Files
mysysy/src/backend/RISCv64/Optimize/Peephole.cpp
2025-07-29 21:30:30 +08:00

652 lines
25 KiB
C++
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

#include "Peephole.h"
#include <functional>
namespace sysy {
char PeepholeOptimizer::ID = 0;
bool PeepholeOptimizer::runOnFunction(Function *F, AnalysisManager& AM) {
// This pass works on MachineFunction level, not IR level
return false;
}
void PeepholeOptimizer::runOnMachineFunction(MachineFunction *mfunc) {
if (!mfunc)
return;
using namespace sysy;
// areRegsEqual: 检查两个寄存器操作数是否相等(考虑虚拟和物理寄存器)。
auto areRegsEqual = [](RegOperand *r1, RegOperand *r2) {
if (!r1 || !r2 || r1->isVirtual() != r2->isVirtual()) {
return false;
}
if (r1->isVirtual()) {
return r1->getVRegNum() == r2->getVRegNum();
} else {
return r1->getPReg() == r2->getPReg();
}
};
// 改进的 isRegUsedLater 函数 - 更完整和准确的实现
auto isRegUsedLater =
[&](const std::vector<std::unique_ptr<MachineInstr>> &instrs,
RegOperand *reg, size_t start_idx) -> bool {
for (size_t j = start_idx; j < instrs.size(); ++j) {
auto *instr = instrs[j].get();
auto opcode = instr->getOpcode();
// 检查所有操作数
for (size_t k = 0; k < instr->getOperands().size(); ++k) {
bool isDefOperand = false;
// 更完整的定义操作数判断逻辑
if (k == 0) { // 第一个操作数通常是目标寄存器
switch (opcode) {
// 算术和逻辑指令 - 第一个操作数是定义
case RVOpcodes::MV:
case RVOpcodes::ADDI:
case RVOpcodes::SLLI:
case RVOpcodes::SRLI:
case RVOpcodes::SRAI:
case RVOpcodes::SLTI:
case RVOpcodes::SLTIU:
case RVOpcodes::XORI:
case RVOpcodes::ORI:
case RVOpcodes::ANDI:
case RVOpcodes::ADD:
case RVOpcodes::SUB:
case RVOpcodes::SLL:
case RVOpcodes::SLT:
case RVOpcodes::SLTU:
case RVOpcodes::XOR:
case RVOpcodes::SRL:
case RVOpcodes::SRA:
case RVOpcodes::OR:
case RVOpcodes::AND:
case RVOpcodes::MUL:
case RVOpcodes::DIV:
case RVOpcodes::REM:
case RVOpcodes::LW:
case RVOpcodes::LH:
case RVOpcodes::LB:
case RVOpcodes::LHU:
case RVOpcodes::LBU:
// 存储指令 - 第一个操作数是使用(要存储的值)
case RVOpcodes::SW:
case RVOpcodes::SH:
case RVOpcodes::SB:
// 分支指令 - 第一个操作数是使用
case RVOpcodes::BEQ:
case RVOpcodes::BNE:
case RVOpcodes::BLT:
case RVOpcodes::BGE:
case RVOpcodes::BLTU:
case RVOpcodes::BGEU:
// 跳转指令 - 可能使用寄存器
case RVOpcodes::JALR:
isDefOperand = false;
break;
default:
// 对于未知指令,保守地假设第一个操作数可能是使用
isDefOperand = false;
break;
}
}
// 如果不是定义操作数,检查是否使用了目标寄存器
if (!isDefOperand) {
if (instr->getOperands()[k]->getKind() == MachineOperand::KIND_REG) {
auto *use_reg =
static_cast<RegOperand *>(instr->getOperands()[k].get());
if (areRegsEqual(reg, use_reg))
return true;
}
// 检查内存操作数中的基址寄存器
if (instr->getOperands()[k]->getKind() == MachineOperand::KIND_MEM) {
auto *mem =
static_cast<MemOperand *>(instr->getOperands()[k].get());
if (areRegsEqual(reg, mem->getBase()))
return true;
}
}
}
}
return false;
};
// 检查寄存器是否在指令中被重新定义(用于更精确的分析)
auto isRegRedefinedAt =
[](MachineInstr *instr, RegOperand *reg,
