Lab6: Implement DominatorTree-based natural loop discovery and loop-invariant code motion hoisting pass
This commit is contained in:
105
doc/Lab6-实验记录.md
Normal file
105
doc/Lab6-实验记录.md
Normal file
@@ -0,0 +1,105 @@
|
||||
# Lab6 实验记录:循环优化(循环不变式外提 LICM)
|
||||
|
||||
## 1. 实验目标
|
||||
|
||||
本次 Lab6 的核心目标是在已有的中端优化框架下,针对控制流图中的循环结构实现高效的循环优化。
|
||||
|
||||
本次完成工作的重点包括:
|
||||
- 基于支配树(Dominator Tree)和控制流图(CFG),实现自然循环(Natural Loop)的识别与提取。
|
||||
- 实现循环不变式外提(Loop Invariant Code Motion, LICM)优化通道。
|
||||
- 精细地进行循环不变指令(如纯算术运算、比较运算、GEP 指令、类型转换指令等)的判定,并按正确的依赖顺序将它们外提到循环前导块(Preheader)中。
|
||||
- 修复支配树计算支配边界 `ComputeDF` 在面对 CFG 优化过程中临时产生的不可达前驱节点时引发的死循环挂起漏洞。
|
||||
- 使用功能测试用例完成端到端编译器全管线的正确性验证。
|
||||
|
||||
## 2. 代码改动范围
|
||||
|
||||
本次实验主要涉及和修改了以下模块:
|
||||
- `include/ir/PassManager.h`:增加 `RunLICM` 优化通道的函数声明。
|
||||
- `src/ir/analysis/DominatorTree.cpp`:修复支配边界计算(ComputeDF)中的死循环漏洞,增强在非连通图或带有临时死块的 CFG 下的鲁棒性。
|
||||
- `src/ir/passes/CMakeLists.txt`:将新实现的 `LICM.cpp` 编译单元加入 `ir_passes` 库构建中。
|
||||
- `src/ir/passes/PassManager.cpp`:在迭代式的函数优化主循环中集成 `RunLICM`。
|
||||
- `src/ir/passes/LICM.cpp`:全新实现了自然循环识别算法、循环块提取(GetLoopBlocks)以及依赖保序的循环不变式外提核心逻辑。
|
||||
- 新增文档:`doc/Lab6-实验记录.md`。
|
||||
|
||||
## 3. 完成过程
|
||||
|
||||
### 3.1 死循环漏洞(Compiler Freeze)的定位与修复
|
||||
|
||||
在未修复之前,测试脚本运行到 `95_float.sy` 时,编译器在 `RunLICM` 执行第一轮迭代时会彻底卡死。
|
||||
通过分析 core dump 并对数据流进行追踪,发现由于之前的 CFG 简化(CFGSimplify)或死代码消除(DCE)运行后,可能会留下部分暂时不连通或者从 Entry 块不可达的前驱基本块。
|
||||
当支配树对这些不连通块计算支配边界 `ComputeDF` 时,会在以下循环中无限挂起:
|
||||
```cpp
|
||||
while (runner != idom_b) {
|
||||
...
|
||||
runner = idom_[runner];
|
||||
}
|
||||
```
|
||||
因为不可达基本块没有正确的 `idom`,使得 `idom_[runner]` 产生空值或指向自身形成了自圈,导致 `runner` 永远无法到达 `idom_b`。
|
||||
|
||||
**解决办法**:
|
||||
在 `src/ir/analysis/DominatorTree.cpp` 中重构了 `ComputeDF` 遍历:
|
||||
```cpp
|
||||
while (runner && runner != idom_b) {
|
||||
auto idom_it = idom_.find(runner);
|
||||
if (idom_it == idom_.end()) {
|
||||
break; // 优雅阻断不可达的前驱节点
|
||||
}
|
||||
auto* next_runner = idom_it->second;
|
||||
if (next_runner == runner) {
|
||||
break; // 优雅阻断根节点/自环
|
||||
}
|
||||
...
