1 //===- HexagonBitSimplify.cpp ---------------------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #include "BitTracker.h" 11 #include "HexagonBitTracker.h" 12 #include "HexagonInstrInfo.h" 13 #include "HexagonRegisterInfo.h" 14 #include "HexagonSubtarget.h" 15 #include "llvm/ADT/BitVector.h" 16 #include "llvm/ADT/DenseMap.h" 17 #include "llvm/ADT/GraphTraits.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SmallVector.h" 20 #include "llvm/ADT/StringRef.h" 21 #include "llvm/CodeGen/MachineBasicBlock.h" 22 #include "llvm/CodeGen/MachineDominators.h" 23 #include "llvm/CodeGen/MachineFunction.h" 24 #include "llvm/CodeGen/MachineFunctionPass.h" 25 #include "llvm/CodeGen/MachineInstr.h" 26 #include "llvm/CodeGen/MachineInstrBuilder.h" 27 #include "llvm/CodeGen/MachineOperand.h" 28 #include "llvm/CodeGen/MachineRegisterInfo.h" 29 #include "llvm/CodeGen/TargetRegisterInfo.h" 30 #include "llvm/IR/DebugLoc.h" 31 #include "llvm/MC/MCInstrDesc.h" 32 #include "llvm/Pass.h" 33 #include "llvm/Support/CommandLine.h" 34 #include "llvm/Support/Compiler.h" 35 #include "llvm/Support/Debug.h" 36 #include "llvm/Support/ErrorHandling.h" 37 #include "llvm/Support/MathExtras.h" 38 #include "llvm/Support/raw_ostream.h" 39 #include <algorithm> 40 #include <cassert> 41 #include <cstdint> 42 #include <iterator> 43 #include <limits> 44 #include <utility> 45 #include <vector> 46 47 #define DEBUG_TYPE "hexbit" 48 49 using namespace llvm; 50 51 static cl::opt<bool> PreserveTiedOps("hexbit-keep-tied", cl::Hidden, 52 cl::init(true), cl::desc("Preserve subregisters in tied operands")); 53 static cl::opt<bool> GenExtract("hexbit-extract", cl::Hidden, 54 cl::init(true), cl::desc("Generate extract instructions")); 55 static cl::opt<bool> GenBitSplit("hexbit-bitsplit", cl::Hidden, 56 cl::init(true), cl::desc("Generate bitsplit instructions")); 57 58 static cl::opt<unsigned> MaxExtract("hexbit-max-extract", cl::Hidden, 59 cl::init(std::numeric_limits<unsigned>::max())); 60 static unsigned CountExtract = 0; 61 static cl::opt<unsigned> MaxBitSplit("hexbit-max-bitsplit", cl::Hidden, 62 cl::init(std::numeric_limits<unsigned>::max())); 63 static unsigned CountBitSplit = 0; 64 65 namespace llvm { 66 67 void initializeHexagonBitSimplifyPass(PassRegistry& Registry); 68 FunctionPass *createHexagonBitSimplify(); 69 70 } // end namespace llvm 71 72 namespace { 73 74 // Set of virtual registers, based on BitVector. 75 struct RegisterSet : private BitVector { 76 RegisterSet() = default; 77 explicit RegisterSet(unsigned s, bool t = false) : BitVector(s, t) {} 78 RegisterSet(const RegisterSet &RS) = default; 79 80 using BitVector::clear; 81 using BitVector::count; 82 83 unsigned find_first() const { 84 int First = BitVector::find_first(); 85 if (First < 0) 86 return 0; 87 return x2v(First); 88 } 89 90 unsigned find_next(unsigned Prev) const { 91 int Next = BitVector::find_next(v2x(Prev)); 92 if (Next < 0) 93 return 0; 94 return x2v(Next); 95 } 96 97 RegisterSet &insert(unsigned R) { 98 unsigned Idx = v2x(R); 99 ensure(Idx); 100 return static_cast<RegisterSet&>(BitVector::set(Idx)); 101 } 102 RegisterSet &remove(unsigned R) { 103 unsigned Idx = v2x(R); 104 if (Idx >= size()) 105 return *this; 106 return static_cast<RegisterSet&>(BitVector::reset(Idx)); 107 } 108 109 RegisterSet &insert(const RegisterSet &Rs) { 110 return static_cast<RegisterSet&>(BitVector::operator|=(Rs)); 111 } 112 RegisterSet &remove(const RegisterSet &Rs) { 113 return static_cast<RegisterSet&>(BitVector::reset(Rs)); 114 } 115 116 reference operator[](unsigned R) { 117 unsigned Idx = v2x(R); 118 ensure(Idx); 119 return BitVector::operator[](Idx); 120 } 121 bool operator[](unsigned R) const { 122 unsigned Idx = v2x(R); 123 assert(Idx < size()); 124 return BitVector::operator[](Idx); 125 } 126 bool has(unsigned R) const { 127 unsigned Idx = v2x(R); 128 if (Idx >= size()) 129 return false; 130 return BitVector::test(Idx); 131 } 132 133 bool empty() const { 134 return !BitVector::any(); 135 } 136 bool includes(const RegisterSet &Rs) const { 137 // A.BitVector::test(B) <=> A-B != {} 138 return !Rs.BitVector::test(*this); 139 } 140 bool intersects(const RegisterSet &Rs) const { 141 return BitVector::anyCommon(Rs); 142 } 143 144 private: 145 void ensure(unsigned Idx) { 146 if (size() <= Idx) 147 resize(std::max(Idx+1, 32U)); 148 } 149 150 static inline unsigned v2x(unsigned v) { 151 return TargetRegisterInfo::virtReg2Index(v); 152 } 153 154 static inline unsigned x2v(unsigned x) { 155 return TargetRegisterInfo::index2VirtReg(x); 156 } 157 }; 158 159 struct PrintRegSet { 160 PrintRegSet(const RegisterSet &S, const TargetRegisterInfo *RI) 161 : RS(S), TRI(RI) {} 162 163 friend raw_ostream &operator<< (raw_ostream &OS, 164 const PrintRegSet &P); 165 166 private: 167 const RegisterSet &RS; 168 const TargetRegisterInfo *TRI; 169 }; 170 171 raw_ostream &operator<< (raw_ostream &OS, const PrintRegSet &P) 172 LLVM_ATTRIBUTE_UNUSED; 173 raw_ostream &operator<< (raw_ostream &OS, const PrintRegSet &P) { 174 OS << '{'; 175 for (unsigned R = P.RS.find_first(); R; R = P.RS.find_next(R)) 176 OS << ' ' << printReg(R, P.TRI); 177 OS << " }"; 178 return OS; 179 } 180 181 class Transformation; 182 183 class HexagonBitSimplify : public MachineFunctionPass { 184 public: 185 static char ID; 186 187 HexagonBitSimplify() : MachineFunctionPass(ID) {} 188 189 StringRef getPassName() const override { 190 return "Hexagon bit simplification"; 191 } 192 193 void getAnalysisUsage(AnalysisUsage &AU) const override { 194 AU.addRequired<MachineDominatorTree>(); 195 AU.addPreserved<MachineDominatorTree>(); 196 MachineFunctionPass::getAnalysisUsage(AU); 197 } 198 199 bool runOnMachineFunction(MachineFunction &MF) override; 200 201 static void getInstrDefs(const MachineInstr &MI, RegisterSet &Defs); 202 static void getInstrUses(const MachineInstr &MI, RegisterSet &Uses); 203 static bool isEqual(const BitTracker::RegisterCell &RC1, uint16_t B1, 204 const BitTracker::RegisterCell &RC2, uint16_t B2, uint16_t W); 205 static bool isZero(const BitTracker::RegisterCell &RC, uint16_t B, 206 uint16_t W); 207 static bool getConst(const BitTracker::RegisterCell &RC, uint16_t B, 208 uint16_t W, uint64_t &U); 209 static bool replaceReg(unsigned OldR, unsigned NewR, 210 MachineRegisterInfo &MRI); 211 static bool getSubregMask(const BitTracker::RegisterRef &RR, 212 unsigned &Begin, unsigned &Width, MachineRegisterInfo &MRI); 213 static bool replaceRegWithSub(unsigned OldR, unsigned NewR, 214 unsigned NewSR, MachineRegisterInfo &MRI); 215 static bool replaceSubWithSub(unsigned OldR, unsigned OldSR, 216 unsigned NewR, unsigned NewSR, MachineRegisterInfo &MRI); 217 static bool parseRegSequence(const MachineInstr &I, 218 BitTracker::RegisterRef &SL, BitTracker::RegisterRef &SH, 219 const MachineRegisterInfo &MRI); 220 221 static bool getUsedBitsInStore(unsigned Opc, BitVector &Bits, 222 uint16_t Begin); 223 static bool getUsedBits(unsigned Opc, unsigned OpN, BitVector &Bits, 224 uint16_t Begin, const HexagonInstrInfo &HII); 225 226 static const TargetRegisterClass *getFinalVRegClass( 227 const BitTracker::RegisterRef &RR, MachineRegisterInfo &MRI); 228 static bool isTransparentCopy(const BitTracker::RegisterRef &RD, 229 const BitTracker::RegisterRef &RS, MachineRegisterInfo &MRI); 230 231 private: 232 MachineDominatorTree *MDT = nullptr; 233 234 bool visitBlock(MachineBasicBlock &B, Transformation &T, RegisterSet &AVs); 235 static bool hasTiedUse(unsigned Reg, MachineRegisterInfo &MRI, 236 unsigned NewSub = Hexagon::NoSubRegister); 237 }; 238 239 using HBS = HexagonBitSimplify; 240 241 // The purpose of this class is to provide a common facility to traverse 242 // the function top-down or bottom-up via the dominator tree, and keep 243 // track of the available registers. 244 class Transformation { 245 public: 246 bool TopDown; 247 248 Transformation(bool TD) : TopDown(TD) {} 249 virtual ~Transformation() = default; 250 251 virtual bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) = 0; 252 }; 253 254 } // end anonymous namespace 255 256 char HexagonBitSimplify::ID = 0; 257 258 INITIALIZE_PASS_BEGIN(HexagonBitSimplify, "hexagon-bit-simplify", 259 "Hexagon bit simplification", false, false) 260 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 261 INITIALIZE_PASS_END(HexagonBitSimplify, "hexagon-bit-simplify", 262 "Hexagon bit simplification", false, false) 263 264 bool HexagonBitSimplify::visitBlock(MachineBasicBlock &B, Transformation &T, 265 RegisterSet &AVs) { 266 bool Changed = false; 267 268 if (T.TopDown) 269 Changed = T.processBlock(B, AVs); 270 271 RegisterSet Defs; 272 for (auto &I : B) 273 getInstrDefs(I, Defs); 274 RegisterSet NewAVs = AVs; 275 NewAVs.insert(Defs); 276 277 for (auto *DTN : children<MachineDomTreeNode*>(MDT->getNode(&B))) 278 Changed |= visitBlock(*(DTN->getBlock()), T, NewAVs); 279 280 if (!T.TopDown) 281 Changed |= T.processBlock(B, AVs); 282 283 return Changed; 284 } 285 286 // 287 // Utility functions: 288 // 289 void HexagonBitSimplify::getInstrDefs(const MachineInstr &MI, 290 RegisterSet &Defs) { 291 for (auto &Op : MI.operands()) { 292 if (!Op.isReg() || !Op.isDef()) 293 continue; 294 unsigned R = Op.getReg(); 295 if (!TargetRegisterInfo::isVirtualRegister(R)) 296 continue; 297 Defs.insert(R); 298 } 299 } 300 301 void HexagonBitSimplify::getInstrUses(const MachineInstr &MI, 302 RegisterSet &Uses) { 303 for (auto &Op : MI.operands()) { 304 if (!Op.isReg() || !Op.isUse()) 305 continue; 306 unsigned R = Op.getReg(); 307 if (!TargetRegisterInfo::isVirtualRegister(R)) 308 continue; 309 Uses.insert(R); 310 } 311 } 312 313 // Check if all the bits in range [B, E) in both cells are equal. 314 bool HexagonBitSimplify::isEqual(const BitTracker::RegisterCell &RC1, 315 uint16_t B1, const BitTracker::RegisterCell &RC2, uint16_t B2, 316 uint16_t W) { 317 for (uint16_t i = 0; i < W; ++i) { 318 // If RC1[i] is "bottom", it cannot be proven equal to RC2[i]. 319 if (RC1[B1+i].Type == BitTracker::BitValue::Ref && RC1[B1+i].RefI.Reg == 0) 320 return false; 321 // Same for RC2[i]. 322 if (RC2[B2+i].Type == BitTracker::BitValue::Ref && RC2[B2+i].RefI.Reg == 0) 323 return false; 324 if (RC1[B1+i] != RC2[B2+i]) 325 return false; 326 } 327 return true; 328 } 329 330 bool HexagonBitSimplify::isZero(const BitTracker::RegisterCell &RC, 331 uint16_t B, uint16_t W) { 332 assert(B < RC.width() && B+W <= RC.width()); 333 for (uint16_t i = B; i < B+W; ++i) 334 if (!RC[i].is(0)) 335 return false; 336 return true; 337 } 338 339 bool HexagonBitSimplify::getConst(const BitTracker::RegisterCell &RC, 340 uint16_t B, uint16_t W, uint64_t &U) { 341 assert(B < RC.width() && B+W <= RC.width()); 342 int64_t T = 0; 343 for (uint16_t i = B+W; i > B; --i) { 344 const BitTracker::BitValue &BV = RC[i-1]; 345 T <<= 1; 346 if (BV.is(1)) 347 T |= 1; 348 else if (!BV.is(0)) 349 return false; 350 } 351 U = T; 352 return true; 353 } 354 355 bool HexagonBitSimplify::replaceReg(unsigned OldR, unsigned NewR, 356 MachineRegisterInfo &MRI) { 357 if (!TargetRegisterInfo::isVirtualRegister(OldR) || 358 !TargetRegisterInfo::isVirtualRegister(NewR)) 359 return false; 360 auto Begin = MRI.use_begin(OldR), End = MRI.use_end(); 361 decltype(End) NextI; 362 for (auto I = Begin; I != End; I = NextI) { 363 NextI = std::next(I); 364 I->setReg(NewR); 365 } 366 return Begin != End; 367 } 368 369 bool HexagonBitSimplify::replaceRegWithSub(unsigned OldR, unsigned NewR, 370 unsigned NewSR, MachineRegisterInfo &MRI) { 371 if (!TargetRegisterInfo::isVirtualRegister(OldR) || 372 !TargetRegisterInfo::isVirtualRegister(NewR)) 373 return false; 374 if (hasTiedUse(OldR, MRI, NewSR)) 375 return false; 376 auto Begin = MRI.use_begin(OldR), End = MRI.use_end(); 377 decltype(End) NextI; 378 for (auto I = Begin; I != End; I = NextI) { 379 NextI = std::next(I); 380 I->setReg(NewR); 381 I->setSubReg(NewSR); 382 } 383 return Begin != End; 384 } 385 386 bool HexagonBitSimplify::replaceSubWithSub(unsigned OldR, unsigned OldSR, 387 unsigned NewR, unsigned NewSR, MachineRegisterInfo &MRI) { 388 if (!TargetRegisterInfo::isVirtualRegister(OldR) || 389 !TargetRegisterInfo::isVirtualRegister(NewR)) 390 return false; 391 if (OldSR != NewSR && hasTiedUse(OldR, MRI, NewSR)) 392 return false; 393 auto Begin = MRI.use_begin(OldR), End = MRI.use_end(); 394 decltype(End) NextI; 395 for (auto I = Begin; I != End; I = NextI) { 396 NextI = std::next(I); 397 if (I->getSubReg() != OldSR) 398 continue; 399 I->setReg(NewR); 400 I->setSubReg(NewSR); 401 } 402 return Begin != End; 403 } 404 405 // For a register ref (pair Reg:Sub), set Begin to the position of the LSB 406 // of Sub in Reg, and set Width to the size of Sub in bits. Return true, 407 // if this succeeded, otherwise return false. 408 bool HexagonBitSimplify::getSubregMask(const BitTracker::RegisterRef &RR, 409 unsigned &Begin, unsigned &Width, MachineRegisterInfo &MRI) { 410 const TargetRegisterClass *RC = MRI.getRegClass(RR.Reg); 411 if (RR.Sub == 0) { 412 Begin = 0; 413 Width = MRI.getTargetRegisterInfo()->getRegSizeInBits(*RC); 414 return true; 415 } 416 417 Begin = 0; 418 419 switch (RC->getID()) { 420 case Hexagon::DoubleRegsRegClassID: 421 case Hexagon::HvxWRRegClassID: 422 Width = MRI.getTargetRegisterInfo()->getRegSizeInBits(*RC) / 2; 423 if (RR.Sub == Hexagon::isub_hi || RR.Sub == Hexagon::vsub_hi) 424 Begin = Width; 425 break; 426 default: 427 return false; 428 } 429 return true; 430 } 431 432 433 // For a REG_SEQUENCE, set SL to the low subregister and SH to the high 434 // subregister. 