const std::function<bool(RegOperand *, RegOperand *)> &areRegsEqual)
-> bool {
if (instr->getOperands().empty())
return false;
auto opcode = instr->getOpcode();
// 只有当第一个操作数是定义操作数时才检查
switch (opcode) {
case RVOpcodes::MV:
case RVOpcodes::ADDI:
case RVOpcodes::ADD:
case RVOpcodes::SUB:
case RVOpcodes::MUL:
case RVOpcodes::LW:
// ... 其他定义指令
if (instr->getOperands()[0]->getKind() == MachineOperand::KIND_REG) {
auto *def_reg =
static_cast<RegOperand *>(instr->getOperands()[0].get());
return areRegsEqual(reg, def_reg);
}
break;
default:
break;
}
return false;
};
// 检查是否为存储-加载模式,支持不同大小的访问
auto isStoreLoadPattern = [](MachineInstr *store_instr,
MachineInstr *load_instr) -> bool {
auto store_op = store_instr->getOpcode();
auto load_op = load_instr->getOpcode();
// 检查存储-加载对应关系
return (store_op == RVOpcodes::SW && load_op == RVOpcodes::LW) || // 32位
(store_op == RVOpcodes::SH &&
load_op == RVOpcodes::LH) || // 16位有符号
(store_op == RVOpcodes::SH &&
load_op == RVOpcodes::LHU) || // 16位无符号
(store_op == RVOpcodes::SB &&
load_op == RVOpcodes::LB) || // 8位有符号
(store_op == RVOpcodes::SB &&
load_op == RVOpcodes::LBU) || // 8位无符号
(store_op == RVOpcodes::SD && load_op == RVOpcodes::LD); // 64位
};
// 检查两个内存访问是否访问相同的内存位置
auto areMemoryAccessesEqual =
[&areRegsEqual](MachineInstr *store_instr, MemOperand *store_mem,
MachineInstr *load_instr, MemOperand *load_mem) -> bool {
// 基址寄存器必须相同
if (!areRegsEqual(store_mem->getBase(), load_mem->getBase())) {
return false;
}
// 偏移量必须相同
if (store_mem->getOffset()->getValue() !=
load_mem->getOffset()->getValue()) {
return false;
}
// 检查访问大小是否兼容
auto store_op = store_instr->getOpcode();
auto load_op = load_instr->getOpcode();
// 获取访问大小(字节数)
auto getAccessSize = [](RVOpcodes opcode) -> int {
switch (opcode) {
case RVOpcodes::LB:
case RVOpcodes::LBU:
case RVOpcodes::SB:
return 1; // 8位
case RVOpcodes::LH:
case RVOpcodes::LHU:
case RVOpcodes::SH:
return 2; // 16位
case RVOpcodes::LW:
case RVOpcodes::SW:
return 4; // 32位
case RVOpcodes::LD:
case RVOpcodes::SD:
return 8; // 64位
default:
return -1; // 未知
}
};
int store_size = getAccessSize(store_op);
int load_size = getAccessSize(load_op);
// 只有访问大小完全匹配时才能进行优化
// 这避免了部分重叠访问的复杂情况
return store_size > 0 && store_size == load_size;
};
// 简单的内存别名分析:检查两个内存访问之间是否可能有冲突的内存操作
auto isMemoryAccessSafe =
[&](const std::vector<std::unique_ptr<MachineInstr>> &instrs,
size_t store_idx, size_t load_idx, MemOperand *mem) -> bool {
// 检查存储和加载之间是否有可能影响内存的指令
for (size_t j = store_idx + 1; j < load_idx; ++j) {
auto *between_instr = instrs[j].get();
auto between_op = between_instr->getOpcode();
// 检查是否有其他内存写入操作
switch (between_op) {
case RVOpcodes::SW:
case RVOpcodes::SH:
case RVOpcodes::SB:
case RVOpcodes::SD: {
// 如果有其他存储操作,需要检查是否可能访问相同的内存
if (between_instr->getOperands().size() >= 2 &&
between_instr->getOperands()[1]->getKind() ==