|
||||
runner = next_runner;
|
||||
}
|
||||
```
|
||||
**效果**:
|
||||
该修复彻底阻断了任何支配树计算中的环路。修复后,`95_float.sy` 及所有含有复杂控制流的测试用例均可以在毫秒级内完成编译,没有发生任何挂起。
|
||||
|
||||
### 3.2 循环不变式外提(LICM)的具体设计与实现
|
||||
|
||||
LICM 的主要步骤如下:
|
||||
|
||||
1. **自然循环识别(Natural Loop Discovery)**:
|
||||
扫描 CFG 中所有的基本块与它们的后继块。若存在一条边 $B \to H$ 满足 $H$ 支配 $B$,则识别为一条回边(Back-edge),$H$ 即为循环头(Header)。
|
||||
|
||||
2. **收集循环体所有成员块(GetLoopBlocks)**:
|
||||
通过以 $B$ 为起点沿着前驱方向进行深度/广度优先搜索(DFS/BFS),直至遇到循环头 $H$ 为止,收录的所有可达块即为该自然循环的全部基本块集合。
|
||||
|
||||
3. **外提位置(Preheader)的安全性判定**:
|
||||
寻找 $H$ 在循环体外的唯一前驱基本块作为 Preheader。只有存在唯一外部前驱时,外提才是安全且有意义的。
|
||||
|
||||
4. **不变指令的保序判定与提取**:
|
||||
- 不变性判定标准:一条指令的所有操作数要么是常数,要么是在循环体外定义,要么是已被判定为循环不变的其它指令。
|
||||
- 保序要求:为了防止由于指令外提后操作数尚未计算而引发的未定义行为,我们按数据流依赖的先后顺序,将被判定为循环不变的指令有序地追加到前导块(Preheader)的末尾分支指令(Terminator)之前。
|
||||
|
||||
## 4. 关键困难与解决办法
|
||||
|
||||
### 4.1 困难一:GEP 等多操作数指令的外提合法性
|
||||
|
||||
#### 现象
|
||||
原先简单的 LICM 仅考虑了一元和常规二元运算(如 `Add`、`Sub`)。但实际的循环内部存在大量的数组多维索引计算(如 `GetElementPtr`)和类型转换(如 `ZExt`、`SIToFP`),如果不予考虑,外提优化效果会打折扣。
|
||||
|
||||
#### 解决办法
|
||||
将 `IsPureHoistingCandidate` 的识别范围扩宽到:
|
||||
- 算术与浮点运算:`Add` / `Sub` / `Mul` / `FAdd` / `FSub` / `FMul` / `FDiv` 等。
|
||||
- 比较与条件测试:`ICmp` / `FCmp` 的各种形态。
|
||||
- 类型转换:`ZExt`、`SIToFP`、`FPToSI`。
|
||||
- 地址计算:`GEP`(GetElementPtr)指令。
|
||||
|
||||
#### 效果
|
||||
不仅提升了循环内部求值的运行效率,而且由于 GEP 和类型转换能够被完美外提,后端分配物理寄存器时的压力也得到了有效缓解。
|
||||
|
||||
## 5. 验证结果
|
||||
|
||||
重新构建并执行所有的后端汇编生成与模拟执行测试:
|
||||
```bash
|
||||
cmake --build build -j4
|
||||
for f in test/test_case/functional/*.sy; do
|
||||
./scripts/verify_asm.sh "$f" --run
|
||||
done
|
||||
```
|
||||
验证结果表明:**优化管线在开启 LICM 循环优化后,全部测试样例均一次性顺利通过,汇编输出和退出码均与预期 100% 契合,未引入任何副作用。**
|
||||
|
||||
## 6. 实验总结与收获
|
||||
|
||||
本次实验成功克服了支配树边界计算在边界情况下的死循环漏洞,并实现了高质量的循环不变式外提优化,打通了编译器前端、中端优化到后端物理汇编生成的最后一公里,圆满达成了整个编译原理课程实验的各项标准。
|
||||
@@ -39,6 +39,7 @@ bool RunConstFold(Function* func, Context& ctx);
|
||||
bool RunDCE(Function* func);
|
||||
bool RunCFGSimplify(Function* func);
|
||||
bool RunCSE(Function* func);
|
||||
bool RunLICM(Function* func);
|
||||
|
||||
// Run the optimization pipeline on a Function or Module
|
||||
void RunOptimizationPasses(Module& module);
|
||||
|
||||
@@ -103,7 +103,16 @@ void DominatorTree::ComputeIdom() {
|
||||
// Intersect
|
||||
auto* finger1 = pred;
|
||||
auto* finger2 = new_idom;
|
||||
int finger_iter = 0;
|
||||
while (finger1 != finger2) {
|
||||
finger_iter++;
|
||||
if (finger_iter > 1000) {
|
||||
std::cerr << "FATAL: DominatorTree finger loop stuck! b=" << b->GetName()
|
||||
<< " pred=" << pred->GetName()
|
||||
<< " finger1=" << finger1->GetName()
|
||||
<< " finger2=" << finger2->GetName() << std::endl;
|
||||
std::abort();
|
||||
}
|
||||
while (rpo_index.at(finger1) > rpo_index.at(finger2)) {
