[backend]引入了新的pass,负责消除数组复杂地址访问
This commit is contained in:
221
src/AddressCalculationExpansion.cpp
Normal file
221
src/AddressCalculationExpansion.cpp
Normal file
@@ -0,0 +1,221 @@
|
||||
#include "AddressCalculationExpansion.h"
|
||||
#include <iostream>
|
||||
#include <vector>
|
||||
#include <algorithm> // for std::reverse (if needed, although not used in final version)
|
||||
#include "IR.h" // 确保IR.h包含所有类型和指令的定义,特别是ConstantValue, AllocaInst, GlobalValue
|
||||
#include "IRBuilder.h" // 确保这里也包含 IRBuilder.h
|
||||
|
||||
extern int DEBUG; // 假设有一个全局的DEBUG宏
|
||||
|
||||
namespace sysy {
|
||||
|
||||
bool AddressCalculationExpansion::run() {
|
||||
bool changed = false;
|
||||
|
||||
// 遍历所有函数
|
||||
for (auto& funcPair : pModule->getFunctions()) {
|
||||
Function* func = funcPair.second.get();
|
||||
|
||||
// 遍历函数中的所有基本块
|
||||
for (auto& bb_ptr : func->getBasicBlocks()) {
|
||||
BasicBlock* bb = bb_ptr.get();
|
||||
// 使用list迭代器,因为可能会删除元素
|
||||
for (auto it = bb->getInstructions().begin(); it != bb->getInstructions().end(); ) {
|
||||
Instruction* inst = it->get();
|
||||
|
||||
Value* basePointer = nullptr;
|
||||
Value* valueToStore = nullptr; // 仅用于 StoreInst
|
||||
size_t firstIndexOperandIdx = 0; // 第一个索引操作数在指令操作数列表中的位置
|
||||
size_t numBaseOperands = 0; // Load/Store指令的基础操作数数量
|
||||
|
||||
if (inst->isLoad()) {
|
||||
// LoadInst: %result = load Type, Type* %ptr, Index1, Index2, ...
|
||||
// 操作数: pointer(0), index1(1), index2(2), ...
|
||||
numBaseOperands = 1; // 只有指针操作数
|
||||
basePointer = inst->getOperand(0);
|
||||
firstIndexOperandIdx = 1;
|
||||
} else if (inst->isStore()) {
|
||||
// StoreInst: store Type %val, Type* %ptr, Index1, Index2, ...
|
||||
// 操作数: value(0), pointer(1), index1(2), index2(3), ...
|
||||
numBaseOperands = 2; // value 和 pointer 操作数
|
||||
valueToStore = inst->getOperand(0);
|
||||
basePointer = inst->getOperand(1);
|
||||
firstIndexOperandIdx = 2;
|
||||
} else {
|
||||
++it;
|
||||
continue;
|
||||
}
|
||||
|
||||
// 如果操作数数量 <= 基础操作数数量,说明没有索引,跳过
|
||||
if (inst->getNumOperands() <= numBaseOperands) {
|
||||
++it;
|
||||
continue;
|
||||
}
|
||||
|
||||
// 获取数组的维度信息
|
||||
std::vector<int> dims; // 存储数组的维度大小(整数)
|
||||
|
||||
// 统一获取维度的方式,不再使用 ArrayType
|
||||
if (AllocaInst* allocaInst = dynamic_cast<AllocaInst*>(basePointer)) {
|
||||
// MODIFICATION: 将 getValue() 替换为 getValue<int>()
|
||||
for (const auto& use_ptr : allocaInst->getDims()) {
|
||||
Value* dimValue = use_ptr->getValue();
|
||||
if (ConstantValue* constVal = dynamic_cast<ConstantValue*>(dimValue)) {
|
||||
