Compare commits

...

5 Commits

11 changed files with 525 additions and 1382 deletions

View File

@ -47,7 +47,7 @@ bool AddressCalculationExpansion::run() {
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>());
dims.push_back(constVal->getInt());
} else {
std::cerr << "Warning: AllocaInst dimension is not a constant integer. Skipping GEP expansion for: ";
SysYPrinter::printValue(allocaInst);
@ -101,13 +101,13 @@ bool AddressCalculationExpansion::run() {
}
}
Value* totalOffset = ConstantValue::get(0);
Value* totalOffset = ConstantInteger::get(0);
pBuilder->setPosition(bb, it);
for (size_t i = 0; i < indexOperands.size(); ++i) {
Value* index = indexOperands[i];
int stride = calculateStride(dims, i);
Value* strideConst = ConstantValue::get(stride);
Value* strideConst = ConstantInteger::get(stride);
Type* intType = Type::getIntType();
BinaryInst* currentDimOffsetInst = pBuilder->createBinaryInst(Instruction::kMul, intType, index, strideConst);
BinaryInst* newTotalOffsetInst = pBuilder->createBinaryInst(Instruction::kAdd, intType, totalOffset, currentDimOffsetInst);

View File

@ -23,10 +23,10 @@ add_executable(sysyc
SysYIRPrinter.cpp
SysYIROptPre.cpp
SysYIRAnalyser.cpp
DeadCodeElimination.cpp
# DeadCodeElimination.cpp
AddressCalculationExpansion.cpp
Mem2Reg.cpp
Reg2Mem.cpp
# Mem2Reg.cpp
# Reg2Mem.cpp
RISCv64Backend.cpp
RISCv64ISel.cpp
RISCv64RegAlloc.cpp

View File

@ -1,276 +0,0 @@
#include "DeadCodeElimination.h"
#include <iostream>
extern int DEBUG;
namespace sysy {
void DeadCodeElimination::runDCEPipeline() {
const auto& functions = pModule->getFunctions();
for (const auto& function : functions) {
const auto& func = function.second;
bool changed = true;
while (changed) {
changed = false;
eliminateDeadStores(func.get(), changed);
eliminateDeadLoads(func.get(), changed);
eliminateDeadAllocas(func.get(), changed);
eliminateDeadRedundantLoadStore(func.get(), changed);
eliminateDeadGlobals(changed);
}
}
}
// 消除无用存储 消除条件:
// 存储的目标指针pointer不是全局变量!isGlobal(pointer))。
// 存储的目标指针不是数组参数(!isArr(pointer) 或不在函数参数列表里)。
// 该指针的所有使用者uses仅限 alloca 或 store即没有 load 或其他指令使用它)。
void DeadCodeElimination::eliminateDeadStores(Function* func, bool& changed) {
for (const auto& block : func->getBasicBlocks()) {
auto& instrs = block->getInstructions();
for (auto iter = instrs.begin(); iter != instrs.end();) {
auto inst = iter->get();
if (!inst->isStore()) {
++iter;
continue;
}
auto storeInst = dynamic_cast<StoreInst*>(inst);
auto pointer = storeInst->getPointer();
// 如果是全局变量或者是函数的数组参数
if (isGlobal(pointer) || (isArr(pointer) &&
std::find(func->getEntryBlock()->getArguments().begin(),
func->getEntryBlock()->getArguments().end(),
pointer) != func->getEntryBlock()->getArguments().end())) {
++iter;
continue;
}
bool changetag = true;
for (auto& use : pointer->getUses()) {
// 依次判断store的指针是否被其他指令使用
auto user = use->getUser();
auto userInst = dynamic_cast<Instruction*>(user);
// 如果使用store的指针的指令不是Alloca或Store则不删除
if (userInst != nullptr && !userInst->isAlloca() && !userInst->isStore()) {
changetag = false;
break;
}
}
if (changetag) {
changed = true;
if(DEBUG){
std::cout << "=== Dead Store Found ===\n";
SysYPrinter::printInst(storeInst);
}
usedelete(storeInst);
iter = instrs.erase(iter);
} else {
++iter;
}
}
}
}
// 消除无用加载 消除条件:
// 该指令的结果未被使用inst->getUses().empty())。
void DeadCodeElimination::eliminateDeadLoads(Function* func, bool& changed) {
for (const auto& block : func->getBasicBlocks()) {
auto& instrs = block->getInstructions();
for (auto iter = instrs.begin(); iter != instrs.end();) {
auto inst = iter->get();
if (inst->isBinary() || inst->isUnary() || inst->isLoad()) {
if (inst->getUses().empty()) {
changed = true;
if(DEBUG){
std::cout << "=== Dead Load Binary Unary Found ===\n";
SysYPrinter::printInst(inst);
}
usedelete(inst);
iter = instrs.erase(iter);
continue;
}
}
++iter;
}
}
}
// 消除无用加载 消除条件:
// 该 alloca 未被任何指令使用allocaInst->getUses().empty())。
// 该 alloca 不是函数的参数(不在 entry 块的参数列表里)。
void DeadCodeElimination::eliminateDeadAllocas(Function* func, bool& changed) {
for (const auto& block : func->getBasicBlocks()) {
auto& instrs = block->getInstructions();
for (auto iter = instrs.begin(); iter != instrs.end();) {
auto inst = iter->get();
if (inst->isAlloca()) {
auto allocaInst = dynamic_cast<AllocaInst*>(inst);
if (allocaInst->getUses().empty() &&
std::find(func->getEntryBlock()->getArguments().begin(),
func->getEntryBlock()->getArguments().end(),
allocaInst) == func->getEntryBlock()->getArguments().end()) {
changed = true;
if(DEBUG){
std::cout << "=== Dead Alloca Found ===\n";
SysYPrinter::printInst(inst);
}
usedelete(inst);
iter = instrs.erase(iter);
continue;
}
}
++iter;
}
}
}
void DeadCodeElimination::eliminateDeadIndirectiveAllocas(Function* func, bool& changed) {
// 删除mem2reg时引入的且现在已经没有value使用了的隐式alloca
FunctionAnalysisInfo* funcInfo = pCFA->getFunctionAnalysisInfo(func);
for (auto it = funcInfo->getIndirectAllocas().begin(); it != funcInfo->getIndirectAllocas().end();) {
auto &allocaInst = *it;
if (allocaInst->getUses().empty()) {
changed = true;
if(DEBUG){
std::cout << "=== Dead Indirect Alloca Found ===\n";
SysYPrinter::printInst(allocaInst.get());
}
it = funcInfo->getIndirectAllocas().erase(it);
} else {
++it;
}
}
}
// 该全局变量未被任何指令使用global->getUses().empty())。
void DeadCodeElimination::eliminateDeadGlobals(bool& changed) {
auto& globals = pModule->getGlobals();
for (auto it = globals.begin(); it != globals.end();) {
auto& global = *it;
if (global->getUses().empty()) {
changed = true;
if(DEBUG){
std::cout << "=== Dead Global Found ===\n";
SysYPrinter::printValue(global.get());
}
it = globals.erase(it);
} else {
++it;
}
}
}
// 消除冗余加载和存储 消除条件:
// phi 指令的目标指针仅被该 phi 使用(无其他 store/load 使用)。
// memset 指令的目标指针未被使用pointer->getUses().empty()
// store -> load -> store 模式
void DeadCodeElimination::eliminateDeadRedundantLoadStore(Function* func, bool& changed) {
for (const auto& block : func->getBasicBlocks()) {
auto& instrs = block->getInstructions();
for (auto iter = instrs.begin(); iter != instrs.end();) {
auto inst = iter->get();
if (inst->isPhi()) {
auto phiInst = dynamic_cast<PhiInst*>(inst);
auto pointer = phiInst->getPointer();
bool tag = true;
for (const auto& use : pointer->getUses()) {
auto user = use->getUser();
if (user != inst) {
tag = false;
break;
}
}
/// 如果 pointer 仅被该 phi 使用,可以删除 ph
if (tag) {
changed = true;
usedelete(inst);
iter = instrs.erase(iter);
continue;
}
// 数组指令还不完善不保证memset优化效果
} else if (inst->isMemset()) {
auto memsetInst = dynamic_cast<MemsetInst*>(inst);
auto pointer = memsetInst->getPointer();
if (pointer->getUses().empty()) {
changed = true;
usedelete(inst);
iter = instrs.erase(iter);
continue;
}
}else if(inst->isLoad()) {
if (iter != instrs.begin()) {
auto loadInst = dynamic_cast<LoadInst*>(inst);
auto loadPointer = loadInst->getPointer();
// TODO:store -> load -> store 模式
auto prevIter = std::prev(iter);
auto prevInst = prevIter->get();
if (prevInst->isStore()) {
auto prevStore = dynamic_cast<StoreInst*>(prevInst);
auto prevStorePointer = prevStore->getPointer();
auto prevStoreValue = prevStore->getOperand(0);
// 确保前一个 store 不是数组操作
if (prevStore->getIndices().empty()) {
// 检查后一条指令是否是 store 同一个值
auto nextIter = std::next(iter);
if (nextIter != instrs.end()) {
auto nextInst = nextIter->get();
if (nextInst->isStore()) {
auto nextStore = dynamic_cast<StoreInst*>(nextInst);
auto nextStorePointer = nextStore->getPointer();
auto nextStoreValue = nextStore->getOperand(0);
// 确保后一个 store 不是数组操作
if (nextStore->getIndices().empty()) {
// 判断优化条件:
// 1. prevStore 的指针操作数 == load 的指针操作数
// 2. nextStore 的值操作数 == load 指令本身
if (prevStorePointer == loadPointer &&
nextStoreValue == loadInst) {
// 可以优化直接把prevStorePointer的值存到nextStorePointer
changed = true;
nextStore->setOperand(0, prevStoreValue);
if(DEBUG){
std::cout << "=== Dead Store Load Store Found(now only del Load) ===\n";
SysYPrinter::printInst(prevStore);
SysYPrinter::printInst(loadInst);
SysYPrinter::printInst(nextStore);
}
usedelete(loadInst);
iter = instrs.erase(iter);
// 删除 prevStore 这里是不是可以留给删除无用store处理
// if (prevStore->getUses().empty()) {
// usedelete(prevStore);
// instrs.erase(prevIter); // 删除 prevStore
// }
continue; // 跳过 ++iter因为已经移动迭代器
}
}
}
}
}
}
}
}
++iter;
}
}
}
bool DeadCodeElimination::isGlobal(Value *val){
auto gval = dynamic_cast<GlobalValue *>(val);
return gval != nullptr;
}
bool DeadCodeElimination::isArr(Value *val){
auto aval = dynamic_cast<AllocaInst *>(val);
return aval != nullptr && aval->getNumDims() != 0;
}
void DeadCodeElimination::usedelete(Instruction *instr){
for (auto &use1 : instr->getOperands()) {
auto val1 = use1->getValue();
val1->removeUse(use1);
}
}
} // namespace sysy

