1 //===- HexagonBitSimplify.cpp ---------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "BitTracker.h" 10 #include "HexagonBitTracker.h" 11 #include "HexagonInstrInfo.h" 12 #include "HexagonRegisterInfo.h" 13 #include "HexagonSubtarget.h" 14 #include "llvm/ADT/BitVector.h" 15 #include "llvm/ADT/DenseMap.h" 16 #include "llvm/ADT/GraphTraits.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SmallVector.h" 19 #include "llvm/ADT/StringRef.h" 20 #include "llvm/CodeGen/MachineBasicBlock.h" 21 #include "llvm/CodeGen/MachineDominators.h" 22 #include "llvm/CodeGen/MachineFunction.h" 23 #include "llvm/CodeGen/MachineFunctionPass.h" 24 #include "llvm/CodeGen/MachineInstr.h" 25 #include "llvm/CodeGen/MachineInstrBuilder.h" 26 #include "llvm/CodeGen/MachineOperand.h" 27 #include "llvm/CodeGen/MachineRegisterInfo.h" 28 #include "llvm/CodeGen/TargetRegisterInfo.h" 29 #include "llvm/IR/DebugLoc.h" 30 #include "llvm/InitializePasses.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 Register::virtReg2Index(v); 152 } 153 154 static inline unsigned x2v(unsigned x) { 155 return Register::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(Register OldR, Register NewR, 210 MachineRegisterInfo &MRI); 211 static bool getSubregMask(const BitTracker::RegisterRef &RR, 212 unsigned &Begin, unsigned &Width, MachineRegisterInfo &MRI); 213 static bool replaceRegWithSub(Register OldR, Register NewR, unsigned NewSR, 214 MachineRegisterInfo &MRI); 215 static bool replaceSubWithSub(Register OldR, unsigned OldSR, Register NewR, 216 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 Register R = Op.getReg(); 295 if (!R.isVirtual()) 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 Register R = Op.getReg(); 307 if (!R.isVirtual()) 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(Register OldR, Register NewR, 356 MachineRegisterInfo &MRI) { 357 if (!OldR.isVirtual() || !NewR.isVirtual()) 358 return false; 359 auto Begin = MRI.use_begin(OldR), End = MRI.use_end(); 360 decltype(End) NextI; 361 for (auto I = Begin; I != End; I = NextI) { 362 NextI = std::next(I); 363 I->setReg(NewR); 364 } 365 return Begin != End; 366 } 367 368 bool HexagonBitSimplify::replaceRegWithSub(Register OldR, Register NewR, 369 unsigned NewSR, 370 MachineRegisterInfo &MRI) { 371 if (!OldR.isVirtual() || !NewR.isVirtual()) 372 return false; 373 if (hasTiedUse(OldR, MRI, NewSR)) 374 return false; 375 auto Begin = MRI.use_begin(OldR), End = MRI.use_end(); 376 decltype(End) NextI; 377 for (auto I = Begin; I != End; I = NextI) { 378 NextI = std::next(I); 379 I->setReg(NewR); 380 I->setSubReg(NewSR); 381 } 382 return Begin != End; 383 } 384 385 bool HexagonBitSimplify::replaceSubWithSub(Register OldR, unsigned OldSR, 386 Register NewR, unsigned NewSR, 387 MachineRegisterInfo &MRI) { 388 if (!OldR.isVirtual() || !NewR.isVirtual()) 389 return false; 390 if (OldSR != NewSR && hasTiedUse(OldR, MRI, NewSR)) 391 return false; 392 auto Begin = MRI.use_begin(OldR), End = MRI.use_end(); 393 decltype(End) NextI; 394 for (auto I = Begin; I != End; I = NextI) { 395 NextI = std::next(I); 396 if (I->getSubReg() != OldSR) 397 continue; 398 I->setReg(NewR); 399 I->setSubReg(NewSR); 400 } 401 return Begin != End; 402 } 403 404 // For a register ref (pair Reg:Sub), set Begin to the position of the LSB 405 // of Sub in Reg, and set Width to the size of Sub in bits. Return true, 406 // if this succeeded, otherwise return false. 407 bool HexagonBitSimplify::getSubregMask(const BitTracker::RegisterRef &RR, 408 unsigned &Begin, unsigned &Width, MachineRegisterInfo &MRI) { 409 const TargetRegisterClass *RC = MRI.getRegClass(RR.Reg); 410 if (RR.Sub == 0) { 411 Begin = 0; 412 Width = MRI.getTargetRegisterInfo()->getRegSizeInBits(*RC); 413 return true; 414 } 415 416 Begin = 0; 417 418 switch (RC->getID()) { 419 case Hexagon::DoubleRegsRegClassID: 420 case Hexagon::HvxWRRegClassID: 421 Width = MRI.getTargetRegisterInfo()->getRegSizeInBits(*RC) / 2; 422 if (RR.Sub == Hexagon::isub_hi || RR.Sub == Hexagon::vsub_hi) 423 Begin = Width; 424 break; 425 default: 426 return false; 427 } 428 return true; 429 } 430 431 432 // For a REG_SEQUENCE, set SL to the low subregister and SH to the high 433 // subregister. 434 bool HexagonBitSimplify::parseRegSequence(const MachineInstr &I, 435 BitTracker::RegisterRef &SL, BitTracker::RegisterRef &SH, 436 const MachineRegisterInfo &MRI) { 437 assert(I.getOpcode() == TargetOpcode::REG_SEQUENCE); 438 unsigned Sub1 = I.getOperand(2).getImm(), Sub2 = I.getOperand(4).getImm(); 439 auto &DstRC = *MRI.getRegClass(I.getOperand(0).getReg()); 440 auto &HRI = static_cast<const HexagonRegisterInfo&>( 441 *MRI.getTargetRegisterInfo()); 442 unsigned SubLo = HRI.getHexagonSubRegIndex(DstRC, Hexagon::ps_sub_lo); 443 unsigned SubHi = HRI.getHexagonSubRegIndex(DstRC, Hexagon::ps_sub_hi); 444 assert((Sub1 == SubLo && Sub2 == SubHi) || (Sub1 == SubHi && Sub2 == SubLo)); 445 if (Sub1 == SubLo && Sub2 == SubHi) { 446 SL = I.getOperand(1); 447 SH = I.getOperand(3); 448 return true; 449 } 450 if (Sub1 == SubHi && Sub2 == SubLo) { 451 SH = I.getOperand(1); 452 SL = I.getOperand(3); 453 return true; 454 } 455 return false; 456 } 457 458 // All stores (except 64-bit stores) take a 32-bit register as the source 459 // of the value to be stored. If the instruction stores into a location 460 // that is shorter than 32 bits, some bits of the source register are not 461 // used. For each store instruction, calculate the set of used bits in 462 // the source register, and set appropriate bits in Bits. Return true if 463 // the bits are calculated, false otherwise. 464 bool HexagonBitSimplify::getUsedBitsInStore(unsigned Opc, BitVector &Bits, 465 uint16_t Begin) { 466 using namespace Hexagon; 467 468 switch (Opc) { 469 // Store byte 470 case S2_storerb_io: // memb(Rs32+#s11:0)=Rt32 471 case S2_storerbnew_io: // memb(Rs32+#s11:0)=Nt8.new 472 case S2_pstorerbt_io: // if (Pv4) memb(Rs32+#u6:0)=Rt32 473 case S2_pstorerbf_io: // if (!Pv4) memb(Rs32+#u6:0)=Rt32 474 case S4_pstorerbtnew_io: // if (Pv4.new) memb(Rs32+#u6:0)=Rt32 475 case S4_pstorerbfnew_io: // if (!Pv4.new) memb(Rs32+#u6:0)=Rt32 476 case S2_pstorerbnewt_io: // if (Pv4) memb(Rs32+#u6:0)=Nt8.new 477 case S2_pstorerbnewf_io: // if (!Pv4) memb(Rs32+#u6:0)=Nt8.new 478 case S4_pstorerbnewtnew_io: // if (Pv4.new) memb(Rs32+#u6:0)=Nt8.new 479 case S4_pstorerbnewfnew_io: // if (!Pv4.new) memb(Rs32+#u6:0)=Nt8.new 480 case S2_storerb_pi: // memb(Rx32++#s4:0)=Rt32 481 case S2_storerbnew_pi: // memb(Rx32++#s4:0)=Nt8.new 482 case S2_pstorerbt_pi: // if (Pv4) memb(Rx32++#s4:0)=Rt32 483 case S2_pstorerbf_pi: // if (!Pv4) memb(Rx32++#s4:0)=Rt32 484 case S2_pstorerbtnew_pi: // if (Pv4.new) memb(Rx32++#s4:0)=Rt32 485 case S2_pstorerbfnew_pi: // if (!Pv4.new) memb(Rx32++#s4:0)=Rt32 486 case S2_pstorerbnewt_pi: // if (Pv4) memb(Rx32++#s4:0)=Nt8.new 487 case S2_pstorerbnewf_pi: // if (!Pv4) memb(Rx32++#s4:0)=Nt8.new 488 case S2_pstorerbnewtnew_pi: // if (Pv4.new) memb(Rx32++#s4:0)=Nt8.new 489 case S2_pstorerbnewfnew_pi: // if (!Pv4.new) memb(Rx32++#s4:0)=Nt8.new 490 case S4_storerb_ap: // memb(Re32=#U6)=Rt32 491 case S4_storerbnew_ap: // memb(Re32=#U6)=Nt8.new 492 case S2_storerb_pr: // memb(Rx32++Mu2)=Rt32 493 case S2_storerbnew_pr: // memb(Rx32++Mu2)=Nt8.new 494 case S4_storerb_ur: // memb(Ru32<<#u2+#U6)=Rt32 495 case S4_storerbnew_ur: // memb(Ru32<<#u2+#U6)=Nt8.new 496 case S2_storerb_pbr: // memb(Rx32++Mu2:brev)=Rt32 497 case S2_storerbnew_pbr: // memb(Rx32++Mu2:brev)=Nt8.new 498 case S2_storerb_pci: // memb(Rx32++#s4:0:circ(Mu2))=Rt32 499 case S2_storerbnew_pci: // memb(Rx32++#s4:0:circ(Mu2))=Nt8.new 500 case S2_storerb_pcr: // memb(Rx32++I:circ(Mu2))=Rt32 501 case S2_storerbnew_pcr: // memb(Rx32++I:circ(Mu2))=Nt8.new 502 case S4_storerb_rr: // memb(Rs32+Ru32<<#u2)=Rt32 503 case S4_storerbnew_rr: // memb(Rs32+Ru32<<#u2)=Nt8.new 504 case S4_pstorerbt_rr: // if (Pv4) memb(Rs32+Ru32<<#u2)=Rt32 505 case S4_pstorerbf_rr: // if (!Pv4) memb(Rs32+Ru32<<#u2)=Rt32 506 case S4_pstorerbtnew_rr: // if (Pv4.new) memb(Rs32+Ru32<<#u2)=Rt32 507 case S4_pstorerbfnew_rr: // if (!Pv4.new) memb(Rs32+Ru32<<#u2)=Rt32 508 case S4_pstorerbnewt_rr: // if (Pv4) memb(Rs32+Ru32<<#u2)=Nt8.new 509 case S4_pstorerbnewf_rr: // if (!Pv4) memb(Rs32+Ru32<<#u2)=Nt8.new 510 case S4_pstorerbnewtnew_rr: // if (Pv4.new) memb(Rs32+Ru32<<#u2)=Nt8.new 511 case S4_pstorerbnewfnew_rr: // if (!Pv4.new) memb(Rs32+Ru32<<#u2)=Nt8.new 512 case S2_storerbgp: // memb(gp+#u16:0)=Rt32 513 case S2_storerbnewgp: // memb(gp+#u16:0)=Nt8.new 514 case S4_pstorerbt_abs: // if (Pv4) memb(#u6)=Rt32 515 case S4_pstorerbf_abs: // if (!Pv4) memb(#u6)=Rt32 516 case S4_pstorerbtnew_abs: // if (Pv4.new) memb(#u6)=Rt32 517 case S4_pstorerbfnew_abs: // if (!Pv4.new) memb(#u6)=Rt32 518 case S4_pstorerbnewt_abs: // if (Pv4) memb(#u6)=Nt8.new 519 case S4_pstorerbnewf_abs: // if (!Pv4) memb(#u6)=Nt8.new 520 case S4_pstorerbnewtnew_abs: // if (Pv4.new) memb(#u6)=Nt8.new 521 case S4_pstorerbnewfnew_abs: // if (!Pv4.new) memb(#u6)=Nt8.new 522 Bits.set(Begin, Begin+8); 523 return true; 524 525 // Store low half 526 case S2_storerh_io: // memh(Rs32+#s11:1)=Rt32 527 case S2_storerhnew_io: // memh(Rs32+#s11:1)=Nt8.new 528 case S2_pstorerht_io: // if (Pv4) memh(Rs32+#u6:1)=Rt32 529 case S2_pstorerhf_io: // if (!Pv4) memh(Rs32+#u6:1)=Rt32 530 case S4_pstorerhtnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Rt32 531 case S4_pstorerhfnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Rt32 532 case S2_pstorerhnewt_io: // if (Pv4) memh(Rs32+#u6:1)=Nt8.new 533 case S2_pstorerhnewf_io: // if (!Pv4) memh(Rs32+#u6:1)=Nt8.new 534 case S4_pstorerhnewtnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Nt8.new 535 case S4_pstorerhnewfnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Nt8.new 536 case S2_storerh_pi: // memh(Rx32++#s4:1)=Rt32 537 case S2_storerhnew_pi: // memh(Rx32++#s4:1)=Nt8.new 538 case S2_pstorerht_pi: // if (Pv4) memh(Rx32++#s4:1)=Rt32 539 case S2_pstorerhf_pi: // if (!Pv4) memh(Rx32++#s4:1)=Rt32 540 case S2_pstorerhtnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Rt32 541 case S2_pstorerhfnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Rt32 