435 bool HexagonBitSimplify::parseRegSequence(const MachineInstr &I, 436 BitTracker::RegisterRef &SL, BitTracker::RegisterRef &SH, 437 const MachineRegisterInfo &MRI) { 438 assert(I.getOpcode() == TargetOpcode::REG_SEQUENCE); 439 unsigned Sub1 = I.getOperand(2).getImm(), Sub2 = I.getOperand(4).getImm(); 440 auto &DstRC = *MRI.getRegClass(I.getOperand(0).getReg()); 441 auto &HRI = static_cast<const HexagonRegisterInfo&>( 442 *MRI.getTargetRegisterInfo()); 443 unsigned SubLo = HRI.getHexagonSubRegIndex(DstRC, Hexagon::ps_sub_lo); 444 unsigned SubHi = HRI.getHexagonSubRegIndex(DstRC, Hexagon::ps_sub_hi); 445 assert((Sub1 == SubLo && Sub2 == SubHi) || (Sub1 == SubHi && Sub2 == SubLo)); 446 if (Sub1 == SubLo && Sub2 == SubHi) { 447 SL = I.getOperand(1); 448 SH = I.getOperand(3); 449 return true; 450 } 451 if (Sub1 == SubHi && Sub2 == SubLo) { 452 SH = I.getOperand(1); 453 SL = I.getOperand(3); 454 return true; 455 } 456 return false; 457 } 458 459 // All stores (except 64-bit stores) take a 32-bit register as the source 460 // of the value to be stored. If the instruction stores into a location 461 // that is shorter than 32 bits, some bits of the source register are not 462 // used. For each store instruction, calculate the set of used bits in 463 // the source register, and set appropriate bits in Bits. Return true if 464 // the bits are calculated, false otherwise. 465 bool HexagonBitSimplify::getUsedBitsInStore(unsigned Opc, BitVector &Bits, 466 uint16_t Begin) { 467 using namespace Hexagon; 468 469 switch (Opc) { 470 // Store byte 471 case S2_storerb_io: // memb(Rs32+#s11:0)=Rt32 472 case S2_storerbnew_io: // memb(Rs32+#s11:0)=Nt8.new 473 case S2_pstorerbt_io: // if (Pv4) memb(Rs32+#u6:0)=Rt32 474 case S2_pstorerbf_io: // if (!Pv4) memb(Rs32+#u6:0)=Rt32 475 case S4_pstorerbtnew_io: // if (Pv4.new) memb(Rs32+#u6:0)=Rt32 476 case S4_pstorerbfnew_io: // if (!Pv4.new) memb(Rs32+#u6:0)=Rt32 477 case S2_pstorerbnewt_io: // if (Pv4) memb(Rs32+#u6:0)=Nt8.new 478 case S2_pstorerbnewf_io: // if (!Pv4) memb(Rs32+#u6:0)=Nt8.new 479 case S4_pstorerbnewtnew_io: // if (Pv4.new) memb(Rs32+#u6:0)=Nt8.new 480 case S4_pstorerbnewfnew_io: // if (!Pv4.new) memb(Rs32+#u6:0)=Nt8.new 481 case S2_storerb_pi: // memb(Rx32++#s4:0)=Rt32 482 case S2_storerbnew_pi: // memb(Rx32++#s4:0)=Nt8.new 483 case S2_pstorerbt_pi: // if (Pv4) memb(Rx32++#s4:0)=Rt32 484 case S2_pstorerbf_pi: // if (!Pv4) memb(Rx32++#s4:0)=Rt32 485 case S2_pstorerbtnew_pi: // if (Pv4.new) memb(Rx32++#s4:0)=Rt32 486 case S2_pstorerbfnew_pi: // if (!Pv4.new) memb(Rx32++#s4:0)=Rt32 487 case S2_pstorerbnewt_pi: // if (Pv4) memb(Rx32++#s4:0)=Nt8.new 488 case S2_pstorerbnewf_pi: // if (!Pv4) memb(Rx32++#s4:0)=Nt8.new 489 case S2_pstorerbnewtnew_pi: // if (Pv4.new) memb(Rx32++#s4:0)=Nt8.new 490 case S2_pstorerbnewfnew_pi: // if (!Pv4.new) memb(Rx32++#s4:0)=Nt8.new 491 case S4_storerb_ap: // memb(Re32=#U6)=Rt32 492 case S4_storerbnew_ap: // memb(Re32=#U6)=Nt8.new 493 case S2_storerb_pr: // memb(Rx32++Mu2)=Rt32 494 case S2_storerbnew_pr: // memb(Rx32++Mu2)=Nt8.new 495 case S4_storerb_ur: // memb(Ru32<<#u2+#U6)=Rt32 496 case S4_storerbnew_ur: // memb(Ru32<<#u2+#U6)=Nt8.new 497 case S2_storerb_pbr: // memb(Rx32++Mu2:brev)=Rt32 498 case S2_storerbnew_pbr: // memb(Rx32++Mu2:brev)=Nt8.new 499 case S2_storerb_pci: // memb(Rx32++#s4:0:circ(Mu2))=Rt32 500 case S2_storerbnew_pci: // memb(Rx32++#s4:0:circ(Mu2))=Nt8.new 501 case S2_storerb_pcr: // memb(Rx32++I:circ(Mu2))=Rt32 502 case S2_storerbnew_pcr: // memb(Rx32++I:circ(Mu2))=Nt8.new 503 case S4_storerb_rr: // memb(Rs32+Ru32<<#u2)=Rt32 504 case S4_storerbnew_rr: // memb(Rs32+Ru32<<#u2)=Nt8.new 505 case S4_pstorerbt_rr: // if (Pv4) memb(Rs32+Ru32<<#u2)=Rt32 506 case S4_pstorerbf_rr: // if (!Pv4) memb(Rs32+Ru32<<#u2)=Rt32 507 case S4_pstorerbtnew_rr: // if (Pv4.new) memb(Rs32+Ru32<<#u2)=Rt32 508 case S4_pstorerbfnew_rr: // if (!Pv4.new) memb(Rs32+Ru32<<#u2)=Rt32 509 case S4_pstorerbnewt_rr: // if (Pv4) memb(Rs32+Ru32<<#u2)=Nt8.new 510 case S4_pstorerbnewf_rr: // if (!Pv4) memb(Rs32+Ru32<<#u2)=Nt8.new 511 case S4_pstorerbnewtnew_rr: // if (Pv4.new) memb(Rs32+Ru32<<#u2)=Nt8.new 512 case S4_pstorerbnewfnew_rr: // if (!Pv4.new) memb(Rs32+Ru32<<#u2)=Nt8.new 513 case S2_storerbgp: // memb(gp+#u16:0)=Rt32 514 case S2_storerbnewgp: // memb(gp+#u16:0)=Nt8.new 515 case S4_pstorerbt_abs: // if (Pv4) memb(#u6)=Rt32 516 case S4_pstorerbf_abs: // if (!Pv4) memb(#u6)=Rt32 517 case S4_pstorerbtnew_abs: // if (Pv4.new) memb(#u6)=Rt32 518 case S4_pstorerbfnew_abs: // if (!Pv4.new) memb(#u6)=Rt32 519 case S4_pstorerbnewt_abs: // if (Pv4) memb(#u6)=Nt8.new 520 case S4_pstorerbnewf_abs: // if (!Pv4) memb(#u6)=Nt8.new 521 case S4_pstorerbnewtnew_abs: // if (Pv4.new) memb(#u6)=Nt8.new 522 case S4_pstorerbnewfnew_abs: // if (!Pv4.new) memb(#u6)=Nt8.new 523 Bits.set(Begin, Begin+8); 524 return true; 525 526 // Store low half 527 case S2_storerh_io: // memh(Rs32+#s11:1)=Rt32 528 case S2_storerhnew_io: // memh(Rs32+#s11:1)=Nt8.new 529 case S2_pstorerht_io: // if (Pv4) memh(Rs32+#u6:1)=Rt32 530 case S2_pstorerhf_io: // if (!Pv4) memh(Rs32+#u6:1)=Rt32 531 case S4_pstorerhtnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Rt32 532 case S4_pstorerhfnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Rt32 533 case S2_pstorerhnewt_io: // if (Pv4) memh(Rs32+#u6:1)=Nt8.new 534 case S2_pstorerhnewf_io: // if (!Pv4) memh(Rs32+#u6:1)=Nt8.new 535 case S4_pstorerhnewtnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Nt8.new 536 case S4_pstorerhnewfnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Nt8.new 537 case S2_storerh_pi: // memh(Rx32++#s4:1)=Rt32 538 case S2_storerhnew_pi: // memh(Rx32++#s4:1)=Nt8.new 539 case S2_pstorerht_pi: // if (Pv4) memh(Rx32++#s4:1)=Rt32 540 case S2_pstorerhf_pi: // if (!Pv4) memh(Rx32++#s4:1)=Rt32 541 case S2_pstorerhtnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Rt32 542 case S2_pstorerhfnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Rt32 543 case S2_pstorerhnewt_pi: // if (Pv4) memh(Rx32++#s4:1)=Nt8.new 544 case S2_pstorerhnewf_pi: // if (!Pv4) memh(Rx32++#s4:1)=Nt8.new 545 case S2_pstorerhnewtnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Nt8.new 546 case S2_pstorerhnewfnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Nt8.new 547 case S4_storerh_ap: // memh(Re32=#U6)=Rt32 548 case S4_storerhnew_ap: // memh(Re32=#U6)=Nt8.new 549 case S2_storerh_pr: // memh(Rx32++Mu2)=Rt32 550 case S2_storerhnew_pr: // memh(Rx32++Mu2)=Nt8.new 551 case S4_storerh_ur: // memh(Ru32<<#u2+#U6)=Rt32 552 case S4_storerhnew_ur: // memh(Ru32<<#u2+#U6)=Nt8.new 553 case S2_storerh_pbr: // memh(Rx32++Mu2:brev)=Rt32 554 case S2_storerhnew_pbr: // memh(Rx32++Mu2:brev)=Nt8.new 555 case S2_storerh_pci: // memh(Rx32++#s4:1:circ(Mu2))=Rt32 556 case S2_storerhnew_pci: // memh(Rx32++#s4:1:circ(Mu2))=Nt8.new 557 case S2_storerh_pcr: // memh(Rx32++I:circ(Mu2))=Rt32 558 case S2_storerhnew_pcr: // memh(Rx32++I:circ(Mu2))=Nt8.new 559 case S4_storerh_rr: // memh(Rs32+Ru32<<#u2)=Rt32 560 case S4_pstorerht_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Rt32 561 case S4_pstorerhf_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Rt32 562 case S4_pstorerhtnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Rt32 563 case S4_pstorerhfnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Rt32 564 case S4_storerhnew_rr: // memh(Rs32+Ru32<<#u2)=Nt8.new 565 case S4_pstorerhnewt_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Nt8.new 566 case S4_pstorerhnewf_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Nt8.new 567 case S4_pstorerhnewtnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Nt8.new 568 case S4_pstorerhnewfnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Nt8.new 569 case S2_storerhgp: // memh(gp+#u16:1)=Rt32 570 case S2_storerhnewgp: // memh(gp+#u16:1)=Nt8.new 571 case S4_pstorerht_abs: // if (Pv4) memh(#u6)=Rt32 572 case S4_pstorerhf_abs: // if (!Pv4) memh(#u6)=Rt32 573 case S4_pstorerhtnew_abs: // if (Pv4.new) memh(#u6)=Rt32 574 case S4_pstorerhfnew_abs: // if (!Pv4.new) memh(#u6)=Rt32 575 case S4_pstorerhnewt_abs: // if (Pv4) memh(#u6)=Nt8.new 576 case S4_pstorerhnewf_abs: // if (!Pv4) memh(#u6)=Nt8.new 577 case S4_pstorerhnewtnew_abs: // if (Pv4.new) memh(#u6)=Nt8.new 578 case S4_pstorerhnewfnew_abs: // if (!Pv4.new) memh(#u6)=Nt8.new 579 Bits.set(Begin, Begin+16); 580 return true; 581 582 // Store high half 583 case S2_storerf_io: // memh(Rs32+#s11:1)=Rt.H32 584 case S2_pstorerft_io: // if (Pv4) memh(Rs32+#u6:1)=Rt.H32 585 case S2_pstorerff_io: // if (!Pv4) memh(Rs32+#u6:1)=Rt.H32 586 case S4_pstorerftnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Rt.H32 587 case S4_pstorerffnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Rt.H32 588 case S2_storerf_pi: // memh(Rx32++#s4:1)=Rt.H32 589 case S2_pstorerft_pi: // if (Pv4) memh(Rx32++#s4:1)=Rt.H32 590 case S2_pstorerff_pi: // if (!Pv4) memh(Rx32++#s4:1)=Rt.H32 591 case S2_pstorerftnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Rt.H32 592 case S2_pstorerffnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Rt.H32 593 case S4_storerf_ap: // memh(Re32=#U6)=Rt.H32 594 case S2_storerf_pr: // memh(Rx32++Mu2)=Rt.H32 595 case S4_storerf_ur: // memh(Ru32<<#u2+#U6)=Rt.H32 596 case S2_storerf_pbr: // memh(Rx32++Mu2:brev)=Rt.H32 597 case S2_storerf_pci: // memh(Rx32++#s4:1:circ(Mu2))=Rt.H32 598 case S2_storerf_pcr: // memh(Rx32++I:circ(Mu2))=Rt.H32 599 case S4_storerf_rr: // memh(Rs32+Ru32<<#u2)=Rt.H32 600 case S4_pstorerft_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Rt.H32 601 case S4_pstorerff_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Rt.H32 602 case S4_pstorerftnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Rt.H32 603 case S4_pstorerffnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Rt.H32 604 case S2_storerfgp: // memh(gp+#u16:1)=Rt.H32 605 case S4_pstorerft_abs: // if (Pv4) memh(#u6)=Rt.H32 606 case S4_pstorerff_abs: // if (!Pv4) memh(#u6)=Rt.H32 607 case S4_pstorerftnew_abs: // if (Pv4.new) memh(#u6)=Rt.H32 608 case S4_pstorerffnew_abs: // if (!Pv4.new) memh(#u6)=Rt.H32 609 Bits.set(Begin+16, Begin+32); 610 return true; 611 } 612 613 return false; 614 } 615 616 // For an instruction with opcode Opc, calculate the set of bits that it 617 // uses in a register in operand OpN. This only calculates the set of used 618 // bits for cases where it does not depend on any operands (as is the case 619 // in shifts, for example). For concrete instructions from a program, the 620 // operand may be a subregister of a larger register, while Bits would 621 // correspond to the larger register in its entirety. Because of that, 622 // the parameter Begin can be used to indicate which bit of Bits should be 623 // considered the LSB of the operand. 624 bool HexagonBitSimplify::getUsedBits(unsigned Opc, unsigned OpN, 625 BitVector &Bits, uint16_t Begin, const HexagonInstrInfo &HII) { 626 using namespace Hexagon; 627 628 const MCInstrDesc &D = HII.get(Opc); 629 if (D.mayStore()) { 630 if (OpN == D.getNumOperands()-1) 631 return getUsedBitsInStore(Opc, Bits, Begin); 632 return false; 633 } 634 635 switch (Opc) { 636 // One register source. Used bits: R1[0-7]. 637 case A2_sxtb: 638 case A2_zxtb: 639 case A4_cmpbeqi: 640 case A4_cmpbgti: 641 case A4_cmpbgtui: 642 if (OpN == 1) { 643 Bits.set(Begin, Begin+8); 644 return true; 645 } 646 break; 647 648 // One register source. Used bits: R1[0-15]. 649 case A2_aslh: 650 case A2_sxth: 651 case A2_zxth: 652 case A4_cmpheqi: 653 case A4_cmphgti: 654 case A4_cmphgtui: 655 if (OpN == 1) { 656 Bits.set(Begin, Begin+16); 657 return true; 658 } 659 break; 660 661 // One register source. Used bits: R1[16-31]. 662 case A2_asrh: 663 if (OpN == 1) { 664 Bits.set(Begin+16, Begin+32); 665 return true; 666 } 667 break; 668 669 // Two register sources. Used bits: R1[0-7], R2[0-7]. 670 case A4_cmpbeq: 671 case A4_cmpbgt: 672 case A4_cmpbgtu: 673 if (OpN == 1) { 674 Bits.set(Begin, Begin+8); 675 return true; 676 } 677 break; 678 679 // Two register sources. Used bits: R1[0-15], R2[0-15]. 680 case A4_cmpheq: 681 case A4_cmphgt: 682 case A4_cmphgtu: 683 case A2_addh_h16_ll: 684 case A2_addh_h16_sat_ll: 685 case A2_addh_l16_ll: 686 case A2_addh_l16_sat_ll: 687 case A2_combine_ll: 688 case A2_subh_h16_ll: 689 case A2_subh_h16_sat_ll: 690 case A2_subh_l16_ll: 691 case A2_subh_l16_sat_ll: 692 case M2_mpy_acc_ll_s0: 693 case M2_mpy_acc_ll_s1: 694 case M2_mpy_acc_sat_ll_s0: 695 case M2_mpy_acc_sat_ll_s1: 696 case M2_mpy_ll_s0: 697 case M2_mpy_ll_s1: 698 case M2_mpy_nac_ll_s0: 699 case M2_mpy_nac_ll_s1: 700 case M2_mpy_nac_sat_ll_s0: 701 case M2_mpy_nac_sat_ll_s1: 702 case M2_mpy_rnd_ll_s0: 703 case M2_mpy_rnd_ll_s1: 704 case M2_mpy_sat_ll_s0: 705 case M2_mpy_sat_ll_s1: 706 case M2_mpy_sat_rnd_ll_s0: 707 case M2_mpy_sat_rnd_ll_s1: 708 case M2_mpyd_acc_ll_s0: 709 case M2_mpyd_acc_ll_s1: 710 case M2_mpyd_ll_s0: 711 case M2_mpyd_ll_s1: 712 case M2_mpyd_nac_ll_s0: 713 case M2_mpyd_nac_ll_s1: 714 case M2_mpyd_rnd_ll_s0: 715 case M2_mpyd_rnd_ll_s1: 716 case M2_mpyu_acc_ll_s0: 717 case M2_mpyu_acc_ll_s1: 718 case M2_mpyu_ll_s0: 719 case M2_mpyu_ll_s1: 720 case M2_mpyu_nac_ll_s0: 721 case M2_mpyu_nac_ll_s1: 722 case M2_mpyud_acc_ll_s0: 723 case M2_mpyud_acc_ll_s1: 724 case M2_mpyud_ll_s0: 725 case M2_mpyud_ll_s1: 726 case M2_mpyud_nac_ll_s0: 727 case M2_mpyud_nac_ll_s1: 728 if (OpN == 1 || OpN == 2) { 729 Bits.set(Begin, Begin+16); 730 return true; 731 } 732 break; 733 734 // Two register sources. Used bits: R1[0-15], R2[16-31]. 735 case A2_addh_h16_lh: 736 case A2_addh_h16_sat_lh: 737 case A2_combine_lh: 738 case A2_subh_h16_lh: 739 case A2_subh_h16_sat_lh: 740 case M2_mpy_acc_lh_s0: 741 case M2_mpy_acc_lh_s1: 742 case M2_mpy_acc_sat_lh_s0: 743 case M2_mpy_acc_sat_lh_s1: 744 case M2_mpy_lh_s0: 745 case M2_mpy_lh_s1: 746 case M2_mpy_nac_lh_s0: 747 case M2_mpy_nac_lh_s1: 748 case M2_mpy_nac_sat_lh_s0: 749 case M2_mpy_nac_sat_lh_s1: 750 case M2_mpy_rnd_lh_s0: 751 case M2_mpy_rnd_lh_s1: 752 case M2_mpy_sat_lh_s0: 753 case M2_mpy_sat_lh_s1: 754 case M2_mpy_sat_rnd_lh_s0: 755 case M2_mpy_sat_rnd_lh_s1: 756 case M2_mpyd_acc_lh_s0: 757 case M2_mpyd_acc_lh_s1: 758 case M2_mpyd_lh_s0: 759 case M2_mpyd_lh_s1: 760 case M2_mpyd_nac_lh_s0: 761 case M2_mpyd_nac_lh_s1: 762 case M2_mpyd_rnd_lh_s0: 763 case M2_mpyd_rnd_lh_s1: 764 case M2_mpyu_acc_lh_s0: 765 case M2_mpyu_acc_lh_s1: 766 case M2_mpyu_lh_s0: 767 case M2_mpyu_lh_s1: 768 case M2_mpyu_nac_lh_s0: 769 case M2_mpyu_nac_lh_s1: 770 case M2_mpyud_acc_lh_s0: 771 case M2_mpyud_acc_lh_s1: 772 case M2_mpyud_lh_s0: 773 case M2_mpyud_lh_s1: 774 case M2_mpyud_nac_lh_s0: 775 case M2_mpyud_nac_lh_s1: 776 // These four are actually LH. 777 case A2_addh_l16_hl: 778 case A2_addh_l16_sat_hl: 779 case A2_subh_l16_hl: 780 case A2_subh_l16_sat_hl: 781 if (OpN == 1) { 782 Bits.set(Begin, Begin+16); 783 return true; 784 } 785 if (OpN == 2) { 786 Bits.set(Begin+16, Begin+32); 787 return true; 788 } 789 break; 790 791 // Two register sources, used bits: R1[16-31], R2[0-15]. 