View File

@ -102,30 +102,54 @@ void Value::replaceAllUsesWith(Value *value) {
uses.clear();
}
ConstantValue* ConstantValue::get(int value) {
static std::map<int, std::unique_ptr<ConstantValue>> intConstants;
auto iter = intConstants.find(value);
if (iter != intConstants.end()) {
return iter->second.get();
// Implementations for static members
std::unordered_map<ConstantValueKey, ConstantValue*, ConstantValueHash, ConstantValueEqual> ConstantValue::mConstantPool;
std::unordered_map<Type*, UndefinedValue*> UndefinedValue::UndefValues;
ConstantValue* ConstantValue::get(Type* type, ConstantValVariant val) {
ConstantValueKey key = {type, val};
auto it = mConstantPool.find(key);
if (it != mConstantPool.end()) {
return it->second;
}
auto inst = new ConstantValue(value);
assert(inst);
auto result = intConstants.emplace(value, inst);
return result.first->second.get();
ConstantValue* newConstant = nullptr;
if (std::holds_alternative<int>(val)) {
newConstant = new ConstantInteger(type, std::get<int>(val));
} else if (std::holds_alternative<float>(val)) {
newConstant = new ConstantFloating(type, std::get<float>(val));
} else {
assert(false && "Unsupported ConstantValVariant type");
}
mConstantPool[key] = newConstant;
return newConstant;
}
ConstantValue* ConstantValue::get(float value) {
static std::map<float, std::unique_ptr<ConstantValue>> floatConstants;
auto iter = floatConstants.find(value);
if (iter != floatConstants.end()) {
return iter->second.get();
}
auto inst = new ConstantValue(value);
assert(inst);
auto result = floatConstants.emplace(value, inst);
return result.first->second.get();
ConstantInteger* ConstantInteger::get(Type* type, int val) {
return dynamic_cast<ConstantInteger*>(ConstantValue::get(type, val));
}
ConstantFloating* ConstantFloating::get(Type* type, float val) {
return dynamic_cast<ConstantFloating*>(ConstantValue::get(type, val));
}
UndefinedValue* UndefinedValue::get(Type* type) {
assert(!type->isVoid() && "Cannot get UndefinedValue of void type!");
auto it = UndefValues.find(type);
if (it != UndefValues.end()) {
return it->second;
}
UndefinedValue* newUndef = new UndefinedValue(type);
UndefValues[type] = newUndef;
return newUndef;
}
auto Function::getCalleesWithNoExternalAndSelf() -> std::set<Function *> {
std::set<Function *> result;
for (auto callee : callees) {
@ -545,6 +569,83 @@ void User::replaceOperand(unsigned index, Value *value) {
value->addUse(use);
}
/**
* phi相关函数
*/
Value* PhiInst::getvalfromBlk(BasicBlock* blk){
refreshB2VMap();
if( blk2val.find(blk) != blk2val.end()) {
return blk2val.at(blk);
}
return nullptr;
}
BasicBlock* PhiInst::getBlkfromVal(Value* val){
// 返回第一个值对应的基本块
for(unsigned i = 0; i < vsize; i++) {
if(getValue(i) == val) {
return getBlock(i);
}
}
return nullptr;
}
void PhiInst::delValue(Value* val){
//根据value删除对应的基本块和值
unsigned i = 0;
BasicBlock* blk = getBlkfromVal(val);
for(i = 0; i < vsize; i++) {
if(getValue(i) == val) {
break;
}
}
removeOperand(2 * i + 1); // 删除blk
removeOperand(2 * i); // 删除val
vsize--;
blk2val.erase(blk); // 删除blk2val映射
}
void PhiInst::delBlk(BasicBlock* blk){
//根据Blk删除对应的基本块和值
unsigned i = 0;
Value* val = getvalfromBlk(blk);
for(i = 0; i < vsize; i++) {
if(getBlock(i) == blk) {
break;
}
}
removeOperand(2 * i + 1); // 删除blk
removeOperand(2 * i); // 删除val
vsize--;
blk2val.erase(blk); // 删除blk2val映射
}
void PhiInst::replaceBlk(BasicBlock* newBlk, unsigned k){
refreshB2VMap();
Value* val = blk2val.at(getBlock(k));
// 替换基本块
setOperand(2 * k + 1, newBlk);
// 替换blk2val映射
blk2val.erase(getBlock(k));
blk2val.emplace(newBlk, val);
}
void PhiInst::replaceold2new(BasicBlock* oldBlk, BasicBlock* newBlk){
refreshB2VMap();
Value* val = blk2val.at(oldBlk);
// 替换基本块
delBlk(oldBlk);
addIncoming(val, newBlk);
}
void PhiInst::refreshB2VMap(){
blk2val.clear();
for(unsigned i = 0; i < vsize; i++) {
blk2val.emplace(getBlock(i), getValue(i));
}
}
CallInst::CallInst(Function *callee, const std::vector<Value *> &args, BasicBlock *parent, const std::string &name)
: Instruction(kCall, callee->getReturnType(), parent, name) {
addOperand(callee);