542 case S2_pstorerhnewt_pi: // if (Pv4) memh(Rx32++#s4:1)=Nt8.new 543 case S2_pstorerhnewf_pi: // if (!Pv4) memh(Rx32++#s4:1)=Nt8.new 544 case S2_pstorerhnewtnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Nt8.new 545 case S2_pstorerhnewfnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Nt8.new 546 case S4_storerh_ap: // memh(Re32=#U6)=Rt32 547 case S4_storerhnew_ap: // memh(Re32=#U6)=Nt8.new 548 case S2_storerh_pr: // memh(Rx32++Mu2)=Rt32 549 case S2_storerhnew_pr: // memh(Rx32++Mu2)=Nt8.new 550 case S4_storerh_ur: // memh(Ru32<<#u2+#U6)=Rt32 551 case S4_storerhnew_ur: // memh(Ru32<<#u2+#U6)=Nt8.new 552 case S2_storerh_pbr: // memh(Rx32++Mu2:brev)=Rt32 553 case S2_storerhnew_pbr: // memh(Rx32++Mu2:brev)=Nt8.new 554 case S2_storerh_pci: // memh(Rx32++#s4:1:circ(Mu2))=Rt32 555 case S2_storerhnew_pci: // memh(Rx32++#s4:1:circ(Mu2))=Nt8.new 556 case S2_storerh_pcr: // memh(Rx32++I:circ(Mu2))=Rt32 557 case S2_storerhnew_pcr: // memh(Rx32++I:circ(Mu2))=Nt8.new 558 case S4_storerh_rr: // memh(Rs32+Ru32<<#u2)=Rt32 559 case S4_pstorerht_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Rt32 560 case S4_pstorerhf_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Rt32 561 case S4_pstorerhtnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Rt32 562 case S4_pstorerhfnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Rt32 563 case S4_storerhnew_rr: // memh(Rs32+Ru32<<#u2)=Nt8.new 564 case S4_pstorerhnewt_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Nt8.new 565 case S4_pstorerhnewf_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Nt8.new 566 case S4_pstorerhnewtnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Nt8.new 567 case S4_pstorerhnewfnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Nt8.new 568 case S2_storerhgp: // memh(gp+#u16:1)=Rt32 569 case S2_storerhnewgp: // memh(gp+#u16:1)=Nt8.new 570 case S4_pstorerht_abs: // if (Pv4) memh(#u6)=Rt32 571 case S4_pstorerhf_abs: // if (!Pv4) memh(#u6)=Rt32 572 case S4_pstorerhtnew_abs: // if (Pv4.new) memh(#u6)=Rt32 573 case S4_pstorerhfnew_abs: // if (!Pv4.new) memh(#u6)=Rt32 574 case S4_pstorerhnewt_abs: // if (Pv4) memh(#u6)=Nt8.new 575 case S4_pstorerhnewf_abs: // if (!Pv4) memh(#u6)=Nt8.new 576 case S4_pstorerhnewtnew_abs: // if (Pv4.new) memh(#u6)=Nt8.new 577 case S4_pstorerhnewfnew_abs: // if (!Pv4.new) memh(#u6)=Nt8.new 578 Bits.set(Begin, Begin+16); 579 return true; 580 581 // Store high half 582 case S2_storerf_io: // memh(Rs32+#s11:1)=Rt.H32 583 case S2_pstorerft_io: // if (Pv4) memh(Rs32+#u6:1)=Rt.H32 584 case S2_pstorerff_io: // if (!Pv4) memh(Rs32+#u6:1)=Rt.H32 585 case S4_pstorerftnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Rt.H32 586 case S4_pstorerffnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Rt.H32 587 case S2_storerf_pi: // memh(Rx32++#s4:1)=Rt.H32 588 case S2_pstorerft_pi: // if (Pv4) memh(Rx32++#s4:1)=Rt.H32 589 case S2_pstorerff_pi: // if (!Pv4) memh(Rx32++#s4:1)=Rt.H32 590 case S2_pstorerftnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Rt.H32 591 case S2_pstorerffnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Rt.H32 592 case S4_storerf_ap: // memh(Re32=#U6)=Rt.H32 593 case S2_storerf_pr: // memh(Rx32++Mu2)=Rt.H32 594 case S4_storerf_ur: // memh(Ru32<<#u2+#U6)=Rt.H32 595 case S2_storerf_pbr: // memh(Rx32++Mu2:brev)=Rt.H32 596 case S2_storerf_pci: // memh(Rx32++#s4:1:circ(Mu2))=Rt.H32 597 case S2_storerf_pcr: // memh(Rx32++I:circ(Mu2))=Rt.H32 598 case S4_storerf_rr: // memh(Rs32+Ru32<<#u2)=Rt.H32 599 case S4_pstorerft_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Rt.H32 600 case S4_pstorerff_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Rt.H32 601 case S4_pstorerftnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Rt.H32 602 case S4_pstorerffnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Rt.H32 603 case S2_storerfgp: // memh(gp+#u16:1)=Rt.H32 604 case S4_pstorerft_abs: // if (Pv4) memh(#u6)=Rt.H32 605 case S4_pstorerff_abs: // if (!Pv4) memh(#u6)=Rt.H32 606 case S4_pstorerftnew_abs: // if (Pv4.new) memh(#u6)=Rt.H32 607 case S4_pstorerffnew_abs: // if (!Pv4.new) memh(#u6)=Rt.H32 608 Bits.set(Begin+16, Begin+32); 609 return true; 610 } 611 612 return false; 613 } 614 615 // For an instruction with opcode Opc, calculate the set of bits that it 616 // uses in a register in operand OpN. This only calculates the set of used 617 // bits for cases where it does not depend on any operands (as is the case 618 // in shifts, for example). For concrete instructions from a program, the 619 // operand may be a subregister of a larger register, while Bits would 620 // correspond to the larger register in its entirety. Because of that, 621 // the parameter Begin can be used to indicate which bit of Bits should be 622 // considered the LSB of the operand. 623 bool HexagonBitSimplify::getUsedBits(unsigned Opc, unsigned OpN, 624 BitVector &Bits, uint16_t Begin, const HexagonInstrInfo &HII) { 625 using namespace Hexagon; 626 627 const MCInstrDesc &D = HII.get(Opc); 628 if (D.mayStore()) { 629 if (OpN == D.getNumOperands()-1) 630 return getUsedBitsInStore(Opc, Bits, Begin); 631 return false; 632 } 633 634 switch (Opc) { 635 // One register source. Used bits: R1[0-7]. 636 case A2_sxtb: 637 case A2_zxtb: 638 case A4_cmpbeqi: 639 case A4_cmpbgti: 640 case A4_cmpbgtui: 641 if (OpN == 1) { 642 Bits.set(Begin, Begin+8); 643 return true; 644 } 645 break; 646 647 // One register source. Used bits: R1[0-15]. 648 case A2_aslh: 649 case A2_sxth: 650 case A2_zxth: 651 case A4_cmpheqi: 652 case A4_cmphgti: 653 case A4_cmphgtui: 654 if (OpN == 1) { 655 Bits.set(Begin, Begin+16); 656 return true; 657 } 658 break; 659 660 // One register source. Used bits: R1[16-31]. 661 case A2_asrh: 662 if (OpN == 1) { 663 Bits.set(Begin+16, Begin+32); 664 return true; 665 } 666 break; 667 668 // Two register sources. Used bits: R1[0-7], R2[0-7]. 669 case A4_cmpbeq: 670 case A4_cmpbgt: 671 case A4_cmpbgtu: 672 if (OpN == 1) { 673 Bits.set(Begin, Begin+8); 674 return true; 675 } 676 break; 677 678 // Two register sources. Used bits: R1[0-15], R2[0-15]. 679 case A4_cmpheq: 680 case A4_cmphgt: 681 case A4_cmphgtu: 682 case A2_addh_h16_ll: 683 case A2_addh_h16_sat_ll: 684 case A2_addh_l16_ll: 685 case A2_addh_l16_sat_ll: 686 case A2_combine_ll: 687 case A2_subh_h16_ll: 688 case A2_subh_h16_sat_ll: 689 case A2_subh_l16_ll: 690 case A2_subh_l16_sat_ll: 691 case M2_mpy_acc_ll_s0: 692 case M2_mpy_acc_ll_s1: 693 case M2_mpy_acc_sat_ll_s0: 694 case M2_mpy_acc_sat_ll_s1: 695 case M2_mpy_ll_s0: 696 case M2_mpy_ll_s1: 697 case M2_mpy_nac_ll_s0: 698 case M2_mpy_nac_ll_s1: 699 case M2_mpy_nac_sat_ll_s0: 700 case M2_mpy_nac_sat_ll_s1: 701 case M2_mpy_rnd_ll_s0: 702 case M2_mpy_rnd_ll_s1: 703 case M2_mpy_sat_ll_s0: 704 case M2_mpy_sat_ll_s1: 705 case M2_mpy_sat_rnd_ll_s0: 706 case M2_mpy_sat_rnd_ll_s1: 707 case M2_mpyd_acc_ll_s0: 708 case M2_mpyd_acc_ll_s1: 709 case M2_mpyd_ll_s0: 710 case M2_mpyd_ll_s1: 711 case M2_mpyd_nac_ll_s0: 712 case M2_mpyd_nac_ll_s1: 713 case M2_mpyd_rnd_ll_s0: 714 case M2_mpyd_rnd_ll_s1: 715 case M2_mpyu_acc_ll_s0: 716 case M2_mpyu_acc_ll_s1: 717 case M2_mpyu_ll_s0: 718 case M2_mpyu_ll_s1: 719 case M2_mpyu_nac_ll_s0: 720 case M2_mpyu_nac_ll_s1: 721 case M2_mpyud_acc_ll_s0: 722 case M2_mpyud_acc_ll_s1: 723 case M2_mpyud_ll_s0: 724 case M2_mpyud_ll_s1: 725 case M2_mpyud_nac_ll_s0: 726 case M2_mpyud_nac_ll_s1: 727 if (OpN == 1 || OpN == 2) { 728 Bits.set(Begin, Begin+16); 729 return true; 730 } 731 break; 732 733 // Two register sources. Used bits: R1[0-15], R2[16-31]. 734 case A2_addh_h16_lh: 735 case A2_addh_h16_sat_lh: 736 case A2_combine_lh: 737 case A2_subh_h16_lh: 738 case A2_subh_h16_sat_lh: 739 case M2_mpy_acc_lh_s0: 740 case M2_mpy_acc_lh_s1: 741 case M2_mpy_acc_sat_lh_s0: 742 case M2_mpy_acc_sat_lh_s1: 743 case M2_mpy_lh_s0: 744 case M2_mpy_lh_s1: 745 case M2_mpy_nac_lh_s0: 746 case M2_mpy_nac_lh_s1: 747 case M2_mpy_nac_sat_lh_s0: 748 case M2_mpy_nac_sat_lh_s1: 749 case M2_mpy_rnd_lh_s0: 750 case M2_mpy_rnd_lh_s1: 751 case M2_mpy_sat_lh_s0: 752 case M2_mpy_sat_lh_s1: 753 case M2_mpy_sat_rnd_lh_s0: 754 case M2_mpy_sat_rnd_lh_s1: 755 case M2_mpyd_acc_lh_s0: 756 case M2_mpyd_acc_lh_s1: 757 case M2_mpyd_lh_s0: 758 case M2_mpyd_lh_s1: 759 case M2_mpyd_nac_lh_s0: 760 case M2_mpyd_nac_lh_s1: 761 case M2_mpyd_rnd_lh_s0: 762 case M2_mpyd_rnd_lh_s1: 763 case M2_mpyu_acc_lh_s0: 764 case M2_mpyu_acc_lh_s1: 765 case M2_mpyu_lh_s0: 766 case M2_mpyu_lh_s1: 767 case M2_mpyu_nac_lh_s0: 768 case M2_mpyu_nac_lh_s1: 769 case M2_mpyud_acc_lh_s0: 770 case M2_mpyud_acc_lh_s1: 771 case M2_mpyud_lh_s0: 772 case M2_mpyud_lh_s1: 773 case M2_mpyud_nac_lh_s0: 774 case M2_mpyud_nac_lh_s1: 775 // These four are actually LH. 776 case A2_addh_l16_hl: 777 case A2_addh_l16_sat_hl: 778 case A2_subh_l16_hl: 779 case A2_subh_l16_sat_hl: 780 if (OpN == 1) { 781 Bits.set(Begin, Begin+16); 782 return true; 783 } 784 if (OpN == 2) { 785 Bits.set(Begin+16, Begin+32); 786 return true; 787 } 788 break; 789 790 // Two register sources, used bits: R1[16-31], R2[0-15]. 791 case A2_addh_h16_hl: 792 case A2_addh_h16_sat_hl: 793 case A2_combine_hl: 794 case A2_subh_h16_hl: 795 case A2_subh_h16_sat_hl: 796 case M2_mpy_acc_hl_s0: 797 case M2_mpy_acc_hl_s1: 798 case M2_mpy_acc_sat_hl_s0: 799 case M2_mpy_acc_sat_hl_s1: 800 case M2_mpy_hl_s0: 801 case M2_mpy_hl_s1: 802 case M2_mpy_nac_hl_s0: 803 case M2_mpy_nac_hl_s1: 804 case M2_mpy_nac_sat_hl_s0: 805 case M2_mpy_nac_sat_hl_s1: 806 case M2_mpy_rnd_hl_s0: 807 case M2_mpy_rnd_hl_s1: 808 case M2_mpy_sat_hl_s0: 809 case M2_mpy_sat_hl_s1: 810 case M2_mpy_sat_rnd_hl_s0: 811 case M2_mpy_sat_rnd_hl_s1: 812 case M2_mpyd_acc_hl_s0: 813 case M2_mpyd_acc_hl_s1: 814 case M2_mpyd_hl_s0: 815 case M2_mpyd_hl_s1: 816 case M2_mpyd_nac_hl_s0: 817 case M2_mpyd_nac_hl_s1: 818 case M2_mpyd_rnd_hl_s0: 819 case M2_mpyd_rnd_hl_s1: 820 case M2_mpyu_acc_hl_s0: 821 case M2_mpyu_acc_hl_s1: 822 case M2_mpyu_hl_s0: 823 case M2_mpyu_hl_s1: 824 case M2_mpyu_nac_hl_s0: 825 case M2_mpyu_nac_hl_s1: 826 case M2_mpyud_acc_hl_s0: 827 case M2_mpyud_acc_hl_s1: 828 case M2_mpyud_hl_s0: 829 case M2_mpyud_hl_s1: 830 case M2_mpyud_nac_hl_s0: 831 case M2_mpyud_nac_hl_s1: 832 if (OpN == 1) { 833 Bits.set(Begin+16, Begin+32); 834 return true; 835 } 836 if (OpN == 2) { 837 Bits.set(Begin, Begin+16); 838 return true; 839 } 840 break; 841 842 // Two register sources, used bits: R1[16-31], R2[16-31]. 