[midend]重构了src目录

This commit is contained in:
Lixuanwang
2025-07-29 21:30:30 +08:00
parent f5922d0178
commit 09ae47924e
84 changed files with 85 additions and 1715 deletions

View File

@@ -0,0 +1,123 @@
#include "CalleeSavedHandler.h"
#include <set>
#include <algorithm>
namespace sysy {
char CalleeSavedHandler::ID = 0;
bool CalleeSavedHandler::runOnFunction(Function *F, AnalysisManager& AM) {
// This pass works on MachineFunction level, not IR level
return false;
}
void CalleeSavedHandler::runOnMachineFunction(MachineFunction* mfunc) {
// 此 Pass 负责分析、分配栈空间并插入 callee-saved 寄存器的保存/恢复指令。
// 它通过与 FrameInfo 协作,确保为 callee-saved 寄存器分配的空间与局部变量/溢出槽的空间不冲突。
// 这样做可以使生成的 sd/ld 指令能被后续的优化 Pass (如 PostRA-Scheduler) 处理。
StackFrameInfo& frame_info = mfunc->getFrameInfo();
std::set<PhysicalReg> used_callee_saved;
// 1. 扫描所有指令找出被使用的s寄存器 (s1-s11)
for (auto& mbb : mfunc->getBlocks()) {
for (auto& instr : mbb->getInstructions()) {
for (auto& op : instr->getOperands()) {
auto check_and_insert_reg = [&](RegOperand* reg_op) {
if (!reg_op->isVirtual()) {
PhysicalReg preg = reg_op->getPReg();
if (preg >= PhysicalReg::S1 && preg <= PhysicalReg::S11) {
used_callee_saved.insert(preg);
}
}
};
if (op->getKind() == MachineOperand::KIND_REG) {
check_and_insert_reg(static_cast<RegOperand*>(op.get()));
} else if (op->getKind() == MachineOperand::KIND_MEM) {
check_and_insert_reg(static_cast<MemOperand*>(op.get())->getBase());
}
}
}
}
if (used_callee_saved.empty()) {
frame_info.callee_saved_size = 0; // 确保大小被初始化
return; // 无需操作
}
// 2. 计算为 callee-saved 寄存器分配的栈空间
// 这里的关键是,偏移的基准点要在局部变量和溢出槽之下。
int callee_saved_size = used_callee_saved.size() * 8;
frame_info.callee_saved_size = callee_saved_size; // 将大小存入 FrameInfo
// 3. 计算无冲突的栈偏移
// 栈向下增长,所以偏移是负数。
// ra/s0 占用 -8 和 -16。局部变量和溢出区在它们之下。callee-saved 区在更下方。
// 我们使用相对于 s0 的偏移。s0 将指向栈顶 (sp + total_size)。
int base_offset = -16 - frame_info.locals_size - frame_info.spill_size;
// 为了栈帧布局确定性,对寄存器进行排序
std::vector<PhysicalReg> sorted_regs(used_callee_saved.begin(), used_callee_saved.end());
std::sort(sorted_regs.begin(), sorted_regs.end());
// 4. 在函数序言插入保存指令
MachineBasicBlock* entry_block = mfunc->getBlocks().front().get();
auto& entry_instrs = entry_block->getInstructions();
auto prologue_end = entry_instrs.begin();
// 找到序言结束的位置通常是addi s0, sp, size之后但为了让优化器看到我们插在更前面
// 合理的位置是在 IR 指令开始之前,即在任何非序言指令(如第一个标签)之前。
// 为简单起见,我们直接插入到块的开头,后续重排 pass 会处理。
// (更优的实现会寻找一个特定的插入点)
int current_offset = base_offset;
for (PhysicalReg reg : sorted_regs) {
auto sd = std::make_unique<MachineInstr>(RVOpcodes::SD);
sd->addOperand(std::make_unique<RegOperand>(reg));
sd->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0), // 基址为帧指针 s0
std::make_unique<ImmOperand>(current_offset)
));
// 从头部插入,但要放在函数标签之后
entry_instrs.insert(entry_instrs.begin() + 1, std::move(sd));
current_offset -= 8;
}
// 5. 在函数结尾ret之前插入恢复指令使用反向遍历来避免迭代器失效
for (auto& mbb : mfunc->getBlocks()) {
// 使用手动控制的反向循环
for (auto it = mbb->getInstructions().begin(); it != mbb->getInstructions().end(); ++it) {
if ((*it)->getOpcode() == RVOpcodes::RET) {
// 1. 创建一个临时vector来存储所有需要插入的恢复指令
std::vector<std::unique_ptr<MachineInstr>> restore_instrs;
int current_offset_load = base_offset;
// 以相同的顺序(例如 s1, s2, ...)创建恢复指令
for (PhysicalReg reg : sorted_regs) {
auto ld = std::make_unique<MachineInstr>(RVOpcodes::LD);
ld->addOperand(std::make_unique<RegOperand>(reg));
ld->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::S0),
std::make_unique<ImmOperand>(current_offset_load)
));
restore_instrs.push_back(std::move(ld));
current_offset_load -= 8;
}
// 2. 使用 make_move_iterator 一次性将所有恢复指令插入到 RET 指令之前
// 这可以高效地转移指令的所有权,并且只让迭代器失效一次。
if (!restore_instrs.empty()) {
mbb->getInstructions().insert(it,
std::make_move_iterator(restore_instrs.begin()),
std::make_move_iterator(restore_instrs.end())
);
}
// 找到了RET并处理完毕后就可以跳出内层循环继续寻找下一个基本块
break;
}
}
}
}
} // namespace sysy