MachineOperand::KIND_MEM) {
auto *other_mem =
static_cast<MemOperand *>(between_instr->getOperands()[1].get());
// 保守的别名分析:如果使用不同的基址寄存器,假设可能别名
if (!areRegsEqual(mem->getBase(), other_mem->getBase())) {
return false; // 可能的别名,不安全
}
// 如果基址相同但偏移量不同,检查是否重叠
int64_t offset1 = mem->getOffset()->getValue();
int64_t offset2 = other_mem->getOffset()->getValue();
// 获取访问大小来检查重叠
auto getAccessSize = [](RVOpcodes opcode) -> int {
switch (opcode) {
case RVOpcodes::SB:
return 1;
case RVOpcodes::SH:
return 2;
case RVOpcodes::SW:
return 4;
case RVOpcodes::SD:
return 8;
default:
return 4; // 默认假设4字节
}
};
int size1 = getAccessSize(RVOpcodes::SW); // 从原存储指令推断
int size2 = getAccessSize(between_op);
// 检查内存区域是否重叠
bool overlaps =
!(offset1 + size1 <= offset2 || offset2 + size2 <= offset1);
if (overlaps) {
return false; // 内存重叠,不安全
}
}
break;
}
// 函数调用可能有副作用
case RVOpcodes::JAL:
case RVOpcodes::JALR:
return false; // 函数调用可能修改内存,不安全
// 原子操作或其他可能修改内存的指令
// 根据具体的RISC-V扩展添加更多指令
default:
// 对于未知指令,采用保守策略
// 可以根据具体需求调整
break;
}
}
return true; // 没有发现潜在的内存冲突
};
// isPowerOfTwo: 检查数值是否为2的幂次并返回其指数。
auto isPowerOfTwo = [](int64_t n) -> int {
if (n <= 0 || (n & (n - 1)) != 0)
return -1;
int shift = 0;
while (n > 1) {
n >>= 1;
shift++;
}
return shift;
};
for (auto &mbb_uptr : mfunc->getBlocks()) {
auto &mbb = *mbb_uptr;
auto &instrs = mbb.getInstructions();
if (instrs.size() < 2)
continue; // 基本块至少需要两条指令进行窥孔
// 遍历指令序列进行窥孔优化
for (size_t i = 0; i + 1 < instrs.size();) {
auto *mi1 = instrs[i].get();
auto *mi2 = instrs[i + 1].get();
bool changed = false;
// 1. 消除冗余交换移动: mv a, b; mv b, a -> mv a, b
if (mi1->getOpcode() == RVOpcodes::MV &&
mi2->getOpcode() == RVOpcodes::MV) {
if (mi1->getOperands().size() == 2 && mi2->getOperands().size() == 2) {
if (mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_REG &&
mi2->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi2->getOperands()[1]->getKind() == MachineOperand::KIND_REG) {
auto *dst1 = static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *src1 = static_cast<RegOperand *>(mi1->getOperands()[1].get());
auto *dst2 = static_cast<RegOperand *>(mi2->getOperands()[0].get());
auto *src2 = static_cast<RegOperand *>(mi2->getOperands()[1].get());
if (areRegsEqual(dst1, src2) && areRegsEqual(src1, dst2)) {
instrs.erase(instrs.begin() + i + 1); // 移除第二条指令
changed = true;
}
}
}
}
// 2. 冗余加载消除: sw t0, offset(base); lw t1, offset(base) -> 替换或消除
// lw 添加ld sd支持
else if (isStoreLoadPattern(mi1, mi2)) {
if (mi1->getOperands().size() == 2 && mi2->getOperands().size() == 2) {
if (mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_MEM &&
mi2->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi2->getOperands()[1]->getKind() == MachineOperand::KIND_MEM) {
auto *store_val =
static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *store_mem =