dims.push_back(constVal->getValue<int>()); // MODIFICATION: 明确指定模板参数为 int
|
||||
} else {
|
||||
// 如果维度不是 ConstantValue,或者不是整数,这里需要根据你的IR定义进行处理
|
||||
std::cerr << "Warning: AllocaInst dimension is not a constant integer. Skipping GEP expansion for: ";
|
||||
SysYPrinter::printValue(allocaInst);
|
||||
std::cerr << "\n";
|
||||
dims.clear(); // 清空,表示无法获取有效维度
|
||||
break;
|
||||
}
|
||||
}
|
||||
} else if (GlobalValue* globalValue = dynamic_cast<GlobalValue*>(basePointer)) {
|
||||
// 处理 GlobalValue 的维度
|
||||
// 假设 GlobalValue 也有类似 AllocaInst 的 getNumDims() 和 getDim() 方法
|
||||
// 这需要根据你的 GlobalValue 实际定义来验证和实现
|
||||
// 如果 GlobalValue::getType() 是 PointerType,其 baseType 可能是另一个 PointerType
|
||||
// 或者 GlobalValue 自身就存储了维度
|
||||
|
||||
// 暂时跳过 GlobalValue 的 GEP 展开,如果无法确定其维度
|
||||
std::cerr << "Warning: GlobalValue dimension handling needs explicit implementation for GEP expansion. Skipping GEP for: ";
|
||||
SysYPrinter::printValue(globalValue);
|
||||
std::cerr << "\n";
|
||||
++it;
|
||||
continue;
|
||||
|
||||
} else {
|
||||
// 基指针不是 AllocaInst 也不是 GlobalValue,或者无法处理其维度
|
||||
// 这可能是函数参数(指针到数组),或者其他复杂指针
|
||||
std::cerr << "Warning: Base pointer is not AllocaInst/GlobalValue or its array dimensions cannot be determined for GEP expansion. Skipping GEP for: ";
|
||||
SysYPrinter::printValue(basePointer);
|
||||
std::cerr << " in instruction ";
|
||||
SysYPrinter::printInst(inst);
|
||||
std::cerr << "\n";
|
||||
++it;
|
||||
continue;
|
||||
}
|
||||
|
||||
// 如果没有获取到有效的维度信息(例如,AllocaInst 没有维度但我们预期有,或者解析失败)
|
||||
// 且不是一个标量 Load/Store(即有索引)
|
||||
if (dims.empty() && (inst->getNumOperands() > numBaseOperands)) {
|
||||
if (DEBUG) {
|
||||
std::cerr << "ACE Warning: Could not get valid array dimensions for ";
|
||||
SysYPrinter::printValue(basePointer);
|
||||
std::cerr << " in instruction ";
|
||||
SysYPrinter::printInst(inst);
|
||||
std::cerr << " (expected dimensions for indices, but got none).\n";
|
||||
}
|
||||
++it;
|
||||
continue;
|
||||
}
|
||||
|
||||
// 收集所有索引操作数
|
||||
std::vector<Value*> indexOperands;
|
||||
for (size_t i = firstIndexOperandIdx; i < inst->getNumOperands(); ++i) {
|
||||
indexOperands.push_back(inst->getOperand(i));
|
||||
}
|
||||
|
||||
// 检查索引数量是否与维度匹配 (一个简单的健全性检查)
|
||||
// 只有当 basePointer 是 AllocaInst 或 GlobalValue 且有明确维度时才进行此检查
|
||||
if (AllocaInst* allocaInst = dynamic_cast<AllocaInst*>(basePointer)) {
|
||||
if (allocaInst->getNumDims() != indexOperands.size()) {
|
||||
if (DEBUG) {
|
||||
std::cerr << "ACE Warning: Index count (" << indexOperands.size() << ") does not match AllocaInst dimensions (" << allocaInst->getNumDims() << ") for instruction ";