View File

@@ -0,0 +1,168 @@
#include "LegalizeImmediates.h"
#include "RISCv64ISel.h" // 需要包含它以调用 getNewVReg()
#include "RISCv64AsmPrinter.h"
#include <vector>
#include <iostream>
// 声明外部调试控制变量
extern int DEBUG;
extern int DEEPDEBUG;
namespace sysy {
char LegalizeImmediatesPass::ID = 0;
// 辅助函数检查一个立即数是否在RISC-V的12位有符号范围内
static bool isLegalImmediate(int64_t imm) {
return imm >= -2048 && imm <= 2047;
}
void LegalizeImmediatesPass::runOnMachineFunction(MachineFunction* mfunc) {
if (DEBUG) {
std::cerr << "===== Running Legalize Immediates Pass on function: " << mfunc->getName() << " =====\n";
}
// 定义我们保留的、用于暂存的物理寄存器
const PhysicalReg TEMP_REG = PhysicalReg::T5;
// 创建一个临时的AsmPrinter用于打印指令方便调试
RISCv64AsmPrinter temp_printer(mfunc);
if (DEEPDEBUG) {
temp_printer.setStream(std::cerr);
}
for (auto& mbb : mfunc->getBlocks()) {
if (DEEPDEBUG) {
std::cerr << "--- Processing Basic Block: " << mbb->getName() << " ---\n";
}
// 创建一个新的指令列表,用于存放合法化后的指令
std::vector<std::unique_ptr<MachineInstr>> new_instructions;
for (auto& instr_ptr : mbb->getInstructions()) {
if (DEEPDEBUG) {
std::cerr << " Checking: ";
// 打印指令时末尾会带换行符,所以这里不用 std::endl
temp_printer.printInstruction(instr_ptr.get(), true);
}
bool legalized = false; // 标记当前指令是否已被展开处理
switch (instr_ptr->getOpcode()) {
case RVOpcodes::ADDI:
case RVOpcodes::ADDIW: {
auto& operands = instr_ptr->getOperands();
auto imm_op = static_cast<ImmOperand*>(operands.back().get());
if (!isLegalImmediate(imm_op->getValue())) {
if (DEEPDEBUG) {
std::cerr << " >> ILLEGAL immediate (" << imm_op->getValue() << "). Expanding...\n";
}
// 立即数超出范围,需要展开
auto rd_op = std::make_unique<RegOperand>(*static_cast<RegOperand*>(operands[0].get()));
auto rs1_op = std::make_unique<RegOperand>(*static_cast<RegOperand*>(operands[1].get()));
// 1. li t5, immediate
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(TEMP_REG));
li->addOperand(std::make_unique<ImmOperand>(imm_op->getValue()));
// 2. add/addw rd, rs1, t5
auto new_op = (instr_ptr->getOpcode() == RVOpcodes::ADDI) ? RVOpcodes::ADD : RVOpcodes::ADDW;
auto add = std::make_unique<MachineInstr>(new_op);
add->addOperand(std::move(rd_op));
add->addOperand(std::move(rs1_op));
add->addOperand(std::make_unique<RegOperand>(TEMP_REG));
if (DEEPDEBUG) {
std::cerr << " New sequence:\n ";
