::replaceAllUsesWith(Value *value) {
|
||||
// for (auto &use : uses) {
|
||||
// auto user = use->getUser();
|
||||
// assert(user && "Use's user cannot be null");
|
||||
// user->setOperand(use->getIndex(), value);
|
||||
// }
|
||||
// uses.clear();
|
||||
// }
|
||||
void Value::replaceAllUsesWith(Value *value) {
|
||||
// 1. 创建 uses 列表的副本进行迭代。
|
||||
// 这样做是为了避免在迭代过程中,由于 setOperand 间接调用 removeUse 或 addUse
|
||||
// 导致原列表被修改,从而引发迭代器失效问题。
|
||||
std::list<std::shared_ptr<Use>> uses_copy = uses;
|
||||
|
||||
for (auto &use_ptr : uses_copy) { // 遍历副本
|
||||
// 2. 检查 shared_ptr<Use> 本身是否为空。这是最常见的崩溃原因之一。
|
||||
if (use_ptr == nullptr) {
|
||||
std::cerr << "Warning: Encountered a null std::shared_ptr<Use> in Value::uses list. Skipping this entry." << std::endl;
|
||||
// 在一个健康的 IR 中,这种情况不应该发生。如果经常出现,说明你的 Use 创建或管理有问题。
|
||||
continue; // 跳过空的智能指针
|
||||
}
|
||||
|
||||
// 3. 检查 Use 对象内部的 User* 是否为空。
|
||||
User* user_val = use_ptr->getUser();
|
||||
if (user_val == nullptr) {
|
||||
std::cerr << "Warning: Use object (" << use_ptr.get() << ") has a null User* in replaceAllUsesWith. Skipping this entry. This indicates IR corruption." << std::endl;
|
||||
// 同样,在一个健康的 IR 中,Use 对象的 User* 不应该为空。
|
||||
continue; // 跳过用户指针为空的 Use 对象
|
||||
}
|
||||
|
||||
// 如果走到这里,use_ptr 和 user_val 都是有效的,可以安全调用 setOperand
|
||||
user_val->setOperand(use_ptr->getIndex(), value);
|
||||
for (auto &use : uses) {
|
||||
use->getUser()->setOperand(use->getIndex(), value);
|
||||
}
|
||||
|
||||
// 4. 处理完所有 use 之后,清空原始的 uses 列表。
|
||||
// replaceAllUsesWith 的目的就是将所有使用关系从当前 Value 转移走,
|
||||
// 所以最后清空列表是正确的。
|
||||
uses.clear();
|
||||
uses.clear();
|
||||
}
|
||||
|
||||
|
||||
// Implementations for static members
|
||||
|
||||
std::unordered_map<ConstantValueKey, ConstantValue*, ConstantValueHash, ConstantValueEqual> ConstantValue::mConstantPool;
|
||||
@@ -214,45 +182,373 @@ auto Function::getCalleesWithNoExternalAndSelf() -> std::set<Function *> {
|
||||
}
|
||||
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++;
|
||||
}
|
||||
|
||||
void Value::removeAllUses() {
|
||||
while (!uses.empty()) {
|
||||
auto use = uses.back();
|
||||
uses.pop_back();
|
||||
if (use && use->getUser()) {
|
||||
auto user = use->getUser();
|
||||
int index = use->getIndex();
|
||||
user->removeOperand(index); ///< 从User中移除该操作数
|
||||
} else {
|
||||
// 如果use或user为null,输出警告信息
|
||||
assert(use != nullptr && "Use cannot be null");
|
||||
assert(use->getUser() != nullptr && "Use's user cannot be null");
|
||||
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));
|
||||
}
|
||||
}
|
||||
uses.clear();
|
||||
}
|
||||
|
||||
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;
|
||||
}
|
||||
|
||||
// TODO:复制GEP指令
|
||||
|
||||
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 ¶m : blocks.front()->getArguments()) {
|
||||
// newFunction->getEntryBlock()->insertArgument(dynamic_cast<AllocaInst *>(oldNewValueMap.at(param)));
|
||||
// }
|
||||
for (const auto &arg : arguments) {
|
||||
auto newArg = dynamic_cast<Argument *>(oldNewValueMap.at(arg));
|
||||
if (newArg != nullptr) {
|
||||
newFunction->insertArgument(newArg);
|
||||
}
|
||||
}
|
||||
|
||||
return newFunction;
|
||||
}
|
||||
/**
|
||||
* 设置操作数
|
||||
*/
|
||||
void User::setOperand(unsigned index, Value *newvalue) {
|
||||
if (index >= operands.size()) {
|
||||
std::cerr << "index=" << index << ", but mOperands max size=" << operands.size() << std::endl;
|
||||
assert(index < operands.size());
|
||||
}
|
||||
std::shared_ptr<Use> olduse = operands[index];
|
||||
Value *oldValue = olduse->getValue();
|
||||
if (oldValue != newvalue) {
|
||||
// 如果新值和旧值不同,先移除旧值的使用关系
|
||||
oldValue->removeUse(olduse);
|
||||
// 设置新的操作数
|
||||
operands[index] = std::make_shared<Use>(index, this, newvalue);
|
||||
newvalue->addUse(operands[index]);
|
||||
}
|
||||
else {
|
||||
// 如果新值和旧值相同,直接更新use的索引
|
||||
operands[index]->setValue(newvalue);
|
||||
}
|
||||
void User::setOperand(unsigned index, Value *value) {
|
||||
assert(index < getNumOperands());
|
||||
operands[index]->setValue(value);
|
||||
value->addUse(operands[index]);
|
||||
}
|
||||
/**
|
||||
* 替换操作数
|
||||
@@ -265,128 +561,81 @@ void User::replaceOperand(unsigned index, Value *value) {
|
||||
value->addUse(use);
|
||||
}
|
||||
|
||||
/**
|
||||
* 移除操作数
|
||||
*/
|
||||
void User::removeOperand(unsigned index) {
|
||||
assert(index < getNumOperands() && "Index out of range in removeOperand");
|
||||
std::shared_ptr<Use> useToRemove = operands.at(index);
|
||||
Value *valueToRemove = useToRemove->getValue();
|
||||
if(valueToRemove) {
|
||||
if(valueToRemove == this) {
|
||||
std::cerr << "Cannot remove operand that is the same as the User itself." << std::endl;
|
||||
}
|
||||
valueToRemove->removeUse(useToRemove);
|
||||
}
|
||||
operands.erase(operands.begin() + index);
|
||||
unsigned newIndex = 0;
|
||||
for(auto it = operands.begin(); it != operands.end(); ++it, ++newIndex) {
|
||||
(*it)->setIndex(newIndex); // 更新剩余操作数的索引
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* phi相关函数
|
||||
*/
|
||||
|
||||
Value* PhiInst::getValfromBlk(BasicBlock* blk) {
|
||||
refreshMap();
|
||||
Value* PhiInst::getvalfromBlk(BasicBlock* blk){
|
||||
refreshB2VMap();
|
||||
if( blk2val.find(blk) != blk2val.end()) {
|
||||
return blk2val.at(blk);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
BasicBlock* PhiInst::getBlkfromVal(Value* val) {
|
||||
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(getIncomingValue(i) == val) {
|
||||
return getIncomingBlock(i);
|
||||
if(getValue(i) == val) {
|
||||
return getBlock(i);
|
||||
}
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
void PhiInst::removeIncomingValue(Value* val){
|
||||
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);
|
||||
if(blk == nullptr) {
|
||||
return; // 如果val没有对应的基本块,直接返回
|
||||
}
|
||||
for(i = 0; i < vsize; i++) {
|
||||
if(getIncomingValue(i) == val) {
|
||||
if(getValue(i) == val) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
removeIncoming(i);
|
||||
removeOperand(2 * i + 1); // 删除blk
|
||||
removeOperand(2 * i); // 删除val
|
||||
vsize--;
|
||||
blk2val.erase(blk); // 删除blk2val映射
|
||||
}
|
||||
|
||||
void PhiInst::removeIncomingBlock(BasicBlock* blk){
|
||||
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);
|
||||
if(val == nullptr) {
|
||||
return; // 如果blk没有对应的值,直接返回
|
||||
}
|
||||
Value* val = getvalfromBlk(blk);
|
||||
for(i = 0; i < vsize; i++) {
|
||||
if(getIncomingBlock(i) == blk) {
|
||||
if(getBlock(i) == blk) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
removeIncoming(i);
|
||||
removeOperand(2 * i + 1); // 删除blk
|
||||
removeOperand(2 * i); // 删除val
|
||||
vsize--;
|
||||
blk2val.erase(blk); // 删除blk2val映射
|
||||
}
|
||||
|
||||
void PhiInst::setIncomingValue(unsigned k, Value* val) {
|
||||
assert(k < vsize && "PhiInst: index out of range");
|
||||
assert(val != nullptr && "PhiInst: value cannot be null");
|
||||
refreshMap();
|
||||
blk2val.erase(getIncomingBlock(k));
|
||||
setOperand(2 * k, val);
|
||||
blk2val[getIncomingBlock(k)] = val;
|
||||
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::setIncomingBlock(unsigned k, BasicBlock* blk) {
|
||||
assert(k < vsize && "PhiInst: index out of range");
|
||||
assert(blk != nullptr && "PhiInst: block cannot be null");
|
||||
refreshMap();
|
||||
auto oldVal = getIncomingValue(k);
|
||||
blk2val.erase(getIncomingBlock(k));
|
||||
setOperand(2 * k + 1, blk);
|
||||
blk2val[blk] = oldVal;
|
||||
}
|
||||
|
||||
void PhiInst::replaceIncomingValue(Value* newVal, Value* oldVal) {
|
||||
refreshMap();
|
||||
assert(blk2val.find(getBlkfromVal(oldVal)) != blk2val.end() && "PhiInst: oldVal not found in blk2val");
|
||||
auto blk = getBlkfromVal(oldVal);
|
||||
removeIncomingValue(oldVal);
|
||||
addIncoming(newVal, blk);
|
||||
}
|
||||
|
||||
void PhiInst::replaceIncomingBlock(BasicBlock* newBlk, BasicBlock* oldBlk) {
|
||||
refreshMap();
|
||||
assert(blk2val.find(oldBlk) != blk2val.end() && "PhiInst: oldBlk not found in blk2val");
|
||||
auto val = blk2val[oldBlk];
|
||||
removeIncomingBlock(oldBlk);
|
||||
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::replaceIncomingValue(Value *oldValue, Value *newValue, BasicBlock *newBlock) {
|
||||
refreshMap();
|
||||
assert(blk2val.find(getBlkfromVal(oldValue)) != blk2val.end() && "PhiInst: oldValue not found in blk2val");
|
||||
auto oldBlock = getBlkfromVal(oldValue);
|
||||
removeIncomingValue(oldValue);
|
||||
addIncoming(newValue, newBlock);
|
||||
}
|
||||
|
||||
void PhiInst::replaceIncomingBlock(BasicBlock *oldBlock, BasicBlock *newBlock, Value *newValue) {
|
||||
refreshMap();
|
||||
assert(blk2val.find(oldBlock) != blk2val.end() && "PhiInst: oldBlock not found in blk2val");
|
||||
auto oldValue = blk2val[oldBlock];
|
||||
removeIncomingBlock(oldBlock);
|
||||
addIncoming(newValue, newBlock);
|
||||
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)
|
||||
|
||||
@@ -1,30 +1,19 @@
|
||||
#include "Dom.h"
|
||||
#include <algorithm> // for std::set_intersection, std::reverse
|
||||
#include <iostream> // for debug output
|
||||
#include <limits> // for std::numeric_limits
|
||||
#include <limits> // for std::numeric_limits
|
||||
#include <queue>
|
||||
#include <functional> // for std::function
|
||||
#include <map>
|
||||
#include <vector>
|
||||
#include <set>
|
||||
|
||||
namespace sysy {
|
||||
|
||||
// ==============================================================
|
||||
// DominatorTreeAnalysisPass 的静态ID
|
||||
// ==============================================================
|
||||
// 初始化 支配树静态 ID
|
||||
void *DominatorTreeAnalysisPass::ID = (void *)&DominatorTreeAnalysisPass::ID;
|
||||
|
||||
// ==============================================================
|
||||
// DominatorTree 结果类的实现
|
||||
// ==============================================================
|
||||
|
||||
// 构造函数:初始化关联函数,但不进行计算
|
||||
DominatorTree::DominatorTree(Function *F) : AssociatedFunction(F) {
|
||||
// 构造时不需要计算,在分析遍运行里计算并填充
|
||||
// 构造时可以不计算,在分析遍运行里计算并填充
|
||||
}
|
||||
|
||||
// Getter 方法 (保持不变)
|
||||
const std::set<BasicBlock *> *DominatorTree::getDominators(BasicBlock *BB) const {
|
||||
auto it = Dominators.find(BB);
|
||||
if (it != Dominators.end()) {
|
||||
@@ -49,437 +38,164 @@ const std::set<BasicBlock *> *DominatorTree::getDominanceFrontier(BasicBlock *BB
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
const std::set<BasicBlock *> *DominatorTree::getDominatorTreeChildren(BasicBlock *BB) const {
|
||||
auto it = DominatorTreeChildren.find(BB);
|
||||
if (it != DominatorTreeChildren.end()) {
|
||||
return &(it->second);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
// 辅助函数:打印 BasicBlock 集合 (保持不变)
|
||||
void printBBSet(const std::string &prefix, const std::set<BasicBlock *> &s) {
|
||||
if (!DEBUG)
|
||||
return;
|
||||
std::cout << prefix << "{";
|
||||
bool first = true;
|
||||
for (const auto &bb : s) {
|
||||
if (!first)
|
||||
std::cout << ", ";
|
||||
std::cout << bb->getName();
|
||||
first = false;
|
||||
}
|
||||
std::cout << "}" << std::endl;
|
||||
}
|
||||
|
||||
// 辅助函数:计算逆后序遍历 (RPO) - 保持不变
|
||||
std::vector<BasicBlock*> DominatorTree::computeReversePostOrder(Function* F) {
|
||||
std::vector<BasicBlock*> postOrder;
|
||||
std::set<BasicBlock*> visited;
|
||||
|
||||
std::function<void(BasicBlock*)> dfs_rpo =
|
||||
[&](BasicBlock* bb) {
|
||||
visited.insert(bb);
|
||||
for (BasicBlock* succ : bb->getSuccessors()) {
|
||||
if (visited.find(succ) == visited.end()) {
|
||||
dfs_rpo(succ);
|
||||
}
|
||||
}
|
||||
postOrder.push_back(bb);
|
||||
};
|
||||
|
||||
dfs_rpo(F->getEntryBlock());
|
||||
std::reverse(postOrder.begin(), postOrder.end());
|
||||
|
||||
if (DEBUG) {
|
||||
std::cout << "--- Computed RPO: ";
|
||||
for (BasicBlock* bb : postOrder) {
|
||||
std::cout << bb->getName() << " ";
|
||||
}
|
||||
std::cout << "---" << std::endl;
|
||||
const std::set<BasicBlock*>* DominatorTree::getDominatorTreeChildren(BasicBlock* BB) const {
|
||||
auto it = DominatorTreeChildren.find(BB);
|
||||
if (it != DominatorTreeChildren.end()) {
|
||||
return &(it->second);
|
||||
}
|
||||
return postOrder;
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
// computeDominators 方法 (保持不变,因为它它是独立于IDom算法的)
|
||||
void DominatorTree::computeDominators(Function *F) {
|
||||
if (DEBUG)
|
||||
std::cout << "--- Computing Dominators ---" << std::endl;
|
||||
|
||||
// 经典的迭代算法计算支配者集合
|
||||
// TODO: 可以替换为更高效的算法,如 Lengauer-Tarjan 算法
|
||||
BasicBlock *entryBlock = F->getEntryBlock();
|
||||
std::vector<BasicBlock*> bbs_rpo = computeReversePostOrder(F);
|
||||
|
||||
for (BasicBlock *bb : bbs_rpo) {
|
||||
for (const auto &bb_ptr : F->getBasicBlocks()) {
|
||||
BasicBlock *bb = bb_ptr.get();
|
||||
if (bb == entryBlock) {
|
||||
Dominators[bb].clear();
|
||||
Dominators[bb].insert(bb);
|
||||
if (DEBUG) std::cout << "Init Dominators[" << bb->getName() << "]: {" << bb->getName() << "}" << std::endl;
|
||||
} else {
|
||||
Dominators[bb].clear();
|
||||
for (BasicBlock *all_bb : bbs_rpo) {
|
||||
Dominators[bb].insert(all_bb);
|
||||
}
|
||||
if (DEBUG) {
|
||||
std::cout << "Init Dominators[" << bb->getName() << "]: ";
|
||||
printBBSet("", Dominators[bb]);
|
||||
for (const auto &all_bb_ptr : F->getBasicBlocks()) {
|
||||
Dominators[bb].insert(all_bb_ptr.get());
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
bool changed = true;
|
||||
int iteration = 0;
|
||||
while (changed) {
|
||||
changed = false;
|
||||
iteration++;
|
||||
if (DEBUG) std::cout << "Iteration " << iteration << std::endl;
|
||||
|
||||
for (BasicBlock *bb : bbs_rpo) {
|
||||
if (bb == entryBlock) continue;
|
||||
for (const auto &bb_ptr : F->getBasicBlocks()) {
|
||||
BasicBlock *bb = bb_ptr.get();
|
||||
if (bb == entryBlock)
|
||||
continue;
|
||||
|
||||
std::set<BasicBlock *> newDom;
|
||||
bool firstPredProcessed = false;
|
||||
|
||||
bool firstPred = true;
|
||||
for (BasicBlock *pred : bb->getPredecessors()) {
|
||||
if(DEBUG){
|
||||
std::cout << " Processing predecessor: " << pred->getName() << std::endl;
|
||||
}
|
||||
if (!firstPredProcessed) {
|
||||
newDom = Dominators[pred];
|
||||
firstPredProcessed = true;
|
||||
if (Dominators.count(pred)) {
|
||||
if (firstPred) {
|
||||
newDom = Dominators[pred];
|
||||
firstPred = false;
|
||||
} else {
|
||||
std::set<BasicBlock *> intersection;
|
||||
std::set_intersection(newDom.begin(), newDom.end(), Dominators[pred].begin(), Dominators[pred].end(),
|
||||
std::inserter(intersection, intersection.begin()));
|
||||
newDom = intersection;
|
||||
std::set<BasicBlock *> intersection;
|
||||
std::set_intersection(newDom.begin(), newDom.end(), Dominators[pred].begin(), Dominators[pred].end(),
|
||||
std::inserter(intersection, intersection.begin()));
|
||||
newDom = intersection;
|
||||
}
|
||||
}
|
||||
}
|
||||
newDom.insert(bb);
|
||||
|
||||
if (newDom != Dominators[bb]) {
|
||||
if (DEBUG) {
|
||||
std::cout << " Dominators[" << bb->getName() << "] changed from ";
|
||||
printBBSet("", Dominators[bb]);
|
||||
std::cout << " to ";
|
||||
printBBSet("", newDom);
|
||||
}
|
||||
Dominators[bb] = newDom;
|
||||
changed = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
if (DEBUG)
|
||||
std::cout << "--- Dominators Computation Finished ---" << std::endl;
|
||||
}
|
||||
|
||||
// ==============================================================
|
||||
// Lengauer-Tarjan 算法辅助数据结构和函数 (私有成员)
|
||||
// ==============================================================
|
||||
|
||||
// DFS 遍历,填充 dfnum_map, vertex_vec, parent_map
|
||||
// 对应用户代码的 dfs 函数
|
||||
void DominatorTree::dfs_lt_helper(BasicBlock* u) {
|
||||
dfnum_map[u] = df_counter;
|
||||
if (df_counter >= vertex_vec.size()) { // 动态调整大小
|
||||
vertex_vec.resize(df_counter + 1);
|
||||
}
|
||||
vertex_vec[df_counter] = u;
|
||||
if (DEBUG) std::cout << " DFS: Visiting " << u->getName() << ", dfnum = " << df_counter << std::endl;
|
||||
df_counter++;
|
||||
|
||||
for (BasicBlock* v : u->getSuccessors()) {
|
||||
if (dfnum_map.find(v) == dfnum_map.end()) { // 如果 v 未访问过
|
||||
parent_map[v] = u;
|
||||
if (DEBUG) std::cout << " DFS: Setting parent[" << v->getName() << "] = " << u->getName() << std::endl;
|
||||
dfs_lt_helper(v);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 并查集:找到集合的代表,并进行路径压缩
|
||||
// 同时更新 label,确保 label[i] 总是指向其祖先链中 sdom_map 最小的节点
|
||||
// 对应用户代码的 find 函数,也包含了 eval 的逻辑
|
||||
BasicBlock* DominatorTree::evalAndCompress_lt_helper(BasicBlock* i) {
|
||||
if (DEBUG) std::cout << " Eval: Processing " << i->getName() << std::endl;
|
||||
// 如果 i 是根 (ancestor_map[i] == nullptr)
|
||||
if (ancestor_map.find(i) == ancestor_map.end() || ancestor_map[i] == nullptr) {
|
||||
if (DEBUG) std::cout << " Eval: " << i->getName() << " is root, returning itself." << std::endl;
|
||||
return i; // 根节点自身就是路径上sdom最小的,因为它没有祖先
|
||||
}
|
||||
|
||||
// 如果 i 的祖先不是根,则递归查找并进行路径压缩
|
||||
BasicBlock* root_ancestor = evalAndCompress_lt_helper(ancestor_map[i]);
|
||||
|
||||
// 路径压缩时,根据 sdom_map 比较并更新 label_map
|
||||
// 确保 label_map[i] 存储的是 i 到 root_ancestor 路径上 sdom_map 最小的节点
|
||||
// 注意:这里的 ancestor_map[i] 已经被递归调用压缩过一次了,所以是root_ancestor的旧路径
|
||||
// 应该比较的是 label_map[ancestor_map[i]] 和 label_map[i]
|
||||
if (sdom_map.count(label_map[ancestor_map[i]]) && // 确保 label_map[ancestor_map[i]] 存在 sdom
|
||||
sdom_map.count(label_map[i]) && // 确保 label_map[i] 存在 sdom
|
||||
dfnum_map[sdom_map[label_map[ancestor_map[i]]]] < dfnum_map[sdom_map[label_map[i]]]) {
|
||||
if (DEBUG) std::cout << " Eval: Updating label for " << i->getName() << " from "
|
||||
<< label_map[i]->getName() << " to " << label_map[ancestor_map[i]]->getName() << std::endl;
|
||||
label_map[i] = label_map[ancestor_map[i]];
|
||||
}
|
||||
|
||||
ancestor_map[i] = root_ancestor; // 执行路径压缩:将 i 直接指向其所属集合的根
|
||||
if (DEBUG) std::cout << " Eval: Path compression for " << i->getName() << ", new ancestor = "
|
||||
<< (root_ancestor ? root_ancestor->getName() : "nullptr") << std::endl;
|
||||
|
||||
return label_map[i]; // <-- **将这里改为返回 label_map[i]**
|
||||
}
|
||||
|
||||
// Link 函数:将 v 加入 u 的 DFS 树子树中 (实际上是并查集操作)
|
||||
// 对应用户代码的 fa[u] = fth[u];
|
||||
void DominatorTree::link_lt_helper(BasicBlock* u_parent, BasicBlock* v_child) {
|
||||
ancestor_map[v_child] = u_parent; // 设置并查集父节点
|
||||
label_map[v_child] = v_child; // 初始化 label 为自身
|
||||
if (DEBUG) std::cout << " Link: " << v_child->getName() << " linked to " << u_parent->getName() << std::endl;
|
||||
}
|
||||
|
||||
// ==============================================================
|
||||
// Lengauer-Tarjan 算法实现 computeIDoms
|
||||
// ==============================================================
|
||||
void DominatorTree::computeIDoms(Function *F) {
|
||||
if (DEBUG) std::cout << "--- Computing Immediate Dominators (IDoms) using Lengauer-Tarjan ---" << std::endl;
|
||||
// 采用与之前类似的简化实现。TODO:Lengauer-Tarjan等算法。
|
||||
BasicBlock *entryBlock = F->getEntryBlock();
|
||||
IDoms[entryBlock] = nullptr;
|
||||
|
||||
BasicBlock *entryBlock = F->getEntryBlock();
|
||||
for (const auto &bb_ptr : F->getBasicBlocks()) {
|
||||
BasicBlock *bb = bb_ptr.get();
|
||||
if (bb == entryBlock)
|
||||
continue;
|
||||
|
||||
// 1. 初始化所有 LT 相关的数据结构
|
||||
dfnum_map.clear();
|
||||
vertex_vec.clear();
|
||||
parent_map.clear();
|
||||
sdom_map.clear();
|
||||
idom_map.clear();
|
||||
bucket_map.clear();
|
||||
ancestor_map.clear();
|
||||
label_map.clear();
|
||||
df_counter = 0; // DFS 计数器从 0 开始
|
||||
BasicBlock *currentIDom = nullptr;
|
||||
const std::set<BasicBlock *> *domsOfBB = getDominators(bb);
|
||||
if (!domsOfBB)
|
||||
continue;
|
||||
|
||||
// 预分配 vertex_vec 的大小,避免频繁resize
|
||||
vertex_vec.resize(F->getBasicBlocks().size() + 1);
|
||||
// 在 DFS 遍历之前,先为所有基本块初始化 sdom 和 label
|
||||
// 这是 Lengauer-Tarjan 算法的要求,确保所有节点在 Phase 2 开始前都在 map 中
|
||||
for (auto &bb_ptr : F->getBasicBlocks()) {
|
||||
BasicBlock* bb = bb_ptr.get();
|
||||
sdom_map[bb] = bb; // sdom(bb) 初始化为 bb 自身
|
||||
label_map[bb] = bb; // label(bb) 初始化为 bb 自身 (用于 Union-Find 的路径压缩)
|
||||
for (BasicBlock *D : *domsOfBB) {
|
||||
if (D == bb)
|
||||
continue;
|
||||
|
||||
bool isCandidateIDom = true;
|
||||
for (BasicBlock *candidate : *domsOfBB) {
|
||||
if (candidate == bb || candidate == D)
|
||||
continue;
|
||||
const std::set<BasicBlock *> *domsOfCandidate = getDominators(candidate);
|
||||
if (domsOfCandidate && domsOfCandidate->count(D) == 0 && domsOfBB->count(candidate)) {
|
||||
isCandidateIDom = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
if (isCandidateIDom) {
|
||||
currentIDom = D;
|
||||
break;
|
||||
}
|
||||
}
|
||||
// 确保入口块也被正确初始化(如果它不在 F->getBasicBlocks() 的正常迭代中)
|
||||
sdom_map[entryBlock] = entryBlock;
|
||||
label_map[entryBlock] = entryBlock;
|
||||
// Phase 1: DFS 遍历并预处理
|
||||
// 对应用户代码的 dfs(st)
|
||||
dfs_lt_helper(entryBlock);
|
||||
idom_map[entryBlock] = nullptr; // 入口块没有即时支配者
|
||||
if (DEBUG) std::cout << " IDom[" << entryBlock->getName() << "] = nullptr" << std::endl;
|
||||
|
||||
if (DEBUG) std::cout << " Sdom[" << entryBlock->getName() << "] = " << entryBlock->getName() << std::endl;
|
||||
|
||||
// 初始化并查集的祖先和 label
|
||||
for (auto const& [bb_key, dfn_val] : dfnum_map) {
|
||||
ancestor_map[bb_key] = nullptr; // 初始为独立集合的根
|
||||
label_map[bb_key] = bb_key; // 初始 label 为自身
|
||||
}
|
||||
|
||||
if (DEBUG) {
|
||||
std::cout << " --- DFS Phase Complete ---" << std::endl;
|
||||
std::cout << " dfnum_map:" << std::endl;
|
||||
for (auto const& [bb, dfn] : dfnum_map) {
|
||||
std::cout << " " << bb->getName() << " -> " << dfn << std::endl;
|
||||
}
|
||||