|
||||
SysYPrinter::printInst(inst);
|
||||
std::cerr << "\n";
|
||||
}
|
||||
++it;
|
||||
continue;
|
||||
}
|
||||
}
|
||||
// TODO: 对 GlobalValue 也做类似检查
|
||||
|
||||
|
||||
// 最终线性偏移量 (字节数)
|
||||
// 初始化为0,因为所有偏移都是相对于基地址
|
||||
Value* totalOffset = ConstantValue::get(0);
|
||||
|
||||
// 在当前Load/Store指令之前设置插入点
|
||||
pBuilder->setPosition(bb, it);
|
||||
|
||||
// 遍历所有索引来计算地址
|
||||
// 累加每个维度贡献的偏移量
|
||||
for (size_t i = 0; i < indexOperands.size(); ++i) {
|
||||
Value* index = indexOperands[i];
|
||||
|
||||
// 计算当前维度索引的步长(字节数)
|
||||
int stride = calculateStride(dims, i);
|
||||
Value* strideConst = ConstantValue::get(stride);
|
||||
|
||||
// 1. 计算当前索引的偏移量: %current_dim_offset = index * stride
|
||||
Type* intType = Type::getIntType(); // 假设索引和偏移量都是int类型
|
||||
BinaryInst* currentDimOffsetInst = pBuilder->createBinaryInst(Instruction::kMul, intType, index, strideConst);
|
||||
// UD 链管理由 IRBuilder 负责,不需要手动 addUse
|
||||
|
||||
// 2. 将当前维度偏移量加到累计总偏移量上: %new_total_offset = total_offset + current_dim_offset
|
||||
BinaryInst* newTotalOffsetInst = pBuilder->createBinaryInst(Instruction::kAdd, intType, totalOffset, currentDimOffsetInst);
|
||||
// UD 链管理由 IRBuilder 负责,不需要手动 addUse
|
||||
|
||||
totalOffset = newTotalOffsetInst; // 更新累计总偏移量
|
||||
}
|
||||
|
||||
// 创建新的LoadInst或StoreInst,将计算出的总偏移量作为唯一的偏移量参数
|
||||
Instruction* newInst = nullptr;
|
||||
// 新 Load/Store 指令的索引列表,只包含计算出的总偏移量
|
||||
std::vector<Value*> newIndices = {totalOffset}; // 将 totalOffset 作为索引
|
||||
|
||||
if (inst->isLoad()) {
|
||||
newInst = pBuilder->createLoadInst(basePointer, newIndices);
|
||||
// 替换所有使用旧LoadInst结果的地方
|
||||
inst->replaceAllUsesWith(newInst);
|
||||
} else { // StoreInst
|
||||
newInst = pBuilder->createStoreInst(valueToStore, basePointer, newIndices);
|
||||
}
|
||||
|
||||
// 删除原始的LoadInst或StoreInst
|
||||
// 注意:这里需要先让迭代器指向下一个指令,再删除当前指令
|
||||
Instruction* oldInst = it->get();
|
||||
++it; // 提前移动迭代器
|
||||
|
||||
// 清理旧指令的UD链,修正 getOperandAsUse 的使用
|
||||
// Instruction 没有 getOperandAsUse()。我们通过 getOperand(i) 获取 Value*,
|
||||
// 然后从该 Value* 的 uses 列表中找到并移除对应 oldInst 的 Use。
|
||||
for(size_t i = 0; i < oldInst->getNumOperands(); ++i) {
|
||||
Value* operandValue = oldInst->getOperand(i);
|
||||
if (operandValue) {
|
||||
for (auto use_it = operandValue->getUses().begin(); use_it != operandValue->getUses().end(); ++use_it) {
|
||||
if ((*use_it)->getUser() == oldInst && (*use_it)->getIndex() == i) {
|
||||
operandValue->removeUse(*use_it);
|
||||
break; // 找到并移除了,可以跳出内层循环
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 从基本块中删除旧指令
|
||||