temp_printer.printInstruction(li.get(), true);
std::cerr << " ";
temp_printer.printInstruction(add.get(), true);
}
new_instructions.push_back(std::move(li));
new_instructions.push_back(std::move(add));
legalized = true;
}
break;
}
// 处理所有内存加载/存储指令
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: {
auto& operands = instr_ptr->getOperands();
auto mem_op = static_cast<MemOperand*>(operands.back().get());
auto offset_op = mem_op->getOffset();
if (!isLegalImmediate(offset_op->getValue())) {
if (DEEPDEBUG) {
std::cerr << " >> ILLEGAL immediate offset (" << offset_op->getValue() << "). Expanding...\n";
}
// 偏移量超出范围,需要展开
auto data_reg_op = std::make_unique<RegOperand>(*static_cast<RegOperand*>(operands[0].get()));
auto base_reg_op = std::make_unique<RegOperand>(*mem_op->getBase());
// 1. li t5, offset
auto li = std::make_unique<MachineInstr>(RVOpcodes::LI);
li->addOperand(std::make_unique<RegOperand>(TEMP_REG));
li->addOperand(std::make_unique<ImmOperand>(offset_op->getValue()));
// 2. add t5, base_reg, t5 (计算最终地址结果也放在t5)
auto add = std::make_unique<MachineInstr>(RVOpcodes::ADD);
add->addOperand(std::make_unique<RegOperand>(TEMP_REG));
add->addOperand(std::move(base_reg_op));
add->addOperand(std::make_unique<RegOperand>(TEMP_REG));
// 3. lw/sw data_reg, 0(t5)
auto mem_instr = std::make_unique<MachineInstr>(instr_ptr->getOpcode());
mem_instr->addOperand(std::move(data_reg_op));
mem_instr->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(TEMP_REG),
std::make_unique<ImmOperand>(0)
));
if (DEEPDEBUG) {
std::cerr << " New sequence:\n ";
temp_printer.printInstruction(li.get(), true);
std::cerr << " ";
temp_printer.printInstruction(add.get(), true);
std::cerr << " ";
temp_printer.printInstruction(mem_instr.get(), true);
}
new_instructions.push_back(std::move(li));
new_instructions.push_back(std::move(add));
new_instructions.push_back(std::move(mem_instr));
legalized = true;
}
break;
}
default:
// 其他指令不需要处理
break;
}
if (!legalized) {
if (DEEPDEBUG) {
std::cerr << " -- Immediate is legal. Skipping.\n";
}
// 如果当前指令不需要合法化,直接将其移动到新列表中
new_instructions.push_back(std::move(instr_ptr));
}
}
// 用新的、已合法化的指令列表替换旧的列表
mbb->getInstructions() = std::move(new_instructions);
}
if (DEBUG) {
std::cerr << "===== Finished Legalize Immediates Pass =====\n\n";
}
}
} // namespace sysy

