Merge branch 'backend' into SCCP

This commit is contained in:
rain2133
2025-07-20 13:00:15 +08:00
346 changed files with 46985 additions and 3851 deletions

View File

@@ -0,0 +1,160 @@
#include "AddressCalculationExpansion.h"
#include <iostream>
#include <vector>
#include "IR.h"
#include "IRBuilder.h"
extern int DEBUG;
namespace sysy {
bool AddressCalculationExpansion::run() {
bool changed = false;
for (auto& funcPair : pModule->getFunctions()) {
Function* func = funcPair.second.get();
for (auto& bb_ptr : func->getBasicBlocks()) {
BasicBlock* bb = bb_ptr.get();
for (auto it = bb->getInstructions().begin(); it != bb->getInstructions().end(); ) {
Instruction* inst = it->get();
Value* basePointer = nullptr;
Value* valueToStore = nullptr;
size_t firstIndexOperandIdx = 0;
size_t numBaseOperands = 0;
if (inst->isLoad()) {
numBaseOperands = 1;
basePointer = inst->getOperand(0);
firstIndexOperandIdx = 1;
} else if (inst->isStore()) {
numBaseOperands = 2;
valueToStore = inst->getOperand(0);
basePointer = inst->getOperand(1);
firstIndexOperandIdx = 2;
} else {
++it;
continue;
}
if (inst->getNumOperands() <= numBaseOperands) {
++it;
continue;
}
std::vector<int> dims;
if (AllocaInst* allocaInst = dynamic_cast<AllocaInst*>(basePointer)) {
for (const auto& use_ptr : allocaInst->getDims()) {
Value* dimValue = use_ptr->getValue();
if (ConstantValue* constVal = dynamic_cast<ConstantValue*>(dimValue)) {
dims.push_back(constVal->getInt());
} else {
std::cerr << "Warning: AllocaInst dimension is not a constant integer. Skipping GEP expansion for: ";
SysYPrinter::printValue(allocaInst);
std::cerr << "\n";
dims.clear();
break;
}
}
} else if (GlobalValue* globalValue = dynamic_cast<GlobalValue*>(basePointer)) {
std::cerr << "Warning: GlobalValue dimension handling needs explicit implementation for GEP expansion. Skipping GEP for: ";
SysYPrinter::printValue(globalValue);
std::cerr << "\n";
++it;
continue;
} else {
std::cerr << "Warning: Base pointer is not AllocaInst/GlobalValue or its array dimensions cannot be determined for GEP expansion. Skipping GEP for: ";
SysYPrinter::printValue(basePointer);
std::cerr << " in instruction ";
SysYPrinter::printInst(inst);
std::cerr << "\n";
++it;
continue;
}
if (dims.empty() && (inst->getNumOperands() > numBaseOperands)) {
if (DEBUG) {
std::cerr << "ACE Warning: Could not get valid array dimensions for ";
SysYPrinter::printValue(basePointer);
std::cerr << " in instruction ";
SysYPrinter::printInst(inst);
std::cerr << " (expected dimensions for indices, but got none).\n";
}
++it;
continue;
}
std::vector<Value*> indexOperands;
for (size_t i = firstIndexOperandIdx; i < inst->getNumOperands(); ++i) {
indexOperands.push_back(inst->getOperand(i));
}
if (AllocaInst* allocaInst = dynamic_cast<AllocaInst*>(basePointer)) {
if (allocaInst->getNumDims() != indexOperands.size()) {
if (DEBUG) {
std::cerr << "ACE Warning: Index count (" << indexOperands.size() << ") does not match AllocaInst dimensions (" << allocaInst->getNumDims() << ") for instruction ";
SysYPrinter::printInst(inst);
std::cerr << "\n";
}
++it;
continue;
}
}
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 = 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);
totalOffset = newTotalOffsetInst;
}
// 计算有效地址effective_address = basePointer + totalOffset
Value* effective_address = pBuilder->createBinaryInst(Instruction::kAdd, basePointer->getType(), basePointer, totalOffset);
// 创建新的 LoadInst 或 StoreInstindices 为空
Instruction* newInst = nullptr;
if (inst->isLoad()) {
newInst = pBuilder->createLoadInst(effective_address, {});
inst->replaceAllUsesWith(newInst);
} else { // StoreInst
newInst = pBuilder->createStoreInst(valueToStore, effective_address, {});
}
Instruction* oldInst = it->get();
++it;
for (size_t i = 0; i < oldInst->getNumOperands(); ++i) {
Value* operandValue = oldInst->getOperand(i);
if (operandValue) {
for (auto use_it = operandValue->getUses().begin(); use_it != operandValue->getUses().end(); ++use_it) {
if ((*use_it)->getUser() == oldInst && (*use_it)->getIndex() == i) {
operandValue->removeUse(*use_it);
break;
}
}
}
}
bb->getInstructions().erase(std::prev(it));
changed = true;
if (DEBUG) {
std::cerr << "ACE: Computed effective address:\n";
SysYPrinter::printInst(dynamic_cast<Instruction*>(effective_address));
std::cerr << "ACE: New Load/Store instruction:\n";
SysYPrinter::printInst(newInst);
std::cerr << "--------------------------------\n";
}
}
}
}
return changed;
}
} // namespace sysy

View File

@@ -23,10 +23,15 @@ add_executable(sysyc
SysYIRPrinter.cpp
SysYIRCFGOpt.cpp
SysYIRAnalyser.cpp
DeadCodeElimination.cpp
Mem2Reg.cpp
Reg2Mem.cpp
# DeadCodeElimination.cpp
AddressCalculationExpansion.cpp
# Mem2Reg.cpp
# Reg2Mem.cpp
RISCv64Backend.cpp
RISCv64ISel.cpp
RISCv64RegAlloc.cpp
RISCv64AsmPrinter.cpp
RISCv64Passes.cpp
)
# 设置 include 路径,包含 ANTLR 运行时库和项目头文件

