[midend-LoopAnalysis]为项目添加别名分析遍,副作用分析遍,循环分析遍,循环特征分析遍

This commit is contained in:
rain2133
2025-08-08 00:56:50 +08:00
parent bd02f5f1eb
commit b1a46b7d58
12 changed files with 2166 additions and 53 deletions

View File

@ -0,0 +1,208 @@
#pragma once
#include "IR.h"
#include "Pass.h"
#include <map>
#include <set>
#include <vector>
#include <memory>
namespace sysy {
// 前向声明
class MemoryLocation;
class AliasAnalysisResult;
/**
* @brief 别名关系类型
* 按风险等级递增排序
*/
enum class AliasType {
NO_ALIAS = 0, // 确定无别名 (不同的局部数组)
SELF_ALIAS = 1, // 自别名 (同一数组的不同索引)
POSSIBLE_ALIAS = 2, // 可能有别名 (函数参数数组)
UNKNOWN_ALIAS = 3 // 未知 (保守估计)
};
/**
* @brief 内存位置信息
* 描述一个内存访问的基础信息
*/
struct MemoryLocation {
Value* basePointer; // 基指针 (剥离GEP后的真实基址)
Value* accessPointer; // 访问指针 (包含索引信息)
// 分类信息
bool isLocalArray; // 是否为局部数组
bool isFunctionParameter; // 是否为函数参数
bool isGlobalArray; // 是否为全局数组
// 索引信息
std::vector<Value*> indices; // GEP索引列表
bool hasConstantIndices; // 是否为常量索引
bool hasLoopVariableIndex; // 是否包含循环变量
int constantOffset; // 常量偏移量 (仅当全部为常量时有效)
// 访问模式
bool hasReads; // 是否有读操作
bool hasWrites; // 是否有写操作
std::vector<Instruction*> accessInsts; // 所有访问指令
MemoryLocation(Value* base, Value* access)
: basePointer(base), accessPointer(access),
isLocalArray(false), isFunctionParameter(false), isGlobalArray(false),
hasConstantIndices(false), hasLoopVariableIndex(false), constantOffset(0),
hasReads(false), hasWrites(false) {}
};
/**
* @brief 别名分析结果
* 存储一个函数的完整别名分析信息
*/
class AliasAnalysisResult : public AnalysisResultBase {
public:
AliasAnalysisResult(Function *F) : AssociatedFunction(F) {}
~AliasAnalysisResult() override = default;
// ========== 基础查询接口 ==========
/**
* 查询两个指针之间的别名关系
*/
AliasType queryAlias(Value* ptr1, Value* ptr2) const;
/**
* 查询指针的内存位置信息
*/
const MemoryLocation* getMemoryLocation(Value* ptr) const;
/**
* 获取所有内存位置
*/
const std::map<Value*, std::unique_ptr<MemoryLocation>>& getAllMemoryLocations() const {
return LocationMap;
}
// ========== 高级查询接口 ==========
/**
* 检查指针是否为局部数组
*/
bool isLocalArray(Value* ptr) const;
/**
* 检查指针是否为函数参数数组
*/
bool isFunctionParameter(Value* ptr) const;
/**
* 检查指针是否为全局数组
*/
bool isGlobalArray(Value* ptr) const;
/**
* 检查指针是否使用常量索引
*/
bool hasConstantAccess(Value* ptr) const;
// ========== 统计接口 ==========
/**
* 获取各类别名类型的统计信息
*/
struct Statistics {
int totalQueries;
int noAlias;
int selfAlias;
int possibleAlias;
int unknownAlias;
int localArrays;
int functionParameters;
int globalArrays;
int constantAccesses;
};
Statistics getStatistics() const;
/**
* 打印别名分析结果 (调试用)
*/
void print() const;
// ========== 内部方法 ==========
void addMemoryLocation(std::unique_ptr<MemoryLocation> location);
void addAliasRelation(Value* ptr1, Value* ptr2, AliasType type);
// ========== 公开数据成员 (供Pass使用) ==========
std::map<Value*, std::unique_ptr<MemoryLocation>> LocationMap; // 内存位置映射
std::map<std::pair<Value*, Value*>, AliasType> AliasMap; // 别名关系缓存
private:
Function *AssociatedFunction; // 关联的函数
// 分类存储
std::vector<Argument*> ArrayParameters; // 数组参数
std::vector<AllocaInst*> LocalArrays; // 局部数组
std::set<GlobalValue*> AccessedGlobals; // 访问的全局变量
};
/**
* @brief SysY语言特化的别名分析Pass
* 针对SysY语言特性优化的别名分析实现
*/
class SysYAliasAnalysisPass : public AnalysisPass {
public:
// 唯一的 Pass ID
static void *ID;
SysYAliasAnalysisPass() : AnalysisPass("SysYAliasAnalysis", Pass::Granularity::Function) {}
// 实现 getPassID
void *getPassID() const override { return &ID; }
// 核心运行方法
bool runOnFunction(Function *F, AnalysisManager &AM) override;
// 获取分析结果
std::unique_ptr<AnalysisResultBase> getResult() override { return std::move(CurrentResult); }
private:
std::unique_ptr<AliasAnalysisResult> CurrentResult; // 当前函数的分析结果
// ========== 主要分析流程 ==========
void collectMemoryAccesses(Function* F); // 收集内存访问
void buildAliasRelations(Function* F); // 构建别名关系
void optimizeForSysY(Function* F); // SysY特化优化
// ========== 内存位置分析 ==========
std::unique_ptr<MemoryLocation> createMemoryLocation(Value* ptr);
Value* getBasePointer(Value* ptr); // 获取基指针
void analyzeMemoryType(MemoryLocation* location); // 分析内存类型
void analyzeIndexPattern(MemoryLocation* location); // 分析索引模式
// ========== 别名关系推断 ==========
AliasType analyzeAliasBetween(MemoryLocation* loc1, MemoryLocation* loc2);
AliasType compareLocalArrays(MemoryLocation* loc1, MemoryLocation* loc2);
AliasType compareParameters(MemoryLocation* loc1, MemoryLocation* loc2);
AliasType compareWithGlobal(MemoryLocation* loc1, MemoryLocation* loc2);
AliasType compareMixedTypes(MemoryLocation* loc1, MemoryLocation* loc2);
// ========== SysY特化优化 ==========
void applySysYConstraints(Function* F); // 应用SysY语言约束
void optimizeParameterAnalysis(Function* F); // 优化参数分析
void optimizeArrayAccessAnalysis(Function* F); // 优化数组访问分析
// ========== 辅助方法 ==========
bool isConstantValue(Value* val); // 是否为常量
bool hasLoopVariableInIndices(const std::vector<Value*>& indices, Function* F);
int calculateConstantOffset(const std::vector<Value*>& indices);
void printStatistics() const; // 打印统计信息
};
} // namespace sysy