792 case A2_addh_h16_hl: 793 case A2_addh_h16_sat_hl: 794 case A2_combine_hl: 795 case A2_subh_h16_hl: 796 case A2_subh_h16_sat_hl: 797 case M2_mpy_acc_hl_s0: 798 case M2_mpy_acc_hl_s1: 799 case M2_mpy_acc_sat_hl_s0: 800 case M2_mpy_acc_sat_hl_s1: 801 case M2_mpy_hl_s0: 802 case M2_mpy_hl_s1: 803 case M2_mpy_nac_hl_s0: 804 case M2_mpy_nac_hl_s1: 805 case M2_mpy_nac_sat_hl_s0: 806 case M2_mpy_nac_sat_hl_s1: 807 case M2_mpy_rnd_hl_s0: 808 case M2_mpy_rnd_hl_s1: 809 case M2_mpy_sat_hl_s0: 810 case M2_mpy_sat_hl_s1: 811 case M2_mpy_sat_rnd_hl_s0: 812 case M2_mpy_sat_rnd_hl_s1: 813 case M2_mpyd_acc_hl_s0: 814 case M2_mpyd_acc_hl_s1: 815 case M2_mpyd_hl_s0: 816 case M2_mpyd_hl_s1: 817 case M2_mpyd_nac_hl_s0: 818 case M2_mpyd_nac_hl_s1: 819 case M2_mpyd_rnd_hl_s0: 820 case M2_mpyd_rnd_hl_s1: 821 case M2_mpyu_acc_hl_s0: 822 case M2_mpyu_acc_hl_s1: 823 case M2_mpyu_hl_s0: 824 case M2_mpyu_hl_s1: 825 case M2_mpyu_nac_hl_s0: 826 case M2_mpyu_nac_hl_s1: 827 case M2_mpyud_acc_hl_s0: 828 case M2_mpyud_acc_hl_s1: 829 case M2_mpyud_hl_s0: 830 case M2_mpyud_hl_s1: 831 case M2_mpyud_nac_hl_s0: 832 case M2_mpyud_nac_hl_s1: 833 if (OpN == 1) { 834 Bits.set(Begin+16, Begin+32); 835 return true; 836 } 837 if (OpN == 2) { 838 Bits.set(Begin, Begin+16); 839 return true; 840 } 841 break; 842 843 // Two register sources, used bits: R1[16-31], R2[16-31]. 844 case A2_addh_h16_hh: 845 case A2_addh_h16_sat_hh: 846 case A2_combine_hh: 847 case A2_subh_h16_hh: 848 case A2_subh_h16_sat_hh: 849 case M2_mpy_acc_hh_s0: 850 case M2_mpy_acc_hh_s1: 851 case M2_mpy_acc_sat_hh_s0: 852 case M2_mpy_acc_sat_hh_s1: 853 case M2_mpy_hh_s0: 854 case M2_mpy_hh_s1: 855 case M2_mpy_nac_hh_s0: 856 case M2_mpy_nac_hh_s1: 857 case M2_mpy_nac_sat_hh_s0: 858 case M2_mpy_nac_sat_hh_s1: 859 case M2_mpy_rnd_hh_s0: 860 case M2_mpy_rnd_hh_s1: 861 case M2_mpy_sat_hh_s0: 862 case M2_mpy_sat_hh_s1: 863 case M2_mpy_sat_rnd_hh_s0: 864 case M2_mpy_sat_rnd_hh_s1: 865 case M2_mpyd_acc_hh_s0: 866 case M2_mpyd_acc_hh_s1: 867 case M2_mpyd_hh_s0: 868 case M2_mpyd_hh_s1: 869 case M2_mpyd_nac_hh_s0: 870 case M2_mpyd_nac_hh_s1: 871 case M2_mpyd_rnd_hh_s0: 872 case M2_mpyd_rnd_hh_s1: 873 case M2_mpyu_acc_hh_s0: 874 case M2_mpyu_acc_hh_s1: 875 case M2_mpyu_hh_s0: 876 case M2_mpyu_hh_s1: 877 case M2_mpyu_nac_hh_s0: 878 case M2_mpyu_nac_hh_s1: 879 case M2_mpyud_acc_hh_s0: 880 case M2_mpyud_acc_hh_s1: 881 case M2_mpyud_hh_s0: 882 case M2_mpyud_hh_s1: 883 case M2_mpyud_nac_hh_s0: 884 case M2_mpyud_nac_hh_s1: 885 if (OpN == 1 || OpN == 2) { 886 Bits.set(Begin+16, Begin+32); 887 return true; 888 } 889 break; 890 } 891 892 return false; 893 } 894 895 // Calculate the register class that matches Reg:Sub. For example, if 896 // %1 is a double register, then %1:isub_hi would match the "int" 897 // register class. 898 const TargetRegisterClass *HexagonBitSimplify::getFinalVRegClass( 899 const BitTracker::RegisterRef &RR, MachineRegisterInfo &MRI) { 900 if (!TargetRegisterInfo::isVirtualRegister(RR.Reg)) 901 return nullptr; 902 auto *RC = MRI.getRegClass(RR.Reg); 903 if (RR.Sub == 0) 904 return RC; 905 auto &HRI = static_cast<const HexagonRegisterInfo&>( 906 *MRI.getTargetRegisterInfo()); 907 908 auto VerifySR = [&HRI] (const TargetRegisterClass *RC, unsigned Sub) -> void { 909 (void)HRI; 910 assert(Sub == HRI.getHexagonSubRegIndex(*RC, Hexagon::ps_sub_lo) || 911 Sub == HRI.getHexagonSubRegIndex(*RC, Hexagon::ps_sub_hi)); 912 }; 913 914 switch (RC->getID()) { 915 case Hexagon::DoubleRegsRegClassID: 916 VerifySR(RC, RR.Sub); 917 return &Hexagon::IntRegsRegClass; 918 case Hexagon::HvxWRRegClassID: 919 VerifySR(RC, RR.Sub); 920 return &Hexagon::HvxVRRegClass; 921 } 922 return nullptr; 923 } 924 925 // Check if RD could be replaced with RS at any possible use of RD. 926 // For example a predicate register cannot be replaced with a integer 927 // register, but a 64-bit register with a subregister can be replaced 928 // with a 32-bit register. 929 bool HexagonBitSimplify::isTransparentCopy(const BitTracker::RegisterRef &RD, 930 const BitTracker::RegisterRef &RS, MachineRegisterInfo &MRI) { 931 if (!TargetRegisterInfo::isVirtualRegister(RD.Reg) || 932 !TargetRegisterInfo::isVirtualRegister(RS.Reg)) 933 return false; 934 // Return false if one (or both) classes are nullptr. 935 auto *DRC = getFinalVRegClass(RD, MRI); 936 if (!DRC) 937 return false; 938 939 return DRC == getFinalVRegClass(RS, MRI); 940 } 941 942 bool HexagonBitSimplify::hasTiedUse(unsigned Reg, MachineRegisterInfo &MRI, 943 unsigned NewSub) { 944 if (!PreserveTiedOps) 945 return false; 946 return llvm::any_of(MRI.use_operands(Reg), 947 [NewSub] (const MachineOperand &Op) -> bool { 948 return Op.getSubReg() != NewSub && Op.isTied(); 949 }); 950 } 951 952 namespace { 953 954 class DeadCodeElimination { 955 public: 956 DeadCodeElimination(MachineFunction &mf, MachineDominatorTree &mdt) 957 : MF(mf), HII(*MF.getSubtarget<HexagonSubtarget>().getInstrInfo()), 958 MDT(mdt), MRI(mf.getRegInfo()) {} 959 960 bool run() { 961 return runOnNode(MDT.getRootNode()); 962 } 963 964 private: 965 bool isDead(unsigned R) const; 966 bool runOnNode(MachineDomTreeNode *N); 967 968 MachineFunction &MF; 969 const HexagonInstrInfo &HII; 970 MachineDominatorTree &MDT; 971 MachineRegisterInfo &MRI; 972 }; 973 974 } // end anonymous namespace 975 976 bool DeadCodeElimination::isDead(unsigned R) const { 977 for (auto I = MRI.use_begin(R), E = MRI.use_end(); I != E; ++I) { 978 MachineInstr *UseI = I->getParent(); 979 if (UseI->isDebugValue()) 980 continue; 981 if (UseI->isPHI()) { 982 assert(!UseI->getOperand(0).getSubReg()); 983 unsigned DR = UseI->getOperand(0).getReg(); 984 if (DR == R) 985 continue; 986 } 987 return false; 988 } 989 return true; 990 } 991 992 bool DeadCodeElimination::runOnNode(MachineDomTreeNode *N) { 993 bool Changed = false; 994 995 for (auto *DTN : children<MachineDomTreeNode*>(N)) 996 Changed |= runOnNode(DTN); 997 998 MachineBasicBlock *B = N->getBlock(); 999 std::vector<MachineInstr*> Instrs; 1000 for (auto I = B->rbegin(), E = B->rend(); I != E; ++I) 1001 Instrs.push_back(&*I); 1002 1003 for (auto MI : Instrs) { 1004 unsigned Opc = MI->getOpcode(); 1005 // Do not touch lifetime markers. This is why the target-independent DCE 1006 // cannot be used. 1007 if (Opc == TargetOpcode::LIFETIME_START || 1008 Opc == TargetOpcode::LIFETIME_END) 1009 continue; 1010 bool Store = false; 1011 if (MI->isInlineAsm()) 1012 continue; 1013 // Delete PHIs if possible. 1014 if (!MI->isPHI() && !MI->isSafeToMove(nullptr, Store)) 1015 continue; 1016 1017 bool AllDead = true; 1018 SmallVector<unsigned,2> Regs; 1019 for (auto &Op : MI->operands()) { 1020 if (!Op.isReg() || !Op.isDef()) 1021 continue; 1022 unsigned R = Op.getReg(); 1023 if (!TargetRegisterInfo::isVirtualRegister(R) || !isDead(R)) { 1024 AllDead = false; 1025 break; 1026 } 1027 Regs.push_back(R); 1028 } 1029 if (!AllDead) 1030 continue; 1031 1032 B->erase(MI); 1033 for (unsigned i = 0, n = Regs.size(); i != n; ++i) 1034 MRI.markUsesInDebugValueAsUndef(Regs[i]); 1035 Changed = true; 1036 } 1037 1038 return Changed; 1039 } 1040 1041 namespace { 1042 1043 // Eliminate redundant instructions 1044 // 1045 // This transformation will identify instructions where the output register 1046 // is the same as one of its input registers. This only works on instructions 1047 // that define a single register (unlike post-increment loads, for example). 1048 // The equality check is actually more detailed: the code calculates which 1049 // bits of the output are used, and only compares these bits with the input 1050 // registers. 1051 // If the output matches an input, the instruction is replaced with COPY. 1052 // The copies will be removed by another transformation. 1053 class RedundantInstrElimination : public Transformation { 1054 public: 1055 RedundantInstrElimination(BitTracker &bt, const HexagonInstrInfo &hii, 1056 const HexagonRegisterInfo &hri, MachineRegisterInfo &mri) 1057 : Transformation(true), HII(hii), HRI(hri), MRI(mri), BT(bt) {} 1058 1059 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; 1060 1061 private: 1062 bool isLossyShiftLeft(const MachineInstr &MI, unsigned OpN, 1063 unsigned &LostB, unsigned &LostE); 1064 bool isLossyShiftRight(const MachineInstr &MI, unsigned OpN, 1065 unsigned &LostB, unsigned &LostE); 1066 bool computeUsedBits(unsigned Reg, BitVector &Bits); 1067 bool computeUsedBits(const MachineInstr &MI, unsigned OpN, BitVector &Bits, 1068 uint16_t Begin); 1069 bool usedBitsEqual(BitTracker::RegisterRef RD, BitTracker::RegisterRef RS); 1070 1071 const HexagonInstrInfo &HII; 1072 const HexagonRegisterInfo &HRI; 1073 MachineRegisterInfo &MRI; 1074 BitTracker &BT; 1075 }; 1076 1077 } // end anonymous namespace 1078 1079 // Check if the instruction is a lossy shift left, where the input being 1080 // shifted is the operand OpN of MI. If true, [LostB, LostE) is the range 1081 // of bit indices that are lost. 1082 bool RedundantInstrElimination::isLossyShiftLeft(const MachineInstr &MI, 1083 unsigned OpN, unsigned &LostB, unsigned &LostE) { 1084 using namespace Hexagon; 1085 1086 unsigned Opc = MI.getOpcode(); 1087 unsigned ImN, RegN, Width; 1088 switch (Opc) { 1089 case S2_asl_i_p: 1090 ImN = 2; 1091 RegN = 1; 1092 Width = 64; 1093 break; 1094 case S2_asl_i_p_acc: 1095 case S2_asl_i_p_and: 1096 case S2_asl_i_p_nac: 1097 case S2_asl_i_p_or: 1098 case S2_asl_i_p_xacc: 1099 ImN = 3; 1100 RegN = 2; 1101 Width = 64; 1102 break; 1103 case S2_asl_i_r: 1104 ImN = 2; 1105 RegN = 1; 1106 Width = 32; 1107 break; 1108 case S2_addasl_rrri: 1109 case S4_andi_asl_ri: 1110 case S4_ori_asl_ri: 1111 case S4_addi_asl_ri: 1112 case S4_subi_asl_ri: 1113 case S2_asl_i_r_acc: 1114 case S2_asl_i_r_and: 1115 case S2_asl_i_r_nac: 1116 case S2_asl_i_r_or: 1117 case S2_asl_i_r_sat: 1118 case S2_asl_i_r_xacc: 1119 ImN = 3; 1120 RegN = 2; 1121 Width = 32; 1122 break; 1123 default: 1124 return false; 1125 } 1126 1127 if (RegN != OpN) 1128 return false; 1129 1130 assert(MI.getOperand(ImN).isImm()); 1131 unsigned S = MI.getOperand(ImN).getImm(); 1132 if (S == 0) 1133 return false; 1134 LostB = Width-S; 1135 LostE = Width; 1136 return true; 1137 } 1138 1139 // Check if the instruction is a lossy shift right, where the input being 1140 // shifted is the operand OpN of MI. If true, [LostB, LostE) is the range 1141 // of bit indices that are lost. 1142 bool RedundantInstrElimination::isLossyShiftRight(const MachineInstr &MI, 1143 unsigned OpN, unsigned &LostB, unsigned &LostE) { 1144 using namespace Hexagon; 1145 1146 unsigned Opc = MI.getOpcode(); 1147 unsigned ImN, RegN; 1148 switch (Opc) { 1149 case S2_asr_i_p: 1150 case S2_lsr_i_p: 1151 ImN = 2; 1152 RegN = 1; 1153 break; 1154 case S2_asr_i_p_acc: 1155 case S2_asr_i_p_and: 1156 case S2_asr_i_p_nac: 1157 case S2_asr_i_p_or: 1158 case S2_lsr_i_p_acc: 1159 case S2_lsr_i_p_and: 1160 case S2_lsr_i_p_nac: 1161 case S2_lsr_i_p_or: 1162 case S2_lsr_i_p_xacc: 1163 ImN = 3; 1164 RegN = 2; 1165 break; 1166 case S2_asr_i_r: 1167 case S2_lsr_i_r: 1168 ImN = 2; 1169 RegN = 1; 1170 break; 1171 case S4_andi_lsr_ri: 1172 case S4_ori_lsr_ri: 1173 case S4_addi_lsr_ri: 1174 case S4_subi_lsr_ri: 1175 case S2_asr_i_r_acc: 1176 case S2_asr_i_r_and: 1177 case S2_asr_i_r_nac: 1178 case S2_asr_i_r_or: 1179 case S2_lsr_i_r_acc: 1180 case S2_lsr_i_r_and: 1181 case S2_lsr_i_r_nac: 1182 case S2_lsr_i_r_or: 1183 case S2_lsr_i_r_xacc: 1184 ImN = 3; 1185 RegN = 2; 1186 break; 1187 1188 default: 1189 return false; 1190 } 1191 1192 if (RegN != OpN) 1193 return false; 1194 1195 assert(MI.getOperand(ImN).isImm()); 1196 unsigned S = MI.getOperand(ImN).getImm(); 1197 LostB = 0; 1198 LostE = S; 1199 return true; 1200 } 1201 1202 // Calculate the bit vector that corresponds to the used bits of register Reg. 1203 // The vector Bits has the same size, as the size of Reg in bits. If the cal- 1204 // culation fails (i.e. the used bits are unknown), it returns false. Other- 1205 // wise, it returns true and sets the corresponding bits in Bits. 1206 bool RedundantInstrElimination::computeUsedBits(unsigned Reg, BitVector &Bits) { 1207 BitVector Used(Bits.size()); 1208 RegisterSet Visited; 1209 std::vector<unsigned> Pending; 1210 Pending.push_back(Reg); 1211 1212 for (unsigned i = 0; i < Pending.size(); ++i) { 1213 unsigned R = Pending[i]; 1214 if (Visited.has(R)) 1215 continue; 1216 Visited.insert(R); 1217 for (auto I = MRI.use_begin(R), E = MRI.use_end(); I != E; ++I) { 1218 BitTracker::RegisterRef UR = *I; 1219 unsigned B, W; 1220 if (!HBS::getSubregMask(UR, B, W, MRI)) 1221 return false; 1222 MachineInstr &UseI = *I->getParent(); 1223 if (UseI.isPHI() || UseI.isCopy()) { 1224 unsigned DefR = UseI.getOperand(0).getReg(); 1225 if (!TargetRegisterInfo::isVirtualRegister(DefR)) 1226 return false; 1227 Pending.push_back(DefR); 1228 } else { 1229 if (!computeUsedBits(UseI, I.getOperandNo(), Used, B)) 1230 return false; 1231 } 1232 } 1233 } 1234 Bits |= Used; 1235 return true; 1236 } 1237 1238 // Calculate the bits used by instruction MI in a register in operand OpN. 1239 // Return true/false if the calculation succeeds/fails. If is succeeds, set 1240 // used bits in Bits. This function does not reset any bits in Bits, so 1241 // subsequent calls over different instructions will result in the union 1242 // of the used bits in all these instructions. 