843 case A2_addh_h16_hh: 844 case A2_addh_h16_sat_hh: 845 case A2_combine_hh: 846 case A2_subh_h16_hh: 847 case A2_subh_h16_sat_hh: 848 case M2_mpy_acc_hh_s0: 849 case M2_mpy_acc_hh_s1: 850 case M2_mpy_acc_sat_hh_s0: 851 case M2_mpy_acc_sat_hh_s1: 852 case M2_mpy_hh_s0: 853 case M2_mpy_hh_s1: 854 case M2_mpy_nac_hh_s0: 855 case M2_mpy_nac_hh_s1: 856 case M2_mpy_nac_sat_hh_s0: 857 case M2_mpy_nac_sat_hh_s1: 858 case M2_mpy_rnd_hh_s0: 859 case M2_mpy_rnd_hh_s1: 860 case M2_mpy_sat_hh_s0: 861 case M2_mpy_sat_hh_s1: 862 case M2_mpy_sat_rnd_hh_s0: 863 case M2_mpy_sat_rnd_hh_s1: 864 case M2_mpyd_acc_hh_s0: 865 case M2_mpyd_acc_hh_s1: 866 case M2_mpyd_hh_s0: 867 case M2_mpyd_hh_s1: 868 case M2_mpyd_nac_hh_s0: 869 case M2_mpyd_nac_hh_s1: 870 case M2_mpyd_rnd_hh_s0: 871 case M2_mpyd_rnd_hh_s1: 872 case M2_mpyu_acc_hh_s0: 873 case M2_mpyu_acc_hh_s1: 874 case M2_mpyu_hh_s0: 875 case M2_mpyu_hh_s1: 876 case M2_mpyu_nac_hh_s0: 877 case M2_mpyu_nac_hh_s1: 878 case M2_mpyud_acc_hh_s0: 879 case M2_mpyud_acc_hh_s1: 880 case M2_mpyud_hh_s0: 881 case M2_mpyud_hh_s1: 882 case M2_mpyud_nac_hh_s0: 883 case M2_mpyud_nac_hh_s1: 884 if (OpN == 1 || OpN == 2) { 885 Bits.set(Begin+16, Begin+32); 886 return true; 887 } 888 break; 889 } 890 891 return false; 892 } 893 894 // Calculate the register class that matches Reg:Sub. For example, if 895 // %1 is a double register, then %1:isub_hi would match the "int" 896 // register class. 897 const TargetRegisterClass *HexagonBitSimplify::getFinalVRegClass( 898 const BitTracker::RegisterRef &RR, MachineRegisterInfo &MRI) { 899 if (!RR.Reg.isVirtual()) 900 return nullptr; 901 auto *RC = MRI.getRegClass(RR.Reg); 902 if (RR.Sub == 0) 903 return RC; 904 auto &HRI = static_cast<const HexagonRegisterInfo&>( 905 *MRI.getTargetRegisterInfo()); 906 907 auto VerifySR = [&HRI] (const TargetRegisterClass *RC, unsigned Sub) -> void { 908 (void)HRI; 909 assert(Sub == HRI.getHexagonSubRegIndex(*RC, Hexagon::ps_sub_lo) || 910 Sub == HRI.getHexagonSubRegIndex(*RC, Hexagon::ps_sub_hi)); 911 }; 912 913 switch (RC->getID()) { 914 case Hexagon::DoubleRegsRegClassID: 915 VerifySR(RC, RR.Sub); 916 return &Hexagon::IntRegsRegClass; 917 case Hexagon::HvxWRRegClassID: 918 VerifySR(RC, RR.Sub); 919 return &Hexagon::HvxVRRegClass; 920 } 921 return nullptr; 922 } 923 924 // Check if RD could be replaced with RS at any possible use of RD. 925 // For example a predicate register cannot be replaced with a integer 926 // register, but a 64-bit register with a subregister can be replaced 927 // with a 32-bit register. 928 bool HexagonBitSimplify::isTransparentCopy(const BitTracker::RegisterRef &RD, 929 const BitTracker::RegisterRef &RS, MachineRegisterInfo &MRI) { 930 if (!RD.Reg.isVirtual() || !RS.Reg.isVirtual()) 931 return false; 932 // Return false if one (or both) classes are nullptr. 933 auto *DRC = getFinalVRegClass(RD, MRI); 934 if (!DRC) 935 return false; 936 937 return DRC == getFinalVRegClass(RS, MRI); 938 } 939 940 bool HexagonBitSimplify::hasTiedUse(unsigned Reg, MachineRegisterInfo &MRI, 941 unsigned NewSub) { 942 if (!PreserveTiedOps) 943 return false; 944 return llvm::any_of(MRI.use_operands(Reg), 945 [NewSub] (const MachineOperand &Op) -> bool { 946 return Op.getSubReg() != NewSub && Op.isTied(); 947 }); 948 } 949 950 namespace { 951 952 class DeadCodeElimination { 953 public: 954 DeadCodeElimination(MachineFunction &mf, MachineDominatorTree &mdt) 955 : MF(mf), HII(*MF.getSubtarget<HexagonSubtarget>().getInstrInfo()), 956 MDT(mdt), MRI(mf.getRegInfo()) {} 957 958 bool run() { 959 return runOnNode(MDT.getRootNode()); 960 } 961 962 private: 963 bool isDead(unsigned R) const; 964 bool runOnNode(MachineDomTreeNode *N); 965 966 MachineFunction &MF; 967 const HexagonInstrInfo &HII; 968 MachineDominatorTree &MDT; 969 MachineRegisterInfo &MRI; 970 }; 971 972 } // end anonymous namespace 973 974 bool DeadCodeElimination::isDead(unsigned R) const { 975 for (const MachineOperand &MO : MRI.use_operands(R)) { 976 const MachineInstr *UseI = MO.getParent(); 977 if (UseI->isDebugValue()) 978 continue; 979 if (UseI->isPHI()) { 980 assert(!UseI->getOperand(0).getSubReg()); 981 Register DR = UseI->getOperand(0).getReg(); 982 if (DR == R) 983 continue; 984 } 985 return false; 986 } 987 return true; 988 } 989 990 bool DeadCodeElimination::runOnNode(MachineDomTreeNode *N) { 991 bool Changed = false; 992 993 for (auto *DTN : children<MachineDomTreeNode*>(N)) 994 Changed |= runOnNode(DTN); 995 996 MachineBasicBlock *B = N->getBlock(); 997 std::vector<MachineInstr*> Instrs; 998 for (auto I = B->rbegin(), E = B->rend(); I != E; ++I) 999 Instrs.push_back(&*I); 1000 1001 for (auto MI : Instrs) { 1002 unsigned Opc = MI->getOpcode(); 1003 // Do not touch lifetime markers. This is why the target-independent DCE 1004 // cannot be used. 1005 if (Opc == TargetOpcode::LIFETIME_START || 1006 Opc == TargetOpcode::LIFETIME_END) 1007 continue; 1008 bool Store = false; 1009 if (MI->isInlineAsm()) 1010 continue; 1011 // Delete PHIs if possible. 1012 if (!MI->isPHI() && !MI->isSafeToMove(nullptr, Store)) 1013 continue; 1014 1015 bool AllDead = true; 1016 SmallVector<unsigned,2> Regs; 1017 for (auto &Op : MI->operands()) { 1018 if (!Op.isReg() || !Op.isDef()) 1019 continue; 1020 Register R = Op.getReg(); 1021 if (!R.isVirtual() || !isDead(R)) { 1022 AllDead = false; 1023 break; 1024 } 1025 Regs.push_back(R); 1026 } 1027 if (!AllDead) 1028 continue; 1029 1030 B->erase(MI); 1031 for (unsigned i = 0, n = Regs.size(); i != n; ++i) 1032 MRI.markUsesInDebugValueAsUndef(Regs[i]); 1033 Changed = true; 1034 } 1035 1036 return Changed; 1037 } 1038 1039 namespace { 1040 1041 // Eliminate redundant instructions 1042 // 1043 // This transformation will identify instructions where the output register 1044 // is the same as one of its input registers. This only works on instructions 1045 // that define a single register (unlike post-increment loads, for example). 1046 // The equality check is actually more detailed: the code calculates which 1047 // bits of the output are used, and only compares these bits with the input 1048 // registers. 1049 // If the output matches an input, the instruction is replaced with COPY. 1050 // The copies will be removed by another transformation. 1051 class RedundantInstrElimination : public Transformation { 1052 public: 1053 RedundantInstrElimination(BitTracker &bt, const HexagonInstrInfo &hii, 1054 const HexagonRegisterInfo &hri, MachineRegisterInfo &mri) 1055 : Transformation(true), HII(hii), HRI(hri), MRI(mri), BT(bt) {} 1056 1057 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; 1058 1059 private: 1060 bool isLossyShiftLeft(const MachineInstr &MI, unsigned OpN, 1061 unsigned &LostB, unsigned &LostE); 1062 bool isLossyShiftRight(const MachineInstr &MI, unsigned OpN, 1063 unsigned &LostB, unsigned &LostE); 1064 bool computeUsedBits(unsigned Reg, BitVector &Bits); 1065 bool computeUsedBits(const MachineInstr &MI, unsigned OpN, BitVector &Bits, 1066 uint16_t Begin); 1067 bool usedBitsEqual(BitTracker::RegisterRef RD, BitTracker::RegisterRef RS); 1068 1069 const HexagonInstrInfo &HII; 1070 const HexagonRegisterInfo &HRI; 1071 MachineRegisterInfo &MRI; 1072 BitTracker &BT; 1073 }; 1074 1075 } // end anonymous namespace 1076 1077 // Check if the instruction is a lossy shift left, where the input being 1078 // shifted is the operand OpN of MI. If true, [LostB, LostE) is the range 1079 // of bit indices that are lost. 1080 bool RedundantInstrElimination::isLossyShiftLeft(const MachineInstr &MI, 1081 unsigned OpN, unsigned &LostB, unsigned &LostE) { 1082 using namespace Hexagon; 1083 1084 unsigned Opc = MI.getOpcode(); 1085 unsigned ImN, RegN, Width; 1086 switch (Opc) { 1087 case S2_asl_i_p: 1088 ImN = 2; 1089 RegN = 1; 1090 Width = 64; 1091 break; 1092 case S2_asl_i_p_acc: 1093 case S2_asl_i_p_and: 1094 case S2_asl_i_p_nac: 1095 case S2_asl_i_p_or: 1096 case S2_asl_i_p_xacc: 1097 ImN = 3; 1098 RegN = 2; 1099 Width = 64; 1100 break; 1101 case S2_asl_i_r: 1102 ImN = 2; 1103 RegN = 1; 1104 Width = 32; 1105 break; 1106 case S2_addasl_rrri: 1107 case S4_andi_asl_ri: 1108 case S4_ori_asl_ri: 1109 case S4_addi_asl_ri: 1110 case S4_subi_asl_ri: 1111 case S2_asl_i_r_acc: 1112 case S2_asl_i_r_and: 1113 case S2_asl_i_r_nac: 1114 case S2_asl_i_r_or: 1115 case S2_asl_i_r_sat: 1116 case S2_asl_i_r_xacc: 1117 ImN = 3; 1118 RegN = 2; 1119 Width = 32; 1120 break; 1121 default: 1122 return false; 1123 } 1124 1125 if (RegN != OpN) 1126 return false; 1127 1128 assert(MI.getOperand(ImN).isImm()); 1129 unsigned S = MI.getOperand(ImN).getImm(); 1130 if (S == 0) 1131 return false; 1132 LostB = Width-S; 1133 LostE = Width; 1134 return true; 1135 } 1136 1137 // Check if the instruction is a lossy shift right, where the input being 1138 // shifted is the operand OpN of MI. If true, [LostB, LostE) is the range 1139 // of bit indices that are lost. 1140 bool RedundantInstrElimination::isLossyShiftRight(const MachineInstr &MI, 1141 unsigned OpN, unsigned &LostB, unsigned &LostE) { 1142 using namespace Hexagon; 1143 1144 unsigned Opc = MI.getOpcode(); 1145 unsigned ImN, RegN; 1146 switch (Opc) { 1147 case S2_asr_i_p: 1148 case S2_lsr_i_p: 1149 ImN = 2; 1150 RegN = 1; 1151 break; 1152 case S2_asr_i_p_acc: 1153 case S2_asr_i_p_and: 1154 case S2_asr_i_p_nac: 1155 case S2_asr_i_p_or: 1156 case S2_lsr_i_p_acc: 1157 case S2_lsr_i_p_and: 1158 case S2_lsr_i_p_nac: 1159 case S2_lsr_i_p_or: 1160 case S2_lsr_i_p_xacc: 1161 ImN = 3; 1162 RegN = 2; 1163 break; 1164 case S2_asr_i_r: 1165 case S2_lsr_i_r: 1166 ImN = 2; 1167 RegN = 1; 1168 break; 1169 case S4_andi_lsr_ri: 1170 case S4_ori_lsr_ri: 1171 case S4_addi_lsr_ri: 1172 case S4_subi_lsr_ri: 1173 case S2_asr_i_r_acc: 1174 case S2_asr_i_r_and: 1175 case S2_asr_i_r_nac: 1176 case S2_asr_i_r_or: 1177 case S2_lsr_i_r_acc: 1178 case S2_lsr_i_r_and: 1179 case S2_lsr_i_r_nac: 1180 case S2_lsr_i_r_or: 