View File

@ -4,8 +4,10 @@
#include "IR.h" // 包含 IR 定义
#include "Pass.h" // 包含 Pass 框架
#include <algorithm>
#include <functional>
#include <map>
#include <memory>
#include <optional>
#include <queue> // 用于循环体块的逆向遍历
#include <set>
#include <vector>
@ -92,6 +94,65 @@ public:
return backEdgeCount == 1;
}
/**
* 获取所有出口目标块 (循环外接收循环出口边的块)
* 使用场景: 循环后置处理、phi节点分析
*/
std::vector<BasicBlock*> getExitTargetBlocks() const {
std::set<BasicBlock*> exitTargetSet;
for (BasicBlock* bb : LoopBlocks) {
for (BasicBlock* succ : bb->getSuccessors()) {
if (!contains(succ)) {
exitTargetSet.insert(succ);
}
}
}
return std::vector<BasicBlock*>(exitTargetSet.begin(), exitTargetSet.end());
}
/**
* 计算循环的"深度"相对于指定的祖先循环
* 使用场景: 相对深度计算、嵌套分析
*/
int getRelativeDepth(Loop* ancestor) const {
if (this == ancestor) return 0;
int depth = 0;
Loop* current = this->ParentLoop;
while (current && current != ancestor) {
depth++;
current = current->ParentLoop;
}
return current == ancestor ? depth : -1; // -1表示不是祖先关系
}
/**
* 检查循环是否包含函数调用
* 使用场景: 内联决策、副作用分析
*/
bool containsFunctionCalls() const {
for (BasicBlock* bb : LoopBlocks) {
for (auto& inst : bb->getInstructions()) {
if (dynamic_cast<CallInst*>(inst.get())) {
return true;
}
}
}
return false;
}
/**
* 估算循环的"热度" (基于嵌套深度和大小)
* 使用场景: 优化优先级、资源分配
*/
double getLoopHotness() const {
// 简单的热度估算: 深度权重 + 大小惩罚
double hotness = std::pow(2.0, Level); // 深度越深越热
hotness /= std::sqrt(LoopBlocks.size()); // 大小越大相对热度降低
return hotness;
}
// --- 供 LoopAnalysisPass 内部调用的方法,用于构建 Loop 对象 ---
void addBlock(BasicBlock *BB) { LoopBlocks.insert(BB); }
void addExitBlock(BasicBlock *BB) { ExitBlocks.insert(BB); }
@ -112,81 +173,345 @@ private:
/**
* @brief 循环分析结果类。
* 包含一个函数中所有识别出的循环。
* 包含一个函数中所有识别出的循环,并提供高效的查询缓存机制
*/
class LoopAnalysisResult : public AnalysisResultBase {
public:
LoopAnalysisResult(Function *F) : AssociatedFunction(F) {}
~LoopAnalysisResult() override = default;
// ========== 缓存统计结构 ==========
struct CacheStats {
size_t innermostLoopsCached;
size_t outermostLoopsCached;
size_t loopsByDepthCached;
size_t containingLoopsCached;
size_t allNestedLoopsCached;
size_t totalCachedQueries;
};
private:
// ========== 高频查询缓存 (必须缓存) ==========
mutable std::optional<std::vector<Loop*>> cachedInnermostLoops;
mutable std::optional<std::vector<Loop*>> cachedOutermostLoops;
mutable std::optional<int> cachedMaxDepth;
mutable std::optional<size_t> cachedLoopCount;
mutable std::map<int, std::vector<Loop*>> cachedLoopsByDepth;
// ========== 中频查询缓存 (选择性缓存) ==========
mutable std::map<BasicBlock*, Loop*> cachedInnermostContainingLoop;
mutable std::map<Loop*, std::set<Loop*>> cachedAllNestedLoops; // 递归嵌套
mutable std::map<BasicBlock*, std::vector<Loop*>> cachedAllContainingLoops;
// ========== 缓存状态管理 ==========
mutable bool cacheValid = true;
// 内部辅助方法
void invalidateCache() const {
cachedInnermostLoops.reset();
cachedOutermostLoops.reset();
cachedMaxDepth.reset();
cachedLoopCount.reset();
cachedLoopsByDepth.clear();
cachedInnermostContainingLoop.clear();
cachedAllNestedLoops.clear();
cachedAllContainingLoops.clear();
cacheValid = false;
}
void ensureCacheValid() const {
if (!cacheValid) {
// 重新计算基础缓存
computeBasicCache();
cacheValid = true;
}
}
void computeBasicCache() const {
// 计算最内层循环
if (!cachedInnermostLoops) {
cachedInnermostLoops = std::vector<Loop*>();
for (const auto& loop : AllLoops) {
if (loop->isInnermost()) {
cachedInnermostLoops->push_back(loop.get());
}
}
}
// 计算最外层循环
if (!cachedOutermostLoops) {
cachedOutermostLoops = std::vector<Loop*>();
for (const auto& loop : AllLoops) {
if (loop->isOutermost()) {
cachedOutermostLoops->push_back(loop.get());
}
}
}
// 计算最大深度
if (!cachedMaxDepth) {
int maxDepth = 0;
for (const auto& loop : AllLoops) {
maxDepth = std::max(maxDepth, loop->getLoopDepth());
}
cachedMaxDepth = maxDepth;
}
// 计算循环总数
if (!cachedLoopCount) {
cachedLoopCount = AllLoops.size();
}
}
public:
// ========== 基础接口 (保持向后兼容,但增加缓存失效) ==========
// 添加一个识别出的循环到结果中
void addLoop(std::unique_ptr<Loop> loop) {
invalidateCache(); // 添加新循环时失效缓存
AllLoops.push_back(std::move(loop));
// 也可以选择将 Loop* 存储到 map 中,方便通过 header 查找
LoopMap[AllLoops.back()->getHeader()] = AllLoops.back().get();
}
// 获取所有识别出的循环unique_ptr 管理内存)
const std::vector<std::unique_ptr<Loop>> &getAllLoops() const { return AllLoops; }
// 获取所有最外层循环
const std::vector<Loop *> &getOutermostLoops() const { return OutermostLoops; }
const std::vector<Loop *> &getInnermostLoops() const { return InnermostLoops; }
// 获取循环总数
size_t getLoopCount() const { return AllLoops.size(); }
// 获取最大循环嵌套深度
int getMaxLoopDepth() const {
int maxDepth = 0;
for (const auto& loop : AllLoops) {
if (loop->getLoopDepth() > maxDepth) {
maxDepth = loop->getLoopDepth();
// ========== 高频查询接口 (缓存优化) ==========
/**
* 获取所有最内层循环 - 循环优化的主要目标
* 使用场景: 循环展开、向量化、循环不变量外提
*/
const std::vector<Loop*>& getInnermostLoops() const {
ensureCacheValid();
if (!cachedInnermostLoops) {
cachedInnermostLoops = std::vector<Loop*>();
for (const auto& loop : AllLoops) {
if (loop->isInnermost()) {
cachedInnermostLoops->push_back(loop.get());
}
}
}
return maxDepth;
return *cachedInnermostLoops;
}
/**
* 获取所有最外层循环
* 使用场景: 循环树遍历、整体优化策略
*/
const std::vector<Loop*>& getOutermostLoops() const {
ensureCacheValid();
if (!cachedOutermostLoops) {
cachedOutermostLoops = std::vector<Loop*>();
for (const auto& loop : AllLoops) {
if (loop->isOutermost()) {
cachedOutermostLoops->push_back(loop.get());
}
}
}
return *cachedOutermostLoops;
}
/**
* 获取指定深度的所有循环
* 使用场景: 分层优化、循环展开决策、并行化分析
*/
const std::vector<Loop*>& getLoopsAtDepth(int depth) const {
ensureCacheValid();
if (cachedLoopsByDepth.find(depth) == cachedLoopsByDepth.end()) {
std::vector<Loop*> result;
for (const auto& loop : AllLoops) {
if (loop->getLoopDepth() == depth) {
result.push_back(loop.get());
}
}
cachedLoopsByDepth[depth] = std::move(result);
}
return cachedLoopsByDepth[depth];
}
/**
* 获取最大循环嵌套深度
* 使用场景: 优化预算分配、编译时间控制
*/
int getMaxLoopDepth() const {
ensureCacheValid();
if (!cachedMaxDepth) {
int maxDepth = 0;
for (const auto& loop : AllLoops) {
maxDepth = std::max(maxDepth, loop->getLoopDepth());
}
cachedMaxDepth = maxDepth;
}
return *cachedMaxDepth;
}
/**
* 获取循环总数
* 使用场景: 统计信息、优化决策
*/
size_t getLoopCount() const {
ensureCacheValid();
if (!cachedLoopCount) {
cachedLoopCount = AllLoops.size();
}
return *cachedLoopCount;
}
// 获取指定深度的循环数量
size_t getLoopCountAtDepth(int depth) const {
size_t count = 0;
for (const auto& loop : AllLoops) {
if (loop->getLoopDepth() == depth) {
count++;
}
}
return count;
return getLoopsAtDepth(depth).size();
}
// 检查函数是否包含循环
bool hasLoops() const { return !AllLoops.empty(); }
// ========== 中频查询接口 (选择性缓存) ==========
/**
* 获取包含指定基本块的最内层循环
* 使用场景: 活跃性分析、寄存器分配、指令调度
*/
Loop* getInnermostContainingLoop(BasicBlock* BB) const {
ensureCacheValid();
if (cachedInnermostContainingLoop.find(BB) == cachedInnermostContainingLoop.end()) {
Loop* result = nullptr;
int maxDepth = -1;
for (const auto& loop : AllLoops) {
if (loop->contains(BB) && loop->getLoopDepth() > maxDepth) {
result = loop.get();
maxDepth = loop->getLoopDepth();
}
}
cachedInnermostContainingLoop[BB] = result;
}
return cachedInnermostContainingLoop[BB];
}
/**
* 获取包含指定基本块的所有循环 (从外到内排序)
* 使用场景: 循环间优化、依赖分析
*/
const std::vector<Loop*>& getAllContainingLoops(BasicBlock* BB) const {
ensureCacheValid();
if (cachedAllContainingLoops.find(BB) == cachedAllContainingLoops.end()) {
std::vector<Loop*> result;
for (const auto& loop : AllLoops) {
if (loop->contains(BB)) {
result.push_back(loop.get());
}
}
// 按深度排序 (外层到内层)
std::sort(result.begin(), result.end(),
[](Loop* a, Loop* b) { return a->getLoopDepth() < b->getLoopDepth(); });
cachedAllContainingLoops[BB] = std::move(result);
}
return cachedAllContainingLoops[BB];
}
/**
* 获取指定循环的所有嵌套子循环 (递归)
* 使用场景: 循环树分析、嵌套优化
*/
const std::set<Loop*>& getAllNestedLoops(Loop* loop) const {
ensureCacheValid();
if (cachedAllNestedLoops.find(loop) == cachedAllNestedLoops.end()) {
std::set<Loop*> result;
std::function<void(Loop*)> collectNested = [&](Loop* current) {
for (Loop* nested : current->getNestedLoops()) {
result.insert(nested);
collectNested(nested); // 递归收集
}
};
collectNested(loop);
cachedAllNestedLoops[loop] = std::move(result);
}
return cachedAllNestedLoops[loop];
}
// ========== 低频查询接口 (按需计算,不缓存) ==========
/**
* 检查两个循环是否有嵌套关系
* 使用场景: 循环间依赖分析
*/
bool isNestedLoop(Loop* inner, Loop* outer) const {
if (inner == outer) return false;
Loop* current = inner->getParentLoop();
while (current) {
if (current == outer) return true;
current = current->getParentLoop();
}
return false;
}
/**
* 获取两个循环的最近公共祖先循环
* 使用场景: 循环融合分析、优化范围确定
*/
Loop* getLowestCommonAncestor(Loop* loop1, Loop* loop2) const {
if (!loop1 || !loop2) return nullptr;
if (loop1 == loop2) return loop1;
// 收集loop1的所有祖先
std::set<Loop*> ancestors1;
Loop* current = loop1;
while (current) {
ancestors1.insert(current);
current = current->getParentLoop();
}
// 查找loop2祖先链中第一个在ancestors1中的循环
current = loop2;
while (current) {
if (ancestors1.count(current)) {
return current;
}
current = current->getParentLoop();
}
return nullptr; // 没有公共祖先
}
// 通过循环头获取 Loop 对象
Loop *getLoopForHeader(BasicBlock *header) const {
auto it = LoopMap.find(header);
return (it != LoopMap.end()) ? it->second : nullptr;
}
// 通过某个基本块获取包含它的最内层循环
// 通过某个基本块获取包含它的最内层循环 (向后兼容接口)
Loop *getLoopContainingBlock(BasicBlock *BB) const {
// 遍历所有循环,找到包含 BB 且层级最深的循环
Loop *innermostContainingLoop = nullptr;
for (const auto &loop_ptr : AllLoops) {
if (loop_ptr->contains(BB)) {
if (!innermostContainingLoop || loop_ptr->getLoopLevel() > innermostContainingLoop->getLoopLevel()) {
innermostContainingLoop = loop_ptr.get();
}
}
}
return innermostContainingLoop;
return getInnermostContainingLoop(BB);
}
// --- 供 LoopAnalysisPass 内部调用的方法,用于构建 LoopAnalysisResult 对象 ---
void addOutermostLoop(Loop *loop) { OutermostLoops.push_back(loop); }
void addInnermostLoop(Loop *loop) { InnermostLoops.push_back(loop); }
void clearOutermostLoops() { OutermostLoops.clear(); }
void clearInnermostLoops() { InnermostLoops.clear(); }
// ========== 缓存管理接口 ==========
/**
* 手动失效缓存 (当IR结构改变时调用)
*/
void invalidateQueryCache() const {
invalidateCache();
}
/**
* 获取缓存统计信息 (用于性能调试)
*/
CacheStats getCacheStats() const {
CacheStats stats = {};
stats.innermostLoopsCached = cachedInnermostLoops.has_value() ? 1 : 0;
stats.outermostLoopsCached = cachedOutermostLoops.has_value() ? 1 : 0;
stats.loopsByDepthCached = cachedLoopsByDepth.size();
stats.containingLoopsCached = cachedInnermostContainingLoop.size();
stats.allNestedLoopsCached = cachedAllNestedLoops.size();
stats.totalCachedQueries = stats.innermostLoopsCached + stats.outermostLoopsCached +
stats.loopsByDepthCached + stats.containingLoopsCached +
stats.allNestedLoopsCached;
return stats;
}
// --- 保留的内部接口 ---
// 注意:由于使用了缓存机制,不再需要手动维护最外层和最内层循环列表
// 打印分析结果
void print() const;
@ -197,8 +522,7 @@ private:
Function *AssociatedFunction; // 结果关联的函数
std::vector<std::unique_ptr<Loop>> AllLoops; // 所有识别出的循环
std::map<BasicBlock *, Loop *> LoopMap; // 循环头到 Loop* 的映射,方便查找
std::vector<Loop *> OutermostLoops; // 最外层循环列表
std::vector<Loop *> InnermostLoops; // 最内层循环的列表
// 注意: 最外层和最内层循环列表已移除,现在通过缓存机制动态计算
};
/**