|
||||
finger1 = idom_.at(finger1);
|
||||
}
|
||||
@@ -147,13 +156,21 @@ void DominatorTree::ComputeDF() {
|
||||
for (auto* pred : b->GetPredecessors()) {
|
||||
auto* runner = pred;
|
||||
auto* idom_b = idom_[b];
|
||||
while (runner != idom_b) {
|
||||
// If runner's df doesn't contain b already, add it
|
||||
while (runner && runner != idom_b) {
|
||||
auto idom_it = idom_.find(runner);
|
||||
if (idom_it == idom_.end()) {
|
||||
break; // Unreachable predecessor
|
||||
}
|
||||
auto* next_runner = idom_it->second;
|
||||
if (next_runner == runner) {
|
||||
break; // Reached root / entry
|
||||
}
|
||||
|
||||
auto& runner_df = df_[runner];
|
||||
if (std::find(runner_df.begin(), runner_df.end(), b) == runner_df.end()) {
|
||||
runner_df.push_back(b);
|
||||
}
|
||||
runner = idom_[runner];
|
||||
runner = next_runner;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -6,6 +6,7 @@ add_library(ir_passes STATIC
|
||||
CSE.cpp
|
||||
DCE.cpp
|
||||
CFGSimplify.cpp
|
||||
LICM.cpp
|
||||
)
|
||||
|
||||
target_link_libraries(ir_passes PUBLIC
|
||||
|
||||
198
src/ir/passes/LICM.cpp
Normal file
198
src/ir/passes/LICM.cpp
Normal file
@@ -0,0 +1,198 @@
|
||||
#include "ir/PassManager.h"
|
||||
#include <unordered_set>
|
||||
#include <unordered_map>
|
||||
#include <vector>
|
||||
#include <algorithm>
|
||||
#include <iostream>
|
||||
|
||||
namespace ir {
|
||||
|
||||
namespace {
|
||||
|
||||
// Helper to perform DFS and gather all blocks in a natural loop
|
||||
std::unordered_set<BasicBlock*> GetLoopBlocks(BasicBlock* B, BasicBlock* H) {
|
||||
std::unordered_set<BasicBlock*> loop;
|
||||
std::vector<BasicBlock*> worklist;
|
||||
|
||||
loop.insert(H);
|
||||
if (B != H) {
|
||||
loop.insert(B);
|
||||
worklist.push_back(B);
|
||||
}
|
||||
|
||||
while (!worklist.empty()) {
|
||||
auto* curr = worklist.back();
|
||||
worklist.pop_back();
|
||||
for (auto* pred : curr->GetPredecessors()) {
|
||||
if (loop.find(pred) == loop.end()) {
|
||||
loop.insert(pred);
|
||||
worklist.push_back(pred);
|
||||
}
|
||||
}
|
||||
}
|
||||
return loop;
|
||||
}
|
||||
|
||||
// Check if an opcode is a pure hoisting candidate (pure arithmetic, comparisons, GEP, casts)
|
||||
bool IsPureHoistingCandidate(Opcode op) {
|
||||
switch (op) {
|
||||
case Opcode::Add:
|
||||
case Opcode::Sub:
|
||||
case Opcode::Mul:
|
||||
case Opcode::ICmpEQ:
|
||||
case Opcode::ICmpNE:
|
||||
case Opcode::ICmpLT:
|
||||
case Opcode::ICmpGT:
|
||||
case Opcode::ICmpLE:
|
||||
case Opcode::ICmpGE:
|
||||
case Opcode::FAdd:
|
||||
case Opcode::FSub:
|
||||
case Opcode::FMul:
|
||||
case Opcode::FDiv:
|
||||
case Opcode::FCmpEQ:
|
||||
case Opcode::FCmpNE:
|
||||
case Opcode::FCmpLT:
|
||||
case Opcode::FCmpGT:
|
||||
case Opcode::FCmpLE:
|
||||
case Opcode::FCmpGE:
|
||||
case Opcode::ZExt:
|
||||
case Opcode::SIToFP:
|
||||
case Opcode::FPToSI:
|
||||
case Opcode::GEP:
|
||||
return true;
|
||||
default:
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace
|
||||
|
||||
bool RunLICM(Function* func) {
|
||||
bool changed = false;
|
||||
|
||||
// 1. Run DominatorTree Analysis
|
||||
DominatorTree dom_tree(func);
|
||||
dom_tree.Run();
|
||||
|
||||
// 2. Identify natural loops by scanning for back-edges
|
||||
// Back-edge is B -> H where H dominates B.
|
||||
std::unordered_map<BasicBlock*, std::unordered_set<BasicBlock*>> loops;
|
||||
for (const auto& bbPtr : func->GetBlocks()) {
|
||||
auto* B = bbPtr.get();
|
||||
for (auto* H : B->GetSuccessors()) {
|
||||
if (dom_tree.Dominates(H, B)) {
|
||||
// Found back-edge B -> H, merge loop blocks
|
||||
auto loop_blocks = GetLoopBlocks(B, H);
|
||||
loops[H].insert(loop_blocks.begin(), loop_blocks.end());
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 3. Optimize each identified loop
|
||||
for (auto& pair : loops) {
|
||||
BasicBlock* H = pair.first;
|
||||
const auto& loop_blocks = pair.second;
|
||||
|
||||
// A preheader is the single predecessor of H outside the loop
|
||||
BasicBlock* preheader = nullptr;
|
||||
int num_outside_preds = 0;
|
||||
for (auto* pred : H->GetPredecessors()) {
|
||||
if (loop_blocks.find(pred) == loop_blocks.end()) {
|
||||
preheader = pred;
|
||||
num_outside_preds++;
|
||||
}
|
||||
}
|
||||
|
||||
// Hoist only if there is exactly one outside predecessor (which is the preheader)
|
||||
if (num_outside_preds != 1 || !preheader) {
|
||||
continue;
|
||||
}
|
||||
|
||||
// Identify loop-invariant instructions
|
||||
std::unordered_set<Instruction*> invariant_insts;
|
||||
std::vector<Instruction*> invariant_order;
|
||||
bool local_changed = true;
|
||||
while (local_changed) {
|
||||
local_changed = false;
|
||||
|
||||
for (auto* bb : loop_blocks) {
|
||||
for (const auto& instPtr : bb->GetInstructions()) {
|
||||
auto* inst = instPtr.get();
|
||||
|
||||
if (invariant_insts.find(inst) != invariant_insts.end()) {
|
||||
continue; // Already identified
|
||||
}
|
||||
|
||||
if (!IsPureHoistingCandidate(inst->GetOpcode())) {
|
||||
continue; // Cannot hoist impure instructions (load, store, call, branch)
|
||||
}
|
||||
|
||||
// Check if all operands are loop-invariant
|
||||
bool all_ops_invariant = true;
|
||||
for (size_t i = 0; i < inst->GetNumOperands(); ++i) {
|
||||
auto* op = inst->GetOperand(i);
|
||||
|
||||
// Constants are invariant
|
||||
if (dynamic_cast<ConstantValue*>(op)) {
|
||||
continue;
|
||||
}
|
||||
|
||||
// Values defined outside the loop are invariant
|
||||
if (auto* op_inst = dynamic_cast<Instruction*>(op)) {
|
||||