std::cout << " vertex_vec (by dfnum):" << std::endl;
|
||||
for (size_t k = 0; k < df_counter; ++k) {
|
||||
if (vertex_vec[k]) std::cout << " [" << k << "] -> " << vertex_vec[k]->getName() << std::endl;
|
||||
}
|
||||
std::cout << " parent_map:" << std::endl;
|
||||
for (auto const& [child, parent] : parent_map) {
|
||||
std::cout << " " << child->getName() << " -> " << (parent ? parent->getName() : "nullptr") << std::endl;
|
||||
}
|
||||
std::cout << " ------------------------" << std::endl;
|
||||
}
|
||||
|
||||
|
||||
// Phase 2: 计算半支配者 (sdom)
|
||||
// 对应用户代码的 for (int i = dfc; i >= 2; --i) 循环的上半部分
|
||||
// 按照 DFS 编号递减的顺序遍历所有节点 (除了 entryBlock,它的 DFS 编号是 0)
|
||||
if (DEBUG) std::cout << "--- Phase 2: Computing Semi-Dominators (sdom) ---" << std::endl;
|
||||
for (int i = df_counter - 1; i >= 1; --i) { // 从 DFS 编号最大的节点开始,到 1
|
||||
BasicBlock* w = vertex_vec[i]; // 当前处理的节点
|
||||
if (DEBUG) std::cout << " Processing node w: " << w->getName() << " (dfnum=" << i << ")" << std::endl;
|
||||
|
||||
|
||||
// 对于 w 的每个前驱 v
|
||||
for (BasicBlock* v : w->getPredecessors()) {
|
||||
if (DEBUG) std::cout << " Considering predecessor v: " << v->getName() << std::endl;
|
||||
// 如果前驱 v 未被 DFS 访问过 (即不在 dfnum_map 中),则跳过
|
||||
if (dfnum_map.find(v) == dfnum_map.end()) {
|
||||
if (DEBUG) std::cout << " Predecessor " << v->getName() << " not in DFS tree, skipping." << std::endl;
|
||||
continue;
|
||||
}
|
||||
|
||||
// 调用 evalAndCompress 来找到 v 在其 DFS 树祖先链上具有最小 sdom 的节点
|
||||
BasicBlock* u_with_min_sdom_on_path = evalAndCompress_lt_helper(v);
|
||||
if (DEBUG) std::cout << " Eval(" << v->getName() << ") returned "
|
||||
<< u_with_min_sdom_on_path->getName() << std::endl;
|
||||
if (DEBUG && sdom_map.count(u_with_min_sdom_on_path) && sdom_map.count(w)) {
|
||||
std::cout << " Comparing sdom: dfnum[" << sdom_map[u_with_min_sdom_on_path]->getName() << "] (" << dfnum_map[sdom_map[u_with_min_sdom_on_path]]
|
||||
<< ") vs dfnum[" << sdom_map[w]->getName() << "] (" << dfnum_map[sdom_map[w]] << ")" << std::endl;
|
||||
}
|
||||
// 比较 sdom(u) 和 sdom(w)
|
||||
if (sdom_map.count(u_with_min_sdom_on_path) && sdom_map.count(w) &&
|
||||
dfnum_map[sdom_map[u_with_min_sdom_on_path]] < dfnum_map[sdom_map[w]]) {
|
||||
if (DEBUG) std::cout << " Updating sdom[" << w->getName() << "] from "
|
||||
<< sdom_map[w]->getName() << " to "
|
||||
<< sdom_map[u_with_min_sdom_on_path]->getName() << std::endl;
|
||||
sdom_map[w] = sdom_map[u_with_min_sdom_on_path]; // 更新 sdom(w)
|
||||
if (DEBUG) std::cout << " Sdom update applied. New sdom[" << w->getName() << "] = " << sdom_map[w]->getName() << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
// 将 w 加入 sdom(w) 对应的桶中
|
||||
bucket_map[sdom_map[w]].push_back(w);
|
||||
if (DEBUG) std::cout << " Adding " << w->getName() << " to bucket of sdom(" << w->getName() << "): "
|
||||
<< sdom_map[w]->getName() << std::endl;
|
||||
|
||||
// 将 w 的父节点加入并查集 (link 操作)
|
||||
if (parent_map.count(w) && parent_map[w] != nullptr) {
|
||||
link_lt_helper(parent_map[w], w);
|
||||
}
|
||||
|
||||
// Phase 3-part 1: 处理 parent[w] 的桶中所有节点,确定部分 idom
|
||||
if (parent_map.count(w) && parent_map[w] != nullptr) {
|
||||
BasicBlock* p = parent_map[w]; // p 是 w 的父节点
|
||||
if (DEBUG) std::cout << " Processing bucket for parent " << p->getName() << std::endl;
|
||||
|
||||
// 注意:这里需要复制桶的内容,因为原始桶在循环中会被clear
|
||||
std::vector<BasicBlock*> nodes_in_p_bucket_copy = bucket_map[p];
|
||||
for (BasicBlock* y : nodes_in_p_bucket_copy) {
|
||||
if (DEBUG) std::cout << " Processing node y from bucket: " << y->getName() << std::endl;
|
||||
// 找到 y 在其 DFS 树祖先链上具有最小 sdom 的节点
|
||||
BasicBlock* u = evalAndCompress_lt_helper(y);
|
||||
if (DEBUG) std::cout << " Eval(" << y->getName() << ") returned " << u->getName() << std::endl;
|
||||
|
||||
// 确定 idom(y)
|
||||
// if sdom(eval(y)) == sdom(parent(w)), then idom(y) = parent(w)
|
||||
// else idom(y) = eval(y)
|
||||
if (sdom_map.count(u) && sdom_map.count(p) &&
|
||||
dfnum_map[sdom_map[u]] < dfnum_map[sdom_map[p]]) {
|
||||
idom_map[y] = u; // 确定的 idom
|
||||
if (DEBUG) std::cout << " IDom[" << y->getName() << "] set to " << u->getName() << std::endl;
|
||||
} else {
|
||||
idom_map[y] = p; // p 是 y 的 idom
|
||||
if (DEBUG) std::cout << " IDom[" << y->getName() << "] set to " << p->getName() << std::endl;
|
||||
}
|
||||
}
|
||||
bucket_map[p].clear(); // 清空桶,防止重复处理
|
||||
if (DEBUG) std::cout << " Cleared bucket for parent " << p->getName() << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
// Phase 3-part 2: 最终确定 idom (处理那些 idom != sdom 的节点)
|
||||
if (DEBUG) std::cout << "--- Phase 3: Finalizing Immediate Dominators (idom) ---" << std::endl;
|
||||
for (int i = 1; i < df_counter; ++i) { // 从 DFS 编号最小的节点 (除了 entryBlock) 开始
|
||||
BasicBlock* w = vertex_vec[i];
|
||||
if (DEBUG) std::cout << " Finalizing node w: " << w->getName() << std::endl;
|
||||
if (idom_map.count(w) && sdom_map.count(w) && idom_map[w] != sdom_map[w]) {
|
||||
// idom[w] 的 idom 是其真正的 idom
|
||||
if (DEBUG) std::cout << " idom[" << w->getName() << "] (" << idom_map[w]->getName()
|
||||
<< ") != sdom[" << w->getName() << "] (" << sdom_map[w]->getName() << ")" << std::endl;
|
||||
if (idom_map.count(idom_map[w])) {
|
||||
idom_map[w] = idom_map[idom_map[w]];
|
||||
if (DEBUG) std::cout << " Updating idom[" << w->getName() << "] to idom(idom(w)): "
|
||||
<< idom_map[w]->getName() << std::endl;
|
||||
} else {
|
||||
if (DEBUG) std::cout << " Warning: idom(idom(" << w->getName() << ")) not found, leaving idom[" << w->getName() << "] as is." << std::endl;
|
||||
}
|
||||
}
|
||||
if (DEBUG) {
|
||||
std::cout << " Final IDom[" << w->getName() << "] = " << (idom_map[w] ? idom_map[w]->getName() : "nullptr") << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
// 将计算结果从 idom_map 存储到 DominatorTree 的成员变量 IDoms 中
|
||||
IDoms = idom_map;
|
||||
|
||||
if (DEBUG) std::cout << "--- Immediate Dominators Computation Finished ---" << std::endl;
|
||||
IDoms[bb] = currentIDom;
|
||||
}
|
||||
}
|
||||
|
||||
// ==============================================================
|
||||
// computeDominanceFrontiers 和 computeDominatorTreeChildren (保持不变)
|
||||
// ==============================================================
|
||||
|
||||
void DominatorTree::computeDominanceFrontiers(Function *F) {
|
||||
if (DEBUG)
|
||||
std::cout << "--- Computing Dominance Frontiers ---" << std::endl;
|
||||
|
||||
// 经典的支配边界计算算法
|
||||
for (const auto &bb_ptr_X : F->getBasicBlocks()) {
|
||||
BasicBlock *X = bb_ptr_X.get();
|
||||
DominanceFrontiers[X].clear();
|
||||
|
||||
for (BasicBlock *Y : X->getSuccessors()) {
|
||||
const std::set<BasicBlock *> *domsOfY = getDominators(Y);
|
||||
if (domsOfY && domsOfY->find(X) == domsOfY->end()) {
|
||||
DominanceFrontiers[X].insert(Y);
|
||||
}
|
||||
}
|
||||
|
||||
const std::set<BasicBlock *> *domsOfX = getDominators(X);
|
||||
if (!domsOfX)
|
||||
continue;
|
||||
for (const auto &bb_ptr_Z : F->getBasicBlocks()) {
|
||||
BasicBlock *Z = bb_ptr_Z.get();
|
||||
const std::set<BasicBlock *> *domsOfZ = getDominators(Z);
|
||||
|
||||
if (!domsOfZ || domsOfZ->find(X) == domsOfZ->end()) { // Z 不被 X 支配
|
||||
if (Z == X)
|
||||
continue;
|
||||
}
|
||||
const std::set<BasicBlock *> *domsOfZ = getDominators(Z);
|
||||
if (domsOfZ && domsOfZ->count(X) && Z != X) {
|
||||
|
||||
for (BasicBlock *Y : Z->getSuccessors()) {
|
||||
const std::set<BasicBlock *> *domsOfY = getDominators(Y);
|
||||
// 如果 Y == X,或者 Y 不被 X 严格支配 (即 Y 不被 X 支配)
|
||||
if (Y == X || (domsOfY && domsOfY->find(X) == domsOfY->end())) {
|
||||
DominanceFrontiers[X].insert(Y);
|
||||
for (BasicBlock *Y : Z->getSuccessors()) {
|
||||
const std::set<BasicBlock *> *domsOfY = getDominators(Y);
|
||||
if (domsOfY && domsOfY->find(X) == domsOfY->end()) {
|
||||
DominanceFrontiers[X].insert(Y);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
if (DEBUG) {
|
||||
std::cout << " DF(" << X->getName() << "): ";
|
||||
printBBSet("", DominanceFrontiers[X]);
|
||||
}
|
||||
}
|
||||
if (DEBUG)
|
||||
std::cout << "--- Dominance Frontiers Computation Finished ---" << std::endl;
|
||||
}
|
||||
|
||||
void DominatorTree::computeDominatorTreeChildren(Function *F) {
|
||||
if (DEBUG)
|
||||
std::cout << "--- Computing Dominator Tree Children ---" << std::endl;
|
||||
// 首先清空,确保重新计算时是空的
|
||||
for (auto &bb_ptr : F->getBasicBlocks()) {
|
||||
DominatorTreeChildren[bb_ptr.get()].clear();
|
||||
}
|
||||
|
||||
for (auto &bb_ptr : F->getBasicBlocks()) {
|
||||
BasicBlock *B = bb_ptr.get();
|
||||
BasicBlock *A = getImmediateDominator(B); // A 是 B 的即时支配者
|
||||
|
||||
if (A) { // 如果 B 有即时支配者 A (即 B 不是入口块)
|
||||
DominatorTreeChildren[A].insert(B);
|
||||
if (DEBUG) {
|
||||
std::cout << " " << B->getName() << " is child of " << A->getName() << std::endl;
|
||||
auto it = getImmediateDominator(B);
|
||||
if (it != nullptr) {
|
||||
BasicBlock *A = it;
|
||||
if (A) {
|
||||
DominatorTreeChildren[A].insert(B);
|
||||
}
|
||||
}
|
||||
}
|
||||
if (DEBUG)
|
||||
std::cout << "--- Dominator Tree Children Computation Finished ---" << std::endl;
|
||||
}
|
||||
|
||||
// ==============================================================
|
||||
// DominatorTreeAnalysisPass 的实现 (保持不变)
|
||||
// DominatorTreeAnalysisPass 的实现
|
||||
// ==============================================================
|
||||
|
||||
bool DominatorTreeAnalysisPass::runOnFunction(Function *F, AnalysisManager &AM) {
|
||||
// 每次运行时清空旧数据,确保重新计算
|
||||
CurrentDominatorTree = std::make_unique<DominatorTree>(F);
|
||||
|
||||
bool DominatorTreeAnalysisPass::runOnFunction(Function* F, AnalysisManager &AM) {
|
||||
CurrentDominatorTree = std::make_unique<DominatorTree>(F);
|
||||
CurrentDominatorTree->computeDominators(F);
|
||||
CurrentDominatorTree->computeIDoms(F); // 修正后的LT算法
|
||||
CurrentDominatorTree->computeIDoms(F);
|
||||
CurrentDominatorTree->computeDominanceFrontiers(F);
|
||||
CurrentDominatorTree->computeDominatorTreeChildren(F);
|
||||
return false;
|
||||
}
|
||||
|
||||
std::unique_ptr<AnalysisResultBase> DominatorTreeAnalysisPass::getResult() {
|
||||
// 返回计算好的 DominatorTree 实例,所有权转移给 AnalysisManager
|
||||
return std::move(CurrentDominatorTree);
|
||||
}
|
||||
|
||||
|
||||
@@ -1,79 +0,0 @@
|
||||
#include "BuildCFG.h"
|
||||
#include "Dom.h"
|
||||
#include "Liveness.h"
|
||||
#include <iostream>
|
||||
#include <queue>
|
||||
#include <set>
|
||||
|
||||
namespace sysy {
|
||||
|
||||
void *BuildCFG::ID = (void *)&BuildCFG::ID; // 定义唯一的 Pass ID
|
||||
|
||||
// 声明Pass的分析使用
|
||||
void BuildCFG::getAnalysisUsage(std::set<void *> &analysisDependencies, std::set<void *> &analysisInvalidations) const {
|
||||
// BuildCFG不依赖其他分析
|
||||
// analysisDependencies.insert(&DominatorTreeAnalysisPass::ID); // 错误的例子
|
||||
|
||||
// BuildCFG会使所有依赖于CFG的分析结果失效,所以它必须声明这些失效
|
||||
analysisInvalidations.insert(&DominatorTreeAnalysisPass::ID);
|
||||
analysisInvalidations.insert(&LivenessAnalysisPass::ID);
|
||||
}
|
||||
|
||||
bool BuildCFG::runOnFunction(Function *F, AnalysisManager &AM) {
|
||||
if (DEBUG) {
|
||||
std::cout << "Running BuildCFG pass on function: " << F->getName() << std::endl;
|
||||
}
|
||||
|
||||
bool changed = false;
|
||||
|
||||
// 1. 清空所有基本块的前驱和后继列表
|
||||
for (auto &bb : F->getBasicBlocks()) {
|
||||
bb->clearPredecessors();
|
||||
bb->clearSuccessors();
|
||||
}
|
||||
|
||||
// 2. 遍历每个基本块,重建CFG
|
||||
for (auto &bb : F->getBasicBlocks()) {
|
||||
// 获取基本块的最后一条指令
|
||||
auto &inst = *bb->terminator();
|
||||
Instruction *termInst = inst.get();
|
||||
// 确保基本块有终结指令
|
||||
if (!termInst) {
|
||||
continue;
|
||||
}
|
||||
|
||||
// 根据终结指令类型,建立前驱后继关系
|
||||
if (termInst->isBranch()) {
|
||||
// 无条件跳转
|
||||
if (termInst->isUnconditional()) {
|
||||
auto brInst = dynamic_cast<UncondBrInst *>(termInst);
|
||||
BasicBlock *succ = dynamic_cast<BasicBlock *>(brInst->getBlock());
|
||||
assert(succ && "Branch instruction's target must be a BasicBlock");
|
||||
bb->addSuccessor(succ);
|
||||
succ->addPredecessor(bb.get());
|
||||
changed = true;
|
||||
|
||||
// 条件跳转
|
||||
} else if (termInst->isConditional()) {
|
||||
auto brInst = dynamic_cast<CondBrInst *>(termInst);
|
||||
BasicBlock *trueSucc = dynamic_cast<BasicBlock *>(brInst->getThenBlock());
|
||||
BasicBlock *falseSucc = dynamic_cast<BasicBlock *>(brInst->getElseBlock());
|
||||
|
||||
assert(trueSucc && falseSucc && "Branch instruction's targets must be BasicBlocks");
|
||||
|
||||
bb->addSuccessor(trueSucc);
|
||||
trueSucc->addPredecessor(bb.get());
|
||||
bb->addSuccessor(falseSucc);
|
||||
falseSucc->addPredecessor(bb.get());
|
||||
changed = true;
|
||||
}
|
||||
} else if (auto retInst = dynamic_cast<ReturnInst *>(termInst)) {
|
||||
// RetInst没有后继,无需处理
|
||||
// ...
|
||||
}
|
||||
}
|
||||
|
||||
return changed;
|
||||
}
|
||||
|
||||
} // namespace sysy
|
||||
241
src/midend/Pass/Optimize/ConstPropagation.cpp
Normal file
241
src/midend/Pass/Optimize/ConstPropagation.cpp
Normal file
@@ -0,0 +1,241 @@
|
||||
#include "Pass/Optimize/ConstPropagation.h"
|
||||
#include "IR.h"
|
||||
#include "Pass.h"
|
||||
#include <climits>
|
||||
#include <cmath>
|
||||
|
||||
namespace sysy {
|
||||
|
||||
char ConstPropagation::ID = 0;
|
||||
|
||||
bool ConstPropagation::runOnFunction(Function *func, AnalysisManager &am) {
|
||||
bool changed = false;
|
||||
bool localChanged = true;
|
||||
|
||||
while (localChanged) {
|
||||
localChanged = false;
|
||||
|
||||
for (auto &bb : func->getBasicBlocks()) {
|
||||
for (auto instIter = bb->getInstructions().begin();
|
||||
instIter != bb->getInstructions().end();) {
|
||||
auto &inst = *instIter;
|
||||
bool shouldAdvanceIter = true;
|
||||
|
||||
// 处理二元运算指令
|
||||
if (auto *binaryInst = dynamic_cast<BinaryInst *>(inst.get())) {
|
||||
auto *lhs = binaryInst->getLhs();
|
||||
auto *rhs = binaryInst->getRhs();
|
||||
|
||||
auto *lhsConst = dynamic_cast<ConstantValue *>(lhs);
|
||||
auto *rhsConst = dynamic_cast<ConstantValue *>(rhs);
|
||||
|
||||
if (lhsConst && rhsConst) {
|
||||
ConstantValue *newConst = nullptr;
|
||||
|
||||
try {
|
||||
if (lhs->isInt() && rhs->isInt()) {
|
||||
int l = lhsConst->getInt();
|
||||
int r = rhsConst->getInt();
|
||||
int result;
|
||||
bool validOperation = true;
|
||||
|
||||
switch (binaryInst->getKind()) {
|
||||
case Instruction::kAdd:
|
||||
// 检查加法溢出
|
||||
if ((r > 0 && l > INT_MAX - r) || (r < 0 && l < INT_MIN - r)) {
|
||||
validOperation = false;
|
||||
} else {
|
||||
result = l + r;
|
||||
}
|
||||
break;
|
||||
case Instruction::kSub:
|
||||
// 检查减法溢出
|
||||
if ((r < 0 && l > INT_MAX + r) || (r > 0 && l < INT_MIN + r)) {
|
||||
validOperation = false;
|
||||
} else {
|
||||
result = l - r;
|
||||
}
|
||||
break;
|
||||
case Instruction::kMul:
|
||||
// 检查乘法溢出
|
||||
if (l != 0 && r != 0 &&
|
||||
(std::abs(l) > INT_MAX / std::abs(r))) {
|
||||
validOperation = false;
|
||||
} else {
|
||||
result = l * r;
|
||||
}
|
||||
break;
|
||||
case Instruction::kDiv:
|
||||
if (r == 0) {
|
||||
validOperation = false;
|
||||
} else {
|
||||
result = l / r;
|
||||
}
|
||||
break;
|
||||
case Instruction::kRem:
|
||||
if (r == 0) {
|
||||
validOperation = false;
|
||||
} else {
|
||||
result = l % r;
|
||||
}
|
||||
break;
|
||||
case Instruction::kICmpEQ: result = (l == r) ? 1 : 0; break;
|
||||
case Instruction::kICmpNE: result = (l != r) ? 1 : 0; break;
|
||||
case Instruction::kICmpLT: result = (l < r) ? 1 : 0; break;
|
||||
case Instruction::kICmpGT: result = (l > r) ? 1 : 0; break;
|
||||
case Instruction::kICmpLE: result = (l <= r) ? 1 : 0; break;
|
||||
case Instruction::kICmpGE: result = (l >= r) ? 1 : 0; break;
|
||||
case Instruction::kAnd: result = (l && r) ? 1 : 0; break;
|
||||
case Instruction::kOr: result = (l || r) ? 1 : 0; break;
|
||||
default:
|
||||
validOperation = false;
|
||||
}
|
||||
|
||||
if (validOperation) {
|
||||
if (binaryInst->isCmp() || binaryInst->getKind() == Instruction::kAnd ||
|
||||
binaryInst->getKind() == Instruction::kOr) {
|
||||
newConst = ConstantInteger::get(Type::getIntType(), result);
|
||||
} else {
|
||||
newConst = ConstantInteger::get(result);
|
||||
}
|
||||
}
|
||||
} else if (lhs->isFloat() && rhs->isFloat()) {
|
||||
float l = lhsConst->getFloat();
|
||||
float r = rhsConst->getFloat();
|
||||
bool validOperation = true;
|
||||
|
||||
switch (binaryInst->getKind()) {
|
||||
case Instruction::kFAdd: {
|
||||
float result = l + r;
|
||||
if (std::isfinite(result)) {
|
||||
newConst = ConstantFloating::get(result);
|
||||
} else {
|
||||
validOperation = false;
|
||||
}
|
||||
break;
|
||||
}
|
||||
case Instruction::kFSub: {
|
||||
float result = l - r;
|
||||
if (std::isfinite(result)) {
|
||||
newConst = ConstantFloating::get(result);
|
||||
} else {
|
||||
validOperation = false;
|
||||
}
|
||||
break;
|
||||
}
|
||||
case Instruction::kFMul: {
|
||||
float result = l * r;
|
||||
if (std::isfinite(result)) {
|
||||
newConst = ConstantFloating::get(result);
|
||||
} else {
|
||||
validOperation = false;
|
||||
}
|
||||
break;
|
||||
}
|
||||
case Instruction::kFDiv: {
|
||||
if (std::abs(r) < std::numeric_limits<float>::epsilon()) {
|
||||
validOperation = false;
|
||||
} else {
|
||||
float result = l / r;
|
||||
if (std::isfinite(result)) {
|
||||
newConst = ConstantFloating::get(result);
|
||||
} else {
|
||||
validOperation = false;
|
||||
}
|
||||
}
|
||||
break;
|
||||
}
|
||||
case Instruction::kFCmpEQ:
|
||||
newConst = ConstantInteger::get(Type::getIntType(), (l == r) ? 1 : 0);
|
||||
break;
|
||||
case Instruction::kFCmpNE:
|
||||
newConst = ConstantInteger::get(Type::getIntType(), (l != r) ? 1 : 0);
|
||||
break;
|
||||
case Instruction::kFCmpLT:
|
||||
newConst = ConstantInteger::get(Type::getIntType(), (l < r) ? 1 : 0);
|
||||
break;
|
||||
case Instruction::kFCmpGT:
|
||||
newConst = ConstantInteger::get(Type::getIntType(), (l > r) ? 1 : 0);
|
||||
break;
|
||||
case Instruction::kFCmpLE:
|
||||
newConst = ConstantInteger::get(Type::getIntType(), (l <= r) ? 1 : 0);
|
||||
break;
|
||||
case Instruction::kFCmpGE:
|
||||
newConst = ConstantInteger::get(Type::getIntType(), (l >= r) ? 1 : 0);
|
||||
break;
|
||||
default:
|
||||
validOperation = false;
|
||||
}
|
||||
}
|
||||
} catch (...) {
|
||||
// 捕获可能的异常,跳过优化
|
||||
newConst = nullptr;
|
||||
}
|
||||
|
||||
if (newConst) {
|
||||
binaryInst->replaceAllUsesWith(newConst);
|
||||
instIter = bb->getInstructions().erase(instIter);
|
||||
shouldAdvanceIter = false;
|
||||
localChanged = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
// 处理一元运算指令
|
||||
else if (auto *unaryInst = dynamic_cast<UnaryInst *>(inst.get())) {
|
||||
auto *operand = unaryInst->getOperand();
|
||||
auto *operandConst = dynamic_cast<ConstantValue *>(operand);
|
||||
|
||||
if (operandConst) {
|
||||
ConstantValue *newConst = nullptr;
|
||||
|
||||
if (operand->isInt()) {
|
||||
int val = operandConst->getInt();
|
||||
|
||||
switch (unaryInst->getKind()) {
|
||||
case Instruction::kNeg:
|
||||
if (val != INT_MIN) { // 避免溢出
|
||||
newConst = ConstantInteger::get(-val);
|
||||
}
|
||||
break;
|
||||
case Instruction::kNot:
|
||||
newConst = ConstantInteger::get(Type::getIntType(), (!val) ? 1 : 0);
|
||||
break;
|
||||
default:
|
||||
break;
|
||||
}
|
||||
} else if (operand->isFloat()) {
|
||||
float val = operandConst->getFloat();
|
||||
|
||||
switch (unaryInst->getKind()) {
|
||||
case Instruction::kFNeg:
|
||||
newConst = ConstantFloating::get(-val);
|
||||
break;
|
||||
default:
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (newConst) {
|
||||
unaryInst->replaceAllUsesWith(newConst);
|
||||
instIter = bb->getInstructions().erase(instIter);
|
||||
shouldAdvanceIter = false;
|
||||
localChanged = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (shouldAdvanceIter) {
|
||||
++instIter;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (localChanged) {
|
||||
changed = true;
|
||||
}
|
||||
}
|
||||
|
||||
return changed;
|
||||
}
|
||||
|
||||
} // namespace sysy
|
||||
@@ -51,8 +51,10 @@ void DCEContext::run(Function *func, AnalysisManager *AM, bool &changed) {
|
||||
// 如果指令不在活跃集合中,则删除它。
|
||||
// 分支和返回指令由 isAlive 处理,并会被保留。
|
||||
if (alive_insts.count(currentInst) == 0) {
|
||||
instIter = SysYIROptUtils::usedelete(instIter); // 删除后返回下一个迭代器
|
||||
// 删除指令,保留用户风格的 SysYIROptUtils::usedelete 和 erase
|
||||
changed = true; // 标记 IR 已被修改
|
||||
SysYIROptUtils::usedelete(currentInst);
|
||||
instIter = basicBlock->getInstructions().erase(instIter); // 删除后返回下一个迭代器
|
||||
} else {
|
||||
++instIter; // 指令活跃,移动到下一个
|
||||
}
|
||||
|
||||
@@ -1,143 +0,0 @@
|
||||
#include "../../include/midend/Pass/Optimize/LargeArrayToGlobal.h"
|
||||
#include "../../IR.h"
|
||||
#include <unordered_map>
|
||||
#include <sstream>
|
||||
#include <string>
|
||||
|
||||
namespace sysy {
|
||||
|
||||
// Helper function to convert type to string
|
||||
static std::string typeToString(Type *type) {
|
||||
if (!type) return "null";
|
||||
|
||||
switch (type->getKind()) {
|
||||
case Type::kInt:
|
||||
return "int";
|
||||
case Type::kFloat:
|
||||
return "float";
|
||||
case Type::kPointer:
|
||||
return "ptr";
|
||||
case Type::kArray: {
|
||||
auto *arrayType = type->as<ArrayType>();
|
||||
return "[" + std::to_string(arrayType->getNumElements()) + " x " +
|
||||
typeToString(arrayType->getElementType()) + "]";
|
||||
}
|
||||
default:
|
||||
return "unknown";
|
||||
}
|
||||
}
|
||||
|
||||
void *LargeArrayToGlobalPass::ID = &LargeArrayToGlobalPass::ID;
|
||||
|
||||
bool LargeArrayToGlobalPass::runOnModule(Module *M, AnalysisManager &AM) {
|
||||
bool changed = false;
|
||||
|
||||