View File

@ -0,0 +1,301 @@
#pragma once
#include "Dom.h" // 支配树分析依赖
#include "Loop.h" // 循环分析依赖
#include "Liveness.h" // 活跃性分析依赖
#include "AliasAnalysis.h" // 别名分析依赖
#include "IR.h" // IR定义
#include "Pass.h" // Pass框架
#include <algorithm>
#include <map>
#include <memory>
#include <optional>
#include <set>
#include <vector>
namespace sysy {
// 前向声明
class LoopCharacteristicsResult;
/**
* @brief 循环特征信息结构
* 存储单个循环的各种特征信息
*/
struct LoopCharacteristics {
Loop* loop; // 关联的循环对象
// ========== 归纳变量分析 ==========
std::vector<Value*> basicInductionVars; // 基本归纳变量 (i = phi(init, i+step))
std::vector<Value*> derivedInductionVars; // 派生归纳变量 (j = i * scale + offset)
std::map<Value*, int> inductionSteps; // 归纳变量的步长
std::map<Value*, Value*> inductionInits; // 归纳变量的初始值
// ========== 循环不变量分析 ==========
std::set<Value*> loopInvariants; // 循环不变量 (循环内定义但值不变)
std::set<Instruction*> invariantInsts; // 不变指令 (可以外提的指令)
// ========== 循环边界分析 ==========
std::optional<int> staticTripCount; // 静态可确定的循环次数
Value* dynamicTripCountExpr; // 动态循环次数表达式
bool hasKnownBounds; // 是否有已知边界
Value* lowerBound; // 循环下界
Value* upperBound; // 循环上界
// ========== 循环形式分析 ==========
bool isCountingLoop; // 是否为计数循环 (for i=0; i<n; i++)
bool isSimpleForLoop; // 是否为简单for循环
bool hasComplexControlFlow; // 是否有复杂控制流 (break, continue)
bool isInnermost; // 是否为最内层循环
bool isParallel; // 是否可并行化
// ========== 内存访问模式 ==========
struct MemoryAccessPattern {
bool isSequential; // 是否顺序访问 (a[i], a[i+1], ...)
bool isStrided; // 是否跨步访问 (a[2*i], a[3*i], ...)
int stride; // 访问步长
std::vector<Instruction*> loadInsts; // load指令列表
std::vector<Instruction*> storeInsts; // store指令列表
// 使用外部别名分析结果
AliasType aliasType; // 别名类型(来自别名分析)
bool isArrayParameter; // 是否为数组参数访问
bool isGlobalArray; // 是否为全局数组访问
bool hasConstantIndices; // 是否使用常量索引
};
std::map<Value*, MemoryAccessPattern> memoryPatterns; // 内存访问模式
// ========== 循环优化提示 ==========
bool benefitsFromUnrolling; // 是否适合循环展开
bool benefitsFromVectorization; // 是否适合向量化
bool benefitsFromTiling; // 是否适合分块
int suggestedUnrollFactor; // 建议的展开因子
// ========== 性能特征 ==========
size_t instructionCount; // 循环体指令数
size_t memoryOperationCount; // 内存操作数
size_t arithmeticOperationCount; // 算术操作数
double computeToMemoryRatio; // 计算与内存操作比率
// 构造函数
LoopCharacteristics(Loop* l) : loop(l), dynamicTripCountExpr(nullptr),
hasKnownBounds(false), lowerBound(nullptr), upperBound(nullptr),
isCountingLoop(false), isSimpleForLoop(false), hasComplexControlFlow(false),
isInnermost(false), isParallel(false), benefitsFromUnrolling(false),
benefitsFromVectorization(false), benefitsFromTiling(false),
suggestedUnrollFactor(1), instructionCount(0), memoryOperationCount(0),
arithmeticOperationCount(0), computeToMemoryRatio(0.0) {}
};
/**
* @brief 循环特征分析结果类
* 包含函数中所有循环的特征信息,并提供查询接口
*/
class LoopCharacteristicsResult : public AnalysisResultBase {
public:
LoopCharacteristicsResult(Function *F) : AssociatedFunction(F) {}
~LoopCharacteristicsResult() override = default;
// ========== 基础接口 ==========
/**
* 添加循环特征信息
*/
void addLoopCharacteristics(std::unique_ptr<LoopCharacteristics> characteristics) {
auto* loop = characteristics->loop;
CharacteristicsMap[loop] = std::move(characteristics);
}
/**
* 获取指定循环的特征信息
*/
const LoopCharacteristics* getCharacteristics(Loop* loop) const {
auto it = CharacteristicsMap.find(loop);
return (it != CharacteristicsMap.end()) ? it->second.get() : nullptr;
}
/**
* 获取所有循环特征信息
*/
const std::map<Loop*, std::unique_ptr<LoopCharacteristics>>& getAllCharacteristics() const {
return CharacteristicsMap;
}
// ========== 查询接口 ==========
/**
* 获取所有计数循环
*/
std::vector<Loop*> getCountingLoops() const {
std::vector<Loop*> result;
for (const auto& [loop, chars] : CharacteristicsMap) {
if (chars->isCountingLoop) {
result.push_back(loop);
}
}
return result;
}
/**
* 获取所有可向量化循环
*/
std::vector<Loop*> getVectorizableLoops() const {
std::vector<Loop*> result;
for (const auto& [loop, chars] : CharacteristicsMap) {
if (chars->benefitsFromVectorization) {
result.push_back(loop);
}
}
return result;
}
/**
* 获取所有适合展开的循环
*/
std::vector<Loop*> getUnrollCandidateLoops() const {
std::vector<Loop*> result;
for (const auto& [loop, chars] : CharacteristicsMap) {
if (chars->benefitsFromUnrolling) {
result.push_back(loop);
}
}
return result;
}
/**
* 获取所有可并行化循环
*/
std::vector<Loop*> getParallelizableLoops() const {
std::vector<Loop*> result;
for (const auto& [loop, chars] : CharacteristicsMap) {
if (chars->isParallel) {
result.push_back(loop);
}
}
return result;
}
/**
* 获取所有有静态已知循环次数的循环
*/
std::vector<Loop*> getStaticBoundLoops() const {
std::vector<Loop*> result;
for (const auto& [loop, chars] : CharacteristicsMap) {
if (chars->staticTripCount.has_value()) {
result.push_back(loop);
}
}
return result;
}
/**
* 根据热度排序循环 (用于优化优先级)
*/
std::vector<Loop*> getLoopsByHotness() const {
std::vector<Loop*> result;
for (const auto& [loop, chars] : CharacteristicsMap) {
result.push_back(loop);
}
// 按循环热度排序 (嵌套深度 + 循环次数 + 指令数)
std::sort(result.begin(), result.end(), [](Loop* a, Loop* b) {
double hotnessA = a->getLoopHotness();
double hotnessB = b->getLoopHotness();
return hotnessA > hotnessB; // 降序排列
});
return result;
}
// ========== 统计接口 ==========
/**
* 获取优化候选统计
*/
struct OptimizationStats {
size_t totalLoops;
size_t countingLoops;
size_t vectorizableLoops;
size_t unrollCandidates;
size_t parallelizableLoops;
size_t staticBoundLoops;
double avgInstructionCount;
double avgComputeMemoryRatio;
};
OptimizationStats getOptimizationStats() const {
OptimizationStats stats = {};
stats.totalLoops = CharacteristicsMap.size();
size_t totalInstructions = 0;
double totalComputeMemoryRatio = 0.0;
for (const auto& [loop, chars] : CharacteristicsMap) {
if (chars->isCountingLoop) stats.countingLoops++;
if (chars->benefitsFromVectorization) stats.vectorizableLoops++;
if (chars->benefitsFromUnrolling) stats.unrollCandidates++;
if (chars->isParallel) stats.parallelizableLoops++;
if (chars->staticTripCount.has_value()) stats.staticBoundLoops++;
totalInstructions += chars->instructionCount;
totalComputeMemoryRatio += chars->computeToMemoryRatio;
}
if (stats.totalLoops > 0) {
stats.avgInstructionCount = static_cast<double>(totalInstructions) / stats.totalLoops;
stats.avgComputeMemoryRatio = totalComputeMemoryRatio / stats.totalLoops;
}
return stats;
}
// 打印分析结果
void print() const;
private:
Function *AssociatedFunction; // 关联的函数
std::map<Loop*, std::unique_ptr<LoopCharacteristics>> CharacteristicsMap; // 循环特征映射
};
/**
* @brief 循环特征分析遍
* 基于循环分析结果,分析每个循环的特征信息,为优化决策提供依据
*/
class LoopCharacteristicsPass : public AnalysisPass {
public:
// 唯一的 Pass ID
static void *ID;
LoopCharacteristicsPass() : AnalysisPass("LoopCharacteristics", Pass::Granularity::Function) {}
// 实现 getPassID
void *getPassID() const override { return &ID; }
// 核心运行方法
bool runOnFunction(Function *F, AnalysisManager &AM) override;
// 获取分析结果
std::unique_ptr<AnalysisResultBase> getResult() override { return std::move(CurrentResult); }
private:
std::unique_ptr<LoopCharacteristicsResult> CurrentResult; // 当前函数的分析结果
// 内部分析方法
void analyzeLoop(Loop* loop, LoopCharacteristics* characteristics, AnalysisManager &AM);
void identifyInductionVariables(Loop* loop, LoopCharacteristics* characteristics);
void identifyLoopInvariants(Loop* loop, LoopCharacteristics* characteristics);
void analyzeLoopBounds(Loop* loop, LoopCharacteristics* characteristics);
void analyzeLoopForm(Loop* loop, LoopCharacteristics* characteristics);
void analyzeMemoryAccessPatterns(Loop* loop, LoopCharacteristics* characteristics);
void evaluateOptimizationOpportunities(Loop* loop, LoopCharacteristics* characteristics);
void computePerformanceMetrics(Loop* loop, LoopCharacteristics* characteristics);
// 辅助方法
bool isInductionVariable(Value* val, Loop* loop);
bool isLoopInvariant(Value* val, Loop* loop);
bool hasLoopCarriedDependence(Loop* loop);
int estimateUnrollFactor(Loop* loop);
bool benefitsFromVectorization(Loop* loop);
};
} // namespace sysy