1181 case S2_lsr_i_r_xacc: 1182 ImN = 3; 1183 RegN = 2; 1184 break; 1185 1186 default: 1187 return false; 1188 } 1189 1190 if (RegN != OpN) 1191 return false; 1192 1193 assert(MI.getOperand(ImN).isImm()); 1194 unsigned S = MI.getOperand(ImN).getImm(); 1195 LostB = 0; 1196 LostE = S; 1197 return true; 1198 } 1199 1200 // Calculate the bit vector that corresponds to the used bits of register Reg. 1201 // The vector Bits has the same size, as the size of Reg in bits. If the cal- 1202 // culation fails (i.e. the used bits are unknown), it returns false. Other- 1203 // wise, it returns true and sets the corresponding bits in Bits. 1204 bool RedundantInstrElimination::computeUsedBits(unsigned Reg, BitVector &Bits) { 1205 BitVector Used(Bits.size()); 1206 RegisterSet Visited; 1207 std::vector<unsigned> Pending; 1208 Pending.push_back(Reg); 1209 1210 for (unsigned i = 0; i < Pending.size(); ++i) { 1211 unsigned R = Pending[i]; 1212 if (Visited.has(R)) 1213 continue; 1214 Visited.insert(R); 1215 for (auto I = MRI.use_begin(R), E = MRI.use_end(); I != E; ++I) { 1216 BitTracker::RegisterRef UR = *I; 1217 unsigned B, W; 1218 if (!HBS::getSubregMask(UR, B, W, MRI)) 1219 return false; 1220 MachineInstr &UseI = *I->getParent(); 1221 if (UseI.isPHI() || UseI.isCopy()) { 1222 Register DefR = UseI.getOperand(0).getReg(); 1223 if (!DefR.isVirtual()) 1224 return false; 1225 Pending.push_back(DefR); 1226 } else { 1227 if (!computeUsedBits(UseI, I.getOperandNo(), Used, B)) 1228 return false; 1229 } 1230 } 1231 } 1232 Bits |= Used; 1233 return true; 1234 } 1235 1236 // Calculate the bits used by instruction MI in a register in operand OpN. 1237 // Return true/false if the calculation succeeds/fails. If is succeeds, set 1238 // used bits in Bits. This function does not reset any bits in Bits, so 1239 // subsequent calls over different instructions will result in the union 1240 // of the used bits in all these instructions. 1241 // The register in question may be used with a sub-register, whereas Bits 1242 // holds the bits for the entire register. To keep track of that, the 1243 // argument Begin indicates where in Bits is the lowest-significant bit 1244 // of the register used in operand OpN. For example, in instruction: 1245 // %1 = S2_lsr_i_r %2:isub_hi, 10 1246 // the operand 1 is a 32-bit register, which happens to be a subregister 1247 // of the 64-bit register %2, and that subregister starts at position 32. 1248 // In this case Begin=32, since Bits[32] would be the lowest-significant bit 1249 // of %2:isub_hi. 1250 bool RedundantInstrElimination::computeUsedBits(const MachineInstr &MI, 1251 unsigned OpN, BitVector &Bits, uint16_t Begin) { 1252 unsigned Opc = MI.getOpcode(); 1253 BitVector T(Bits.size()); 1254 bool GotBits = HBS::getUsedBits(Opc, OpN, T, Begin, HII); 1255 // Even if we don't have bits yet, we could still provide some information 1256 // if the instruction is a lossy shift: the lost bits will be marked as 1257 // not used. 1258 unsigned LB, LE; 1259 if (isLossyShiftLeft(MI, OpN, LB, LE) || isLossyShiftRight(MI, OpN, LB, LE)) { 1260 assert(MI.getOperand(OpN).isReg()); 1261 BitTracker::RegisterRef RR = MI.getOperand(OpN); 1262 const TargetRegisterClass *RC = HBS::getFinalVRegClass(RR, MRI); 1263 uint16_t Width = HRI.getRegSizeInBits(*RC); 1264 1265 if (!GotBits) 1266 T.set(Begin, Begin+Width); 1267 assert(LB <= LE && LB < Width && LE <= Width); 1268 T.reset(Begin+LB, Begin+LE); 1269 GotBits = true; 1270 } 1271 if (GotBits) 1272 Bits |= T; 1273 return GotBits; 1274 } 1275 1276 // Calculates the used bits in RD ("defined register"), and checks if these 1277 // bits in RS ("used register") and RD are identical. 1278 bool RedundantInstrElimination::usedBitsEqual(BitTracker::RegisterRef RD, 1279 BitTracker::RegisterRef RS) { 1280 const BitTracker::RegisterCell &DC = BT.lookup(RD.Reg); 1281 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); 1282 1283 unsigned DB, DW; 1284 if (!HBS::getSubregMask(RD, DB, DW, MRI)) 1285 return false; 1286 unsigned SB, SW; 1287 if (!HBS::getSubregMask(RS, SB, SW, MRI)) 1288 return false; 1289 if (SW != DW) 1290 return false; 1291 1292 BitVector Used(DC.width()); 1293 if (!computeUsedBits(RD.Reg, Used)) 1294 return false; 1295 1296 for (unsigned i = 0; i != DW; ++i) 1297 if (Used[i+DB] && DC[DB+i] != SC[SB+i]) 1298 return false; 1299 return true; 1300 } 1301 1302 bool RedundantInstrElimination::processBlock(MachineBasicBlock &B, 1303 const RegisterSet&) { 1304 if (!BT.reached(&B)) 1305 return false; 1306 bool Changed = false; 1307 1308 for (auto I = B.begin(), E = B.end(); I != E; ++I) { 1309 MachineInstr *MI = &*I; 1310 1311 if (MI->getOpcode() == TargetOpcode::COPY) 1312 continue; 1313 if (MI->isPHI() || MI->hasUnmodeledSideEffects() || MI->isInlineAsm()) 1314 continue; 1315 unsigned NumD = MI->getDesc().getNumDefs(); 1316 if (NumD != 1) 1317 continue; 1318 1319 BitTracker::RegisterRef RD = MI->getOperand(0); 1320 if (!BT.has(RD.Reg)) 1321 continue; 1322 const BitTracker::RegisterCell &DC = BT.lookup(RD.Reg); 1323 auto At = MachineBasicBlock::iterator(MI); 1324 1325 // Find a source operand that is equal to the result. 1326 for (auto &Op : MI->uses()) { 1327 if (!Op.isReg()) 1328 continue; 1329 BitTracker::RegisterRef RS = Op; 1330 if (!BT.has(RS.Reg)) 1331 continue; 1332 if (!HBS::isTransparentCopy(RD, RS, MRI)) 1333 continue; 1334 1335 unsigned BN, BW; 1336 if (!HBS::getSubregMask(RS, BN, BW, MRI)) 1337 continue; 1338 1339 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); 1340 if (!usedBitsEqual(RD, RS) && !HBS::isEqual(DC, 0, SC, BN, BW)) 1341 continue; 1342 1343 // If found, replace the instruction with a COPY. 1344 const DebugLoc &DL = MI->getDebugLoc(); 1345 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); 1346 Register NewR = MRI.createVirtualRegister(FRC); 1347 MachineInstr *CopyI = 1348 BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR) 1349 .addReg(RS.Reg, 0, RS.Sub); 1350 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); 1351 // This pass can create copies between registers that don't have the 1352 // exact same values. Updating the tracker has to involve updating 1353 // all dependent cells. Example: 1354 // %1 = inst %2 ; %1 != %2, but used bits are equal 1355 // 1356 // %3 = copy %2 ; <- inserted 1357 // ... = %3 ; <- replaced from %2 1358 // Indirectly, we can create a "copy" between %1 and %2 even 1359 // though their exact values do not match. 1360 BT.visit(*CopyI); 1361 Changed = true; 1362 break; 1363 } 1364 } 1365 1366 return Changed; 1367 } 1368 1369 namespace { 1370 1371 // Recognize instructions that produce constant values known at compile-time. 1372 // Replace them with register definitions that load these constants directly. 1373 class ConstGeneration : public Transformation { 1374 public: 1375 ConstGeneration(BitTracker &bt, const HexagonInstrInfo &hii, 1376 MachineRegisterInfo &mri) 1377 : Transformation(true), HII(hii), MRI(mri), BT(bt) {} 1378 1379 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; 1380 static bool isTfrConst(const MachineInstr &MI); 1381 1382 private: 1383 Register genTfrConst(const TargetRegisterClass *RC, int64_t C, 1384 MachineBasicBlock &B, MachineBasicBlock::iterator At, 1385 DebugLoc &DL); 1386 1387 const HexagonInstrInfo &HII; 1388 MachineRegisterInfo &MRI; 1389 BitTracker &BT; 1390 }; 1391 1392 } // end anonymous namespace 1393 1394 bool ConstGeneration::isTfrConst(const MachineInstr &MI) { 1395 unsigned Opc = MI.getOpcode(); 1396 switch (Opc) { 1397 case Hexagon::A2_combineii: 1398 case Hexagon::A4_combineii: 1399 case Hexagon::A2_tfrsi: 1400 case Hexagon::A2_tfrpi: 1401 case Hexagon::PS_true: 1402 case Hexagon::PS_false: 1403 case Hexagon::CONST32: 1404 case Hexagon::CONST64: 1405 return true; 1406 } 1407 return false; 1408 } 1409 1410 // Generate a transfer-immediate instruction that is appropriate for the 1411 // register class and the actual value being transferred. 1412 Register ConstGeneration::genTfrConst(const TargetRegisterClass *RC, int64_t C, 1413 MachineBasicBlock &B, 1414 MachineBasicBlock::iterator At, 1415 DebugLoc &DL) { 1416 Register 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 MachineFunction *MF = B.getParent(); 1440 auto &HST = MF->getSubtarget<HexagonSubtarget>(); 1441 1442 // Disable CONST64 for tiny core since it takes a LD resource. 1443 if (!HST.isTinyCore() || 1444 MF->getFunction().hasOptSize()) { 1445 BuildMI(B, At, DL, HII.get(Hexagon::CONST64), Reg) 1446 .addImm(C); 1447 return Reg; 1448 } 1449 } 1450 1451 if (RC == &Hexagon::PredRegsRegClass) { 1452 unsigned Opc; 1453 if (C == 0) 1454 Opc = Hexagon::PS_false; 1455 else if ((C & 0xFF) == 0xFF) 1456 Opc = Hexagon::PS_true; 1457 else 1458 return 0; 1459 BuildMI(B, At, DL, HII.get(Opc), Reg); 1460 return Reg; 1461 } 1462 1463 return 0; 1464 } 1465 1466 bool ConstGeneration::processBlock(MachineBasicBlock &B, const RegisterSet&) { 1467 if (!BT.reached(&B)) 1468 return false; 1469 bool Changed = false; 1470 RegisterSet Defs; 1471 1472 for (auto I = B.begin(), E = B.end(); I != E; ++I) { 1473 if (isTfrConst(*I)) 1474 continue; 1475 Defs.clear(); 1476 HBS::getInstrDefs(*I, Defs); 1477 if (Defs.count() != 1) 1478 continue; 1479 Register DR = Defs.find_first(); 1480 if (!DR.isVirtual()) 1481 continue; 1482 uint64_t U; 1483 const BitTracker::RegisterCell &DRC = BT.lookup(DR); 1484 if (HBS::getConst(DRC, 0, DRC.width(), U)) { 1485 int64_t C = U; 1486 DebugLoc DL = I->getDebugLoc(); 1487 auto At = I->isPHI() ? B.getFirstNonPHI() : I; 1488 Register ImmReg = genTfrConst(MRI.getRegClass(DR), C, B, At, DL); 1489 if (ImmReg) { 1490 HBS::replaceReg(DR, ImmReg, MRI); 1491 BT.put(ImmReg, DRC); 1492 Changed = true; 1493 } 1494 } 1495 } 1496 return Changed; 1497 } 1498 1499 namespace { 1500 1501 // Identify pairs of available registers which hold identical values. 1502 // In such cases, only one of them needs to be calculated, the other one 1503 // will be defined as a copy of the first. 