static_cast<MemOperand *>(mi1->getOperands()[1].get());
auto *load_val =
static_cast<RegOperand *>(mi2->getOperands()[0].get());
auto *load_mem =
static_cast<MemOperand *>(mi2->getOperands()[1].get());
// 检查内存访问是否匹配(基址、偏移量和访问大小)
if (areMemoryAccessesEqual(mi1, store_mem, mi2, load_mem)) {
// 进行简单的内存别名分析
if (isMemoryAccessSafe(instrs, i, i + 1, store_mem)) {
if (areRegsEqual(store_val, load_val)) {
// sw r1, mem; lw r1, mem -> 消除冗余的lw
instrs.erase(instrs.begin() + i + 1);
changed = true;
} else {
// sw r1, mem; lw r2, mem -> 替换lw为mv r2, r1
auto newInstr = std::make_unique<MachineInstr>(RVOpcodes::MV);
newInstr->addOperand(std::make_unique<RegOperand>(*load_val));
newInstr->addOperand(
std::make_unique<RegOperand>(*store_val));
instrs[i + 1] = std::move(newInstr);
changed = true;
}
}
}
}
}
}
// 3. 强度削减: mul y, x, 2^n -> slli y, x, n
else if (mi1->getOpcode() == RVOpcodes::MUL &&
mi1->getOperands().size() == 3) {
auto *dst_op = mi1->getOperands()[0].get();
auto *src1_op = mi1->getOperands()[1].get();
auto *src2_op = mi1->getOperands()[2].get();
if (dst_op->getKind() == MachineOperand::KIND_REG) {
auto *dst_reg = static_cast<RegOperand *>(dst_op);
RegOperand *src_reg = nullptr;
int shift = -1;
if (src1_op->getKind() == MachineOperand::KIND_REG &&
src2_op->getKind() == MachineOperand::KIND_IMM) {
shift =
isPowerOfTwo(static_cast<ImmOperand *>(src2_op)->getValue());
if (shift >= 0)
src_reg = static_cast<RegOperand *>(src1_op);
} else if (src1_op->getKind() == MachineOperand::KIND_IMM &&
src2_op->getKind() == MachineOperand::KIND_REG) {
shift =
isPowerOfTwo(static_cast<ImmOperand *>(src1_op)->getValue());
if (shift >= 0)
src_reg = static_cast<RegOperand *>(src2_op);
}
if (src_reg && shift >= 0 &&
shift <= 31) { // RISC-V 移位量限制 (0-31)
auto newInstr = std::make_unique<MachineInstr>(RVOpcodes::SLLI);
newInstr->addOperand(std::make_unique<RegOperand>(*dst_reg));
newInstr->addOperand(std::make_unique<RegOperand>(*src_reg));
newInstr->addOperand(std::make_unique<ImmOperand>(shift));
instrs[i] = std::move(newInstr);
changed = true;
}
}
}
// 4. 地址计算优化: addi dst, base, imm1; lw/sw val, imm2(dst) -> lw/sw
// val, (imm1+imm2)(base)
else if (mi1->getOpcode() == RVOpcodes::ADDI &&
mi1->getOperands().size() == 3) {
auto opcode2 = mi2->getOpcode();
if (opcode2 == RVOpcodes::LW || opcode2 == RVOpcodes::SW) {
if (mi2->getOperands().size() == 2 &&
mi2->getOperands()[1]->getKind() == MachineOperand::KIND_MEM &&
mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[2]->getKind() == MachineOperand::KIND_IMM) {
auto *addi_dst =
static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *addi_base =
static_cast<RegOperand *>(mi1->getOperands()[1].get());
auto *addi_imm =
static_cast<ImmOperand *>(mi1->getOperands()[2].get());
auto *mem_op =
static_cast<MemOperand *>(mi2->getOperands()[1].get());