View File

@ -0,0 +1,126 @@
#pragma once
#include "Pass.h"
#include "IR.h"
#include "AliasAnalysis.h"
#include <unordered_set>
#include <unordered_map>
namespace sysy {
// 副作用类型枚举
enum class SideEffectType {
NO_SIDE_EFFECT, // 无副作用
MEMORY_WRITE, // 内存写入store、memset
FUNCTION_CALL, // 函数调用(可能有任意副作用)
IO_OPERATION, // I/O操作printf、scanf等
UNKNOWN // 未知副作用
};
// 副作用信息结构
struct SideEffectInfo {
SideEffectType type = SideEffectType::NO_SIDE_EFFECT;
bool mayModifyGlobal = false; // 可能修改全局变量
bool mayModifyMemory = false; // 可能修改内存
bool mayCallFunction = false; // 可能调用函数
bool isPure = true; // 是否为纯函数(无副作用且结果只依赖参数)
// 合并两个副作用信息
SideEffectInfo merge(const SideEffectInfo& other) const {
SideEffectInfo result;
result.type = (type == SideEffectType::NO_SIDE_EFFECT) ? other.type : type;
result.mayModifyGlobal = mayModifyGlobal || other.mayModifyGlobal;
result.mayModifyMemory = mayModifyMemory || other.mayModifyMemory;
result.mayCallFunction = mayCallFunction || other.mayCallFunction;
result.isPure = isPure && other.isPure;
return result;
}
};
// 副作用分析结果类
class SideEffectAnalysisResult : public AnalysisResultBase {
private:
// 指令级别的副作用信息
std::unordered_map<Instruction*, SideEffectInfo> instructionSideEffects;
// 函数级别的副作用信息
std::unordered_map<Function*, SideEffectInfo> functionSideEffects;
// 已知的SysY标准库函数副作用信息
std::unordered_map<std::string, SideEffectInfo> knownFunctions;
public:
SideEffectAnalysisResult();
virtual ~SideEffectAnalysisResult() noexcept override = default;
// 获取指令的副作用信息
const SideEffectInfo& getInstructionSideEffect(Instruction* inst) const;
// 获取函数的副作用信息
const SideEffectInfo& getFunctionSideEffect(Function* func) const;
// 设置指令的副作用信息
void setInstructionSideEffect(Instruction* inst, const SideEffectInfo& info);
// 设置函数的副作用信息
void setFunctionSideEffect(Function* func, const SideEffectInfo& info);
// 检查指令是否有副作用
bool hasSideEffect(Instruction* inst) const;
// 检查指令是否可能修改内存
bool mayModifyMemory(Instruction* inst) const;
// 检查指令是否可能修改全局状态
bool mayModifyGlobal(Instruction* inst) const;
// 检查函数是否为纯函数
bool isPureFunction(Function* func) const;
// 获取已知函数的副作用信息
const SideEffectInfo* getKnownFunctionSideEffect(const std::string& funcName) const;
// 初始化已知函数的副作用信息
void initializeKnownFunctions();
private:
};
// 副作用分析遍类
class SysYSideEffectAnalysisPass : public AnalysisPass {
public:
// 静态成员作为该遍的唯一ID
static void* ID;
SysYSideEffectAnalysisPass() : AnalysisPass("SysYSideEffectAnalysis", Granularity::Function) {}
// 在函数上运行分析
bool runOnFunction(Function* F, AnalysisManager& AM) override;
// 获取分析结果
std::unique_ptr<AnalysisResultBase> getResult() override;
// Pass 基类中的纯虚函数,必须实现
void* getPassID() const override { return &ID; }
private:
// 分析结果
std::unique_ptr<SideEffectAnalysisResult> result;
// 别名分析结果(在整个函数分析过程中重复使用)
AliasAnalysisResult* aliasAnalysis = nullptr;
// 分析单个指令的副作用
SideEffectInfo analyzeInstruction(Instruction* inst, AnalysisManager& AM);
// 分析函数调用指令的副作用
SideEffectInfo analyzeCallInstruction(CallInst* call, AnalysisManager& AM);
// 分析存储指令的副作用
SideEffectInfo analyzeStoreInstruction(StoreInst* store, AnalysisManager& AM);
// 分析内存设置指令的副作用
SideEffectInfo analyzeMemsetInstruction(MemsetInst* memset, AnalysisManager& AM);
};
} // namespace sysy

View File

@ -151,17 +151,21 @@ public:
}
AnalysisPass *analysisPass = static_cast<AnalysisPass *>(basePass.get());
if(DEBUG){
std::cout << "Running Analysis Pass: " << analysisPass->getName() << "\n";
}
// 根据分析遍的粒度处理
switch (analysisPass->getGranularity()) {
case Pass::Granularity::Module: {
// 检查是否已存在有效结果
auto it = moduleCachedResults.find(analysisID);
if (it != moduleCachedResults.end()) {
if(DEBUG) {
std::cout << "Using cached result for Analysis Pass: " << analysisPass->getName() << "\n";
}
return static_cast<T *>(it->second.get()); // 返回缓存结果
}
// 只有在实际运行时才打印调试信息
if(DEBUG){
std::cout << "Running Analysis Pass: " << analysisPass->getName() << "\n";
}
// 运行模块级分析遍
if (!pModuleRef) {
std::cerr << "Error: Module reference not set for AnalysisManager to run Module Pass.\n";
@ -183,8 +187,16 @@ public:
// 检查是否已存在有效结果
auto it = functionCachedResults.find({F, analysisID});
if (it != functionCachedResults.end()) {
if(DEBUG) {
std::cout << "Using cached result for Analysis Pass: " << analysisPass->getName() << " (Function: " << F->getName() << ")\n";
}
return static_cast<T *>(it->second.get()); // 返回缓存结果
}
// 只有在实际运行时才打印调试信息
if(DEBUG){
std::cout << "Running Analysis Pass: " << analysisPass->getName() << "\n";
std::cout << "Function: " << F->getName() << "\n";
}
// 运行函数级分析遍
analysisPass->runOnFunction(F, *this);
// 获取结果并缓存
@ -202,8 +214,16 @@ public:
// 检查是否已存在有效结果
auto it = basicBlockCachedResults.find({BB, analysisID});
if (it != basicBlockCachedResults.end()) {
if(DEBUG) {
std::cout << "Using cached result for Analysis Pass: " << analysisPass->getName() << " (BasicBlock: " << BB->getName() << ")\n";
}
return static_cast<T *>(it->second.get()); // 返回缓存结果
}
// 只有在实际运行时才打印调试信息
if(DEBUG){
std::cout << "Running Analysis Pass: " << analysisPass->getName() << "\n";
std::cout << "BasicBlock: " << BB->getName() << "\n";
}
// 运行基本块级分析遍
analysisPass->runOnBasicBlock(BB, *this);
// 获取结果并缓存

View File

@ -7,6 +7,9 @@ add_library(midend_lib STATIC
Pass/Analysis/Dom.cpp
Pass/Analysis/Liveness.cpp
Pass/Analysis/Loop.cpp
Pass/Analysis/LoopCharacteristics.cpp
Pass/Analysis/AliasAnalysis.cpp
Pass/Analysis/SideEffectAnalysis.cpp
Pass/Optimize/DCE.cpp
Pass/Optimize/Mem2Reg.cpp
Pass/Optimize/Reg2Mem.cpp