if (!M) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// Collect all alloca instructions from all functions
|
||||
std::vector<std::pair<AllocaInst*, Function*>> allocasToConvert;
|
||||
|
||||
for (auto &funcPair : M->getFunctions()) {
|
||||
Function *F = funcPair.second.get();
|
||||
if (!F || F->getBasicBlocks().begin() == F->getBasicBlocks().end()) {
|
||||
continue;
|
||||
}
|
||||
|
||||
for (auto &BB : F->getBasicBlocks()) {
|
||||
for (auto &inst : BB->getInstructions()) {
|
||||
if (auto *alloca = dynamic_cast<AllocaInst*>(inst.get())) {
|
||||
Type *allocatedType = alloca->getAllocatedType();
|
||||
|
||||
// Calculate the size of the allocated type
|
||||
unsigned size = calculateTypeSize(allocatedType);
|
||||
|
||||
// Debug: print size information
|
||||
std::cout << "LargeArrayToGlobalPass: Found alloca with size " << size
|
||||
<< " for type " << typeToString(allocatedType) << std::endl;
|
||||
|
||||
// Convert arrays of 1KB (1024 bytes) or larger to global variables
|
||||
if (size >= 1024) {
|
||||
std::cout << "LargeArrayToGlobalPass: Converting array of size " << size << " to global" << std::endl;
|
||||
allocasToConvert.emplace_back(alloca, F);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Convert the collected alloca instructions to global variables
|
||||
for (auto [alloca, F] : allocasToConvert) {
|
||||
convertAllocaToGlobal(alloca, F, M);
|
||||
changed = true;
|
||||
}
|
||||
|
||||
return changed;
|
||||
}
|
||||
|
||||
unsigned LargeArrayToGlobalPass::calculateTypeSize(Type *type) {
|
||||
if (!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() * calculateTypeSize(arrayType->getElementType());
|
||||
}
|
||||
default:
|
||||
return 0;
|
||||
}
|
||||
}
|
||||
|
||||
void LargeArrayToGlobalPass::convertAllocaToGlobal(AllocaInst *alloca, Function *F, Module *M) {
|
||||
Type *allocatedType = alloca->getAllocatedType();
|
||||
|
||||
// Create a unique name for the global variable
|
||||
std::string globalName = generateUniqueGlobalName(alloca, F);
|
||||
|
||||
// Create the global variable - GlobalValue expects pointer type
|
||||
Type *pointerType = Type::getPointerType(allocatedType);
|
||||
GlobalValue *globalVar = M->createGlobalValue(globalName, pointerType);
|
||||
|
||||
if (!globalVar) {
|
||||
return;
|
||||
}
|
||||
|
||||
// Replace all uses of the alloca with the global variable
|
||||
alloca->replaceAllUsesWith(globalVar);
|
||||
|
||||
// Remove the alloca instruction from its basic block
|
||||
for (auto &BB : F->getBasicBlocks()) {
|
||||
auto &instructions = BB->getInstructions();
|
||||
for (auto it = instructions.begin(); it != instructions.end(); ++it) {
|
||||
if (it->get() == alloca) {
|
||||
instructions.erase(it);
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
std::string LargeArrayToGlobalPass::generateUniqueGlobalName(AllocaInst *alloca, Function *F) {
|
||||
std::string baseName = alloca->getName();
|
||||
if (baseName.empty()) {
|
||||
baseName = "array";
|
||||
}
|
||||
|
||||
// Ensure uniqueness by appending function name and counter
|
||||
static std::unordered_map<std::string, int> nameCounter;
|
||||
std::string key = F->getName() + "." + baseName;
|
||||
|
||||
int counter = nameCounter[key]++;
|
||||
std::ostringstream oss;
|
||||
oss << key << "." << counter;
|
||||
|
||||
return oss.str();
|
||||
}
|
||||
|
||||
} // namespace sysy
|
||||
@@ -60,7 +60,7 @@ void Mem2RegContext::run(Function *func, AnalysisManager *AM) {
|
||||
}
|
||||
|
||||
// 从入口基本块开始,对支配树进行 DFS 遍历,进行变量重命名
|
||||
renameVariables(func->getEntryBlock()); // 第一个参数 alloca 在这里不使用,因为是递归入口点
|
||||
renameVariables(nullptr, func->getEntryBlock()); // 第一个参数 alloca 在这里不使用,因为是递归入口点
|
||||
|
||||
// --------------------------------------------------------------------
|
||||
// 阶段4: 清理
|
||||
@@ -209,21 +209,16 @@ void Mem2RegContext::insertPhis(AllocaInst *alloca, const std::unordered_set<Bas
|
||||
}
|
||||
|
||||
// 对支配树进行深度优先遍历,重命名变量并替换 load/store 指令
|
||||
// 移除了 AllocaInst *currentAlloca 参数,因为这个函数是为整个基本块处理所有可提升的 Alloca
|
||||
void Mem2RegContext::renameVariables(BasicBlock *currentBB) {
|
||||
// 1. 在函数开始时,记录每个 promotableAlloca 的当前栈深度。
|
||||
// 这将用于在函数返回时精确地回溯栈状态。
|
||||
std::map<AllocaInst *, size_t> originalStackSizes;
|
||||
for (auto alloca : promotableAllocas) {
|
||||
originalStackSizes[alloca] = allocaToValueStackMap[alloca].size();
|
||||
}
|
||||
void Mem2RegContext::renameVariables(AllocaInst *currentAlloca, BasicBlock *currentBB) {
|
||||
// 维护一个局部栈,用于存储当前基本块中为 Phi 和 Store 创建的 SSA 值,以便在退出时弹出
|
||||
std::stack<Value *> localStackPushed;
|
||||
|
||||
// --------------------------------------------------------------------
|
||||
// 处理当前基本块的指令
|
||||
// --------------------------------------------------------------------
|
||||
for (auto instIter = currentBB->getInstructions().begin(); instIter != currentBB->getInstructions().end();) {
|
||||
Instruction *inst = instIter->get();
|
||||
bool instDeleted = false;
|
||||
bool instDeleted = false;
|
||||
|
||||
// 处理 Phi 指令 (如果是当前 alloca 的 Phi)
|
||||
if (auto phiInst = dynamic_cast<PhiInst *>(inst)) {
|
||||
@@ -232,69 +227,52 @@ void Mem2RegContext::renameVariables(BasicBlock *currentBB) {
|
||||
if (allocaToPhiMap[alloca].count(currentBB) && allocaToPhiMap[alloca][currentBB] == phiInst) {
|
||||
// 为 Phi 指令的输出创建一个新的 SSA 值,并压入值栈
|
||||
allocaToValueStackMap[alloca].push(phiInst);
|
||||
if (DEBUG) {
|
||||
std::cout << "Mem2Reg: Pushed Phi " << (phiInst->getName().empty() ? "anonymous" : phiInst->getName()) << " for alloca " << alloca->getName()
|
||||
<< ". Stack size: " << allocaToValueStackMap[alloca].size() << std::endl;
|
||||
}
|
||||
break; // 找到对应的 alloca,处理下一个指令
|
||||
localStackPushed.push(phiInst); // 记录以便弹出
|
||||
break; // 找到对应的 alloca,处理下一个指令
|
||||
}
|
||||
}
|
||||
}
|
||||
// 处理 LoadInst
|
||||
else if (auto loadInst = dynamic_cast<LoadInst *>(inst)) {
|
||||
// 检查这个 LoadInst 是否是为某个可提升的 alloca
|
||||
for (auto alloca : promotableAllocas) {
|
||||
// 检查 LoadInst 的指针是否直接是 alloca,或者是指向 alloca 的 GEP
|
||||
Value *ptrOperand = loadInst->getPointer();
|
||||
if (ptrOperand == alloca || (dynamic_cast<GetElementPtrInst *>(ptrOperand) &&
|
||||
dynamic_cast<GetElementPtrInst *>(ptrOperand)->getBasePointer() == alloca)) {
|
||||
if (loadInst->getPointer() == alloca) {
|
||||
// loadInst->getPointer() 返回 AllocaInst*
|
||||
// 将 LoadInst 的所有用途替换为当前 alloca 值栈顶部的 SSA 值
|
||||
assert(!allocaToValueStackMap[alloca].empty() && "Value stack empty for alloca during load replacement!");
|
||||
if (DEBUG) {
|
||||
std::cout << "Mem2Reg: Replacing load "
|
||||
<< (ptrOperand->getName().empty() ? "anonymous" : ptrOperand->getName()) << " with SSA value "
|
||||
<< (allocaToValueStackMap[alloca].top()->getName().empty()
|
||||
? "anonymous"
|
||||
: allocaToValueStackMap[alloca].top()->getName())
|
||||
<< " for alloca " << alloca->getName() << std::endl;
|
||||
std::cout << "Mem2Reg: allocaToValueStackMap[" << alloca->getName()
|
||||
<< "] size: " << allocaToValueStackMap[alloca].size() << std::endl;
|
||||
}
|
||||
loadInst->replaceAllUsesWith(allocaToValueStackMap[alloca].top());
|
||||
instIter = SysYIROptUtils::usedelete(instIter);
|
||||
// instIter = currentBB->force_delete_inst(loadInst); // 删除 LoadInst
|
||||
SysYIROptUtils::usedelete(loadInst); // 仅删除 use 关系
|
||||
instIter = currentBB->getInstructions().erase(instIter); // 删除 LoadInst
|
||||
instDeleted = true;
|
||||
// std::cerr << "Mem2Reg: Replaced load " << loadInst->name() << " with SSA value." << std::endl;
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
// 处理 StoreInst
|
||||
else if (auto storeInst = dynamic_cast<StoreInst *>(inst)) {
|
||||
// 检查这个 StoreInst 是否是为某个可提升的 alloca
|
||||
for (auto alloca : promotableAllocas) {
|
||||
// 检查 StoreInst 的指针是否直接是 alloca,或者是指向 alloca 的 GEP
|
||||
Value *ptrOperand = storeInst->getPointer();
|
||||
if (ptrOperand == alloca || (dynamic_cast<GetElementPtrInst *>(ptrOperand) &&
|
||||
dynamic_cast<GetElementPtrInst *>(ptrOperand)->getBasePointer() == alloca)) {
|
||||
if (DEBUG) {
|
||||
std::cout << "Mem2Reg: Replacing store to "
|
||||
<< (ptrOperand->getName().empty() ? "anonymous" : ptrOperand->getName()) << " with SSA value "
|
||||
<< (storeInst->getValue()->getName().empty() ? "anonymous" : storeInst->getValue()->getName())
|
||||
<< " for alloca " << alloca->getName() << std::endl;
|
||||
std::cout << "Mem2Reg: allocaToValueStackMap[" << alloca->getName()
|
||||
<< "] size before push: " << allocaToValueStackMap[alloca].size() << std::endl;
|
||||
}
|
||||
if (storeInst->getPointer() == alloca) {
|
||||
// 假设 storeInst->getPointer() 返回 AllocaInst*
|
||||
// 将 StoreInst 存储的值作为新的 SSA 值,压入值栈
|
||||
allocaToValueStackMap[alloca].push(storeInst->getValue());
|
||||
instIter = SysYIROptUtils::usedelete(instIter);
|
||||
localStackPushed.push(storeInst->getValue()); // 记录以便弹出
|
||||
SysYIROptUtils::usedelete(storeInst);
|
||||
instIter = currentBB->getInstructions().erase(instIter); // 删除 StoreInst
|
||||
instDeleted = true;
|
||||
if (DEBUG) {
|
||||
std::cout << "Mem2Reg: allocaToValueStackMap[" << alloca->getName()
|
||||
<< "] size after push: " << allocaToValueStackMap[alloca].size() << std::endl;
|
||||
}
|
||||
// std::cerr << "Mem2Reg: Replaced store to " << storeInst->ptr()->name() << " with SSA value." << std::endl;
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (!instDeleted) {
|
||||
++instIter; // 如果指令没有被删除,移动到下一个
|
||||
}
|
||||
}
|
||||
|
||||
// --------------------------------------------------------------------
|
||||
// 处理后继基本块的 Phi 指令参数
|
||||
// --------------------------------------------------------------------
|
||||
@@ -309,57 +287,38 @@ void Mem2RegContext::renameVariables(BasicBlock *currentBB) {
|
||||
// 参数值是当前 alloca 值栈顶部的 SSA 值
|
||||
assert(!allocaToValueStackMap[alloca].empty() && "Value stack empty for alloca when setting phi operand!");
|
||||
phiInst->addIncoming(allocaToValueStackMap[alloca].top(), currentBB);
|
||||
if (DEBUG) {
|
||||
std::cout << "Mem2Reg: Added incoming arg to Phi "
|
||||
<< (phiInst->getName().empty() ? "anonymous" : phiInst->getName()) << " from "
|
||||
<< currentBB->getName() << " with value "
|
||||
<< (allocaToValueStackMap[alloca].top()->getName().empty()
|
||||
? "anonymous"
|
||||
: allocaToValueStackMap[alloca].top()->getName())
|
||||
<< std::endl;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// --------------------------------------------------------------------
|
||||
// 递归访问支配树的子节点
|
||||
// --------------------------------------------------------------------
|
||||
const std::set<BasicBlock *> *dominatedBlocks = dt->getDominatorTreeChildren(currentBB);
|
||||
if (dominatedBlocks) { // 检查是否存在子节点
|
||||
if(DEBUG){
|
||||
std::cout << "Mem2Reg: Processing dominated blocks for " << currentBB->getName() << std::endl;
|
||||
for (auto dominatedBB : *dominatedBlocks) {
|
||||
std::cout << "Mem2Reg: Dominated block: " << (dominatedBB ? dominatedBB->getName() : "null") << std::endl;
|
||||
}
|
||||
}
|
||||
if(dominatedBlocks){
|
||||
for (auto dominatedBB : *dominatedBlocks) {
|
||||
if (dominatedBB) { // 确保子块有效
|
||||
if (DEBUG) {
|
||||
std::cout << "Mem2Reg: Recursively renaming variables in dominated block: " << dominatedBB->getName()
|
||||
<< std::endl;
|
||||
}
|
||||
renameVariables(dominatedBB); // 递归调用,不再传递 currentAlloca
|
||||
if (dominatedBB) {
|
||||
std::cout << "Mem2Reg: Recursively renaming variables in dominated block: " << dominatedBB->getName() << std::endl;
|
||||
renameVariables(currentAlloca, dominatedBB);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
// --------------------------------------------------------------------
|
||||
// 退出基本块时,弹出在此块中压入值栈的 SSA 值,恢复栈到进入该块时的状态
|
||||
// 退出基本块时,弹出在此块中压入值栈的 SSA 值
|
||||
// --------------------------------------------------------------------
|
||||
for (auto alloca : promotableAllocas) {
|
||||
while (allocaToValueStackMap[alloca].size() > originalStackSizes[alloca]) {
|
||||
if (DEBUG) {
|
||||
std::cout << "Mem2Reg: Popping value "
|
||||
<< (allocaToValueStackMap[alloca].top()->getName().empty()
|
||||
? "anonymous"
|
||||
: allocaToValueStackMap[alloca].top()->getName())
|
||||
<< " for alloca " << alloca->getName() << ". Stack size: " << allocaToValueStackMap[alloca].size()
|
||||
<< " -> " << (allocaToValueStackMap[alloca].size() - 1) << std::endl;
|
||||
while (!localStackPushed.empty()) {
|
||||
Value *val = localStackPushed.top();
|
||||
localStackPushed.pop();
|
||||
// 找到是哪个 alloca 对应的栈
|
||||
for (auto alloca : promotableAllocas) {
|
||||
if (!allocaToValueStackMap[alloca].empty() && allocaToValueStackMap[alloca].top() == val) {
|
||||
allocaToValueStackMap[alloca].pop();
|
||||
break;
|
||||
}
|
||||
allocaToValueStackMap[alloca].pop();
|
||||
}
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
// 删除所有原始的 AllocaInst、LoadInst 和 StoreInst
|
||||
@@ -368,6 +327,7 @@ void Mem2RegContext::cleanup() {
|
||||
if (alloca && alloca->getParent()) {
|
||||
// 删除 alloca 指令本身
|
||||
SysYIROptUtils::usedelete(alloca);
|
||||
alloca->getParent()->removeInst(alloca); // 从基本块中删除 alloca
|
||||
|
||||
// std::cerr << "Mem2Reg: Deleted alloca " << alloca->name() << std::endl;
|
||||
}
|
||||
|
||||
@@ -74,7 +74,7 @@ void Reg2MemContext::allocateMemoryForSSAValues(Function *func) {
|
||||
// 默认情况下,将所有参数是提升到内存
|
||||
if (isPromotableToMemory(arg)) {
|
||||
// 参数的类型就是 AllocaInst 需要分配的类型
|
||||
AllocaInst *alloca = builder->createAllocaInst(Type::getPointerType(arg->getType()), arg->getName() + ".reg2mem");
|
||||
AllocaInst *alloca = builder->createAllocaInst(Type::getPointerType(arg->getType()), {}, arg->getName() + ".reg2mem");
|
||||
// 将参数值 store 到 alloca 中 (这是 Mem2Reg 逆转的关键一步)
|
||||
valueToAllocaMap[arg] = alloca;
|
||||
|
||||
@@ -103,7 +103,7 @@ void Reg2MemContext::allocateMemoryForSSAValues(Function *func) {
|
||||
// AllocaInst 应该在入口块,而不是当前指令所在块
|
||||
// 这里我们只是创建,并稍后调整其位置
|
||||
// 通常的做法是在循环结束后统一将 alloca 放到 entryBlock 的顶部
|
||||
AllocaInst *alloca = builder->createAllocaInst(Type::getPointerType(inst.get()->getType()), inst.get()->getName() + ".reg2mem");
|
||||
AllocaInst *alloca = builder->createAllocaInst(Type::getPointerType(inst.get()->getType()), {}, inst.get()->getName() + ".reg2mem");
|
||||
valueToAllocaMap[inst.get()] = alloca;
|
||||
}
|
||||
}
|
||||
@@ -148,8 +148,8 @@ void Reg2MemContext::rewritePhis(Function *func) {
|
||||
// 1. 为 Phi 指令的每个入边,在前驱块的末尾插入 Store 指令
|
||||
// PhiInst 假设有 getIncomingValues() 和 getIncomingBlocks()
|
||||
for (unsigned i = 0; i < phiInst->getNumIncomingValues(); ++i) { // 假设 PhiInst 是通过操作数来管理入边的
|
||||
Value *incomingValue = phiInst->getIncomingValue(i); // 获取入值
|
||||
BasicBlock *incomingBlock = phiInst->getIncomingBlock(i); // 获取对应的入块
|
||||
Value *incomingValue = phiInst->getValue(i); // 获取入值
|
||||
BasicBlock *incomingBlock = phiInst->getBlock(i); // 获取对应的入块
|
||||
|
||||
// 在入块的跳转指令之前插入 StoreInst
|
||||
// 需要找到 incomingBlock 的终结指令 (Terminator Instruction)
|
||||
@@ -181,7 +181,8 @@ void Reg2MemContext::rewritePhis(Function *func) {
|
||||
// 实际删除 Phi 指令
|
||||
for (auto phi : phisToErase) {
|
||||
if (phi && phi->getParent()) {
|
||||
SysYIROptUtils::usedelete(phi);
|
||||
SysYIROptUtils::usedelete(phi); // 清理 use-def 链
|
||||
phi->getParent()->removeInst(phi); // 从基本块中删除
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -1,880 +0,0 @@
|
||||
#include "SCCP.h"
|
||||
#include "Dom.h"
|
||||
#include "Liveness.h"
|
||||
#include <algorithm>
|
||||
#include <cassert>
|
||||
#include <cmath> // For std::fmod, std::fabs
|
||||
#include <limits> // For std::numeric_limits
|
||||
|
||||
namespace sysy {
|
||||
|
||||
// Pass ID for SCCP
|
||||
void *SCCP::ID = (void *)&SCCP::ID;
|
||||
|
||||
// SCCPContext methods
|
||||
SSAPValue SCCPContext::Meet(const SSAPValue &a, const SSAPValue &b) {
|
||||
if (a.state == LatticeVal::Bottom || b.state == LatticeVal::Bottom) {
|
||||
return SSAPValue(LatticeVal::Bottom);
|
||||
}
|
||||
if (a.state == LatticeVal::Top) {
|
||||
return b;
|
||||
}
|
||||
if (b.state == LatticeVal::Top) {
|
||||
return a;
|
||||
}
|
||||
// Both are constants
|
||||
if (a.constant_type != b.constant_type) {
|
||||
return SSAPValue(LatticeVal::Bottom); // 不同类型的常量,结果为 Bottom
|
||||
}
|
||||
if (a.constantVal == b.constantVal) {
|
||||
return a; // 相同常量
|
||||
}
|
||||
return SSAPValue(LatticeVal::Bottom); // 相同类型但值不同,结果为 Bottom
|
||||
}
|
||||
|
||||
SSAPValue SCCPContext::GetValueState(Value *v) {
|
||||
if (auto constVal = dynamic_cast<ConstantValue *>(v)) {
|
||||
// 特殊处理 UndefinedValue:将其视为 Bottom
|
||||
if (dynamic_cast<UndefinedValue *>(constVal)) {
|
||||
return SSAPValue(LatticeVal::Bottom);
|
||||
}
|
||||
// 处理常规的 ConstantInteger 和 ConstantFloating
|
||||
if (constVal->getType()->isInt()) {
|
||||
return SSAPValue(constVal->getInt());
|
||||
} else if (constVal->getType()->isFloat()) {
|
||||
return SSAPValue(constVal->getFloat());
|
||||
} else {
|
||||
// 对于其他 ConstantValue 类型(例如,ConstantArray 等),
|
||||
// 如果它们的具体值不能用于标量常量传播,则保守地视为 Bottom。
|
||||
return SSAPValue(LatticeVal::Bottom);
|
||||
}
|
||||
}
|
||||
if (valueState.count(v)) {
|
||||
return valueState[v];
|
||||
}
|
||||
return SSAPValue(); // 默认初始化为 Top
|
||||
}
|
||||
|
||||
void SCCPContext::UpdateState(Value *v, SSAPValue newState) {
|
||||
SSAPValue oldState = GetValueState(v);
|
||||
if (newState != oldState) {
|
||||
if (DEBUG) {
|
||||
std::cout << "Updating state for " << v->getName() << " from (";
|
||||
if (oldState.state == LatticeVal::Top)
|
||||
std::cout << "Top";
|
||||
else if (oldState.state == LatticeVal::Constant) {
|
||||
if (oldState.constant_type == ValueType::Integer)
|
||||
std::cout << "Const<int>(" << std::get<int>(oldState.constantVal) << ")";
|
||||
else
|
||||
std::cout << "Const<float>(" << std::get<float>(oldState.constantVal) << ")";
|
||||
} else
|
||||
std::cout << "Bottom";
|
||||
std::cout << ") to (";
|
||||
if (newState.state == LatticeVal::Top)
|
||||
std::cout << "Top";
|
||||
else if (newState.state == LatticeVal::Constant) {
|
||||
if (newState.constant_type == ValueType::Integer)
|
||||
std::cout << "Const<int>(" << std::get<int>(newState.constantVal) << ")";
|
||||
else
|
||||
std::cout << "Const<float>(" << std::get<float>(newState.constantVal) << ")";
|
||||
} else
|
||||
std::cout << "Bottom";
|
||||
std::cout << ")" << std::endl;
|
||||
}
|
||||
|
||||
valueState[v] = newState;
|
||||
// 如果状态发生变化,将所有使用者添加到指令工作列表
|
||||
for (auto &use_ptr : v->getUses()) {
|
||||
if (auto userInst = dynamic_cast<Instruction *>(use_ptr->getUser())) {
|
||||
instWorkList.push(userInst);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void SCCPContext::AddEdgeToWorkList(BasicBlock *fromBB, BasicBlock *toBB) {
|
||||
// 检查边是否已经访问过,防止重复处理
|
||||
if (visitedCFGEdges.count({fromBB, toBB})) {
|
||||
return;
|
||||
}
|
||||
visitedCFGEdges.insert({fromBB, toBB});
|
||||
|
||||
if (DEBUG) {
|
||||
std::cout << "Adding edge to worklist: " << fromBB->getName() << " -> " << toBB->getName() << std::endl;
|
||||
}
|
||||
edgeWorkList.push({fromBB, toBB});
|
||||
}
|
||||
|
||||
void SCCPContext::MarkBlockExecutable(BasicBlock *block) {
|
||||
if (executableBlocks.insert(block).second) { // insert 返回 pair,second 为 true 表示插入成功
|
||||
if (DEBUG) {
|
||||
std::cout << "Marking block " << block->getName() << " as executable." << std::endl;
|
||||
}
|
||||
// 将新可执行块中的所有指令添加到指令工作列表
|
||||
for (auto &inst_ptr : block->getInstructions()) {
|
||||
instWorkList.push(inst_ptr.get());
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 辅助函数:对二元操作进行常量折叠
|
||||
SSAPValue SCCPContext::ComputeConstant(BinaryInst *binaryInst, SSAPValue lhsVal, SSAPValue rhsVal) {
|
||||
// 确保操作数是常量
|
||||
if (lhsVal.state != LatticeVal::Constant || rhsVal.state != LatticeVal::Constant) {
|
||||
return SSAPValue(LatticeVal::Bottom); // 如果不是常量,则不能折叠
|
||||
}
|
||||
|
||||
// 处理整数运算 (kAdd, kSub, kMul, kDiv, kRem, kICmp*, kAnd, kOr)
|
||||
if (lhsVal.constant_type == ValueType::Integer && rhsVal.constant_type == ValueType::Integer) {
|
||||
int lhs = std::get<int>(lhsVal.constantVal);
|
||||
int rhs = std::get<int>(rhsVal.constantVal);
|
||||
int result = 0;
|
||||
|
||||
switch (binaryInst->getKind()) {
|
||||
case Instruction::kAdd:
|
||||
result = lhs + rhs;
|
||||
break;
|
||||
case Instruction::kSub:
|
||||
result = lhs - rhs;
|
||||
break;
|
||||
case Instruction::kMul:
|
||||
result = lhs * rhs;
|
||||
break;
|
||||
case Instruction::kDiv:
|
||||
if (rhs == 0)
|
||||
return SSAPValue(LatticeVal::Bottom); // 除零
|
||||
result = lhs / rhs;
|
||||
break;
|
||||
case Instruction::kRem:
|
||||
if (rhs == 0)
|
||||
return SSAPValue(LatticeVal::Bottom); // 模零
|
||||
result = lhs % rhs;
|
||||
break;
|
||||
case Instruction::kICmpEQ:
|
||||
result = (lhs == rhs);
|
||||
break;
|
||||
case Instruction::kICmpNE:
|
||||
result = (lhs != rhs);
|
||||
break;
|
||||
case Instruction::kICmpLT:
|
||||
result = (lhs < rhs);
|
||||
break;
|
||||
case Instruction::kICmpGT:
|
||||
result = (lhs > rhs);
|
||||
break;
|
||||
case Instruction::kICmpLE:
|
||||
result = (lhs <= rhs);
|
||||
break;
|
||||
case Instruction::kICmpGE:
|
||||
result = (lhs >= rhs);
|
||||
break;
|
||||
case Instruction::kAnd:
|
||||
result = (lhs && rhs);
|
||||
break;
|
||||
case Instruction::kOr:
|
||||
result = (lhs || rhs);
|
||||
break;
|
||||
default:
|
||||
return SSAPValue(LatticeVal::Bottom); // 未知或不匹配的二元操作
|
||||
}
|
||||
return SSAPValue(result);
|
||||
}
|
||||
// 处理浮点运算 (kFAdd, kFSub, kFMul, kFDiv, kFCmp*)
|
||||
else if (lhsVal.constant_type == ValueType::Float && rhsVal.constant_type == ValueType::Float) {
|