225
src/RISCv64AsmPrinter.cpp Normal file
View File

@@ -0,0 +1,225 @@
#include "RISCv64AsmPrinter.h"
#include "RISCv64ISel.h"
#include <stdexcept>
namespace sysy {
// 检查是否为内存加载/存储指令,以处理特殊的打印格式
bool isMemoryOp(RVOpcodes opcode) {
switch (opcode) {
case RVOpcodes::LB: case RVOpcodes::LH: case RVOpcodes::LW: case RVOpcodes::LD:
case RVOpcodes::LBU: case RVOpcodes::LHU: case RVOpcodes::LWU:
case RVOpcodes::SB: case RVOpcodes::SH: case RVOpcodes::SW: case RVOpcodes::SD:
return true;
default:
return false;
}
}
RISCv64AsmPrinter::RISCv64AsmPrinter(MachineFunction* mfunc) : MFunc(mfunc) {}
void RISCv64AsmPrinter::run(std::ostream& os) {
OS = &os;
*OS << ".globl " << MFunc->getName() << "\n";
*OS << MFunc->getName() << ":\n";
printPrologue();
for (auto& mbb : MFunc->getBlocks()) {
printBasicBlock(mbb.get());
}
}
void RISCv64AsmPrinter::printPrologue() {
StackFrameInfo& frame_info = MFunc->getFrameInfo();
// 序言需要为保存ra和s0预留16字节
int total_stack_size = frame_info.locals_size + frame_info.spill_size + 16;
int aligned_stack_size = (total_stack_size + 15) & ~15;
frame_info.total_size = aligned_stack_size;
if (aligned_stack_size > 0) {
*OS << " addi sp, sp, -" << aligned_stack_size << "\n";
*OS << " sd ra, " << (aligned_stack_size - 8) << "(sp)\n";
*OS << " sd s0, " << (aligned_stack_size - 16) << "(sp)\n";
*OS << " mv s0, sp\n";
}
// 忠实还原保存函数入口参数的逻辑
Function* F = MFunc->getFunc();
if (F && F->getEntryBlock()) {
int arg_idx = 0;
RISCv64ISel* isel = MFunc->getISel();
for (AllocaInst* alloca_for_param : F->getEntryBlock()->getArguments()) {
if (arg_idx >= 8) break;
unsigned vreg = isel->getVReg(alloca_for_param);
if (frame_info.alloca_offsets.count(vreg)) {
int offset = frame_info.alloca_offsets.at(vreg);
auto arg_reg = static_cast<PhysicalReg>(static_cast<int>(PhysicalReg::A0) + arg_idx);
*OS << " sw " << regToString(arg_reg) << ", " << offset << "(s0)\n";
}
arg_idx++;
}
}
}
void RISCv64AsmPrinter::printEpilogue() {
int aligned_stack_size = MFunc->getFrameInfo().total_size;
if (aligned_stack_size > 0) {
*OS << " ld ra, " << (aligned_stack_size - 8) << "(sp)\n";
*OS << " ld s0, " << (aligned_stack_size - 16) << "(sp)\n";
*OS << " addi sp, sp, " << aligned_stack_size << "\n";
}
}
void RISCv64AsmPrinter::printBasicBlock(MachineBasicBlock* mbb) {
if (!mbb->getName().empty()) {
*OS << mbb->getName() << ":\n";
}
for (auto& instr : mbb->getInstructions()) {
printInstruction(instr.get());
}
}
void RISCv64AsmPrinter::printInstruction(MachineInstr* instr) {
auto opcode = instr->getOpcode();
if (opcode == RVOpcodes::RET) {
printEpilogue();
}
if (opcode != RVOpcodes::LABEL) {
*OS << " ";
}
switch (opcode) {
case RVOpcodes::ADD: *OS << "add "; break; case RVOpcodes::ADDI: *OS << "addi "; break;
case RVOpcodes::ADDW: *OS << "addw "; break; case RVOpcodes::ADDIW: *OS << "addiw "; break;
case RVOpcodes::SUB: *OS << "sub "; break; case RVOpcodes::SUBW: *OS << "subw "; break;
case RVOpcodes::MUL: *OS << "mul "; break; case RVOpcodes::MULW: *OS << "mulw "; break;
case RVOpcodes::DIV: *OS << "div "; break; case RVOpcodes::DIVW: *OS << "divw "; break;
case RVOpcodes::REM: *OS << "rem "; break; case RVOpcodes::REMW: *OS << "remw "; break;
case RVOpcodes::XOR: *OS << "xor "; break; case RVOpcodes::XORI: *OS << "xori "; break;
case RVOpcodes::OR: *OS << "or "; break; case RVOpcodes::ORI: *OS << "ori "; break;
case RVOpcodes::AND: *OS << "and "; break; case RVOpcodes::ANDI: *OS << "andi "; break;
case RVOpcodes::SLL: *OS << "sll "; break; case RVOpcodes::SLLI: *OS << "slli "; break;
case RVOpcodes::SLLW: *OS << "sllw "; break; case RVOpcodes::SLLIW: *OS << "slliw "; break;
case RVOpcodes::SRL: *OS << "srl "; break; case RVOpcodes::SRLI: *OS << "srli "; break;
case RVOpcodes::SRLW: *OS << "srlw "; break; case RVOpcodes::SRLIW: *OS << "srliw "; break;
case RVOpcodes::SRA: *OS << "sra "; break; case RVOpcodes::SRAI: *OS << "srai "; break;
case RVOpcodes::SRAW: *OS << "sraw "; break; case RVOpcodes::SRAIW: *OS << "sraiw "; break;
case RVOpcodes::SLT: *OS << "slt "; break; case RVOpcodes::SLTI: *OS << "slti "; break;
case RVOpcodes::SLTU: *OS << "sltu "; break; case RVOpcodes::SLTIU: *OS << "sltiu "; break;
case RVOpcodes::LW: *OS << "lw "; break; case RVOpcodes::LH: *OS << "lh "; break;
case RVOpcodes::LB: *OS << "lb "; break; case RVOpcodes::LWU: *OS << "lwu "; break;
case RVOpcodes::LHU: *OS << "lhu "; break; case RVOpcodes::LBU: *OS << "lbu "; break;
case RVOpcodes::SW: *OS << "sw "; break; case RVOpcodes::SH: *OS << "sh "; break;
case RVOpcodes::SB: *OS << "sb "; break; case RVOpcodes::LD: *OS << "ld "; break;
case RVOpcodes::SD: *OS << "sd "; break;
case RVOpcodes::J: *OS << "j "; break; case RVOpcodes::JAL: *OS << "jal "; break;
case RVOpcodes::JALR: *OS << "jalr "; break; case RVOpcodes::RET: *OS << "ret"; break;
case RVOpcodes::BEQ: *OS << "beq "; break; case RVOpcodes::BNE: *OS << "bne "; break;
case RVOpcodes::BLT: *OS << "blt "; break; case RVOpcodes::BGE: *OS << "bge "; break;
case RVOpcodes::BLTU: *OS << "bltu "; break; case RVOpcodes::BGEU: *OS << "bgeu "; break;
case RVOpcodes::LI: *OS << "li "; break; case RVOpcodes::LA: *OS << "la "; break;
case RVOpcodes::MV: *OS << "mv "; break; case RVOpcodes::NEG: *OS << "neg "; break;
case RVOpcodes::NEGW: *OS << "negw "; break; case RVOpcodes::SEQZ: *OS << "seqz "; break;
case RVOpcodes::SNEZ: *OS << "snez "; break;
case RVOpcodes::CALL: *OS << "call "; break;
case RVOpcodes::LABEL:
printOperand(instr->getOperands()[0].get());
*OS << ":";
break;
case RVOpcodes::FRAME_LOAD:
case RVOpcodes::FRAME_STORE:
// These should have been eliminated by RegAlloc
throw std::runtime_error("FRAME pseudo-instruction not eliminated before AsmPrinter");
default:
throw std::runtime_error("Unknown opcode in AsmPrinter");
}
const auto& operands = instr->getOperands();
if (!operands.empty()) {
if (isMemoryOp(opcode)) {
printOperand(operands[0].get());
*OS << ", ";
printOperand(operands[1].get());
} else {
for (size_t i = 0; i < operands.size(); ++i) {
printOperand(operands[i].get());
if (i < operands.size() - 1) {
*OS << ", ";
}
}
}
}
*OS << "\n";
}
void RISCv64AsmPrinter::printOperand(MachineOperand* op) {
if (!op) return;
switch(op->getKind()) {
case MachineOperand::KIND_REG: {
auto reg_op = static_cast<RegOperand*>(op);
if (reg_op->isVirtual()) {
*OS << "%vreg" << reg_op->getVRegNum();
} else {
*OS << regToString(reg_op->getPReg());
}
break;
}
case MachineOperand::KIND_IMM:
*OS << static_cast<ImmOperand*>(op)->getValue();
break;
case MachineOperand::KIND_LABEL:
*OS << static_cast<LabelOperand*>(op)->getName();
break;
case MachineOperand::KIND_MEM: {
auto mem_op = static_cast<MemOperand*>(op);
printOperand(mem_op->getOffset());
*OS << "(";
printOperand(mem_op->getBase());
*OS << ")";
break;
}
}
}
std::string RISCv64AsmPrinter::regToString(PhysicalReg reg) {
switch (reg) {
case PhysicalReg::ZERO: return "x0"; case PhysicalReg::RA: return "ra";
case PhysicalReg::SP: return "sp"; case PhysicalReg::GP: return "gp";
case PhysicalReg::TP: return "tp"; case PhysicalReg::T0: return "t0";
case PhysicalReg::T1: return "t1"; case PhysicalReg::T2: return "t2";
case PhysicalReg::S0: return "s0"; case PhysicalReg::S1: return "s1";
case PhysicalReg::A0: return "a0"; case PhysicalReg::A1: return "a1";
case PhysicalReg::A2: return "a2"; case PhysicalReg::A3: return "a3";
case PhysicalReg::A4: return "a4"; case PhysicalReg::A5: return "a5";
case PhysicalReg::A6: return "a6"; case PhysicalReg::A7: return "a7";
case PhysicalReg::S2: return "s2"; case PhysicalReg::S3: return "s3";
case PhysicalReg::S4: return "s4"; case PhysicalReg::S5: return "s5";
case PhysicalReg::S6: return "s6"; case PhysicalReg::S7: return "s7";
case PhysicalReg::S8: return "s8"; case PhysicalReg::S9: return "s9";
case PhysicalReg::S10: return "s10"; case PhysicalReg::S11: return "s11";
case PhysicalReg::T3: return "t3"; case PhysicalReg::T4: return "t4";
case PhysicalReg::T5: return "t5"; case PhysicalReg::T6: return "t6";
case PhysicalReg::F0: return "f0"; case PhysicalReg::F1: return "f1";
case PhysicalReg::F2: return "f2"; case PhysicalReg::F3: return "f3";
case PhysicalReg::F4: return "f4"; case PhysicalReg::F5: return "f5";
case PhysicalReg::F6: return "f6"; case PhysicalReg::F7: return "f7";
case PhysicalReg::F8: return "f8"; case PhysicalReg::F9: return "f9";
case PhysicalReg::F10: return "f10"; case PhysicalReg::F11: return "f11";
case PhysicalReg::F12: return "f12"; case PhysicalReg::F13: return "f13";
case PhysicalReg::F14: return "f14"; case PhysicalReg::F15: return "f15";
case PhysicalReg::F16: return "f16"; case PhysicalReg::F17: return "f17";
case PhysicalReg::F18: return "f18"; case PhysicalReg::F19: return "f19";
case PhysicalReg::F20: return "f20"; case PhysicalReg::F21: return "f21";
case PhysicalReg::F22: return "f22"; case PhysicalReg::F23: return "f23";
case PhysicalReg::F24: return "f24"; case PhysicalReg::F25: return "f25";
case PhysicalReg::F26: return "f26"; case PhysicalReg::F27: return "f27";
case PhysicalReg::F28: return "f28"; case PhysicalReg::F29: return "f29";
case PhysicalReg::F30: return "f30"; case PhysicalReg::F31: return "f31";
default: return "UNKNOWN_REG";
}
}
} // namespace sysy