View File

@ -0,0 +1,380 @@
#include "AliasAnalysis.h"
#include "SysYIRPrinter.h"
#include <iostream>
extern int DEBUG;
namespace sysy {
// 静态成员初始化
void *SysYAliasAnalysisPass::ID = (void *)&SysYAliasAnalysisPass::ID;
// ========== AliasAnalysisResult 实现 ==========
void AliasAnalysisResult::print() const {
std::cout << "---- Alias Analysis Results for Function: " << AssociatedFunction->getName() << " ----\n";
// 打印内存位置信息
std::cout << " Memory Locations (" << LocationMap.size() << "):\n";
for (const auto& pair : LocationMap) {
const auto& loc = pair.second;
std::cout << " - Base: " << loc->basePointer->getName();
std::cout << " (Type: ";
if (loc->isLocalArray) std::cout << "Local";
else if (loc->isFunctionParameter) std::cout << "Parameter";
else if (loc->isGlobalArray) std::cout << "Global";
else std::cout << "Unknown";
std::cout << ")\n";
}
// 打印别名关系
std::cout << " Alias Relations (" << AliasMap.size() << "):\n";
for (const auto& pair : AliasMap) {
std::cout << " - (" << pair.first.first->getName() << ", " << pair.first.second->getName() << "): ";
switch (pair.second) {
case AliasType::NO_ALIAS: std::cout << "No Alias"; break;
case AliasType::SELF_ALIAS: std::cout << "Self Alias"; break;
case AliasType::POSSIBLE_ALIAS: std::cout << "Possible Alias"; break;
case AliasType::UNKNOWN_ALIAS: std::cout << "Unknown Alias"; break;
}
std::cout << "\n";
}
std::cout << "-----------------------------------------------------------\n";
}
AliasType AliasAnalysisResult::queryAlias(Value* ptr1, Value* ptr2) const {
auto key = std::make_pair(ptr1, ptr2);
auto it = AliasMap.find(key);
if (it != AliasMap.end()) {
return it->second;
}
// 尝试反向查找
key = std::make_pair(ptr2, ptr1);
it = AliasMap.find(key);
if (it != AliasMap.end()) {
return it->second;
}
return AliasType::UNKNOWN_ALIAS; // 保守估计
}
const MemoryLocation* AliasAnalysisResult::getMemoryLocation(Value* ptr) const {
auto it = LocationMap.find(ptr);
return (it != LocationMap.end()) ? it->second.get() : nullptr;
}
bool AliasAnalysisResult::isLocalArray(Value* ptr) const {
const MemoryLocation* loc = getMemoryLocation(ptr);
return loc && loc->isLocalArray;
}
bool AliasAnalysisResult::isFunctionParameter(Value* ptr) const {
const MemoryLocation* loc = getMemoryLocation(ptr);
return loc && loc->isFunctionParameter;
}
bool AliasAnalysisResult::isGlobalArray(Value* ptr) const {
const MemoryLocation* loc = getMemoryLocation(ptr);
return loc && loc->isGlobalArray;
}
bool AliasAnalysisResult::hasConstantAccess(Value* ptr) const {
const MemoryLocation* loc = getMemoryLocation(ptr);
return loc && loc->hasConstantIndices;
}
AliasAnalysisResult::Statistics AliasAnalysisResult::getStatistics() const {
Statistics stats = {0};
stats.totalQueries = AliasMap.size();
for (auto& pair : AliasMap) {
switch (pair.second) {
case AliasType::NO_ALIAS: stats.noAlias++; break;
case AliasType::SELF_ALIAS: stats.selfAlias++; break;
case AliasType::POSSIBLE_ALIAS: stats.possibleAlias++; break;
case AliasType::UNKNOWN_ALIAS: stats.unknownAlias++; break;
}
}
for (auto& loc : LocationMap) {
if (loc.second->isLocalArray) stats.localArrays++;
if (loc.second->isFunctionParameter) stats.functionParameters++;
if (loc.second->isGlobalArray) stats.globalArrays++;
if (loc.second->hasConstantIndices) stats.constantAccesses++;
}
return stats;
}
// void AliasAnalysisResult::print() const {
// std::cout << "=== Alias Analysis Results ===" << std::endl;
// auto stats = getStatistics();
// std::cout << "Total queries: " << stats.totalQueries << std::endl;
// std::cout << "No alias: " << stats.noAlias << std::endl;
// std::cout << "Self alias: " << stats.selfAlias << std::endl;
// std::cout << "Possible alias: " << stats.possibleAlias << std::endl;
// std::cout << "Unknown alias: " << stats.unknownAlias << std::endl;
// std::cout << "Local arrays: " << stats.localArrays << std::endl;
// std::cout << "Function parameters: " << stats.functionParameters << std::endl;
// std::cout << "Global arrays: " << stats.globalArrays << std::endl;
// std::cout << "Constant accesses: " << stats.constantAccesses << std::endl;
// }
void AliasAnalysisResult::addMemoryLocation(std::unique_ptr<MemoryLocation> location) {
Value* ptr = location->accessPointer;
LocationMap[ptr] = std::move(location);
}
void AliasAnalysisResult::addAliasRelation(Value* ptr1, Value* ptr2, AliasType type) {
auto key = std::make_pair(ptr1, ptr2);
AliasMap[key] = type;
}
// ========== SysYAliasAnalysisPass 实现 ==========
bool SysYAliasAnalysisPass::runOnFunction(Function *F, AnalysisManager &AM) {
if (DEBUG) {
std::cout << "Running SysY Alias Analysis on function: " << F->getName() << std::endl;
}
// 创建分析结果
CurrentResult = std::make_unique<AliasAnalysisResult>(F);
// 执行主要分析步骤
collectMemoryAccesses(F);
buildAliasRelations(F);
optimizeForSysY(F);
if (DEBUG) {
CurrentResult->print();
}
return false; // 分析遍不修改IR
}
void SysYAliasAnalysisPass::collectMemoryAccesses(Function* F) {
// 收集函数中所有内存访问指令
for (auto& bb : F->getBasicBlocks()) {
for (auto& inst : bb->getInstructions()) {
Value* ptr = nullptr;
if (auto* loadInst = dynamic_cast<LoadInst*>(inst.get())) {
ptr = loadInst->getPointer();
} else if (auto* storeInst = dynamic_cast<StoreInst*>(inst.get())) {
ptr = storeInst->getPointer();
}
if (ptr) {
// 创建内存位置信息
auto location = createMemoryLocation(ptr);
location->accessInsts.push_back(inst.get());
// 更新读写标记
if (dynamic_cast<LoadInst*>(inst.get())) {
location->hasReads = true;
} else {
location->hasWrites = true;
}
CurrentResult->addMemoryLocation(std::move(location));
}
}
}
}
void SysYAliasAnalysisPass::buildAliasRelations(Function *F) {
// 构建所有内存访问之间的别名关系
auto& locationMap = CurrentResult->LocationMap;
std::vector<Value*> allPointers;
for (auto& pair : locationMap) {
allPointers.push_back(pair.first);
}
// 两两比较所有指针
for (size_t i = 0; i < allPointers.size(); ++i) {
for (size_t j = i + 1; j < allPointers.size(); ++j) {
Value* ptr1 = allPointers[i];
Value* ptr2 = allPointers[j];
MemoryLocation* loc1 = locationMap[ptr1].get();
MemoryLocation* loc2 = locationMap[ptr2].get();
AliasType aliasType = analyzeAliasBetween(loc1, loc2);
CurrentResult->addAliasRelation(ptr1, ptr2, aliasType);
}
}
}
void SysYAliasAnalysisPass::optimizeForSysY(Function* F) {
// SysY特化优化
applySysYConstraints(F);
optimizeParameterAnalysis(F);
optimizeArrayAccessAnalysis(F);
}
std::unique_ptr<MemoryLocation> SysYAliasAnalysisPass::createMemoryLocation(Value* ptr) {
Value* basePtr = getBasePointer(ptr);
auto location = std::make_unique<MemoryLocation>(basePtr, ptr);
// 分析内存类型和索引模式
analyzeMemoryType(location.get());
analyzeIndexPattern(location.get());
return location;
}
Value* SysYAliasAnalysisPass::getBasePointer(Value* ptr) {
// 递归剥离GEP指令找到真正的基指针
if (auto* gepInst = dynamic_cast<GetElementPtrInst*>(ptr)) {
return getBasePointer(gepInst->getBasePointer());
}
return ptr;
}
void SysYAliasAnalysisPass::analyzeMemoryType(MemoryLocation* location) {
Value* base = location->basePointer;
// 检查内存类型
if (dynamic_cast<AllocaInst*>(base)) {
location->isLocalArray = true;
} else if (dynamic_cast<Argument*>(base)) {
location->isFunctionParameter = true;
} else if (dynamic_cast<GlobalValue*>(base)) {
location->isGlobalArray = true;
}
}
void SysYAliasAnalysisPass::analyzeIndexPattern(MemoryLocation* location) {
// 分析GEP指令的索引模式
if (auto* gepInst = dynamic_cast<GetElementPtrInst*>(location->accessPointer)) {
// 收集所有索引
for (unsigned i = 0; i < gepInst->getNumIndices(); ++i) {
Value* index = gepInst->getIndex(i);
location->indices.push_back(index);
// 检查是否为常量索引
if (!isConstantValue(index)) {
location->hasConstantIndices = false;
}
}
// 如果没有非常量索引,则为常量访问
if (location->indices.empty()) {
location->hasConstantIndices = true;
}
// 检查是否包含循环变量
Function* containingFunc = nullptr;
if (auto* inst = dynamic_cast<Instruction*>(location->basePointer)) {
containingFunc = inst->getParent()->getParent();
} else if (auto* arg = dynamic_cast<Argument*>(location->basePointer)) {
containingFunc = arg->getParent();
}
if (containingFunc) {
location->hasLoopVariableIndex = hasLoopVariableInIndices(location->indices, containingFunc);
}
// 计算常量偏移
if (location->hasConstantIndices) {
location->constantOffset = calculateConstantOffset(location->indices);
}
}
}
AliasType SysYAliasAnalysisPass::analyzeAliasBetween(MemoryLocation* loc1, MemoryLocation* loc2) {
// 分析两个内存位置之间的别名关系
// 1. 相同基指针的自别名
if (loc1->basePointer == loc2->basePointer) {
return AliasType::SELF_ALIAS;
}
// 2. 不同类型的内存位置
if ((loc1->isLocalArray && loc2->isLocalArray)) {
return compareLocalArrays(loc1, loc2);
}
if ((loc1->isFunctionParameter && loc2->isFunctionParameter)) {
return compareParameters(loc1, loc2);
}
if ((loc1->isGlobalArray || loc2->isGlobalArray)) {
return compareWithGlobal(loc1, loc2);
}
return compareMixedTypes(loc1, loc2);
}
AliasType SysYAliasAnalysisPass::compareLocalArrays(MemoryLocation* loc1, MemoryLocation* loc2) {
// 不同局部数组不别名
return AliasType::NO_ALIAS;
}
AliasType SysYAliasAnalysisPass::compareParameters(MemoryLocation* loc1, MemoryLocation* loc2) {
// 不同函数参数可能别名
return AliasType::POSSIBLE_ALIAS;
}
AliasType SysYAliasAnalysisPass::compareWithGlobal(MemoryLocation* loc1, MemoryLocation* loc2) {
// 涉及全局数组的访问,保守估计
return AliasType::POSSIBLE_ALIAS;
}
AliasType SysYAliasAnalysisPass::compareMixedTypes(MemoryLocation* loc1, MemoryLocation* loc2) {
// 混合类型访问,保守估计
return AliasType::UNKNOWN_ALIAS;
}
void SysYAliasAnalysisPass::applySysYConstraints(Function* F) {
// 应用SysY语言特定的约束
// SysY没有指针运算简化别名分析
}
void SysYAliasAnalysisPass::optimizeParameterAnalysis(Function* F) {
// 优化参数别名分析
}
void SysYAliasAnalysisPass::optimizeArrayAccessAnalysis(Function* F) {
// 优化数组访问别名分析
}
bool SysYAliasAnalysisPass::isConstantValue(Value* val) {
return dynamic_cast<ConstantInteger*>(val) != nullptr; // 简化,只检查整数常量
}
bool SysYAliasAnalysisPass::hasLoopVariableInIndices(const std::vector<Value*>& indices, Function* F) {
// 简化版本:检查索引是否包含循环变量
for (Value* index : indices) {
if (!isConstantValue(index)) {
return true; // 保守估计
}
}
return false;
}
int SysYAliasAnalysisPass::calculateConstantOffset(const std::vector<Value*>& indices) {
int offset = 0;
for (Value* index : indices) {
if (auto* constInt = dynamic_cast<ConstantInteger*>(index)) {
// ConstantInteger的getVal()返回variant需要提取int值
auto val = constInt->getVal();
if (std::holds_alternative<int>(val)) {
offset += std::get<int>(val);
}
}
}
return offset;
}
void SysYAliasAnalysisPass::printStatistics() const {
if (CurrentResult) {
CurrentResult->print();
}
}
} // namespace sysy