1504 class CopyGeneration : public Transformation { 1505 public: 1506 CopyGeneration(BitTracker &bt, const HexagonInstrInfo &hii, 1507 const HexagonRegisterInfo &hri, MachineRegisterInfo &mri) 1508 : Transformation(true), HII(hii), HRI(hri), MRI(mri), BT(bt) {} 1509 1510 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; 1511 1512 private: 1513 bool findMatch(const BitTracker::RegisterRef &Inp, 1514 BitTracker::RegisterRef &Out, const RegisterSet &AVs); 1515 1516 const HexagonInstrInfo &HII; 1517 const HexagonRegisterInfo &HRI; 1518 MachineRegisterInfo &MRI; 1519 BitTracker &BT; 1520 RegisterSet Forbidden; 1521 }; 1522 1523 // Eliminate register copies RD = RS, by replacing the uses of RD with 1524 // with uses of RS. 1525 class CopyPropagation : public Transformation { 1526 public: 1527 CopyPropagation(const HexagonRegisterInfo &hri, MachineRegisterInfo &mri) 1528 : Transformation(false), HRI(hri), MRI(mri) {} 1529 1530 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; 1531 1532 static bool isCopyReg(unsigned Opc, bool NoConv); 1533 1534 private: 1535 bool propagateRegCopy(MachineInstr &MI); 1536 1537 const HexagonRegisterInfo &HRI; 1538 MachineRegisterInfo &MRI; 1539 }; 1540 1541 } // end anonymous namespace 1542 1543 /// Check if there is a register in AVs that is identical to Inp. If so, 1544 /// set Out to the found register. The output may be a pair Reg:Sub. 1545 bool CopyGeneration::findMatch(const BitTracker::RegisterRef &Inp, 1546 BitTracker::RegisterRef &Out, const RegisterSet &AVs) { 1547 if (!BT.has(Inp.Reg)) 1548 return false; 1549 const BitTracker::RegisterCell &InpRC = BT.lookup(Inp.Reg); 1550 auto *FRC = HBS::getFinalVRegClass(Inp, MRI); 1551 unsigned B, W; 1552 if (!HBS::getSubregMask(Inp, B, W, MRI)) 1553 return false; 1554 1555 for (Register R = AVs.find_first(); R; R = AVs.find_next(R)) { 1556 if (!BT.has(R) || Forbidden[R]) 1557 continue; 1558 const BitTracker::RegisterCell &RC = BT.lookup(R); 1559 unsigned RW = RC.width(); 1560 if (W == RW) { 1561 if (FRC != MRI.getRegClass(R)) 1562 continue; 1563 if (!HBS::isTransparentCopy(R, Inp, MRI)) 1564 continue; 1565 if (!HBS::isEqual(InpRC, B, RC, 0, W)) 1566 continue; 1567 Out.Reg = R; 1568 Out.Sub = 0; 1569 return true; 1570 } 1571 // Check if there is a super-register, whose part (with a subregister) 1572 // is equal to the input. 1573 // Only do double registers for now. 1574 if (W*2 != RW) 1575 continue; 1576 if (MRI.getRegClass(R) != &Hexagon::DoubleRegsRegClass) 1577 continue; 1578 1579 if (HBS::isEqual(InpRC, B, RC, 0, W)) 1580 Out.Sub = Hexagon::isub_lo; 1581 else if (HBS::isEqual(InpRC, B, RC, W, W)) 1582 Out.Sub = Hexagon::isub_hi; 1583 else 1584 continue; 1585 Out.Reg = R; 1586 if (HBS::isTransparentCopy(Out, Inp, MRI)) 1587 return true; 1588 } 1589 return false; 1590 } 1591 1592 bool CopyGeneration::processBlock(MachineBasicBlock &B, 1593 const RegisterSet &AVs) { 1594 if (!BT.reached(&B)) 1595 return false; 1596 RegisterSet AVB(AVs); 1597 bool Changed = false; 1598 RegisterSet Defs; 1599 1600 for (auto I = B.begin(), E = B.end(); I != E; ++I, AVB.insert(Defs)) { 1601 Defs.clear(); 1602 HBS::getInstrDefs(*I, Defs); 1603 1604 unsigned Opc = I->getOpcode(); 1605 if (CopyPropagation::isCopyReg(Opc, false) || 1606 ConstGeneration::isTfrConst(*I)) 1607 continue; 1608 1609 DebugLoc DL = I->getDebugLoc(); 1610 auto At = I->isPHI() ? B.getFirstNonPHI() : I; 1611 1612 for (Register R = Defs.find_first(); R; R = Defs.find_next(R)) { 1613 BitTracker::RegisterRef MR; 1614 auto *FRC = HBS::getFinalVRegClass(R, MRI); 1615 1616 if (findMatch(R, MR, AVB)) { 1617 Register NewR = MRI.createVirtualRegister(FRC); 1618 BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR) 1619 .addReg(MR.Reg, 0, MR.Sub); 1620 BT.put(BitTracker::RegisterRef(NewR), BT.get(MR)); 1621 HBS::replaceReg(R, NewR, MRI); 1622 Forbidden.insert(R); 1623 continue; 1624 } 1625 1626 if (FRC == &Hexagon::DoubleRegsRegClass || 1627 FRC == &Hexagon::HvxWRRegClass) { 1628 // Try to generate REG_SEQUENCE. 1629 unsigned SubLo = HRI.getHexagonSubRegIndex(*FRC, Hexagon::ps_sub_lo); 1630 unsigned SubHi = HRI.getHexagonSubRegIndex(*FRC, Hexagon::ps_sub_hi); 1631 BitTracker::RegisterRef TL = { R, SubLo }; 1632 BitTracker::RegisterRef TH = { R, SubHi }; 1633 BitTracker::RegisterRef ML, MH; 1634 if (findMatch(TL, ML, AVB) && findMatch(TH, MH, AVB)) { 1635 auto *FRC = HBS::getFinalVRegClass(R, MRI); 1636 Register NewR = MRI.createVirtualRegister(FRC); 1637 BuildMI(B, At, DL, HII.get(TargetOpcode::REG_SEQUENCE), NewR) 1638 .addReg(ML.Reg, 0, ML.Sub) 1639 .addImm(SubLo) 1640 .addReg(MH.Reg, 0, MH.Sub) 1641 .addImm(SubHi); 1642 BT.put(BitTracker::RegisterRef(NewR), BT.get(R)); 1643 HBS::replaceReg(R, NewR, MRI); 1644 Forbidden.insert(R); 1645 } 1646 } 1647 } 1648 } 1649 1650 return Changed; 1651 } 1652 1653 bool CopyPropagation::isCopyReg(unsigned Opc, bool NoConv) { 1654 switch (Opc) { 1655 case TargetOpcode::COPY: 1656 case TargetOpcode::REG_SEQUENCE: 1657 case Hexagon::A4_combineir: 1658 case Hexagon::A4_combineri: 1659 return true; 1660 case Hexagon::A2_tfr: 1661 case Hexagon::A2_tfrp: 1662 case Hexagon::A2_combinew: 1663 case Hexagon::V6_vcombine: 1664 return NoConv; 1665 default: 1666 break; 1667 } 1668 return false; 1669 } 1670 1671 bool CopyPropagation::propagateRegCopy(MachineInstr &MI) { 1672 bool Changed = false; 1673 unsigned Opc = MI.getOpcode(); 1674 BitTracker::RegisterRef RD = MI.getOperand(0); 1675 assert(MI.getOperand(0).getSubReg() == 0); 1676 1677 switch (Opc) { 1678 case TargetOpcode::COPY: 1679 case Hexagon::A2_tfr: 1680 case Hexagon::A2_tfrp: { 1681 BitTracker::RegisterRef RS = MI.getOperand(1); 1682 if (!HBS::isTransparentCopy(RD, RS, MRI)) 1683 break; 1684 if (RS.Sub != 0) 1685 Changed = HBS::replaceRegWithSub(RD.Reg, RS.Reg, RS.Sub, MRI); 1686 else 1687 Changed = HBS::replaceReg(RD.Reg, RS.Reg, MRI); 1688 break; 1689 } 1690 case TargetOpcode::REG_SEQUENCE: { 1691 BitTracker::RegisterRef SL, SH; 1692 if (HBS::parseRegSequence(MI, SL, SH, MRI)) { 1693 const TargetRegisterClass &RC = *MRI.getRegClass(RD.Reg); 1694 unsigned SubLo = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_lo); 1695 unsigned SubHi = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_hi); 1696 Changed = HBS::replaceSubWithSub(RD.Reg, SubLo, SL.Reg, SL.Sub, MRI); 1697 Changed |= HBS::replaceSubWithSub(RD.Reg, SubHi, SH.Reg, SH.Sub, MRI); 1698 } 1699 break; 1700 } 1701 case Hexagon::A2_combinew: 1702 case Hexagon::V6_vcombine: { 1703 const TargetRegisterClass &RC = *MRI.getRegClass(RD.Reg); 1704 unsigned SubLo = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_lo); 1705 unsigned SubHi = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_hi); 1706 BitTracker::RegisterRef RH = MI.getOperand(1), RL = MI.getOperand(2); 1707 Changed = HBS::replaceSubWithSub(RD.Reg, SubLo, RL.Reg, RL.Sub, MRI); 1708 Changed |= HBS::replaceSubWithSub(RD.Reg, SubHi, RH.Reg, RH.Sub, MRI); 1709 break; 1710 } 1711 case Hexagon::A4_combineir: 1712 case Hexagon::A4_combineri: { 1713 unsigned SrcX = (Opc == Hexagon::A4_combineir) ? 2 : 1; 1714 unsigned Sub = (Opc == Hexagon::A4_combineir) ? Hexagon::isub_lo 1715 : Hexagon::isub_hi; 1716 BitTracker::RegisterRef RS = MI.getOperand(SrcX); 1717 Changed = HBS::replaceSubWithSub(RD.Reg, Sub, RS.Reg, RS.Sub, MRI); 1718 break; 1719 } 1720 } 1721 return Changed; 1722 } 1723 1724 bool CopyPropagation::processBlock(MachineBasicBlock &B, const RegisterSet&) { 1725 std::vector<MachineInstr*> Instrs; 1726 for (MachineInstr &MI : llvm::reverse(B)) 1727 Instrs.push_back(&MI); 1728 1729 bool Changed = false; 1730 for (auto I : Instrs) { 1731 unsigned Opc = I->getOpcode(); 1732 if (!CopyPropagation::isCopyReg(Opc, true)) 1733 continue; 1734 Changed |= propagateRegCopy(*I); 1735 } 1736 1737 return Changed; 1738 } 1739 1740 namespace { 1741 1742 // Recognize patterns that can be simplified and replace them with the 1743 // simpler forms. 1744 // This is by no means complete 1745 class BitSimplification : public Transformation { 1746 public: 1747 BitSimplification(BitTracker &bt, const MachineDominatorTree &mdt, 1748 const HexagonInstrInfo &hii, const HexagonRegisterInfo &hri, 1749 MachineRegisterInfo &mri, MachineFunction &mf) 1750 : Transformation(true), MDT(mdt), HII(hii), HRI(hri), MRI(mri), 1751 MF(mf), BT(bt) {} 1752 1753 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; 1754 1755 private: 1756 struct RegHalf : public BitTracker::RegisterRef { 1757 bool Low; // Low/High halfword. 1758 }; 1759 1760 bool matchHalf(unsigned SelfR, const BitTracker::RegisterCell &RC, 1761 unsigned B, RegHalf &RH); 1762 bool validateReg(BitTracker::RegisterRef R, unsigned Opc, unsigned OpNum); 1763 1764 bool matchPackhl(unsigned SelfR, const BitTracker::RegisterCell &RC, 1765 BitTracker::RegisterRef &Rs, BitTracker::RegisterRef &Rt); 1766 unsigned getCombineOpcode(bool HLow, bool LLow); 1767 1768 bool genStoreUpperHalf(MachineInstr *MI); 1769 bool genStoreImmediate(MachineInstr *MI); 1770 bool genPackhl(MachineInstr *MI, BitTracker::RegisterRef RD, 1771 const BitTracker::RegisterCell &RC); 1772 bool genExtractHalf(MachineInstr *MI, BitTracker::RegisterRef RD, 1773 const BitTracker::RegisterCell &RC); 1774 bool genCombineHalf(MachineInstr *MI, BitTracker::RegisterRef RD, 1775 const BitTracker::RegisterCell &RC); 1776 bool genExtractLow(MachineInstr *MI, BitTracker::RegisterRef RD, 1777 const BitTracker::RegisterCell &RC); 1778 bool genBitSplit(MachineInstr *MI, BitTracker::RegisterRef RD, 1779 const BitTracker::RegisterCell &RC, const RegisterSet &AVs); 1780 bool simplifyTstbit(MachineInstr *MI, BitTracker::RegisterRef RD, 1781 const BitTracker::RegisterCell &RC); 1782 bool simplifyExtractLow(MachineInstr *MI, BitTracker::RegisterRef RD, 1783 const BitTracker::RegisterCell &RC, const RegisterSet &AVs); 1784 bool simplifyRCmp0(MachineInstr *MI, BitTracker::RegisterRef RD); 1785 1786 // Cache of created instructions to avoid creating duplicates. 1787 // XXX Currently only used by genBitSplit. 