File diff suppressed because it is too large Load Diff

635
src/RISCv64ISel.cpp Normal file
View File

@@ -0,0 +1,635 @@
#include "RISCv64ISel.h"
#include <stdexcept>
#include <set>
#include <functional>
#include <cmath> // For std::fabs
#include <limits> // For std::numeric_limits
namespace sysy {
// DAG节点定义 (内部实现)
struct RISCv64ISel::DAGNode {
enum NodeKind { CONSTANT, LOAD, STORE, BINARY, CALL, RETURN, BRANCH, ALLOCA_ADDR, UNARY, MEMSET };
NodeKind kind;
Value* value = nullptr;
std::vector<DAGNode*> operands;
std::vector<DAGNode*> users;
DAGNode(NodeKind k) : kind(k) {}
};
RISCv64ISel::RISCv64ISel() : vreg_counter(0), local_label_counter(0) {}
// 为一个IR Value获取或分配一个新的虚拟寄存器
unsigned RISCv64ISel::getVReg(Value* val) {
if (!val) {
throw std::runtime_error("Cannot get vreg for a null Value.");
}
if (vreg_map.find(val) == vreg_map.end()) {
if (vreg_counter == 0) {
vreg_counter = 1; // vreg 0 保留
}
vreg_map[val] = vreg_counter++;
}
return vreg_map.at(val);
}
// 主入口函数
std::unique_ptr<MachineFunction> RISCv64ISel::runOnFunction(Function* func) {
F = func;
if (!F) return nullptr;
MFunc = std::make_unique<MachineFunction>(F, this);
vreg_map.clear();
bb_map.clear();
vreg_counter = 0;
local_label_counter = 0;
select();
return std::move(MFunc);
}
// 指令选择主流程
void RISCv64ISel::select() {
for (const auto& bb_ptr : F->getBasicBlocks()) {
auto mbb = std::make_unique<MachineBasicBlock>(bb_ptr->getName(), MFunc.get());
bb_map[bb_ptr.get()] = mbb.get();
MFunc->addBlock(std::move(mbb));
}
if (F->getEntryBlock()) {
for (auto* arg_alloca : F->getEntryBlock()->getArguments()) {
getVReg(arg_alloca);
}
}
for (const auto& bb_ptr : F->getBasicBlocks()) {
selectBasicBlock(bb_ptr.get());
}
for (const auto& bb_ptr : F->getBasicBlocks()) {
CurMBB = bb_map.at(bb_ptr.get());
for (auto succ : bb_ptr->getSuccessors()) {
CurMBB->successors.push_back(bb_map.at(succ));
}
for (auto pred : bb_ptr->getPredecessors()) {
CurMBB->predecessors.push_back(bb_map.at(pred));
}
}
}
// 处理单个基本块
void RISCv64ISel::selectBasicBlock(BasicBlock* bb) {
CurMBB = bb_map.at(bb);
auto dag = build_dag(bb);
std::map<Value*, DAGNode*> value_to_node;
for(const auto& node : dag) {
if (node->value) {
value_to_node[node->value] = node.get();
}
}
std::set<DAGNode*> selected_nodes;
std::function<void(DAGNode*)> select_recursive =
[&](DAGNode* node) {
if (!node || selected_nodes.count(node)) return;
for (auto operand : node->operands) {
select_recursive(operand);
}
selectNode(node);
selected_nodes.insert(node);
};
for (const auto& inst_ptr : bb->getInstructions()) {
DAGNode* node_to_select = nullptr;
if (value_to_node.count(inst_ptr.get())) {
node_to_select = value_to_node.at(inst_ptr.get());
} else {
for(const auto& node : dag) {
if(node->value == inst_ptr.get()) {
node_to_select = node.get();
break;
}
}
}
if(node_to_select) {
select_recursive(node_to_select);
}
}
}
// 核心函数为DAG节点选择并生成MachineInstr (忠实移植版)
void RISCv64ISel::selectNode(DAGNode* node) {
switch (node->kind) {
case DAGNode::CONSTANT:
case DAGNode::ALLOCA_ADDR:
if (node->value) getVReg(node->value);
break;
case DAGNode::LOAD: {
auto dest_vreg = getVReg(node->value);
Value* ptr_val = node->operands[0]->value;
if (auto alloca = dynamic_cast<AllocaInst*>(ptr_val)) {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::FRAME_LOAD);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(getVReg(alloca)));
CurMBB->addInstruction(std::move(instr));
} else if (auto global = dynamic_cast<GlobalValue*>(ptr_val)) {
auto addr_vreg = getNewVReg();
auto la = std::make_unique<MachineInstr>(RVOpcodes::LA);
la->addOperand(std::make_unique<RegOperand>(addr_vreg));
la->addOperand(std::make_unique<LabelOperand>(global->getName()));
CurMBB->addInstruction(std::move(la));
auto lw = std::make_unique<MachineInstr>(RVOpcodes::LW);
lw->addOperand(std::make_unique<RegOperand>(dest_vreg));
lw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(addr_vreg),
std::make_unique<ImmOperand>(0)
));
CurMBB->addInstruction(std::move(lw));
} else {
auto ptr_vreg = getVReg(ptr_val);
auto lw = std::make_unique<MachineInstr>(RVOpcodes::LW);
lw->addOperand(std::make_unique<RegOperand>(dest_vreg));
lw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(ptr_vreg),
std::make_unique<ImmOperand>(0)
));
CurMBB->addInstruction(std::move(lw));
}
break;
}
case DAGNode::STORE: {
Value* val_to_store = node->operands[0]->value;
Value* ptr_val = node->operands[1]->value;
if (auto val_const = dynamic_cast<ConstantValue*>(val_to_store)) {
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(getVReg(val_const)));
li->addOperand(std::make_unique<ImmOperand>(val_const->getInt()));
CurMBB->addInstruction(std::move(li));
}
auto val_vreg = getVReg(val_to_store);
if (auto alloca = dynamic_cast<AllocaInst*>(ptr_val)) {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::FRAME_STORE);
instr->addOperand(std::make_unique<RegOperand>(val_vreg));
instr->addOperand(std::make_unique<RegOperand>(getVReg(alloca)));
CurMBB->addInstruction(std::move(instr));
} else if (auto global = dynamic_cast<GlobalValue*>(ptr_val)) {
auto addr_vreg = getNewVReg();
auto la = std::make_unique<MachineInstr>(RVOpcodes::LA);
la->addOperand(std::make_unique<RegOperand>(addr_vreg));
la->addOperand(std::make_unique<LabelOperand>(global->getName()));
CurMBB->addInstruction(std::move(la));
auto sw = std::make_unique<MachineInstr>(RVOpcodes::SW);
sw->addOperand(std::make_unique<RegOperand>(val_vreg));
sw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(addr_vreg),
std::make_unique<ImmOperand>(0)
));
CurMBB->addInstruction(std::move(sw));
} else {
auto ptr_vreg = getVReg(ptr_val);
auto sw = std::make_unique<MachineInstr>(RVOpcodes::SW);
sw->addOperand(std::make_unique<RegOperand>(val_vreg));
sw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(ptr_vreg),
std::make_unique<ImmOperand>(0)
));
CurMBB->addInstruction(std::move(sw));
}
break;
}
case DAGNode::BINARY: {
auto bin = dynamic_cast<BinaryInst*>(node->value);
Value* lhs = bin->getLhs();
Value* rhs = bin->getRhs();
auto load_val_if_const = [&](Value* val) {
if (auto c = dynamic_cast<ConstantValue*>(val)) {
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(getVReg(c)));
li->addOperand(std::make_unique<ImmOperand>(c->getInt()));
CurMBB->addInstruction(std::move(li));
}
};
load_val_if_const(lhs);
load_val_if_const(rhs);
auto dest_vreg = getVReg(bin);
auto lhs_vreg = getVReg(lhs);
auto rhs_vreg = getVReg(rhs);
if (bin->getKind() == BinaryInst::kAdd) {
if (auto rhs_const = dynamic_cast<ConstantValue*>(rhs)) {
if (rhs_const->getInt() >= -2048 && rhs_const->getInt() < 2048) {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::ADDIW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<ImmOperand>(rhs_const->getInt()));
CurMBB->addInstruction(std::move(instr));
return;
}
}
}
switch (bin->getKind()) {
case BinaryInst::kAdd: {
RVOpcodes opcode = (lhs->getType()->isPointer() || rhs->getType()->isPointer()) ? RVOpcodes::ADD : RVOpcodes::ADDW;
auto instr = std::make_unique<MachineInstr>(opcode);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case BinaryInst::kSub: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SUBW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case BinaryInst::kMul: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::MULW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case Instruction::kDiv: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::DIVW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case Instruction::kRem: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::REMW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case BinaryInst::kICmpEQ: {
auto sub = std::make_unique<MachineInstr>(RVOpcodes::SUBW);
sub->addOperand(std::make_unique<RegOperand>(dest_vreg));
sub->addOperand(std::make_unique<RegOperand>(lhs_vreg));
sub->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(sub));
auto seqz = std::make_unique<MachineInstr>(RVOpcodes::SEQZ);
seqz->addOperand(std::make_unique<RegOperand>(dest_vreg));
seqz->addOperand(std::make_unique<RegOperand>(dest_vreg));
CurMBB->addInstruction(std::move(seqz));
break;
}
case BinaryInst::kICmpNE: {
auto sub = std::make_unique<MachineInstr>(RVOpcodes::SUBW);
sub->addOperand(std::make_unique<RegOperand>(dest_vreg));
sub->addOperand(std::make_unique<RegOperand>(lhs_vreg));
sub->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(sub));
auto snez = std::make_unique<MachineInstr>(RVOpcodes::SNEZ);
snez->addOperand(std::make_unique<RegOperand>(dest_vreg));
snez->addOperand(std::make_unique<RegOperand>(dest_vreg));
CurMBB->addInstruction(std::move(snez));
break;
}
case BinaryInst::kICmpLT: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SLT);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case BinaryInst::kICmpGT: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SLT);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(rhs_vreg));
instr->addOperand(std::make_unique<RegOperand>(lhs_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case BinaryInst::kICmpLE: {
auto slt = std::make_unique<MachineInstr>(RVOpcodes::SLT);
slt->addOperand(std::make_unique<RegOperand>(dest_vreg));
slt->addOperand(std::make_unique<RegOperand>(rhs_vreg));
slt->addOperand(std::make_unique<RegOperand>(lhs_vreg));
CurMBB->addInstruction(std::move(slt));
auto xori = std::make_unique<MachineInstr>(RVOpcodes::XORI);
xori->addOperand(std::make_unique<RegOperand>(dest_vreg));
xori->addOperand(std::make_unique<RegOperand>(dest_vreg));
xori->addOperand(std::make_unique<ImmOperand>(1));
CurMBB->addInstruction(std::move(xori));
break;
}
case BinaryInst::kICmpGE: {
auto slt = std::make_unique<MachineInstr>(RVOpcodes::SLT);
slt->addOperand(std::make_unique<RegOperand>(dest_vreg));
slt->addOperand(std::make_unique<RegOperand>(lhs_vreg));
slt->addOperand(std::make_unique<RegOperand>(rhs_vreg));
CurMBB->addInstruction(std::move(slt));
auto xori = std::make_unique<MachineInstr>(RVOpcodes::XORI);
xori->addOperand(std::make_unique<RegOperand>(dest_vreg));
xori->addOperand(std::make_unique<RegOperand>(dest_vreg));
xori->addOperand(std::make_unique<ImmOperand>(1));
CurMBB->addInstruction(std::move(xori));
break;
}
default:
throw std::runtime_error("Unsupported binary instruction in ISel");
}
break;
}
case DAGNode::UNARY: {
auto unary = dynamic_cast<UnaryInst*>(node->value);
auto dest_vreg = getVReg(unary);
auto src_vreg = getVReg(unary->getOperand());
switch (unary->getKind()) {
case UnaryInst::kNeg: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SUBW);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::ZERO));
instr->addOperand(std::make_unique<RegOperand>(src_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
case UnaryInst::kNot: {
auto instr = std::make_unique<MachineInstr>(RVOpcodes::SEQZ);
instr->addOperand(std::make_unique<RegOperand>(dest_vreg));
instr->addOperand(std::make_unique<RegOperand>(src_vreg));
CurMBB->addInstruction(std::move(instr));
break;
}
default:
throw std::runtime_error("Unsupported unary instruction in ISel");
}
break;
}
case DAGNode::CALL: {
auto call = dynamic_cast<CallInst*>(node->value);
for (size_t i = 0; i < node->operands.size() && i < 8; ++i) {
DAGNode* arg_node = node->operands[i];
auto arg_preg = static_cast<PhysicalReg>(static_cast<int>(PhysicalReg::A0) + i);
if (arg_node->kind == DAGNode::CONSTANT) {
if (auto const_val = dynamic_cast<ConstantValue*>(arg_node->value)) {
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(arg_preg));