auto *mem_base = mem_op->getBase();
auto *mem_imm = mem_op->getOffset();
// 检查 ADDI 的目标寄存器是否是内存操作的基址
if (areRegsEqual(addi_dst, mem_base)) {
// 改进的使用检查:考虑寄存器可能在后续被重新定义的情况
bool canOptimize = true;
// 检查从 i+2 开始的指令
for (size_t j = i + 2; j < instrs.size(); ++j) {
auto *later_instr = instrs[j].get();
// 如果寄存器被重新定义,那么它后面的使用就不相关了
if (isRegRedefinedAt(later_instr, addi_dst, areRegsEqual)) {
break; // 寄存器被重新定义,可以安全优化
}
// 如果寄存器被使用,则不能优化
if (isRegUsedLater(instrs, addi_dst, j)) {
canOptimize = false;
break;
}
}
if (canOptimize) {
int64_t new_offset = addi_imm->getValue() + mem_imm->getValue();
// 检查新偏移量是否符合 RISC-V 12位有符号立即数范围
if (new_offset >= -2048 && new_offset <= 2047) {
auto new_mem_op = std::make_unique<MemOperand>(
std::make_unique<RegOperand>(*addi_base),
std::make_unique<ImmOperand>(new_offset));
mi2->getOperands()[1] = std::move(new_mem_op);
instrs.erase(instrs.begin() + i);
changed = true;
}
}
}
}
}
}
// 5. 冗余移动指令消除: mv x, y; op z, x, ... -> op z, y, ... (如果 x
// 之后不再使用)
else if (mi1->getOpcode() == RVOpcodes::MV &&
mi1->getOperands().size() == 2) {
if (mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_REG) {
auto *mv_dst = static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *mv_src = static_cast<RegOperand *>(mi1->getOperands()[1].get());
// 检查第二条指令是否使用了 mv 的目标寄存器
std::vector<size_t> use_positions;
for (size_t k = 1; k < mi2->getOperands().size(); ++k) {
if (mi2->getOperands()[k]->getKind() == MachineOperand::KIND_REG) {
auto *use_reg =
static_cast<RegOperand *>(mi2->getOperands()[k].get());
if (areRegsEqual(mv_dst, use_reg)) {
use_positions.push_back(k);
}
}
// 也检查内存操作数中的基址寄存器
else if (mi2->getOperands()[k]->getKind() ==
MachineOperand::KIND_MEM) {
auto *mem =
static_cast<MemOperand *>(mi2->getOperands()[k].get());
if (areRegsEqual(mv_dst, mem->getBase())) {
// 对于内存操作数我们需要创建新的MemOperand
auto new_mem = std::make_unique<MemOperand>(
std::make_unique<RegOperand>(*mv_src),
std::make_unique<ImmOperand>(mem->getOffset()->getValue()));
mi2->getOperands()[k] = std::move(new_mem);
use_positions.push_back(k); // 标记已处理
}
}
}
if (!use_positions.empty()) {
// 改进的后续使用检查
bool canOptimize = true;
for (size_t j = i + 2; j < instrs.size(); ++j) {
auto *later_instr = instrs[j].get();
// 如果寄存器被重新定义,后续使用就不相关了
if (isRegRedefinedAt(later_instr, mv_dst, areRegsEqual)) {
break;
}
// 检查是否还有其他使用
if (isRegUsedLater(instrs, mv_dst, j)) {
canOptimize = false;
break;
}
}
if (canOptimize) {
// 替换所有寄存器使用(内存操作数已在上面处理)
for (size_t pos : use_positions) {
if (mi2->getOperands()[pos]->getKind() ==
MachineOperand::KIND_REG) {
mi2->getOperands()[pos] =
std::make_unique<RegOperand>(*mv_src);
}
}
instrs.erase(instrs.begin() + i);
changed = true;
}
}
}
}
// 6. 连续加法指令合并: addi t1, t0, imm1; addi t2, t1, imm2 -> addi t2,
// t0, (imm1+imm2)
else if (mi1->getOpcode() == RVOpcodes::ADDI &&