// 使用 BasicBlock 的 list 的 erase 方法
|
||||
bb->getInstructions().erase(std::prev(it)); // it 已经向前移动,所以需要 prev(it) 指向旧指令的位置
|
||||
|
||||
changed = true;
|
||||
if (DEBUG) {
|
||||
std::cerr << "ACE: Expanded array access. New insts leading to final offset:\n";
|
||||
// 检查 totalOffset 是否是 Instruction 类型才能打印
|
||||
if (Instruction* offsetInst = dynamic_cast<Instruction*>(totalOffset)) {
|
||||
SysYPrinter::printInst(offsetInst);
|
||||
} else {
|
||||
SysYPrinter::printValue(totalOffset); // 如果是常量,打印Value
|
||||
}
|
||||
std::cerr << "ACE: New Load/Store instruction:\n";
|
||||
SysYPrinter::printInst(newInst);
|
||||
std::cerr << "--------------------------------\n";
|
||||
}
|
||||
// 迭代器it已经提前移动,无需再次++it
|
||||
}
|
||||
}
|
||||
}
|
||||
return changed;
|
||||
}
|
||||
|
||||
} // namespace sysy
|
||||
@@ -24,6 +24,7 @@ add_executable(sysyc
|
||||
SysYIROptPre.cpp
|
||||
SysYIRAnalyser.cpp
|
||||
DeadCodeElimination.cpp
|
||||
AddressCalculationExpansion.cpp
|
||||
Mem2Reg.cpp
|
||||
Reg2Mem.cpp
|
||||
RISCv64Backend.cpp
|
||||
|
||||
59
src/include/AddressCalculationExpansion.h
Normal file
59
src/include/AddressCalculationExpansion.h
Normal file
@@ -0,0 +1,59 @@
|
||||
#pragma once
|
||||
|
||||
#include "IR.h" // 假设IR.h包含了Module, Function, BasicBlock, Instruction, Value, IRBuilder, Type等定义
|
||||
#include "IRBuilder.h" // 需要IRBuilder来创建新指令
|
||||
#include "SysYIRPrinter.h" // 新增: 用于调试输出
|
||||
#include <memory>
|
||||
#include <string>
|
||||
#include <unordered_map>
|
||||
#include <vector>
|
||||
#include <list> // 用于迭代和修改指令列表
|
||||
#include <algorithm> // for std::reverse (if needed, although not used in final version)
|
||||
#include <iostream> // MODIFICATION: 用于警告输出
|
||||
|
||||
namespace sysy {
|
||||
|
||||
/**
|
||||
* @brief AddressCalculationExpansion Pass
|
||||
*
|
||||
* 这是一个IR优化Pass,用于将LoadInst和StoreInst中包含的多维数组索引
|
||||
* 显式地转换为IR中的BinaryInst(乘法和加法)序列,并生成带有线性偏移量的
|
||||
* LoadInst/StoreInst。
|
||||
*
|
||||
* 目的:确保在寄存器分配之前,所有中间地址计算的结果都有明确的IR指令和对应的虚拟寄存器,
|
||||
* 从而避免在后端DAG构建时临时创建值而导致寄存器分配缺失的问题。
|
||||
*
|
||||
* SysY语言特性:
|
||||
* - 无指针类型(所有数组访问的基地址是alloca或global的AllocaType/ArrayType)
|
||||
* - 数据类型只有int和float,且都占用4字节。
|
||||
* - LoadInst和StoreInst直接接受多个索引作为额外操作数。
|
||||
*/
|
||||
class AddressCalculationExpansion {
|
||||
private:
|
||||
Module* pModule;
|
||||
IRBuilder* pBuilder; // 用于在IR中插入新指令
|
||||
|
||||
// 数组元素的固定大小,根据SysY特性,int和float都是4字节
|
||||
static const int ELEMENT_SIZE = 4;
|
||||
|
||||
// 辅助函数:根据数组的维度信息和当前索引的维度,计算该索引的步长(字节数)
|
||||
// dims: 包含所有维度大小的vector,例如 {2, 3, 4}
|
||||
// currentDimIndex: 当前正在处理的索引在 dims 中的位置 (0, 1, 2...)