li->addOperand(std::make_unique<ImmOperand>(const_val->getInt()));
CurMBB->addInstruction(std::move(li));
}
} else {
auto src_vreg = getVReg(arg_node->value);
auto mv = std::make_unique<MachineInstr>(RVOpcodes::MV);
mv->addOperand(std::make_unique<RegOperand>(arg_preg));
mv->addOperand(std::make_unique<RegOperand>(src_vreg));
CurMBB->addInstruction(std::move(mv));
}
}
auto call_instr = std::make_unique<MachineInstr>(RVOpcodes::CALL);
call_instr->addOperand(std::make_unique<LabelOperand>(call->getCallee()->getName()));
CurMBB->addInstruction(std::move(call_instr));
if (!call->getType()->isVoid()) {
auto mv_instr = std::make_unique<MachineInstr>(RVOpcodes::MV);
mv_instr->addOperand(std::make_unique<RegOperand>(getVReg(call)));
mv_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::A0));
CurMBB->addInstruction(std::move(mv_instr));
}
break;
}
case DAGNode::RETURN: {
auto ret_inst_ir = dynamic_cast<ReturnInst*>(node->value);
if (ret_inst_ir && ret_inst_ir->hasReturnValue()) {
Value* ret_val = ret_inst_ir->getReturnValue();
if (auto const_val = dynamic_cast<ConstantValue*>(ret_val)) {
auto li_instr = std::make_unique<MachineInstr>(RVOpcodes::LI);
li_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::A0));
li_instr->addOperand(std::make_unique<ImmOperand>(const_val->getInt()));
CurMBB->addInstruction(std::move(li_instr));
} else {
auto mv_instr = std::make_unique<MachineInstr>(RVOpcodes::MV);
mv_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::A0));
mv_instr->addOperand(std::make_unique<RegOperand>(getVReg(ret_val)));
CurMBB->addInstruction(std::move(mv_instr));
}
}
auto ret_mi = std::make_unique<MachineInstr>(RVOpcodes::RET);
CurMBB->addInstruction(std::move(ret_mi));
break;
}
case DAGNode::BRANCH: {
if (auto cond_br = dynamic_cast<CondBrInst*>(node->value)) {
auto br_instr = std::make_unique<MachineInstr>(RVOpcodes::BNE);
br_instr->addOperand(std::make_unique<RegOperand>(getVReg(cond_br->getCondition())));
br_instr->addOperand(std::make_unique<RegOperand>(PhysicalReg::ZERO));
br_instr->addOperand(std::make_unique<LabelOperand>(cond_br->getThenBlock()->getName()));
CurMBB->addInstruction(std::move(br_instr));
} else if (auto uncond_br = dynamic_cast<UncondBrInst*>(node->value)) {
auto j_instr = std::make_unique<MachineInstr>(RVOpcodes::J);
j_instr->addOperand(std::make_unique<LabelOperand>(uncond_br->getBlock()->getName()));
CurMBB->addInstruction(std::move(j_instr));
}
break;
}
case DAGNode::MEMSET: {
auto memset = dynamic_cast<MemsetInst*>(node->value);
auto r_dest_addr = getVReg(memset->getPointer());
auto r_num_bytes = getVReg(memset->getSize());
auto r_value_byte = getVReg(memset->getValue());
auto r_counter = getNewVReg();
auto r_end_addr = getNewVReg();
auto r_current_addr = getNewVReg();
auto r_temp_val = getNewVReg();
auto add_instr = [&](RVOpcodes op, unsigned rd, unsigned rs1, unsigned rs2) {
auto i = std::make_unique<MachineInstr>(op);
i->addOperand(std::make_unique<RegOperand>(rd));
i->addOperand(std::make_unique<RegOperand>(rs1));
i->addOperand(std::make_unique<RegOperand>(rs2));
CurMBB->addInstruction(std::move(i));
};
auto addi_instr = [&](RVOpcodes op, unsigned rd, unsigned rs1, int64_t imm) {
auto i = std::make_unique<MachineInstr>(op);
i->addOperand(std::make_unique<RegOperand>(rd));
i->addOperand(std::make_unique<RegOperand>(rs1));
i->addOperand(std::make_unique<ImmOperand>(imm));
CurMBB->addInstruction(std::move(i));
};
auto store_instr = [&](RVOpcodes op, unsigned src, unsigned base, int64_t off) {
auto i = std::make_unique<MachineInstr>(op);
i->addOperand(std::make_unique<RegOperand>(src));
i->addOperand(std::make_unique<MemOperand>(std::make_unique<RegOperand>(base), std::make_unique<ImmOperand>(off)));
CurMBB->addInstruction(std::move(i));
};
auto branch_instr = [&](RVOpcodes op, unsigned rs1, unsigned rs2, const std::string& label) {
auto i = std::make_unique<MachineInstr>(op);
i->addOperand(std::make_unique<RegOperand>(rs1));
i->addOperand(std::make_unique<RegOperand>(rs2));
i->addOperand(std::make_unique<LabelOperand>(label));
CurMBB->addInstruction(std::move(i));
};
auto jump_instr = [&](const std::string& label) {
auto i = std::make_unique<MachineInstr>(RVOpcodes::J);
i->addOperand(std::make_unique<LabelOperand>(label));
CurMBB->addInstruction(std::move(i));
};
auto label_instr = [&](const std::string& name) {
auto i = std::make_unique<MachineInstr>(RVOpcodes::LABEL);
i->addOperand(std::make_unique<LabelOperand>(name));
CurMBB->addInstruction(std::move(i));
};
int unique_id = this->local_label_counter++;
std::string loop_start_label = MFunc->getName() + "_memset_loop_start_" + std::to_string(unique_id);
std::string loop_end_label = MFunc->getName() + "_memset_loop_end_" + std::to_string(unique_id);
std::string remainder_label = MFunc->getName() + "_memset_remainder_" + std::to_string(unique_id);
std::string done_label = MFunc->getName() + "_memset_done_" + std::to_string(unique_id);
addi_instr(RVOpcodes::ANDI, r_temp_val, r_value_byte, 255);
addi_instr(RVOpcodes::SLLI, r_value_byte, r_temp_val, 8);
add_instr(RVOpcodes::OR, r_temp_val, r_temp_val, r_value_byte);
addi_instr(RVOpcodes::SLLI, r_value_byte, r_temp_val, 16);
add_instr(RVOpcodes::OR, r_temp_val, r_temp_val, r_value_byte);
addi_instr(RVOpcodes::SLLI, r_value_byte, r_temp_val, 32);
add_instr(RVOpcodes::OR, r_temp_val, r_temp_val, r_value_byte);
add_instr(RVOpcodes::ADD, r_end_addr, r_dest_addr, r_num_bytes);
auto mv = std::make_unique<MachineInstr>(RVOpcodes::MV);
mv->addOperand(std::make_unique<RegOperand>(r_current_addr));
mv->addOperand(std::make_unique<RegOperand>(r_dest_addr));
CurMBB->addInstruction(std::move(mv));
addi_instr(RVOpcodes::ANDI, r_counter, r_num_bytes, -8);
add_instr(RVOpcodes::ADD, r_counter, r_dest_addr, r_counter);
label_instr(loop_start_label);
branch_instr(RVOpcodes::BGEU, r_current_addr, r_counter, loop_end_label);
store_instr(RVOpcodes::SD, r_temp_val, r_current_addr, 0);
addi_instr(RVOpcodes::ADDI, r_current_addr, r_current_addr, 8);
jump_instr(loop_start_label);
label_instr(loop_end_label);
label_instr(remainder_label);
branch_instr(RVOpcodes::BGEU, r_current_addr, r_end_addr, done_label);
store_instr(RVOpcodes::SB, r_temp_val, r_current_addr, 0);
addi_instr(RVOpcodes::ADDI, r_current_addr, r_current_addr, 1);
jump_instr(remainder_label);
label_instr(done_label);
break;
}
default:
throw std::runtime_error("Unsupported DAGNode kind in ISel");
}
}
// 以下是忠实移植的DAG构建函数
RISCv64ISel::DAGNode* RISCv64ISel::create_node(int kind_int, Value* val, std::map<Value*, DAGNode*>& value_to_node, std::vector<std::unique_ptr<DAGNode>>& nodes_storage) {
auto kind = static_cast<DAGNode::NodeKind>(kind_int);
if (val && value_to_node.count(val) && kind != DAGNode::STORE && kind != DAGNode::RETURN && kind != DAGNode::BRANCH && kind != DAGNode::MEMSET) {
return value_to_node[val];
}
auto node = std::make_unique<DAGNode>(kind);
node->value = val;
DAGNode* raw_node_ptr = node.get();
nodes_storage.push_back(std::move(node));
if (val && !val->getType()->isVoid() && (dynamic_cast<Instruction*>(val) || dynamic_cast<GlobalValue*>(val))) {
value_to_node[val] = raw_node_ptr;
}
return raw_node_ptr;
}
RISCv64ISel::DAGNode* RISCv64ISel::get_operand_node(Value* val_ir, std::map<Value*, DAGNode*>& value_to_node, std::vector<std::unique_ptr<DAGNode>>& nodes_storage) {
if (value_to_node.count(val_ir)) {
return value_to_node[val_ir];
} else if (dynamic_cast<ConstantValue*>(val_ir)) {
return create_node(DAGNode::CONSTANT, val_ir, value_to_node, nodes_storage);
} else if (dynamic_cast<GlobalValue*>(val_ir)) {
return create_node(DAGNode::CONSTANT, val_ir, value_to_node, nodes_storage);
} else if (dynamic_cast<AllocaInst*>(val_ir)) {
return create_node(DAGNode::ALLOCA_ADDR, val_ir, value_to_node, nodes_storage);
}
return create_node(DAGNode::LOAD, val_ir, value_to_node, nodes_storage);
}
std::vector<std::unique_ptr<RISCv64ISel::DAGNode>> RISCv64ISel::build_dag(BasicBlock* bb) {
std::vector<std::unique_ptr<DAGNode>> nodes_storage;
std::map<Value*, DAGNode*> value_to_node;
for (const auto& inst_ptr : bb->getInstructions()) {
Instruction* inst = inst_ptr.get();
if (auto alloca = dynamic_cast<AllocaInst*>(inst)) {
create_node(DAGNode::ALLOCA_ADDR, alloca, value_to_node, nodes_storage);
} else if (auto store = dynamic_cast<StoreInst*>(inst)) {
auto store_node = create_node(DAGNode::STORE, store, value_to_node, nodes_storage);
store_node->operands.push_back(get_operand_node(store->getValue(), value_to_node, nodes_storage));
store_node->operands.push_back(get_operand_node(store->getPointer(), value_to_node, nodes_storage));
} else if (auto memset = dynamic_cast<MemsetInst*>(inst)) {
auto memset_node = create_node(DAGNode::MEMSET, memset, value_to_node, nodes_storage);
memset_node->operands.push_back(get_operand_node(memset->getPointer(), value_to_node, nodes_storage));
memset_node->operands.push_back(get_operand_node(memset->getBegin(), value_to_node, nodes_storage));
memset_node->operands.push_back(get_operand_node(memset->getSize(), value_to_node, nodes_storage));
memset_node->operands.push_back(get_operand_node(memset->getValue(), value_to_node, nodes_storage));
} else if (auto load = dynamic_cast<LoadInst*>(inst)) {
auto load_node = create_node(DAGNode::LOAD, load, value_to_node, nodes_storage);
load_node->operands.push_back(get_operand_node(load->getPointer(), value_to_node, nodes_storage));
} else if (auto bin = dynamic_cast<BinaryInst*>(inst)) {
if(value_to_node.count(bin)) continue;
if (bin->getKind() == BinaryInst::kSub) {
if (auto const_lhs = dynamic_cast<ConstantValue*>(bin->getLhs())) {
if (const_lhs->getInt() == 0) {
auto unary_node = create_node(DAGNode::UNARY, bin, value_to_node, nodes_storage);
unary_node->operands.push_back(get_operand_node(bin->getRhs(), value_to_node, nodes_storage));
continue;
}
}
}
auto bin_node = create_node(DAGNode::BINARY, bin, value_to_node, nodes_storage);
bin_node->operands.push_back(get_operand_node(bin->getLhs(), value_to_node, nodes_storage));
bin_node->operands.push_back(get_operand_node(bin->getRhs(), value_to_node, nodes_storage));
} else if (auto un = dynamic_cast<UnaryInst*>(inst)) {
if(value_to_node.count(un)) continue;
auto unary_node = create_node(DAGNode::UNARY, un, value_to_node, nodes_storage);
unary_node->operands.push_back(get_operand_node(un->getOperand(), value_to_node, nodes_storage));
} else if (auto call = dynamic_cast<CallInst*>(inst)) {
if(value_to_node.count(call)) continue;
auto call_node = create_node(DAGNode::CALL, call, value_to_node, nodes_storage);
for (auto arg : call->getArguments()) {
call_node->operands.push_back(get_operand_node(arg->getValue(), value_to_node, nodes_storage));
}
} else if (auto ret = dynamic_cast<ReturnInst*>(inst)) {
auto ret_node = create_node(DAGNode::RETURN, ret, value_to_node, nodes_storage);
if (ret->hasReturnValue()) {
ret_node->operands.push_back(get_operand_node(ret->getReturnValue(), value_to_node, nodes_storage));
}
} else if (auto cond_br = dynamic_cast<CondBrInst*>(inst)) {
auto br_node = create_node(DAGNode::BRANCH, cond_br, value_to_node, nodes_storage);
br_node->operands.push_back(get_operand_node(cond_br->getCondition(), value_to_node, nodes_storage));
} else if (auto uncond_br = dynamic_cast<UncondBrInst*>(inst)) {
create_node(DAGNode::BRANCH, uncond_br, value_to_node, nodes_storage);
}
}
return nodes_storage;
}
} // namespace sysy