View File

@ -1,801 +0,0 @@
#include "Mem2Reg.h"
#include <algorithm>
#include <cassert>
#include <iterator>
#include <memory>
#include <queue>
#include <stack>
#include <string>
#include <unordered_map>
#include <utility>
#include "IR.h"
#include "SysYIRAnalyser.h"
#include "SysYIRPrinter.h"
namespace sysy {
// 计算给定变量的定义块集合的迭代支配边界
// TODO优化Semi-Naive IDF
std::unordered_set<BasicBlock *> Mem2Reg::computeIterDf(const std::unordered_set<BasicBlock *> &blocks) {
std::unordered_set<BasicBlock *> workList;
std::unordered_set<BasicBlock *> ret_list;
workList.insert(blocks.begin(), blocks.end());
while (!workList.empty()) {
auto n = workList.begin();
BlockAnalysisInfo* blockInfo = controlFlowAnalysis->getBlockAnalysisInfo(*n);
auto DFs = blockInfo->getDomFrontiers();
for (auto c : DFs) {
// 如果c不在ret_list中则将其加入ret_list和workList
// 这里的c是n的支配边界
// 也就是n的支配边界中的块
// 需要注意的是,支配边界是一个集合,所以可能会有重复
if (ret_list.count(c) == 0U) {
ret_list.emplace(c);
workList.emplace(c);
}
}
workList.erase(n);
}
return ret_list;
}
/**
* 计算value2Blocks的映射包括value2AllocBlocks、value2DefBlocks以及value2UseBlocks
* 其中value2DefBlocks可用于计算迭代支配边界来插入相应变量的phi结点
* 这里的value2AllocBlocks、value2DefBlocks和value2UseBlocks改变了函数级别的分析信息
*/
auto Mem2Reg::computeValue2Blocks() -> void {
SysYPrinter printer(pModule); // 初始化打印机
// std::cout << "===== Start computeValue2Blocks =====" << std::endl;
auto &functions = pModule->getFunctions();
for (const auto &function : functions) {
auto func = function.second.get();
// std::cout << "\nProcessing function: " << func->getName() << std::endl;
FunctionAnalysisInfo* funcInfo = controlFlowAnalysis->getFunctionAnalysisInfo(func);
if (!funcInfo) {
std::cerr << "ERROR: No analysis info for function " << func->getName() << std::endl;
continue;
}
auto basicBlocks = func->getBasicBlocks();
// std::cout << "BasicBlocks count: " << basicBlocks.size() << std::endl;
for (auto &it : basicBlocks) {
auto basicBlock = it.get();
// std::cout << "\nProcessing BB: " << basicBlock->getName() << std::endl;
// printer.printBlock(basicBlock); // 打印基本块内容
auto &instrs = basicBlock->getInstructions();
for (auto &instr : instrs) {
// std::cout << " Analyzing instruction: ";
// printer.printInst(instr.get());
// std::cout << std::endl;
if (instr->isAlloca()) {
if (!(isArr(instr.get()) || isGlobal(instr.get()))) {
// std::cout << " Found alloca: ";
// printer.printInst(instr.get());
// std::cout << " -> Adding to allocBlocks" << std::endl;
funcInfo->addValue2AllocBlocks(instr.get(), basicBlock);
} else {
// std::cout << " Skip array/global alloca: ";
// printer.printInst(instr.get());
// std::cout << std::endl;
}
}
else if (instr->isStore()) {
auto val = instr->getOperand(1);
// std::cout << " Store target: ";
// printer.printInst(dynamic_cast<Instruction *>(val));
if (!(isArr(val) || isGlobal(val))) {
// std::cout << " Adding store to defBlocks for value: ";
// printer.printInst(dynamic_cast<Instruction *>(instr.get()));
// std::cout << std::endl;
// 将store的目标值添加到defBlocks中
funcInfo->addValue2DefBlocks(val, basicBlock);
} else {
// std::cout << " Skip array/global store" << std::endl;
}
}
else if (instr->isLoad()) {
auto val = instr->getOperand(0);
// std::cout << " Load source: ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// std::cout << std::endl;
if (!(isArr(val) || isGlobal(val))) {
// std::cout << " Adding load to useBlocks for value: ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// std::cout << std::endl;
funcInfo->addValue2UseBlocks(val, basicBlock);
} else {
// std::cout << " Skip array/global load" << std::endl;
}
}
}
}
// 打印分析结果
// std::cout << "\nAnalysis results for function " << func->getName() << ":" << std::endl;
// auto &allocMap = funcInfo->getValue2AllocBlocks();
// std::cout << "AllocBlocks (" << allocMap.size() << "):" << std::endl;
// for (auto &[val, bb] : allocMap) {
// std::cout << " ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// std::cout << " in BB: " << bb->getName() << std::endl;
// }
// auto &defMap = funcInfo->getValue2DefBlocks();
// std::cout << "DefBlocks (" << defMap.size() << "):" << std::endl;
// for (auto &[val, bbs] : defMap) {
// std::cout << " ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// for (const auto &[bb, count] : bbs) {
// std::cout << " in BB: " << bb->getName() << " (count: " << count << ")";
// }
// }
// auto &useMap = funcInfo->getValue2UseBlocks();
// std::cout << "UseBlocks (" << useMap.size() << "):" << std::endl;
// for (auto &[val, bbs] : useMap) {
// std::cout << " ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// for (const auto &[bb, count] : bbs) {
// std::cout << " in BB: " << bb->getName() << " (count: " << count << ")";
// }
// }
}
// std::cout << "===== End computeValue2Blocks =====" << std::endl;
}
/**
* @brief 级联关系的顺带消除用于llvm mem2reg类预优化1
*
* 采用队列进行模拟从某种程度上来看其实可以看作是UD链的反向操作
*
* @param [in] instr store指令使用的指令
* @param [in] changed 不动点法的判断标准,地址传递
* @param [in] func 指令所在函数
* @param [in] block 指令所在基本块
* @param [in] instrs 基本块所在指令集合,地址传递
* @return 无返回值,但满足条件的情况下会对指令进行删除
*/
auto Mem2Reg::cascade(Instruction *instr, bool &changed, Function *func, BasicBlock *block,
std::list<std::unique_ptr<Instruction>> &instrs) -> void {
if (instr != nullptr) {
if (instr->isUnary() || instr->isBinary() || instr->isLoad()) {
std::queue<Instruction *> toRemove;
toRemove.push(instr);
while (!toRemove.empty()) {
auto top = toRemove.front();
toRemove.pop();
auto operands = top->getOperands();
for (const auto &operand : operands) {
auto elem = dynamic_cast<Instruction *>(operand->getValue());
if (elem != nullptr) {
if ((elem->isUnary() || elem->isBinary() || elem->isLoad()) && elem->getUses().size() == 1 &&
elem->getUses().front()->getUser() == top) {
toRemove.push(elem);
} else if (elem->isAlloca()) {
// value2UseBlock中该block对应次数-1如果该变量的该useblock中count减为0了则意味着
// 该block其他地方也没用到该alloc了故从value2UseBlock中删除
FunctionAnalysisInfo* funcInfo = controlFlowAnalysis->getFunctionAnalysisInfo(func);
auto res = funcInfo->removeValue2UseBlock(elem, block);
// 只要有一次返回了true就说明有变化
if (res) {
changed = true;
}
}
}
}
auto tofind =
std::find_if(instrs.begin(), instrs.end(), [&top](const auto &instr) { return instr.get() == top; });
assert(tofind != instrs.end());
usedelete(tofind->get());
instrs.erase(tofind);
}
}
}
}
/**
* llvm mem2reg预优化1: 删除不含load的alloc和store
*
* 1. 删除不含load的alloc和store
* 2. 删除store指令之前的用于作store指令第0个操作数的那些级联指令就冗余了也要删除
* 3. 删除之后可能有些变量的load使用恰好又没有了因此再次从第一步开始循环这里使用不动点法
*
* 由于删除了级联关系,所以这里的方法有点儿激进;
* 同时也考虑了级联关系时如果调用了函数可能会有side effect所以没有删除调用函数的级联关系
* 而且关于函数参数的alloca不会在指令中删除也不会在value2Alloca中删除;
* 同样地我们不考虑数组和global不过这里的代码是基于value2blocks的在value2blocks中已经考虑了所以不用显式指明
*=
*/
auto Mem2Reg::preOptimize1() -> void {
SysYPrinter printer(pModule); // 初始化打印机
auto &functions = pModule->getFunctions();
// std::cout << "===== Start preOptimize1 =====" << std::endl;
for (const auto &function : functions) {
auto func = function.second.get();
// std::cout << "\nProcessing function: " << func->getName() << std::endl;
FunctionAnalysisInfo* funcInfo = controlFlowAnalysis->getFunctionAnalysisInfo(func);
if (!funcInfo) {
// std::cerr << "ERROR: No analysis info for function " << func->getName() << std::endl;
continue;
}
auto &vToDefB = funcInfo->getValue2DefBlocks();
auto &vToUseB = funcInfo->getValue2UseBlocks();
auto &vToAllocB = funcInfo->getValue2AllocBlocks();
// 打印初始状态
// std::cout << "Initial allocas: " << vToAllocB.size() << std::endl;
// for (auto &[val, bb] : vToAllocB) {
// std::cout << " Alloca: ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// std::cout << " in BB: " << bb->getName() << std::endl;
// }
// 阶段1删除无store的alloca
// std::cout << "\nPhase 1: Remove unused allocas" << std::endl;
for (auto iter = vToAllocB.begin(); iter != vToAllocB.end();) {
auto val = iter->first;
auto bb = iter->second;
// std::cout << "Checking alloca: ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// std::cout << " in BB: " << bb->getName() << std::endl;
// 如果该alloca没有对应的store指令且不在函数参数中
// 这里的vToDefB是value2DefBlocksvToUseB是value2UseBlocks
// 打印vToDefB
// std::cout << "DefBlocks (" << vToDefB.size() << "):" << std::endl;
// for (auto &[val, bbs] : vToDefB) {
// std::cout << " ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// for (const auto &[bb, count] : bbs) {
// std::cout << " in BB: " << bb->getName() << " (count: " << count << ")" << std::endl;
// }
// }
// std::cout << vToDefB.count(val) << std::endl;
if (vToDefB.count(val) == 0U &&
std::find(func->getEntryBlock()->getArguments().begin(),
func->getEntryBlock()->getArguments().end(),
val) == func->getEntryBlock()->getArguments().end()) {
// std::cout << " Removing unused alloca: ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// std::cout << std::endl;
auto tofind = std::find_if(bb->getInstructions().begin(),
bb->getInstructions().end(),
[val](const auto &instr) {
return instr.get() == val;
});
if (tofind == bb->getInstructions().end()) {
// std::cerr << "ERROR: Alloca not found in BB!" << std::endl;
++iter;
continue;
}
usedelete(tofind->get());
bb->getInstructions().erase(tofind);
iter = vToAllocB.erase(iter);
} else {
++iter;
}
}
// 阶段2删除无load的store
// std::cout << "\nPhase 2: Remove dead stores" << std::endl;
bool changed = true;
int iteration = 0;
while (changed) {
changed = false;
iteration++;
// std::cout << "\nIteration " << iteration << std::endl;
for (auto iter = vToDefB.begin(); iter != vToDefB.end();) {
auto val = iter->first;
// std::cout << "Checking value: ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// std::cout << std::endl;
if (vToUseB.count(val) == 0U) {
// std::cout << " Found dead store for value: ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// std::cout << std::endl;
auto blocks = funcInfo->getDefBlocksByValue(val);
for (auto block : blocks) {
// std::cout << " Processing BB: " << block->getName() << std::endl;
// printer.printBlock(block); // 打印基本块内容
auto &instrs = block->getInstructions();
for (auto it = instrs.begin(); it != instrs.end();) {
if ((*it)->isStore() && (*it)->getOperand(1) == val) {
// std::cout << " Removing store: ";
// printer.printInst(it->get());
std::cout << std::endl;
auto valUsedByStore = dynamic_cast<Instruction *>((*it)->getOperand(0));
usedelete(it->get());
if (valUsedByStore != nullptr &&
valUsedByStore->getUses().size() == 1 &&
valUsedByStore->getUses().front()->getUser() == (*it).get()) {
// std::cout << " Cascade deleting: ";
// printer.printInst(valUsedByStore);
// std::cout << std::endl;
cascade(valUsedByStore, changed, func, block, instrs);
}
it = instrs.erase(it);
changed = true;
} else {
++it;
}
}
}
// 删除对应的alloca
if (std::find(func->getEntryBlock()->getArguments().begin(),
func->getEntryBlock()->getArguments().end(),
val) == func->getEntryBlock()->getArguments().end()) {
auto bb = funcInfo->getAllocBlockByValue(val);
if (bb != nullptr) {
// std::cout << " Removing alloca: ";
// printer.printInst(dynamic_cast<Instruction *>(val));
// std::cout << " in BB: " << bb->getName() << std::endl;
funcInfo->removeValue2AllocBlock(val);
auto tofind = std::find_if(bb->getInstructions().begin(),
bb->getInstructions().end(),
[val](const auto &instr) {
return instr.get() == val;
});
if (tofind != bb->getInstructions().end()) {
usedelete(tofind->get());
bb->getInstructions().erase(tofind);
} else {
std::cerr << "ERROR: Alloca not found in BB!" << std::endl;
}
}
}
iter = vToDefB.erase(iter);
} else {
++iter;
}
}
}
}
// std::cout << "===== End preOptimize1 =====" << std::endl;
}
/**
* llvm mem2reg预优化2: 针对某个变量的Defblocks只有一个块的情况
*
* 1. 该基本块最后一次对该变量的store指令后的所有对该变量的load指令都可以替换为该基本块最后一次store指令的第0个操作数
* 2. 以该基本块为必经结点的结点集合中的对该变量的load指令都可以替换为该基本块最后一次对该变量的store指令的第0个操作数
* 3.
* 如果对该变量的所有load均替换掉了删除该基本块中最后一次store指令如果这个store指令是唯一的define那么再删除alloca指令不删除参数的alloca
* 4.
* 如果对该value的所有load都替换掉了对于该变量剩下还有store的话就转换成了preOptimize1的情况再调用preOptimize1进行删除