1243 // The register in question may be used with a sub-register, whereas Bits 1244 // holds the bits for the entire register. To keep track of that, the 1245 // argument Begin indicates where in Bits is the lowest-significant bit 1246 // of the register used in operand OpN. For example, in instruction: 1247 // %1 = S2_lsr_i_r %2:isub_hi, 10 1248 // the operand 1 is a 32-bit register, which happens to be a subregister 1249 // of the 64-bit register %2, and that subregister starts at position 32. 1250 // In this case Begin=32, since Bits[32] would be the lowest-significant bit 1251 // of %2:isub_hi. 1252 bool RedundantInstrElimination::computeUsedBits(const MachineInstr &MI, 1253 unsigned OpN, BitVector &Bits, uint16_t Begin) { 1254 unsigned Opc = MI.getOpcode(); 1255 BitVector T(Bits.size()); 1256 bool GotBits = HBS::getUsedBits(Opc, OpN, T, Begin, HII); 1257 // Even if we don't have bits yet, we could still provide some information 1258 // if the instruction is a lossy shift: the lost bits will be marked as 1259 // not used. 1260 unsigned LB, LE; 1261 if (isLossyShiftLeft(MI, OpN, LB, LE) || isLossyShiftRight(MI, OpN, LB, LE)) { 1262 assert(MI.getOperand(OpN).isReg()); 1263 BitTracker::RegisterRef RR = MI.getOperand(OpN); 1264 const TargetRegisterClass *RC = HBS::getFinalVRegClass(RR, MRI); 1265 uint16_t Width = HRI.getRegSizeInBits(*RC); 1266 1267 if (!GotBits) 1268 T.set(Begin, Begin+Width); 1269 assert(LB <= LE && LB < Width && LE <= Width); 1270 T.reset(Begin+LB, Begin+LE); 1271 GotBits = true; 1272 } 1273 if (GotBits) 1274 Bits |= T; 1275 return GotBits; 1276 } 1277 1278 // Calculates the used bits in RD ("defined register"), and checks if these 1279 // bits in RS ("used register") and RD are identical. 1280 bool RedundantInstrElimination::usedBitsEqual(BitTracker::RegisterRef RD, 1281 BitTracker::RegisterRef RS) { 1282 const BitTracker::RegisterCell &DC = BT.lookup(RD.Reg); 1283 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); 1284 1285 unsigned DB, DW; 1286 if (!HBS::getSubregMask(RD, DB, DW, MRI)) 1287 return false; 1288 unsigned SB, SW; 1289 if (!HBS::getSubregMask(RS, SB, SW, MRI)) 1290 return false; 1291 if (SW != DW) 1292 return false; 1293 1294 BitVector Used(DC.width()); 1295 if (!computeUsedBits(RD.Reg, Used)) 1296 return false; 1297 1298 for (unsigned i = 0; i != DW; ++i) 1299 if (Used[i+DB] && DC[DB+i] != SC[SB+i]) 1300 return false; 1301 return true; 1302 } 1303 1304 bool RedundantInstrElimination::processBlock(MachineBasicBlock &B, 1305 const RegisterSet&) { 1306 if (!BT.reached(&B)) 1307 return false; 1308 bool Changed = false; 1309 1310 for (auto I = B.begin(), E = B.end(), NextI = I; I != E; ++I) { 1311 NextI = std::next(I); 1312 MachineInstr *MI = &*I; 1313 1314 if (MI->getOpcode() == TargetOpcode::COPY) 1315 continue; 1316 if (MI->isPHI() || MI->hasUnmodeledSideEffects() || MI->isInlineAsm()) 1317 continue; 1318 unsigned NumD = MI->getDesc().getNumDefs(); 1319 if (NumD != 1) 1320 continue; 1321 1322 BitTracker::RegisterRef RD = MI->getOperand(0); 1323 if (!BT.has(RD.Reg)) 1324 continue; 1325 const BitTracker::RegisterCell &DC = BT.lookup(RD.Reg); 1326 auto At = MachineBasicBlock::iterator(MI); 1327 1328 // Find a source operand that is equal to the result. 1329 for (auto &Op : MI->uses()) { 1330 if (!Op.isReg()) 1331 continue; 1332 BitTracker::RegisterRef RS = Op; 1333 if (!BT.has(RS.Reg)) 1334 continue; 1335 if (!HBS::isTransparentCopy(RD, RS, MRI)) 1336 continue; 1337 1338 unsigned BN, BW; 1339 if (!HBS::getSubregMask(RS, BN, BW, MRI)) 1340 continue; 1341 1342 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); 1343 if (!usedBitsEqual(RD, RS) && !HBS::isEqual(DC, 0, SC, BN, BW)) 1344 continue; 1345 1346 // If found, replace the instruction with a COPY. 1347 const DebugLoc &DL = MI->getDebugLoc(); 1348 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); 1349 unsigned NewR = MRI.createVirtualRegister(FRC); 1350 MachineInstr *CopyI = 1351 BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR) 1352 .addReg(RS.Reg, 0, RS.Sub); 1353 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); 1354 // This pass can create copies between registers that don't have the 1355 // exact same values. Updating the tracker has to involve updating 1356 // all dependent cells. Example: 1357 // %1 = inst %2 ; %1 != %2, but used bits are equal 1358 // 1359 // %3 = copy %2 ; <- inserted 1360 // ... = %3 ; <- replaced from %2 1361 // Indirectly, we can create a "copy" between %1 and %2 even 1362 // though their exact values do not match. 1363 BT.visit(*CopyI); 1364 Changed = true; 1365 break; 1366 } 1367 } 1368 1369 return Changed; 1370 } 1371 1372 namespace { 1373 1374 // Recognize instructions that produce constant values known at compile-time. 1375 // Replace them with register definitions that load these constants directly. 1376 class ConstGeneration : public Transformation { 1377 public: 1378 ConstGeneration(BitTracker &bt, const HexagonInstrInfo &hii, 1379 MachineRegisterInfo &mri) 1380 : Transformation(true), HII(hii), MRI(mri), BT(bt) {} 1381 1382 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; 1383 static bool isTfrConst(const MachineInstr &MI); 1384 1385 private: 1386 unsigned genTfrConst(const TargetRegisterClass *RC, int64_t C, 1387 MachineBasicBlock &B, MachineBasicBlock::iterator At, DebugLoc &DL); 1388 1389 const HexagonInstrInfo &HII; 1390 MachineRegisterInfo &MRI; 1391 BitTracker &BT; 1392 }; 1393 1394 } // end anonymous namespace 1395 1396 bool ConstGeneration::isTfrConst(const MachineInstr &MI) { 1397 unsigned Opc = MI.getOpcode(); 1398 switch (Opc) { 1399 case Hexagon::A2_combineii: 1400 case Hexagon::A4_combineii: 1401 case Hexagon::A2_tfrsi: 1402 case Hexagon::A2_tfrpi: 1403 case Hexagon::PS_true: 1404 case Hexagon::PS_false: 1405 case Hexagon::CONST32: 1406 case Hexagon::CONST64: 1407 return true; 1408 } 1409 return false; 1410 } 1411 1412 // Generate a transfer-immediate instruction that is appropriate for the 1413 // register class and the actual value being transferred. 1414 unsigned ConstGeneration::genTfrConst(const TargetRegisterClass *RC, int64_t C, 1415 MachineBasicBlock &B, MachineBasicBlock::iterator At, DebugLoc &DL) { 1416 unsigned Reg = MRI.createVirtualRegister(RC); 1417 if (RC == &Hexagon::IntRegsRegClass) { 1418 BuildMI(B, At, DL, HII.get(Hexagon::A2_tfrsi), Reg) 1419 .addImm(int32_t(C)); 1420 return Reg; 1421 } 1422 1423 if (RC == &Hexagon::DoubleRegsRegClass) { 1424 if (isInt<8>(C)) { 1425 BuildMI(B, At, DL, HII.get(Hexagon::A2_tfrpi), Reg) 1426 .addImm(C); 1427 return Reg; 1428 } 1429 1430 unsigned Lo = Lo_32(C), Hi = Hi_32(C); 1431 if (isInt<8>(Lo) || isInt<8>(Hi)) { 1432 unsigned Opc = isInt<8>(Lo) ? Hexagon::A2_combineii 1433 : Hexagon::A4_combineii; 1434 BuildMI(B, At, DL, HII.get(Opc), Reg) 1435 .addImm(int32_t(Hi)) 1436 .addImm(int32_t(Lo)); 1437 return Reg; 1438 } 1439 1440 BuildMI(B, At, DL, HII.get(Hexagon::CONST64), Reg) 1441 .addImm(C); 1442 return Reg; 1443 } 1444 1445 if (RC == &Hexagon::PredRegsRegClass) { 1446 unsigned Opc; 1447 if (C == 0) 1448 Opc = Hexagon::PS_false; 1449 else if ((C & 0xFF) == 0xFF) 1450 Opc = Hexagon::PS_true; 1451 else 1452 return 0; 1453 BuildMI(B, At, DL, HII.get(Opc), Reg); 1454 return Reg; 1455 } 1456 1457 return 0; 1458 } 1459 1460 bool ConstGeneration::processBlock(MachineBasicBlock &B, const RegisterSet&) { 1461 if (!BT.reached(&B)) 1462 return false; 1463 bool Changed = false; 1464 RegisterSet Defs; 1465 1466 for (auto I = B.begin(), E = B.end(); I != E; ++I) { 1467 if (isTfrConst(*I)) 1468 continue; 1469 Defs.clear(); 1470 HBS::getInstrDefs(*I, Defs); 1471 if (Defs.count() != 1) 1472 continue; 1473 unsigned DR = Defs.find_first(); 1474 if (!TargetRegisterInfo::isVirtualRegister(DR)) 1475 continue; 1476 uint64_t U; 1477 const BitTracker::RegisterCell &DRC = BT.lookup(DR); 1478 if (HBS::getConst(DRC, 0, DRC.width(), U)) { 1479 int64_t C = U; 1480 DebugLoc DL = I->getDebugLoc(); 1481 auto At = I->isPHI() ? B.getFirstNonPHI() : I; 1482 unsigned ImmReg = genTfrConst(MRI.getRegClass(DR), C, B, At, DL); 1483 if (ImmReg) { 1484 HBS::replaceReg(DR, ImmReg, MRI); 1485 BT.put(ImmReg, DRC); 1486 Changed = true; 1487 } 1488 } 1489 } 1490 return Changed; 1491 } 1492 1493 namespace { 1494 1495 // Identify pairs of available registers which hold identical values. 1496 // In such cases, only one of them needs to be calculated, the other one 1497 // will be defined as a copy of the first. 1498 class CopyGeneration : public Transformation { 1499 public: 1500 CopyGeneration(BitTracker &bt, const HexagonInstrInfo &hii, 1501 const HexagonRegisterInfo &hri, MachineRegisterInfo &mri) 1502 : Transformation(true), HII(hii), HRI(hri), MRI(mri), BT(bt) {} 1503 1504 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; 1505 1506 private: 1507 bool findMatch(const BitTracker::RegisterRef &Inp, 1508 BitTracker::RegisterRef &Out, const RegisterSet &AVs); 1509 1510 const HexagonInstrInfo &HII; 1511 const HexagonRegisterInfo &HRI; 1512 MachineRegisterInfo &MRI; 1513 BitTracker &BT; 1514 RegisterSet Forbidden; 1515 }; 1516 1517 // Eliminate register copies RD = RS, by replacing the uses of RD with 1518 // with uses of RS. 1519 class CopyPropagation : public Transformation { 1520 public: 1521 CopyPropagation(const HexagonRegisterInfo &hri, MachineRegisterInfo &mri) 1522 : Transformation(false), HRI(hri), MRI(mri) {} 1523 1524 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; 1525 1526 static bool isCopyReg(unsigned Opc, bool NoConv); 1527 1528 private: 1529 bool propagateRegCopy(MachineInstr &MI); 1530 1531 const HexagonRegisterInfo &HRI; 1532 MachineRegisterInfo &MRI; 1533 }; 1534 1535 } // end anonymous namespace 1536 1537 /// Check if there is a register in AVs that is identical to Inp. If so, 1538 /// set Out to the found register. The output may be a pair Reg:Sub. 1539 bool CopyGeneration::findMatch(const BitTracker::RegisterRef &Inp, 1540 BitTracker::RegisterRef &Out, const RegisterSet &AVs) { 1541 if (!BT.has(Inp.Reg)) 1542 return false; 1543 const BitTracker::RegisterCell &InpRC = BT.lookup(Inp.Reg); 1544 auto *FRC = HBS::getFinalVRegClass(Inp, MRI); 1545 unsigned B, W; 1546 if (!HBS::getSubregMask(Inp, B, W, MRI)) 1547 return false; 1548 1549 for (unsigned R = AVs.find_first(); R; R = AVs.find_next(R)) { 1550 if (!BT.has(R) || Forbidden[R]) 1551 continue; 1552 const BitTracker::RegisterCell &RC = BT.lookup(R); 1553 unsigned RW = RC.width(); 1554 if (W == RW) { 1555 if (FRC != MRI.getRegClass(R)) 1556 continue; 1557 if (!HBS::isTransparentCopy(R, Inp, MRI)) 1558 continue; 1559 if (!HBS::isEqual(InpRC, B, RC, 0, W)) 1560 continue; 1561 Out.Reg = R; 1562 Out.Sub = 0; 1563 return true; 1564 } 1565 // Check if there is a super-register, whose part (with a subregister) 1566 // is equal to the input. 1567 // Only do double registers for now. 1568 if (W*2 != RW) 1569 continue; 1570 if (MRI.getRegClass(R) != &Hexagon::DoubleRegsRegClass) 1571 continue; 1572 1573 if (HBS::isEqual(InpRC, B, RC, 0, W)) 1574 Out.Sub = Hexagon::isub_lo; 1575 else if (HBS::isEqual(InpRC, B, RC, W, W)) 1576 Out.Sub = Hexagon::isub_hi; 1577 else 1578 continue; 1579 Out.Reg = R; 1580 if (HBS::isTransparentCopy(Out, Inp, MRI)) 1581 return true; 1582 } 1583 return false; 1584 } 1585 1586 bool CopyGeneration::processBlock(MachineBasicBlock &B, 1587 const RegisterSet &AVs) { 1588 if (!BT.reached(&B)) 1589 return false; 1590 RegisterSet AVB(AVs); 1591 bool Changed = false; 1592 RegisterSet Defs; 1593 1594 for (auto I = B.begin(), E = B.end(), NextI = I; I != E; 1595 ++I, AVB.insert(Defs)) { 1596 NextI = std::next(I); 1597 Defs.clear(); 1598 HBS::getInstrDefs(*I, Defs); 1599 1600 unsigned Opc = I->getOpcode(); 1601 if (CopyPropagation::isCopyReg(Opc, false) || 1602 ConstGeneration::isTfrConst(*I)) 1603 continue; 1604 1605 DebugLoc DL = I->getDebugLoc(); 1606 auto At = I->isPHI() ? B.getFirstNonPHI() : I; 1607 1608 for (unsigned R = Defs.find_first(); R; R = Defs.find_next(R)) { 1609 BitTracker::RegisterRef MR; 1610 auto *FRC = HBS::getFinalVRegClass(R, MRI); 1611 1612 if (findMatch(R, MR, AVB)) { 1613 unsigned NewR = MRI.createVirtualRegister(FRC); 1614 BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR) 1615 .addReg(MR.Reg, 0, MR.Sub); 1616 BT.put(BitTracker::RegisterRef(NewR), BT.get(MR)); 1617 HBS::replaceReg(R, NewR, MRI); 1618 Forbidden.insert(R); 1619 continue; 1620 } 1621 1622 if (FRC == &Hexagon::DoubleRegsRegClass || 1623 FRC == &Hexagon::HvxWRRegClass) { 1624 // Try to generate REG_SEQUENCE. 1625 unsigned SubLo = HRI.getHexagonSubRegIndex(*FRC, Hexagon::ps_sub_lo); 1626 unsigned SubHi = HRI.getHexagonSubRegIndex(*FRC, Hexagon::ps_sub_hi); 1627 BitTracker::RegisterRef TL = { R, SubLo }; 1628 BitTracker::RegisterRef TH = { R, SubHi }; 1629 BitTracker::RegisterRef ML, MH; 1630 if (findMatch(TL, ML, AVB) && findMatch(TH, MH, AVB)) { 1631 auto *FRC = HBS::getFinalVRegClass(R, MRI); 1632 unsigned NewR = MRI.createVirtualRegister(FRC); 1633 BuildMI(B, At, DL, HII.get(TargetOpcode::REG_SEQUENCE), NewR) 1634 .addReg(ML.Reg, 0, ML.Sub) 1635 .addImm(SubLo) 1636 .addReg(MH.Reg, 0, MH.Sub) 1637 .addImm(SubHi); 1638 BT.put(BitTracker::RegisterRef(NewR), BT.get(R)); 1639 HBS::replaceReg(R, NewR, MRI); 1640 Forbidden.insert(R); 1641 } 1642 } 1643 } 1644 } 1645 1646 return Changed; 1647 } 1648 1649 bool CopyPropagation::isCopyReg(unsigned Opc, bool NoConv) { 1650 switch (Opc) { 1651 case TargetOpcode::COPY: 1652 case TargetOpcode::REG_SEQUENCE: 1653 case Hexagon::A4_combineir: 1654 case Hexagon::A4_combineri: 1655 return true; 1656 case Hexagon::A2_tfr: 1657 case Hexagon::A2_tfrp: 1658 case Hexagon::A2_combinew: 1659 case Hexagon::V6_vcombine: 1660 return NoConv; 1661 default: 1662 break; 1663 } 1664 return false; 1665 } 1666 1667 bool CopyPropagation::propagateRegCopy(MachineInstr &MI) { 1668 bool Changed = false; 1669 unsigned Opc = MI.getOpcode(); 1670 BitTracker::RegisterRef RD = MI.getOperand(0); 1671 assert(MI.getOperand(0).getSubReg() == 0); 1672 1673 switch (Opc) { 1674 case TargetOpcode::COPY: 1675 case Hexagon::A2_tfr: 1676 case Hexagon::A2_tfrp: { 1677 BitTracker::RegisterRef RS = MI.getOperand(1); 1678 if (!HBS::isTransparentCopy(RD, RS, MRI)) 1679 break; 1680 if (RS.Sub != 0) 1681 Changed = HBS::replaceRegWithSub(RD.Reg, RS.Reg, RS.Sub, MRI); 1682 else 1683 Changed = HBS::replaceReg(RD.Reg, RS.Reg, MRI); 1684 break; 1685 } 1686 case TargetOpcode::REG_SEQUENCE: { 1687 BitTracker::RegisterRef SL, SH; 1688 if (HBS::parseRegSequence(MI, SL, SH, MRI)) { 1689 const TargetRegisterClass &RC = *MRI.getRegClass(RD.Reg); 1690 unsigned SubLo = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_lo); 1691 unsigned SubHi = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_hi); 1692 Changed = HBS::replaceSubWithSub(RD.Reg, SubLo, SL.Reg, SL.Sub, MRI); 1693 Changed |= HBS::replaceSubWithSub(RD.Reg, SubHi, SH.Reg, SH.Sub, MRI); 1694 } 1695 break; 1696 } 1697 case Hexagon::A2_combinew: 1698 case Hexagon::V6_vcombine: { 1699 const TargetRegisterClass &RC = *MRI.getRegClass(RD.Reg); 1700 unsigned SubLo = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_lo); 1701 unsigned SubHi = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_hi); 1702 BitTracker::RegisterRef RH = MI.getOperand(1), RL = MI.getOperand(2); 1703 Changed = HBS::replaceSubWithSub(RD.Reg, SubLo, RL.Reg, RL.Sub, MRI); 1704 Changed |= HBS::replaceSubWithSub(RD.Reg, SubHi, RH.Reg, RH.Sub, MRI); 1705 break; 1706 } 1707 case Hexagon::A4_combineir: 1708 case Hexagon::A4_combineri: { 1709 unsigned SrcX = (Opc == Hexagon::A4_combineir) ? 