54
src/RISCv64Passes.cpp Normal file
View File

@@ -0,0 +1,54 @@
#include "RISCv64Passes.h"
#include <iostream>
namespace sysy {
// --- 寄存器分配前优化 ---
void PreRA_Scheduler::runOnMachineFunction(MachineFunction* mfunc) {
// TODO: 在此实现寄存器分配前的指令调度。
// 遍历mfunc中的每一个MachineBasicBlock。
// 对每个基本块内的MachineInstr列表进行重排。
//
// 实现思路:
// 1. 分析每个基本块内指令的数据依赖关系,构建依赖图(DAG)。
// 2. 根据目标处理器的流水线特性(指令延迟等),使用列表调度等算法对指令进行重排。
// 3. 此时操作的是虚拟寄存器,只存在真依赖,调度自由度最大。
//
// std::cout << "Running Pre-RA Instruction Scheduler..." << std::endl;
}
// --- 寄存器分配后优化 ---
void PeepholeOptimizer::runOnMachineFunction(MachineFunction* mfunc) {
// TODO: 在此实现窥孔优化。
// 遍历mfunc中的每一个MachineBasicBlock。
// 对每个基本块内的MachineInstr列表进行扫描和替换。
//
// 实现思路:
// 1. 维护一个大小固定例如3-5条指令的滑动窗口。
// 2. 识别特定的冗余模式,例如:
// - `mv a0, a1` 后紧跟 `mv a1, a0` (可消除的交换)
// - `sw t0, 12(s0)` 后紧跟 `lw t1, 12(s0)` (冗余加载)
// - 强度削减: `mul x, x, 2` -> `slli x, x, 1`
// 3. 识别后直接修改MachineInstr列表删除、替换或插入指令
//
// std::cout << "Running Post-RA Peephole Optimizer..." << std::endl;
}
void PostRA_Scheduler::runOnMachineFunction(MachineFunction* mfunc) {
// TODO: 在此实现寄存器分配后的局部指令调度。
// 遍历mfunc中的每一个MachineBasicBlock。
// 重点关注由寄存器分配器插入的spill/fill代码。
//
// 实现思路:
// 1. 识别出用于spill/fill的lw/sw指令。
// 2. 在不违反数据依赖(包括物理寄存器引入的伪依赖)的前提下,
// 尝试将lw指令向上移动使其与使用它的指令之间有足够的距离以隐藏访存延迟。
// 3. 同样可以尝试将sw指令向下移动。
//
// std::cout << "Running Post-RA Local Scheduler..." << std::endl;
}
} // namespace sysy