mi2->getOpcode() == RVOpcodes::ADDI) {
if (mi1->getOperands().size() == 3 && mi2->getOperands().size() == 3) {
if (mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[2]->getKind() == MachineOperand::KIND_IMM &&
mi2->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi2->getOperands()[1]->getKind() == MachineOperand::KIND_REG &&
mi2->getOperands()[2]->getKind() == MachineOperand::KIND_IMM) {
auto *addi1_dst =
static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *addi1_src =
static_cast<RegOperand *>(mi1->getOperands()[1].get());
auto *addi1_imm =
static_cast<ImmOperand *>(mi1->getOperands()[2].get());
auto *addi2_dst =
static_cast<RegOperand *>(mi2->getOperands()[0].get());
auto *addi2_src =
static_cast<RegOperand *>(mi2->getOperands()[1].get());
auto *addi2_imm =
static_cast<ImmOperand *>(mi2->getOperands()[2].get());
// 检查第一个ADDI的目标是否是第二个ADDI的源
if (areRegsEqual(addi1_dst, addi2_src)) {
// 改进的中间寄存器使用检查
bool canOptimize = true;
for (size_t j = i + 2; j < instrs.size(); ++j) {
auto *later_instr = instrs[j].get();
// 如果中间寄存器被重新定义,后续使用不相关
if (isRegRedefinedAt(later_instr, addi1_dst, areRegsEqual)) {
break;
}
// 检查是否有其他使用
if (isRegUsedLater(instrs, addi1_dst, j)) {
canOptimize = false;
break;
}
}
if (canOptimize) {
int64_t new_imm = addi1_imm->getValue() + addi2_imm->getValue();
// 检查新立即数范围
if (new_imm >= -2048 && new_imm <= 2047) {
auto newInstr =
std::make_unique<MachineInstr>(RVOpcodes::ADDI);
newInstr->addOperand(
std::make_unique<RegOperand>(*addi2_dst));
newInstr->addOperand(
std::make_unique<RegOperand>(*addi1_src));
newInstr->addOperand(std::make_unique<ImmOperand>(new_imm));
instrs[i + 1] = std::move(newInstr);
instrs.erase(instrs.begin() + i);
changed = true;
}
}
}
}
}
}
// 7. ADD with zero optimization: add r1, r2, zero -> mv r1, r2
else if (mi1->getOpcode() == RVOpcodes::ADD &&
mi1->getOperands().size() == 3) {
if (mi1->getOperands()[0]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[1]->getKind() == MachineOperand::KIND_REG &&
mi1->getOperands()[2]->getKind() == MachineOperand::KIND_REG) {
auto *add_dst =
static_cast<RegOperand *>(mi1->getOperands()[0].get());
auto *add_src1 =
static_cast<RegOperand *>(mi1->getOperands()[1].get());
auto *add_src2 =
static_cast<RegOperand *>(mi1->getOperands()[2].get());
// 检查第二个源操作数是否为ZERO寄存器
if (!add_src2->isVirtual() &&
add_src2->getPReg() == PhysicalReg::ZERO) {
// 创建新的 MV 指令
auto newInstr = std::make_unique<MachineInstr>(RVOpcodes::MV);
newInstr->addOperand(std::make_unique<RegOperand>(*add_dst));
newInstr->addOperand(std::make_unique<RegOperand>(*add_src1));
instrs[i] = std::move(newInstr);
changed = true;
}
}
}
// 根据是否发生变化调整遍历索引
if (!changed) {
++i; // 没有优化,继续检查下一对指令
} else {
// 发生变化,适当回退以捕获新的优化机会。
// 这是一种安全的回退策略,可以触发连锁优化,且不会导致无限循环。
if (i > 0) {
--i;
}
}
}
}
}
} // namespace sysy