View File

@ -290,16 +290,14 @@ bool LoopAnalysisPass::runOnFunction(Function *F, AnalysisManager &AM) {
}
}
// 3. 计算循环层级 (Level) 并填充最外层/最内层循环列表
// 3. 计算循环层级 (Level)
std::queue<Loop *> q_level;
// 查找所有最外层循环没有父循环的设置其层级为0并加入队列
CurrentResult->clearOutermostLoops(); // 清空最外层循环列表
for (const auto &loop_ptr : allLoops) {
if (loop_ptr->isOutermost()) {
loop_ptr->setLoopLevel(0);
q_level.push(loop_ptr.get());
CurrentResult->addOutermostLoop(loop_ptr.get());
}
}
@ -314,12 +312,24 @@ bool LoopAnalysisPass::runOnFunction(Function *F, AnalysisManager &AM) {
}
}
// 填充最内层循环列表
CurrentResult->clearInnermostLoops(); // 清空最内层循环列表
for (const auto &loop_ptr : allLoops) {
if (loop_ptr->isInnermost()) {
CurrentResult->addInnermostLoop(loop_ptr.get());
if (DEBUG) {
std::cout << "Loop Analysis completed for function: " << F->getName() << std::endl;
std::cout << "Total loops found: " << CurrentResult->getLoopCount() << std::endl;
std::cout << "Max loop depth: " << CurrentResult->getMaxLoopDepth() << std::endl;
std::cout << "Innermost loops: " << CurrentResult->getInnermostLoops().size() << std::endl;
std::cout << "Outermost loops: " << CurrentResult->getOutermostLoops().size() << std::endl;
// 打印各深度的循环分布
for (int depth = 1; depth <= CurrentResult->getMaxLoopDepth(); ++depth) {
int count = CurrentResult->getLoopCountAtDepth(depth);
if (count > 0) {
std::cout << "Loops at depth " << depth << ": " << count << std::endl;
}
}
// 输出缓存统计
auto cacheStats = CurrentResult->getCacheStats();
std::cout << "Cache statistics - Total cached queries: " << cacheStats.totalCachedQueries << std::endl;
}
return changed;