322
src/RISCv64RegAlloc.cpp Normal file
View File

@@ -0,0 +1,322 @@
#include "RISCv64RegAlloc.h"
#include "RISCv64ISel.h"
#include <algorithm>
#include <vector>
namespace sysy {
RISCv64RegAlloc::RISCv64RegAlloc(MachineFunction* mfunc) : MFunc(mfunc) {
allocable_int_regs = {
PhysicalReg::T0, PhysicalReg::T1, PhysicalReg::T2, PhysicalReg::T3,
PhysicalReg::T4, PhysicalReg::T5, PhysicalReg::T6,
PhysicalReg::A0, PhysicalReg::A1, PhysicalReg::A2, PhysicalReg::A3,
PhysicalReg::A4, PhysicalReg::A5, PhysicalReg::A6, PhysicalReg::A7,
PhysicalReg::S0, PhysicalReg::S1, PhysicalReg::S2, PhysicalReg::S3,
PhysicalReg::S4, PhysicalReg::S5, PhysicalReg::S6, PhysicalReg::S7,
PhysicalReg::S8, PhysicalReg::S9, PhysicalReg::S10, PhysicalReg::S11,
};
}
void RISCv64RegAlloc::run() {
eliminateFrameIndices();
analyzeLiveness();
buildInterferenceGraph();
colorGraph();
rewriteFunction();
}
void RISCv64RegAlloc::eliminateFrameIndices() {
StackFrameInfo& frame_info = MFunc->getFrameInfo();
int current_offset = 0;
Function* F = MFunc->getFunc();
RISCv64ISel* isel = MFunc->getISel();
for (auto& bb : F->getBasicBlocks()) {
for (auto& inst : bb->getInstructions()) {
if (auto alloca = dynamic_cast<AllocaInst*>(inst.get())) {
int size = 4;
if (!alloca->getDims().empty()) {
int num_elements = 1;
for (const auto& dim_use : alloca->getDims()) {
if (auto const_dim = dynamic_cast<ConstantValue*>(dim_use->getValue())) {
num_elements *= const_dim->getInt();
}
}
size *= num_elements;
}
current_offset += size;
unsigned alloca_vreg = isel->getVReg(alloca);
frame_info.alloca_offsets[alloca_vreg] = -current_offset;
}
}
}
frame_info.locals_size = current_offset;
for (auto& mbb : MFunc->getBlocks()) {
std::vector<std::unique_ptr<MachineInstr>> new_instructions;
for (auto& instr_ptr : mbb->getInstructions()) {
if (instr_ptr->getOpcode() == RVOpcodes::FRAME_LOAD) {
auto& operands = instr_ptr->getOperands();
unsigned dest_vreg = static_cast<RegOperand*>(operands[0].get())->getVRegNum();
unsigned alloca_vreg = static_cast<RegOperand*>(operands[1].get())->getVRegNum();
int offset = frame_info.alloca_offsets.at(alloca_vreg);
auto addr_vreg = isel->getNewVReg();
auto addi = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
addi->addOperand(std::make_unique<RegOperand>(addr_vreg));
addi->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
addi->addOperand(std::make_unique<ImmOperand>(offset));
new_instructions.push_back(std::move(addi));
auto lw = std::make_unique<MachineInstr>(RVOpcodes::LW);
lw->addOperand(std::make_unique<RegOperand>(dest_vreg));
lw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(addr_vreg),
std::make_unique<ImmOperand>(0)));
new_instructions.push_back(std::move(lw));
} else if (instr_ptr->getOpcode() == RVOpcodes::FRAME_STORE) {
auto& operands = instr_ptr->getOperands();
unsigned src_vreg = static_cast<RegOperand*>(operands[0].get())->getVRegNum();
unsigned alloca_vreg = static_cast<RegOperand*>(operands[1].get())->getVRegNum();
int offset = frame_info.alloca_offsets.at(alloca_vreg);
auto addr_vreg = isel->getNewVReg();
auto addi = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
addi->addOperand(std::make_unique<RegOperand>(addr_vreg));
addi->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
addi->addOperand(std::make_unique<ImmOperand>(offset));
new_instructions.push_back(std::move(addi));
auto sw = std::make_unique<MachineInstr>(RVOpcodes::SW);
sw->addOperand(std::make_unique<RegOperand>(src_vreg));
sw->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(addr_vreg),
std::make_unique<ImmOperand>(0)));
new_instructions.push_back(std::move(sw));
} else {
new_instructions.push_back(std::move(instr_ptr));
}
}
mbb->getInstructions() = std::move(new_instructions);
}
}
void RISCv64RegAlloc::getInstrUseDef(MachineInstr* instr, LiveSet& use, LiveSet& def) {
bool is_def = true;
auto opcode = instr->getOpcode();
// 预定义def和use规则
if (opcode == RVOpcodes::SW || opcode == RVOpcodes::SD ||
opcode == RVOpcodes::BEQ || opcode == RVOpcodes::BNE ||
opcode == RVOpcodes::BLT || opcode == RVOpcodes::BGE ||
opcode == RVOpcodes::RET || opcode == RVOpcodes::J) {
is_def = false;
}
if (opcode == RVOpcodes::CALL) {
// CALL会杀死所有调用者保存寄存器这是一个简化处理
// 同时也使用了传入a0-a7的参数
}
for (const auto& op : instr->getOperands()) {
if (op->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<RegOperand*>(op.get());
if (reg_op->isVirtual()) {
if (is_def) {
def.insert(reg_op->getVRegNum());
is_def = false;
} else {
use.insert(reg_op->getVRegNum());
}
}
} else if (op->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand*>(op.get());
if (mem_op->getBase()->isVirtual()) {
use.insert(mem_op->getBase()->getVRegNum());
}
}
}
}
void RISCv64RegAlloc::analyzeLiveness() {
bool changed = true;
while (changed) {
changed = false;
for (auto it = MFunc->getBlocks().rbegin(); it != MFunc->getBlocks().rend(); ++it) {
auto& mbb = *it;
LiveSet live_out;
for (auto succ : mbb->successors) {
if (!succ->getInstructions().empty()) {
auto first_instr = succ->getInstructions().front().get();
if (live_in_map.count(first_instr)) {
live_out.insert(live_in_map.at(first_instr).begin(), live_in_map.at(first_instr).end());
}
}
}
for (auto instr_it = mbb->getInstructions().rbegin(); instr_it != mbb->getInstructions().rend(); ++instr_it) {
MachineInstr* instr = instr_it->get();
LiveSet old_live_in = live_in_map[instr];
live_out_map[instr] = live_out;
LiveSet use, def;
getInstrUseDef(instr, use, def);
LiveSet live_in = use;
LiveSet diff = live_out;
for (auto vreg : def) {
diff.erase(vreg);
}
live_in.insert(diff.begin(), diff.end());
live_in_map[instr] = live_in;
live_out = live_in;
if (live_in_map[instr] != old_live_in) {
changed = true;
}
}
}
}
}
void RISCv64RegAlloc::buildInterferenceGraph() {
std::set<unsigned> all_vregs;
for (auto& mbb : MFunc->getBlocks()) {
for(auto& instr : mbb->getInstructions()) {
LiveSet use, def;
getInstrUseDef(instr.get(), use, def);
for(auto u : use) all_vregs.insert(u);
for(auto d : def) all_vregs.insert(d);
}
}
for (auto vreg : all_vregs) { interference_graph[vreg] = {}; }
for (auto& mbb : MFunc->getBlocks()) {
for (auto& instr : mbb->getInstructions()) {
LiveSet def, use;
getInstrUseDef(instr.get(), use, def);
const LiveSet& live_out = live_out_map.at(instr.get());
for (unsigned d : def) {
for (unsigned l : live_out) {
if (d != l) {
interference_graph[d].insert(l);
interference_graph[l].insert(d);
}
}
}
}
}
}
void RISCv64RegAlloc::colorGraph() {
std::vector<unsigned> sorted_vregs;
for (auto const& [vreg, neighbors] : interference_graph) {
sorted_vregs.push_back(vreg);
}
std::sort(sorted_vregs.begin(), sorted_vregs.end(), [&](unsigned a, unsigned b) {
return interference_graph[a].size() > interference_graph[b].size();
});
for (unsigned vreg : sorted_vregs) {
std::set<PhysicalReg> used_colors;
for (unsigned neighbor : interference_graph.at(vreg)) {
if (color_map.count(neighbor)) {
used_colors.insert(color_map.at(neighbor));
}
}
bool colored = false;
for (PhysicalReg preg : allocable_int_regs) {
if (used_colors.find(preg) == used_colors.end()) {
color_map[vreg] = preg;
colored = true;
break;
}
}
if (!colored) {
spilled_vregs.insert(vreg);
}
}
}
void RISCv64RegAlloc::rewriteFunction() {
StackFrameInfo& frame_info = MFunc->getFrameInfo();
int current_offset = frame_info.locals_size;
for (unsigned vreg : spilled_vregs) {
current_offset += 4;
frame_info.spill_offsets[vreg] = -current_offset;
}
frame_info.spill_size = current_offset - frame_info.locals_size;
for (auto& mbb : MFunc->getBlocks()) {
std::vector<std::unique_ptr<MachineInstr>> new_instructions;
for (auto& instr_ptr : mbb->getInstructions()) {
LiveSet use, def;
getInstrUseDef(instr_ptr.get(), use, def);
for (unsigned vreg : use) {
if (spilled_vregs.count(vreg)) {
int offset = frame_info.spill_offsets.at(vreg);
auto load = std::make_unique<MachineInstr>(RVOpcodes::LW);
load->addOperand(std::make_unique<RegOperand>(vreg));
load->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0),
std::make_unique<ImmOperand>(offset)
));
new_instructions.push_back(std::move(load));
}
}
new_instructions.push_back(std::move(instr_ptr));
for (unsigned vreg : def) {
if (spilled_vregs.count(vreg)) {
int offset = frame_info.spill_offsets.at(vreg);
auto store = std::make_unique<MachineInstr>(RVOpcodes::SW);
store->addOperand(std::make_unique<RegOperand>(vreg));
store->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0),
std::make_unique<ImmOperand>(offset)
));
new_instructions.push_back(std::move(store));
}
}
}
mbb->getInstructions() = std::move(new_instructions);
}
for (auto& mbb : MFunc->getBlocks()) {
for (auto& instr_ptr : mbb->getInstructions()) {
for (auto& op_ptr : instr_ptr->getOperands()) {
if(op_ptr->getKind() == MachineOperand::KIND_REG) {
auto reg_op = static_cast<RegOperand*>(op_ptr.get());
if (reg_op->isVirtual()) {
unsigned vreg = reg_op->getVRegNum();
if (color_map.count(vreg)) {
reg_op->setPReg(color_map.at(vreg));
} else if (spilled_vregs.count(vreg)) {
reg_op->setPReg(PhysicalReg::T6); // 溢出统一用t6
}
}
} else if (op_ptr->getKind() == MachineOperand::KIND_MEM) {
auto mem_op = static_cast<MemOperand*>(op_ptr.get());
auto base_reg_op = mem_op->getBase();
if(base_reg_op->isVirtual()){
unsigned vreg = base_reg_op->getVRegNum();
if(color_map.count(vreg)) {
base_reg_op->setPReg(color_map.at(vreg));
} else if (spilled_vregs.count(vreg)) {
base_reg_op->setPReg(PhysicalReg::T6);
}
}
}
}
}
}
}
} // namespace sysy

View File

@@ -0,0 +1,59 @@
#pragma once
#include "IR.h" // 假设IR.h包含了Module, Function, BasicBlock, Instruction, Value, IRBuilder, Type等定义
#include "IRBuilder.h" // 需要IRBuilder来创建新指令
#include "SysYIRPrinter.h" // 新增: 用于调试输出
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>
#include <list> // 用于迭代和修改指令列表
#include <algorithm> // for std::reverse (if needed, although not used in final version)
#include <iostream> // MODIFICATION: 用于警告输出
namespace sysy {
/**
* @brief AddressCalculationExpansion Pass
*
* 这是一个IR优化Pass用于将LoadInst和StoreInst中包含的多维数组索引
* 显式地转换为IR中的BinaryInst乘法和加法序列并生成带有线性偏移量的
* LoadInst/StoreInst。
*
* 目的确保在寄存器分配之前所有中间地址计算的结果都有明确的IR指令和对应的虚拟寄存器
* 从而避免在后端DAG构建时临时创建值而导致寄存器分配缺失的问题。
*
* SysY语言特性
* - 无指针类型所有数组访问的基地址是alloca或global的AllocaType/ArrayType
* - 数据类型只有int和float且都占用4字节。
* - LoadInst和StoreInst直接接受多个索引作为额外操作数。
*/
class AddressCalculationExpansion {
private:
Module* pModule;
IRBuilder* pBuilder; // 用于在IR中插入新指令
// 数组元素的固定大小根据SysY特性int和float都是4字节
static const int ELEMENT_SIZE = 4;
// 辅助函数:根据数组的维度信息和当前索引的维度,计算该索引的步长(字节数)
// dims: 包含所有维度大小的vector例如 {2, 3, 4}
// currentDimIndex: 当前正在处理的索引在 dims 中的位置 (0, 1, 2...)
int calculateStride(const std::vector<int>& dims, size_t currentDimIndex) {
int stride = ELEMENT_SIZE; // 最内层元素大小 (4字节)
// 乘以当前维度之后的所有维度的大小
for (size_t i = currentDimIndex + 1; i < dims.size(); ++i) {
stride *= dims[i];
}
return stride;
}
public:
AddressCalculationExpansion(Module* module, IRBuilder* builder)
: pModule(module), pBuilder(builder) {}
// 运行此Pass
bool run();
};
} // namespace sysy

View File

@@ -0,0 +1,32 @@
#ifndef RISCV64_ASMPRINTER_H
#define RISCV64_ASMPRINTER_H
#include "RISCv64LLIR.h"
#include <iostream>
namespace sysy {
class RISCv64AsmPrinter {
public:
RISCv64AsmPrinter(MachineFunction* mfunc);
// 主入口
void run(std::ostream& os);
private:
// 打印各个部分
void printPrologue();
void printEpilogue();
void printBasicBlock(MachineBasicBlock* mbb);
void printInstruction(MachineInstr* instr);
// 辅助函数
std::string regToString(PhysicalReg reg);
void printOperand(MachineOperand* op);
MachineFunction* MFunc;
std::ostream* OS;
};
} // namespace sysy
#endif // RISCV64_ASMPRINTER_H