2 : 1; 1710 unsigned Sub = (Opc == Hexagon::A4_combineir) ? Hexagon::isub_lo 1711 : Hexagon::isub_hi; 1712 BitTracker::RegisterRef RS = MI.getOperand(SrcX); 1713 Changed = HBS::replaceSubWithSub(RD.Reg, Sub, RS.Reg, RS.Sub, MRI); 1714 break; 1715 } 1716 } 1717 return Changed; 1718 } 1719 1720 bool CopyPropagation::processBlock(MachineBasicBlock &B, const RegisterSet&) { 1721 std::vector<MachineInstr*> Instrs; 1722 for (auto I = B.rbegin(), E = B.rend(); I != E; ++I) 1723 Instrs.push_back(&*I); 1724 1725 bool Changed = false; 1726 for (auto I : Instrs) { 1727 unsigned Opc = I->getOpcode(); 1728 if (!CopyPropagation::isCopyReg(Opc, true)) 1729 continue; 1730 Changed |= propagateRegCopy(*I); 1731 } 1732 1733 return Changed; 1734 } 1735 1736 namespace { 1737 1738 // Recognize patterns that can be simplified and replace them with the 1739 // simpler forms. 1740 // This is by no means complete 1741 class BitSimplification : public Transformation { 1742 public: 1743 BitSimplification(BitTracker &bt, const MachineDominatorTree &mdt, 1744 const HexagonInstrInfo &hii, const HexagonRegisterInfo &hri, 1745 MachineRegisterInfo &mri, MachineFunction &mf) 1746 : Transformation(true), MDT(mdt), HII(hii), HRI(hri), MRI(mri), 1747 MF(mf), BT(bt) {} 1748 1749 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; 1750 1751 private: 1752 struct RegHalf : public BitTracker::RegisterRef { 1753 bool Low; // Low/High halfword. 1754 }; 1755 1756 bool matchHalf(unsigned SelfR, const BitTracker::RegisterCell &RC, 1757 unsigned B, RegHalf &RH); 1758 bool validateReg(BitTracker::RegisterRef R, unsigned Opc, unsigned OpNum); 1759 1760 bool matchPackhl(unsigned SelfR, const BitTracker::RegisterCell &RC, 1761 BitTracker::RegisterRef &Rs, BitTracker::RegisterRef &Rt); 1762 unsigned getCombineOpcode(bool HLow, bool LLow); 1763 1764 bool genStoreUpperHalf(MachineInstr *MI); 1765 bool genStoreImmediate(MachineInstr *MI); 1766 bool genPackhl(MachineInstr *MI, BitTracker::RegisterRef RD, 1767 const BitTracker::RegisterCell &RC); 1768 bool genExtractHalf(MachineInstr *MI, BitTracker::RegisterRef RD, 1769 const BitTracker::RegisterCell &RC); 1770 bool genCombineHalf(MachineInstr *MI, BitTracker::RegisterRef RD, 1771 const BitTracker::RegisterCell &RC); 1772 bool genExtractLow(MachineInstr *MI, BitTracker::RegisterRef RD, 1773 const BitTracker::RegisterCell &RC); 1774 bool genBitSplit(MachineInstr *MI, BitTracker::RegisterRef RD, 1775 const BitTracker::RegisterCell &RC, const RegisterSet &AVs); 1776 bool simplifyTstbit(MachineInstr *MI, BitTracker::RegisterRef RD, 1777 const BitTracker::RegisterCell &RC); 1778 bool simplifyExtractLow(MachineInstr *MI, BitTracker::RegisterRef RD, 1779 const BitTracker::RegisterCell &RC, const RegisterSet &AVs); 1780 bool simplifyRCmp0(MachineInstr *MI, BitTracker::RegisterRef RD); 1781 1782 // Cache of created instructions to avoid creating duplicates. 1783 // XXX Currently only used by genBitSplit. 1784 std::vector<MachineInstr*> NewMIs; 1785 1786 const MachineDominatorTree &MDT; 1787 const HexagonInstrInfo &HII; 1788 const HexagonRegisterInfo &HRI; 1789 MachineRegisterInfo &MRI; 1790 MachineFunction &MF; 1791 BitTracker &BT; 1792 }; 1793 1794 } // end anonymous namespace 1795 1796 // Check if the bits [B..B+16) in register cell RC form a valid halfword, 1797 // i.e. [0..16), [16..32), etc. of some register. If so, return true and 1798 // set the information about the found register in RH. 1799 bool BitSimplification::matchHalf(unsigned SelfR, 1800 const BitTracker::RegisterCell &RC, unsigned B, RegHalf &RH) { 1801 // XXX This could be searching in the set of available registers, in case 1802 // the match is not exact. 1803 1804 // Match 16-bit chunks, where the RC[B..B+15] references exactly one 1805 // register and all the bits B..B+15 match between RC and the register. 1806 // This is meant to match "v1[0-15]", where v1 = { [0]:0 [1-15]:v1... }, 1807 // and RC = { [0]:0 [1-15]:v1[1-15]... }. 1808 bool Low = false; 1809 unsigned I = B; 1810 while (I < B+16 && RC[I].num()) 1811 I++; 1812 if (I == B+16) 1813 return false; 1814 1815 unsigned Reg = RC[I].RefI.Reg; 1816 unsigned P = RC[I].RefI.Pos; // The RefI.Pos will be advanced by I-B. 1817 if (P < I-B) 1818 return false; 1819 unsigned Pos = P - (I-B); 1820 1821 if (Reg == 0 || Reg == SelfR) // Don't match "self". 1822 return false; 1823 if (!TargetRegisterInfo::isVirtualRegister(Reg)) 1824 return false; 1825 if (!BT.has(Reg)) 1826 return false; 1827 1828 const BitTracker::RegisterCell &SC = BT.lookup(Reg); 1829 if (Pos+16 > SC.width()) 1830 return false; 1831 1832 for (unsigned i = 0; i < 16; ++i) { 1833 const BitTracker::BitValue &RV = RC[i+B]; 1834 if (RV.Type == BitTracker::BitValue::Ref) { 1835 if (RV.RefI.Reg != Reg) 1836 return false; 1837 if (RV.RefI.Pos != i+Pos) 1838 return false; 1839 continue; 1840 } 1841 if (RC[i+B] != SC[i+Pos]) 1842 return false; 1843 } 1844 1845 unsigned Sub = 0; 1846 switch (Pos) { 1847 case 0: 1848 Sub = Hexagon::isub_lo; 1849 Low = true; 1850 break; 1851 case 16: 1852 Sub = Hexagon::isub_lo; 1853 Low = false; 1854 break; 1855 case 32: 1856 Sub = Hexagon::isub_hi; 1857 Low = true; 1858 break; 1859 case 48: 1860 Sub = Hexagon::isub_hi; 1861 Low = false; 1862 break; 1863 default: 1864 return false; 1865 } 1866 1867 RH.Reg = Reg; 1868 RH.Sub = Sub; 1869 RH.Low = Low; 1870 // If the subregister is not valid with the register, set it to 0. 1871 if (!HBS::getFinalVRegClass(RH, MRI)) 1872 RH.Sub = 0; 1873 1874 return true; 1875 } 1876 1877 bool BitSimplification::validateReg(BitTracker::RegisterRef R, unsigned Opc, 1878 unsigned OpNum) { 1879 auto *OpRC = HII.getRegClass(HII.get(Opc), OpNum, &HRI, MF); 1880 auto *RRC = HBS::getFinalVRegClass(R, MRI); 1881 return OpRC->hasSubClassEq(RRC); 1882 } 1883 1884 // Check if RC matches the pattern of a S2_packhl. If so, return true and 1885 // set the inputs Rs and Rt. 1886 bool BitSimplification::matchPackhl(unsigned SelfR, 1887 const BitTracker::RegisterCell &RC, BitTracker::RegisterRef &Rs, 1888 BitTracker::RegisterRef &Rt) { 1889 RegHalf L1, H1, L2, H2; 1890 1891 if (!matchHalf(SelfR, RC, 0, L2) || !matchHalf(SelfR, RC, 16, L1)) 1892 return false; 1893 if (!matchHalf(SelfR, RC, 32, H2) || !matchHalf(SelfR, RC, 48, H1)) 1894 return false; 1895 1896 // Rs = H1.L1, Rt = H2.L2 1897 if (H1.Reg != L1.Reg || H1.Sub != L1.Sub || H1.Low || !L1.Low) 1898 return false; 1899 if (H2.Reg != L2.Reg || H2.Sub != L2.Sub || H2.Low || !L2.Low) 1900 return false; 1901 1902 Rs = H1; 1903 Rt = H2; 1904 return true; 1905 } 1906 1907 unsigned BitSimplification::getCombineOpcode(bool HLow, bool LLow) { 1908 return HLow ? LLow ? Hexagon::A2_combine_ll 1909 : Hexagon::A2_combine_lh 1910 : LLow ? Hexagon::A2_combine_hl 1911 : Hexagon::A2_combine_hh; 1912 } 1913 1914 // If MI stores the upper halfword of a register (potentially obtained via 1915 // shifts or extracts), replace it with a storerf instruction. This could 1916 // cause the "extraction" code to become dead. 1917 bool BitSimplification::genStoreUpperHalf(MachineInstr *MI) { 1918 unsigned Opc = MI->getOpcode(); 1919 if (Opc != Hexagon::S2_storerh_io) 1920 return false; 1921 1922 MachineOperand &ValOp = MI->getOperand(2); 1923 BitTracker::RegisterRef RS = ValOp; 1924 if (!BT.has(RS.Reg)) 1925 return false; 1926 const BitTracker::RegisterCell &RC = BT.lookup(RS.Reg); 1927 RegHalf H; 1928 if (!matchHalf(0, RC, 0, H)) 1929 return false; 1930 if (H.Low) 1931 return false; 1932 MI->setDesc(HII.get(Hexagon::S2_storerf_io)); 1933 ValOp.setReg(H.Reg); 1934 ValOp.setSubReg(H.Sub); 1935 return true; 1936 } 1937 1938 // If MI stores a value known at compile-time, and the value is within a range 1939 // that avoids using constant-extenders, replace it with a store-immediate. 1940 bool BitSimplification::genStoreImmediate(MachineInstr *MI) { 1941 unsigned Opc = MI->getOpcode(); 1942 unsigned Align = 0; 1943 switch (Opc) { 1944 case Hexagon::S2_storeri_io: 1945 Align++; 1946 LLVM_FALLTHROUGH; 1947 case Hexagon::S2_storerh_io: 1948 Align++; 1949 LLVM_FALLTHROUGH; 1950 case Hexagon::S2_storerb_io: 1951 break; 1952 default: 1953 return false; 1954 } 1955 1956 // Avoid stores to frame-indices (due to an unknown offset). 1957 if (!MI->getOperand(0).isReg()) 1958 return false; 1959 MachineOperand &OffOp = MI->getOperand(1); 1960 if (!OffOp.isImm()) 1961 return false; 1962 1963 int64_t Off = OffOp.getImm(); 1964 // Offset is u6:a. Sadly, there is no isShiftedUInt(n,x). 1965 if (!isUIntN(6+Align, Off) || (Off & ((1<<Align)-1))) 1966 return false; 1967 // Source register: 1968 BitTracker::RegisterRef RS = MI->getOperand(2); 1969 if (!BT.has(RS.Reg)) 1970 return false; 1971 const BitTracker::RegisterCell &RC = BT.lookup(RS.Reg); 1972 uint64_t U; 1973 if (!HBS::getConst(RC, 0, RC.width(), U)) 1974 return false; 1975 1976 // Only consider 8-bit values to avoid constant-extenders. 1977 int V; 1978 switch (Opc) { 1979 case Hexagon::S2_storerb_io: 1980 V = int8_t(U); 1981 break; 1982 case Hexagon::S2_storerh_io: 1983 V = int16_t(U); 1984 break; 1985 case Hexagon::S2_storeri_io: 1986 V = int32_t(U); 1987 break; 1988 } 1989 if (!isInt<8>(V)) 1990 return false; 1991 1992 MI->RemoveOperand(2); 1993 switch (Opc) { 1994 case Hexagon::S2_storerb_io: 1995 MI->setDesc(HII.get(Hexagon::S4_storeirb_io)); 1996 break; 1997 case Hexagon::S2_storerh_io: 1998 MI->setDesc(HII.get(Hexagon::S4_storeirh_io)); 1999 break; 2000 case Hexagon::S2_storeri_io: 2001 MI->setDesc(HII.get(Hexagon::S4_storeiri_io)); 2002 break; 2003 } 2004 MI->addOperand(MachineOperand::CreateImm(V)); 2005 return true; 2006 } 2007 2008 // If MI is equivalent o S2_packhl, generate the S2_packhl. MI could be the 2009 // last instruction in a sequence that results in something equivalent to 2010 // the pack-halfwords. The intent is to cause the entire sequence to become 2011 // dead. 2012 bool BitSimplification::genPackhl(MachineInstr *MI, 2013 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { 2014 unsigned Opc = MI->getOpcode(); 2015 if (Opc == Hexagon::S2_packhl) 2016 return false; 2017 BitTracker::RegisterRef Rs, Rt; 2018 if (!matchPackhl(RD.Reg, RC, Rs, Rt)) 2019 return false; 2020 if (!validateReg(Rs, Hexagon::S2_packhl, 1) || 2021 !validateReg(Rt, Hexagon::S2_packhl, 2)) 2022 return false; 2023 2024 MachineBasicBlock &B = *MI->getParent(); 2025 unsigned NewR = MRI.createVirtualRegister(&Hexagon::DoubleRegsRegClass); 2026 DebugLoc DL = MI->getDebugLoc(); 2027 auto At = MI->isPHI() ? B.getFirstNonPHI() 2028 : MachineBasicBlock::iterator(MI); 2029 BuildMI(B, At, DL, HII.get(Hexagon::S2_packhl), NewR) 2030 .addReg(Rs.Reg, 0, Rs.Sub) 2031 .addReg(Rt.Reg, 0, Rt.Sub); 2032 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); 2033 BT.put(BitTracker::RegisterRef(NewR), RC); 2034 return true; 2035 } 2036 2037 // If MI produces halfword of the input in the low half of the output, 2038 // replace it with zero-extend or extractu. 2039 bool BitSimplification::genExtractHalf(MachineInstr *MI, 2040 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { 2041 RegHalf L; 2042 // Check for halfword in low 16 bits, zeros elsewhere. 2043 if (!matchHalf(RD.Reg, RC, 0, L) || !HBS::isZero(RC, 16, 16)) 2044 return false; 2045 2046 unsigned Opc = MI->getOpcode(); 2047 MachineBasicBlock &B = *MI->getParent(); 2048 DebugLoc DL = MI->getDebugLoc(); 2049 2050 // Prefer zxth, since zxth can go in any slot, while extractu only in 2051 // slots 2 and 3. 2052 unsigned NewR = 0; 2053 auto At = MI->isPHI() ? B.getFirstNonPHI() 2054 : MachineBasicBlock::iterator(MI); 2055 if (L.Low && Opc != Hexagon::A2_zxth) { 2056 if (validateReg(L, Hexagon::A2_zxth, 1)) { 2057 NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 2058 BuildMI(B, At, DL, HII.get(Hexagon::A2_zxth), NewR) 2059 .addReg(L.Reg, 0, L.Sub); 2060 } 2061 } else if (!L.Low && Opc != Hexagon::S2_lsr_i_r) { 2062 if (validateReg(L, Hexagon::S2_lsr_i_r, 1)) { 2063 NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 2064 BuildMI(B, MI, DL, HII.get(Hexagon::S2_lsr_i_r), NewR) 2065 .addReg(L.Reg, 0, L.Sub) 2066 .addImm(16); 2067 } 2068 } 2069 if (NewR == 0) 2070 return false; 2071 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); 2072 BT.put(BitTracker::RegisterRef(NewR), RC); 2073 return true; 2074 } 2075 2076 // If MI is equivalent to a combine(.L/.H, .L/.H) replace with with the 2077 // combine. 