View File

@ -0,0 +1,454 @@
#include "LoopCharacteristics.h"
#include "Dom.h"
#include "Loop.h"
#include "Liveness.h"
#include <iostream>
#include <cmath>
// 使用全局调试开关
extern int DEBUG;
namespace sysy {
// 定义 Pass 的唯一 ID
void *LoopCharacteristicsPass::ID = (void *)&LoopCharacteristicsPass::ID;
void LoopCharacteristicsResult::print() const {
if (!DEBUG) return; // 只有在 DEBUG 模式下才打印
std::cout << "\n--- Loop Characteristics Analysis Results for Function: "
<< AssociatedFunction->getName() << " ---" << std::endl;
if (CharacteristicsMap.empty()) {
std::cout << " No loop characteristics found." << std::endl;
return;
}
// 打印统计信息
auto stats = getOptimizationStats();
std::cout << "\n=== Optimization Statistics ===" << std::endl;
std::cout << "Total Loops: " << stats.totalLoops << std::endl;
std::cout << "Counting Loops: " << stats.countingLoops << std::endl;
std::cout << "Vectorizable Loops: " << stats.vectorizableLoops << std::endl;
std::cout << "Unroll Candidates: " << stats.unrollCandidates << std::endl;
std::cout << "Parallelizable Loops: " << stats.parallelizableLoops << std::endl;
std::cout << "Static Bound Loops: " << stats.staticBoundLoops << std::endl;
std::cout << "Avg Instructions per Loop: " << stats.avgInstructionCount << std::endl;
std::cout << "Avg Compute/Memory Ratio: " << stats.avgComputeMemoryRatio << std::endl;
// 按热度排序并打印循环特征
auto loopsByHotness = getLoopsByHotness();
std::cout << "\n=== Loop Characteristics (by hotness) ===" << std::endl;
for (auto* loop : loopsByHotness) {
auto* chars = getCharacteristics(loop);
if (!chars) continue;
std::cout << "\n--- Loop: " << loop->getName() << " (Hotness: "
<< loop->getLoopHotness() << ") ---" << std::endl;
std::cout << " Level: " << loop->getLoopLevel() << std::endl;
std::cout << " Blocks: " << loop->getLoopSize() << std::endl;
std::cout << " Instructions: " << chars->instructionCount << std::endl;
std::cout << " Memory Operations: " << chars->memoryOperationCount << std::endl;
std::cout << " Compute/Memory Ratio: " << chars->computeToMemoryRatio << std::endl;
// 循环形式
std::cout << " Form: ";
if (chars->isCountingLoop) std::cout << "Counting ";
if (chars->isSimpleForLoop) std::cout << "SimpleFor ";
if (chars->isInnermost) std::cout << "Innermost ";
if (chars->hasComplexControlFlow) std::cout << "Complex ";
std::cout << std::endl;
// 边界信息
if (chars->staticTripCount.has_value()) {
std::cout << " Static Trip Count: " << *chars->staticTripCount << std::endl;
}
if (chars->hasKnownBounds) {
std::cout << " Has Known Bounds: Yes" << std::endl;
}
// 优化机会
std::cout << " Optimization Opportunities: ";
if (chars->benefitsFromUnrolling)
std::cout << "Unroll(factor=" << chars->suggestedUnrollFactor << ") ";
if (chars->benefitsFromVectorization) std::cout << "Vectorize ";
if (chars->benefitsFromTiling) std::cout << "Tile ";
if (chars->isParallel) std::cout << "Parallelize ";
std::cout << std::endl;
// 归纳变量
if (!chars->basicInductionVars.empty()) {
std::cout << " Basic Induction Vars: " << chars->basicInductionVars.size() << std::endl;
}
if (!chars->derivedInductionVars.empty()) {
std::cout << " Derived Induction Vars: " << chars->derivedInductionVars.size() << std::endl;
}
// 循环不变量
if (!chars->loopInvariants.empty()) {
std::cout << " Loop Invariants: " << chars->loopInvariants.size() << std::endl;
}
if (!chars->invariantInsts.empty()) {
std::cout << " Hoistable Instructions: " << chars->invariantInsts.size() << std::endl;
}
}
std::cout << "-----------------------------------------------" << std::endl;
}
bool LoopCharacteristicsPass::runOnFunction(Function *F, AnalysisManager &AM) {
if (F->getBasicBlocks().empty()) {
CurrentResult = std::make_unique<LoopCharacteristicsResult>(F);
return false; // 空函数
}
if (DEBUG)
std::cout << "Running LoopCharacteristicsPass on function: " << F->getName() << std::endl;
// 获取循环分析结果 - 这是我们的核心依赖
auto* loopAnalysisResult = AM.getAnalysisResult<LoopAnalysisResult, LoopAnalysisPass>(F);
if (!loopAnalysisResult) {
std::cerr << "Error: LoopAnalysisResult not available for function " << F->getName() << std::endl;
CurrentResult = std::make_unique<LoopCharacteristicsResult>(F);
return false;
}
// 如果没有循环,直接返回
if (!loopAnalysisResult->hasLoops()) {
CurrentResult = std::make_unique<LoopCharacteristicsResult>(F);
return false;
}
CurrentResult = std::make_unique<LoopCharacteristicsResult>(F);
// 分析每个循环的特征
for (const auto& loop_ptr : loopAnalysisResult->getAllLoops()) {
Loop* loop = loop_ptr.get();
auto characteristics = std::make_unique<LoopCharacteristics>(loop);
// 执行各种特征分析
analyzeLoop(loop, characteristics.get(), AM);
// 添加到结果中
CurrentResult->addLoopCharacteristics(std::move(characteristics));
}
if (DEBUG) {
std::cout << "LoopCharacteristicsPass completed for function: " << F->getName() << std::endl;
auto stats = CurrentResult->getOptimizationStats();
std::cout << "Analyzed " << stats.totalLoops << " loops, found "
<< stats.vectorizableLoops << " vectorizable, "
<< stats.unrollCandidates << " unrollable" << std::endl;
}
return false; // 特征分析不修改IR
}
void LoopCharacteristicsPass::analyzeLoop(Loop* loop, LoopCharacteristics* characteristics, AnalysisManager &AM) {
if (DEBUG)
std::cout << " Analyzing characteristics of loop: " << loop->getName() << std::endl;
// 按顺序执行各种分析
computePerformanceMetrics(loop, characteristics);
analyzeLoopForm(loop, characteristics);
identifyInductionVariables(loop, characteristics);
identifyLoopInvariants(loop, characteristics);
analyzeLoopBounds(loop, characteristics);
analyzeMemoryAccessPatterns(loop, characteristics);
evaluateOptimizationOpportunities(loop, characteristics);
}
void LoopCharacteristicsPass::computePerformanceMetrics(Loop* loop, LoopCharacteristics* characteristics) {
size_t totalInsts = 0;
size_t memoryOps = 0;
size_t arithmeticOps = 0;
// 遍历循环中的所有指令
for (BasicBlock* bb : loop->getBlocks()) {
for (auto& inst : bb->getInstructions()) {
totalInsts++;
// 分类指令类型
if (dynamic_cast<LoadInst*>(inst.get()) || dynamic_cast<StoreInst*>(inst.get())) {
memoryOps++;
} else if (dynamic_cast<BinaryInst*>(inst.get())) {
// 检查是否为算术运算
auto* binInst = dynamic_cast<BinaryInst*>(inst.get());
// 简化:假设所有二元运算都是算术运算
arithmeticOps++;
}
}
}
characteristics->instructionCount = totalInsts;
characteristics->memoryOperationCount = memoryOps;
characteristics->arithmeticOperationCount = arithmeticOps;
// 计算计算与内存操作比率
if (memoryOps > 0) {
characteristics->computeToMemoryRatio = static_cast<double>(arithmeticOps) / memoryOps;
} else {
characteristics->computeToMemoryRatio = arithmeticOps; // 纯计算循环
}
}
void LoopCharacteristicsPass::analyzeLoopForm(Loop* loop, LoopCharacteristics* characteristics) {
// 基本形式判断
characteristics->isInnermost = loop->isInnermost();
// 检查是否为简单循环 (只有一个回边)
bool isSimple = loop->isSimpleLoop();
characteristics->isSimpleForLoop = isSimple;
// 检查复杂控制流 (多个出口表示可能有break/continue)
auto exitingBlocks = loop->getExitingBlocks();
characteristics->hasComplexControlFlow = exitingBlocks.size() > 1;
// 初步判断是否为计数循环 (需要更复杂的分析)
// 简化版本:如果是简单循环且是最内层,很可能是计数循环
characteristics->isCountingLoop = isSimple && loop->isInnermost() && exitingBlocks.size() == 1;
}
void LoopCharacteristicsPass::identifyInductionVariables(Loop* loop, LoopCharacteristics* characteristics) {
// 寻找基本归纳变量
BasicBlock* header = loop->getHeader();
// 遍历循环头的phi指令寻找归纳变量模式
for (auto& inst : header->getInstructions()) {
auto* phiInst = dynamic_cast<PhiInst*>(inst.get());
if (!phiInst) continue;
// 检查phi指令是否符合归纳变量模式
if (isInductionVariable(phiInst, loop)) {
characteristics->basicInductionVars.push_back(phiInst);
// 分析步长 (简化版本)
characteristics->inductionSteps[phiInst] = 1; // 默认步长为1
if (DEBUG)
std::cout << " Found basic induction variable: " << phiInst->getName() << std::endl;
}
}
// 寻找派生归纳变量 (基于基本归纳变量的线性表达式)
for (BasicBlock* bb : loop->getBlocks()) {
for (auto& inst : bb->getInstructions()) {
// 检查是否为基于归纳变量的计算
if (auto* binInst = dynamic_cast<BinaryInst*>(inst.get())) {
// 简化:检查操作数是否包含基本归纳变量
for (Value* basicIV : characteristics->basicInductionVars) {
// 这里需要更复杂的分析来确定派生关系
// 暂时简化处理
}
}
}
}
}
void LoopCharacteristicsPass::identifyLoopInvariants(Loop* loop, LoopCharacteristics* characteristics) {
// 收集循环不变量
for (BasicBlock* bb : loop->getBlocks()) {
for (auto& inst : bb->getInstructions()) {
Value* val = inst.get();
// 跳过phi指令和终结指令
if (dynamic_cast<PhiInst*>(val)) {
continue;
}
// 检查是否为终结指令
if (auto* instPtr = dynamic_cast<Instruction*>(val)) {
if (instPtr->isTerminator()) {
continue;
}
}
if (isLoopInvariant(val, loop)) {
characteristics->loopInvariants.insert(val);
characteristics->invariantInsts.insert(static_cast<Instruction*>(val));
if (DEBUG)
std::cout << " Found loop invariant: " << val->getName() << std::endl;
}
}
}
}
void LoopCharacteristicsPass::analyzeLoopBounds(Loop* loop, LoopCharacteristics* characteristics) {
// 简化的边界分析
// 在实际实现中,需要分析循环的条件表达式来确定边界
// 检查是否有静态可确定的循环次数
if (characteristics->isCountingLoop && !characteristics->basicInductionVars.empty()) {
// 简化:如果是计数循环且有基本归纳变量,尝试确定循环次数
// 这里需要更复杂的符号执行或约束求解
// 暂时设置一个保守估计
if (characteristics->instructionCount < 10) {
characteristics->staticTripCount = 100; // 假设小循环执行100次
characteristics->hasKnownBounds = true;
}
}
}
void LoopCharacteristicsPass::analyzeMemoryAccessPatterns(Loop* loop, LoopCharacteristics* characteristics) {
// 使用外部别名分析结果 - 大幅简化版本
std::map<Value*, std::vector<Instruction*>> accessMap;
// 收集所有内存访问
for (BasicBlock* bb : loop->getBlocks()) {
for (auto& inst : bb->getInstructions()) {
if (auto* loadInst = dynamic_cast<LoadInst*>(inst.get())) {
Value* ptr = loadInst->getPointer();
accessMap[ptr].push_back(loadInst);
} else if (auto* storeInst = dynamic_cast<StoreInst*>(inst.get())) {
Value* ptr = storeInst->getPointer();
accessMap[ptr].push_back(storeInst);
}
}
}
// 分析每个内存位置的访问模式
for (auto& [ptr, accesses] : accessMap) {
LoopCharacteristics::MemoryAccessPattern pattern;
// 初始化基本字段
pattern.isSequential = true; // 简化:假设大部分访问是顺序的
pattern.isStrided = false;
pattern.stride = 1;
// 使用别名分析结果 (简化:设置默认值,实际应该查询别名分析)
pattern.aliasType = AliasType::UNKNOWN_ALIAS; // 保守默认值
pattern.isArrayParameter = false;
pattern.isGlobalArray = false;
pattern.hasConstantIndices = true;
// 分类load和store
for (Instruction* inst : accesses) {
if (dynamic_cast<LoadInst*>(inst)) {
pattern.loadInsts.push_back(inst);
} else {
pattern.storeInsts.push_back(inst);
}
}
characteristics->memoryPatterns[ptr] = pattern;
if (DEBUG && (static_cast<int>(pattern.aliasType) >= 2)) { // POSSIBLE_ALIAS及以上
std::cout << " Found potential aliasing for memory access, type: "
<< static_cast<int>(pattern.aliasType) << std::endl;
}
}
}
void LoopCharacteristicsPass::evaluateOptimizationOpportunities(Loop* loop, LoopCharacteristics* characteristics) {
// 评估循环展开机会
characteristics->benefitsFromUnrolling =
characteristics->isInnermost &&
characteristics->instructionCount > 3 &&
characteristics->instructionCount < 50 &&
!characteristics->hasComplexControlFlow;
if (characteristics->benefitsFromUnrolling) {
characteristics->suggestedUnrollFactor = estimateUnrollFactor(loop);
}
// 评估向量化机会
characteristics->benefitsFromVectorization = benefitsFromVectorization(loop);
// 评估并行化机会
characteristics->isParallel =
!hasLoopCarriedDependence(loop) &&
characteristics->isCountingLoop;
// 评估分块机会 (主要针对嵌套循环)
characteristics->benefitsFromTiling =
!loop->isInnermost() &&
characteristics->memoryOperationCount > characteristics->arithmeticOperationCount;
}
// ========== 辅助方法实现 ==========
bool LoopCharacteristicsPass::isInductionVariable(Value* val, Loop* loop) {
// 简化的归纳变量检测
auto* phiInst = dynamic_cast<PhiInst*>(val);
if (!phiInst) return false;
// 检查phi指令是否在循环头
if (phiInst->getParent() != loop->getHeader()) return false;
// 检查是否有来自循环内的更新
for (auto& [incomingBB, incomingVal] : phiInst->getIncomingValues()) {
if (loop->contains(incomingBB)) {
// 简化:如果有来自循环内的值,认为可能是归纳变量
return true;
}
}
return false;
}
bool LoopCharacteristicsPass::isLoopInvariant(Value* val, Loop* loop) {
auto* inst = dynamic_cast<Instruction*>(val);
if (!inst) return true; // 非指令(如常量)认为是不变的
// 如果指令不在循环内定义,则是不变的
if (!loop->contains(inst->getParent())) {
return true;
}
// 检查操作数是否都是循环不变的
// 简化版本如果是load指令且指针是不变的认为可能是不变的
if (auto* loadInst = dynamic_cast<LoadInst*>(inst)) {
Value* ptr = loadInst->getPointer();
return isLoopInvariant(ptr, loop);
}
// 简化:对于其他指令,保守地认为是变化的
return false;
}
bool LoopCharacteristicsPass::hasLoopCarriedDependence(Loop* loop) {
// 简化的依赖分析
// 检查是否有写后读或写后写依赖跨越循环迭代
std::set<Value*> writtenVars;
std::set<Value*> readVars;
for (BasicBlock* bb : loop->getBlocks()) {
for (auto& inst : bb->getInstructions()) {
if (auto* storeInst = dynamic_cast<StoreInst*>(inst.get())) {
writtenVars.insert(storeInst->getPointer());
} else if (auto* loadInst = dynamic_cast<LoadInst*>(inst.get())) {
readVars.insert(loadInst->getPointer());
}
}
}
// 简化:如果有写后读到同一变量,假设存在依赖
for (Value* written : writtenVars) {
if (readVars.count(written)) {
return true; // 可能存在依赖
}
}
return false; // 保守估计:没有明显依赖
}
int LoopCharacteristicsPass::estimateUnrollFactor(Loop* loop) {
// 基于循环体大小估算展开因子
if (loop->getLoopSize() <= 2) return 8; // 很小的循环
if (loop->getLoopSize() <= 5) return 4; // 小循环
if (loop->getLoopSize() <= 10) return 2; // 中等循环
return 1; // 大循环不建议展开
}
bool LoopCharacteristicsPass::benefitsFromVectorization(Loop* loop) {
// 简化的向量化收益评估
return loop->isInnermost() && // 最内层循环
loop->isSimpleLoop() && // 简单循环结构
!hasLoopCarriedDependence(loop); // 没有明显的依赖
}
} // namespace sysy