*
* 同样不考虑数组和全局变量因为这些变量不会被mem2reg优化在value2blocks中已经考虑了所以不用显式指明
* 替换的操作采用了UD链进行简化和效率的提升
*
*/
auto Mem2Reg::preOptimize2() -> void {
auto &functions = pModule->getFunctions();
for (const auto &function : functions) {
auto func = function.second.get();
FunctionAnalysisInfo* funcInfo = controlFlowAnalysis->getFunctionAnalysisInfo(func);
auto values = funcInfo->getValuesOfDefBlock();
for (auto val : values) {
auto blocks = funcInfo->getDefBlocksByValue(val);
// 该val只有一个defining block
if (blocks.size() == 1) {
auto block = *blocks.begin();
auto &instrs = block->getInstructions();
auto rit = std::find_if(instrs.rbegin(), instrs.rend(),
[val](const auto &instr) { return instr->isStore() && instr->getOperand(1) == val; });
// 注意reverse_iterator求base后是指向下一个指令因此要减一才是原来的指令
assert(rit != instrs.rend());
auto it = --rit.base();
auto propogationVal = (*it)->getOperand(0);
// 其实该块中it后对该val的load指令也可以替换掉了
for (auto curit = std::next(it); curit != instrs.end();) {
if ((*curit)->isLoad() && (*curit)->getOperand(0) == val) {
curit->get()->replaceAllUsesWith(propogationVal);
usedelete(curit->get());
curit = instrs.erase(curit);
funcInfo->removeValue2UseBlock(val, block);
} else {
++curit;
}
}
// 在支配树后继结点中替换load指令的操作数
BlockAnalysisInfo* blockInfo = controlFlowAnalysis->getBlockAnalysisInfo(block);
std::vector<BasicBlock *> blkchildren;
// 获取该块的支配树后继结点
std::queue<BasicBlock *> q;
auto sdoms = blockInfo->getSdoms();
for (auto sdom : sdoms) {
q.push(sdom);
blkchildren.push_back(sdom);
}
while (!q.empty()) {
auto blk = q.front();
q.pop();
BlockAnalysisInfo* blkInfo = controlFlowAnalysis->getBlockAnalysisInfo(blk);
for (auto sdom : blkInfo->getSdoms()) {
q.push(sdom);
blkchildren.push_back(sdom);
}
}
for (auto child : blkchildren) {
auto &childInstrs = child->getInstructions();
for (auto childIter = childInstrs.begin(); childIter != childInstrs.end();) {
if ((*childIter)->isLoad() && (*childIter)->getOperand(0) == val) {
childIter->get()->replaceAllUsesWith(propogationVal);
usedelete(childIter->get());
childIter = childInstrs.erase(childIter);
funcInfo->removeValue2UseBlock(val, child);
} else {
++childIter;
}
}
}
// 如果对该val的所有load均替换掉了那么对于该val的defining block中的最后一个define也可以删除了
// 同时该块中前面对于该val的define也变成死代码了可调用preOptimize1进行删除
if (funcInfo->getUseBlocksByValue(val).empty()) {
usedelete(it->get());
instrs.erase(it);
auto change = funcInfo->removeValue2DefBlock(val, block);
if (change) {
// 如果define是唯一的且不是函数参数的alloca直接删alloca
if (std::find(func->getEntryBlock()->getArguments().begin(), func->getEntryBlock()->getArguments().end(),
val) == func->getEntryBlock()->getArguments().end()) {
auto bb = funcInfo->getAllocBlockByValue(val);
assert(bb != nullptr);
auto tofind = std::find_if(bb->getInstructions().begin(), bb->getInstructions().end(),
[val](const auto &instr) { return instr.get() == val; });
usedelete(tofind->get());
bb->getInstructions().erase(tofind);
funcInfo->removeValue2AllocBlock(val);
}
} else {
// 如果该变量还有其他的define那么前面的define也变成死代码了
assert(!funcInfo->getDefBlocksByValue(val).empty());
assert(funcInfo->getUseBlocksByValue(val).empty());
preOptimize1();
}
}
}
}
}
}
/**
* @brief llvm mem2reg类预优化3针对某个变量的所有读写都在同一个块中的情况
*
* 1. 将每一个load替换成前一个store的值并删除该load
* 2. 如果在load前没有对该变量的store则不删除该load
* 3. 如果一个store后没有任何对改变量的load则删除该store
*
* @note 额外说明第二点不用显式处理因为我们的方法是从找到第一个store开始
* 第三点其实可以更激进一步地理解即每次替换了load之后它对应地那个store也可以删除了同时注意这里不要使用preoptimize1进行处理因为他们的级联关系是有用的即用来求load的替换值
* 同样地我们这里不考虑数组和全局变量因为这些变量不会被mem2reg优化不过这里在计算value2DefBlocks时已经跳过了所以不需要再显式处理了
* 替换的操作采用了UD链进行简化和效率的提升
*
* @param [in] void
* @return 无返回值,但满足条件的情况下会对指令的操作数进行替换以及对指令进行删除
*/
auto Mem2Reg::preOptimize3() -> void {
auto &functions = pModule->getFunctions();
for (const auto &function : functions) {
auto func = function.second.get();
FunctionAnalysisInfo* funcInfo = controlFlowAnalysis->getFunctionAnalysisInfo(func);
auto values = funcInfo->getValuesOfDefBlock();
for (auto val : values) {
auto sblocks = funcInfo->getDefBlocksByValue(val);
auto lblocks = funcInfo->getUseBlocksByValue(val);
if (sblocks.size() == 1 && lblocks.size() == 1 && *sblocks.begin() == *lblocks.begin()) {
auto block = *sblocks.begin();
auto &instrs = block->getInstructions();
auto it = std::find_if(instrs.begin(), instrs.end(),
[val](const auto &instr) { return instr->isStore() && instr->getOperand(1) == val; });
while (it != instrs.end()) {
auto propogationVal = (*it)->getOperand(0);
auto last = std::find_if(std::next(it), instrs.end(), [val](const auto &instr) {
return instr->isStore() && instr->getOperand(1) == val;
});
for (auto curit = std::next(it); curit != last;) {
if ((*curit)->isLoad() && (*curit)->getOperand(0) == val) {
curit->get()->replaceAllUsesWith(propogationVal);
usedelete(curit->get());
curit = instrs.erase(curit);
funcInfo->removeValue2UseBlock(val, block);
} else {
++curit;
}
}
// 替换了load之后它对应地那个store也可以删除了
if (!(std::find_if(func->getEntryBlock()->getArguments().begin(), func->getEntryBlock()->getArguments().end(),
[val](const auto &instr) { return instr == val; }) !=
func->getEntryBlock()->getArguments().end()) &&
last == instrs.end()) {
usedelete(it->get());
it = instrs.erase(it);
if (funcInfo->removeValue2DefBlock(val, block)) {
auto bb = funcInfo->getAllocBlockByValue(val);
if (bb != nullptr) {
auto tofind = std::find_if(bb->getInstructions().begin(), bb->getInstructions().end(),
[val](const auto &instr) { return instr.get() == val; });
usedelete(tofind->get());
bb->getInstructions().erase(tofind);
funcInfo->removeValue2AllocBlock(val);
}
}
}
it = last;
}
}
}
}
}
/**
* 为所有变量的定义块集合的迭代支配边界插入phi结点
*
* insertPhi是mem2reg的核心之一这里是对所有变量的迭代支配边界的phi结点插入无参数也无返回值
* 同样跳过对数组和全局变量的处理因为这些变量不会被mem2reg优化刚好这里在计算value2DefBlocks时已经跳过了所以不需要再显式处理了
* 同时我们进行了剪枝处理只有在基本块入口活跃的变量才插入phi函数
*
*/
auto Mem2Reg::insertPhi() -> void {
auto &functions = pModule->getFunctions();
for (const auto &function : functions) {
auto func = function.second.get();
FunctionAnalysisInfo* funcInfo = controlFlowAnalysis->getFunctionAnalysisInfo(func);
const auto &vToDefB = funcInfo->getValue2DefBlocks();
for (const auto &map_pair : vToDefB) {
// 首先为每个变量找到迭代支配边界
auto val = map_pair.first;
auto blocks = funcInfo->getDefBlocksByValue(val);
auto itDFs = computeIterDf(blocks);
// 然后在每个变量相应的迭代支配边界上插入phi结点
for (auto basicBlock : itDFs) {
const auto &actiTable = activeVarAnalysis->getActiveTable();
auto dval = dynamic_cast<User *>(val);
// 只有在基本块入口活跃的变量才插入phi函数
if (actiTable.at(basicBlock).front().count(dval) != 0U) {
pBuilder->createPhiInst(val->getType(), val, basicBlock);
}
}
}
}
}
/**
* 重命名
*
* 重命名是mem2reg的核心之二这里是对单个块的重命名递归实现
* 同样跳过对数组和全局变量的处理因为这些变量不会被mem2reg优化
*
*/
auto Mem2Reg::rename(BasicBlock *block, std::unordered_map<Value *, int> &count,
std::unordered_map<Value *, std::stack<Instruction *>> &stacks) -> void {
auto &instrs = block->getInstructions();
std::unordered_map<Value *, int> valPop;
// 第一大步:对块中的所有指令遍历处理
for (auto iter = instrs.begin(); iter != instrs.end();) {
auto instr = iter->get();
// 对于load指令变量用最新的那个
if (instr->isLoad()) {
auto val = instr->getOperand(0);
if (!(isArr(val) || isGlobal(val))) {
if (!stacks[val].empty()) {
instr->replaceOperand(0, stacks[val].top());
}
}
}
// 然后对于define的情况看alloca、store和phi指令
if (instr->isDefine()) {
if (instr->isAlloca()) {
// alloca指令名字不改了命名就按xx_1x_2...来就行
auto val = instr;
if (!(isArr(val) || isGlobal(val))) {
++valPop[val];
stacks[val].push(val);
++count[val];
}
} else if (instr->isPhi()) {
// Phi指令也是一条特殊的define指令
auto val = dynamic_cast<PhiInst *>(instr)->getMapVal();
if (!(isArr(val) || isGlobal(val))) {
auto i = count[val];
if (i == 0) {
// 对还未alloca就有phi的指令的处理直接删除
usedelete(iter->get());
iter = instrs.erase(iter);
continue;
}
auto newname = dynamic_cast<Instruction *>(val)->getName() + "_" + std::to_string(i);
auto newalloca = pBuilder->createAllocaInstWithoutInsert(val->getType(), {}, block, newname);
FunctionAnalysisInfo* ParentfuncInfo = controlFlowAnalysis->getFunctionAnalysisInfo(block->getParent());
ParentfuncInfo->addIndirectAlloca(newalloca);
instr->replaceOperand(0, newalloca);
++valPop[val];
stacks[val].push(newalloca);
++count[val];
}
} else {
// store指令看operand的名字我们的实现是规定变量在operand的第二位用一个新的alloca x_i代替
auto val = instr->getOperand(1);
if (!(isArr(val) || isGlobal(val))) {
auto i = count[val];
auto newname = dynamic_cast<Instruction *>(val)->getName() + "_" + std::to_string(i);
auto newalloca = pBuilder->createAllocaInstWithoutInsert(val->getType(), {}, block, newname);
FunctionAnalysisInfo* ParentfuncInfo = controlFlowAnalysis->getFunctionAnalysisInfo(block->getParent());
ParentfuncInfo->addIndirectAlloca(newalloca);
// block->getParent()->addIndirectAlloca(newalloca);
instr->replaceOperand(1, newalloca);
++valPop[val];
stacks[val].push(newalloca);
++count[val];
}
}
}
++iter;
}
// 第二大步把所有CFG中的该块的successor的phi指令的相应operand确定
for (auto succ : block->getSuccessors()) {
auto position = getPredIndex(block, succ);
for (auto &instr : succ->getInstructions()) {
if (instr->isPhi()) {
auto val = dynamic_cast<PhiInst *>(instr.get())->getMapVal();
if (!stacks[val].empty()) {
instr->replaceOperand(position + 1, stacks[val].top());
}
} else {
// phi指令是添加在块的最前面的因此过了之后就不会有phi了直接break
break;
}
}
}
// 第三大步递归支配树的后继支配树才能表示define-use关系
BlockAnalysisInfo* blockInfo = controlFlowAnalysis->getBlockAnalysisInfo(block);
for (auto sdom : blockInfo->getSdoms()) {
rename(sdom, count, stacks);
}
// 第四大步遍历块中的所有指令如果涉及到define就弹栈这一步是必要的可以从递归的整体性来思考原因
// 注意这里count没清理因为平级之间计数仍然是一直增加的但是stack要清理因为define-use关系来自直接
// 支配结点而不是平级之间,不清理栈会被污染
// 提前优化知道变量对应的要弹栈的次数就可以了没必要遍历所有instr.
for (auto val_pair : valPop) {
auto val = val_pair.first;
for (int i = 0; i < val_pair.second; ++i) {
stacks[val].pop();
}
}
}
/**
* 重命名所有块
*
* 调用rename自上而下实现所有rename
*
*/
auto Mem2Reg::renameAll() -> void {
auto &functions = pModule->getFunctions();
for (const auto &function : functions) {
auto func = function.second.get();
// 对于每个function都要SSA化所以count和stacks定义在这并初始化
std::unordered_map<Value *, int> count;
std::unordered_map<Value *, std::stack<Instruction *>> stacks;
FunctionAnalysisInfo* funcInfo = controlFlowAnalysis->getFunctionAnalysisInfo(func);
for (const auto &map_pair : funcInfo->getValue2DefBlocks()) {
auto val = map_pair.first;
count[val] = 0;
}
rename(func->getEntryBlock(), count, stacks);
}
}
/**
* mem2reg对外的接口
*
* 静态单一赋值 + mem2reg等pass的逻辑组合
*
*/
auto Mem2Reg::mem2regPipeline() -> void {
// 首先进行mem2reg的前置分析
controlFlowAnalysis->clear();
controlFlowAnalysis->runControlFlowAnalysis();
// 活跃变量分析
activeVarAnalysis->clear();
dataFlowAnalysisUtils.addBackwardAnalyzer(activeVarAnalysis);
dataFlowAnalysisUtils.backwardAnalyze(pModule);
// 计算所有valueToBlocks的定义映射
computeValue2Blocks();
// SysYPrinter printer(pModule);
// 参考llvm的mem2reg遍在插入phi结点之前先做些优化
preOptimize1();
// printer.printIR();
preOptimize2();
// printer.printIR();
// 优化三 可能会针对局部变量优化而删除整个块的alloca/store
preOptimize3();
//再进行活跃变量分析
// 报错?
// printer.printIR();
dataFlowAnalysisUtils.backwardAnalyze(pModule);
// 为所有变量插入phi结点
insertPhi();
// 重命名
renameAll();
}
/**
* 计算块n是块s的第几个前驱
*
* helperfunction没有返回值但是会将dom和other的交集赋值给dom
*
*/
auto Mem2Reg::getPredIndex(BasicBlock *n, BasicBlock *s) -> int {
int index = 0;
for (auto elem : s->getPredecessors()) {
if (elem == n) {
break;
}
++index;
}
assert(index < static_cast<int>(s->getPredecessors().size()) && "n is not a predecessor of s.");
return index;
}
/**
* 判断一个value是不是全局变量
*/
auto Mem2Reg::isGlobal(Value *val) -> bool {
auto gval = dynamic_cast<GlobalValue *>(val);
return gval != nullptr;
}
/**
* 判断一个value是不是数组
*/
auto Mem2Reg::isArr(Value *val) -> bool {
auto aval = dynamic_cast<AllocaInst *>(val);
return aval != nullptr && aval->getNumDims() != 0;
}
/**
* 删除一个指令的operand对应的value的该条use
*/
auto Mem2Reg::usedelete(Instruction *instr) -> void {
for (auto &use : instr->getOperands()) {
auto val = use->getValue();
val->removeUse(use);
}
}
} // namespace sysy