1788 std::vector<MachineInstr*> NewMIs; 1789 1790 const MachineDominatorTree &MDT; 1791 const HexagonInstrInfo &HII; 1792 const HexagonRegisterInfo &HRI; 1793 MachineRegisterInfo &MRI; 1794 MachineFunction &MF; 1795 BitTracker &BT; 1796 }; 1797 1798 } // end anonymous namespace 1799 1800 // Check if the bits [B..B+16) in register cell RC form a valid halfword, 1801 // i.e. [0..16), [16..32), etc. of some register. If so, return true and 1802 // set the information about the found register in RH. 1803 bool BitSimplification::matchHalf(unsigned SelfR, 1804 const BitTracker::RegisterCell &RC, unsigned B, RegHalf &RH) { 1805 // XXX This could be searching in the set of available registers, in case 1806 // the match is not exact. 1807 1808 // Match 16-bit chunks, where the RC[B..B+15] references exactly one 1809 // register and all the bits B..B+15 match between RC and the register. 1810 // This is meant to match "v1[0-15]", where v1 = { [0]:0 [1-15]:v1... }, 1811 // and RC = { [0]:0 [1-15]:v1[1-15]... }. 1812 bool Low = false; 1813 unsigned I = B; 1814 while (I < B+16 && RC[I].num()) 1815 I++; 1816 if (I == B+16) 1817 return false; 1818 1819 Register Reg = RC[I].RefI.Reg; 1820 unsigned P = RC[I].RefI.Pos; // The RefI.Pos will be advanced by I-B. 1821 if (P < I-B) 1822 return false; 1823 unsigned Pos = P - (I-B); 1824 1825 if (Reg == 0 || Reg == SelfR) // Don't match "self". 1826 return false; 1827 if (!Reg.isVirtual()) 1828 return false; 1829 if (!BT.has(Reg)) 1830 return false; 1831 1832 const BitTracker::RegisterCell &SC = BT.lookup(Reg); 1833 if (Pos+16 > SC.width()) 1834 return false; 1835 1836 for (unsigned i = 0; i < 16; ++i) { 1837 const BitTracker::BitValue &RV = RC[i+B]; 1838 if (RV.Type == BitTracker::BitValue::Ref) { 1839 if (RV.RefI.Reg != Reg) 1840 return false; 1841 if (RV.RefI.Pos != i+Pos) 1842 return false; 1843 continue; 1844 } 1845 if (RC[i+B] != SC[i+Pos]) 1846 return false; 1847 } 1848 1849 unsigned Sub = 0; 1850 switch (Pos) { 1851 case 0: 1852 Sub = Hexagon::isub_lo; 1853 Low = true; 1854 break; 1855 case 16: 1856 Sub = Hexagon::isub_lo; 1857 Low = false; 1858 break; 1859 case 32: 1860 Sub = Hexagon::isub_hi; 1861 Low = true; 1862 break; 1863 case 48: 1864 Sub = Hexagon::isub_hi; 1865 Low = false; 1866 break; 1867 default: 1868 return false; 1869 } 1870 1871 RH.Reg = Reg; 1872 RH.Sub = Sub; 1873 RH.Low = Low; 1874 // If the subregister is not valid with the register, set it to 0. 1875 if (!HBS::getFinalVRegClass(RH, MRI)) 1876 RH.Sub = 0; 1877 1878 return true; 1879 } 1880 1881 bool BitSimplification::validateReg(BitTracker::RegisterRef R, unsigned Opc, 1882 unsigned OpNum) { 1883 auto *OpRC = HII.getRegClass(HII.get(Opc), OpNum, &HRI, MF); 1884 auto *RRC = HBS::getFinalVRegClass(R, MRI); 1885 return OpRC->hasSubClassEq(RRC); 1886 } 1887 1888 // Check if RC matches the pattern of a S2_packhl. If so, return true and 1889 // set the inputs Rs and Rt. 1890 bool BitSimplification::matchPackhl(unsigned SelfR, 1891 const BitTracker::RegisterCell &RC, BitTracker::RegisterRef &Rs, 1892 BitTracker::RegisterRef &Rt) { 1893 RegHalf L1, H1, L2, H2; 1894 1895 if (!matchHalf(SelfR, RC, 0, L2) || !matchHalf(SelfR, RC, 16, L1)) 1896 return false; 1897 if (!matchHalf(SelfR, RC, 32, H2) || !matchHalf(SelfR, RC, 48, H1)) 1898 return false; 1899 1900 // Rs = H1.L1, Rt = H2.L2 1901 if (H1.Reg != L1.Reg || H1.Sub != L1.Sub || H1.Low || !L1.Low) 1902 return false; 1903 if (H2.Reg != L2.Reg || H2.Sub != L2.Sub || H2.Low || !L2.Low) 1904 return false; 1905 1906 Rs = H1; 1907 Rt = H2; 1908 return true; 1909 } 1910 1911 unsigned BitSimplification::getCombineOpcode(bool HLow, bool LLow) { 1912 return HLow ? LLow ? Hexagon::A2_combine_ll 1913 : Hexagon::A2_combine_lh 1914 : LLow ? Hexagon::A2_combine_hl 1915 : Hexagon::A2_combine_hh; 1916 } 1917 1918 // If MI stores the upper halfword of a register (potentially obtained via 1919 // shifts or extracts), replace it with a storerf instruction. This could 1920 // cause the "extraction" code to become dead. 1921 bool BitSimplification::genStoreUpperHalf(MachineInstr *MI) { 1922 unsigned Opc = MI->getOpcode(); 1923 if (Opc != Hexagon::S2_storerh_io) 1924 return false; 1925 1926 MachineOperand &ValOp = MI->getOperand(2); 1927 BitTracker::RegisterRef RS = ValOp; 1928 if (!BT.has(RS.Reg)) 1929 return false; 1930 const BitTracker::RegisterCell &RC = BT.lookup(RS.Reg); 1931 RegHalf H; 1932 if (!matchHalf(0, RC, 0, H)) 1933 return false; 1934 if (H.Low) 1935 return false; 1936 MI->setDesc(HII.get(Hexagon::S2_storerf_io)); 1937 ValOp.setReg(H.Reg); 1938 ValOp.setSubReg(H.Sub); 1939 return true; 1940 } 1941 1942 // If MI stores a value known at compile-time, and the value is within a range 1943 // that avoids using constant-extenders, replace it with a store-immediate. 1944 bool BitSimplification::genStoreImmediate(MachineInstr *MI) { 1945 unsigned Opc = MI->getOpcode(); 1946 unsigned Align = 0; 1947 switch (Opc) { 1948 case Hexagon::S2_storeri_io: 1949 Align++; 1950 LLVM_FALLTHROUGH; 1951 case Hexagon::S2_storerh_io: 1952 Align++; 1953 LLVM_FALLTHROUGH; 1954 case Hexagon::S2_storerb_io: 1955 break; 1956 default: 1957 return false; 1958 } 1959 1960 // Avoid stores to frame-indices (due to an unknown offset). 1961 if (!MI->getOperand(0).isReg()) 1962 return false; 1963 MachineOperand &OffOp = MI->getOperand(1); 1964 if (!OffOp.isImm()) 1965 return false; 1966 1967 int64_t Off = OffOp.getImm(); 1968 // Offset is u6:a. Sadly, there is no isShiftedUInt(n,x). 1969 if (!isUIntN(6+Align, Off) || (Off & ((1<<Align)-1))) 1970 return false; 1971 // Source register: 1972 BitTracker::RegisterRef RS = MI->getOperand(2); 1973 if (!BT.has(RS.Reg)) 1974 return false; 1975 const BitTracker::RegisterCell &RC = BT.lookup(RS.Reg); 1976 uint64_t U; 1977 if (!HBS::getConst(RC, 0, RC.width(), U)) 1978 return false; 1979 1980 // Only consider 8-bit values to avoid constant-extenders. 1981 int V; 1982 switch (Opc) { 1983 case Hexagon::S2_storerb_io: 1984 V = int8_t(U); 1985 break; 1986 case Hexagon::S2_storerh_io: 1987 V = int16_t(U); 1988 break; 1989 case Hexagon::S2_storeri_io: 1990 V = int32_t(U); 1991 break; 1992 default: 1993 // Opc is already checked above to be one of the three store instructions. 1994 // This silences a -Wuninitialized false positive on GCC 5.4. 1995 llvm_unreachable("Unexpected store opcode"); 1996 } 1997 if (!isInt<8>(V)) 1998 return false; 1999 2000 MI->RemoveOperand(2); 2001 switch (Opc) { 2002 case Hexagon::S2_storerb_io: 2003 MI->setDesc(HII.get(Hexagon::S4_storeirb_io)); 2004 break; 2005 case Hexagon::S2_storerh_io: 2006 MI->setDesc(HII.get(Hexagon::S4_storeirh_io)); 2007 break; 2008 case Hexagon::S2_storeri_io: 2009 MI->setDesc(HII.get(Hexagon::S4_storeiri_io)); 2010 break; 2011 } 2012 MI->addOperand(MachineOperand::CreateImm(V)); 2013 return true; 2014 } 2015 2016 // If MI is equivalent o S2_packhl, generate the S2_packhl. MI could be the 2017 // last instruction in a sequence that results in something equivalent to 2018 // the pack-halfwords. The intent is to cause the entire sequence to become 2019 // dead. 2020 bool BitSimplification::genPackhl(MachineInstr *MI, 2021 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { 2022 unsigned Opc = MI->getOpcode(); 2023 if (Opc == Hexagon::S2_packhl) 2024 return false; 2025 BitTracker::RegisterRef Rs, Rt; 2026 if (!matchPackhl(RD.Reg, RC, Rs, Rt)) 2027 return false; 2028 if (!validateReg(Rs, Hexagon::S2_packhl, 1) || 2029 !validateReg(Rt, Hexagon::S2_packhl, 2)) 2030 return false; 2031 2032 MachineBasicBlock &B = *MI->getParent(); 2033 Register NewR = MRI.createVirtualRegister(&Hexagon::DoubleRegsRegClass); 2034 DebugLoc DL = MI->getDebugLoc(); 2035 auto At = MI->isPHI() ? B.getFirstNonPHI() 2036 : MachineBasicBlock::iterator(MI); 2037 BuildMI(B, At, DL, HII.get(Hexagon::S2_packhl), NewR) 2038 .addReg(Rs.Reg, 0, Rs.Sub) 2039 .addReg(Rt.Reg, 0, Rt.Sub); 2040 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); 2041 BT.put(BitTracker::RegisterRef(NewR), RC); 2042 return true; 2043 } 2044 2045 // If MI produces halfword of the input in the low half of the output, 2046 // replace it with zero-extend or extractu. 2047 bool BitSimplification::genExtractHalf(MachineInstr *MI, 2048 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { 2049 RegHalf L; 2050 // Check for halfword in low 16 bits, zeros elsewhere. 2051 if (!matchHalf(RD.Reg, RC, 0, L) || !HBS::isZero(RC, 16, 16)) 2052 return false; 2053 2054 unsigned Opc = MI->getOpcode(); 2055 MachineBasicBlock &B = *MI->getParent(); 2056 DebugLoc DL = MI->getDebugLoc(); 2057 2058 // Prefer zxth, since zxth can go in any slot, while extractu only in 2059 // slots 2 and 3. 2060 unsigned NewR = 0; 2061 auto At = MI->isPHI() ? B.getFirstNonPHI() 2062 : MachineBasicBlock::iterator(MI); 2063 if (L.Low && Opc != Hexagon::A2_zxth) { 2064 if (validateReg(L, Hexagon::A2_zxth, 1)) { 2065 NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 2066 BuildMI(B, At, DL, HII.get(Hexagon::A2_zxth), NewR) 2067 .addReg(L.Reg, 0, L.Sub); 2068 } 2069 } else if (!L.Low && Opc != Hexagon::S2_lsr_i_r) { 2070 if (validateReg(L, Hexagon::S2_lsr_i_r, 1)) { 2071 NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 2072 BuildMI(B, MI, DL, HII.get(Hexagon::S2_lsr_i_r), NewR) 2073 .addReg(L.Reg, 0, L.Sub) 2074 .addImm(16); 2075 } 2076 } 2077 if (NewR == 0) 2078 return false; 2079 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); 2080 BT.put(BitTracker::RegisterRef(NewR), RC); 2081 return true; 2082 } 2083 2084 // If MI is equivalent to a combine(.L/.H, .L/.H) replace with with the 2085 // combine. 2086 bool BitSimplification::genCombineHalf(MachineInstr *MI, 2087 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { 2088 RegHalf L, H; 2089 // Check for combine h/l 2090 if (!matchHalf(RD.Reg, RC, 0, L) || !matchHalf(RD.Reg, RC, 16, H)) 2091 return false; 2092 // Do nothing if this is just a reg copy. 2093 if (L.Reg == H.Reg && L.Sub == H.Sub && !H.Low && L.Low) 2094 return false; 2095 2096 unsigned Opc = MI->getOpcode(); 2097 unsigned COpc = getCombineOpcode(H.Low, L.Low); 2098 if (COpc == Opc) 2099 return false; 2100 if (!validateReg(H, COpc, 1) || !validateReg(L, COpc, 2)) 2101 return false; 2102 2103 MachineBasicBlock &B = *MI->getParent(); 2104 DebugLoc DL = MI->getDebugLoc(); 2105 Register NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 2106 auto At = MI->isPHI() ? B.getFirstNonPHI() 2107 : MachineBasicBlock::iterator(MI); 2108 BuildMI(B, At, DL, HII.get(COpc), NewR) 2109 .addReg(H.Reg, 0, H.Sub) 2110 .addReg(L.Reg, 0, L.Sub); 2111 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); 2112 BT.put(BitTracker::RegisterRef(NewR), RC); 2113 return true; 2114 } 2115 2116 // If MI resets high bits of a register and keeps the lower ones, replace it 2117 // with zero-extend byte/half, and-immediate, or extractu, as appropriate. 