||||
float lhs = std::get<float>(lhsVal.constantVal);
|
||||
float rhs = std::get<float>(rhsVal.constantVal);
|
||||
float f_result = 0.0f;
|
||||
int i_result = 0; // For comparison results
|
||||
|
||||
switch (binaryInst->getKind()) {
|
||||
case Instruction::kFAdd:
|
||||
f_result = lhs + rhs;
|
||||
break;
|
||||
case Instruction::kFSub:
|
||||
f_result = lhs - rhs;
|
||||
break;
|
||||
case Instruction::kFMul:
|
||||
f_result = lhs * rhs;
|
||||
break;
|
||||
case Instruction::kFDiv:
|
||||
if (rhs == 0.0f)
|
||||
return SSAPValue(LatticeVal::Bottom); // 除零
|
||||
f_result = lhs / rhs;
|
||||
break;
|
||||
// kRem 不支持浮点数,但如果你的 IR 定义了浮点模运算,需要使用 std::fmod
|
||||
case Instruction::kFCmpEQ:
|
||||
i_result = (lhs == rhs);
|
||||
return SSAPValue(i_result);
|
||||
case Instruction::kFCmpNE:
|
||||
i_result = (lhs != rhs);
|
||||
return SSAPValue(i_result);
|
||||
case Instruction::kFCmpLT:
|
||||
i_result = (lhs < rhs);
|
||||
return SSAPValue(i_result);
|
||||
case Instruction::kFCmpGT:
|
||||
i_result = (lhs > rhs);
|
||||
return SSAPValue(i_result);
|
||||
case Instruction::kFCmpLE:
|
||||
i_result = (lhs <= rhs);
|
||||
return SSAPValue(i_result);
|
||||
case Instruction::kFCmpGE:
|
||||
i_result = (lhs >= rhs);
|
||||
return SSAPValue(i_result);
|
||||
default:
|
||||
return SSAPValue(LatticeVal::Bottom); // 未知或不匹配的浮点二元操作
|
||||
}
|
||||
return SSAPValue(f_result);
|
||||
}
|
||||
|
||||
return SSAPValue(LatticeVal::Bottom); // 类型不匹配或不支持的类型组合
|
||||
}
|
||||
|
||||
// 辅助函数:对一元操作进行常量折叠
|
||||
SSAPValue SCCPContext::ComputeConstant(UnaryInst *unaryInst, SSAPValue operandVal) {
|
||||
if (operandVal.state != LatticeVal::Constant) {
|
||||
return SSAPValue(LatticeVal::Bottom);
|
||||
}
|
||||
|
||||
if (operandVal.constant_type == ValueType::Integer) {
|
||||
int val = std::get<int>(operandVal.constantVal);
|
||||
switch (unaryInst->getKind()) {
|
||||
case Instruction::kAdd:
|
||||
return SSAPValue(val);
|
||||
case Instruction::kNeg:
|
||||
return SSAPValue(-val);
|
||||
case Instruction::kNot:
|
||||
return SSAPValue(!val);
|
||||
default:
|
||||
return SSAPValue(LatticeVal::Bottom);
|
||||
}
|
||||
} else if (operandVal.constant_type == ValueType::Float) {
|
||||
float val = std::get<float>(operandVal.constantVal);
|
||||
switch (unaryInst->getKind()) {
|
||||
case Instruction::kAdd:
|
||||
return SSAPValue(val);
|
||||
case Instruction::kFNeg:
|
||||
return SSAPValue(-val);
|
||||
case Instruction::kFNot:
|
||||
return SSAPValue(static_cast<int>(val == 0.0f)); // 浮点数非,0.0f 为真,其他为假
|
||||
default:
|
||||
return SSAPValue(LatticeVal::Bottom);
|
||||
}
|
||||
}
|
||||
return SSAPValue(LatticeVal::Bottom);
|
||||
}
|
||||
|
||||
// 辅助函数:处理单条指令
|
||||
void SCCPContext::ProcessInstruction(Instruction *inst) {
|
||||
SSAPValue oldState = GetValueState(inst);
|
||||
SSAPValue newState;
|
||||
|
||||
if (!executableBlocks.count(inst->getParent())) {
|
||||
// 如果指令所在的块不可执行,其值应保持 Top
|
||||
// 除非它之前已经是 Bottom,因为 Bottom 是单调的
|
||||
if (oldState.state != LatticeVal::Bottom) {
|
||||
newState = SSAPValue(); // Top
|
||||
} else {
|
||||
newState = oldState; // 保持 Bottom
|
||||
}
|
||||
UpdateState(inst, newState);
|
||||
return; // 不处理不可达块中的指令的实际值
|
||||
}
|
||||
|
||||
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::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:
|
||||
case Instruction::kAnd:
|
||||
case Instruction::kOr: {
|
||||
BinaryInst *binaryInst = static_cast<BinaryInst *>(inst);
|
||||
SSAPValue lhs = GetValueState(binaryInst->getOperand(0));
|
||||
SSAPValue rhs = GetValueState(binaryInst->getOperand(1));
|
||||
// 如果任一操作数是 Bottom,结果就是 Bottom
|
||||
if (lhs.state == LatticeVal::Bottom || rhs.state == LatticeVal::Bottom) {
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
} else if (lhs.state == LatticeVal::Top || rhs.state == LatticeVal::Top) {
|
||||
newState = SSAPValue(); // Top
|
||||
} else { // 都是常量
|
||||
newState = ComputeConstant(binaryInst, lhs, rhs);
|
||||
}
|
||||
break;
|
||||
}
|
||||
case Instruction::kNeg:
|
||||
case Instruction::kNot:
|
||||
case Instruction::kFNeg:
|
||||
case Instruction::kFNot: {
|
||||
UnaryInst *unaryInst = static_cast<UnaryInst *>(inst);
|
||||
SSAPValue operand = GetValueState(unaryInst->getOperand());
|
||||
if (operand.state == LatticeVal::Bottom) {
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
} else if (operand.state == LatticeVal::Top) {
|
||||
newState = SSAPValue(); // Top
|
||||
} else { // 是常量
|
||||
newState = ComputeConstant(unaryInst, operand);
|
||||
}
|
||||
break;
|
||||
}
|
||||
// 直接处理类型转换指令
|
||||
case Instruction::kFtoI: {
|
||||
SSAPValue operand = GetValueState(inst->getOperand(0));
|
||||
if (operand.state == LatticeVal::Constant && operand.constant_type == ValueType::Float) {
|
||||
newState = SSAPValue(static_cast<int>(std::get<float>(operand.constantVal)));
|
||||
} else if (operand.state == LatticeVal::Bottom) {
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
} else { // Top
|
||||
newState = SSAPValue();
|
||||
}
|
||||
break;
|
||||
}
|
||||
case Instruction::kItoF: {
|
||||
SSAPValue operand = GetValueState(inst->getOperand(0));
|
||||
if (operand.state == LatticeVal::Constant && operand.constant_type == ValueType::Integer) {
|
||||
newState = SSAPValue(static_cast<float>(std::get<int>(operand.constantVal)));
|
||||
} else if (operand.state == LatticeVal::Bottom) {
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
} else { // Top
|
||||
newState = SSAPValue();
|
||||
}
|
||||
break;
|
||||
}
|
||||
case Instruction::kBitFtoI: {
|
||||
SSAPValue operand = GetValueState(inst->getOperand(0));
|
||||
if (operand.state == LatticeVal::Constant && operand.constant_type == ValueType::Float) {
|
||||
float fval = std::get<float>(operand.constantVal);
|
||||
newState = SSAPValue(*reinterpret_cast<int *>(&fval));
|
||||
} else if (operand.state == LatticeVal::Bottom) {
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
} else { // Top
|
||||
newState = SSAPValue();
|
||||
}
|
||||
break;
|
||||
}
|
||||
case Instruction::kBitItoF: {
|
||||
SSAPValue operand = GetValueState(inst->getOperand(0));
|
||||
if (operand.state == LatticeVal::Constant && operand.constant_type == ValueType::Integer) {
|
||||
int ival = std::get<int>(operand.constantVal);
|
||||
newState = SSAPValue(*reinterpret_cast<float *>(&ival));
|
||||
} else if (operand.state == LatticeVal::Bottom) {
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
} else { // Top
|
||||
newState = SSAPValue();
|
||||
}
|
||||
break;
|
||||
}
|
||||
case Instruction::kLoad: {
|
||||
// 对于 Load 指令,除非我们有特殊的别名分析,否则假定为 Bottom
|
||||
// 或者如果它加载的是一个已知常量地址的全局常量
|
||||
Value *ptr = inst->getOperand(0);
|
||||
if (auto globalVal = dynamic_cast<GlobalValue *>(ptr)) {
|
||||
// 如果 GlobalValue 有初始化器,并且它是常量,我们可以传播
|
||||
// 这需要额外的逻辑来检查 globalVal 的初始化器
|
||||
// 暂时保守地设置为 Bottom
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
} else {
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
}
|
||||
break;
|
||||
}
|
||||
case Instruction::kStore:
|
||||
// Store 指令不产生值,其 SSAPValue 不重要
|
||||
newState = SSAPValue(); // 保持 Top
|
||||
break;
|
||||
case Instruction::kCall:
|
||||
// 大多数 Call 指令都假定为 Bottom,除非是纯函数且所有参数都是常量
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
break;
|
||||
case Instruction::kGetElementPtr: {
|
||||
// GEP 指令计算地址,通常其结果值(地址指向的内容)是 Bottom
|
||||
// 除非所有索引和基指针都是常量,指向一个确定常量值的内存位置
|
||||
bool all_ops_constant = true;
|
||||
for (unsigned i = 0; i < inst->getNumOperands(); ++i) {
|
||||
if (GetValueState(inst->getOperand(i)).state != LatticeVal::Constant) {
|
||||
all_ops_constant = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
// 即使地址是常量,地址处的内容通常不是。所以通常是 Bottom
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
break;
|
||||
}
|
||||
case Instruction::kPhi: {
|
||||
PhiInst *phi = static_cast<PhiInst *>(inst);
|
||||
SSAPValue phiResult = SSAPValue(); // 初始为 Top
|
||||
|
||||
for (unsigned i = 0; i < phi->getNumIncomingValues(); ++i) {
|
||||
Value *incomingVal = phi->getIncomingValue(i);
|
||||
BasicBlock *incomingBlock = phi->getIncomingBlock(i);
|
||||
|
||||
if (executableBlocks.count(incomingBlock)) { // 仅考虑可执行前驱
|
||||
phiResult = Meet(phiResult, GetValueState(incomingVal));
|
||||
if (phiResult.state == LatticeVal::Bottom)
|
||||
break; // 如果已经 Bottom,则提前退出
|
||||
}
|
||||
}
|
||||
newState = phiResult;
|
||||
break;
|
||||
}
|
||||
case Instruction::kAlloca: // 对应 kAlloca
|
||||
// Alloca 分配内存,返回一个指针,其内容是 Bottom
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
break;
|
||||
case Instruction::kBr: // 对应 kBr
|
||||
case Instruction::kCondBr: // 对应 kCondBr
|
||||
case Instruction::kReturn: // 对应 kReturn
|
||||
case Instruction::kUnreachable: // 对应 kUnreachable
|
||||
// 终结符指令不产生值
|
||||
newState = SSAPValue(); // 保持 Top
|
||||
break;
|
||||
case Instruction::kMemset:
|
||||
// Memset 不产生值,但有副作用,不进行常量传播
|
||||
newState = SSAPValue(LatticeVal::Bottom);
|
||||
break;
|
||||
default:
|
||||
if (DEBUG) {
|
||||
std::cout << "Unimplemented instruction kind in SCCP: " << inst->getKind() << std::endl;
|
||||
}
|
||||
newState = SSAPValue(LatticeVal::Bottom); // 未知指令保守处理为 Bottom
|
||||
break;
|
||||
}
|
||||
UpdateState(inst, newState);
|
||||
|
||||
// 特殊处理终结符指令,影响 CFG 边的可达性
|
||||
if (inst->isTerminator()) {
|
||||
if (inst->isBranch()) {
|
||||
|
||||
if (inst->isCondBr()) { // 使用 kCondBr
|
||||
CondBrInst *branchInst = static_cast<CondBrInst *>(inst);
|
||||
SSAPValue condVal = GetValueState(branchInst->getOperand(0));
|
||||
if (condVal.state == LatticeVal::Constant) {
|
||||
bool condition_is_true = false;
|
||||
if (condVal.constant_type == ValueType::Integer) {
|
||||
condition_is_true = (std::get<int>(condVal.constantVal) != 0);
|
||||
} else if (condVal.constant_type == ValueType::Float) {
|
||||
condition_is_true = (std::get<float>(condVal.constantVal) != 0.0f);
|
||||
}
|
||||
|
||||
if (condition_is_true) {
|
||||
AddEdgeToWorkList(branchInst->getParent(), branchInst->getThenBlock());
|
||||
} else {
|
||||
AddEdgeToWorkList(branchInst->getParent(), branchInst->getElseBlock());
|
||||
}
|
||||
} else { // 条件是 Top 或 Bottom,两条路径都可能
|
||||
AddEdgeToWorkList(branchInst->getParent(), branchInst->getThenBlock());
|
||||
AddEdgeToWorkList(branchInst->getParent(), branchInst->getElseBlock());
|
||||
}
|
||||
} else { // 无条件分支 (kBr)
|
||||
UncondBrInst *branchInst = static_cast<UncondBrInst *>(inst);
|
||||
AddEdgeToWorkList(branchInst->getParent(), branchInst->getBlock());
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 辅助函数:处理单条控制流边
|
||||
void SCCPContext::ProcessEdge(const std::pair<BasicBlock *, BasicBlock *> &edge) {
|
||||
BasicBlock *fromBB = edge.first;
|
||||
BasicBlock *toBB = edge.second;
|
||||
|
||||
MarkBlockExecutable(toBB);
|
||||
|
||||
// 对于目标块中的所有 Phi 指令,重新评估其值,因为可能有新的前驱被激活
|
||||
for (auto &inst_ptr : toBB->getInstructions()) {
|
||||
if (dynamic_cast<PhiInst *>(inst_ptr.get())) {
|
||||
instWorkList.push(inst_ptr.get());
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 阶段1: 常量传播与折叠
|
||||
bool SCCPContext::PropagateConstants(Function *func) {
|
||||
bool changed = false;
|
||||
|
||||
// 初始化:所有值 Top,所有块不可执行
|
||||
for (auto &bb_ptr : func->getBasicBlocks()) {
|
||||
executableBlocks.erase(bb_ptr.get());
|
||||
for (auto &inst_ptr : bb_ptr->getInstructions()) {
|
||||
valueState[inst_ptr.get()] = SSAPValue(); // Top
|
||||
}
|
||||
}
|
||||
|
||||
// 标记入口块为可执行
|
||||
if (!func->getBasicBlocks().empty()) {
|
||||
MarkBlockExecutable(func->getEntryBlock());
|
||||
}
|
||||
|
||||
// 主循环:处理工作列表直到不动点
|
||||
while (!instWorkList.empty() || !edgeWorkList.empty()) {
|
||||
while (!edgeWorkList.empty()) {
|
||||
ProcessEdge(edgeWorkList.front());
|
||||
edgeWorkList.pop();
|
||||
}
|
||||
|
||||
while (!instWorkList.empty()) {
|
||||
Instruction *inst = instWorkList.front();
|
||||
instWorkList.pop();
|
||||
ProcessInstruction(inst);
|
||||
}
|
||||
}
|
||||
|
||||
// 应用常量替换和死代码消除
|
||||
std::vector<Instruction *> instsToDelete;
|
||||
for (auto &bb_ptr : func->getBasicBlocks()) {
|
||||
BasicBlock *bb = bb_ptr.get();
|
||||
if (!executableBlocks.count(bb)) {
|
||||
// 整个块是死块,标记所有指令删除
|
||||
for (auto &inst_ptr : bb->getInstructions()) {
|
||||
instsToDelete.push_back(inst_ptr.get());
|
||||
}
|
||||
changed = true;
|
||||
continue;
|
||||
}
|
||||
for (auto it = bb->begin(); it != bb->end();) {
|
||||
Instruction *inst = it->get();
|
||||
SSAPValue ssaPVal = GetValueState(inst);
|
||||
|
||||
if (ssaPVal.state == LatticeVal::Constant) {
|
||||
ConstantValue *constVal = nullptr;
|
||||
if (ssaPVal.constant_type == ValueType::Integer) {
|
||||
constVal = ConstantInteger::get(std::get<int>(ssaPVal.constantVal));
|
||||
} else if (ssaPVal.constant_type == ValueType::Float) {
|
||||
constVal = ConstantFloating::get(std::get<float>(ssaPVal.constantVal));
|
||||
} else {
|
||||
constVal = UndefinedValue::get(inst->getType()); // 不应发生
|
||||
}
|
||||
|
||||
if (DEBUG) {
|
||||
std::cout << "Replacing " << inst->getName() << " with constant ";
|
||||
if (ssaPVal.constant_type == ValueType::Integer)
|
||||
std::cout << std::get<int>(ssaPVal.constantVal);
|
||||
else
|
||||
std::cout << std::get<float>(ssaPVal.constantVal);
|
||||
std::cout << std::endl;
|
||||
}
|
||||
inst->replaceAllUsesWith(constVal);
|
||||
instsToDelete.push_back(inst);
|
||||
++it;
|
||||
changed = true;
|
||||
} else {
|
||||
// 如果操作数是常量,直接替换为常量值(常量折叠)
|
||||
for (unsigned i = 0; i < inst->getNumOperands(); ++i) {
|
||||
Value *operand = inst->getOperand(i);
|
||||
SSAPValue opVal = GetValueState(operand);
|
||||
if (opVal.state == LatticeVal::Constant) {
|
||||
ConstantValue *constOp = nullptr;
|
||||
if (opVal.constant_type == ValueType::Integer) {
|
||||
constOp = ConstantInteger::get(std::get<int>(opVal.constantVal));
|
||||
} else if (opVal.constant_type == ValueType::Float) {
|
||||
constOp = ConstantFloating::get(std::get<float>(opVal.constantVal));
|
||||
} else {
|
||||
constOp = UndefinedValue::get(operand->getType());
|
||||
}
|
||||
|
||||
if (constOp != operand) {
|
||||
inst->setOperand(i, constOp);
|
||||
changed = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
++it;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 实际删除指令
|
||||
// TODO: 删除的逻辑需要考虑修改
|
||||
for (Instruction *inst : instsToDelete) {
|
||||
// 在尝试删除之前,先检查指令是否仍然附加到其父基本块。
|
||||
// 如果它已经没有父块,可能说明它已被其他方式处理或已处于无效状态。
|
||||
if (inst->getParent() != nullptr) {
|
||||
// 调用负责完整删除的函数,该函数应负责清除uses并将其从父块中移除。
|
||||
SysYIROptUtils::usedelete(inst);
|
||||
}
|
||||
else {
|
||||
// 指令已不属于任何父块,无需再次删除。
|
||||
if (DEBUG) {
|
||||
std::cerr << "Info: Instruction " << inst->getName() << " was already detached or is not in a parent block." << std::endl;
|
||||
}
|
||||
}
|
||||
}
|
||||
return changed;
|
||||
}
|
||||
|
||||
// 阶段2: 控制流简化
|
||||
bool SCCPContext::SimplifyControlFlow(Function *func) {
|
||||
bool changed = false;
|
||||
|
||||
// 重新确定可达块,因为 PropagateConstants 可能改变了分支条件
|
||||
std::unordered_set<BasicBlock *> newReachableBlocks = FindReachableBlocks(func);
|
||||
|
||||
// 移除不可达块
|
||||
std::vector<BasicBlock *> blocksToDelete;
|
||||
for (auto &bb_ptr : func->getBasicBlocks()) {
|
||||
if (bb_ptr.get() == func->getEntryBlock())
|
||||
continue; // 入口块不能删除
|
||||
if (newReachableBlocks.find(bb_ptr.get()) == newReachableBlocks.end()) {
|
||||
blocksToDelete.push_back(bb_ptr.get());
|
||||
changed = true;
|
||||
}
|
||||
}
|
||||
for (BasicBlock *bb : blocksToDelete) {
|
||||
RemoveDeadBlock(bb, func);
|
||||
}
|
||||
|
||||
// 简化分支指令
|
||||
for (auto &bb_ptr : func->getBasicBlocks()) {
|
||||
BasicBlock *bb = bb_ptr.get();
|
||||
if (!newReachableBlocks.count(bb))
|
||||
continue; // 只处理可达块
|
||||
|
||||
Instruction *terminator = bb->terminator()->get();
|
||||
if (terminator->isBranch()) {
|
||||
|
||||
if (terminator->isCondBr()) { // 检查是否是条件分支 (kCondBr)
|
||||
CondBrInst *branchInst = static_cast<CondBrInst *>(terminator);
|
||||
SSAPValue condVal = GetValueState(branchInst->getOperand(0));
|
||||
if (condVal.state == LatticeVal::Constant) {
|
||||
bool condition_is_true = false;
|
||||
if (condVal.constant_type == ValueType::Integer) {
|
||||
condition_is_true = (std::get<int>(condVal.constantVal) != 0);
|
||||
} else if (condVal.constant_type == ValueType::Float) {
|
||||
condition_is_true = (std::get<float>(condVal.constantVal) != 0.0f);
|
||||
}
|
||||
SimplifyBranch(branchInst, condition_is_true);
|
||||
changed = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return changed;
|
||||
}
|
||||
|
||||
// 查找所有可达的基本块 (基于常量条件)
|
||||
std::unordered_set<BasicBlock *> SCCPContext::FindReachableBlocks(Function *func) {
|
||||
std::unordered_set<BasicBlock *> reachable;
|
||||
std::queue<BasicBlock *> q;
|
||||
|
||||
if (func->getEntryBlock()) {
|
||||
q.push(func->getEntryBlock());
|
||||
reachable.insert(func->getEntryBlock());
|
||||
}
|
||||
|
||||
while (!q.empty()) {
|
||||
BasicBlock *currentBB = q.front();
|
||||
q.pop();
|
||||
|
||||
Instruction *terminator = currentBB->terminator()->get();
|
||||
if (!terminator)
|
||||
continue;
|
||||
|
||||
if (terminator->isBranch()) {
|
||||
if (terminator->isCondBr()) { // 检查是否是条件分支 (kCondBr)
|
||||
CondBrInst *branchInst = static_cast<CondBrInst *>(terminator);
|
||||
SSAPValue condVal = GetValueState(branchInst->getOperand(0));
|
||||
if (condVal.state == LatticeVal::Constant) {
|
||||
bool condition_is_true = false;
|
||||
if (condVal.constant_type == ValueType::Integer) {
|
||||
condition_is_true = (std::get<int>(condVal.constantVal) != 0);
|
||||
} else if (condVal.constant_type == ValueType::Float) {
|
||||
condition_is_true = (std::get<float>(condVal.constantVal) != 0.0f);
|
||||
}
|
||||
|
||||
if (condition_is_true) {
|
||||
BasicBlock *trueBlock = branchInst->getThenBlock();
|
||||
if (reachable.find(trueBlock) == reachable.end()) {
|
||||
reachable.insert(trueBlock);
|
||||
q.push(trueBlock);
|
||||
}
|
||||
} else {
|
||||
BasicBlock *falseBlock = branchInst->getElseBlock();
|
||||
if (reachable.find(falseBlock) == reachable.end()) {
|
||||
reachable.insert(falseBlock);
|
||||
q.push(falseBlock);
|
||||
}
|
||||
}
|
||||
} else { // 条件是 Top 或 Bottom,两条路径都可达
|
||||
for (auto succ : branchInst->getSuccessors()) {
|
||||
if (reachable.find(succ) == reachable.end()) {
|
||||
reachable.insert(succ);
|
||||
q.push(succ);
|
||||
}
|
||||
}
|
||||
}
|
||||
} else { // 无条件分支 (kBr)
|
||||
UncondBrInst *branchInst = static_cast<UncondBrInst *>(terminator);
|
||||
BasicBlock *targetBlock = branchInst->getBlock();
|
||||
if (reachable.find(targetBlock) == reachable.end()) {
|
||||
reachable.insert(targetBlock);
|
||||
q.push(targetBlock);
|
||||
}
|
||||
}
|
||||
} else if (terminator->isReturn() || terminator->isUnreachable()) {
|
||||
// ReturnInst 没有后继,不需要处理
|
||||
// UnreachableInst 也没有后继,不需要处理
|
||||
}
|
||||
}
|
||||
return reachable;
|
||||
}
|
||||
|
||||
// 移除死块
|
||||
void SCCPContext::RemoveDeadBlock(BasicBlock *bb, Function *func) {
|
||||
if (DEBUG) {
|
||||
std::cout << "Removing dead block: " << bb->getName() << std::endl;
|
||||
}
|
||||
// 首先更新其所有前驱的终结指令,移除指向死块的边
|
||||
std::vector<BasicBlock *> preds_to_update;
|
||||
for (auto &pred : bb->getPredecessors()) {
|
||||
if (pred != nullptr) { // 检查是否为空指针
|
||||
preds_to_update.push_back(pred);
|
||||
}
|
||||
}
|
||||
for (BasicBlock *pred : preds_to_update) {
|
||||
if (executableBlocks.count(pred)) {
|
||||
UpdateTerminator(pred, bb);
|
||||
}
|
||||
}
|
||||
|
||||
// 移除其后继的 Phi 节点的入边
|
||||
std::vector<BasicBlock *> succs_to_update;
|
||||
for (auto succ : bb->getSuccessors()) {
|
||||
succs_to_update.push_back(succ);
|
||||
}
|
||||
for (BasicBlock *succ : succs_to_update) {
|
||||
RemovePhiIncoming(succ, bb);
|
||||
succ->removePredecessor(bb);
|
||||
}
|
||||
|
||||
func->removeBasicBlock(bb); // 从函数中移除基本块
|
||||
}
|
||||
|
||||
// 简化分支(将条件分支替换为无条件分支)
|
||||
void SCCPContext::SimplifyBranch(CondBrInst *brInst, bool condVal) {
|
||||
BasicBlock *parentBB = brInst->getParent();
|
||||
BasicBlock *trueBlock = brInst->getThenBlock();
|
||||
BasicBlock *falseBlock = brInst->getElseBlock();
|
||||
|
||||
if (DEBUG) {
|
||||
std::cout << "Simplifying branch in " << parentBB->getName() << ": cond is " << (condVal ? "true" : "false")
|
||||
<< std::endl;
|
||||
}
|
||||
|
||||
builder->setPosition(parentBB, parentBB->findInstIterator(brInst));
|
||||
if (condVal) { // 条件为真,跳转到真分支
|
||||
builder->createUncondBrInst(trueBlock); // 插入无条件分支 kBr
|
||||
SysYIROptUtils::usedelete(brInst); // 移除旧的条件分支指令
|
||||
parentBB->removeSuccessor(falseBlock);
|
||||
falseBlock->removePredecessor(parentBB);
|
||||
RemovePhiIncoming(falseBlock, parentBB);
|
||||
} else { // 条件为假,跳转到假分支
|
||||
builder->createUncondBrInst(falseBlock); // 插入无条件分支 kBr
|
||||
SysYIROptUtils::usedelete(brInst); // 移除旧的条件分支指令
|
||||
parentBB->removeSuccessor(trueBlock);
|
||||
trueBlock->removePredecessor(parentBB);
|
||||
RemovePhiIncoming(trueBlock, parentBB);
|
||||
}
|
||||
}
|
||||
|
||||
// 更新前驱块的终结指令(当一个后继块被移除时)
|
||||
void SCCPContext::UpdateTerminator(BasicBlock *predBB, BasicBlock *removedSucc) {
|
||||
Instruction *terminator = predBB->terminator()->get();
|
||||
if (!terminator)
|
||||
return;
|
||||
|
||||
if (terminator->isBranch()) {
|
||||
if (terminator->isCondBr()) { // 如果是条件分支
|
||||
CondBrInst *branchInst = static_cast<CondBrInst *>(terminator);
|
||||
if (branchInst->getThenBlock() == removedSucc) {
|
||||
if (DEBUG) {
|
||||
std::cout << "Updating cond br in " << predBB->getName() << ": True block (" << removedSucc->getName()
|
||||
<< ") removed. Converting to Br to " << branchInst->getElseBlock()->getName() << std::endl;