View File

@ -1,129 +0,0 @@
#include "Reg2Mem.h"
#include <cstddef>
#include <iostream>
#include <list>
#include <memory>
namespace sysy {
/**
* 删除phi节点
* 删除phi节点后可能会生成冗余存储代码
*/
void Reg2Mem::DeletePhiInst(){
auto &functions = pModule->getFunctions();
for (auto &function : functions) {
auto basicBlocks = function.second->getBasicBlocks();
for (auto &basicBlock : basicBlocks) {
for (auto iter = basicBlock->begin(); iter != basicBlock->end();) {
auto &instruction = *iter;
if (instruction->isPhi()) {
auto predBlocks = basicBlock->getPredecessors();
// 寻找源和目的
// 目的就是phi指令的第一个操作数
// 源就是phi指令的后续操作数
auto destination = instruction->getOperand(0);
int predBlockindex = 0;
for (auto &predBlock : predBlocks) {
++predBlockindex;
// 判断前驱块儿只有一个后继还是多个后继
// 如果有多个
auto source = instruction->getOperand(predBlockindex);
if (source == destination) {
continue;
}
// std::cout << predBlock->getNumSuccessors() << std::endl;
if (predBlock->getNumSuccessors() > 1) {
// 创建一个basicblock
auto newbasicBlock = function.second->addBasicBlock();
std::stringstream ss;
ss << " phidel.L" << pBuilder->getLabelIndex();
newbasicBlock->setName(ss.str());
ss.str("");
// // 修改前驱后继关系
basicBlock->replacePredecessor(predBlock, newbasicBlock);
// predBlock = newbasicBlock;
newbasicBlock->addPredecessor(predBlock);
newbasicBlock->addSuccessor(basicBlock.get());
predBlock->removeSuccessor(basicBlock.get());
predBlock->addSuccessor(newbasicBlock);
// std::cout << "the block name is " << basicBlock->getName() << std::endl;
// for (auto pb : basicBlock->getPredecessors()) {
// // newbasicBlock->addPredecessor(pb);
// std::cout << pb->getName() << std::endl;
// }
// sysy::BasicBlock::conectBlocks(newbasicBlock, static_cast<BasicBlock *>(basicBlock.get()));
// 若后为跳转指令,应该修改跳转指令所到达的位置
auto thelastinst = predBlock->end();
(--thelastinst);
if (thelastinst->get()->isConditional() || thelastinst->get()->isUnconditional()) { // 如果是跳转指令
auto opnum = thelastinst->get()->getNumOperands();
for (size_t i = 0; i < opnum; i++) {
if (thelastinst->get()->getOperand(i) == basicBlock.get()) {
thelastinst->get()->replaceOperand(i, newbasicBlock);
}
}
}
// 在新块中插入store指令
pBuilder->setPosition(newbasicBlock, newbasicBlock->end());
// pBuilder->createStoreInst(source, destination);
if (source->isInt() || source->isFloat()) {
pBuilder->createStoreInst(source, destination);
} else {
auto loadInst = pBuilder->createLoadInst(source);
pBuilder->createStoreInst(loadInst, destination);
}
// pBuilder->createMoveInst(Instruction::kMove, destination->getType(), destination, source,
// newbasicBlock);
pBuilder->setPosition(newbasicBlock, newbasicBlock->end());
pBuilder->createUncondBrInst(basicBlock.get(), {});
} else {
// 如果前驱块只有一个后继
auto thelastinst = predBlock->end();
(--thelastinst);
// std::cout << predBlock->getName() << std::endl;
// std::cout << thelastinst->get() << std::endl;
// std::cout << "First point 11 " << std::endl;
if (thelastinst->get()->isConditional() || thelastinst->get()->isUnconditional()) {
// 在跳转语句前insert st指令
pBuilder->setPosition(predBlock, thelastinst);
} else {
pBuilder->setPosition(predBlock, predBlock->end());
}
if (source->isInt() || source->isFloat()) {
pBuilder->createStoreInst(source, destination);
} else {
auto loadInst = pBuilder->createLoadInst(source);
pBuilder->createStoreInst(loadInst, destination);
}
}
}
// 删除phi指令
auto &instructions = basicBlock->getInstructions();
usedelete(iter->get());
iter = instructions.erase(iter);
if (basicBlock->getNumInstructions() == 0) {
if (basicBlock->getNumSuccessors() == 1) {
pBuilder->setPosition(basicBlock.get(), basicBlock->end());
pBuilder->createUncondBrInst(basicBlock->getSuccessors()[0], {});
}
}
} else {
break;
}
}
}
}
}
void Reg2Mem::usedelete(Instruction *instr) {
for (auto &use : instr->getOperands()) {
auto val = use->getValue();
val->removeUse(use);
}
}
} // namespace sysy