2118 bool BitSimplification::genExtractLow(MachineInstr *MI, 2119 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { 2120 unsigned Opc = MI->getOpcode(); 2121 switch (Opc) { 2122 case Hexagon::A2_zxtb: 2123 case Hexagon::A2_zxth: 2124 case Hexagon::S2_extractu: 2125 return false; 2126 } 2127 if (Opc == Hexagon::A2_andir && MI->getOperand(2).isImm()) { 2128 int32_t Imm = MI->getOperand(2).getImm(); 2129 if (isInt<10>(Imm)) 2130 return false; 2131 } 2132 2133 if (MI->hasUnmodeledSideEffects() || MI->isInlineAsm()) 2134 return false; 2135 unsigned W = RC.width(); 2136 while (W > 0 && RC[W-1].is(0)) 2137 W--; 2138 if (W == 0 || W == RC.width()) 2139 return false; 2140 unsigned NewOpc = (W == 8) ? Hexagon::A2_zxtb 2141 : (W == 16) ? Hexagon::A2_zxth 2142 : (W < 10) ? Hexagon::A2_andir 2143 : Hexagon::S2_extractu; 2144 MachineBasicBlock &B = *MI->getParent(); 2145 DebugLoc DL = MI->getDebugLoc(); 2146 2147 for (auto &Op : MI->uses()) { 2148 if (!Op.isReg()) 2149 continue; 2150 BitTracker::RegisterRef RS = Op; 2151 if (!BT.has(RS.Reg)) 2152 continue; 2153 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); 2154 unsigned BN, BW; 2155 if (!HBS::getSubregMask(RS, BN, BW, MRI)) 2156 continue; 2157 if (BW < W || !HBS::isEqual(RC, 0, SC, BN, W)) 2158 continue; 2159 if (!validateReg(RS, NewOpc, 1)) 2160 continue; 2161 2162 Register NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 2163 auto At = MI->isPHI() ? B.getFirstNonPHI() 2164 : MachineBasicBlock::iterator(MI); 2165 auto MIB = BuildMI(B, At, DL, HII.get(NewOpc), NewR) 2166 .addReg(RS.Reg, 0, RS.Sub); 2167 if (NewOpc == Hexagon::A2_andir) 2168 MIB.addImm((1 << W) - 1); 2169 else if (NewOpc == Hexagon::S2_extractu) 2170 MIB.addImm(W).addImm(0); 2171 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); 2172 BT.put(BitTracker::RegisterRef(NewR), RC); 2173 return true; 2174 } 2175 return false; 2176 } 2177 2178 bool BitSimplification::genBitSplit(MachineInstr *MI, 2179 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC, 2180 const RegisterSet &AVs) { 2181 if (!GenBitSplit) 2182 return false; 2183 if (MaxBitSplit.getNumOccurrences()) { 2184 if (CountBitSplit >= MaxBitSplit) 2185 return false; 2186 } 2187 2188 unsigned Opc = MI->getOpcode(); 2189 switch (Opc) { 2190 case Hexagon::A4_bitsplit: 2191 case Hexagon::A4_bitspliti: 2192 return false; 2193 } 2194 2195 unsigned W = RC.width(); 2196 if (W != 32) 2197 return false; 2198 2199 auto ctlz = [] (const BitTracker::RegisterCell &C) -> unsigned { 2200 unsigned Z = C.width(); 2201 while (Z > 0 && C[Z-1].is(0)) 2202 --Z; 2203 return C.width() - Z; 2204 }; 2205 2206 // Count the number of leading zeros in the target RC. 2207 unsigned Z = ctlz(RC); 2208 if (Z == 0 || Z == W) 2209 return false; 2210 2211 // A simplistic analysis: assume the source register (the one being split) 2212 // is fully unknown, and that all its bits are self-references. 2213 const BitTracker::BitValue &B0 = RC[0]; 2214 if (B0.Type != BitTracker::BitValue::Ref) 2215 return false; 2216 2217 unsigned SrcR = B0.RefI.Reg; 2218 unsigned SrcSR = 0; 2219 unsigned Pos = B0.RefI.Pos; 2220 2221 // All the non-zero bits should be consecutive bits from the same register. 2222 for (unsigned i = 1; i < W-Z; ++i) { 2223 const BitTracker::BitValue &V = RC[i]; 2224 if (V.Type != BitTracker::BitValue::Ref) 2225 return false; 2226 if (V.RefI.Reg != SrcR || V.RefI.Pos != Pos+i) 2227 return false; 2228 } 2229 2230 // Now, find the other bitfield among AVs. 2231 for (unsigned S = AVs.find_first(); S; S = AVs.find_next(S)) { 2232 // The number of leading zeros here should be the number of trailing 2233 // non-zeros in RC. 2234 unsigned SRC = MRI.getRegClass(S)->getID(); 2235 if (SRC != Hexagon::IntRegsRegClassID && 2236 SRC != Hexagon::DoubleRegsRegClassID) 2237 continue; 2238 if (!BT.has(S)) 2239 continue; 2240 const BitTracker::RegisterCell &SC = BT.lookup(S); 2241 if (SC.width() != W || ctlz(SC) != W-Z) 2242 continue; 2243 // The Z lower bits should now match SrcR. 2244 const BitTracker::BitValue &S0 = SC[0]; 2245 if (S0.Type != BitTracker::BitValue::Ref || S0.RefI.Reg != SrcR) 2246 continue; 2247 unsigned P = S0.RefI.Pos; 2248 2249 if (Pos <= P && (Pos + W-Z) != P) 2250 continue; 2251 if (P < Pos && (P + Z) != Pos) 2252 continue; 2253 // The starting bitfield position must be at a subregister boundary. 2254 if (std::min(P, Pos) != 0 && std::min(P, Pos) != 32) 2255 continue; 2256 2257 unsigned I; 2258 for (I = 1; I < Z; ++I) { 2259 const BitTracker::BitValue &V = SC[I]; 2260 if (V.Type != BitTracker::BitValue::Ref) 2261 break; 2262 if (V.RefI.Reg != SrcR || V.RefI.Pos != P+I) 2263 break; 2264 } 2265 if (I != Z) 2266 continue; 2267 2268 // Generate bitsplit where S is defined. 2269 if (MaxBitSplit.getNumOccurrences()) 2270 CountBitSplit++; 2271 MachineInstr *DefS = MRI.getVRegDef(S); 2272 assert(DefS != nullptr); 2273 DebugLoc DL = DefS->getDebugLoc(); 2274 MachineBasicBlock &B = *DefS->getParent(); 2275 auto At = DefS->isPHI() ? B.getFirstNonPHI() 2276 : MachineBasicBlock::iterator(DefS); 2277 if (MRI.getRegClass(SrcR)->getID() == Hexagon::DoubleRegsRegClassID) 2278 SrcSR = (std::min(Pos, P) == 32) ? Hexagon::isub_hi : Hexagon::isub_lo; 2279 if (!validateReg({SrcR,SrcSR}, Hexagon::A4_bitspliti, 1)) 2280 continue; 2281 unsigned ImmOp = Pos <= P ? W-Z : Z; 2282 2283 // Find an existing bitsplit instruction if one already exists. 2284 unsigned NewR = 0; 2285 for (MachineInstr *In : NewMIs) { 2286 if (In->getOpcode() != Hexagon::A4_bitspliti) 2287 continue; 2288 MachineOperand &Op1 = In->getOperand(1); 2289 if (Op1.getReg() != SrcR || Op1.getSubReg() != SrcSR) 2290 continue; 2291 if (In->getOperand(2).getImm() != ImmOp) 2292 continue; 2293 // Check if the target register is available here. 2294 MachineOperand &Op0 = In->getOperand(0); 2295 MachineInstr *DefI = MRI.getVRegDef(Op0.getReg()); 2296 assert(DefI != nullptr); 2297 if (!MDT.dominates(DefI, &*At)) 2298 continue; 2299 2300 // Found one that can be reused. 2301 assert(Op0.getSubReg() == 0); 2302 NewR = Op0.getReg(); 2303 break; 2304 } 2305 if (!NewR) { 2306 NewR = MRI.createVirtualRegister(&Hexagon::DoubleRegsRegClass); 2307 auto NewBS = BuildMI(B, At, DL, HII.get(Hexagon::A4_bitspliti), NewR) 2308 .addReg(SrcR, 0, SrcSR) 2309 .addImm(ImmOp); 2310 NewMIs.push_back(NewBS); 2311 } 2312 if (Pos <= P) { 2313 HBS::replaceRegWithSub(RD.Reg, NewR, Hexagon::isub_lo, MRI); 2314 HBS::replaceRegWithSub(S, NewR, Hexagon::isub_hi, MRI); 2315 } else { 2316 HBS::replaceRegWithSub(S, NewR, Hexagon::isub_lo, MRI); 2317 HBS::replaceRegWithSub(RD.Reg, NewR, Hexagon::isub_hi, MRI); 2318 } 2319 return true; 2320 } 2321 2322 return false; 2323 } 2324 2325 // Check for tstbit simplification opportunity, where the bit being checked 2326 // can be tracked back to another register. For example: 2327 // %2 = S2_lsr_i_r %1, 5 2328 // %3 = S2_tstbit_i %2, 0 2329 // => 2330 // %3 = S2_tstbit_i %1, 5 2331 bool BitSimplification::simplifyTstbit(MachineInstr *MI, 2332 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { 2333 unsigned Opc = MI->getOpcode(); 2334 if (Opc != Hexagon::S2_tstbit_i) 2335 return false; 2336 2337 unsigned BN = MI->getOperand(2).getImm(); 2338 BitTracker::RegisterRef RS = MI->getOperand(1); 2339 unsigned F, W; 2340 DebugLoc DL = MI->getDebugLoc(); 2341 if (!BT.has(RS.Reg) || !HBS::getSubregMask(RS, F, W, MRI)) 2342 return false; 2343 MachineBasicBlock &B = *MI->getParent(); 2344 auto At = MI->isPHI() ? B.getFirstNonPHI() 2345 : MachineBasicBlock::iterator(MI); 2346 2347 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); 2348 const BitTracker::BitValue &V = SC[F+BN]; 2349 if (V.Type == BitTracker::BitValue::Ref && V.RefI.Reg != RS.Reg) { 2350 const TargetRegisterClass *TC = MRI.getRegClass(V.RefI.Reg); 2351 // Need to map V.RefI.Reg to a 32-bit register, i.e. if it is 2352 // a double register, need to use a subregister and adjust bit 2353 // number. 2354 unsigned P = std::numeric_limits<unsigned>::max(); 2355 BitTracker::RegisterRef RR(V.RefI.Reg, 0); 2356 if (TC == &Hexagon::DoubleRegsRegClass) { 2357 P = V.RefI.Pos; 2358 RR.Sub = Hexagon::isub_lo; 2359 if (P >= 32) { 2360 P -= 32; 2361 RR.Sub = Hexagon::isub_hi; 2362 } 2363 } else if (TC == &Hexagon::IntRegsRegClass) { 2364 P = V.RefI.Pos; 2365 } 2366 if (P != std::numeric_limits<unsigned>::max()) { 2367 Register NewR = MRI.createVirtualRegister(&Hexagon::PredRegsRegClass); 2368 BuildMI(B, At, DL, HII.get(Hexagon::S2_tstbit_i), NewR) 2369 .addReg(RR.Reg, 0, RR.Sub) 2370 .addImm(P); 2371 HBS::replaceReg(RD.Reg, NewR, MRI); 2372 BT.put(NewR, RC); 2373 return true; 2374 } 2375 } else if (V.is(0) || V.is(1)) { 2376 Register NewR = MRI.createVirtualRegister(&Hexagon::PredRegsRegClass); 2377 unsigned NewOpc = V.is(0) ? Hexagon::PS_false : Hexagon::PS_true; 2378 BuildMI(B, At, DL, HII.get(NewOpc), NewR); 2379 HBS::replaceReg(RD.Reg, NewR, MRI); 2380 return true; 2381 } 2382 2383 return false; 2384 } 2385 2386 // Detect whether RD is a bitfield extract (sign- or zero-extended) of 2387 // some register from the AVs set. Create a new corresponding instruction 2388 // at the location of MI. The intent is to recognize situations where 2389 // a sequence of instructions performs an operation that is equivalent to 2390 // an extract operation, such as a shift left followed by a shift right. 2391 bool BitSimplification::simplifyExtractLow(MachineInstr *MI, 2392 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC, 2393 const RegisterSet &AVs) { 2394 if (!GenExtract) 2395 return false; 2396 if (MaxExtract.getNumOccurrences()) { 2397 if (CountExtract >= MaxExtract) 2398 return false; 2399 CountExtract++; 2400 } 2401 2402 unsigned W = RC.width(); 2403 unsigned RW = W; 2404 unsigned Len; 2405 bool Signed; 2406 2407 // The code is mostly class-independent, except for the part that generates 2408 // the extract instruction, and establishes the source register (in case it 2409 // needs to use a subregister). 