2078 bool BitSimplification::genCombineHalf(MachineInstr *MI, 2079 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { 2080 RegHalf L, H; 2081 // Check for combine h/l 2082 if (!matchHalf(RD.Reg, RC, 0, L) || !matchHalf(RD.Reg, RC, 16, H)) 2083 return false; 2084 // Do nothing if this is just a reg copy. 2085 if (L.Reg == H.Reg && L.Sub == H.Sub && !H.Low && L.Low) 2086 return false; 2087 2088 unsigned Opc = MI->getOpcode(); 2089 unsigned COpc = getCombineOpcode(H.Low, L.Low); 2090 if (COpc == Opc) 2091 return false; 2092 if (!validateReg(H, COpc, 1) || !validateReg(L, COpc, 2)) 2093 return false; 2094 2095 MachineBasicBlock &B = *MI->getParent(); 2096 DebugLoc DL = MI->getDebugLoc(); 2097 unsigned NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 2098 auto At = MI->isPHI() ? B.getFirstNonPHI() 2099 : MachineBasicBlock::iterator(MI); 2100 BuildMI(B, At, DL, HII.get(COpc), NewR) 2101 .addReg(H.Reg, 0, H.Sub) 2102 .addReg(L.Reg, 0, L.Sub); 2103 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); 2104 BT.put(BitTracker::RegisterRef(NewR), RC); 2105 return true; 2106 } 2107 2108 // If MI resets high bits of a register and keeps the lower ones, replace it 2109 // with zero-extend byte/half, and-immediate, or extractu, as appropriate. 2110 bool BitSimplification::genExtractLow(MachineInstr *MI, 2111 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { 2112 unsigned Opc = MI->getOpcode(); 2113 switch (Opc) { 2114 case Hexagon::A2_zxtb: 2115 case Hexagon::A2_zxth: 2116 case Hexagon::S2_extractu: 2117 return false; 2118 } 2119 if (Opc == Hexagon::A2_andir && MI->getOperand(2).isImm()) { 2120 int32_t Imm = MI->getOperand(2).getImm(); 2121 if (isInt<10>(Imm)) 2122 return false; 2123 } 2124 2125 if (MI->hasUnmodeledSideEffects() || MI->isInlineAsm()) 2126 return false; 2127 unsigned W = RC.width(); 2128 while (W > 0 && RC[W-1].is(0)) 2129 W--; 2130 if (W == 0 || W == RC.width()) 2131 return false; 2132 unsigned NewOpc = (W == 8) ? Hexagon::A2_zxtb 2133 : (W == 16) ? Hexagon::A2_zxth 2134 : (W < 10) ? Hexagon::A2_andir 2135 : Hexagon::S2_extractu; 2136 MachineBasicBlock &B = *MI->getParent(); 2137 DebugLoc DL = MI->getDebugLoc(); 2138 2139 for (auto &Op : MI->uses()) { 2140 if (!Op.isReg()) 2141 continue; 2142 BitTracker::RegisterRef RS = Op; 2143 if (!BT.has(RS.Reg)) 2144 continue; 2145 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); 2146 unsigned BN, BW; 2147 if (!HBS::getSubregMask(RS, BN, BW, MRI)) 2148 continue; 2149 if (BW < W || !HBS::isEqual(RC, 0, SC, BN, W)) 2150 continue; 2151 if (!validateReg(RS, NewOpc, 1)) 2152 continue; 2153 2154 unsigned NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 2155 auto At = MI->isPHI() ? B.getFirstNonPHI() 2156 : MachineBasicBlock::iterator(MI); 2157 auto MIB = BuildMI(B, At, DL, HII.get(NewOpc), NewR) 2158 .addReg(RS.Reg, 0, RS.Sub); 2159 if (NewOpc == Hexagon::A2_andir) 2160 MIB.addImm((1 << W) - 1); 2161 else if (NewOpc == Hexagon::S2_extractu) 2162 MIB.addImm(W).addImm(0); 2163 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); 2164 BT.put(BitTracker::RegisterRef(NewR), RC); 2165 return true; 2166 } 2167 return false; 2168 } 2169 2170 bool BitSimplification::genBitSplit(MachineInstr *MI, 2171 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC, 2172 const RegisterSet &AVs) { 2173 if (!GenBitSplit) 2174 return false; 2175 if (MaxBitSplit.getNumOccurrences()) { 2176 if (CountBitSplit >= MaxBitSplit) 2177 return false; 2178 } 2179 2180 unsigned Opc = MI->getOpcode(); 2181 switch (Opc) { 2182 case Hexagon::A4_bitsplit: 2183 case Hexagon::A4_bitspliti: 2184 return false; 2185 } 2186 2187 unsigned W = RC.width(); 2188 if (W != 32) 2189 return false; 2190 2191 auto ctlz = [] (const BitTracker::RegisterCell &C) -> unsigned { 2192 unsigned Z = C.width(); 2193 while (Z > 0 && C[Z-1].is(0)) 2194 --Z; 2195 return C.width() - Z; 2196 }; 2197 2198 // Count the number of leading zeros in the target RC. 2199 unsigned Z = ctlz(RC); 2200 if (Z == 0 || Z == W) 2201 return false; 2202 2203 // A simplistic analysis: assume the source register (the one being split) 2204 // is fully unknown, and that all its bits are self-references. 2205 const BitTracker::BitValue &B0 = RC[0]; 2206 if (B0.Type != BitTracker::BitValue::Ref) 2207 return false; 2208 2209 unsigned SrcR = B0.RefI.Reg; 2210 unsigned SrcSR = 0; 2211 unsigned Pos = B0.RefI.Pos; 2212 2213 // All the non-zero bits should be consecutive bits from the same register. 2214 for (unsigned i = 1; i < W-Z; ++i) { 2215 const BitTracker::BitValue &V = RC[i]; 2216 if (V.Type != BitTracker::BitValue::Ref) 2217 return false; 2218 if (V.RefI.Reg != SrcR || V.RefI.Pos != Pos+i) 2219 return false; 2220 } 2221 2222 // Now, find the other bitfield among AVs. 2223 for (unsigned S = AVs.find_first(); S; S = AVs.find_next(S)) { 2224 // The number of leading zeros here should be the number of trailing 2225 // non-zeros in RC. 2226 if (!BT.has(S)) 2227 continue; 2228 const BitTracker::RegisterCell &SC = BT.lookup(S); 2229 if (SC.width() != W || ctlz(SC) != W-Z) 2230 continue; 2231 // The Z lower bits should now match SrcR. 2232 const BitTracker::BitValue &S0 = SC[0]; 2233 if (S0.Type != BitTracker::BitValue::Ref || S0.RefI.Reg != SrcR) 2234 continue; 2235 unsigned P = S0.RefI.Pos; 2236 2237 if (Pos <= P && (Pos + W-Z) != P) 2238 continue; 2239 if (P < Pos && (P + Z) != Pos) 2240 continue; 2241 // The starting bitfield position must be at a subregister boundary. 2242 if (std::min(P, Pos) != 0 && std::min(P, Pos) != 32) 2243 continue; 2244 2245 unsigned I; 2246 for (I = 1; I < Z; ++I) { 2247 const BitTracker::BitValue &V = SC[I]; 2248 if (V.Type != BitTracker::BitValue::Ref) 2249 break; 2250 if (V.RefI.Reg != SrcR || V.RefI.Pos != P+I) 2251 break; 2252 } 2253 if (I != Z) 2254 continue; 2255 2256 // Generate bitsplit where S is defined. 2257 if (MaxBitSplit.getNumOccurrences()) 2258 CountBitSplit++; 2259 MachineInstr *DefS = MRI.getVRegDef(S); 2260 assert(DefS != nullptr); 2261 DebugLoc DL = DefS->getDebugLoc(); 2262 MachineBasicBlock &B = *DefS->getParent(); 2263 auto At = DefS->isPHI() ? B.getFirstNonPHI() 2264 : MachineBasicBlock::iterator(DefS); 2265 if (MRI.getRegClass(SrcR)->getID() == Hexagon::DoubleRegsRegClassID) 2266 SrcSR = (std::min(Pos, P) == 32) ? Hexagon::isub_hi : Hexagon::isub_lo; 2267 if (!validateReg({SrcR,SrcSR}, Hexagon::A4_bitspliti, 1)) 2268 continue; 2269 unsigned ImmOp = Pos <= P ? W-Z : Z; 2270 2271 // Find an existing bitsplit instruction if one already exists. 2272 unsigned NewR = 0; 2273 for (MachineInstr *In : NewMIs) { 2274 if (In->getOpcode() != Hexagon::A4_bitspliti) 2275 continue; 2276 MachineOperand &Op1 = In->getOperand(1); 2277 if (Op1.getReg() != SrcR || Op1.getSubReg() != SrcSR) 2278 continue; 2279 if (In->getOperand(2).getImm() != ImmOp) 2280 continue; 2281 // Check if the target register is available here. 2282 MachineOperand &Op0 = In->getOperand(0); 2283 MachineInstr *DefI = MRI.getVRegDef(Op0.getReg()); 2284 assert(DefI != nullptr); 2285 if (!MDT.dominates(DefI, &*At)) 2286 continue; 2287 2288 // Found one that can be reused. 2289 assert(Op0.getSubReg() == 0); 2290 NewR = Op0.getReg(); 2291 break; 2292 } 2293 if (!NewR) { 2294 NewR = MRI.createVirtualRegister(&Hexagon::DoubleRegsRegClass); 2295 auto NewBS = BuildMI(B, At, DL, HII.get(Hexagon::A4_bitspliti), NewR) 2296 .addReg(SrcR, 0, SrcSR) 2297 .addImm(ImmOp); 2298 NewMIs.push_back(NewBS); 2299 } 2300 if (Pos <= P) { 2301 HBS::replaceRegWithSub(RD.Reg, NewR, Hexagon::isub_lo, MRI); 2302 HBS::replaceRegWithSub(S, NewR, Hexagon::isub_hi, MRI); 2303 } else { 2304 HBS::replaceRegWithSub(S, NewR, Hexagon::isub_lo, MRI); 2305 HBS::replaceRegWithSub(RD.Reg, NewR, Hexagon::isub_hi, MRI); 2306 } 2307 return true; 2308 } 2309 2310 return false; 2311 } 2312 2313 // Check for tstbit simplification opportunity, where the bit being checked 2314 // can be tracked back to another register. For example: 2315 // %2 = S2_lsr_i_r %1, 5 2316 // %3 = S2_tstbit_i %2, 0 2317 // => 2318 // %3 = S2_tstbit_i %1, 5 2319 bool BitSimplification::simplifyTstbit(MachineInstr *MI, 2320 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { 2321 unsigned Opc = MI->getOpcode(); 2322 if (Opc != Hexagon::S2_tstbit_i) 2323 return false; 2324 2325 unsigned BN = MI->getOperand(2).getImm(); 2326 BitTracker::RegisterRef RS = MI->getOperand(1); 2327 unsigned F, W; 2328 DebugLoc DL = MI->getDebugLoc(); 2329 if (!BT.has(RS.Reg) || !HBS::getSubregMask(RS, F, W, MRI)) 2330 return false; 2331 MachineBasicBlock &B = *MI->getParent(); 2332 auto At = MI->isPHI() ? B.getFirstNonPHI() 2333 : MachineBasicBlock::iterator(MI); 2334 2335 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); 2336 const BitTracker::BitValue &V = SC[F+BN]; 2337 if (V.Type == BitTracker::BitValue::Ref && V.RefI.Reg != RS.Reg) { 2338 const TargetRegisterClass *TC = MRI.getRegClass(V.RefI.Reg); 2339 // Need to map V.RefI.Reg to a 32-bit register, i.e. if it is 2340 // a double register, need to use a subregister and adjust bit 2341 // number. 2342 unsigned P = std::numeric_limits<unsigned>::max(); 2343 BitTracker::RegisterRef RR(V.RefI.Reg, 0); 2344 if (TC == &Hexagon::DoubleRegsRegClass) { 2345 P = V.RefI.Pos; 2346 RR.Sub = Hexagon::isub_lo; 2347 if (P >= 32) { 2348 P -= 32; 2349 RR.Sub = Hexagon::isub_hi; 2350 } 2351 } else if (TC == &Hexagon::IntRegsRegClass) { 2352 P = V.RefI.Pos; 2353 } 2354 if (P != std::numeric_limits<unsigned>::max()) { 2355 unsigned NewR = MRI.createVirtualRegister(&Hexagon::PredRegsRegClass); 2356 BuildMI(B, At, DL, HII.get(Hexagon::S2_tstbit_i), NewR) 2357 .addReg(RR.Reg, 0, RR.Sub) 2358 .addImm(P); 2359 HBS::replaceReg(RD.Reg, NewR, MRI); 2360 BT.put(NewR, RC); 2361 return true; 2362 } 2363 } else if (V.is(0) || V.is(1)) { 2364 unsigned NewR = MRI.createVirtualRegister(&Hexagon::PredRegsRegClass); 2365 unsigned NewOpc = V.is(0) ? Hexagon::PS_false : Hexagon::PS_true; 2366 BuildMI(B, At, DL, HII.get(NewOpc), NewR); 2367 HBS::replaceReg(RD.Reg, NewR, MRI); 2368 return true; 2369 } 2370 2371 return false; 2372 } 2373 2374 // Detect whether RD is a bitfield extract (sign- or zero-extended) of 2375 // some register from the AVs set. Create a new corresponding instruction 2376 // at the location of MI. The intent is to recognize situations where 2377 // a sequence of instructions performs an operation that is equivalent to 2378 // an extract operation, such as a shift left followed by a shift right. 2379 bool BitSimplification::simplifyExtractLow(MachineInstr *MI, 2380 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC, 2381 const RegisterSet &AVs) { 2382 if (!GenExtract) 2383 return false; 2384 if (MaxExtract.getNumOccurrences()) { 2385 if (CountExtract >= MaxExtract) 2386 return false; 2387 CountExtract++; 2388 } 2389 2390 unsigned W = RC.width(); 2391 unsigned RW = W; 2392 unsigned Len; 2393 bool Signed; 2394 2395 // The code is mostly class-independent, except for the part that generates 2396 // the extract instruction, and establishes the source register (in case it 2397 // needs to use a subregister). 2398 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); 2399 if (FRC != &Hexagon::IntRegsRegClass && FRC != &Hexagon::DoubleRegsRegClass) 2400 return false; 2401 assert(RD.Sub == 0); 2402 2403 // Observation: 2404 // If the cell has a form of 00..0xx..x with k zeros and n remaining 2405 // bits, this could be an extractu of the n bits, but it could also be 2406 // an extractu of a longer field which happens to have 0s in the top 2407 // bit positions. 2408 // The same logic applies to sign-extended fields. 2409 // 2410 // Do not check for the extended extracts, since it would expand the 2411 // search space quite a bit. The search may be expensive as it is. 2412 2413 const BitTracker::BitValue &TopV = RC[W-1]; 2414 2415 // Eliminate candidates that have self-referential bits, since they 2416 // cannot be extracts from other registers. Also, skip registers that 2417 // have compile-time constant values. 2418 bool IsConst = true; 2419 for (unsigned I = 0; I != W; ++I) { 2420 const BitTracker::BitValue &V = RC[I]; 2421 if (V.Type == BitTracker::BitValue::Ref && V.RefI.Reg == RD.Reg) 2422 return false; 2423 IsConst = IsConst && (V.is(0) || V.is(1)); 2424 } 2425 if (IsConst) 2426 return false; 2427 2428 if (TopV.is(0) || TopV.is(1)) { 2429 bool S = TopV.is(1); 2430 for (--W; W > 0 && RC[W-1].is(S); --W) 2431 ; 2432 Len = W; 2433 Signed = S; 2434 // The sign bit must be a part of the field being extended. 2435 if (Signed) 2436 ++Len; 2437 } else { 2438 // This could still be a sign-extended extract. 2439 assert(TopV.Type == BitTracker::BitValue::Ref); 2440 if (TopV.RefI.Reg == RD.Reg || TopV.RefI.Pos == W-1) 2441 return false; 2442 for (--W; W > 0 && RC[W-1] == TopV; --W) 2443 ; 2444 // The top bits of RC are copies of TopV. One occurrence of TopV will 2445 // be a part of the field. 2446 Len = W + 1; 2447 Signed = true; 2448 } 2449 2450 // This would be just a copy. It should be handled elsewhere. 2451 if (Len == RW) 2452 return false; 2453 2454 LLVM_DEBUG({ 2455 dbgs() << __func__ << " on reg: " << printReg(RD.Reg, &HRI, RD.Sub) 2456 << ", MI: " << *MI; 2457 dbgs() << "Cell: " << RC << '\n'; 2458 dbgs() << "Expected bitfield size: " << Len << " bits, " 2459 << (Signed ? "sign" : "zero") << "-extended\n"; 2460 }); 2461 2462 bool Changed = false; 2463 2464 for (unsigned R = AVs.find_first(); R != 0; R = AVs.find_next(R)) { 2465 if (!BT.has(R)) 2466 continue; 2467 const BitTracker::RegisterCell &SC = BT.lookup(R); 2468 unsigned SW = SC.width(); 2469 2470 // The source can be longer than the destination, as long as its size is 2471 // a multiple of the size of the destination. Also, we would need to be 2472 // able to refer to the subregister in the source that would be of the 2473 // same size as the destination, but only check the sizes here. 2474 if (SW < RW || (SW % RW) != 0) 2475 continue; 2476 2477 // The field can start at any offset in SC as long as it contains Len 2478 // bits and does not cross subregister boundary (if the source register 2479 // is longer than the destination). 