View File

@ -10,10 +10,9 @@
#include <sstream>
#include <string>
#include <vector>
using namespace std;
#include "SysYIRGenerator.h"
using namespace std;
namespace sysy {
/*
@ -130,30 +129,111 @@ std::any SysYIRGenerator::visitVarDecl(SysYParser::VarDeclContext *ctx) {
delete root;
if (dims.empty()) {
builder.createStoreInst(values.getValue(0), alloca);
} else {
// 对于多维数组使用memset初始化
// 计算每个维度的大小
// 这里的values.getNumbers()返回的是每个维度的大小
// 这里的values.getValues()返回的是每个维度对应的值
// 例如对于一个二维数组values.getNumbers()可能是[3, 4]表示3行4列
// values.getValues()可能是[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]
// 对于每个维度使用memset将对应的值填充到数组中
// 这里的alloca是一个指向数组的指针
const std::vector<unsigned int> & counterNumbers = values.getNumbers();
const std::vector<sysy::Value *> & counterValues = values.getValues();
unsigned begin = 0;
for (size_t i = 0; i < counterNumbers.size(); i++) {
} else{
// **数组变量初始化**
const std::vector<sysy::Value *> &counterValues = values.getValues();
// 计算数组的**总元素数量**和**总字节大小**
int numElements = 1;
// 存储每个维度的实际整数大小,用于索引计算
std::vector<int> dimSizes;
for (Value *dimVal : dims) {
if (ConstantInteger *constInt = dynamic_cast<ConstantInteger *>(dimVal)) {
int dimSize = constInt->getInt();
numElements *= dimSize;
dimSizes.push_back(dimSize);
}
// TODO else 错误处理:数组维度必须是常量(对于静态分配)
}
unsigned int elementSizeInBytes = type->getSize(); // 获取单个元素的大小(字节)
unsigned int totalSizeInBytes = numElements * elementSizeInBytes;
// **判断是否可以进行全零初始化优化**
bool allValuesAreZero = false;
if (counterValues.empty()) { // 例如 int arr[3] = {}; 或 int arr[3][4] = {};
allValuesAreZero = true;
}
else {
allValuesAreZero = true;
for (Value *val : counterValues){
if (ConstantInteger *constInt = dynamic_cast<ConstantInteger *>(val)){
if (constInt->getInt() != 0){
allValuesAreZero = false;
break;
}
}
else{
// 如果值不是常量,我们通常不能确定它是否为零,所以不进行 memset 优化
allValuesAreZero = false;
break;
}
}
}
if (allValuesAreZero) {
// 如果所有初始化值都是零(或没有明确初始化但语法允许),使用 memset 优化
builder.createMemsetInst(
alloca, ConstantValue::get(static_cast<int>(begin)),
ConstantValue::get(static_cast<int>(counterNumbers[i])),
counterValues[i]);
begin += counterNumbers[i];
alloca, // 目标数组的起始地址
ConstantInteger::get(0), // 偏移量通常为0后续删除
ConstantInteger::get(totalSizeInBytes),
ConstantInteger::get(0)); // 填充的总字节数
}
else {
// **逐元素存储:遍历所有初始值,并为每个值生成一个 store 指令**
for (size_t k = 0; k < counterValues.size(); ++k) {
// 用于存储当前元素的索引列表
std::vector<Value *> currentIndices;
int tempLinearIndex = k; // 临时线性索引,用于计算多维索引
// **将线性索引转换为多维索引**
// 这个循环从最内层维度开始倒推,计算每个维度的索引
// 假设是行主序row-major order这是 C/C++ 数组的标准存储方式
for (int dimIdx = dimSizes.size() - 1; dimIdx >= 0; --dimIdx)
{
// 计算当前维度的索引,并插入到列表的最前面
currentIndices.insert(currentIndices.begin(),
ConstantInteger::get(static_cast<int>(tempLinearIndex % dimSizes[dimIdx])));
// 更新线性索引,用于计算下一个更高维度的索引
tempLinearIndex /= dimSizes[dimIdx];
}
// **生成 store 指令,传入值、基指针和计算出的索引列表**
// 你的 builder.createStoreInst 签名需要能够接受这些参数
// 假设你的 builder.createStoreInst(Value *val, Value *ptr, const std::vector<Value *> &indices, ...)
builder.createStoreInst(counterValues[k], alloca, currentIndices);
}
}
}
}
else
{ // **如果没有显式初始化值,默认对数组进行零初始化**
if (!dims.empty())
{ // 只有数组才需要默认的零初始化
int numElements = 1;
for (Value *dimVal : dims)
{
if (ConstantInteger *constInt = dynamic_cast<ConstantInteger *>(dimVal))
{
numElements *= constInt->getInt();
}
}
unsigned int elementSizeInBytes = type->getSize();
unsigned int totalSizeInBytes = numElements * elementSizeInBytes;
// 使用 memset 将整个数组清零
builder.createMemsetInst(
alloca,
ConstantInteger::get(0),
ConstantInteger::get(totalSizeInBytes),
ConstantInteger::get(0)
); // 填充的总字节数
}
// 标量变量如果没有初始化值,通常不生成额外的初始化指令,因为其内存已分配但未赋值。
}
module->addVariable(name, alloca);
}
return std::any();
}
@ -218,7 +298,7 @@ std::any SysYIRGenerator::visitFuncDef(SysYParser::FuncDefContext *ctx){
paramNames.push_back(param->Ident()->getText());
std::vector<Value *> dims = {};
if (!param->LBRACK().empty()) {
dims.push_back(ConstantValue::get(-1)); // 第一个维度不确定
dims.push_back(ConstantInteger::get(-1)); // 第一个维度不确定
for (const auto &exp : param->exp()) {
dims.push_back(std::any_cast<Value *>(visitExp(exp)));
}
@ -247,9 +327,9 @@ std::any SysYIRGenerator::visitFuncDef(SysYParser::FuncDefContext *ctx){
if(HasReturnInst == false) {
// 如果没有return语句则默认返回0
if (returnType != Type::getVoidType()) {
Value* returnValue = ConstantValue::get(0);
Value* returnValue = ConstantInteger::get(0);
if (returnType == Type::getFloatType()) {
returnValue = ConstantValue::get(0.0f);
returnValue = ConstantFloating::get(0.0f);
}
builder.createReturnInst(returnValue);
} else {
@ -286,9 +366,9 @@ std::any SysYIRGenerator::visitAssignStmt(SysYParser::AssignStmtContext *ctx) {
ConstantValue * constValue = dynamic_cast<ConstantValue *>(value);
if (constValue != nullptr) {
if (variableType == Type::getFloatType()) {
value = ConstantValue::get(static_cast<float>(constValue->getInt()));
value = ConstantInteger::get(static_cast<float>(constValue->getInt()));
} else {
value = ConstantValue::get(static_cast<int>(constValue->getFloat()));
value = ConstantFloating::get(static_cast<int>(constValue->getFloat()));
}
} else {
if (variableType == Type::getFloatType()) {
@ -478,9 +558,9 @@ std::any SysYIRGenerator::visitReturnStmt(SysYParser::ReturnStmtContext *ctx) {
ConstantValue * constValue = dynamic_cast<ConstantValue *>(returnValue);
if (constValue != nullptr) {
if (funcType == Type::getFloatType()) {
returnValue = ConstantValue::get(static_cast<float>(constValue->getInt()));
returnValue = ConstantInteger::get(static_cast<float>(constValue->getInt()));
} else {
returnValue = ConstantValue::get(static_cast<int>(constValue->getFloat()));
returnValue = ConstantFloating::get(static_cast<int>(constValue->getFloat()));
}
} else {
if (funcType == Type::getFloatType()) {
@ -560,10 +640,10 @@ std::any SysYIRGenerator::visitPrimaryExp(SysYParser::PrimaryExpContext *ctx) {
std::any SysYIRGenerator::visitNumber(SysYParser::NumberContext *ctx) {
if (ctx->ILITERAL() != nullptr) {
int value = std::stol(ctx->ILITERAL()->getText(), nullptr, 0);
return static_cast<Value *>(ConstantValue::get(value));
return static_cast<Value *>(ConstantInteger::get(value));
} else if (ctx->FLITERAL() != nullptr) {
float value = std::stof(ctx->FLITERAL()->getText());
return static_cast<Value *>(ConstantValue::get(value));
return static_cast<Value *>(ConstantFloating::get(value));
}
throw std::runtime_error("Unknown number type.");
return std::any(); // 不会到达这里
@ -599,9 +679,9 @@ std::any SysYIRGenerator::visitCall(SysYParser::CallContext *ctx) {
ConstantValue * constValue = dynamic_cast<ConstantValue *>(args[i]);
if (constValue != nullptr) {
if (params[i]->getType() == Type::getPointerType(Type::getFloatType())) {
args[i] = ConstantValue::get(static_cast<float>(constValue->getInt()));
args[i] = ConstantInteger::get(static_cast<float>(constValue->getInt()));
} else {
args[i] = ConstantValue::get(static_cast<int>(constValue->getFloat()));
args[i] = ConstantFloating::get(static_cast<int>(constValue->getFloat()));
}
} else {
if (params[i]->getType() == Type::getPointerType(Type::getFloatType())) {
@ -629,9 +709,9 @@ std::any SysYIRGenerator::visitUnaryExp(SysYParser::UnaryExpContext *ctx) {
ConstantValue * constValue = dynamic_cast<ConstantValue *>(value);
if (constValue != nullptr) {
if (constValue->isFloat()) {
result = ConstantValue::get(-constValue->getFloat());
result = ConstantFloating::get(-constValue->getFloat());
} else {
result = ConstantValue::get(-constValue->getInt());
result = ConstantInteger::get(-constValue->getInt());
}
} else if (value != nullptr) {
if (value->getType() == Type::getIntType()) {
@ -648,9 +728,9 @@ std::any SysYIRGenerator::visitUnaryExp(SysYParser::UnaryExpContext *ctx) {
if (constValue != nullptr) {
if (constValue->isFloat()) {
result =
ConstantValue::get(1 - (constValue->getFloat() != 0.0F ? 1 : 0));
ConstantFloating::get(1 - (constValue->getFloat() != 0.0F ? 1 : 0));
} else {
result = ConstantValue::get(1 - (constValue->getInt() != 0 ? 1 : 0));
result = ConstantInteger::get(1 - (constValue->getInt() != 0 ? 1 : 0));
}
} else if (value != nullptr) {
if (value->getType() == Type::getIntType()) {
@ -692,13 +772,13 @@ std::any SysYIRGenerator::visitMulExp(SysYParser::MulExpContext *ctx) {
if (operandType != floatType) {
ConstantValue * constValue = dynamic_cast<ConstantValue *>(operand);
if (constValue != nullptr)
operand = ConstantValue::get(static_cast<float>(constValue->getInt()));
operand = ConstantFloating::get(static_cast<float>(constValue->getInt()));
else
operand = builder.createIToFInst(operand);
} else if (resultType != floatType) {
ConstantValue* constResult = dynamic_cast<ConstantValue *>(result);
if (constResult != nullptr)
result = ConstantValue::get(static_cast<float>(constResult->getInt()));
result = ConstantFloating::get(static_cast<float>(constResult->getInt()));
else
result = builder.createIToFInst(result);
}
@ -707,14 +787,14 @@ std::any SysYIRGenerator::visitMulExp(SysYParser::MulExpContext *ctx) {
ConstantValue* constOperand = dynamic_cast<ConstantValue *>(operand);
if (opType == SysYParser::MUL) {
if ((constOperand != nullptr) && (constResult != nullptr)) {
result = ConstantValue::get(constResult->getFloat() *
result = ConstantFloating::get(constResult->getFloat() *
constOperand->getFloat());
} else {
result = builder.createFMulInst(result, operand);
}
} else if (opType == SysYParser::DIV) {
if ((constOperand != nullptr) && (constResult != nullptr)) {
result = ConstantValue::get(constResult->getFloat() /
result = ConstantFloating::get(constResult->getFloat() /
constOperand->getFloat());
} else {
result = builder.createFDivInst(result, operand);
@ -729,17 +809,17 @@ std::any SysYIRGenerator::visitMulExp(SysYParser::MulExpContext *ctx) {