View File

@@ -0,0 +1,118 @@
#include "PrologueEpilogueInsertion.h"
namespace sysy {
char PrologueEpilogueInsertionPass::ID = 0;
void PrologueEpilogueInsertionPass::runOnMachineFunction(MachineFunction* mfunc) {
StackFrameInfo& frame_info = mfunc->getFrameInfo();
// 完全遵循 AsmPrinter 中的计算逻辑
int total_stack_size = frame_info.locals_size +
frame_info.spill_size +
frame_info.callee_saved_size +
16; // 为 ra 和 s0 固定的16字节
int aligned_stack_size = (total_stack_size + 15) & ~15;
frame_info.total_size = aligned_stack_size;
// 只有在需要分配栈空间时才生成指令
if (aligned_stack_size > 0) {
// --- 1. 插入序言 ---
MachineBasicBlock* entry_block = mfunc->getBlocks().front().get();
auto& entry_instrs = entry_block->getInstructions();
std::vector<std::unique_ptr<MachineInstr>> prologue_instrs;
// 严格按照 AsmPrinter 的打印顺序来创建和组织指令
// 1. addi sp, sp, -aligned_stack_size
auto alloc_stack = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
alloc_stack->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
alloc_stack->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
alloc_stack->addOperand(std::make_unique<ImmOperand>(-aligned_stack_size));
prologue_instrs.push_back(std::move(alloc_stack));
// 2. sd ra, (aligned_stack_size - 8)(sp)
auto save_ra = std::make_unique<MachineInstr>(RVOpcodes::SD);
save_ra->addOperand(std::make_unique<RegOperand>(PhysicalReg::RA));
save_ra->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::SP),
std::make_unique<ImmOperand>(aligned_stack_size - 8)
));
prologue_instrs.push_back(std::move(save_ra));
// 3. sd s0, (aligned_stack_size - 16)(sp)
auto save_fp = std::make_unique<MachineInstr>(RVOpcodes::SD);
save_fp->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
save_fp->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::SP),
std::make_unique<ImmOperand>(aligned_stack_size - 16)
));
prologue_instrs.push_back(std::move(save_fp));
// 4. addi s0, sp, aligned_stack_size
auto set_fp = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
set_fp->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
set_fp->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
set_fp->addOperand(std::make_unique<ImmOperand>(aligned_stack_size));
prologue_instrs.push_back(std::move(set_fp));
// 确定插入点(在函数名标签之后)
auto insert_pos = entry_instrs.begin();
// [重要] 这里我们不再需要跳过LABEL因为AsmPrinter将不再打印函数名标签
// 第一个基本块的标签就是函数入口
// 一次性将所有序言指令插入
if (!prologue_instrs.empty()) {
entry_instrs.insert(insert_pos,
std::make_move_iterator(prologue_instrs.begin()),
std::make_move_iterator(prologue_instrs.end()));
}
// --- 2. 插入尾声 ---
for (auto& mbb : mfunc->getBlocks()) {
for (auto it = mbb->getInstructions().begin(); it != mbb->getInstructions().end(); ++it) {
if ((*it)->getOpcode() == RVOpcodes::RET) {
std::vector<std::unique_ptr<MachineInstr>> epilogue_instrs;
// 同样严格按照 AsmPrinter 的打印顺序
// 1. ld ra, (aligned_stack_size - 8)(sp)
auto restore_ra = std::make_unique<MachineInstr>(RVOpcodes::LD);
restore_ra->addOperand(std::make_unique<RegOperand>(PhysicalReg::RA));
restore_ra->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::SP),
std::make_unique<ImmOperand>(aligned_stack_size - 8)
));
epilogue_instrs.push_back(std::move(restore_ra));
// 2. ld s0, (aligned_stack_size - 16)(sp)
auto restore_fp = std::make_unique<MachineInstr>(RVOpcodes::LD);
restore_fp->addOperand(std::make_unique<RegOperand>(PhysicalReg::S0));
restore_fp->addOperand(std::make_unique<MemOperand>(
std::make_unique<RegOperand>(PhysicalReg::SP),
std::make_unique<ImmOperand>(aligned_stack_size - 16)
));
epilogue_instrs.push_back(std::move(restore_fp));
// 3. addi sp, sp, aligned_stack_size
auto dealloc_stack = std::make_unique<MachineInstr>(RVOpcodes::ADDI);
dealloc_stack->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
dealloc_stack->addOperand(std::make_unique<RegOperand>(PhysicalReg::SP));
dealloc_stack->addOperand(std::make_unique<ImmOperand>(aligned_stack_size));
epilogue_instrs.push_back(std::move(dealloc_stack));
if (!epilogue_instrs.empty()) {
mbb->getInstructions().insert(it,
std::make_move_iterator(epilogue_instrs.begin()),
std::make_move_iterator(epilogue_instrs.end()));
}
// 处理完一个基本块中的RET后迭代器已失效需跳出
goto next_block;
}
}
next_block:;
}
}
}
} // namespace sysy