2480 unsigned Off = 0; 2481 while (Off <= SW-Len) { 2482 unsigned OE = (Off+Len)/RW; 2483 if (OE != Off/RW) { 2484 // The assumption here is that if the source (R) is longer than the 2485 // destination, then the destination is a sequence of words of 2486 // size RW, and each such word in R can be accessed via a subregister. 2487 // 2488 // If the beginning and the end of the field cross the subregister 2489 // boundary, advance to the next subregister. 2490 Off = OE*RW; 2491 continue; 2492 } 2493 if (HBS::isEqual(RC, 0, SC, Off, Len)) 2494 break; 2495 ++Off; 2496 } 2497 2498 if (Off > SW-Len) 2499 continue; 2500 2501 // Found match. 2502 unsigned ExtOpc = 0; 2503 if (Off == 0) { 2504 if (Len == 8) 2505 ExtOpc = Signed ? Hexagon::A2_sxtb : Hexagon::A2_zxtb; 2506 else if (Len == 16) 2507 ExtOpc = Signed ? Hexagon::A2_sxth : Hexagon::A2_zxth; 2508 else if (Len < 10 && !Signed) 2509 ExtOpc = Hexagon::A2_andir; 2510 } 2511 if (ExtOpc == 0) { 2512 ExtOpc = 2513 Signed ? (RW == 32 ? Hexagon::S4_extract : Hexagon::S4_extractp) 2514 : (RW == 32 ? Hexagon::S2_extractu : Hexagon::S2_extractup); 2515 } 2516 unsigned SR = 0; 2517 // This only recognizes isub_lo and isub_hi. 2518 if (RW != SW && RW*2 != SW) 2519 continue; 2520 if (RW != SW) 2521 SR = (Off/RW == 0) ? Hexagon::isub_lo : Hexagon::isub_hi; 2522 Off = Off % RW; 2523 2524 if (!validateReg({R,SR}, ExtOpc, 1)) 2525 continue; 2526 2527 // Don't generate the same instruction as the one being optimized. 2528 if (MI->getOpcode() == ExtOpc) { 2529 // All possible ExtOpc's have the source in operand(1). 2530 const MachineOperand &SrcOp = MI->getOperand(1); 2531 if (SrcOp.getReg() == R) 2532 continue; 2533 } 2534 2535 DebugLoc DL = MI->getDebugLoc(); 2536 MachineBasicBlock &B = *MI->getParent(); 2537 unsigned NewR = MRI.createVirtualRegister(FRC); 2538 auto At = MI->isPHI() ? B.getFirstNonPHI() 2539 : MachineBasicBlock::iterator(MI); 2540 auto MIB = BuildMI(B, At, DL, HII.get(ExtOpc), NewR) 2541 .addReg(R, 0, SR); 2542 switch (ExtOpc) { 2543 case Hexagon::A2_sxtb: 2544 case Hexagon::A2_zxtb: 2545 case Hexagon::A2_sxth: 2546 case Hexagon::A2_zxth: 2547 break; 2548 case Hexagon::A2_andir: 2549 MIB.addImm((1u << Len) - 1); 2550 break; 2551 case Hexagon::S4_extract: 2552 case Hexagon::S2_extractu: 2553 case Hexagon::S4_extractp: 2554 case Hexagon::S2_extractup: 2555 MIB.addImm(Len) 2556 .addImm(Off); 2557 break; 2558 default: 2559 llvm_unreachable("Unexpected opcode"); 2560 } 2561 2562 HBS::replaceReg(RD.Reg, NewR, MRI); 2563 BT.put(BitTracker::RegisterRef(NewR), RC); 2564 Changed = true; 2565 break; 2566 } 2567 2568 return Changed; 2569 } 2570 2571 bool BitSimplification::simplifyRCmp0(MachineInstr *MI, 2572 BitTracker::RegisterRef RD) { 2573 unsigned Opc = MI->getOpcode(); 2574 if (Opc != Hexagon::A4_rcmpeqi && Opc != Hexagon::A4_rcmpneqi) 2575 return false; 2576 MachineOperand &CmpOp = MI->getOperand(2); 2577 if (!CmpOp.isImm() || CmpOp.getImm() != 0) 2578 return false; 2579 2580 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); 2581 if (FRC != &Hexagon::IntRegsRegClass && FRC != &Hexagon::DoubleRegsRegClass) 2582 return false; 2583 assert(RD.Sub == 0); 2584 2585 MachineBasicBlock &B = *MI->getParent(); 2586 const DebugLoc &DL = MI->getDebugLoc(); 2587 auto At = MI->isPHI() ? B.getFirstNonPHI() 2588 : MachineBasicBlock::iterator(MI); 2589 bool KnownZ = true; 2590 bool KnownNZ = false; 2591 2592 BitTracker::RegisterRef SR = MI->getOperand(1); 2593 if (!BT.has(SR.Reg)) 2594 return false; 2595 const BitTracker::RegisterCell &SC = BT.lookup(SR.Reg); 2596 unsigned F, W; 2597 if (!HBS::getSubregMask(SR, F, W, MRI)) 2598 return false; 2599 2600 for (uint16_t I = F; I != F+W; ++I) { 2601 const BitTracker::BitValue &V = SC[I]; 2602 if (!V.is(0)) 2603 KnownZ = false; 2604 if (V.is(1)) 2605 KnownNZ = true; 2606 } 2607 2608 auto ReplaceWithConst = [&] (int C) { 2609 unsigned NewR = MRI.createVirtualRegister(FRC); 2610 BuildMI(B, At, DL, HII.get(Hexagon::A2_tfrsi), NewR) 2611 .addImm(C); 2612 HBS::replaceReg(RD.Reg, NewR, MRI); 2613 BitTracker::RegisterCell NewRC(W); 2614 for (uint16_t I = 0; I != W; ++I) { 2615 NewRC[I] = BitTracker::BitValue(C & 1); 2616 C = unsigned(C) >> 1; 2617 } 2618 BT.put(BitTracker::RegisterRef(NewR), NewRC); 2619 return true; 2620 }; 2621 2622 auto IsNonZero = [] (const MachineOperand &Op) { 2623 if (Op.isGlobal() || Op.isBlockAddress()) 2624 return true; 2625 if (Op.isImm()) 2626 return Op.getImm() != 0; 2627 if (Op.isCImm()) 2628 return !Op.getCImm()->isZero(); 2629 if (Op.isFPImm()) 2630 return !Op.getFPImm()->isZero(); 2631 return false; 2632 }; 2633 2634 auto IsZero = [] (const MachineOperand &Op) { 2635 if (Op.isGlobal() || Op.isBlockAddress()) 2636 return false; 2637 if (Op.isImm()) 2638 return Op.getImm() == 0; 2639 if (Op.isCImm()) 2640 return Op.getCImm()->isZero(); 2641 if (Op.isFPImm()) 2642 return Op.getFPImm()->isZero(); 2643 return false; 2644 }; 2645 2646 // If the source register is known to be 0 or non-0, the comparison can 2647 // be folded to a load of a constant. 2648 if (KnownZ || KnownNZ) { 2649 assert(KnownZ != KnownNZ && "Register cannot be both 0 and non-0"); 2650 return ReplaceWithConst(KnownZ == (Opc == Hexagon::A4_rcmpeqi)); 2651 } 2652 2653 // Special case: if the compare comes from a C2_muxii, then we know the 2654 // two possible constants that can be the source value. 2655 MachineInstr *InpDef = MRI.getVRegDef(SR.Reg); 2656 if (!InpDef) 2657 return false; 2658 if (SR.Sub == 0 && InpDef->getOpcode() == Hexagon::C2_muxii) { 2659 MachineOperand &Src1 = InpDef->getOperand(2); 2660 MachineOperand &Src2 = InpDef->getOperand(3); 2661 // Check if both are non-zero. 2662 bool KnownNZ1 = IsNonZero(Src1), KnownNZ2 = IsNonZero(Src2); 2663 if (KnownNZ1 && KnownNZ2) 2664 return ReplaceWithConst(Opc == Hexagon::A4_rcmpneqi); 2665 // Check if both are zero. 2666 bool KnownZ1 = IsZero(Src1), KnownZ2 = IsZero(Src2); 2667 if (KnownZ1 && KnownZ2) 2668 return ReplaceWithConst(Opc == Hexagon::A4_rcmpeqi); 2669 2670 // If for both operands we know that they are either 0 or non-0, 2671 // replace the comparison with a C2_muxii, using the same predicate 2672 // register, but with operands substituted with 0/1 accordingly. 2673 if ((KnownZ1 || KnownNZ1) && (KnownZ2 || KnownNZ2)) { 2674 unsigned NewR = MRI.createVirtualRegister(FRC); 2675 BuildMI(B, At, DL, HII.get(Hexagon::C2_muxii), NewR) 2676 .addReg(InpDef->getOperand(1).getReg()) 2677 .addImm(KnownZ1 == (Opc == Hexagon::A4_rcmpeqi)) 2678 .addImm(KnownZ2 == (Opc == Hexagon::A4_rcmpeqi)); 2679 HBS::replaceReg(RD.Reg, NewR, MRI); 2680 // Create a new cell with only the least significant bit unknown. 2681 BitTracker::RegisterCell NewRC(W); 2682 NewRC[0] = BitTracker::BitValue::self(); 2683 NewRC.fill(1, W, BitTracker::BitValue::Zero); 2684 BT.put(BitTracker::RegisterRef(NewR), NewRC); 2685 return true; 2686 } 2687 } 2688 2689 return false; 2690 } 2691 2692 bool BitSimplification::processBlock(MachineBasicBlock &B, 2693 const RegisterSet &AVs) { 2694 if (!BT.reached(&B)) 2695 return false; 2696 bool Changed = false; 2697 RegisterSet AVB = AVs; 2698 RegisterSet Defs; 2699 2700 for (auto I = B.begin(), E = B.end(); I != E; ++I, AVB.insert(Defs)) { 2701 MachineInstr *MI = &*I; 2702 Defs.clear(); 2703 HBS::getInstrDefs(*MI, Defs); 2704 2705 unsigned Opc = MI->getOpcode(); 2706 if (Opc == TargetOpcode::COPY || Opc == TargetOpcode::REG_SEQUENCE) 2707 continue; 2708 2709 if (MI->mayStore()) { 2710 bool T = genStoreUpperHalf(MI); 2711 T = T || genStoreImmediate(MI); 2712 Changed |= T; 2713 continue; 2714 } 2715 2716 if (Defs.count() != 1) 2717 continue; 2718 const MachineOperand &Op0 = MI->getOperand(0); 2719 if (!Op0.isReg() || !Op0.isDef()) 2720 continue; 2721 BitTracker::RegisterRef RD = Op0; 2722 if (!BT.has(RD.Reg)) 2723 continue; 2724 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); 2725 const BitTracker::RegisterCell &RC = BT.lookup(RD.Reg); 2726 2727 if (FRC->getID() == Hexagon::DoubleRegsRegClassID) { 2728 bool T = genPackhl(MI, RD, RC); 2729 T = T || simplifyExtractLow(MI, RD, RC, AVB); 2730 Changed |= T; 2731 continue; 2732 } 2733 2734 if (FRC->getID() == Hexagon::IntRegsRegClassID) { 2735 bool T = genBitSplit(MI, RD, RC, AVB); 2736 T = T || simplifyExtractLow(MI, RD, RC, AVB); 2737 T = T || genExtractHalf(MI, RD, RC); 2738 T = T || genCombineHalf(MI, RD, RC); 2739 T = T || genExtractLow(MI, RD, RC); 2740 T = T || simplifyRCmp0(MI, RD); 2741 Changed |= T; 2742 continue; 2743 } 2744 2745 if (FRC->getID() == Hexagon::PredRegsRegClassID) { 2746 bool T = simplifyTstbit(MI, RD, RC); 2747 Changed |= T; 2748 continue; 2749 } 2750 } 2751 return Changed; 2752 } 2753 2754 bool HexagonBitSimplify::runOnMachineFunction(MachineFunction &MF) { 2755 if (skipFunction(MF.getFunction())) 2756 return false; 2757 2758 auto &HST = MF.getSubtarget<HexagonSubtarget>(); 2759 auto &HRI = *HST.getRegisterInfo(); 2760 auto &HII = *HST.getInstrInfo(); 2761 2762 MDT = &getAnalysis<MachineDominatorTree>(); 2763 MachineRegisterInfo &MRI = MF.getRegInfo(); 2764 bool Changed; 2765 2766 Changed = DeadCodeElimination(MF, *MDT).run(); 2767 2768 const HexagonEvaluator HE(HRI, MRI, HII, MF); 2769 BitTracker BT(HE, MF); 2770 LLVM_DEBUG(BT.trace(true)); 2771 BT.run(); 2772 2773 MachineBasicBlock &Entry = MF.front(); 2774 2775 RegisterSet AIG; // Available registers for IG. 2776 ConstGeneration ImmG(BT, HII, MRI); 2777 Changed |= visitBlock(Entry, ImmG, AIG); 2778 2779 RegisterSet ARE; // Available registers for RIE. 2780 RedundantInstrElimination RIE(BT, HII, HRI, MRI); 2781 bool Ried = visitBlock(Entry, RIE, ARE); 2782 if (Ried) { 2783 Changed = true; 2784 BT.run(); 2785 } 2786 2787 RegisterSet ACG; // Available registers for CG. 2788 CopyGeneration CopyG(BT, HII, HRI, MRI); 2789 Changed |= visitBlock(Entry, CopyG, ACG); 2790 2791 RegisterSet ACP; // Available registers for CP. 2792 CopyPropagation CopyP(HRI, MRI); 2793 Changed |= visitBlock(Entry, CopyP, ACP); 2794 2795 Changed = DeadCodeElimination(MF, *MDT).run() || Changed; 2796 2797 BT.run(); 2798 RegisterSet ABS; // Available registers for BS. 2799 BitSimplification BitS(BT, *MDT, HII, HRI, MRI, MF); 2800 Changed |= visitBlock(Entry, BitS, ABS); 2801 2802 Changed = DeadCodeElimination(MF, *MDT).run() || Changed; 2803 2804 if (Changed) { 2805 for (auto &B : MF) 2806 for (auto &I : B) 2807 I.clearKillInfo(); 2808 DeadCodeElimination(MF, *MDT).run(); 2809 } 2810 return Changed; 2811 } 2812 2813 // Recognize loops where the code at the end of the loop matches the code 2814 // before the entry of the loop, and the matching code is such that is can 2815 // be simplified. This pass relies on the bit simplification above and only 2816 // prepares code in a way that can be handled by the bit simplifcation. 2817 // 2818 // This is the motivating testcase (and explanation): 2819 // 2820 // { 2821 // loop0(.LBB0_2, r1) // %for.body.preheader 2822 // r5:4 = memd(r0++#8) 2823 // } 2824 // { 2825 // r3 = lsr(r4, #16) 2826 // r7:6 = combine(r5, r5) 2827 // } 2828 // { 2829 // r3 = insert(r5, #16, #16) 2830 // r7:6 = vlsrw(r7:6, #16) 2831 // } 2832 // .LBB0_2: 2833 // { 2834 // memh(r2+#4) = r5 2835 // memh(r2+#6) = r6 # R6 is really R5.H 2836 // } 2837 // { 2838 // r2 = add(r2, #8) 2839 // memh(r2+#0) = r4 2840 // memh(r2+#2) = r3 # R3 is really R4.H 2841 // } 2842 // { 2843 // r5:4 = memd(r0++#8) 2844 // } 2845 // { # "Shuffling" code that sets up R3 and R6 2846 // r3 = lsr(r4, #16) # so that their halves can be stored in the 2847 // r7:6 = combine(r5, r5) # next iteration. This could be folded into 2848 // } # the stores if the code was at the beginning 2849 // { # of the loop iteration. Since the same code 2850 // r3 = insert(r5, #16, #16) # precedes the loop, it can actually be moved 2851 // r7:6 = vlsrw(r7:6, #16) # there. 2852 // }:endloop0 2853 // 2854 // 2855 // The outcome: 2856 // 2857 // { 2858 // loop0(.LBB0_2, r1) 2859 // r5:4 = memd(r0++#8) 2860 // } 2861 // .LBB0_2: 2862 // { 2863 // memh(r2+#4) = r5 2864 // memh(r2+#6) = r5.h 2865 // } 2866 // { 2867 // r2 = add(r2, #8) 2868 // memh(r2+#0) = r4 2869 // memh(r2+#2) = r4.h 2870 // } 2871 // { 2872 // r5:4 = memd(r0++#8) 2873 // }:endloop0 2874 2875 namespace llvm { 2876 2877 FunctionPass *createHexagonLoopRescheduling(); 2878 void initializeHexagonLoopReschedulingPass(PassRegistry&); 2879 2880 } // end namespace llvm 2881 2882 namespace { 2883 2884 class HexagonLoopRescheduling : public MachineFunctionPass { 2885 public: 2886 static char ID; 2887 2888 HexagonLoopRescheduling() : MachineFunctionPass(ID) { 2889 initializeHexagonLoopReschedulingPass(*PassRegistry::getPassRegistry()); 2890 } 2891 2892 bool runOnMachineFunction(MachineFunction &MF) override; 2893 2894 private: 2895 const HexagonInstrInfo *HII = nullptr; 2896 const HexagonRegisterInfo *HRI = nullptr; 2897 MachineRegisterInfo *MRI = nullptr; 2898 BitTracker *BTP = nullptr; 2899 2900 struct LoopCand { 2901 LoopCand(MachineBasicBlock *lb, MachineBasicBlock *pb, 2902 MachineBasicBlock *eb) : LB(lb), PB(pb), EB(eb) {} 2903 2904 MachineBasicBlock *LB, *PB, *EB; 2905 }; 2906 using InstrList = std::vector<MachineInstr *>; 2907 struct InstrGroup { 2908 BitTracker::RegisterRef Inp, Out; 2909 InstrList Ins; 2910 }; 2911 struct PhiInfo { 2912 PhiInfo(MachineInstr &P, MachineBasicBlock &B); 2913 2914 unsigned DefR; 2915 BitTracker::RegisterRef LR, PR; // Loop Register, Preheader Register 2916 MachineBasicBlock *LB, *PB; // Loop Block, Preheader Block 2917 }; 2918 2919 static unsigned getDefReg(const MachineInstr *MI); 2920 bool isConst(unsigned Reg) const; 2921 bool isBitShuffle(const MachineInstr *MI, unsigned DefR) const; 2922 bool isStoreInput(const MachineInstr *MI, unsigned DefR) const; 2923 bool isShuffleOf(unsigned OutR, unsigned InpR) const; 2924 bool isSameShuffle(unsigned OutR1, unsigned InpR1, unsigned OutR2, 2925 unsigned &InpR2) const; 2926 void moveGroup(InstrGroup &G, MachineBasicBlock &LB, MachineBasicBlock &PB, 2927 MachineBasicBlock::iterator At, unsigned