View File

@@ -3,118 +3,23 @@
#include "IR.h"
#include <string>
#include <vector>
#include <map>
#include <set>
#include <memory>
#include <iostream>
#include <functional> // For std::function
extern int DEBUG;
extern int DEEPDEBUG;
namespace sysy {
// RISCv64CodeGen 现在是一个高层驱动器
class RISCv64CodeGen {
public:
enum class PhysicalReg {
ZERO, RA, SP, GP, TP, T0, T1, T2, S0, S1, A0, A1, A2, A3, A4, A5, A6, A7, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, T3, T4, T5, T6,
F0, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12, F13, F14, F15,F16, F17, F18, F19, F20, F21, F22, F23, F24, F25, F26, F27, F28, F29, F30, F31
};
// Move DAGNode and RegAllocResult to public section
struct DAGNode {
enum NodeKind { CONSTANT, LOAD, STORE, BINARY, CALL, RETURN, BRANCH, ALLOCA_ADDR, UNARY };
NodeKind kind;
Value* value = nullptr; // For IR Value
std::string inst; // Generated RISC-V instruction(s) for this node
std::string result_vreg; // Virtual register assigned to this node's result
std::vector<DAGNode*> operands;
std::vector<DAGNode*> users; // For debugging and potentially optimizations
DAGNode(NodeKind k) : kind(k) {}
// Debugging / helper
std::string getNodeKindString() const {
switch (kind) {
case CONSTANT: return "CONSTANT";
case LOAD: return "LOAD";
case STORE: return "STORE";
case BINARY: return "BINARY";
case CALL: return "CALL";
case RETURN: return "RETURN";
case BRANCH: return "BRANCH";
case ALLOCA_ADDR: return "ALLOCA_ADDR";
case UNARY: return "UNARY";
default: return "UNKNOWN";
}
}
};
struct RegAllocResult {
std::map<std::string, PhysicalReg> vreg_to_preg; // Virtual register to Physical Register mapping
std::map<Value*, int> stack_map; // Value (AllocaInst) to stack offset
int stack_size = 0; // Total stack frame size for locals and spills
};
RISCv64CodeGen(Module* mod) : module(mod) {}
// 唯一的公共入口点
std::string code_gen();
std::string module_gen();
std::string function_gen(Function* func);
// 修改 basicBlock_gen 的声明,添加 int block_idx 参数
std::string basicBlock_gen(BasicBlock* bb, const RegAllocResult& alloc, int block_idx);
// DAG related
std::vector<std::unique_ptr<DAGNode>> build_dag(BasicBlock* bb);
void select_instructions(DAGNode* node, const RegAllocResult& alloc);
// 改变 emit_instructions 的参数,使其可以直接添加汇编指令到 main ss
void emit_instructions(DAGNode* node, std::stringstream& ss, const RegAllocResult& alloc, std::set<DAGNode*>& emitted_nodes);
// Register Allocation related
std::map<Instruction*, std::set<std::string>> liveness_analysis(Function* func);
std::map<std::string, std::set<std::string>> build_interference_graph(
const std::map<Instruction*, std::set<std::string>>& live_sets);
void color_graph(std::map<std::string, PhysicalReg>& vreg_to_preg,
const std::map<std::string, std::set<std::string>>& interference_graph);
RegAllocResult register_allocation(Function* func);
void eliminate_phi(Function* func); // Phi elimination is typically done before DAG building
// Utility
std::string reg_to_string(PhysicalReg reg);
void print_dag(const std::vector<std::unique_ptr<DAGNode>>& dag, const std::string& bb_name);
private:
static const std::vector<PhysicalReg> allocable_regs;
std::map<Value*, std::string> value_vreg_map; // Maps IR Value* to its virtual register name
// 模块级代码生成
std::string module_gen();
// 函数级代码生成 (实现新的流水线)
std::string function_gen(Function* func);
Module* module;
int vreg_counter = 0; // Counter for unique virtual register names
int alloca_offset_counter = 0; // Counter for alloca offsets
// 新增一个成员变量来存储当前函数的所有 DAGNode以确保其生命周期贯穿整个函数代码生成
// 这样可以在多个 BasicBlock_gen 调用中访问到完整的 DAG 节点
std::vector<std::unique_ptr<DAGNode>> current_function_dag_nodes;
// 为空标签定义一个伪名称前缀,加上块索引以确保唯一性
const std::string ENTRY_BLOCK_PSEUDO_NAME = "entry_block_";
// !!! 修改get_operand_node 辅助函数现在需要传入 value_to_node 和 nodes_storage 的引用
// 因为它们是 build_dag 局部管理的
DAGNode* get_operand_node(
Value* val_ir,
std::map<Value*, DAGNode*>& value_to_node,
std::vector<std::unique_ptr<DAGNode>>& nodes_storage
);
// !!! 新增create_node 辅助函数也需要传入 value_to_node 和 nodes_storage 的引用
// 并且它应该不再是 lambda而是一个真正的成员函数
DAGNode* create_node(
DAGNode::NodeKind kind,
Value* val,
std::map<Value*, DAGNode*>& value_to_node,
std::vector<std::unique_ptr<DAGNode>>& nodes_storage
);
std::vector<std::unique_ptr<Instruction>> temp_instructions_storage; // 用于存储 build_dag 中创建的临时 BinaryInst
};
} // namespace sysy

49
src/include/RISCv64ISel.h Normal file
View File

@@ -0,0 +1,49 @@
#ifndef RISCV64_ISEL_H
#define RISCV64_ISEL_H
#include "RISCv64LLIR.h"
namespace sysy {
class RISCv64ISel {
public:
RISCv64ISel();
// 模块主入口将一个高层IR函数转换为底层LLIR函数
std::unique_ptr<MachineFunction> runOnFunction(Function* func);
// 公开接口以便后续模块如RegAlloc可以查询或创建vreg
unsigned getVReg(Value* val);
unsigned getNewVReg() { return vreg_counter++; }
private:
// DAG节点定义作为ISel的内部实现细节
struct DAGNode;
// 指令选择主流程
void select();
// 为单个基本块生成指令
void selectBasicBlock(BasicBlock* bb);
// 核心函数为DAG节点选择并生成MachineInstr
void selectNode(DAGNode* node);
// DAG 构建相关函数 (从原RISCv64Backend迁移)
std::vector<std::unique_ptr<DAGNode>> build_dag(BasicBlock* bb);
DAGNode* get_operand_node(Value* val_ir, std::map<Value*, DAGNode*>&, std::vector<std::unique_ptr<DAGNode>>&);
DAGNode* create_node(int kind, Value* val, std::map<Value*, DAGNode*>&, std::vector<std::unique_ptr<DAGNode>>&);
// 状态
Function* F; // 当前处理的高层IR函数
std::unique_ptr<MachineFunction> MFunc; // 正在构建的底层LLIR函数
MachineBasicBlock* CurMBB; // 当前正在处理的机器基本块
// 映射关系
std::map<Value*, unsigned> vreg_map;
std::map<const BasicBlock*, MachineBasicBlock*> bb_map;
unsigned vreg_counter;
int local_label_counter;
};
} // namespace sysy
#endif // RISCV64_ISEL_H

200
src/include/RISCv64LLIR.h Normal file
View File

@@ -0,0 +1,200 @@
#ifndef RISCV64_LLIR_H
#define RISCV64_LLIR_H
#include "IR.h" // 确保包含了您自己的IR头文件
#include <string>
#include <vector>
#include <memory>
#include <cstdint>
#include <map>
// 前向声明,避免循环引用
namespace sysy {
class Function;
class RISCv64ISel;
}
namespace sysy {
// 物理寄存器定义
enum class PhysicalReg {
ZERO, RA, SP, GP, TP, T0, T1, T2, S0, S1, A0, A1, A2, A3, A4, A5, A6, A7, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, T3, T4, T5, T6,
F0, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12, F13, F14, F15,F16, F17, F18, F19, F20, F21, F22, F23, F24, F25, F26, F27, F28, F29, F30, F31
};
// RISC-V 指令操作码枚举
enum class RVOpcodes {
// 算术指令
ADD, ADDI, ADDW, ADDIW, SUB, SUBW, MUL, MULW, DIV, DIVW, REM, REMW,
// 逻辑指令
XOR, XORI, OR, ORI, AND, ANDI,
// 移位指令
SLL, SLLI, SLLW, SLLIW, SRL, SRLI, SRLW, SRLIW, SRA, SRAI, SRAW, SRAIW,
// 比较指令
SLT, SLTI, SLTU, SLTIU,
// 内存访问指令
LW, LH, LB, LWU, LHU, LBU, SW, SH, SB, LD, SD,
// 控制流指令
J, JAL, JALR, RET,
BEQ, BNE, BLT, BGE, BLTU, BGEU,
// 伪指令
LI, LA, MV, NEG, NEGW, SEQZ, SNEZ,
// 函数调用
CALL,
// 特殊标记,非指令
LABEL,
// 新增伪指令,用于解耦栈帧处理
FRAME_LOAD, // 从栈帧加载 (AllocaInst)
FRAME_STORE, // 保存到栈帧 (AllocaInst)
};
class MachineOperand;
class RegOperand;
class ImmOperand;
class LabelOperand;
class MemOperand;
class MachineInstr;
class MachineBasicBlock;
class MachineFunction;
// 操作数基类
class MachineOperand {
public:
enum OperandKind { KIND_REG, KIND_IMM, KIND_LABEL, KIND_MEM };
MachineOperand(OperandKind kind) : kind(kind) {}
virtual ~MachineOperand() = default;
OperandKind getKind() const { return kind; }
private:
OperandKind kind;
};
// 寄存器操作数
class RegOperand : public MachineOperand {
public:
// 构造虚拟寄存器
RegOperand(unsigned vreg_num)
: MachineOperand(KIND_REG), vreg_num(vreg_num), is_virtual(true) {}
// 构造物理寄存器
RegOperand(PhysicalReg preg)
: MachineOperand(KIND_REG), preg(preg), is_virtual(false) {}
bool isVirtual() const { return is_virtual; }
unsigned getVRegNum() const { return vreg_num; }
PhysicalReg getPReg() const { return preg; }
void setPReg(PhysicalReg new_preg) {
preg = new_preg;
is_virtual = false;
}
private:
unsigned vreg_num = 0;
PhysicalReg preg = PhysicalReg::ZERO;
bool is_virtual;
};
// 立即数操作数
class ImmOperand : public MachineOperand {
public:
ImmOperand(int64_t value) : MachineOperand(KIND_IMM), value(value) {}
int64_t getValue() const { return value; }
private:
int64_t value;
};
// 标签操作数
class LabelOperand : public MachineOperand {
public:
LabelOperand(const std::string& name) : MachineOperand(KIND_LABEL), name(name) {}
const std::string& getName() const { return name; }
private:
std::string name;
};
// 内存操作数, 表示 offset(base_reg)
class MemOperand : public MachineOperand {
public:
MemOperand(std::unique_ptr<RegOperand> base, std::unique_ptr<ImmOperand> offset)
: MachineOperand(KIND_MEM), base(std::move(base)), offset(std::move(offset)) {}
RegOperand* getBase() const { return base.get(); }
ImmOperand* getOffset() const { return offset.get(); }
private:
std::unique_ptr<RegOperand> base;
std::unique_ptr<ImmOperand> offset;
};
// 机器指令
class MachineInstr {
public:
MachineInstr(RVOpcodes opcode) : opcode(opcode) {}
RVOpcodes getOpcode() const { return opcode; }
const std::vector<std::unique_ptr<MachineOperand>>& getOperands() const { return operands; }
std::vector<std::unique_ptr<MachineOperand>>& getOperands() { return operands; }
void addOperand(std::unique_ptr<MachineOperand> operand) {
operands.push_back(std::move(operand));
}
private:
RVOpcodes opcode;
std::vector<std::unique_ptr<MachineOperand>> operands;
};
// 机器基本块
class MachineBasicBlock {
public:
MachineBasicBlock(const std::string& name, MachineFunction* parent)
: name(name), parent(parent) {}
const std::string& getName() const { return name; }
MachineFunction* getParent() const { return parent; }
const std::vector<std::unique_ptr<MachineInstr>>& getInstructions() const { return instructions; }
std::vector<std::unique_ptr<MachineInstr>>& getInstructions() { return instructions; }
void addInstruction(std::unique_ptr<MachineInstr> instr) {
instructions.push_back(std::move(instr));
}
std::vector<MachineBasicBlock*> successors;
std::vector<MachineBasicBlock*> predecessors;
private:
std::string name;
std::vector<std::unique_ptr<MachineInstr>> instructions;
MachineFunction* parent;
};
// 栈帧信息
struct StackFrameInfo {
int locals_size = 0; // 仅为AllocaInst分配的大小
int spill_size = 0; // 仅为溢出分配的大小
int total_size = 0; // 总大小
std::map<unsigned, int> alloca_offsets; // <AllocaInst的vreg, 栈偏移>
std::map<unsigned, int> spill_offsets; // <溢出vreg, 栈偏移>
};
// 机器函数
class MachineFunction {
public:
MachineFunction(Function* func, RISCv64ISel* isel) : F(func), name(func->getName()), isel(isel) {}
Function* getFunc() const { return F; }
RISCv64ISel* getISel() const { return isel; }
const std::string& getName() const { return name; }
StackFrameInfo& getFrameInfo() { return frame_info; }
const std::vector<std::unique_ptr<MachineBasicBlock>>& getBlocks() const { return blocks; }
std::vector<std::unique_ptr<MachineBasicBlock>>& getBlocks() { return blocks; }
void addBlock(std::unique_ptr<MachineBasicBlock> block) {
blocks.push_back(std::move(block));
}
private:
Function* F;
RISCv64ISel* isel; // 指向创建它的ISel用于获取vreg映射等信息
std::string name;
std::vector<std::unique_ptr<MachineBasicBlock>> blocks;
StackFrameInfo frame_info;
};
} // namespace sysy
#endif // RISCV64_LLIR_H