View File

@ -0,0 +1,281 @@
#include "SideEffectAnalysis.h"
#include "AliasAnalysis.h"
#include "SysYIRPrinter.h"
#include <iostream>
namespace sysy {
// 副作用分析遍的静态 ID
void* SysYSideEffectAnalysisPass::ID = (void*)&SysYSideEffectAnalysisPass::ID;
// ======================================================================
// SideEffectAnalysisResult 类的实现
// ======================================================================
SideEffectAnalysisResult::SideEffectAnalysisResult() {
initializeKnownFunctions();
}
const SideEffectInfo& SideEffectAnalysisResult::getInstructionSideEffect(Instruction* inst) const {
auto it = instructionSideEffects.find(inst);
if (it != instructionSideEffects.end()) {
return it->second;
}
// 返回默认的无副作用信息
static SideEffectInfo noEffect;
return noEffect;
}
const SideEffectInfo& SideEffectAnalysisResult::getFunctionSideEffect(Function* func) const {
auto it = functionSideEffects.find(func);
if (it != functionSideEffects.end()) {
return it->second;
}
// 返回默认的无副作用信息
static SideEffectInfo noEffect;
return noEffect;
}
void SideEffectAnalysisResult::setInstructionSideEffect(Instruction* inst, const SideEffectInfo& info) {
instructionSideEffects[inst] = info;
}
void SideEffectAnalysisResult::setFunctionSideEffect(Function* func, const SideEffectInfo& info) {
functionSideEffects[func] = info;
}
bool SideEffectAnalysisResult::hasSideEffect(Instruction* inst) const {
const auto& info = getInstructionSideEffect(inst);
return info.type != SideEffectType::NO_SIDE_EFFECT;
}
bool SideEffectAnalysisResult::mayModifyMemory(Instruction* inst) const {
const auto& info = getInstructionSideEffect(inst);
return info.mayModifyMemory;
}
bool SideEffectAnalysisResult::mayModifyGlobal(Instruction* inst) const {
const auto& info = getInstructionSideEffect(inst);
return info.mayModifyGlobal;
}
bool SideEffectAnalysisResult::isPureFunction(Function* func) const {
const auto& info = getFunctionSideEffect(func);
return info.isPure;
}
void SideEffectAnalysisResult::initializeKnownFunctions() {
// SysY标准库函数的副作用信息
// I/O函数 - 有副作用
SideEffectInfo ioEffect;
ioEffect.type = SideEffectType::IO_OPERATION;
ioEffect.mayModifyGlobal = true;
ioEffect.mayModifyMemory = true;
ioEffect.mayCallFunction = true;
ioEffect.isPure = false;
// knownFunctions["printf"] = ioEffect;
// knownFunctions["scanf"] = ioEffect;
knownFunctions["getint"] = ioEffect;
knownFunctions["getch"] = ioEffect;
knownFunctions["getfloat"] = ioEffect;
knownFunctions["getarray"] = ioEffect;
knownFunctions["getfarray"] = ioEffect;
knownFunctions["putint"] = ioEffect;
knownFunctions["putch"] = ioEffect;
knownFunctions["putfloat"] = ioEffect;
knownFunctions["putarray"] = ioEffect;
knownFunctions["putfarray"] = ioEffect;
// 时间函数 - 有副作用
SideEffectInfo timeEffect;
timeEffect.type = SideEffectType::FUNCTION_CALL;
timeEffect.mayModifyGlobal = true;
timeEffect.mayModifyMemory = false;
timeEffect.mayCallFunction = true;
timeEffect.isPure = false;
knownFunctions["_sysy_starttime"] = timeEffect;
knownFunctions["_sysy_stoptime"] = timeEffect;
}
const SideEffectInfo* SideEffectAnalysisResult::getKnownFunctionSideEffect(const std::string& funcName) const {
auto it = knownFunctions.find(funcName);
return (it != knownFunctions.end()) ? &it->second : nullptr;
}
// ======================================================================
// SysYSideEffectAnalysisPass 类的实现
// ======================================================================
bool SysYSideEffectAnalysisPass::runOnFunction(Function* F, AnalysisManager& AM) {
if (DEBUG) {
std::cout << "Running SideEffect analysis on function: " << F->getName() << std::endl;
}
// 创建分析结果构造函数中已经调用了initializeKnownFunctions
result = std::make_unique<SideEffectAnalysisResult>();
// 获取别名分析结果,在整个函数分析过程中重复使用
aliasAnalysis = AM.getAnalysisResult<AliasAnalysisResult, SysYAliasAnalysisPass>(F);
// 分析函数中的每条指令
SideEffectInfo functionSideEffect;
for (auto& BB : F->getBasicBlocks()) {
for (auto& I : BB->getInstructions_Range()) {
Instruction* inst = I.get();
SideEffectInfo instEffect = analyzeInstruction(inst, AM);
// 记录指令的副作用信息
result->setInstructionSideEffect(inst, instEffect);
// 合并到函数级别的副作用信息中
functionSideEffect = functionSideEffect.merge(instEffect);
}
}
// 记录函数级别的副作用信息
result->setFunctionSideEffect(F, functionSideEffect);
if (DEBUG) {
std::cout << "---- Side Effect Analysis Results for Function: " << F->getName() << " ----\n";
for (auto& BB : F->getBasicBlocks()) {
for (auto& I : BB->getInstructions_Range()) {
Instruction* inst = I.get();
const auto& info = result->getInstructionSideEffect(inst);
SysYPrinter::printInst(inst);
std::cout << " -> Side Effect: ";
switch (info.type) {
case SideEffectType::NO_SIDE_EFFECT: std::cout << "None"; break;
case SideEffectType::MEMORY_WRITE: std::cout << "Memory Write"; break;
case SideEffectType::FUNCTION_CALL: std::cout << "Function Call"; break;
case SideEffectType::IO_OPERATION: std::cout << "I/O Operation"; break;
case SideEffectType::UNKNOWN: std::cout << "Unknown"; break;
}
std::cout << " (Modifies Global: " << (info.mayModifyGlobal ? "Yes" : "No")
<< ", Modifies Memory: " << (info.mayModifyMemory ? "Yes" : "No")
<< ", Is Pure: " << (info.isPure ? "Yes" : "No") << ")\n";
}
}
std::cout << "------------------------------------------------------------------\n";
}
return false; // Analysis passes return false since they don't modify the IR
}
std::unique_ptr<AnalysisResultBase> SysYSideEffectAnalysisPass::getResult() {
return std::move(result);
}
SideEffectInfo SysYSideEffectAnalysisPass::analyzeInstruction(Instruction* inst, AnalysisManager& AM) {
SideEffectInfo info;
// 根据指令类型进行分析
if (inst->isCall()) {
return analyzeCallInstruction(static_cast<CallInst*>(inst), AM);
} else if (inst->isStore()) {
return analyzeStoreInstruction(static_cast<StoreInst*>(inst), AM);
} else if (inst->isMemset()) {
return analyzeMemsetInstruction(static_cast<MemsetInst*>(inst), AM);
} else if (inst->isBranch() || inst->isReturn()) {
// 控制流指令无副作用,但必须保留
info.type = SideEffectType::NO_SIDE_EFFECT;
info.isPure = true;
} else {
// 其他指令(算术、逻辑、比较等)通常无副作用
info.type = SideEffectType::NO_SIDE_EFFECT;
info.isPure = true;
}
return info;
}
SideEffectInfo SysYSideEffectAnalysisPass::analyzeCallInstruction(CallInst* call, AnalysisManager& AM) {
SideEffectInfo info;
// 获取被调用的函数
Function* calledFunc = call->getCallee();
if (!calledFunc) {
// 间接调用,保守处理
info.type = SideEffectType::UNKNOWN;
info.mayModifyGlobal = true;
info.mayModifyMemory = true;
info.mayCallFunction = true;
info.isPure = false;
return info;
}
std::string funcName = calledFunc->getName();
// 检查是否为已知的标准库函数
const SideEffectInfo* knownInfo = result->getKnownFunctionSideEffect(funcName);
if (knownInfo) {
return *knownInfo;
}
// 对于用户定义的函数,检查是否已经分析过
const SideEffectInfo& funcEffect = result->getFunctionSideEffect(calledFunc);
if (funcEffect.type != SideEffectType::NO_SIDE_EFFECT || !funcEffect.isPure) {
return funcEffect;
}
// 对于未分析的用户函数,保守处理
info.type = SideEffectType::FUNCTION_CALL;
info.mayModifyGlobal = true;
info.mayModifyMemory = true;
info.mayCallFunction = true;
info.isPure = false;
return info;
}
SideEffectInfo SysYSideEffectAnalysisPass::analyzeStoreInstruction(StoreInst* store, AnalysisManager& AM) {
SideEffectInfo info;
info.type = SideEffectType::MEMORY_WRITE;
info.mayModifyMemory = true;
info.isPure = false;
// 使用缓存的别名分析结果
if (aliasAnalysis) {
Value* storePtr = store->getPointer();
// 如果存储到全局变量或可能别名的位置,则可能修改全局状态
if (!aliasAnalysis->isLocalArray(storePtr)) {
info.mayModifyGlobal = true;
}
} else {
// 没有别名分析结果,保守处理
info.mayModifyGlobal = true;
}
return info;
}
SideEffectInfo SysYSideEffectAnalysisPass::analyzeMemsetInstruction(MemsetInst* memset, AnalysisManager& AM) {
SideEffectInfo info;
info.type = SideEffectType::MEMORY_WRITE;
info.mayModifyMemory = true;
info.isPure = false;
// 使用缓存的别名分析结果
if (aliasAnalysis) {
Value* memsetPtr = memset->getPointer();
// 如果memset操作全局变量或可能别名的位置则可能修改全局状态
if (!aliasAnalysis->isLocalArray(memsetPtr)) {
info.mayModifyGlobal = true;
}
} else {
// 没有别名分析结果,保守处理
info.mayModifyGlobal = true;
}
return info;
}
} // namespace sysy

View File

@ -1,6 +1,9 @@
#include "Dom.h"
#include "Liveness.h"
#include "Loop.h"
#include "LoopCharacteristics.h"
#include "AliasAnalysis.h"
#include "SideEffectAnalysis.h"
#include "SysYIRCFGOpt.h"
#include "SysYIRPrinter.h"
#include "DCE.h"
@ -38,7 +41,10 @@ void PassManager::runOptimizationPipeline(Module* moduleIR, IRBuilder* builderIR
// 注册分析遍
registerAnalysisPass<sysy::DominatorTreeAnalysisPass>();
registerAnalysisPass<sysy::LivenessAnalysisPass>();
registerAnalysisPass<SysYAliasAnalysisPass>(); // 别名分析 (优先级高)
registerAnalysisPass<SysYSideEffectAnalysisPass>(); // 副作用分析 (依赖别名分析)
registerAnalysisPass<LoopAnalysisPass>();
registerAnalysisPass<LoopCharacteristicsPass>(); // 循环特征分析依赖别名分析
// 注册优化遍
registerOptimizationPass<SysYDelInstAfterBrPass>();