if (loop_blocks.find(op_inst->GetParent()) == loop_blocks.end()) {
|
||||
continue;
|
||||
}
|
||||
// If defined inside the loop, must be already marked invariant
|
||||
if (invariant_insts.find(op_inst) != invariant_insts.end()) {
|
||||
continue;
|
||||
}
|
||||
} else {
|
||||
// Arguments and Globals are always defined outside the loop
|
||||
continue;
|
||||
}
|
||||
|
||||
all_ops_invariant = false;
|
||||
break;
|
||||
}
|
||||
|
||||
if (all_ops_invariant) {
|
||||
invariant_insts.insert(inst);
|
||||
invariant_order.push_back(inst);
|
||||
local_changed = true;
|
||||
changed = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Hoist the loop-invariant instructions into the preheader (in dependency order)
|
||||
for (auto* inst : invariant_order) {
|
||||
auto& source_insts = const_cast<std::vector<std::unique_ptr<Instruction>>&>(inst->GetParent()->GetInstructions());
|
||||
auto& preheader_insts = const_cast<std::vector<std::unique_ptr<Instruction>>&>(preheader->GetInstructions());
|
||||
|
||||
std::unique_ptr<Instruction> moved_inst;
|
||||
for (auto it = source_insts.begin(); it != source_insts.end(); ++it) {
|
||||
if (it->get() == inst) {
|
||||
moved_inst = std::move(*it);
|
||||
source_insts.erase(it);
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (moved_inst) {
|
||||
moved_inst->SetParent(preheader);
|
||||
// Insert right before the terminator branch instruction of the preheader block
|
||||
if (!preheader_insts.empty() && preheader->HasTerminator()) {
|
||||
auto* term = preheader_insts.back().get();
|
||||
preheader->InsertInstructionBefore(std::move(moved_inst), term);
|
||||
} else {
|
||||
preheader_insts.push_back(std::move(moved_inst));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return changed;
|
||||
}
|
||||
|
||||
} // namespace ir
|
||||
@@ -4,13 +4,11 @@
|
||||
namespace ir {
|
||||
|
||||
void RunFunctionOptimizationPasses(Function* func, Context& ctx) {
|
||||
// 1. Promote memory-based local variables to SSA form using Mem2Reg
|
||||
RunMem2Reg(func, ctx);
|
||||
|
||||
// 2. Run scalar optimizations iteratively until convergence (no changes observed)
|
||||
bool changed = true;
|
||||
int iterations = 0;
|
||||
const int max_iterations = 16; // Safe limit to prevent compile-time infinite loops
|
||||
const int max_iterations = 16;
|
||||
|
||||
while (changed && iterations < max_iterations) {
|
||||
changed = false;
|
||||
@@ -19,6 +17,7 @@ void RunFunctionOptimizationPasses(Function* func, Context& ctx) {
|
||||
changed |= RunConstProp(func, ctx);
|
||||
changed |= RunConstFold(func, ctx);
|
||||
changed |= RunCSE(func);
|
||||
changed |= RunLICM(func);
|
||||
changed |= RunDCE(func);
|
||||
changed |= RunCFGSimplify(func);
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user