|
||||
}
|
||||
builder->setPosition(predBB, predBB->findInstIterator(branchInst));
|
||||
builder->createUncondBrInst(branchInst->getElseBlock());
|
||||
SysYIROptUtils::usedelete(branchInst);
|
||||
predBB->removeSuccessor(removedSucc);
|
||||
} else if (branchInst->getElseBlock() == removedSucc) {
|
||||
if (DEBUG) {
|
||||
std::cout << "Updating cond br in " << predBB->getName() << ": False block (" << removedSucc->getName()
|
||||
<< ") removed. Converting to Br to " << branchInst->getThenBlock()->getName() << std::endl;
|
||||
}
|
||||
builder->setPosition(predBB, predBB->findInstIterator(branchInst));
|
||||
builder->createUncondBrInst(branchInst->getThenBlock());
|
||||
SysYIROptUtils::usedelete(branchInst);
|
||||
predBB->removeSuccessor(removedSucc);
|
||||
}
|
||||
} else { // 无条件分支 (kBr)
|
||||
UncondBrInst *branchInst = static_cast<UncondBrInst *>(terminator);
|
||||
if (branchInst->getBlock() == removedSucc) {
|
||||
if (DEBUG) {
|
||||
std::cout << "Updating unconditional br in " << predBB->getName() << ": Target block ("
|
||||
<< removedSucc->getName() << ") removed. Replacing with Unreachable." << std::endl;
|
||||
}
|
||||
SysYIROptUtils::usedelete(branchInst);
|
||||
predBB->removeSuccessor(removedSucc);
|
||||
builder->setPosition(predBB, predBB->end());
|
||||
builder->createUnreachableInst();
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 移除 Phi 节点的入边(当其前驱块被移除时)
|
||||
void SCCPContext::RemovePhiIncoming(BasicBlock *phiParentBB, BasicBlock *removedPred) { // 修正 removedPred 类型
|
||||
std::vector<Instruction *> insts_to_check;
|
||||
for (auto &inst_ptr : phiParentBB->getInstructions()) {
|
||||
insts_to_check.push_back(inst_ptr.get());
|
||||
}
|
||||
|
||||
for (Instruction *inst : insts_to_check) {
|
||||
if (auto phi = dynamic_cast<PhiInst *>(inst)) {
|
||||
phi->removeIncomingBlock(removedPred);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 运行 SCCP 优化
|
||||
void SCCPContext::run(Function *func, AnalysisManager &AM) {
|
||||
bool changed_constant_propagation = PropagateConstants(func);
|
||||
bool changed_control_flow = SimplifyControlFlow(func);
|
||||
|
||||
// 如果任何一个阶段修改了 IR,标记分析结果为失效
|
||||
if (changed_constant_propagation || changed_control_flow) {
|
||||
// AM.invalidate(); // 假设有这样的方法来使所有分析结果失效
|
||||
}
|
||||
}
|
||||
|
||||
// SCCP Pass methods
|
||||
bool SCCP::runOnFunction(Function *F, AnalysisManager &AM) {
|
||||
if (DEBUG) {
|
||||
std::cout << "Running SCCP on function: " << F->getName() << std::endl;
|
||||
}
|
||||
SCCPContext context(builder);
|
||||
context.run(F, AM);
|
||||
return true;
|
||||
}
|
||||
|
||||
void SCCP::getAnalysisUsage(std::set<void *> &analysisDependencies, std::set<void *> &analysisInvalidations) const {
|
||||
// analysisInvalidations.insert(nullptr); // 表示使所有默认分析失效
|
||||
analysisInvalidations.insert(&DominatorTreeAnalysisPass::ID); // 支配树可能受影响
|
||||
analysisInvalidations.insert(&LivenessAnalysisPass::ID); // 活跃性分析很可能失效
|
||||
}
|
||||
|
||||
} // namespace sysy
|
||||
@@ -1,12 +1,12 @@
|
||||
#include "SysYIRCFGOpt.h"
|
||||
#include "SysYIROptUtils.h"
|
||||
#include <cassert>
|
||||
#include <iostream>
|
||||
#include <list>
|
||||
#include <map>
|
||||
#include <memory>
|
||||
#include <queue> // 引入队列,SysYDelNoPreBLock需要
|
||||
#include <string>
|
||||
#include <iostream>
|
||||
#include <queue> // 引入队列,SysYDelNoPreBLock需要
|
||||
|
||||
namespace sysy {
|
||||
|
||||
@@ -18,6 +18,7 @@ void *SysYBlockMergePass::ID = (void *)&SysYBlockMergePass::ID;
|
||||
void *SysYAddReturnPass::ID = (void *)&SysYAddReturnPass::ID;
|
||||
void *SysYCondBr2BrPass::ID = (void *)&SysYCondBr2BrPass::ID;
|
||||
|
||||
|
||||
// ======================================================================
|
||||
// SysYCFGOptUtils: 辅助工具类,包含实际的CFG优化逻辑
|
||||
// ======================================================================
|
||||
@@ -25,42 +26,40 @@ void *SysYCondBr2BrPass::ID = (void *)&SysYCondBr2BrPass::ID;
|
||||
// 删除br后的无用指令
|
||||
bool SysYCFGOptUtils::SysYDelInstAfterBr(Function *func) {
|
||||
bool changed = false;
|
||||
|
||||
|
||||
auto basicBlocks = func->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 ((*iter)->isTerminator()) {
|
||||
if ((*iter)->isTerminator()){
|
||||
Branch = true;
|
||||
Branchiter = iter;
|
||||
break;
|
||||
}
|
||||
}
|
||||
if (Branchiter != instructions.end())
|
||||
++Branchiter;
|
||||
if (Branchiter != instructions.end()) ++Branchiter;
|
||||
while (Branchiter != instructions.end()) {
|
||||
changed = true;
|
||||
Branchiter = instructions.erase(Branchiter);
|
||||
}
|
||||
|
||||
if (Branch) { // 更新前驱后继关系
|
||||
auto thelastinstinst = basicBlock->terminator();
|
||||
|
||||
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()) {
|
||||
auto brinst = dynamic_cast<UncondBrInst *>(thelastinstinst->get());
|
||||
BasicBlock *branchBlock = dynamic_cast<BasicBlock *>(brinst->getBlock());
|
||||
BasicBlock* branchBlock = dynamic_cast<BasicBlock *>(thelastinstinst->get()->getOperand(0));
|
||||
basicBlock->addSuccessor(branchBlock);
|
||||
branchBlock->addPredecessor(basicBlock.get());
|
||||
} else if (thelastinstinst->get()->isConditional()) {
|
||||
auto brinst = dynamic_cast<CondBrInst *>(thelastinstinst->get());
|
||||
BasicBlock *thenBlock = dynamic_cast<BasicBlock *>(brinst->getThenBlock());
|
||||
BasicBlock *elseBlock = dynamic_cast<BasicBlock *>(brinst->getElseBlock());
|
||||
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());
|
||||
@@ -76,27 +75,29 @@ bool SysYCFGOptUtils::SysYDelInstAfterBr(Function *func) {
|
||||
bool SysYCFGOptUtils::SysYBlockMerge(Function *func) {
|
||||
bool changed = false;
|
||||
|
||||
for (auto blockiter = func->getBasicBlocks().begin(); blockiter != func->getBasicBlocks().end();) {
|
||||
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];
|
||||
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->terminator();
|
||||
auto thelastinstinst = block->end();
|
||||
(--thelastinstinst);
|
||||
if (thelastinstinst->get()->isUnconditional()) {
|
||||
thelastinstinst = SysYIROptUtils::usedelete(thelastinstinst);
|
||||
SysYIROptUtils::usedelete(thelastinstinst->get());
|
||||
thelastinstinst = block->getInstructions().erase(thelastinstinst);
|
||||
} else if (thelastinstinst->get()->isConditional()) {
|
||||
// 按道理不会走到这个分支
|
||||
// 如果是条件分支,查看then else是否相同
|
||||
auto brinst = dynamic_cast<CondBrInst *>(thelastinstinst->get());
|
||||
if (brinst->getThenBlock() == brinst->getElseBlock()) {
|
||||
thelastinstinst = SysYIROptUtils::usedelete(thelastinstinst);
|
||||
// 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()) {
|
||||
SysYIROptUtils::usedelete(thelastinstinst->get());
|
||||
thelastinstinst = block->getInstructions().erase(thelastinstinst);
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -105,7 +106,7 @@ bool SysYCFGOptUtils::SysYBlockMerge(Function *func) {
|
||||
for (auto institer = nextBlock->begin(); institer != nextBlock->end();) {
|
||||
institer->get()->setParent(block);
|
||||
block->getInstructions().emplace_back(institer->release());
|
||||
institer = nextBlock->getInstructions().erase(institer);
|
||||
institer = nextBlock->getInstructions().erase(institer);
|
||||
}
|
||||
// 更新前驱后继关系,类似树节点操作
|
||||
block->removeSuccessor(nextBlock);
|
||||
@@ -136,434 +137,323 @@ bool SysYCFGOptUtils::SysYBlockMerge(Function *func) {
|
||||
|
||||
// 删除无前驱块,兼容SSA后的处理
|
||||
bool SysYCFGOptUtils::SysYDelNoPreBLock(Function *func) {
|
||||
bool changed = false; // 标记是否有基本块被删除
|
||||
std::set<BasicBlock *> reachableBlocks; // 用于存储所有可达的基本块
|
||||
std::queue<BasicBlock *> blockQueue; // BFS 遍历队列
|
||||
|
||||
BasicBlock *entryBlock = func->getEntryBlock();
|
||||
if (entryBlock) { // 确保函数有入口块
|
||||
reachableBlocks.insert(entryBlock); // 将入口块标记为可达
|
||||
blockQueue.push(entryBlock); // 入口块入队
|
||||
}
|
||||
// 如果没有入口块(比如一个空函数),则没有块是可达的,所有块都将被删除。
|
||||
|
||||
while (!blockQueue.empty()) { // BFS 遍历:只要队列不空
|
||||
BasicBlock *currentBlock = blockQueue.front();
|
||||
blockQueue.pop(); // 取出当前块
|
||||
|
||||
for (auto &succ : currentBlock->getSuccessors()) { // 遍历当前块的所有后继
|
||||
// 如果后继块不在 reachableBlocks 中(即尚未被访问过)
|
||||
if (reachableBlocks.find(succ) == reachableBlocks.end()) {
|
||||
reachableBlocks.insert(succ); // 标记为可达
|
||||
blockQueue.push(succ); // 入队,以便继续遍历
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
std::vector<BasicBlock *> blocksToDelete; // 用于存储所有不可达的基本块
|
||||
|
||||
for (auto &blockPtr : func->getBasicBlocks()) {
|
||||
BasicBlock *block = blockPtr.get();
|
||||
// 如果当前块不在 reachableBlocks 集合中,说明它是不可达的
|
||||
if (reachableBlocks.find(block) == reachableBlocks.end()) {
|
||||
blocksToDelete.push_back(block); // 将其加入待删除列表
|
||||
changed = true; // 只要找到一个不可达块,就说明函数发生了改变
|
||||
}
|
||||
}
|
||||
|
||||
for (BasicBlock *unreachableBlock : blocksToDelete) {
|
||||
// 遍历不可达块中的所有指令,并删除它们
|
||||
for (auto instIter = unreachableBlock->getInstructions().begin();
|
||||
instIter != unreachableBlock->getInstructions().end();) {
|
||||
instIter = SysYIROptUtils::usedelete(instIter);
|
||||
}
|
||||
}
|
||||
|
||||
for (BasicBlock *unreachableBlock : blocksToDelete) {
|
||||
for (BasicBlock *succBlock : unreachableBlock->getSuccessors()) {
|
||||
// 只有当后继块自身是可达的(没有被删除)时才需要处理
|
||||
if (reachableBlocks.count(succBlock)) {
|
||||
for (auto &phiInstPtr : succBlock->getInstructions()) {
|
||||
// Phi 指令总是在基本块的开头。一旦遇到非 Phi 指令即可停止。
|
||||
if (phiInstPtr->getKind() != Instruction::kPhi) {
|
||||
break;
|
||||
}
|
||||
// 将这个 Phi 节点中来自不可达前驱(unreachableBlock)的输入参数删除
|
||||
dynamic_cast<PhiInst *>(phiInstPtr.get())->removeIncomingBlock(unreachableBlock);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for (auto blockIter = func->getBasicBlocks().begin(); blockIter != func->getBasicBlocks().end();) {
|
||||
BasicBlock *currentBlock = blockIter->get();
|
||||
// 如果当前块不在可达块集合中,则将其从函数中移除
|
||||
if (reachableBlocks.find(currentBlock) == reachableBlocks.end()) {
|
||||
// func->removeBasicBlock 应该返回下一个有效的迭代器
|
||||
func->removeBasicBlock((blockIter++)->get());
|
||||
} else {
|
||||
blockIter++; // 如果可达,则移动到下一个块
|
||||
}
|
||||
}
|
||||
|
||||
return changed;
|
||||
}
|
||||
|
||||
bool SysYCFGOptUtils::SysYDelEmptyBlock(Function *func, IRBuilder *pBuilder) {
|
||||
|
||||
bool changed = false;
|
||||
|
||||
// 步骤 1: 识别并映射所有符合“空块”定义的基本块及其目标后继
|
||||
// 使用 std::map 来存储 <空块, 空块跳转目标>
|
||||
// 这样可以处理空块链:A -> B -> C,如果 B 是空块,A 应该跳到 C
|
||||
std::map<BasicBlock *, BasicBlock *> emptyBlockRedirectMap;
|
||||
|
||||
// 为了避免在遍历 func->getBasicBlocks() 时修改它导致迭代器失效,
|
||||
// 我们先收集所有的基本块。
|
||||
std::vector<BasicBlock *> allBlocks;
|
||||
for (auto &blockPtr : func->getBasicBlocks()) {
|
||||
allBlocks.push_back(blockPtr.get());
|
||||
for (auto &block : func->getBasicBlocks()) {
|
||||
block->setreachableFalse();
|
||||
}
|
||||
|
||||
for (BasicBlock *block : allBlocks) {
|
||||
// 入口块通常不应该被认为是空块并删除,除非它没有实际指令且只有一个后继,
|
||||
// 但为了安全起见,通常会跳过入口块的删除。
|
||||
// 如果入口块是空的,它应该被合并到它的后继,但处理起来更复杂,这里先不处理入口块为空的情况
|
||||
if (block == func->getEntryBlock()) {
|
||||
continue;
|
||||
}
|
||||
|
||||
// 检查基本块是否是空的:除了Phi指令外,只包含一个终止指令 (Terminator)
|
||||
// 且该终止指令必须是无条件跳转。
|
||||
// 空块必须只有一个后继才能被简化
|
||||
if (block->getNumSuccessors() == 1) {
|
||||
bool hasNonPhiNonTerminator = false;
|
||||
// 遍历除了最后一个指令之外的指令
|
||||
for (auto instIter = block->getInstructions().begin(); instIter != block->getInstructions().end();) {
|
||||
// 如果是终止指令(例如 br, ret),且不是最后一个指令,则该块有问题
|
||||
if ((*instIter)->isTerminator() && instIter != block->terminator()) {
|
||||
hasNonPhiNonTerminator = true;
|
||||
break;
|
||||
}
|
||||
// 如果不是 Phi 指令且不是终止指令
|
||||
if (!(*instIter)->isPhi() && !(*instIter)->isTerminator()) {
|
||||
hasNonPhiNonTerminator = true;
|
||||
break;
|
||||
}
|
||||
++instIter;
|
||||
if (!hasNonPhiNonTerminator &&
|
||||
instIter == block->getInstructions().end()) { // 如果块中只有 Phi 指令和一个 Terminator
|
||||
// 确保最后一个指令是无条件跳转
|
||||
auto lastInst = block->terminator()->get();
|
||||
if (lastInst && lastInst->isUnconditional()) {
|
||||
emptyBlockRedirectMap[block] = block->getSuccessors().front();
|
||||
}
|
||||
}
|
||||
// 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);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 步骤 2: 遍历 emptyBlockRedirectMap,处理空块链
|
||||
// 确保每个空块都直接重定向到其最终的非空后继块
|
||||
for (auto const &[emptyBlock, directSucc] : emptyBlockRedirectMap) {
|
||||
BasicBlock *targetBlock = directSucc;
|
||||
// 沿着空块链一直找到最终的非空块目标
|
||||
while (emptyBlockRedirectMap.count(targetBlock)) {
|
||||
targetBlock = emptyBlockRedirectMap[targetBlock];
|
||||
}
|
||||
emptyBlockRedirectMap[emptyBlock] = targetBlock; // 更新映射到最终目标
|
||||
}
|
||||
|
||||
// 步骤 3: 遍历所有基本块,重定向其终止指令,绕过空块
|
||||
// 注意:这里需要再次遍历所有块,包括可能成为新目标的块
|
||||
for (BasicBlock *currentBlock : allBlocks) {
|
||||
// 如果 currentBlock 本身就是个空块,它会通过其前驱的重定向被处理,这里跳过
|
||||
if (emptyBlockRedirectMap.count(currentBlock)) {
|
||||
continue;
|
||||
}
|
||||
|
||||
// 获取当前块的最后一个指令(终止指令)
|
||||
if (currentBlock->getInstructions().empty()) {
|
||||
// 理论上,除了入口块和可能被合并的空块外,所有块都应该有终止指令
|
||||
// 如果这里碰到空块,可能是逻辑错误或者需要特殊处理
|
||||
continue;
|
||||
}
|
||||
|
||||
std::function<Value *(Value *, BasicBlock *)> getUltimateSourceValue = [&](Value *val,
|
||||
BasicBlock *currentDefBlock) -> Value * {
|
||||
// 如果值不是指令,例如常量或函数参数,则它本身就是最终来源
|
||||
if (auto instr = dynamic_cast<Instruction *>(val)) { // Assuming Value* has a method to check if it's an instruction
|
||||
return val;
|
||||
}
|
||||
|
||||
Instruction *inst = dynamic_cast<Instruction *>(val);
|
||||
// 如果定义指令不在任何空块中,它就是最终来源
|
||||
if (!emptyBlockRedirectMap.count(currentDefBlock)) {
|
||||
return val;
|
||||
}
|
||||
|
||||
// 如果是 Phi 指令,且它在空块中,则继续追溯其在空块链中前驱的传入值
|
||||
if (inst->getKind() == Instruction::kPhi) {
|
||||
PhiInst *phi = dynamic_cast<PhiInst *>(inst);
|
||||
// 查找哪个前驱是空块链中的上一个块
|
||||
for (size_t i = 0; i < phi->getNumOperands(); i += 2) {
|
||||
BasicBlock *incomingBlock = dynamic_cast<BasicBlock *>(phi->getOperand(i + 1));
|
||||
// 检查 incomingBlock 是否是当前空块的前驱,且也在空块映射中(或就是 P)
|
||||
// 找到在空块链中导致 currentDefBlock 的那个前驱块
|
||||
if (emptyBlockRedirectMap.count(incomingBlock) || incomingBlock == currentBlock) {
|
||||
// 递归追溯该传入值
|
||||
return getUltimateSourceValue(phi->getValfromBlk(incomingBlock), incomingBlock);
|
||||
}
|
||||
}
|
||||
}
|
||||
// 如果是其他指令或者无法追溯到Phi链,则认为它在空块中产生,无法安全传播,返回null或原值
|
||||
// 在严格的空块定义下,除了Phi和Terminator,不应有其他指令产生值。
|
||||
return val; // Fallback: If not a Phi, or unable to trace, return itself (may be dangling)
|
||||
};
|
||||
|
||||
auto lastInst = currentBlock->getInstructions().back().get();
|
||||
|
||||
if (lastInst->isUnconditional()) { // 无条件跳转
|
||||
UncondBrInst *brInst = dynamic_cast<UncondBrInst *>(lastInst);
|
||||
BasicBlock *oldTarget = dynamic_cast<BasicBlock *>(brInst->getBlock()); // 原始跳转目标
|
||||
|
||||
if (emptyBlockRedirectMap.count(oldTarget)) { // 如果目标是空块
|
||||
BasicBlock *newTarget = emptyBlockRedirectMap[oldTarget]; // 获取最终目标
|
||||
|
||||
// 更新 CFG 关系
|
||||
currentBlock->removeSuccessor(oldTarget);
|
||||
oldTarget->removePredecessor(currentBlock);
|
||||
|
||||
brInst->replaceOperand(0, newTarget); // 更新跳转指令的操作数
|
||||
currentBlock->addSuccessor(newTarget);
|
||||
newTarget->addPredecessor(currentBlock);
|
||||
|
||||
changed = true; // 标记发生改变
|
||||
|
||||
for (auto &phiInstPtr : newTarget->getInstructions()) {
|
||||
if (phiInstPtr->getKind() == Instruction::kPhi) {
|
||||
PhiInst *phiInst = dynamic_cast<PhiInst *>(phiInstPtr.get());
|
||||
BasicBlock *actualEmptyPredecessorOfS = nullptr;
|
||||
for (size_t i = 0; i < phiInst->getNumOperands(); i += 2) {
|
||||
BasicBlock *incomingBlock = dynamic_cast<BasicBlock *>(phiInst->getOperand(i + 1));
|
||||
if (incomingBlock && emptyBlockRedirectMap.count(incomingBlock) &&
|
||||
emptyBlockRedirectMap[incomingBlock] == newTarget) {
|
||||
actualEmptyPredecessorOfS = incomingBlock;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (actualEmptyPredecessorOfS) {
|
||||
// 获取 Phi 节点原本从 actualEmptyPredecessorOfS 接收的值
|
||||
Value *valueFromEmptyPredecessor = phiInst->getValfromBlk(actualEmptyPredecessorOfS);
|
||||
|
||||
// 追溯这个值,找到它在非空块中的最终来源
|
||||
// currentBlock 是 P
|
||||
// oldTarget 是 E1 (链的起点)
|
||||
// actualEmptyPredecessorOfS 是 En (链的终点,S 的前驱)
|
||||
Value *ultimateSourceValue = getUltimateSourceValue(valueFromEmptyPredecessor, actualEmptyPredecessorOfS);
|
||||
|
||||
// 替换 Phi 节点的传入块和传入值
|
||||
if (ultimateSourceValue) { // 确保成功追溯到有效来源
|
||||
// phiInst->replaceIncoming(actualEmptyPredecessorOfS, currentBlock, ultimateSourceValue);
|
||||
phiInst->replaceIncomingBlock(actualEmptyPredecessorOfS, currentBlock, ultimateSourceValue);
|
||||
} else {
|
||||
assert(false && "[DelEmptyBlock] Unable to trace a valid source for Phi instruction");
|
||||
// 无法追溯到有效来源,这可能是个错误或特殊情况
|
||||
// 此时可能需要移除该 Phi 项,或者插入一个 undef 值
|
||||
phiInst->getValfromBlk(actualEmptyPredecessorOfS);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
} else if (lastInst->getKind() == Instruction::kCondBr) { // 条件跳转
|
||||
CondBrInst *condBrInst = dynamic_cast<CondBrInst *>(lastInst);
|
||||
BasicBlock *oldThenTarget = dynamic_cast<BasicBlock *>(condBrInst->getThenBlock());
|
||||
BasicBlock *oldElseTarget = dynamic_cast<BasicBlock *>(condBrInst->getElseBlock());
|
||||
|
||||
bool thenPathChanged = false;
|
||||
bool elsePathChanged = false;
|
||||
|
||||
// 处理 Then 分支
|
||||
if (emptyBlockRedirectMap.count(oldThenTarget)) {
|
||||
BasicBlock *newThenTarget = emptyBlockRedirectMap[oldThenTarget];
|
||||
condBrInst->replaceOperand(1, newThenTarget); // 更新跳转指令操作数
|
||||
|
||||
currentBlock->removeSuccessor(oldThenTarget);
|
||||
oldThenTarget->removePredecessor(currentBlock);
|
||||
currentBlock->addSuccessor(newThenTarget);
|
||||
newThenTarget->addPredecessor(currentBlock);
|
||||
thenPathChanged = true;
|
||||
changed = true;
|
||||
|
||||
// 处理新 Then 目标块中的 Phi 指令
|
||||
// for (auto &phiInstPtr : newThenTarget->getInstructions()) {
|
||||
// if (phiInstPtr->getKind() == Instruction::kPhi) {
|
||||
// dynamic_cast<PhiInst *>(phiInstPtr.get())->delBlk(oldThenTarget);
|
||||
// } else {
|
||||
// break;
|
||||
// }
|
||||
// }
|
||||
for (auto &phiInstPtr : newThenTarget->getInstructions()) {
|
||||
if (phiInstPtr->getKind() == Instruction::kPhi) {
|
||||
PhiInst *phiInst = dynamic_cast<PhiInst *>(phiInstPtr.get());
|
||||
BasicBlock *actualEmptyPredecessorOfS = nullptr;
|
||||
for (size_t i = 0; i < phiInst->getNumOperands(); i += 2) {
|
||||
BasicBlock *incomingBlock = dynamic_cast<BasicBlock *>(phiInst->getOperand(i + 1));
|
||||
if (incomingBlock && emptyBlockRedirectMap.count(incomingBlock) &&
|
||||
emptyBlockRedirectMap[incomingBlock] == newThenTarget) {
|
||||
actualEmptyPredecessorOfS = incomingBlock;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (actualEmptyPredecessorOfS) {
|
||||
// 获取 Phi 节点原本从 actualEmptyPredecessorOfS 接收的值
|
||||
Value *valueFromEmptyPredecessor = phiInst->getValfromBlk(actualEmptyPredecessorOfS);
|
||||
|
||||
// 追溯这个值,找到它在非空块中的最终来源
|
||||
// currentBlock 是 P
|
||||
// oldTarget 是 E1 (链的起点)
|
||||
// actualEmptyPredecessorOfS 是 En (链的终点,S 的前驱)
|
||||
Value *ultimateSourceValue = getUltimateSourceValue(valueFromEmptyPredecessor, actualEmptyPredecessorOfS);
|
||||
|
||||
// 替换 Phi 节点的传入块和传入值
|
||||
if (ultimateSourceValue) { // 确保成功追溯到有效来源
|
||||
// phiInst->replaceIncoming(actualEmptyPredecessorOfS, currentBlock, ultimateSourceValue);
|
||||
phiInst->replaceIncomingBlock(actualEmptyPredecessorOfS, currentBlock, ultimateSourceValue);
|
||||
} else {
|
||||
assert(false && "[DelEmptyBlock] Unable to trace a valid source for Phi instruction");
|
||||
// 无法追溯到有效来源,这可能是个错误或特殊情况
|
||||
// 此时可能需要移除该 Phi 项,或者插入一个 undef 值
|
||||
phiInst->removeIncomingBlock(actualEmptyPredecessorOfS);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
// 处理 Else 分支
|
||||
if (emptyBlockRedirectMap.count(oldElseTarget)) {
|
||||
BasicBlock *newElseTarget = emptyBlockRedirectMap[oldElseTarget];
|
||||
condBrInst->replaceOperand(2, newElseTarget); // 更新跳转指令操作数
|
||||
|
||||
currentBlock->removeSuccessor(oldElseTarget);
|
||||
oldElseTarget->removePredecessor(currentBlock);
|
||||
currentBlock->addSuccessor(newElseTarget);
|
||||
newElseTarget->addPredecessor(currentBlock);
|
||||
elsePathChanged = true;
|
||||
changed = true;
|
||||
|
||||
// 处理新 Else 目标块中的 Phi 指令
|
||||
// for (auto &phiInstPtr : newElseTarget->getInstructions()) {
|
||||
// if (phiInstPtr->getKind() == Instruction::kPhi) {
|
||||
// dynamic_cast<PhiInst *>(phiInstPtr.get())->delBlk(oldElseTarget);
|
||||
// } else {
|
||||
// break;
|
||||
// }
|
||||
// }
|
||||
for (auto &phiInstPtr : newElseTarget->getInstructions()) {
|
||||
if (phiInstPtr->getKind() == Instruction::kPhi) {
|
||||
PhiInst *phiInst = dynamic_cast<PhiInst *>(phiInstPtr.get());
|
||||
BasicBlock *actualEmptyPredecessorOfS = nullptr;
|
||||
for (size_t i = 0; i < phiInst->getNumOperands(); i += 2) {
|
||||
BasicBlock *incomingBlock = dynamic_cast<BasicBlock *>(phiInst->getOperand(i + 1));
|
||||
if (incomingBlock && emptyBlockRedirectMap.count(incomingBlock) &&
|
||||
emptyBlockRedirectMap[incomingBlock] == newElseTarget) {
|
||||
actualEmptyPredecessorOfS = incomingBlock;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (actualEmptyPredecessorOfS) {
|
||||
// 获取 Phi 节点原本从 actualEmptyPredecessorOfS 接收的值
|
||||
Value *valueFromEmptyPredecessor = phiInst->getValfromBlk(actualEmptyPredecessorOfS);
|
||||
|
||||
// 追溯这个值,找到它在非空块中的最终来源
|
||||
// currentBlock 是 P
|
||||
// oldTarget 是 E1 (链的起点)
|
||||
// actualEmptyPredecessorOfS 是 En (链的终点,S 的前驱)
|
||||