ConstantValue * constOperand = dynamic_cast<ConstantValue *>(operand);
if (opType == SysYParser::MUL) {
if ((constOperand != nullptr) && (constResult != nullptr))
result = ConstantValue::get(constResult->getInt() * constOperand->getInt());
result = ConstantInteger::get(constResult->getInt() * constOperand->getInt());
else
result = builder.createMulInst(result, operand);
} else if (opType == SysYParser::DIV) {
if ((constOperand != nullptr) && (constResult != nullptr))
result = ConstantValue::get(constResult->getInt() / constOperand->getInt());
result = ConstantInteger::get(constResult->getInt() / constOperand->getInt());
else
result = builder.createDivInst(result, operand);
} else {
if ((constOperand != nullptr) && (constResult != nullptr))
result = ConstantValue::get(constResult->getInt() % constOperand->getInt());
result = ConstantInteger::get(constResult->getInt() % constOperand->getInt());
else
result = builder.createRemInst(result, operand);
}
@ -767,13 +847,13 @@ std::any SysYIRGenerator::visitAddExp(SysYParser::AddExpContext *ctx) {
if (operandType != floatType) {
ConstantValue * constOperand = dynamic_cast<ConstantValue *>(operand);
if (constOperand != nullptr)
operand = ConstantValue::get(static_cast<float>(constOperand->getInt()));
operand = ConstantFloating::get(static_cast<float>(constOperand->getInt()));
else
operand = builder.createIToFInst(operand);
} else if (resultType != floatType) {
ConstantValue * constResult = dynamic_cast<ConstantValue *>(result);
if (constResult != nullptr)
result = ConstantValue::get(static_cast<float>(constResult->getInt()));
result = ConstantFloating::get(static_cast<float>(constResult->getInt()));
else
result = builder.createIToFInst(result);
}
@ -782,12 +862,12 @@ std::any SysYIRGenerator::visitAddExp(SysYParser::AddExpContext *ctx) {
ConstantValue * constOperand = dynamic_cast<ConstantValue *>(operand);
if (opType == SysYParser::ADD) {
if ((constResult != nullptr) && (constOperand != nullptr))
result = ConstantValue::get(constResult->getFloat() + constOperand->getFloat());
result = ConstantFloating::get(constResult->getFloat() + constOperand->getFloat());
else
result = builder.createFAddInst(result, operand);
} else {
if ((constResult != nullptr) && (constOperand != nullptr))
result = ConstantValue::get(constResult->getFloat() - constOperand->getFloat());
result = ConstantFloating::get(constResult->getFloat() - constOperand->getFloat());
else
result = builder.createFSubInst(result, operand);
}
@ -796,12 +876,12 @@ std::any SysYIRGenerator::visitAddExp(SysYParser::AddExpContext *ctx) {
ConstantValue * constOperand = dynamic_cast<ConstantValue *>(operand);
if (opType == SysYParser::ADD) {
if ((constResult != nullptr) && (constOperand != nullptr))
result = ConstantValue::get(constResult->getInt() + constOperand->getInt());
result = ConstantInteger::get(constResult->getInt() + constOperand->getInt());
else
result = builder.createAddInst(result, operand);
} else {
if ((constResult != nullptr) && (constOperand != nullptr))
result = ConstantValue::get(constResult->getInt() - constOperand->getInt());
result = ConstantInteger::get(constResult->getInt() - constOperand->getInt());
else
result = builder.createSubInst(result, operand);
}
@ -833,10 +913,10 @@ std::any SysYIRGenerator::visitRelExp(SysYParser::RelExpContext *ctx) {
auto operand2 = constOperand->isFloat() ? constOperand->getFloat()
: constOperand->getInt();
if (opType == SysYParser::LT) result = ConstantValue::get(operand1 < operand2 ? 1 : 0);
else if (opType == SysYParser::GT) result = ConstantValue::get(operand1 > operand2 ? 1 : 0);
else if (opType == SysYParser::LE) result = ConstantValue::get(operand1 <= operand2 ? 1 : 0);
else if (opType == SysYParser::GE) result = ConstantValue::get(operand1 >= operand2 ? 1 : 0);
if (opType == SysYParser::LT) result = ConstantInteger::get(operand1 < operand2 ? 1 : 0);
else if (opType == SysYParser::GT) result = ConstantInteger::get(operand1 > operand2 ? 1 : 0);
else if (opType == SysYParser::LE) result = ConstantInteger::get(operand1 <= operand2 ? 1 : 0);
else if (opType == SysYParser::GE) result = ConstantInteger::get(operand1 >= operand2 ? 1 : 0);
else assert(false);
} else {
@ -848,14 +928,14 @@ std::any SysYIRGenerator::visitRelExp(SysYParser::RelExpContext *ctx) {
if (resultType == floatType || operandType == floatType) {
if (resultType != floatType) {
if (constResult != nullptr)
result = ConstantValue::get(static_cast<float>(constResult->getInt()));
result = ConstantFloating::get(static_cast<float>(constResult->getInt()));
else
result = builder.createIToFInst(result);
}
if (operandType != floatType) {
if (constOperand != nullptr)
operand = ConstantValue::get(static_cast<float>(constOperand->getInt()));
operand = ConstantFloating::get(static_cast<float>(constOperand->getInt()));
else
operand = builder.createIToFInst(operand);
@ -901,8 +981,8 @@ std::any SysYIRGenerator::visitEqExp(SysYParser::EqExpContext *ctx) {
auto operand2 = constOperand->isFloat() ? constOperand->getFloat()
: constOperand->getInt();
if (opType == SysYParser::EQ) result = ConstantValue::get(operand1 == operand2 ? 1 : 0);
else if (opType == SysYParser::NE) result = ConstantValue::get(operand1 != operand2 ? 1 : 0);
if (opType == SysYParser::EQ) result = ConstantInteger::get(operand1 == operand2 ? 1 : 0);
else if (opType == SysYParser::NE) result = ConstantInteger::get(operand1 != operand2 ? 1 : 0);
else assert(false);
} else {
@ -913,13 +993,13 @@ std::any SysYIRGenerator::visitEqExp(SysYParser::EqExpContext *ctx) {
if (resultType == floatType || operandType == floatType) {
if (resultType != floatType) {
if (constResult != nullptr)
result = ConstantValue::get(static_cast<float>(constResult->getInt()));
result = ConstantFloating::get(static_cast<float>(constResult->getInt()));
else
result = builder.createIToFInst(result);
}
if (operandType != floatType) {
if (constOperand != nullptr)
operand = ConstantValue::get(static_cast<float>(constOperand->getInt()));
operand = ConstantFloating::get(static_cast<float>(constOperand->getInt()));
else
operand = builder.createIToFInst(operand);
}
@ -943,9 +1023,9 @@ std::any SysYIRGenerator::visitEqExp(SysYParser::EqExpContext *ctx) {
// 如果只有一个关系表达式则将结果转换为0或1
if (constResult != nullptr) {
if (constResult->isFloat())
result = ConstantValue::get(constResult->getFloat() != 0.0F ? 1 : 0);
result = ConstantInteger::get(constResult->getFloat() != 0.0F ? 1 : 0);
else
result = ConstantValue::get(constResult->getInt() != 0 ? 1 : 0);
result = ConstantInteger::get(constResult->getInt() != 0 ? 1 : 0);
}
}
@ -1013,6 +1093,7 @@ void Utils::tree2Array(Type *type, ArrayValueTree *root,
ValueCounter &result, IRBuilder *builder) {
Value* value = root->getValue();
auto &children = root->getChildren();
// 类型转换
if (value != nullptr) {
if (type == value->getType()) {
result.push_back(value);
@ -1020,14 +1101,14 @@ void Utils::tree2Array(Type *type, ArrayValueTree *root,
if (type == Type::getFloatType()) {
ConstantValue* constValue = dynamic_cast<ConstantValue *>(value);
if (constValue != nullptr)
result.push_back(ConstantValue::get(static_cast<float>(constValue->getInt())));
result.push_back(ConstantFloating::get(static_cast<float>(constValue->getInt())));
else
result.push_back(builder->createIToFInst(value));
} else {
ConstantValue* constValue = dynamic_cast<ConstantValue *>(value);
if (constValue != nullptr)
result.push_back(ConstantValue::get(static_cast<int>(constValue->getFloat())));
result.push_back(ConstantInteger::get(static_cast<int>(constValue->getFloat())));
else
result.push_back(builder->createFtoIInst(value));
@ -1061,9 +1142,9 @@ void Utils::tree2Array(Type *type, ArrayValueTree *root,
int num = blockSize - afterSize + beforeSize;
if (num > 0) {
if (type == Type::getFloatType())
result.push_back(ConstantValue::get(0.0F), num);
result.push_back(ConstantFloating::get(0.0F), num);
else
result.push_back(ConstantValue::get(0), num);
result.push_back(ConstantInteger::get(0), num);
}
}
@ -1101,7 +1182,7 @@ void Utils::initExternalFunction(Module *pModule, IRBuilder *pBuilder) {
funcName, pModule, pBuilder);
paramTypes.push_back(Type::getIntType());
paramNames.emplace_back("x");
paramDims.push_back(std::vector<Value *>{ConstantValue::get(-1)});
paramDims.push_back(std::vector<Value *>{ConstantInteger::get(-1)});
funcName = "getarray";
Utils::createExternalFunction(paramTypes, paramNames, paramDims, returnType,
funcName, pModule, pBuilder);
@ -1117,7 +1198,7 @@ void Utils::initExternalFunction(Module *pModule, IRBuilder *pBuilder) {
returnType = Type::getIntType();
paramTypes.push_back(Type::getFloatType());
paramNames.emplace_back("x");
paramDims.push_back(std::vector<Value *>{ConstantValue::get(-1)});
paramDims.push_back(std::vector<Value *>{ConstantInteger::get(-1)});
funcName = "getfarray";
Utils::createExternalFunction(paramTypes, paramNames, paramDims, returnType,
funcName, pModule, pBuilder);
@ -1141,7 +1222,7 @@ void Utils::initExternalFunction(Module *pModule, IRBuilder *pBuilder) {
paramTypes.push_back(Type::getIntType());
paramDims.clear();
paramDims.emplace_back();
paramDims.push_back(std::vector<Value *>{ConstantValue::get(-1)});
paramDims.push_back(std::vector<Value *>{ConstantInteger::get(-1)});
paramNames.clear();
paramNames.emplace_back("n");
paramNames.emplace_back("a");
@ -1164,7 +1245,7 @@ void Utils::initExternalFunction(Module *pModule, IRBuilder *pBuilder) {
paramTypes.push_back(Type::getFloatType());
paramDims.clear();
paramDims.emplace_back();
paramDims.push_back(std::vector<Value *>{ConstantValue::get(-1)});
paramDims.push_back(std::vector<Value *>{ConstantInteger::get(-1)});
paramNames.clear();
paramNames.emplace_back("n");
paramNames.emplace_back("a");