OldPhiR, unsigned NewPredR); 2928 bool processLoop(LoopCand &C); 2929 }; 2930 2931 } // end anonymous namespace 2932 2933 char HexagonLoopRescheduling::ID = 0; 2934 2935 INITIALIZE_PASS(HexagonLoopRescheduling, "hexagon-loop-resched", 2936 "Hexagon Loop Rescheduling", false, false) 2937 2938 HexagonLoopRescheduling::PhiInfo::PhiInfo(MachineInstr &P, 2939 MachineBasicBlock &B) { 2940 DefR = HexagonLoopRescheduling::getDefReg(&P); 2941 LB = &B; 2942 PB = nullptr; 2943 for (unsigned i = 1, n = P.getNumOperands(); i < n; i += 2) { 2944 const MachineOperand &OpB = P.getOperand(i+1); 2945 if (OpB.getMBB() == &B) { 2946 LR = P.getOperand(i); 2947 continue; 2948 } 2949 PB = OpB.getMBB(); 2950 PR = P.getOperand(i); 2951 } 2952 } 2953 2954 unsigned HexagonLoopRescheduling::getDefReg(const MachineInstr *MI) { 2955 RegisterSet Defs; 2956 HBS::getInstrDefs(*MI, Defs); 2957 if (Defs.count() != 1) 2958 return 0; 2959 return Defs.find_first(); 2960 } 2961 2962 bool HexagonLoopRescheduling::isConst(unsigned Reg) const { 2963 if (!BTP->has(Reg)) 2964 return false; 2965 const BitTracker::RegisterCell &RC = BTP->lookup(Reg); 2966 for (unsigned i = 0, w = RC.width(); i < w; ++i) { 2967 const BitTracker::BitValue &V = RC[i]; 2968 if (!V.is(0) && !V.is(1)) 2969 return false; 2970 } 2971 return true; 2972 } 2973 2974 bool HexagonLoopRescheduling::isBitShuffle(const MachineInstr *MI, 2975 unsigned DefR) const { 2976 unsigned Opc = MI->getOpcode(); 2977 switch (Opc) { 2978 case TargetOpcode::COPY: 2979 case Hexagon::S2_lsr_i_r: 2980 case Hexagon::S2_asr_i_r: 2981 case Hexagon::S2_asl_i_r: 2982 case Hexagon::S2_lsr_i_p: 2983 case Hexagon::S2_asr_i_p: 2984 case Hexagon::S2_asl_i_p: 2985 case Hexagon::S2_insert: 2986 case Hexagon::A2_or: 2987 case Hexagon::A2_orp: 2988 case Hexagon::A2_and: 2989 case Hexagon::A2_andp: 2990 case Hexagon::A2_combinew: 2991 case Hexagon::A4_combineri: 2992 case Hexagon::A4_combineir: 2993 case Hexagon::A2_combineii: 2994 case Hexagon::A4_combineii: 2995 case Hexagon::A2_combine_ll: 2996 case Hexagon::A2_combine_lh: 2997 case Hexagon::A2_combine_hl: 2998 case Hexagon::A2_combine_hh: 2999 return true; 3000 } 3001 return false; 3002 } 3003 3004 bool HexagonLoopRescheduling::isStoreInput(const MachineInstr *MI, 3005 unsigned InpR) const { 3006 for (unsigned i = 0, n = MI->getNumOperands(); i < n; ++i) { 3007 const MachineOperand &Op = MI->getOperand(i); 3008 if (!Op.isReg()) 3009 continue; 3010 if (Op.getReg() == InpR) 3011 return i == n-1; 3012 } 3013 return false; 3014 } 3015 3016 bool HexagonLoopRescheduling::isShuffleOf(unsigned OutR, unsigned InpR) const { 3017 if (!BTP->has(OutR) || !BTP->has(InpR)) 3018 return false; 3019 const BitTracker::RegisterCell &OutC = BTP->lookup(OutR); 3020 for (unsigned i = 0, w = OutC.width(); i < w; ++i) { 3021 const BitTracker::BitValue &V = OutC[i]; 3022 if (V.Type != BitTracker::BitValue::Ref) 3023 continue; 3024 if (V.RefI.Reg != InpR) 3025 return false; 3026 } 3027 return true; 3028 } 3029 3030 bool HexagonLoopRescheduling::isSameShuffle(unsigned OutR1, unsigned InpR1, 3031 unsigned OutR2, unsigned &InpR2) const { 3032 if (!BTP->has(OutR1) || !BTP->has(InpR1) || !BTP->has(OutR2)) 3033 return false; 3034 const BitTracker::RegisterCell &OutC1 = BTP->lookup(OutR1); 3035 const BitTracker::RegisterCell &OutC2 = BTP->lookup(OutR2); 3036 unsigned W = OutC1.width(); 3037 unsigned MatchR = 0; 3038 if (W != OutC2.width()) 3039 return false; 3040 for (unsigned i = 0; i < W; ++i) { 3041 const BitTracker::BitValue &V1 = OutC1[i], &V2 = OutC2[i]; 3042 if (V1.Type != V2.Type || V1.Type == BitTracker::BitValue::One) 3043 return false; 3044 if (V1.Type != BitTracker::BitValue::Ref) 3045 continue; 3046 if (V1.RefI.Pos != V2.RefI.Pos) 3047 return false; 3048 if (V1.RefI.Reg != InpR1) 3049 return false; 3050 if (V2.RefI.Reg == 0 || V2.RefI.Reg == OutR2) 3051 return false; 3052 if (!MatchR) 3053 MatchR = V2.RefI.Reg; 3054 else if (V2.RefI.Reg != MatchR) 3055 return false; 3056 } 3057 InpR2 = MatchR; 3058 return true; 3059 } 3060 3061 void HexagonLoopRescheduling::moveGroup(InstrGroup &G, MachineBasicBlock &LB, 3062 MachineBasicBlock &PB, MachineBasicBlock::iterator At, unsigned OldPhiR, 3063 unsigned NewPredR) { 3064 DenseMap<unsigned,unsigned> RegMap; 3065 3066 const TargetRegisterClass *PhiRC = MRI->getRegClass(NewPredR); 3067 unsigned PhiR = MRI->createVirtualRegister(PhiRC); 3068 BuildMI(LB, At, At->getDebugLoc(), HII->get(TargetOpcode::PHI), PhiR) 3069 .addReg(NewPredR) 3070 .addMBB(&PB) 3071 .addReg(G.Inp.Reg) 3072 .addMBB(&LB); 3073 RegMap.insert(std::make_pair(G.Inp.Reg, PhiR)); 3074 3075 for (unsigned i = G.Ins.size(); i > 0; --i) { 3076 const MachineInstr *SI = G.Ins[i-1]; 3077 unsigned DR = getDefReg(SI); 3078 const TargetRegisterClass *RC = MRI->getRegClass(DR); 3079 unsigned NewDR = MRI->createVirtualRegister(RC); 3080 DebugLoc DL = SI->getDebugLoc(); 3081 3082 auto MIB = BuildMI(LB, At, DL, HII->get(SI->getOpcode()), NewDR); 3083 for (unsigned j = 0, m = SI->getNumOperands(); j < m; ++j) { 3084 const MachineOperand &Op = SI->getOperand(j); 3085 if (!Op.isReg()) { 3086 MIB.add(Op); 3087 continue; 3088 } 3089 if (!Op.isUse()) 3090 continue; 3091 unsigned UseR = RegMap[Op.getReg()]; 3092 MIB.addReg(UseR, 0, Op.getSubReg()); 3093 } 3094 RegMap.insert(std::make_pair(DR, NewDR)); 3095 } 3096 3097 HBS::replaceReg(OldPhiR, RegMap[G.Out.Reg], *MRI); 3098 } 3099 3100 bool HexagonLoopRescheduling::processLoop(LoopCand &C) { 3101 LLVM_DEBUG(dbgs() << "Processing loop in " << printMBBReference(*C.LB) 3102 << "\n"); 3103 std::vector<PhiInfo> Phis; 3104 for (auto &I : *C.LB) { 3105 if (!I.isPHI()) 3106 break; 3107 unsigned PR = getDefReg(&I); 3108 if (isConst(PR)) 3109 continue; 3110 bool BadUse = false, GoodUse = false; 3111 for (auto UI = MRI->use_begin(PR), UE = MRI->use_end(); UI != UE; ++UI) { 3112 MachineInstr *UseI = UI->getParent(); 3113 if (UseI->getParent() != C.LB) { 3114 BadUse = true; 3115 break; 3116 } 3117 if (isBitShuffle(UseI, PR) || isStoreInput(UseI, PR)) 3118 GoodUse = true; 3119 } 3120 if (BadUse || !GoodUse) 3121 continue; 3122 3123 Phis.push_back(PhiInfo(I, *C.LB)); 3124 } 3125 3126 LLVM_DEBUG({ 3127 dbgs() << "Phis: {"; 3128 for (auto &I : Phis) { 3129 dbgs() << ' ' << printReg(I.DefR, HRI) << "=phi(" 3130 << printReg(I.PR.Reg, HRI, I.PR.Sub) << ":b" << I.PB->getNumber() 3131 << ',' << printReg(I.LR.Reg, HRI, I.LR.Sub) << ":b" 3132 << I.LB->getNumber() << ')'; 3133 } 3134 dbgs() << " }\n"; 3135 }); 3136 3137 if (Phis.empty()) 3138 return false; 3139 3140 bool Changed = false; 3141 InstrList ShufIns; 3142 3143 // Go backwards in the block: for each bit shuffling instruction, check 3144 // if that instruction could potentially be moved to the front of the loop: 3145 // the output of the loop cannot be used in a non-shuffling instruction 3146 // in this loop. 3147 for (auto I = C.LB->rbegin(), E = C.LB->rend(); I != E; ++I) { 3148 if (I->isTerminator()) 3149 continue; 3150 if (I->isPHI()) 3151 break; 3152 3153 RegisterSet Defs; 3154 HBS::getInstrDefs(*I, Defs); 3155 if (Defs.count() != 1) 3156 continue; 3157 unsigned DefR = Defs.find_first(); 3158 if (!TargetRegisterInfo::isVirtualRegister(DefR)) 3159 continue; 3160 if (!isBitShuffle(&*I, DefR)) 3161 continue; 3162 3163 bool BadUse = false; 3164 for (auto UI = MRI->use_begin(DefR), UE = MRI->use_end(); UI != UE; ++UI) { 3165 MachineInstr *UseI = UI->getParent(); 3166 if (UseI->getParent() == C.LB) { 3167 if (UseI->isPHI()) { 3168 // If the use is in a phi node in this loop, then it should be 3169 // the value corresponding to the back edge. 3170 unsigned Idx = UI.getOperandNo(); 3171 if (UseI->getOperand(Idx+1).getMBB() != C.LB) 3172 BadUse = true; 3173 } else { 3174 auto F = find(ShufIns, UseI); 3175 if (F == ShufIns.end()) 3176 BadUse = true; 3177 } 3178 } else { 3179 // There is a use outside of the loop, but there is no epilog block 3180 // suitable for a copy-out. 3181 if (C.EB == nullptr) 3182 BadUse = true; 3183 } 3184 if (BadUse) 3185 break; 3186 } 3187 3188 if (BadUse) 3189 continue; 3190 ShufIns.push_back(&*I); 3191 } 3192 3193 // Partition the list of shuffling instructions into instruction groups, 3194 // where each group has to be moved as a whole (i.e. a group is a chain of 3195 // dependent instructions). A group produces a single live output register, 3196 // which is meant to be the input of the loop phi node (although this is 3197 // not checked here yet). It also uses a single register as its input, 3198 // which is some value produced in the loop body. After moving the group 3199 // to the beginning of the loop, that input register would need to be 3200 // the loop-carried register (through a phi node) instead of the (currently 3201 // loop-carried) output register. 3202 using InstrGroupList = std::vector<InstrGroup>; 3203 InstrGroupList Groups; 3204 3205 for (unsigned i = 0, n = ShufIns.size(); i < n; ++i) { 3206 MachineInstr *SI = ShufIns[i]; 3207 if (SI == nullptr) 3208 continue; 3209 3210 InstrGroup G; 3211 G.Ins.push_back(SI); 3212 G.Out.Reg = getDefReg(SI); 3213 RegisterSet Inputs; 3214 HBS::getInstrUses(*SI, Inputs); 3215 3216 for (unsigned j = i+1; j < n; ++j) { 3217 MachineInstr *MI = ShufIns[j]; 3218 if (MI == nullptr) 3219 continue; 3220 RegisterSet Defs; 3221 HBS::getInstrDefs(*MI, Defs); 3222 // If this instruction does not define any pending inputs, skip it. 3223 if (!Defs.intersects(Inputs)) 3224 continue; 3225 // Otherwise, add it to the current group and remove the inputs that 3226 // are defined by MI. 3227 G.Ins.push_back(MI); 3228 Inputs.remove(Defs); 3229 // Then add all registers used by MI. 3230 HBS::getInstrUses(*MI, Inputs); 3231 ShufIns[j] = nullptr; 3232 } 3233 3234 // Only add a group if it requires at most one register. 3235 if (Inputs.count() > 1) 3236 continue; 3237 auto LoopInpEq = [G] (const PhiInfo &P) -> bool { 3238 return G.Out.Reg == P.LR.Reg; 3239 }; 3240 if (llvm::find_if(Phis, LoopInpEq) == Phis.end()) 3241 continue; 3242 3243 G.Inp.Reg = Inputs.find_first(); 3244 Groups.push_back(G); 3245 } 3246 3247 LLVM_DEBUG({ 3248 for (unsigned i = 0, n = Groups.size(); i < n; ++i) { 3249 InstrGroup &G = Groups[i]; 3250 dbgs() << "Group[" << i << "] inp: " 3251 << printReg(G.Inp.Reg, HRI, G.Inp.Sub) 3252 << " out: " << printReg(G.Out.Reg, HRI, G.Out.Sub) << "\n"; 3253 for (unsigned j = 0, m = G.Ins.size(); j < m; ++j) 3254 dbgs() << " " << *G.Ins[j]; 3255 } 3256 }); 3257 3258 for (unsigned i = 0, n = Groups.size(); i < n; ++i) { 3259 InstrGroup &G = Groups[i]; 3260 if (!isShuffleOf(G.Out.Reg, G.Inp.Reg)) 3261 continue; 3262 auto LoopInpEq = [G] (const PhiInfo &P) -> bool { 3263 return G.Out.Reg == P.LR.Reg; 3264 }; 3265 auto F = llvm::find_if(Phis, LoopInpEq); 3266 if (F == Phis.end()) 3267 continue; 3268 unsigned PrehR = 0; 3269 if (!isSameShuffle(G.Out.Reg, G.Inp.Reg, F->PR.Reg, PrehR)) { 3270 const MachineInstr *DefPrehR = MRI->getVRegDef(F->PR.Reg); 3271 unsigned Opc = DefPrehR->getOpcode(); 3272 if (Opc != Hexagon::A2_tfrsi && Opc != Hexagon::A2_tfrpi) 3273 continue; 3274 if (!DefPrehR->getOperand(1).isImm()) 3275 continue; 3276 if (DefPrehR->getOperand(1).getImm() != 0) 3277 continue; 3278 const TargetRegisterClass *RC = MRI->getRegClass(G.Inp.Reg); 3279 if (RC != MRI->getRegClass(F->PR.Reg)) { 3280 PrehR = MRI->createVirtualRegister(RC); 3281 unsigned TfrI = (RC == &Hexagon::IntRegsRegClass) ? Hexagon::A2_tfrsi 3282 : Hexagon::A2_tfrpi; 3283 auto T = C.PB->getFirstTerminator(); 3284 DebugLoc DL = (T != C.PB->end()) ? T->getDebugLoc() : DebugLoc(); 3285 BuildMI(*C.PB, T, DL, HII->get(TfrI), PrehR) 3286 .addImm(0); 3287 } else { 3288 PrehR = F->PR.Reg; 3289 } 3290 } 3291 // isSameShuffle could match with PrehR being of a wider class than 3292 // G.Inp.Reg, for example if G shuffles the low 32 bits of its input, 3293 // it would match for the input being a 32-bit register, and PrehR 3294 // being a 64-bit register (where the low 32 bits match). This could 3295 // be handled, but for now skip these cases. 3296 if (MRI->getRegClass(PrehR) != MRI->getRegClass(G.Inp.Reg)) 3297 continue; 3298 moveGroup(G, *F->LB, *F->PB, F->LB->getFirstNonPHI(), F->DefR, PrehR); 3299 Changed = true; 3300 } 3301 3302 return Changed; 3303 } 3304 3305 bool HexagonLoopRescheduling::runOnMachineFunction(MachineFunction &MF) { 3306 if (skipFunction(MF.getFunction())) 3307 return false; 3308 3309 auto &HST = MF.getSubtarget<HexagonSubtarget>(); 3310 HII = HST.getInstrInfo(); 3311 HRI = HST.getRegisterInfo(); 3312 MRI = &MF.getRegInfo(); 3313 const HexagonEvaluator HE(*HRI, *MRI, *HII, MF); 3314 BitTracker BT(HE, MF); 3315 LLVM_DEBUG(BT.trace(true)); 3316 BT.run(); 3317 BTP = &BT; 3318 3319 std::vector<LoopCand> Cand; 3320 3321 for (auto &B : MF) { 3322 if (B.pred_size() != 2 || B.succ_size() != 2) 3323 continue; 3324 MachineBasicBlock *PB = nullptr; 3325 bool IsLoop = false; 3326 for (auto PI = B.pred_begin(), PE = B.pred_end(); PI != PE; ++PI) { 3327 if (*PI != &B) 3328 PB = *PI; 3329 else 3330 IsLoop = true; 3331 } 3332 if (!IsLoop) 3333 continue; 3334 3335 MachineBasicBlock *EB = nullptr; 3336 for (auto SI = B.succ_begin(), SE = B.succ_end(); SI != SE; ++SI) { 3337 if (*SI == &B) 3338 continue; 3339 // Set EP to the epilog block, if it has only 1 predecessor (i.e. the 3340 // edge from B to EP is non-critical. 3341 if ((*SI)->pred_size() == 1) 3342 EB = *SI; 3343 break; 3344 } 3345 3346 Cand.push_back(LoopCand(&B, PB, EB)); 3347 } 3348 3349 bool Changed = false; 3350 for (auto &C : Cand) 3351 Changed |= processLoop(C); 3352 3353 return Changed; 3354 } 3355 3356 //===----------------------------------------------------------------------===// 3357 // Public Constructor Functions 3358 //===----------------------------------------------------------------------===// 3359 3360 FunctionPass *llvm::createHexagonLoopRescheduling() { 3361 return new HexagonLoopRescheduling(); 3362 } 3363 3364 FunctionPass *llvm::createHexagonBitSimplify() { 3365 return new HexagonBitSimplify(); 3366 } 3367