|
||||
int calculateStride(const std::vector<int>& dims, size_t currentDimIndex) {
|
||||
int stride = ELEMENT_SIZE; // 最内层元素大小 (4字节)
|
||||
// 乘以当前维度之后的所有维度的大小
|
||||
for (size_t i = currentDimIndex + 1; i < dims.size(); ++i) {
|
||||
stride *= dims[i];
|
||||
}
|
||||
return stride;
|
||||
}
|
||||
|
||||
public:
|
||||
AddressCalculationExpansion(Module* module, IRBuilder* builder)
|
||||
: pModule(module), pBuilder(builder) {}
|
||||
|
||||
// 运行此Pass
|
||||
bool run();
|
||||
};
|
||||
|
||||
} // namespace sysy
|
||||
@@ -17,6 +17,7 @@ using namespace antlr4;
|
||||
#include "RISCv64Backend.h"
|
||||
#include "SysYIRAnalyser.h"
|
||||
#include "DeadCodeElimination.h"
|
||||
#include "AddressCalculationExpansion.h"
|
||||
#include "Mem2Reg.h"
|
||||
#include "Reg2Mem.h"
|
||||
|
||||
@@ -143,20 +144,19 @@ int main(int argc, char **argv) {
|
||||
cout << "=== After CFA & AVA (Default) ===\n";
|
||||
SysYPrinter(moduleIR).printIR(); // 临时打印器用于调试
|
||||
}
|
||||
|
||||
DeadCodeElimination dce(moduleIR, &cfa, &ava);
|
||||
dce.runDCEPipeline();
|
||||
if (DEBUG) {
|
||||
cout << "=== After 1st DCE (Default) ===\n";
|
||||
SysYPrinter(moduleIR).printIR();
|
||||
AddressCalculationExpansion ace(moduleIR, builder);
|
||||
if (ace.run()) {
|
||||
cout << "AddressCalculationExpansion made changes.\n";
|
||||
// 如果 ACE 改变了IR,并且 DEBUG 模式开启,可以考虑打印IR
|
||||
if (DEBUG) {
|
||||
cout << "=== After AddressCalculationExpansion ===\n";
|
||||
SysYPrinter(moduleIR).printIR();
|
||||
}
|
||||
} else {
|
||||
cout << "AddressCalculationExpansion made no changes.\n";
|
||||
}
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
// 根据优化级别,执行额外的优化 pass
|
||||
if (optLevel >= 1) {
|
||||
cout << "Applying additional -O" << optLevel << " optimizations...\n";
|
||||
@@ -174,6 +174,12 @@ int main(int argc, char **argv) {
|
||||
// MyCustomOpt2 opt2_pass(moduleIR, builder, &cfa); // 假设需要CFA
|
||||
// opt2_pass.run();
|
||||
// ... 更多 -O1 特有的优化
|
||||
DeadCodeElimination dce(moduleIR, &cfa, &ava);
|
||||
dce.runDCEPipeline();
|
||||
if (DEBUG) {
|
||||
cout << "=== After 1st DCE (Default) ===\n";
|
||||
SysYPrinter(moduleIR).printIR();
|
||||
}
|
||||
|
||||
Mem2Reg mem2reg(moduleIR, builder, &cfa, &ava);
|
||||
mem2reg.mem2regPipeline();
|
||||
@@ -212,7 +218,7 @@ int main(int argc, char **argv) {
|
||||
// 设置 DEBUG 模式(如果指定了 'asmd')
|
||||
if (argStopAfter == "asmd") {
|
||||
DEBUG = 1;
|
||||
DEEPDEBUG = 1;
|
||||
// DEEPDEBUG = 1;
|
||||
}
|
||||
sysy::RISCv64CodeGen codegen(moduleIR); // 传入优化后的 moduleIR
|
||||
string asmCode = codegen.code_gen();
|
||||
|
||||
Reference in New Issue
Block a user