View File

@ -469,9 +469,9 @@ void SysYOptPre::SysYAddReturn() {
pBuilder->setPosition(block.get(), block->end());
// TODO: 如果int float函数缺少返回值是否需要报错
if (func->getReturnType()->isInt()) {
pBuilder->createReturnInst(ConstantValue::get(0));
pBuilder->createReturnInst(ConstantInteger::get(0));
} else if (func->getReturnType()->isFloat()) {
pBuilder->createReturnInst(ConstantValue::get(0.0F));
pBuilder->createReturnInst(ConstantFloating::get(0.0F));
} else {
pBuilder->createReturnInst();
}

View File

@ -268,6 +268,51 @@ class ValueCounter {
} ///< 清空ValueCounter
};
// --- Refactored ConstantValue and related classes start here ---
using ConstantValVariant = std::variant<int, float>;
// Helper for hashing std::variant
struct VariantHash {
template <class T>
std::size_t operator()(const T& val) const {
return std::hash<T>{}(val);
}
std::size_t operator()(const ConstantValVariant& v) const {
return std::visit(*this, v);
}
};
struct ConstantValueKey {
Type* type;
ConstantValVariant val;
bool operator==(const ConstantValueKey& other) const {
// Assuming Type objects are canonicalized, or add Type::isSame()
// If Type::isSame() is not available and Type objects are not canonicalized,
// this comparison might not be robust enough for structural equivalence of types.
return type == other.type && val == other.val;
}
};
struct ConstantValueHash {
std::size_t operator()(const ConstantValueKey& key) const {
std::size_t typeHash = std::hash<Type*>{}(key.type);
std::size_t valHash = VariantHash{}(key.val);
// A simple way to combine hashes
return typeHash ^ (valHash << 1);
}
};
struct ConstantValueEqual {
bool operator()(const ConstantValueKey& lhs, const ConstantValueKey& rhs) const {
// Assuming Type objects are canonicalized (e.g., Type::getIntType() always returns same pointer)
// If not, and Type::isSame() is intended, it should be added to Type class.
return lhs.type == rhs.type && lhs.val == rhs.val;
}
};
/*!
* Static constants known at compile time.
*
@ -276,45 +321,135 @@ class ValueCounter {
* `ConstantValue`并不由指令定义, 也不使用任何Value。它的类型为int/float。
*/
template<class T> struct always_false : std::false_type {};
template<class T> constexpr bool always_false_v = always_false<T>::value;
class ConstantValue : public Value {
protected:
/// 定义字面量类型的聚合类型
union {
int iScalar;
float fScalar;
};
protected:
static std::unordered_map<ConstantValueKey, ConstantValue*, ConstantValueHash, ConstantValueEqual> mConstantPool;
protected:
explicit ConstantValue(int value, const std::string &name = "") : Value(Type::getIntType(), name), iScalar(value) {}
explicit ConstantValue(float value, const std::string &name = "")
: Value(Type::getFloatType(), name), fScalar(value) {}
public:
explicit ConstantValue(Type* type, const std::string& name = "") : Value(type, name) {}
virtual ~ConstantValue() = default;
public:
static ConstantValue* get(int value); ///< 获取一个int类型的ConstValue *其值为value
static ConstantValue* get(float value); ///< 获取一个float类型的ConstValue *其值为value
virtual size_t hash() const = 0;
virtual ConstantValVariant getVal() const = 0;
public:
// Static factory method to get a canonical ConstantValue from the pool
static ConstantValue* get(Type* type, ConstantValVariant val);
// Helper methods to access constant values with appropriate casting
int getInt() const {
assert(isInt());
return iScalar;
} ///< 返回int类型的值
assert(getType()->isInt() && "Calling getInt() on non-integer type");
return std::get<int>(getVal());
}
float getFloat() const {
assert(isFloat());
return fScalar;
} ///< 返回float类型的值
template <typename T>
T getValue() const {
if (std::is_same<T, int>::value && isInt()) {
assert(getType()->isFloat() && "Calling getFloat() on non-float type");
return std::get<float>(getVal());
}
template<typename T>
T getVal() const {
if constexpr (std::is_same_v<T, int>) {
return getInt();
}
if (std::is_same<T, float>::value && isFloat()) {
} else if constexpr (std::is_same_v<T, float>) {
return getFloat();
} else {
// This ensures a compilation error if an unsupported type is used
static_assert(always_false_v<T>, "Unsupported type for ConstantValue::getValue()");
}
throw std::bad_cast(); // 或者其他适当的异常处理
} ///< 返回值getInt和getFloat统一化整数返回整形浮点返回浮点型
}
virtual bool isZero() const = 0;
virtual bool isOne() const = 0;
};
class ConstantInteger : public ConstantValue {
int constVal;
public:
explicit ConstantInteger(Type* type, int val, const std::string& name = "")
: ConstantValue(type, name), constVal(val) {}
size_t hash() const override {
std::size_t typeHash = std::hash<Type*>{}(getType());
std::size_t valHash = std::hash<int>{}(constVal);
return typeHash ^ (valHash << 1);
}
int getInt() const { return constVal; }
ConstantValVariant getVal() const override { return constVal; }
static ConstantInteger* get(Type* type, int val);
static ConstantInteger* get(int val) { return get(Type::getIntType(), val); }
ConstantInteger* getNeg() const {
assert(getType()->isInt() && "Cannot negate non-integer constant");
return ConstantInteger::get(-constVal);
}
bool isZero() const override { return constVal == 0; }
bool isOne() const override { return constVal == 1; }
};
class ConstantFloating : public ConstantValue {
float constFVal;
public:
explicit ConstantFloating(Type* type, float val, const std::string& name = "")
: ConstantValue(type, name), constFVal(val) {}
size_t hash() const override {
std::size_t typeHash = std::hash<Type*>{}(getType());
std::size_t valHash = std::hash<float>{}(constFVal);
return typeHash ^ (valHash << 1);
}
float getFloat() const { return constFVal; }
ConstantValVariant getVal() const override { return constFVal; }
static ConstantFloating* get(Type* type, float val);
static ConstantFloating* get(float val) { return get(Type::getFloatType(), val); }
ConstantFloating* getNeg() const {
assert(getType()->isFloat() && "Cannot negate non-float constant");
return ConstantFloating::get(-constFVal);
}
bool isZero() const override { return constFVal == 0.0f; }
bool isOne() const override { return constFVal == 1.0f; }
};
class UndefinedValue : public ConstantValue {
private:
static std::unordered_map<Type*, UndefinedValue*> UndefValues;
protected:
explicit UndefinedValue(Type* type, const std::string& name = "")
: ConstantValue(type, name) {
assert(!type->isVoid() && "Cannot create UndefinedValue of void type!");
}
public:
static UndefinedValue* get(Type* type);
size_t hash() const override {
return std::hash<Type*>{}(getType());
}
ConstantValVariant getVal() const override {
if (getType()->isInt()) {
return 0; // Return 0 for undefined integer
} else if (getType()->isFloat()) {
return 0.0f; // Return 0.0f for undefined float
}
assert(false && "UndefinedValue has unexpected type for getValue()");
return 0; // Should not be reached
}
bool isZero() const override { return false; }
bool isOne() const override { return false; }
};
// --- End of refactored ConstantValue and related classes ---
class Instruction;
class Function;
class BasicBlock;
@ -562,8 +697,8 @@ class Instruction : public User {
kLa = 0x1UL << 36,
kMemset = 0x1UL << 37,
kGetSubArray = 0x1UL << 38,
// constant
kConstant = 0x1UL << 37,
// 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,
@ -755,24 +890,51 @@ class LaInst : public Instruction {
class PhiInst : public Instruction {
friend class IRBuilder;
friend class Function;
friend class SysySSA;
protected:
Value *map_val; // Phi的旧值
PhiInst(Type *type, Value *lhs, const std::vector<Value *> &rhs, Value *mval, BasicBlock *parent,
std::unordered_map<BasicBlock *, Value *> blk2val; ///< 存储每个基本块对应的值
unsigned vsize; ///< 存储值的数量
PhiInst(Type *type,
const std::vector<Value *> &rhs = {},
const std::vector<BasicBlock*> &Blocks = {},
BasicBlock *parent = nullptr,
const std::string &name = "")
: Instruction(Kind::kPhi, type, parent, name) {
map_val = mval;
addOperand(lhs);
addOperands(rhs);
: 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 < rhs.size(); ++i) {
addOperand(rhs[i]);
blk2val[Blocks[i]] = rhs[i];
}
}
public:
Value* getMapVal() { return map_val; }
Value* getPointer() const { return getOperand(0); }
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
void addIncoming(Value *value, BasicBlock *block) {
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 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()); }
Value* getValue(unsigned index) const { return getOperand(index + 1); }
};
@ -884,7 +1046,7 @@ public:
}
} ///< 根据指令类型进行二元计算eval template模板实现
static BinaryInst* create(Kind kind, Type *type, Value *lhs, Value *rhs, BasicBlock *parent, const std::string &name = "") {
// 后端处理数组访存操作时需要创建计算地址的指令,需要在外部构造 BinaryInst 对象所以写了个public的方法。
// 后端处理数组访存操作时需要创建计算地址的指令,需要在外部构造 BinaryInst 对象
return new BinaryInst(kind, type, lhs, rhs, parent, name);
}
}; // class BinaryInst
@ -1230,12 +1392,15 @@ protected:
if (init.size() == 0) {
unsigned num = 1;
for (unsigned i = 0; i < numDims; i++) {
num *= dynamic_cast<ConstantValue *>(dims[i])->getInt();
// 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(ConstantValue::get(0.0F), num);
init.push_back(ConstantFloating::get(0.0F), num); // Use new constant factory
} else {
init.push_back(ConstantValue::get(0), num);
init.push_back(ConstantInteger::get(0), num); // Use new constant factory
}
}
initValues = init;
@ -1261,8 +1426,11 @@ public:
Value* getByIndices(const std::vector<Value *> &indices) const {
int index = 0;
for (size_t i = 0; i < indices.size(); i++) {
index = dynamic_cast<ConstantValue *>(getDim(i))->getInt() * index +
dynamic_cast<ConstantValue *>(indices[i])->getInt();
// Ensure dims[i] and indices[i] are ConstantInteger and retrieve their values correctly
auto dim_val = dynamic_cast<ConstantInteger*>(getDim(i));
auto idx_val = dynamic_cast<ConstantInteger*>(indices[i]);
assert(dim_val && idx_val && "Dims and indices must be constant integers");
index = dim_val->getInt() * index + idx_val->getInt();
}
return getByIndex(index);
} ///< 通过多维索引indices获取初始值
@ -1303,8 +1471,11 @@ class ConstantVariable : public User, public LVal {
int index = 0;
// 计算偏移量
for (size_t i = 0; i < indices.size(); i++) {
index = dynamic_cast<ConstantValue *>(getDim(i))->getInt() * index +
dynamic_cast<ConstantValue *>(indices[i])->getInt();
// Ensure dims[i] and indices[i] are ConstantInteger and retrieve their values correctly
auto dim_val = dynamic_cast<ConstantInteger*>(getDim(i));
auto idx_val = dynamic_cast<ConstantInteger*>(indices[i]);
assert(dim_val && idx_val && "Dims and indices must be constant integers");
index = dim_val->getInt() * index + idx_val->getInt();
}
return getByIndex(index);

View File

@ -333,15 +333,11 @@ class IRBuilder {
block->getInstructions().emplace(position, inst);
return inst;
} ///< 创建store指令
PhiInst * createPhiInst(Type *type, Value *lhs, BasicBlock *parent, const std::string &name = "") {
auto predNum = parent->getNumPredecessors();
std::vector<Value *> rhs;
for (size_t i = 0; i < predNum; i++) {
rhs.push_back(lhs);
}
auto inst = new PhiInst(type, lhs, rhs, lhs, parent, name);
PhiInst * createPhiInst(Type *type, const std::vector<Value*> &vals = {}, const std::vector<BasicBlock*> &blks = {}, const std::string &name = "") {
auto predNum = block->getNumPredecessors();
auto inst = new PhiInst(type, vals, blks, block, name);
assert(inst);
parent->getInstructions().emplace(parent->begin(), inst);
block->getInstructions().emplace(block->begin(), inst);
return inst;
} ///< 创建Phi指令
};

View File

@ -16,10 +16,10 @@ using namespace antlr4;
#include "SysYIROptPre.h"
#include "RISCv64Backend.h"
#include "SysYIRAnalyser.h"
#include "DeadCodeElimination.h"
// #include "DeadCodeElimination.h"
#include "AddressCalculationExpansion.h"
#include "Mem2Reg.h"
#include "Reg2Mem.h"
// #include "Mem2Reg.h"
// #include "Reg2Mem.h"
using namespace sysy;
@ -174,32 +174,32 @@ 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();
}
// 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();
if (DEBUG) {
cout << "=== After Mem2Reg (Default) ===\n";
SysYPrinter(moduleIR).printIR();
}
// Mem2Reg mem2reg(moduleIR, builder, &cfa, &ava);
// mem2reg.mem2regPipeline();
// if (DEBUG) {
// cout << "=== After Mem2Reg (Default) ===\n";
// SysYPrinter(moduleIR).printIR();
// }
Reg2Mem reg2mem(moduleIR, builder);
reg2mem.DeletePhiInst();
if (DEBUG) {
cout << "=== After Reg2Mem (Default) ===\n";
SysYPrinter(moduleIR).printIR();
}
// Reg2Mem reg2mem(moduleIR, builder);
// reg2mem.DeletePhiInst();
// if (DEBUG) {
// cout << "=== After Reg2Mem (Default) ===\n";
// SysYPrinter(moduleIR).printIR();
// }
dce.runDCEPipeline(); // 第二次 DCE (默认)
if (DEBUG) {
cout << "=== After 2nd DCE (Default) ===\n";
SysYPrinter(moduleIR).printIR();
}
// dce.runDCEPipeline(); // 第二次 DCE (默认)
// if (DEBUG) {
// cout << "=== After 2nd DCE (Default) ===\n";
// SysYPrinter(moduleIR).printIR();
// }
} else {
if (DEBUG) cout << "No additional middle-end optimizations applied for -O" << optLevel << ".\n";
}