Value *ultimateSourceValue = getUltimateSourceValue(valueFromEmptyPredecessor, actualEmptyPredecessorOfS);
|
||||
|
||||
// 替换 Phi 节点的传入块和传入值
|
||||
if (ultimateSourceValue) { // 确保成功追溯到有效来源
|
||||
// phiInst->replaceIncoming(actualEmptyPredecessorOfS, currentBlock, ultimateSourceValue);
|
||||
phiInst->replaceIncomingBlock(actualEmptyPredecessorOfS, currentBlock, ultimateSourceValue);
|
||||
} else {
|
||||
assert(false && "[DelEmptyBlock] Unable to trace a valid source for Phi instruction");
|
||||
// 无法追溯到有效来源,这可能是个错误或特殊情况
|
||||
// 此时可能需要移除该 Phi 项,或者插入一个 undef 值
|
||||
phiInst->removeIncomingBlock(actualEmptyPredecessorOfS);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 额外处理:如果条件跳转的两个分支现在指向同一个块,则可以简化为无条件跳转
|
||||
if (condBrInst->getThenBlock() == condBrInst->getElseBlock()) {
|
||||
BasicBlock *commonTarget = dynamic_cast<BasicBlock *>(condBrInst->getThenBlock());
|
||||
SysYIROptUtils::usedelete(lastInst); // 删除旧的条件跳转指令
|
||||
pBuilder->setPosition(currentBlock, currentBlock->end());
|
||||
pBuilder->createUncondBrInst(commonTarget); // 插入新的无条件跳转指令
|
||||
|
||||
// 更安全地更新 CFG 关系
|
||||
std::set<BasicBlock *> currentSuccessors;
|
||||
currentSuccessors.insert(oldThenTarget);
|
||||
currentSuccessors.insert(oldElseTarget);
|
||||
|
||||
// 移除旧的后继关系
|
||||
for (BasicBlock *succ : currentSuccessors) {
|
||||
currentBlock->removeSuccessor(succ);
|
||||
succ->removePredecessor(currentBlock);
|
||||
}
|
||||
// 添加新的后继关系
|
||||
currentBlock->addSuccessor(commonTarget);
|
||||
commonTarget->addPredecessor(currentBlock);
|
||||
|
||||
changed = true;
|
||||
// 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 instIter = blockIter->get()->getInstructions().begin();
|
||||
instIter != blockIter->get()->getInstructions().end();) {
|
||||
SysYIROptUtils::usedelete(instIter->get());
|
||||
instIter = blockIter->get()->getInstructions().erase(instIter);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 步骤 4: 真正地删除空基本块
|
||||
// 注意:只能在所有跳转和 Phi 指令都更新完毕后才能删除这些块
|
||||
|
||||
for (auto blockIter = func->getBasicBlocks().begin(); blockIter != func->getBasicBlocks().end();) {
|
||||
BasicBlock *currentBlock = blockIter->get();
|
||||
if (emptyBlockRedirectMap.count(currentBlock)) { // 如果在空块映射中
|
||||
// 入口块不应该被删除,即使它符合空块定义,因为函数需要一个入口
|
||||
if (currentBlock == func->getEntryBlock()) {
|
||||
++blockIter;
|
||||
continue;
|
||||
if (!blockIter->get()->getreachable()) {
|
||||
for (auto succblock : blockIter->get()->getSuccessors()) {
|
||||
for (auto &phiinst : succblock->getInstructions()) {
|
||||
if (phiinst->getKind() != Instruction::kPhi) {
|
||||
break;
|
||||
}
|
||||
// 使用 delBlk 方法正确地删除对应于被删除基本块的传入值
|
||||
dynamic_cast<PhiInst *>(phiinst.get())->delBlk(blockIter->get());
|
||||
}
|
||||
}
|
||||
|
||||
// 在删除块之前,确保其内部指令被正确删除(虽然这类块指令很少)
|
||||
for (auto instIter = currentBlock->getInstructions().begin();
|
||||
instIter != currentBlock->getInstructions().end();) {
|
||||
instIter = SysYIROptUtils::usedelete(instIter);
|
||||
}
|
||||
|
||||
// 移除块
|
||||
// 删除不可达基本块,注意迭代器不可达问题
|
||||
func->removeBasicBlock((blockIter++)->get());
|
||||
changed = true;
|
||||
} else {
|
||||
++blockIter;
|
||||
blockIter++;
|
||||
}
|
||||
}
|
||||
|
||||
return changed;
|
||||
}
|
||||
|
||||
// 删除空块
|
||||
bool SysYCFGOptUtils::SysYDelEmptyBlock(Function *func, IRBuilder* pBuilder) {
|
||||
bool changed = false;
|
||||
|
||||
// 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 = func->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 && basicBlock->getNumSuccessors() == 1) // 确保有后继且只有一个
|
||||
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 (这段逻辑在优化遍中可能需要调整,这里是原样保留)
|
||||
// 通常,DelEmptyBlock 应该在BlockMerge之后运行,如果存在完全空块,它会尝试填充一个Br指令。
|
||||
// 但是,它主要目的是重定向跳转。
|
||||
if (distance(basicBlock->begin(), basicBlock->end()) == 0) {
|
||||
if (basicBlock->getNumSuccessors() == 0) {
|
||||
continue;
|
||||
}
|
||||
if (basicBlock->getNumSuccessors() > 1) {
|
||||
// 如果一个空块有多个后继,说明CFG结构有问题或者需要特殊处理,这里简单assert
|
||||
assert(false && "Empty block with multiple successors found during SysYDelEmptyBlock");
|
||||
}
|
||||
// 这里的逻辑有点问题,如果一个块是空的,且只有一个后继,应该直接跳转到后继。
|
||||
// 如果这个块最终被删除了,那么其前驱也需要重定向。
|
||||
// 这个循环的目的是重定向现有的跳转指令,而不是创建新的。
|
||||
// 所以下面的逻辑才是核心。
|
||||
// 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.count(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0)))) {
|
||||
thelastBlockOld = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0));
|
||||
thelastinst->get()->replaceOperand(0, EmptyBlocks[thelastBlockOld]);
|
||||
}
|
||||
|
||||
// 如果有重定向发生
|
||||
if (thelastBlockOld != nullptr) {
|
||||
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());
|
||||
changed = true; // 标记IR被修改
|
||||
}
|
||||
|
||||
|
||||
if (thelastBlockOld != nullptr) {
|
||||
for (auto &InstInNew : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(0))->getInstructions()) {
|
||||
if (InstInNew->isPhi()) {
|
||||
// 使用 delBlk 方法删除 oldBlock 对应的传入值
|
||||
dynamic_cast<PhiInst *>(InstInNew.get())->delBlk(thelastBlockOld);
|
||||
} 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));
|
||||
bool thenChanged = false;
|
||||
bool elseChanged = false;
|
||||
|
||||
|
||||
BasicBlock *thelastBlockOld = nullptr;
|
||||
while (EmptyBlocks.count(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1)))) {
|
||||
thelastBlockOld = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1));
|
||||
thelastinst->get()->replaceOperand(
|
||||
1, EmptyBlocks[dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1))]);
|
||||
thenChanged = true;
|
||||
}
|
||||
|
||||
if (thenChanged) {
|
||||
basicBlock->removeSuccessor(OldThenBlock);
|
||||
OldThenBlock->removePredecessor(basicBlock.get());
|
||||
basicBlock->addSuccessor(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1)));
|
||||
dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1))->addPredecessor(basicBlock.get());
|
||||
changed = true; // 标记IR被修改
|
||||
}
|
||||
|
||||
// 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));
|
||||
SysYIROptUtils::usedelete(thelastinst->get());
|
||||
thelastinst = basicBlock->getInstructions().erase(thelastinst);
|
||||
pBuilder->setPosition(basicBlock.get(), basicBlock->end());
|
||||
pBuilder->createUncondBrInst(thebrBlock, {});
|
||||
changed = true; // 标记IR被修改
|
||||
continue;
|
||||
}
|
||||
|
||||
if (thelastBlockOld != nullptr) {
|
||||
for (auto &InstInNew : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(1))->getInstructions()) {
|
||||
if (InstInNew->isPhi()) {
|
||||
// 使用 delBlk 方法删除 oldBlock 对应的传入值
|
||||
dynamic_cast<PhiInst *>(InstInNew.get())->delBlk(thelastBlockOld);
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
thelastBlockOld = nullptr;
|
||||
while (EmptyBlocks.count(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2)))) {
|
||||
thelastBlockOld = dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2));
|
||||
thelastinst->get()->replaceOperand(
|
||||
2, EmptyBlocks[dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2))]);
|
||||
elseChanged = true;
|
||||
}
|
||||
|
||||
if (elseChanged) {
|
||||
basicBlock->removeSuccessor(OldElseBlock);
|
||||
OldElseBlock->removePredecessor(basicBlock.get());
|
||||
basicBlock->addSuccessor(dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2)));
|
||||
dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2))->addPredecessor(basicBlock.get());
|
||||
changed = true; // 标记IR被修改
|
||||
}
|
||||
|
||||
// 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));
|
||||
SysYIROptUtils::usedelete(thelastinst->get());
|
||||
thelastinst = basicBlock->getInstructions().erase(thelastinst);
|
||||
pBuilder->setPosition(basicBlock.get(), basicBlock->end());
|
||||
pBuilder->createUncondBrInst(thebrBlock, {});
|
||||
changed = true; // 标记IR被修改
|
||||
continue;
|
||||
}
|
||||
|
||||
|
||||
// 如果有重定向发生
|
||||
// 需要更新后继块的前驱关系
|
||||
if (thelastBlockOld != nullptr) {
|
||||
for (auto &InstInNew : dynamic_cast<BasicBlock *>(thelastinst->get()->getOperand(2))->getInstructions()) {
|
||||
if (InstInNew->isPhi()) {
|
||||
// 使用 delBlk 方法删除 oldBlock 对应的传入值
|
||||
dynamic_cast<PhiInst *>(InstInNew.get())->delBlk(thelastBlockOld);
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
} else {
|
||||
// 如果不是终止指令,但有后继 (例如,末尾没有显式终止指令的块)
|
||||
// 这段逻辑可能需要更严谨的CFG检查来确保正确性
|
||||
if (basicBlock->getNumSuccessors() == 1) {
|
||||
// 这里的逻辑似乎是想为没有terminator的块添加一个,但通常这应该在CFG构建阶段完成。
|
||||
// 如果这里仍然执行,确保它符合预期。
|
||||
// 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());
|
||||
// changed = true; // 标记IR被修改
|
||||
// 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;
|
||||
// }
|
||||
// }
|
||||
// }
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 真正的删除空块
|
||||
for (auto iter = func->getBasicBlocks().begin(); iter != func->getBasicBlocks().end();) {
|
||||
|
||||
if (EmptyBlocks.count(iter->get())) {
|
||||
// EntryBlock跳过
|
||||
if (iter->get() == func->getEntryBlock()) {
|
||||
++iter;
|
||||
continue;
|
||||
}
|
||||
|
||||
for (auto instIter = iter->get()->getInstructions().begin();
|
||||
instIter != iter->get()->getInstructions().end();) {
|
||||
SysYIROptUtils::usedelete(instIter->get()); // 仅删除 use 关系
|
||||
// 显式地从基本块中删除指令并更新迭代器
|
||||
instIter = iter->get()->getInstructions().erase(instIter);
|
||||
}
|
||||
// 删除不可达基本块的phi指令的操作数
|
||||
for (auto &succ : iter->get()->getSuccessors()) {
|
||||
for (auto &instinsucc : succ->getInstructions()) {
|
||||
if (instinsucc->isPhi()) {
|
||||
// iter->get() 就是当前被删除的空基本块,它作为前驱连接到这里的Phi指令
|
||||
dynamic_cast<PhiInst *>(instinsucc.get())->delBlk(iter->get());
|
||||
} else {
|
||||
// Phi 指令通常在基本块的开头,如果不是 Phi 指令就停止检查
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
func->removeBasicBlock((iter++)->get());
|
||||
changed = true;
|
||||
} else {
|
||||
++iter;
|
||||
}
|
||||
}
|
||||
|
||||
return changed;
|
||||
}
|
||||
|
||||
// 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)
|
||||
bool SysYCFGOptUtils::SysYAddReturn(Function *func, IRBuilder *pBuilder) {
|
||||
bool SysYCFGOptUtils::SysYAddReturn(Function *func, IRBuilder* pBuilder) {
|
||||
bool changed = false;
|
||||
auto basicBlocks = func->getBasicBlocks();
|
||||
for (auto &block : basicBlocks) {
|
||||
@@ -577,8 +467,7 @@ bool SysYCFGOptUtils::SysYAddReturn(Function *func, IRBuilder *pBuilder) {
|
||||
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;
|
||||
// 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
|
||||
@@ -594,7 +483,7 @@ bool SysYCFGOptUtils::SysYAddReturn(Function *func, IRBuilder *pBuilder) {
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
return changed;
|
||||
}
|
||||
|
||||
@@ -602,18 +491,18 @@ bool SysYCFGOptUtils::SysYAddReturn(Function *func, IRBuilder *pBuilder) {
|
||||
// 主要针对已知条件值的分支转换为无条件分支
|
||||
// 例如 if (cond) { ... } else { ... } 中的 cond 已经
|
||||
// 确定为 true 或 false 的情况
|
||||
bool SysYCFGOptUtils::SysYCondBr2Br(Function *func, IRBuilder *pBuilder) {
|
||||
bool SysYCFGOptUtils::SysYCondBr2Br(Function *func, IRBuilder* pBuilder) {
|
||||
bool changed = false;
|
||||
|
||||
for (auto &basicblock : func->getBasicBlocks()) {
|
||||
if (basicblock->getNumInstructions() == 0)
|
||||
continue;
|
||||
|
||||
auto thelast = basicblock->getInstructions().end();
|
||||
--thelast;
|
||||
|
||||
auto thelast = basicblock->terminator();
|
||||
|
||||
if (thelast->get()->isConditional()) {
|
||||
auto condBrInst = dynamic_cast<CondBrInst *>(thelast->get());
|
||||
ConstantValue *constOperand = dynamic_cast<ConstantValue *>(condBrInst->getCondition());
|
||||
if (thelast->get()->isConditional()){
|
||||
ConstantValue *constOperand = dynamic_cast<ConstantValue *>(thelast->get()->getOperand(0));
|
||||
std::string opname;
|
||||
int constint = 0;
|
||||
float constfloat = 0.0F;
|
||||
@@ -632,31 +521,32 @@ bool SysYCFGOptUtils::SysYCondBr2Br(Function *func, IRBuilder *pBuilder) {
|
||||
if (constfloat_Use || constint_Use) {
|
||||
changed = true;
|
||||
|
||||
auto thenBlock = dynamic_cast<BasicBlock *>(condBrInst->getThenBlock());
|
||||
auto elseBlock = dynamic_cast<BasicBlock *>(condBrInst->getElseBlock());
|
||||
thelast = SysYIROptUtils::usedelete(thelast);
|
||||
auto thenBlock = dynamic_cast<BasicBlock *>(thelast->get()->getOperand(1));
|
||||
auto elseBlock = dynamic_cast<BasicBlock *>(thelast->get()->getOperand(2));
|
||||
SysYIROptUtils::usedelete(thelast->get());
|
||||
thelast = basicblock->getInstructions().erase(thelast);
|
||||
if ((constfloat_Use && constfloat == 1.0F) || (constint_Use && constint == 1)) {
|
||||
// cond为true或非0
|
||||
pBuilder->setPosition(basicblock.get(), basicblock->end());
|
||||
pBuilder->createUncondBrInst(thenBlock);
|
||||
|
||||
pBuilder->createUncondBrInst(thenBlock, {});
|
||||
|
||||
// 更新CFG关系
|
||||
basicblock->removeSuccessor(elseBlock);
|
||||
elseBlock->removePredecessor(basicblock.get());
|
||||
|
||||
|
||||
// 删除elseBlock的phi指令中对应的basicblock.get()的传入值
|
||||
for (auto &phiinst : elseBlock->getInstructions()) {
|
||||
if (phiinst->getKind() != Instruction::kPhi) {
|
||||
break;
|
||||
}
|
||||
// 使用 delBlk 方法删除 basicblock.get() 对应的传入值
|
||||
dynamic_cast<PhiInst *>(phiinst.get())->removeIncomingBlock(basicblock.get());
|
||||
dynamic_cast<PhiInst *>(phiinst.get())->delBlk(basicblock.get());
|
||||
}
|
||||
|
||||
|
||||
} else { // cond为false或0
|
||||
|
||||
pBuilder->setPosition(basicblock.get(), basicblock->end());
|
||||
pBuilder->createUncondBrInst(elseBlock);
|
||||
pBuilder->createUncondBrInst(elseBlock, {});
|
||||
|
||||
// 更新CFG关系
|
||||
basicblock->removeSuccessor(thenBlock);
|
||||
@@ -668,8 +558,9 @@ bool SysYCFGOptUtils::SysYCondBr2Br(Function *func, IRBuilder *pBuilder) {
|
||||
break;
|
||||
}
|
||||
// 使用 delBlk 方法删除 basicblock.get() 对应的传入值
|
||||
dynamic_cast<PhiInst *>(phiinst.get())->removeIncomingBlock(basicblock.get());
|
||||
dynamic_cast<PhiInst *>(phiinst.get())->delBlk(basicblock.get());
|
||||
}
|
||||
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -682,28 +573,28 @@ bool SysYCFGOptUtils::SysYCondBr2Br(Function *func, IRBuilder *pBuilder) {
|
||||
// 独立的CFG优化遍的实现
|
||||
// ======================================================================
|
||||
|
||||
bool SysYDelInstAfterBrPass::runOnFunction(Function *F, AnalysisManager &AM) {
|
||||
bool SysYDelInstAfterBrPass::runOnFunction(Function *F, AnalysisManager& AM) {
|
||||
return SysYCFGOptUtils::SysYDelInstAfterBr(F);
|
||||
}
|
||||
|
||||
bool SysYDelEmptyBlockPass::runOnFunction(Function *F, AnalysisManager &AM) {
|
||||
bool SysYDelEmptyBlockPass::runOnFunction(Function *F, AnalysisManager& AM) {
|
||||
return SysYCFGOptUtils::SysYDelEmptyBlock(F, pBuilder);
|
||||
}
|
||||
|
||||
bool SysYDelNoPreBLockPass::runOnFunction(Function *F, AnalysisManager &AM) {
|
||||
bool SysYDelNoPreBLockPass::runOnFunction(Function *F, AnalysisManager& AM) {
|
||||
return SysYCFGOptUtils::SysYDelNoPreBLock(F);
|
||||
}
|
||||
|
||||
bool SysYBlockMergePass::runOnFunction(Function *F, AnalysisManager &AM) {
|
||||
return SysYCFGOptUtils::SysYBlockMerge(F);
|
||||
bool SysYBlockMergePass::runOnFunction(Function *F, AnalysisManager& AM) {
|
||||
return SysYCFGOptUtils::SysYBlockMerge(F);
|
||||
}
|
||||
|
||||
bool SysYAddReturnPass::runOnFunction(Function *F, AnalysisManager &AM) {
|
||||
bool SysYAddReturnPass::runOnFunction(Function *F, AnalysisManager& AM) {
|
||||
return SysYCFGOptUtils::SysYAddReturn(F, pBuilder);
|
||||
}
|
||||
|
||||
bool SysYCondBr2BrPass::runOnFunction(Function *F, AnalysisManager &AM) {
|
||||
bool SysYCondBr2BrPass::runOnFunction(Function *F, AnalysisManager& AM) {
|
||||
return SysYCFGOptUtils::SysYCondBr2Br(F, pBuilder);
|
||||
}
|
||||
|
||||
} // namespace sysy
|
||||
} // namespace sysy
|
||||
@@ -5,9 +5,7 @@
|
||||
#include "DCE.h"
|
||||
#include "Mem2Reg.h"
|
||||
#include "Reg2Mem.h"
|
||||
#include "SCCP.h"
|
||||
#include "BuildCFG.h"
|
||||
#include "LargeArrayToGlobal.h"
|
||||
#include "ConstPropagation.h"
|
||||
#include "Pass.h"
|
||||
#include <iostream>
|
||||
#include <queue>
|
||||
@@ -37,13 +35,10 @@ void PassManager::runOptimizationPipeline(Module* moduleIR, IRBuilder* builderIR
|
||||
3. 添加优化passid
|
||||
*/
|
||||
// 注册分析遍
|
||||
registerAnalysisPass<DominatorTreeAnalysisPass>();
|
||||
registerAnalysisPass<LivenessAnalysisPass>();
|
||||
registerAnalysisPass<sysy::DominatorTreeAnalysisPass>();
|
||||
registerAnalysisPass<sysy::LivenessAnalysisPass>();
|
||||
|
||||
// 注册优化遍
|
||||
registerOptimizationPass<BuildCFG>();
|
||||
registerOptimizationPass<LargeArrayToGlobalPass>();
|
||||
|
||||
registerOptimizationPass<SysYDelInstAfterBrPass>();
|
||||
registerOptimizationPass<SysYDelNoPreBLockPass>();
|
||||
registerOptimizationPass<SysYBlockMergePass>();
|
||||
@@ -56,23 +51,11 @@ void PassManager::runOptimizationPipeline(Module* moduleIR, IRBuilder* builderIR
|
||||
registerOptimizationPass<Mem2Reg>(builderIR);
|
||||
registerOptimizationPass<Reg2Mem>(builderIR);
|
||||
|
||||
registerOptimizationPass<SCCP>(builderIR);
|
||||
|
||||
if (optLevel >= 1) {
|
||||
//经过设计安排优化遍的执行顺序以及执行逻辑
|
||||
if (DEBUG) std::cout << "Applying -O1 optimizations.\n";
|
||||
if (DEBUG) std::cout << "--- Running custom optimization sequence ---\n";
|
||||
|
||||
if(DEBUG) {
|
||||
std::cout << "=== IR Before CFGOpt Optimizations ===\n";
|
||||
printPasses();
|
||||
}
|
||||
|
||||
this->clearPasses();
|
||||
this->addPass(&BuildCFG::ID);
|
||||
this->addPass(&LargeArrayToGlobalPass::ID);
|
||||
this->run();
|
||||
|
||||
this->clearPasses();
|
||||
this->addPass(&SysYDelInstAfterBrPass::ID);
|
||||
this->addPass(&SysYDelNoPreBLockPass::ID);
|
||||
@@ -82,10 +65,6 @@ void PassManager::runOptimizationPipeline(Module* moduleIR, IRBuilder* builderIR
|
||||
this->addPass(&SysYAddReturnPass::ID);
|
||||
this->run();
|
||||
|
||||
this->clearPasses();
|
||||
this->addPass(&BuildCFG::ID);
|
||||
this->run();
|
||||
|
||||
if(DEBUG) {
|
||||
std::cout << "=== IR After CFGOpt Optimizations ===\n";
|
||||
printPasses();
|
||||
@@ -102,6 +81,7 @@ void PassManager::runOptimizationPipeline(Module* moduleIR, IRBuilder* builderIR
|
||||
|
||||
this->clearPasses();
|
||||
this->addPass(&Mem2Reg::ID);
|
||||
this->addPass(&ConstPropagation::ID);
|
||||
this->run();
|
||||
|
||||
if(DEBUG) {
|
||||
@@ -109,15 +89,6 @@ void PassManager::runOptimizationPipeline(Module* moduleIR, IRBuilder* builderIR
|
||||
printPasses();
|
||||
}
|
||||
|
||||
this->clearPasses();
|
||||
this->addPass(&SCCP::ID);
|
||||
this->run();
|
||||
|
||||
if(DEBUG) {
|
||||
std::cout << "=== IR After SCCP Optimizations ===\n";
|
||||
printPasses();
|
||||
}
|
||||
|
||||
this->clearPasses();
|
||||
this->addPass(&Reg2Mem::ID);
|
||||
this->run();
|
||||
@@ -126,9 +97,7 @@ void PassManager::runOptimizationPipeline(Module* moduleIR, IRBuilder* builderIR
|
||||
std::cout << "=== IR After Reg2Mem Optimizations ===\n";
|
||||
printPasses();
|
||||
}
|
||||
this->clearPasses();
|
||||
this->addPass(&BuildCFG::ID);
|
||||
this->run();
|
||||
|
||||
if (DEBUG) std::cout << "--- Custom optimization sequence finished ---\n";
|
||||
}
|
||||
|
||||
@@ -143,7 +112,6 @@ void PassManager::runOptimizationPipeline(Module* moduleIR, IRBuilder* builderIR
|
||||
SysYPrinter printer(moduleIR);
|
||||
printer.printIR();
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
void PassManager::clearPasses() {
|
||||
|
||||
@@ -15,139 +15,6 @@
|
||||
using namespace std;
|
||||
namespace sysy {
|
||||
|
||||
std::pair<long long, int> calculate_signed_magic(int d) {
|
||||
if (d == 0) throw std::runtime_error("Division by zero");
|
||||
if (d == 1 || d == -1) return {0, 0}; // Not used by strength reduction
|
||||
|
||||
int k = 0;
|
||||
unsigned int ad = (d > 0) ? d : -d;
|
||||
unsigned int temp = ad;
|
||||
while (temp > 0) {
|
||||
temp >>= 1;
|
||||
k++;
|
||||
}
|
||||
if ((ad & (ad - 1)) == 0) { // if power of 2
|
||||
k--;
|
||||
}
|
||||
|
||||
unsigned __int128 m_val = 1;
|
||||
m_val <<= (32 + k - 1);
|
||||
unsigned __int128 m_prime = m_val / ad;
|
||||
long long m = m_prime + 1;
|
||||
|
||||
return {m, k};
|
||||
}
|
||||
|
||||
// 清除因函数调用而失效的表达式缓存(保守策略)
|
||||
void SysYIRGenerator::invalidateExpressionsOnCall() {
|
||||
availableBinaryExpressions.clear();
|
||||
availableUnaryExpressions.clear();
|
||||
availableLoads.clear();
|
||||
availableGEPs.clear();
|
||||
}
|
||||
|
||||
// 在进入新的基本块时清空所有表达式缓存
|
||||
void SysYIRGenerator::enterNewBasicBlock() {
|
||||
availableBinaryExpressions.clear();
|
||||
availableUnaryExpressions.clear();
|
||||
availableLoads.clear();
|
||||
availableGEPs.clear();
|
||||
}
|
||||
|
||||
// 清除因变量赋值而失效的表达式缓存
|
||||
// @param storedAddress: store 指令的目标地址 (例如 AllocaInst* 或 GEPInst*)
|
||||
void SysYIRGenerator::invalidateExpressionsOnStore(Value *storedAddress) {
|
||||
// 遍历二元表达式缓存,移除受影响的条目
|
||||
// 创建一个临时列表来存储要移除的键,避免在迭代时修改容器
|
||||
std::vector<ExpKey> binaryKeysToRemove;
|
||||
for (const auto &pair : availableBinaryExpressions) {
|
||||
// 检查左操作数
|
||||
// 如果左操作数是 LoadInst,并且它从 storedAddress 加载
|
||||
if (auto loadInst = dynamic_cast<LoadInst *>(pair.first.left)) {
|
||||
if (loadInst->getPointer() == storedAddress) {
|
||||
binaryKeysToRemove.push_back(pair.first);
|
||||
continue; // 这个表达式已标记为移除,跳到下一个
|
||||
}
|
||||
}
|
||||
// 如果左操作数本身就是被存储的地址 (例如,将一个地址值直接作为操作数,虽然不常见)
|
||||
if (pair.first.left == storedAddress) {
|
||||
binaryKeysToRemove.push_back(pair.first);
|
||||
continue;
|
||||
}
|
||||
|
||||
// 检查右操作数,逻辑同左操作数
|
||||
if (auto loadInst = dynamic_cast<LoadInst *>(pair.first.right)) {
|
||||
if (loadInst->getPointer() == storedAddress) {
|
||||
binaryKeysToRemove.push_back(pair.first);
|
||||
continue;
|
||||
}
|
||||
}
|
||||