View File

@@ -0,0 +1,61 @@
#ifndef RISCV64_PASSES_H
#define RISCV64_PASSES_H
#include "RISCv64LLIR.h"
namespace sysy {
/**
* @class Pass
* @brief 所有优化Pass的抽象基类 (可选,但推荐)
* * 定义一个通用的接口,所有优化都应该实现它。
*/
class Pass {
public:
virtual ~Pass() = default;
virtual void runOnMachineFunction(MachineFunction* mfunc) = 0;
};
// --- 寄存器分配前优化 ---
/**
* @class PreRA_Scheduler
* @brief 寄存器分配前的指令调度器
* * 在虚拟寄存器上进行操作,此时调度自由度最大,
* 主要目标是隐藏指令延迟,提高流水线效率。
*/
class PreRA_Scheduler : public Pass {
public:
void runOnMachineFunction(MachineFunction* mfunc) override;
};
// --- 寄存器分配后优化 ---
/**
* @class PeepholeOptimizer
* @brief 窥孔优化器
* * 在已分配物理寄存器的指令流上,通过一个小的滑动窗口来查找
* 并替换掉一些冗余或低效的指令模式。
*/
class PeepholeOptimizer : public Pass {
public:
void runOnMachineFunction(MachineFunction* mfunc) override;
};
/**
* @class PostRA_Scheduler
* @brief 寄存器分配后的局部指令调度器
* * 主要目标是优化寄存器分配器插入的spill/fill代码(lw/sw)
* 尝试将加载指令提前,以隐藏其访存延迟。
*/
class PostRA_Scheduler : public Pass {
public:
void runOnMachineFunction(MachineFunction* mfunc) override;
};
} // namespace sysy
#endif // RISCV64_PASSES_H

View File

@@ -0,0 +1,56 @@
#ifndef RISCV64_REGALLOC_H
#define RISCV64_REGALLOC_H
#include "RISCv64LLIR.h"
namespace sysy {
class RISCv64RegAlloc {
public:
RISCv64RegAlloc(MachineFunction* mfunc);
// 模块主入口
void run();
private:
using LiveSet = std::set<unsigned>; // 活跃虚拟寄存器集合
using InterferenceGraph = std::map<unsigned, std::set<unsigned>>;
// 栈帧管理
void eliminateFrameIndices();
// 活跃性分析
void analyzeLiveness();
// 构建干扰图
void buildInterferenceGraph();
// 图着色分配寄存器
void colorGraph();
// 重写函数替换vreg并插入溢出代码
void rewriteFunction();
// 辅助函数获取指令的Use/Def集合
void getInstrUseDef(MachineInstr* instr, LiveSet& use, LiveSet& def);
MachineFunction* MFunc;
// 活跃性分析结果
std::map<const MachineInstr*, LiveSet> live_in_map;
std::map<const MachineInstr*, LiveSet> live_out_map;
// 干扰图
InterferenceGraph interference_graph;
// 图着色结果
std::map<unsigned, PhysicalReg> color_map; // vreg -> preg
std::set<unsigned> spilled_vregs; // 被溢出的vreg集合
// 可用的物理寄存器池
std::vector<PhysicalReg> allocable_int_regs;
};
} // namespace sysy
#endif // RISCV64_REGALLOC_H

View File

@@ -16,9 +16,10 @@ using namespace antlr4;
#include "SysYIRCFGOpt.h"
#include "RISCv64Backend.h"
#include "SysYIRAnalyser.h"
#include "DeadCodeElimination.h"
#include "Mem2Reg.h"
#include "Reg2Mem.h"
// #include "DeadCodeElimination.h"
#include "AddressCalculationExpansion.h"
// #include "Mem2Reg.h"
// #include "Reg2Mem.h"
using namespace sysy;
@@ -124,8 +125,12 @@ int main(int argc, char **argv) {
// 无论最终输出是 IR 还是 ASM只要不是停止在 AST 阶段,都会进入此优化流程。
// optLevel = 0 时,执行默认优化。
// optLevel >= 1 时,执行默认优化 + 额外的 -O1 优化。
cout << "Applying middle-end optimizations (level -O" << optLevel << ")...\n";
if (DEBUG) cout << "Applying middle-end optimizations (level -O" << optLevel << ")...\n";
// 设置 DEBUG 模式(如果指定了 'ird'
if (argStopAfter == "ird") {
DEBUG = 1; // 这里可能需要更精细地控制 DEBUG 的开启时机和范围
}
// 默认优化 pass (在所有优化级别都会执行)
SysYCFGOpt cfgopt(moduleIR, builder);
cfgopt.SysYOptimizateAfterIR();
@@ -139,37 +144,22 @@ int main(int argc, char **argv) {
cout << "=== After CFA & AVA (Default) ===\n";
SysYPrinter(moduleIR).printIR(); // 临时打印器用于调试
}
DeadCodeElimination dce(moduleIR, &cfa, &ava);
dce.runDCEPipeline();
if (DEBUG) {
cout << "=== After 1st DCE (Default) ===\n";
SysYPrinter(moduleIR).printIR();
AddressCalculationExpansion ace(moduleIR, builder);
if (ace.run()) {
if (DEBUG) cout << "AddressCalculationExpansion made changes.\n";
// 如果 ACE 改变了IR并且 DEBUG 模式开启可以考虑打印IR
if (DEBUG) {
cout << "=== After AddressCalculationExpansion ===\n";
SysYPrinter(moduleIR).printIR();
}
} else {
if (DEBUG) cout << "AddressCalculationExpansion made no changes.\n";
}
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();
}
dce.runDCEPipeline(); // 第二次 DCE (默认)
if (DEBUG) {
cout << "=== After 2nd DCE (Default) ===\n";
SysYPrinter(moduleIR).printIR();
}
// 根据优化级别,执行额外的优化 pass
if (optLevel >= 1) {
cout << "Applying additional -O" << optLevel << " optimizations...\n";
if (DEBUG) cout << "Applying additional -O" << optLevel << " optimizations...\n";
// 放置 -O1 及其以上级别要启用的额外优化 pass
// 例如:
// MyNewOptimizationPass newOpt(moduleIR, builder);
@@ -184,17 +174,39 @@ int main(int argc, char **argv) {
// MyCustomOpt2 opt2_pass(moduleIR, builder, &cfa); // 假设需要CFA
// opt2_pass.run();
// ... 更多 -O1 特有的优化
// DeadCodeElimination dce(moduleIR, &cfa, &ava);
// dce.runDCEPipeline();
// if (DEBUG) {
// cout << "=== After 1st DCE (Default) ===\n";
// SysYPrinter(moduleIR).printIR();
// }
// Mem2Reg mem2reg(moduleIR, builder, &cfa, &ava);
// mem2reg.mem2regPipeline();
// 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();
// }
// dce.runDCEPipeline(); // 第二次 DCE (默认)
// if (DEBUG) {
// cout << "=== After 2nd DCE (Default) ===\n";
// SysYPrinter(moduleIR).printIR();
// }
} else {
cout << "No additional middle-end optimizations applied for -O" << optLevel << ".\n";
if (DEBUG) cout << "No additional middle-end optimizations applied for -O" << optLevel << ".\n";
}
// 5. 根据 argStopAfter 决定后续操作
// a) 如果指定停止在 IR 阶段,则打印最终 IR 并退出
if (argStopAfter == "ir" || argStopAfter == "ird") {
// 设置 DEBUG 模式(如果指定了 'ird'
if (argStopAfter == "ird") {
DEBUG = 1; // 这里可能需要更精细地控制 DEBUG 的开启时机和范围
}
// 打印最终 IR
cout << "=== Final IR ===\n";
SysYPrinter printer(moduleIR); // 在这里创建打印器,因为可能之前调试时用过临时打印器
@@ -203,17 +215,16 @@ int main(int argc, char **argv) {
}
// b) 如果未停止在 IR 阶段,则继续生成汇编 (后端)
// 设置 DEBUG 模式(如果指定了 'asmd'
if (argStopAfter == "asmd") {
DEBUG = 1;
// DEEPDEBUG = 1;
}
sysy::RISCv64CodeGen codegen(moduleIR); // 传入优化后的 moduleIR
string asmCode = codegen.code_gen();
// 如果指定停止在 ASM 阶段,则打印/保存汇编并退出
if (argStopAfter == "asm" || argStopAfter == "asmd") {
// 设置 DEBUG 模式(如果指定了 'asmd'
if (argStopAfter == "asmd") {
DEBUG = 1;
DEEPDEBUG = 1;
}
if (!argOutputFilename.empty()) {
ofstream fout(argOutputFilename);
if (not fout.is_open()) {