if (pair.first.right == storedAddress) {
|
||||
binaryKeysToRemove.push_back(pair.first);
|
||||
continue;
|
||||
}
|
||||
}
|
||||
// 实际移除条目
|
||||
for (const auto &key : binaryKeysToRemove) {
|
||||
availableBinaryExpressions.erase(key);
|
||||
}
|
||||
|
||||
// 遍历一元表达式缓存,移除受影响的条目
|
||||
std::vector<UnExpKey> unaryKeysToRemove;
|
||||
for (const auto &pair : availableUnaryExpressions) {
|
||||
// 检查操作数
|
||||
if (auto loadInst = dynamic_cast<LoadInst *>(pair.first.operand)) {
|
||||
if (loadInst->getPointer() == storedAddress) {
|
||||
unaryKeysToRemove.push_back(pair.first);
|
||||
continue;
|
||||
}
|
||||
}
|
||||
if (pair.first.operand == storedAddress) {
|
||||
unaryKeysToRemove.push_back(pair.first);
|
||||
continue;
|
||||
}
|
||||
}
|
||||
// 实际移除条目
|
||||
for (const auto &key : unaryKeysToRemove) {
|
||||
availableUnaryExpressions.erase(key);
|
||||
}
|
||||
availableLoads.erase(storedAddress);
|
||||
|
||||
std::vector<GEPKey> gepKeysToRemove;
|
||||
for (const auto &pair : availableGEPs) {
|
||||
// 检查 GEP 的基指针是否受存储影响
|
||||
if (auto loadInst = dynamic_cast<LoadInst *>(pair.first.basePointer)) {
|
||||
if (loadInst->getPointer() == storedAddress) {
|
||||
gepKeysToRemove.push_back(pair.first);
|
||||
continue; // 标记此GEP为移除,跳过后续检查
|
||||
}
|
||||
}
|
||||
// 如果基指针本身就是存储的目标地址 (不常见,但可能)
|
||||
if (pair.first.basePointer == storedAddress) {
|
||||
gepKeysToRemove.push_back(pair.first);
|
||||
continue;
|
||||
}
|
||||
|
||||
// 检查 GEP 的每个索引是否受存储影响
|
||||
for (const auto &indexVal : pair.first.indices) {
|
||||
if (auto loadInst = dynamic_cast<LoadInst *>(indexVal)) {
|
||||
if (loadInst->getPointer() == storedAddress) {
|
||||
gepKeysToRemove.push_back(pair.first);
|
||||
break; // 标记此GEP为移除,并跳出内部循环
|
||||
}
|
||||
}
|
||||
// 如果索引本身就是存储的目标地址
|
||||
if (indexVal == storedAddress) {
|
||||
gepKeysToRemove.push_back(pair.first);
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
// 实际移除条目
|
||||
for (const auto &key : gepKeysToRemove) {
|
||||
availableGEPs.erase(key);
|
||||
}
|
||||
}
|
||||
|
||||
// std::vector<Value*> BinaryValueStack; ///< 用于存储value的栈
|
||||
// std::vector<int> BinaryOpStack; ///< 用于存储二元表达式的操作符栈
|
||||
@@ -377,56 +244,27 @@ void SysYIRGenerator::compute() {
|
||||
}
|
||||
} else {
|
||||
// 否则,创建相应的IR指令
|
||||
ExpKey currentExpKey(static_cast<BinaryOp>(op), lhs, rhs);
|
||||
auto it = availableBinaryExpressions.find(currentExpKey);
|
||||
|
||||
if (it != availableBinaryExpressions.end()) {
|
||||
// 在缓存中找到,重用结果
|
||||
resultValue = it->second;
|
||||
} else {
|
||||
if (commonType == Type::getIntType()) {
|
||||
switch (op) {
|
||||
case BinaryOp::ADD: resultValue = builder.createAddInst(lhs, rhs); break;
|
||||
case BinaryOp::SUB: resultValue = builder.createSubInst(lhs, rhs); break;
|
||||
case BinaryOp::MUL: resultValue = builder.createMulInst(lhs, rhs); break;
|
||||
case BinaryOp::DIV: {
|
||||
ConstantInteger *rhsConst = dynamic_cast<ConstantInteger *>(rhs);
|
||||
if (rhsConst) {
|
||||
int divisor = rhsConst->getInt();
|
||||
if (divisor > 0 && (divisor & (divisor - 1)) == 0) {
|
||||
int shift = 0;
|
||||
int temp = divisor;
|
||||
while (temp > 1) {
|
||||
temp >>= 1;
|
||||
shift++;
|
||||
}
|
||||
resultValue = builder.createSRAInst(lhs, ConstantInteger::get(shift));
|
||||
} else {
|
||||
resultValue = builder.createDivInst(lhs, rhs);
|
||||
}
|
||||
} else {
|
||||
resultValue = builder.createDivInst(lhs, rhs);
|
||||
}
|
||||
break;
|
||||
}
|
||||
case BinaryOp::MOD: resultValue = builder.createRemInst(lhs, rhs); break;
|
||||
}
|
||||
} else if (commonType == Type::getFloatType()) {
|
||||
switch (op) {
|
||||
case BinaryOp::ADD: resultValue = builder.createFAddInst(lhs, rhs); break;
|
||||
case BinaryOp::SUB: resultValue = builder.createFSubInst(lhs, rhs); break;
|
||||
case BinaryOp::MUL: resultValue = builder.createFMulInst(lhs, rhs); break;
|
||||
case BinaryOp::DIV: resultValue = builder.createFDivInst(lhs, rhs); break;
|
||||
case BinaryOp::MOD:
|
||||
std::cerr << "Error: Modulo operator not supported for float types." << std::endl;
|
||||
return;
|
||||
}
|
||||
} else {
|
||||
std::cerr << "Error: Unsupported type for binary instruction." << std::endl;
|
||||
if (commonType == Type::getIntType()) {
|
||||
switch (op) {
|
||||
case BinaryOp::ADD: resultValue = builder.createAddInst(lhs, rhs); break;
|
||||
case BinaryOp::SUB: resultValue = builder.createSubInst(lhs, rhs); break;
|
||||
case BinaryOp::MUL: resultValue = builder.createMulInst(lhs, rhs); break;
|
||||
case BinaryOp::DIV: resultValue = builder.createDivInst(lhs, rhs); break;
|
||||
case BinaryOp::MOD: resultValue = builder.createRemInst(lhs, rhs); break;
|
||||
}
|
||||
} else if (commonType == Type::getFloatType()) {
|
||||
switch (op) {
|
||||
case BinaryOp::ADD: resultValue = builder.createFAddInst(lhs, rhs); break;
|
||||
case BinaryOp::SUB: resultValue = builder.createFSubInst(lhs, rhs); break;
|
||||
case BinaryOp::MUL: resultValue = builder.createFMulInst(lhs, rhs); break;
|
||||
case BinaryOp::DIV: resultValue = builder.createFDivInst(lhs, rhs); break;
|
||||
case BinaryOp::MOD:
|
||||
std::cerr << "Error: Modulo operator not supported for float types." << std::endl;
|
||||
return;
|
||||
}
|
||||
// 将新创建的指令结果添加到缓存
|
||||
availableBinaryExpressions[currentExpKey] = resultValue;
|
||||
} else {
|
||||
std::cerr << "Error: Unsupported type for binary instruction." << std::endl;
|
||||
return;
|
||||
}
|
||||
}
|
||||
break;
|
||||
@@ -478,45 +316,36 @@ void SysYIRGenerator::compute() {
|
||||
return;
|
||||
}
|
||||
} else {
|
||||
// 否则,创建相应的IR指令 (在这里应用CSE)
|
||||
UnExpKey currentUnExpKey(static_cast<BinaryOp>(op), operand);
|
||||
auto it = availableUnaryExpressions.find(currentUnExpKey);
|
||||
if (it != availableUnaryExpressions.end()) {
|
||||
// 在缓存中找到,重用结果
|
||||
resultValue = it->second;
|
||||
} else {
|
||||
switch (op) {
|
||||
case BinaryOp::PLUS:
|
||||
resultValue = operand; // 一元加指令通常直接返回操作数
|
||||
break;
|
||||
case BinaryOp::NEG: {
|
||||
if (commonType == sysy::Type::getIntType()) {
|
||||
resultValue = builder.createNegInst(operand);
|
||||
} else if (commonType == sysy::Type::getFloatType()) {
|
||||
resultValue = builder.createFNegInst(operand);
|
||||
} else {
|
||||
std::cerr << "Error: Negation not supported for operand type." << std::endl;
|
||||
return;
|
||||
}
|
||||
break;
|
||||
}
|
||||
case BinaryOp::NOT:
|
||||
// 逻辑非
|
||||
if (commonType == sysy::Type::getIntType()) {
|
||||
resultValue = builder.createNotInst(operand);
|
||||
} else if (commonType == sysy::Type::getFloatType()) {
|
||||
resultValue = builder.createFNotInst(operand);
|
||||
} else {
|
||||
std::cerr << "Error: Logical NOT not supported for operand type." << std::endl;
|
||||
return;
|
||||
}
|
||||
break;
|
||||
default:
|
||||
std::cerr << "Error: Unknown unary operator for instructions: " << op << std::endl;
|
||||
return;
|
||||
// 否则,创建相应的IR指令
|
||||
switch (op) {
|
||||
case BinaryOp::PLUS:
|
||||
resultValue = operand; // 一元加指令通常直接返回操作数
|
||||
break;
|
||||
case BinaryOp::NEG: {
|
||||
if (commonType == sysy::Type::getIntType()) {
|
||||
resultValue = builder.createNegInst(operand);
|
||||
} else if (commonType == sysy::Type::getFloatType()) {
|
||||
resultValue = builder.createFNegInst(operand);
|
||||
} else {
|
||||
std::cerr << "Error: Negation not supported for operand type." << std::endl;
|
||||
return;
|
||||
}
|
||||
// 将新创建的指令结果添加到缓存
|
||||
availableUnaryExpressions[currentUnExpKey] = resultValue;
|
||||
break;
|
||||
}
|
||||
case BinaryOp::NOT:
|
||||
// 逻辑非
|
||||
if (commonType == sysy::Type::getIntType()) {
|
||||
resultValue = builder.createNotInst(operand);
|
||||
} else if (commonType == sysy::Type::getFloatType()) {
|
||||
resultValue = builder.createFNotInst(operand);
|
||||
} else {
|
||||
std::cerr << "Error: Logical NOT not supported for operand type." << std::endl;
|
||||
return;
|
||||
}
|
||||
break;
|
||||
default:
|
||||
std::cerr << "Error: Unknown unary operator for instructions: " << op << std::endl;
|
||||
return;
|
||||
}
|
||||
}
|
||||
break;
|
||||
@@ -658,19 +487,7 @@ Value* SysYIRGenerator::getGEPAddressInst(Value* basePointer, const std::vector<
|
||||
// `indices` 向量现在由调用方(如 visitLValue, visitVarDecl, visitAssignStmt)负责完整准备,
|
||||
// 包括是否需要添加初始的 `0` 索引。
|
||||
// 所以这里直接将其传递给 `builder.createGetElementPtrInst`。
|
||||
GEPKey key = {basePointer, indices};
|
||||
|
||||
// 尝试从缓存中查找
|
||||
auto it = availableGEPs.find(key);
|
||||
if (it != availableGEPs.end()) {
|
||||
return it->second; // 缓存命中,返回已有的 GEPInst*
|
||||
}
|
||||
|
||||
// 缓存未命中,创建新的 GEPInst
|
||||
Value* gepInst = builder.createGetElementPtrInst(basePointer, indices); // 假设 builder 提供了 createGEPInst 方法
|
||||
availableGEPs[key] = gepInst; // 将新的 GEPInst* 加入缓存
|
||||
|
||||
return gepInst;
|
||||
return builder.createGetElementPtrInst(basePointer, indices);
|
||||
}
|
||||
|
||||
/*
|
||||
@@ -716,7 +533,7 @@ std::any SysYIRGenerator::visitGlobalConstDecl(SysYParser::GlobalConstDeclContex
|
||||
if (!dims.empty()) { // 如果有维度,说明是数组
|
||||
variableType = buildArrayType(type, dims); // 构建完整的 ArrayType
|
||||
}
|
||||
module->createConstVar(name, Type::getPointerType(variableType), values);
|
||||
module->createConstVar(name, Type::getPointerType(variableType), values, dims);
|
||||
}
|
||||
return std::any();
|
||||
}
|
||||
@@ -745,7 +562,7 @@ std::any SysYIRGenerator::visitGlobalVarDecl(SysYParser::GlobalVarDeclContext *c
|
||||
if (!dims.empty()) { // 如果有维度,说明是数组
|
||||
variableType = buildArrayType(type, dims); // 构建完整的 ArrayType
|
||||
}
|
||||
module->createGlobalValue(name, Type::getPointerType(variableType), values);
|
||||
module->createGlobalValue(name, Type::getPointerType(variableType), dims, values);
|
||||
}
|
||||
return std::any();
|
||||
}
|
||||
@@ -769,13 +586,7 @@ std::any SysYIRGenerator::visitConstDecl(SysYParser::ConstDeclContext *ctx) {
|
||||
|
||||
// 显式地为局部常量在栈上分配空间
|
||||
// alloca 的类型将是指针指向常量类型,例如 `int*` 或 `int[2][3]*`
|
||||
// 将alloca全部集中到entry中
|
||||
auto entry = builder.getBasicBlock()->getParent()->getEntryBlock();
|
||||
auto it = builder.getPosition();
|
||||
auto nowblk = builder.getBasicBlock();
|
||||
builder.setPosition(entry, entry->terminator());
|
||||
AllocaInst *alloca = builder.createAllocaInst(Type::getPointerType(variableType), name);
|
||||
builder.setPosition(nowblk, it);
|
||||
AllocaInst *alloca = builder.createAllocaInst(Type::getPointerType(variableType), {}, name);
|
||||
|
||||
ArrayValueTree *root = std::any_cast<ArrayValueTree *>(constDef->constInitVal()->accept(this));
|
||||
ValueCounter values;
|
||||
@@ -842,44 +653,7 @@ std::any SysYIRGenerator::visitConstDecl(SysYParser::ConstDeclContext *ctx) {
|
||||
Value *currentValue = counterValues[k];
|
||||
unsigned currentRepeatNum = counterNumbers[k];
|
||||
|
||||
// 检查是否是0,并且重复次数足够大(例如 >16),才用 memset
|
||||
if (ConstantInteger *constInt = dynamic_cast<ConstantInteger *>(currentValue)) {
|
||||
if (constInt->getInt() == 0 && currentRepeatNum >= 16) { // 阈值可调整(如16、32等)
|
||||
// 计算 memset 的起始地址(基于当前线性偏移量)
|
||||
std::vector<Value *> memsetStartIndices;
|
||||
int tempLinearIndex = linearIndexOffset;
|
||||
|
||||
// 将线性索引转换为多维索引
|
||||
for (int dimIdx = dimSizes.size() - 1; dimIdx >= 0; --dimIdx) {
|
||||
memsetStartIndices.insert(memsetStartIndices.begin(),
|
||||
ConstantInteger::get(static_cast<int>(tempLinearIndex % dimSizes[dimIdx])));
|
||||
tempLinearIndex /= dimSizes[dimIdx];
|
||||
}
|
||||
|
||||
// 构造 GEP 计算 memset 的起始地址
|
||||
std::vector<Value *> gepIndicesForMemset;
|
||||
gepIndicesForMemset.push_back(ConstantInteger::get(0)); // 跳过 alloca 类型
|
||||
gepIndicesForMemset.insert(gepIndicesForMemset.end(), memsetStartIndices.begin(),
|
||||
memsetStartIndices.end());
|
||||
|
||||
Value *memsetPtr = builder.createGetElementPtrInst(alloca, gepIndicesForMemset);
|
||||
|
||||
// 计算 memset 的字节数 = 元素个数 × 元素大小
|
||||
Type *elementType = type;;
|
||||
uint64_t elementSize = elementType->getSize();
|
||||
Value *size = ConstantInteger::get(currentRepeatNum * elementSize);
|
||||
|
||||
// 生成 memset 指令(假设你的 IRBuilder 有 createMemset 方法)
|
||||
builder.createMemsetInst(memsetPtr, ConstantInteger::get(0), size, ConstantInteger::get(0));
|
||||
|
||||
// 跳过这些已处理的0
|
||||
linearIndexOffset += currentRepeatNum;
|
||||
continue; // 直接进入下一次循环
|
||||
}
|
||||
}
|
||||
|
||||
for (unsigned i = 0; i < currentRepeatNum; ++i) {
|
||||
// 对于非零值,生成对应的 store 指令
|
||||
std::vector<Value *> currentIndices;
|
||||
int tempLinearIndex = linearIndexOffset + i; // 使用偏移量和当前重复次数内的索引
|
||||
|
||||
@@ -932,12 +706,8 @@ std::any SysYIRGenerator::visitVarDecl(SysYParser::VarDeclContext *ctx) {
|
||||
|
||||
// 对于数组,alloca 的类型将是指针指向数组类型,例如 `int[2][3]*`
|
||||
// 对于标量,alloca 的类型将是指针指向标量类型,例如 `int*`
|
||||
auto entry = builder.getBasicBlock()->getParent()->getEntryBlock();
|
||||
auto it = builder.getPosition();
|
||||
auto nowblk = builder.getBasicBlock();
|
||||
builder.setPosition(entry, entry->terminator());
|
||||
AllocaInst *alloca = builder.createAllocaInst(Type::getPointerType(variableType), name);
|
||||
builder.setPosition(nowblk, it);
|
||||
AllocaInst* alloca =
|
||||
builder.createAllocaInst(Type::getPointerType(variableType), {}, name);
|
||||
|
||||
if (varDef->initVal() != nullptr) {
|
||||
ValueCounter values;
|
||||
@@ -991,73 +761,39 @@ std::any SysYIRGenerator::visitVarDecl(SysYParser::VarDeclContext *ctx) {
|
||||
ConstantInteger::get(0));
|
||||
}
|
||||
else {
|
||||
|
||||
|
||||
int linearIndexOffset = 0; // 用于追踪当前处理的线性索引的偏移量
|
||||
for (int k = 0; k < counterValues.size(); ++k) {
|
||||
// 当前 Value 的值和重复次数
|
||||
Value *currentValue = counterValues[k];
|
||||
unsigned currentRepeatNum = counterNumbers[k];
|
||||
// 检查是否是0,并且重复次数足够大(例如 >16),才用 memset
|
||||
if (ConstantInteger *constInt = dynamic_cast<ConstantInteger *>(currentValue)) {
|
||||
if (constInt->getInt() == 0 && currentRepeatNum >= 16) { // 阈值可调整(如16、32等)
|
||||
// 计算 memset 的起始地址(基于当前线性偏移量)
|
||||
std::vector<Value *> memsetStartIndices;
|
||||
int tempLinearIndex = linearIndexOffset;
|
||||
// 当前 Value 的值和重复次数
|
||||
Value* currentValue = counterValues[k];
|
||||
unsigned currentRepeatNum = counterNumbers[k];
|
||||
|
||||
// 将线性索引转换为多维索引
|
||||
for (int dimIdx = dimSizes.size() - 1; dimIdx >= 0; --dimIdx) {
|
||||
memsetStartIndices.insert(memsetStartIndices.begin(),
|
||||
ConstantInteger::get(static_cast<int>(tempLinearIndex % dimSizes[dimIdx])));
|
||||
tempLinearIndex /= dimSizes[dimIdx];
|
||||
}
|
||||
for (unsigned i = 0; i < currentRepeatNum; ++i) {
|
||||
std::vector<Value *> currentIndices;
|
||||
int tempLinearIndex = linearIndexOffset + i; // 使用偏移量和当前重复次数内的索引
|
||||
|
||||
// 构造 GEP 计算 memset 的起始地址
|
||||
std::vector<Value *> gepIndicesForMemset;
|
||||
gepIndicesForMemset.push_back(ConstantInteger::get(0)); // 跳过 alloca 类型
|
||||
gepIndicesForMemset.insert(gepIndicesForMemset.end(), memsetStartIndices.begin(),
|
||||
memsetStartIndices.end());
|
||||
|
||||
Value *memsetPtr = builder.createGetElementPtrInst(alloca, gepIndicesForMemset);
|
||||
|
||||
// 计算 memset 的字节数 = 元素个数 × 元素大小
|
||||
Type *elementType = type;
|
||||
;
|
||||
uint64_t elementSize = elementType->getSize();
|
||||
Value *size = ConstantInteger::get(currentRepeatNum * elementSize);
|
||||
|
||||
// 生成 memset 指令(假设你的 IRBuilder 有 createMemset 方法)
|
||||
builder.createMemsetInst(memsetPtr, ConstantInteger::get(0), size, ConstantInteger::get(0));
|
||||
|
||||
// 跳过这些已处理的0
|
||||
linearIndexOffset += currentRepeatNum;
|
||||
continue; // 直接进入下一次循环
|
||||
// 将线性索引转换为多维索引
|
||||
for (int dimIdx = dimSizes.size() - 1; dimIdx >= 0; --dimIdx) {
|
||||
currentIndices.insert(currentIndices.begin(),
|
||||
ConstantInteger::get(static_cast<int>(tempLinearIndex % dimSizes[dimIdx])));
|
||||
tempLinearIndex /= dimSizes[dimIdx];
|
||||
}
|
||||
|
||||
// 对于局部数组,alloca 本身就是 GEP 的基指针。
|
||||
// GEP 的第一个索引必须是 0,用于“步过”整个数组。
|
||||
std::vector<Value*> gepIndicesForInit;
|
||||
gepIndicesForInit.push_back(ConstantInteger::get(0));
|
||||
gepIndicesForInit.insert(gepIndicesForInit.end(), currentIndices.begin(), currentIndices.end());
|
||||
|
||||
// 计算元素的地址
|
||||
Value* elementAddress = getGEPAddressInst(alloca, gepIndicesForInit);
|
||||
// 生成 store 指令
|
||||
builder.createStoreInst(currentValue, elementAddress);
|
||||
}
|
||||
}
|
||||
for (unsigned i = 0; i < currentRepeatNum; ++i) {
|
||||
std::vector<Value *> currentIndices;
|
||||
int tempLinearIndex = linearIndexOffset + i; // 使用偏移量和当前重复次数内的索引
|
||||
|
||||
// 将线性索引转换为多维索引
|
||||
for (int dimIdx = dimSizes.size() - 1; dimIdx >= 0; --dimIdx) {
|
||||
currentIndices.insert(currentIndices.begin(),
|
||||
ConstantInteger::get(static_cast<int>(tempLinearIndex % dimSizes[dimIdx])));
|
||||
tempLinearIndex /= dimSizes[dimIdx];
|
||||
}
|
||||
|
||||
// 对于局部数组,alloca 本身就是 GEP 的基指针。
|
||||
// GEP 的第一个索引必须是 0,用于“步过”整个数组。
|
||||
std::vector<Value *> gepIndicesForInit;
|
||||
gepIndicesForInit.push_back(ConstantInteger::get(0));
|
||||
gepIndicesForInit.insert(gepIndicesForInit.end(), currentIndices.begin(), currentIndices.end());
|
||||
|
||||
// 计算元素的地址
|
||||
Value *elementAddress = getGEPAddressInst(alloca, gepIndicesForInit);
|
||||
// 生成 store 指令
|
||||
builder.createStoreInst(currentValue, elementAddress);
|
||||
}
|
||||
// 更新线性索引偏移量,以便下一次迭代从正确的位置开始
|
||||
linearIndexOffset += currentRepeatNum;
|
||||
// 更新线性索引偏移量,以便下一次迭代从正确的位置开始
|
||||
linearIndexOffset += currentRepeatNum;
|
||||
}
|
||||
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -1139,8 +875,6 @@ std::any SysYIRGenerator::visitFuncType(SysYParser::FuncTypeContext *ctx) {
|
||||
std::any SysYIRGenerator::visitFuncDef(SysYParser::FuncDefContext *ctx){
|
||||
// 更新作用域
|
||||
module->enterNewScope();
|
||||
// 清除CSE缓存
|
||||
enterNewBasicBlock();
|
||||
|
||||
auto name = ctx->Ident()->getText();
|
||||
std::vector<Type *> paramActualTypes;
|
||||
@@ -1161,7 +895,7 @@ std::any SysYIRGenerator::visitFuncDef(SysYParser::FuncDefContext *ctx){
|
||||
currentParamDims.push_back(ConstantInteger::get(-1)); // 标记第一个维度为未知
|
||||
for (const auto &exp : param->exp()) {
|
||||
// 访问表达式以获取维度大小,这些维度必须是常量
|
||||
Value* dimVal = computeExp(exp);
|
||||
Value* dimVal = std::any_cast<Value *>(visitExp(exp));
|
||||
// 确保维度是常量整数,否则 buildArrayType 会断言失败
|
||||
assert(dynamic_cast<ConstantInteger*>(dimVal) && "Array dimension in parameter must be a constant integer!");
|
||||
currentParamDims.push_back(dimVal);
|
||||
@@ -1216,7 +950,7 @@ std::any SysYIRGenerator::visitFuncDef(SysYParser::FuncDefContext *ctx){
|
||||
auto funcArgs = function->getArguments();
|
||||
std::vector<AllocaInst *> allocas;
|
||||
for (int i = 0; i < paramActualTypes.size(); ++i) {
|
||||
AllocaInst *alloca = builder.createAllocaInst(Type::getPointerType(paramActualTypes[i]), paramNames[i]);
|
||||
AllocaInst *alloca = builder.createAllocaInst(Type::getPointerType(paramActualTypes[i]), {}, paramNames[i]);
|
||||
allocas.push_back(alloca);
|
||||
module->addVariable(paramNames[i], alloca);
|
||||
}
|
||||
@@ -1231,8 +965,7 @@ std::any SysYIRGenerator::visitFuncDef(SysYParser::FuncDefContext *ctx){
|
||||
BasicBlock* funcBodyEntry = function->addBasicBlock("funcBodyEntry_" + name);
|
||||
|
||||
// 从 entryBB 无条件跳转到 funcBodyEntry
|
||||
builder.createUncondBrInst(funcBodyEntry);
|
||||
BasicBlock::conectBlocks(entry, funcBodyEntry); // 连接 entryBB 和 funcBodyEntry
|
||||
builder.createUncondBrInst(funcBodyEntry, {});
|
||||
builder.setPosition(funcBodyEntry,funcBodyEntry->end()); // 将插入点设置到 funcBodyEntry
|
||||
|
||||
for (auto item : ctx->blockStmt()->blockItem()) {
|
||||
@@ -1296,16 +1029,7 @@ std::any SysYIRGenerator::visitAssignStmt(SysYParser::AssignStmtContext *ctx) {
|
||||
if (AllocaInst *alloc = dynamic_cast<AllocaInst *>(variable)) {
|
||||
Type* allocatedType = alloc->getType()->as<PointerType>()->getBaseType();
|
||||
if (allocatedType->isPointer()) {
|
||||
// 尝试从缓存中获取 builder.createLoadInst(alloc) 的结果
|
||||
auto it = availableLoads.find(alloc);
|
||||
if (it != availableLoads.end()) {
|
||||
gepBasePointer = it->second; // 缓存命中,重用
|
||||
} else {
|
||||
gepBasePointer = builder.createLoadInst(alloc); // 缓存未命中,创建新的 LoadInst
|
||||
availableLoads[alloc] = gepBasePointer; // 将结果加入缓存
|
||||
}
|
||||
// --- CSE 结束 ---
|
||||
// gepBasePointer = builder.createLoadInst(alloc);
|
||||
gepBasePointer = builder.createLoadInst(alloc);
|
||||
gepIndices = indices;
|
||||
} else {
|
||||
gepBasePointer = alloc;
|
||||
@@ -1367,9 +1091,9 @@ std::any SysYIRGenerator::visitAssignStmt(SysYParser::AssignStmtContext *ctx) {
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
builder.createStoreInst(RValue, LValue);
|
||||
invalidateExpressionsOnStore(LValue);
|
||||
|
||||
return std::any();
|
||||
}
|
||||
|
||||
@@ -1406,9 +1130,7 @@ std::any SysYIRGenerator::visitIfStmt(SysYParser::IfStmtContext *ctx) {
|
||||
labelstring.str("");
|
||||
function->addBasicBlock(thenBlock);
|
||||
builder.setPosition(thenBlock, thenBlock->end());
|
||||
// CSE清除缓存
|
||||
enterNewBasicBlock();
|
||||
|
||||
|
||||
auto block = dynamic_cast<SysYParser::BlockStmtContext *>(ctx->stmt(0));
|
||||
// 如果是块语句,直接访问
|
||||
// 否则访问语句
|
||||
@@ -1419,7 +1141,7 @@ std::any SysYIRGenerator::visitIfStmt(SysYParser::IfStmtContext *ctx) {
|
||||
ctx->stmt(0)->accept(this);
|
||||
module->leaveScope();
|
||||
}
|
||||
builder.createUncondBrInst(exitBlock);
|
||||
builder.createUncondBrInst(exitBlock, {});
|
||||
BasicBlock::conectBlocks(builder.getBasicBlock(), exitBlock);
|
||||
|
||||
labelstring << "if_else.L" << builder.getLabelIndex();
|
||||
@@ -1427,9 +1149,7 @@ std::any SysYIRGenerator::visitIfStmt(SysYParser::IfStmtContext *ctx) {
|
||||
labelstring.str("");
|
||||
function->addBasicBlock(elseBlock);
|
||||
builder.setPosition(elseBlock, elseBlock->end());
|
||||
// CSE清除缓存
|
||||
enterNewBasicBlock();
|
||||
|
||||
|
||||
block = dynamic_cast<SysYParser::BlockStmtContext *>(ctx->stmt(1));
|
||||
if (block != nullptr) {
|
||||
visitBlockStmt(block);
|
||||
@@ -1438,7 +1158,7 @@ std::any SysYIRGenerator::visitIfStmt(SysYParser::IfStmtContext *ctx) {
|
||||
ctx->stmt(1)->accept(this);
|
||||
module->leaveScope();
|
||||
}
|
||||
builder.createUncondBrInst(exitBlock);
|
||||
builder.createUncondBrInst(exitBlock, {});
|
||||
BasicBlock::conectBlocks(builder.getBasicBlock(), exitBlock);
|
||||
|
||||
labelstring << "if_exit.L" << builder.getLabelIndex();
|
||||
@@ -1446,9 +1166,7 @@ std::any SysYIRGenerator::visitIfStmt(SysYParser::IfStmtContext *ctx) {
|
||||
labelstring.str("");
|
||||
function->addBasicBlock(exitBlock);
|
||||
builder.setPosition(exitBlock, exitBlock->end());
|
||||
// CSE清除缓存
|
||||
enterNewBasicBlock();
|
||||
|
||||
|
||||
} else {
|
||||
builder.pushTrueBlock(thenBlock);
|
||||
builder.pushFalseBlock(exitBlock);
|
||||
@@ -1461,9 +1179,7 @@ std::any SysYIRGenerator::visitIfStmt(SysYParser::IfStmtContext *ctx) {
|
||||
labelstring.str("");
|
||||
function->addBasicBlock(thenBlock);
|
||||
builder.setPosition(thenBlock, thenBlock->end());
|
||||
// CSE清除缓存
|
||||
enterNewBasicBlock();
|
||||
|
||||
|
||||
auto block = dynamic_cast<SysYParser::BlockStmtContext *>(ctx->stmt(0));
|
||||
if (block != nullptr) {
|
||||
visitBlockStmt(block);
|
||||
@@ -1472,7 +1188,7 @@ std::any SysYIRGenerator::visitIfStmt(SysYParser::IfStmtContext *ctx) {
|
||||
ctx->stmt(0)->accept(this);
|
||||
module->leaveScope();
|
||||
}
|
||||
builder.createUncondBrInst(exitBlock);
|
||||
builder.createUncondBrInst(exitBlock, {});
|
||||
BasicBlock::conectBlocks(builder.getBasicBlock(), exitBlock);
|
||||
|
||||
labelstring << "if_exit.L" << builder.getLabelIndex();
|
||||
@@ -1480,9 +1196,6 @@ std::any SysYIRGenerator::visitIfStmt(SysYParser::IfStmtContext *ctx) {
|
||||
labelstring.str("");
|
||||
function->addBasicBlock(exitBlock);
|
||||
builder.setPosition(exitBlock, exitBlock->end());
|
||||
// CSE清除缓存
|
||||
enterNewBasicBlock();
|
||||
|
||||
}
|
||||
return std::any();
|
||||
}
|
||||
@@ -1497,12 +1210,10 @@ std::any SysYIRGenerator::visitWhileStmt(SysYParser::WhileStmtContext *ctx) {
|
||||
labelstring << "while_head.L" << builder.getLabelIndex();
|
||||
BasicBlock *headBlock = function->addBasicBlock(labelstring.str());
|
||||
labelstring.str("");
|
||||
builder.createUncondBrInst(headBlock);
|
||||
builder.createUncondBrInst(headBlock, {});
|
||||
BasicBlock::conectBlocks(curBlock, headBlock);
|
||||
builder.setPosition(headBlock, headBlock->end());
|
||||
// CSE清除缓存
|
||||
enterNewBasicBlock();
|
||||
|
||||
|
||||
BasicBlock* bodyBlock = new BasicBlock(function);
|
||||
BasicBlock* exitBlock = new BasicBlock(function);
|
||||
|
||||
@@ -1518,8 +1229,6 @@ std::any SysYIRGenerator::visitWhileStmt(SysYParser::WhileStmtContext *ctx) {
|
||||
labelstring.str("");
|
||||
function->addBasicBlock(bodyBlock);
|
||||
builder.setPosition(bodyBlock, bodyBlock->end());
|
||||
// CSE清除缓存
|
||||
enterNewBasicBlock();
|
||||
|
||||
builder.pushBreakBlock(exitBlock);
|
||||
builder.pushContinueBlock(headBlock);
|
||||
@@ -1534,7 +1243,7 @@ std::any SysYIRGenerator::visitWhileStmt(SysYParser::WhileStmtContext *ctx) {
|
||||
module->leaveScope();
|
||||
}
|
||||
|
||||
builder.createUncondBrInst(headBlock);
|
||||
builder.createUncondBrInst(headBlock, {});
|
||||
BasicBlock::conectBlocks(builder.getBasicBlock(), exitBlock);
|
||||
builder.popBreakBlock();
|
||||
builder.popContinueBlock();
|
||||
@@ -1544,22 +1253,20 @@ std::any SysYIRGenerator::visitWhileStmt(SysYParser::WhileStmtContext *ctx) {
|
||||
labelstring.str("");
|
||||
function->addBasicBlock(exitBlock);
|
||||
builder.setPosition(exitBlock, exitBlock->end());
|
||||
// CSE清除缓存
|
||||
enterNewBasicBlock();
|
||||
|
||||
|
||||
return std::any();
|
||||
}
|
||||
|
||||
std::any SysYIRGenerator::visitBreakStmt(SysYParser::BreakStmtContext *ctx) {
|
||||
BasicBlock* breakBlock = builder.getBreakBlock();
|
||||
builder.createUncondBrInst(breakBlock);
|
||||
builder.createUncondBrInst(breakBlock, {});
|
||||
BasicBlock::conectBlocks(builder.getBasicBlock(), breakBlock);
|
||||
return std::any();
|
||||
}
|
||||
|
||||
std::any SysYIRGenerator::visitContinueStmt(SysYParser::ContinueStmtContext *ctx) {
|
||||
BasicBlock* continueBlock = builder.getContinueBlock();
|
||||
builder.createUncondBrInst(continueBlock);
|
||||
builder.createUncondBrInst(continueBlock, {});
|
||||
BasicBlock::conectBlocks(builder.getBasicBlock(), continueBlock);
|
||||
return std::any();
|
||||
}
|
||||
@@ -1661,34 +1368,62 @@ std::any SysYIRGenerator::visitLValue(SysYParser::LValueContext *ctx) {
|
||||
|
||||
// 3. 处理可变变量 (AllocaInst/GlobalValue) 或带非常量索引的常量变量
|
||||
// 这里区分标量访问和数组元素/子数组访问
|
||||
Value *targetAddress = nullptr;
|
||||
|
||||
// 检查是否是访问标量变量本身(没有索引,且声明维度为0)
|
||||
if (dims.empty() && declaredNumDims == 0) {
|
||||
// 对于标量变量,直接加载其值。
|
||||
// variable 本身就是指向标量的指针 (e.g., int* %a)
|
||||
if (dynamic_cast<AllocaInst*>(variable) || dynamic_cast<GlobalValue*>(variable)) {
|
||||
targetAddress = variable;
|
||||
value = builder.createLoadInst(variable);
|
||||
} else {
|
||||
// 如果走到这里且不是AllocaInst/GlobalValue,但dims为空且declaredNumDims为0,
|
||||
// 且又不是ConstantVariable (前面已处理),则可能是错误情况。
|
||||
assert(false && "Unhandled scalar variable type in LValue access.");
|
||||
return static_cast<Value*>(nullptr);
|
||||
}
|
||||
} else {
|
||||
// 访问数组元素或子数组(有索引,或变量本身是数组/多维指针)
|
||||
Value* gepBasePointer = nullptr;
|
||||
std::vector<Value*> gepIndices;
|
||||
std::vector<Value*> gepIndices; // 准备传递给 getGEPAddressInst 的索引列表
|
||||
// GEP 的基指针就是变量本身(它是一个指向内存的指针)
|
||||
if (AllocaInst *alloc = dynamic_cast<AllocaInst *>(variable)) {
|
||||
// 情况 A: 局部变量 (AllocaInst)
|
||||
// 获取 AllocaInst 分配的内存的实际类型。
|
||||
// 例如:对于 `int b[10][20];`,`allocatedType` 是 `[10 x [20 x i32]]`。
|
||||
// 对于 `int b[][20]` 的函数参数,其 AllocaInst 存储的是一个指针,
|
||||
// 此时 `allocatedType` 是 `[20 x i32]*`。
|
||||
Type* allocatedType = alloc->getType()->as<PointerType>()->getBaseType();
|
||||
|
||||
if (allocatedType->isPointer()) {
|
||||
gepBasePointer = builder.createLoadInst(alloc);
|
||||
// 如果 AllocaInst 分配的是一个指针类型 (例如,用于存储函数参数的指针,如 int b[][20] 中的 b)
|
||||
// 即 `allocatedType` 是一个指向数组指针的指针 (e.g., [20 x i32]**)
|
||||
// 那么 GEP 的基指针是加载这个指针变量的值。
|
||||
gepBasePointer = builder.createLoadInst(alloc); // 加载出实际的指针值 (e.g., [20 x i32]*)
|
||||
// 对于这种参数指针,用户提供的索引直接作用于它。不需要额外的 0。
|
||||
gepIndices = dims;
|
||||
} else {
|
||||
gepBasePointer = alloc;
|
||||
// 如果 AllocaInst 分配的是实际的数组数据 (例如,int b[10][20] 中的 b)
|
||||
// 那么 AllocaInst 本身就是 GEP 的基指针。
|
||||
// 这里的 `alloc` 是指向数组的指针 (e.g., [10 x [20 x i32]]*)
|
||||
gepBasePointer = alloc; // 类型是 [10 x [20 x i32]]*
|
||||
// 对于这种完整的数组分配,GEP 的第一个索引必须是 0,用于“步过”整个数组。
|
||||
gepIndices.push_back(ConstantInteger::get(0));
|
||||
gepIndices.insert(gepIndices.end(), dims.begin(), dims.end());
|
||||
}
|
||||
} else if (GlobalValue *glob = dynamic_cast<GlobalValue *>(variable)) {
|
||||
gepBasePointer = glob;
|
||||
// 情况 B: 全局变量 (GlobalValue)
|
||||
// GlobalValue 总是指向全局数据的指针。
|
||||
gepBasePointer = glob; // 类型是 [61 x [67 x i32]]*
|
||||
// 对于全局数组,GEP 的第一个索引必须是 0,用于“步过”整个数组。
|
||||
gepIndices.push_back(ConstantInteger::get(0));
|
||||
gepIndices.insert(gepIndices.end(), dims.begin(), dims.end());
|
||||
} else if (ConstantVariable *constV = dynamic_cast<ConstantVariable *>(variable)) {
|
||||
// 情况 C: 常量变量 (ConstantVariable),如果它代表全局数组常量
|
||||
// 假设 ConstantVariable 可以直接作为 GEP 的基指针。
|
||||
gepBasePointer = constV;
|
||||
// 对于常量数组,也需要 0 索引来“步过”整个数组。
|
||||
// 这里可以进一步检查 constV->getType()->as<PointerType>()->getBaseType()->isArray()
|
||||
// 但为了简洁,假设所有 ConstantVariable 作为 GEP 基指针时都需要此 0。
|
||||
gepIndices.push_back(ConstantInteger::get(0));
|
||||
gepIndices.insert(gepIndices.end(), dims.begin(), dims.end());
|
||||
} else {
|
||||
@@ -1696,25 +1431,18 @@ std::any SysYIRGenerator::visitLValue(SysYParser::LValueContext *ctx) {
|
||||
return static_cast<Value *>(nullptr);
|
||||
}
|
||||
|
||||
targetAddress = getGEPAddressInst(gepBasePointer, gepIndices);
|
||||
// 现在调用 getGEPAddressInst,传入正确准备的基指针和索引列表
|
||||
Value *targetAddress = getGEPAddressInst(gepBasePointer, gepIndices);
|
||||
|
||||
}
|
||||
|
||||
// 如果提供的索引数量少于声明的维度数量,则表示访问的是子数组,返回其地址 (无需加载)
|
||||
if (dims.size() < declaredNumDims) {
|
||||
value = targetAddress;
|
||||
} else {
|
||||
// value = builder.createLoadInst(targetAddress);
|
||||
auto it = availableLoads.find(targetAddress);
|
||||
if (it != availableLoads.end()) {
|
||||
value = it->second; // 缓存命中,重用已有的 LoadInst 结果
|
||||
// 如果提供的索引数量少于声明的维度数量,则表示访问的是子数组,返回其地址
|
||||
if (dims.size() < declaredNumDims) {
|
||||
value = targetAddress;
|
||||
} else {
|
||||
// 缓存未命中,创建新的 LoadInst
|
||||
value = builder.createLoadInst(targetAddress);
|
||||
availableLoads[targetAddress] = value; // 将新的 LoadInst 结果加入缓存
|
||||
// 否则,表示访问的是最终的标量元素,加载其值
|
||||
// 假设 createLoadInst 接受 Value* pointer
|
||||
value = builder.createLoadInst(targetAddress);
|
||||
}
|
||||
}
|
||||
|
||||
return value;
|
||||
}
|
||||
|
||||
@@ -1834,7 +1562,6 @@ std::any SysYIRGenerator::visitUnaryExp(SysYParser::UnaryExpContext *ctx) {
|
||||
visitPrimaryExp(ctx->primaryExp());
|
||||
} else if (ctx->call() != nullptr) {
|
||||
BinaryExpStack.push_back(std::any_cast<Value *>(visitCall(ctx->call())));BinaryExpLenStack.back()++;
|
||||
invalidateExpressionsOnCall();
|
||||
} else if (ctx->unaryOp() != nullptr) {
|
||||
// 遇到一元操作符,将其压入 BinaryExpStack
|
||||
auto opNode = dynamic_cast<antlr4::tree::TerminalNode*>(ctx->unaryOp()->children[0]);
|
||||
@@ -2080,7 +1807,7 @@ std::any SysYIRGenerator::visitLAndExp(SysYParser::LAndExpContext *ctx){
|
||||
labelstring.str("");
|
||||
|
||||
auto cond = std::any_cast<Value *>(visitEqExp(ctx->eqExp(i)));
|
||||
builder.createCondBrInst(cond, newtrueBlock, falseBlock);
|
||||
builder.createCondBrInst(cond, newtrueBlock, falseBlock, {}, {});
|
||||
|
||||
BasicBlock::conectBlocks(curBlock, newtrueBlock);
|
||||
BasicBlock::conectBlocks(curBlock, falseBlock);
|
||||
@@ -2090,7 +1817,7 @@ std::any SysYIRGenerator::visitLAndExp(SysYParser::LAndExpContext *ctx){
|
||||
}
|
||||
|
||||
auto cond = std::any_cast<Value *>(visitEqExp(conds.back()));
|
||||
builder.createCondBrInst(cond, trueBlock, falseBlock);
|
||||
builder.createCondBrInst(cond, trueBlock, falseBlock, {}, {});
|
||||
|
||||
BasicBlock::conectBlocks(curBlock, trueBlock);
|
||||
BasicBlock::conectBlocks(curBlock, falseBlock);
|
||||
@@ -2207,7 +1934,7 @@ void Utils::createExternalFunction(
|
||||
for (int i = 0; i < paramTypes.size(); ++i) {
|
||||
auto arg = new Argument(paramTypes[i], function, i, paramNames[i]);
|
||||
auto alloca = pBuilder->createAllocaInst(
|
||||
Type::getPointerType(paramTypes[i]), paramNames[i]);
|
||||
Type::getPointerType(paramTypes[i]), {}, paramNames[i]);
|
||||
function->insertArgument(arg);
|
||||
auto store = pBuilder->createStoreInst(arg, alloca);
|
||||
pModule->addVariable(paramNames[i], alloca);
|
||||
|
||||
@@ -240,8 +240,6 @@ void SysYPrinter::printInst(Instruction *pInst) {
|
||||
case Kind::kMul:
|
||||
case Kind::kDiv:
|
||||
case Kind::kRem:
|
||||
case Kind::kSRA:
|
||||
case Kind::kMulh:
|
||||
case Kind::kFAdd:
|
||||
case Kind::kFSub:
|
||||
case Kind::kFMul:
|
||||
@@ -274,8 +272,6 @@ void SysYPrinter::printInst(Instruction *pInst) {
|
||||
case Kind::kMul: std::cout << "mul"; break;
|
||||
case Kind::kDiv: std::cout << "sdiv"; break;
|
||||
case Kind::kRem: std::cout << "srem"; break;
|
||||
case Kind::kSRA: std::cout << "ashr"; break;
|
||||
case Kind::kMulh: std::cout << "mulh"; break;
|
||||
case Kind::kFAdd: std::cout << "fadd"; break;
|
||||
case Kind::kFSub: std::cout << "fsub"; break;
|
||||
case Kind::kFMul: std::cout << "fmul"; break;
|
||||
@@ -299,12 +295,7 @@ void SysYPrinter::printInst(Instruction *pInst) {
|
||||
|
||||
// Types and operands
|
||||
std::cout << " ";
|
||||
// For comparison operations, print operand types instead of result type
|
||||
if (pInst->getKind() >= Kind::kICmpEQ && pInst->getKind() <= Kind::kFCmpGE) {
|
||||
printType(binInst->getLhs()->getType());
|
||||
} else {
|
||||
printType(binInst->getType());
|
||||
}
|
||||
printType(binInst->getType());
|
||||
std::cout << " ";
|
||||
printValue(binInst->getLhs());
|
||||
std::cout << ", ";
|
||||
@@ -417,12 +408,7 @@ void SysYPrinter::printInst(Instruction *pInst) {
|
||||
}
|
||||
std::cout << std::endl;
|
||||
} break;
|
||||
|
||||
case Kind::kUnreachable: {
|
||||
std::cout << "Unreachable" << std::endl;
|
||||
|
||||
} break;
|
||||
|
||||
|
||||
case Kind::kAlloca: {
|
||||
auto allocaInst = dynamic_cast<AllocaInst *>(pInst);
|
||||
std::cout << "%" << allocaInst->getName() << " = alloca ";
|
||||
@@ -433,6 +419,17 @@ void SysYPrinter::printInst(Instruction *pInst) {
|
||||
auto allocatedType = allocaInst->getAllocatedType();
|
||||
printType(allocatedType);
|
||||
|
||||
// 仍然打印维度信息,如果存在的话
|
||||
if (allocaInst->getNumDims() > 0) {
|
||||
std::cout << ", ";
|
||||
for (size_t i = 0; i < allocaInst->getNumDims(); i++) {
|
||||
if (i > 0) std::cout << ", ";
|
||||
printType(Type::getIntType()); // 维度大小通常是 i32 类型
|
||||
std::cout << " ";
|
||||
printValue(allocaInst->getDim(i));
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << ", align 4" << std::endl;
|
||||
} break;
|
||||
|
||||
@@ -445,6 +442,17 @@ void SysYPrinter::printInst(Instruction *pInst) {
|
||||
std::cout << " ";
|
||||
printValue(loadInst->getPointer()); // 要加载的地址
|
||||
|
||||
// 仍然打印索引信息,如果存在的话
|
||||
if (loadInst->getNumIndices() > 0) {
|
||||
std::cout << ", indices "; // 或者其他分隔符,取决于你期望的格式
|
||||
for (size_t i = 0; i < loadInst->getNumIndices(); i++) {
|
||||
if (i > 0) std::cout << ", ";
|
||||
printType(loadInst->getIndex(i)->getType());
|
||||
std::cout << " ";
|
||||
printValue(loadInst->getIndex(i));
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << ", align 4" << std::endl;
|
||||
} break;
|
||||
|
||||
@@ -459,6 +467,16 @@ void SysYPrinter::printInst(Instruction *pInst) {
|
||||
std::cout << " ";
|
||||
printValue(storeInst->getPointer()); // 目标地址
|
||||
|
||||
// 仍然打印索引信息,如果存在的话
|
||||
if (storeInst->getNumIndices() > 0) {
|
||||
std::cout << ", indices "; // 或者其他分隔符
|
||||
for (size_t i = 0; i < storeInst->getNumIndices(); i++) {
|
||||
if (i > 0) std::cout << ", ";
|
||||
printType(storeInst->getIndex(i)->getType());
|
||||
std::cout << " ";
|
||||
printValue(storeInst->getIndex(i));
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << ", align 4" << std::endl;
|
||||
} break;
|
||||
@@ -517,9 +535,9 @@ void SysYPrinter::printInst(Instruction *pInst) {
|
||||
if (!firstPair) std::cout << ", ";
|
||||
firstPair = false;
|
||||
std::cout << "[ ";
|
||||
printValue(phiInst->getIncomingValue(i));
|
||||
printValue(phiInst->getValue(i));
|
||||
std::cout << ", %";
|
||||
printBlock(phiInst->getIncomingBlock(i));
|
||||
printBlock(phiInst->getBlock(i));
|
||||
std::cout << " ]";
|
||||
}
|
||||
std::cout << std::endl;
|
||||
|
||||
@@ -21,8 +21,6 @@ using namespace sysy;
|
||||
|
||||
int DEBUG = 0;
|
||||
int DEEPDEBUG = 0;
|
||||
int DEEPERDEBUG = 0;
|
||||
int DEBUGLENGTH = 50;
|
||||
|
||||
static string argStopAfter;
|
||||
static string argInputFile;
|
||||
|
||||
1
testdata/functional/00_main.out
vendored
Normal file
1
testdata/functional/00_main.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
3
|
||||
3
testdata/functional/00_main.sy
vendored
Normal file
3
testdata/functional/00_main.sy
vendored
Normal file
@@ -0,0 +1,3 @@
|
||||
int main(){
|
||||
return 3;
|
||||
}
|
||||
1
testdata/functional/01_var_defn2.out
vendored
Normal file
1
testdata/functional/01_var_defn2.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
10
|
||||
8
testdata/functional/01_var_defn2.sy
vendored
Normal file
8
testdata/functional/01_var_defn2.sy
vendored
Normal file
@@ -0,0 +1,8 @@
|
||||
//test domain of global var define and local define
|
||||
int a = 3;
|
||||
int b = 5;
|
||||
|
||||
int main(){
|
||||
int a = 5;
|
||||
return a + b;
|
||||
}
|
||||
1
testdata/functional/02_var_defn3.out
vendored
Normal file
1
testdata/functional/02_var_defn3.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
5
|
||||
8
testdata/functional/02_var_defn3.sy
vendored
Normal file
8
testdata/functional/02_var_defn3.sy
vendored
Normal file
@@ -0,0 +1,8 @@
|
||||
//test local var define
|
||||
int main(){
|
||||
int a, b0, _c;
|
||||
a = 1;
|
||||
b0 = 2;
|
||||
_c = 3;
|
||||
return b0 + _c;
|
||||
}
|
||||
1
testdata/functional/03_arr_defn2.out
vendored
Normal file
1
testdata/functional/03_arr_defn2.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
0
|
||||
4
testdata/functional/03_arr_defn2.sy
vendored
Normal file
4
testdata/functional/03_arr_defn2.sy
vendored
Normal file
@@ -0,0 +1,4 @@
|
||||
int a[10][10];
|
||||
int main(){
|
||||
return 0;
|
||||
}
|
||||
1
testdata/functional/04_arr_defn3.out
vendored
Normal file
1
testdata/functional/04_arr_defn3.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
14
|
||||
9
testdata/functional/04_arr_defn3.sy
vendored
Normal file
9
testdata/functional/04_arr_defn3.sy
vendored
Normal file
@@ -0,0 +1,9 @@
|
||||
//test array define
|
||||
int main(){
|
||||
int a[4][2] = {};
|
||||
int b[4][2] = {1, 2, 3, 4, 5, 6, 7, 8};
|
||||
int c[4][2] = {{1, 2}, {3, 4}, {5, 6}, {7, 8}};
|
||||
int d[4][2] = {1, 2, {3}, {5}, 7 , 8};
|
||||
int e[4][2] = {{d[2][1], c[2][1]}, {3, 4}, {5, 6}, {7, 8}};
|
||||
return e[3][1] + e[0][0] + e[0][1] + a[2][0];
|
||||
}
|
||||
1
testdata/functional/05_arr_defn4.out
vendored
Normal file
1
testdata/functional/05_arr_defn4.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
21
|
||||
9
testdata/functional/05_arr_defn4.sy
vendored
Normal file
9
testdata/functional/05_arr_defn4.sy
vendored
Normal file
@@ -0,0 +1,9 @@
|
||||
int main(){
|
||||
const int a[4][2] = {{1, 2}, {3, 4}, {}, 7};
|
||||
|
||||
int b[4][2] = {};
|
||||
int c[4][2] = {1, 2, 3, 4, 5, 6, 7, 8};
|
||||
int d[3 + 1][2] = {1, 2, {3}, {5}, a[3][0], 8};
|
||||
int e[4][2][1] = {{d[2][1], {c[2][1]}}, {3, 4}, {5, 6}, {7, 8}};
|
||||
return e[3][1][0] + e[0][0][0] + e[0][1][0] + d[3][0];
|
||||
}
|
||||
1
testdata/functional/06_const_var_defn2.out
vendored
Normal file
1
testdata/functional/06_const_var_defn2.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
5
|
||||
6
testdata/functional/06_const_var_defn2.sy
vendored
Normal file
6
testdata/functional/06_const_var_defn2.sy
vendored
Normal file
@@ -0,0 +1,6 @@
|
||||
//test const gloal var define
|
||||
const int a = 10, b = 5;
|
||||
|
||||
int main(){
|
||||
return b;
|
||||
}
|
||||
1
testdata/functional/07_const_var_defn3.out
vendored
Normal file
1
testdata/functional/07_const_var_defn3.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
5
|
||||
5
testdata/functional/07_const_var_defn3.sy
vendored
Normal file
5
testdata/functional/07_const_var_defn3.sy
vendored
Normal file
@@ -0,0 +1,5 @@
|
||||
//test const local var define
|
||||
int main(){
|
||||
const int a = 10, b = 5;
|
||||
return b;
|
||||
}
|
||||
1
testdata/functional/08_const_array_defn.out
vendored
Normal file
1
testdata/functional/08_const_array_defn.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
4
|
||||
5
testdata/functional/08_const_array_defn.sy
vendored
Normal file
5
testdata/functional/08_const_array_defn.sy
vendored
Normal file
@@ -0,0 +1,5 @@
|
||||
const int a[5]={0,1,2,3,4};
|
||||
|
||||
int main(){
|
||||
return a[4];
|
||||
}
|
||||
1
testdata/functional/09_func_defn.out
vendored
Normal file
1
testdata/functional/09_func_defn.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
9
|
||||
11
testdata/functional/09_func_defn.sy
vendored
Normal file
11
testdata/functional/09_func_defn.sy
vendored
Normal file
@@ -0,0 +1,11 @@
|
||||
int a;
|
||||
int func(int p){
|
||||
p = p - 1;
|
||||
return p;
|
||||
}
|
||||
int main(){
|
||||
int b;
|
||||
a = 10;
|
||||
b = func(a);
|
||||
return b;
|
||||
}
|
||||
1
testdata/functional/10_var_defn_func.out
vendored
Normal file
1
testdata/functional/10_var_defn_func.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
4
|
||||
8
testdata/functional/10_var_defn_func.sy
vendored
Normal file
8
testdata/functional/10_var_defn_func.sy
vendored
Normal file
@@ -0,0 +1,8 @@
|
||||
int defn(){
|
||||
return 4;
|
||||
}
|
||||
|
||||
int main(){
|
||||
int a=defn();
|
||||
return a;
|
||||
}
|
||||
1
testdata/functional/11_add2.out
vendored
Normal file
1
testdata/functional/11_add2.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
9
|
||||
7
testdata/functional/11_add2.sy
vendored
Normal file
7
testdata/functional/11_add2.sy
vendored
Normal file
@@ -0,0 +1,7 @@
|
||||
//test add
|
||||
int main(){
|
||||
int a, b;
|
||||
a = 10;
|
||||
b = -1;
|
||||
return a + b;
|
||||
}
|
||||
1
testdata/functional/12_addc.out
vendored
Normal file
1
testdata/functional/12_addc.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
15
|
||||
5
testdata/functional/12_addc.sy
vendored
Normal file
5
testdata/functional/12_addc.sy
vendored
Normal file
@@ -0,0 +1,5 @@
|
||||
//test addc
|
||||
const int a = 10;
|
||||
int main(){
|
||||
return a + 5;
|
||||
}
|
||||
1
testdata/functional/13_sub2.out
vendored
Normal file
1
testdata/functional/13_sub2.out
vendored
Normal file
@@ -0,0 +1 @@
|
||||
248
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user