2410 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); 2411 if (FRC != &Hexagon::IntRegsRegClass && FRC != &Hexagon::DoubleRegsRegClass) 2412 return false; 2413 assert(RD.Sub == 0); 2414 2415 // Observation: 2416 // If the cell has a form of 00..0xx..x with k zeros and n remaining 2417 // bits, this could be an extractu of the n bits, but it could also be 2418 // an extractu of a longer field which happens to have 0s in the top 2419 // bit positions. 2420 // The same logic applies to sign-extended fields. 2421 // 2422 // Do not check for the extended extracts, since it would expand the 2423 // search space quite a bit. The search may be expensive as it is. 2424 2425 const BitTracker::BitValue &TopV = RC[W-1]; 2426 2427 // Eliminate candidates that have self-referential bits, since they 2428 // cannot be extracts from other registers. Also, skip registers that 2429 // have compile-time constant values. 2430 bool IsConst = true; 2431 for (unsigned I = 0; I != W; ++I) { 2432 const BitTracker::BitValue &V = RC[I]; 2433 if (V.Type == BitTracker::BitValue::Ref && V.RefI.Reg == RD.Reg) 2434 return false; 2435 IsConst = IsConst && (V.is(0) || V.is(1)); 2436 } 2437 if (IsConst) 2438 return false; 2439 2440 if (TopV.is(0) || TopV.is(1)) { 2441 bool S = TopV.is(1); 2442 for (--W; W > 0 && RC[W-1].is(S); --W) 2443 ; 2444 Len = W; 2445 Signed = S; 2446 // The sign bit must be a part of the field being extended. 2447 if (Signed) 2448 ++Len; 2449 } else { 2450 // This could still be a sign-extended extract. 2451 assert(TopV.Type == BitTracker::BitValue::Ref); 2452 if (TopV.RefI.Reg == RD.Reg || TopV.RefI.Pos == W-1) 2453 return false; 2454 for (--W; W > 0 && RC[W-1] == TopV; --W) 2455 ; 2456 // The top bits of RC are copies of TopV. One occurrence of TopV will 2457 // be a part of the field. 2458 Len = W + 1; 2459 Signed = true; 2460 } 2461 2462 // This would be just a copy. It should be handled elsewhere. 2463 if (Len == RW) 2464 return false; 2465 2466 LLVM_DEBUG({ 2467 dbgs() << __func__ << " on reg: " << printReg(RD.Reg, &HRI, RD.Sub) 2468 << ", MI: " << *MI; 2469 dbgs() << "Cell: " << RC << '\n'; 2470 dbgs() << "Expected bitfield size: " << Len << " bits, " 2471 << (Signed ? "sign" : "zero") << "-extended\n"; 2472 }); 2473 2474 bool Changed = false; 2475 2476 for (unsigned R = AVs.find_first(); R != 0; R = AVs.find_next(R)) { 2477 if (!BT.has(R)) 2478 continue; 2479 const BitTracker::RegisterCell &SC = BT.lookup(R); 2480 unsigned SW = SC.width(); 2481 2482 // The source can be longer than the destination, as long as its size is 2483 // a multiple of the size of the destination. Also, we would need to be 2484 // able to refer to the subregister in the source that would be of the 2485 // same size as the destination, but only check the sizes here. 2486 if (SW < RW || (SW % RW) != 0) 2487 continue; 2488 2489 // The field can start at any offset in SC as long as it contains Len 2490 // bits and does not cross subregister boundary (if the source register 2491 // is longer than the destination). 2492 unsigned Off = 0; 2493 while (Off <= SW-Len) { 2494 unsigned OE = (Off+Len)/RW; 2495 if (OE != Off/RW) { 2496 // The assumption here is that if the source (R) is longer than the 2497 // destination, then the destination is a sequence of words of 2498 // size RW, and each such word in R can be accessed via a subregister. 2499 // 2500 // If the beginning and the end of the field cross the subregister 2501 // boundary, advance to the next subregister. 2502 Off = OE*RW; 2503 continue; 2504 } 2505 if (HBS::isEqual(RC, 0, SC, Off, Len)) 2506 break; 2507 ++Off; 2508 } 2509 2510 if (Off > SW-Len) 2511 continue; 2512 2513 // Found match. 2514 unsigned ExtOpc = 0; 2515 if (Off == 0) { 2516 if (Len == 8) 2517 ExtOpc = Signed ? Hexagon::A2_sxtb : Hexagon::A2_zxtb; 2518 else if (Len == 16) 2519 ExtOpc = Signed ? Hexagon::A2_sxth : Hexagon::A2_zxth; 2520 else if (Len < 10 && !Signed) 2521 ExtOpc = Hexagon::A2_andir; 2522 } 2523 if (ExtOpc == 0) { 2524 ExtOpc = 2525 Signed ? (RW == 32 ? Hexagon::S4_extract : Hexagon::S4_extractp) 2526 : (RW == 32 ? Hexagon::S2_extractu : Hexagon::S2_extractup); 2527 } 2528 unsigned SR = 0; 2529 // This only recognizes isub_lo and isub_hi. 2530 if (RW != SW && RW*2 != SW) 2531 continue; 2532 if (RW != SW) 2533 SR = (Off/RW == 0) ? Hexagon::isub_lo : Hexagon::isub_hi; 2534 Off = Off % RW; 2535 2536 if (!validateReg({R,SR}, ExtOpc, 1)) 2537 continue; 2538 2539 // Don't generate the same instruction as the one being optimized. 2540 if (MI->getOpcode() == ExtOpc) { 2541 // All possible ExtOpc's have the source in operand(1). 2542 const MachineOperand &SrcOp = MI->getOperand(1); 2543 if (SrcOp.getReg() == R) 2544 continue; 2545 } 2546 2547 DebugLoc DL = MI->getDebugLoc(); 2548 MachineBasicBlock &B = *MI->getParent(); 2549 Register NewR = MRI.createVirtualRegister(FRC); 2550 auto At = MI->isPHI() ? B.getFirstNonPHI() 2551 : MachineBasicBlock::iterator(MI); 2552 auto MIB = BuildMI(B, At, DL, HII.get(ExtOpc), NewR) 2553 .addReg(R, 0, SR); 2554 switch (ExtOpc) { 2555 case Hexagon::A2_sxtb: 2556 case Hexagon::A2_zxtb: 2557 case Hexagon::A2_sxth: 2558 case Hexagon::A2_zxth: 2559 break; 2560 case Hexagon::A2_andir: 2561 MIB.addImm((1u << Len) - 1); 2562 break; 2563 case Hexagon::S4_extract: 2564 case Hexagon::S2_extractu: 2565 case Hexagon::S4_extractp: 2566 case Hexagon::S2_extractup: 2567 MIB.addImm(Len) 2568 .addImm(Off); 2569 break; 2570 default: 2571 llvm_unreachable("Unexpected opcode"); 2572 } 2573 2574 HBS::replaceReg(RD.Reg, NewR, MRI); 2575 BT.put(BitTracker::RegisterRef(NewR), RC); 2576 Changed = true; 2577 break; 2578 } 2579 2580 return Changed; 2581 } 2582 2583 bool BitSimplification::simplifyRCmp0(MachineInstr *MI, 2584 BitTracker::RegisterRef RD) { 2585 unsigned Opc = MI->getOpcode(); 2586 if (Opc != Hexagon::A4_rcmpeqi && Opc != Hexagon::A4_rcmpneqi) 2587 return false; 2588 MachineOperand &CmpOp = MI->getOperand(2); 2589 if (!CmpOp.isImm() || CmpOp.getImm() != 0) 2590 return false; 2591 2592 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); 2593 if (FRC != &Hexagon::IntRegsRegClass && FRC != &Hexagon::DoubleRegsRegClass) 2594 return false; 2595 assert(RD.Sub == 0); 2596 2597 MachineBasicBlock &B = *MI->getParent(); 2598 const DebugLoc &DL = MI->getDebugLoc(); 2599 auto At = MI->isPHI() ? B.getFirstNonPHI() 2600 : MachineBasicBlock::iterator(MI); 2601 bool KnownZ = true; 2602 bool KnownNZ = false; 2603 2604 BitTracker::RegisterRef SR = MI->getOperand(1); 2605 if (!BT.has(SR.Reg)) 2606 return false; 2607 const BitTracker::RegisterCell &SC = BT.lookup(SR.Reg); 2608 unsigned F, W; 2609 if (!HBS::getSubregMask(SR, F, W, MRI)) 2610 return false; 2611 2612 for (uint16_t I = F; I != F+W; ++I) { 2613 const BitTracker::BitValue &V = SC[I]; 2614 if (!V.is(0)) 2615 KnownZ = false; 2616 if (V.is(1)) 2617 KnownNZ = true; 2618 } 2619 2620 auto ReplaceWithConst = [&](int C) { 2621 Register NewR = MRI.createVirtualRegister(FRC); 2622 BuildMI(B, At, DL, HII.get(Hexagon::A2_tfrsi), NewR) 2623 .addImm(C); 2624 HBS::replaceReg(RD.Reg, NewR, MRI); 2625 BitTracker::RegisterCell NewRC(W); 2626 for (uint16_t I = 0; I != W; ++I) { 2627 NewRC[I] = BitTracker::BitValue(C & 1); 2628 C = unsigned(C) >> 1; 2629 } 2630 BT.put(BitTracker::RegisterRef(NewR), NewRC); 2631 return true; 2632 }; 2633 2634 auto IsNonZero = [] (const MachineOperand &Op) { 2635 if (Op.isGlobal() || Op.isBlockAddress()) 2636 return true; 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 auto IsZero = [] (const MachineOperand &Op) { 2647 if (Op.isGlobal() || Op.isBlockAddress()) 2648 return false; 2649 if (Op.isImm()) 2650 return Op.getImm() == 0; 2651 if (Op.isCImm()) 2652 return Op.getCImm()->isZero(); 2653 if (Op.isFPImm()) 2654 return Op.getFPImm()->isZero(); 2655 return false; 2656 }; 2657 2658 // If the source register is known to be 0 or non-0, the comparison can 2659 // be folded to a load of a constant. 2660 if (KnownZ || KnownNZ) { 2661 assert(KnownZ != KnownNZ && "Register cannot be both 0 and non-0"); 2662 return ReplaceWithConst(KnownZ == (Opc == Hexagon::A4_rcmpeqi)); 2663 } 2664 2665 // Special case: if the compare comes from a C2_muxii, then we know the 2666 // two possible constants that can be the source value. 2667 MachineInstr *InpDef = MRI.getVRegDef(SR.Reg); 2668 if (!InpDef) 2669 return false; 2670 if (SR.Sub == 0 && InpDef->getOpcode() == Hexagon::C2_muxii) { 2671 MachineOperand &Src1 = InpDef->getOperand(2); 2672 MachineOperand &Src2 = InpDef->getOperand(3); 2673 // Check if both are non-zero. 2674 bool KnownNZ1 = IsNonZero(Src1), KnownNZ2 = IsNonZero(Src2); 2675 if (KnownNZ1 && KnownNZ2) 2676 return ReplaceWithConst(Opc == Hexagon::A4_rcmpneqi); 2677 // Check if both are zero. 2678 bool KnownZ1 = IsZero(Src1), KnownZ2 = IsZero(Src2); 2679 if (KnownZ1 && KnownZ2) 2680 return ReplaceWithConst(Opc == Hexagon::A4_rcmpeqi); 2681 2682 // If for both operands we know that they are either 0 or non-0, 2683 // replace the comparison with a C2_muxii, using the same predicate 2684 // register, but with operands substituted with 0/1 accordingly. 2685 if ((KnownZ1 || KnownNZ1) && (KnownZ2 || KnownNZ2)) { 2686 Register NewR = MRI.createVirtualRegister(FRC); 2687 BuildMI(B, At, DL, HII.get(Hexagon::C2_muxii), NewR) 2688 .addReg(InpDef->getOperand(1).getReg()) 2689 .addImm(KnownZ1 == (Opc == Hexagon::A4_rcmpeqi)) 2690 .addImm(KnownZ2 == (Opc == Hexagon::A4_rcmpeqi)); 2691 HBS::replaceReg(RD.Reg, NewR, MRI); 2692 // Create a new cell with only the least significant bit unknown. 2693 BitTracker::RegisterCell NewRC(W); 2694 NewRC[0] = BitTracker::BitValue::self(); 2695 NewRC.fill(1, W, BitTracker::BitValue::Zero); 2696 BT.put(BitTracker::RegisterRef(NewR), NewRC); 2697 return true; 2698 } 2699 } 2700 2701 return false; 2702 } 2703 2704 bool BitSimplification::processBlock(MachineBasicBlock &B, 2705 const RegisterSet &AVs) { 2706 if (!BT.reached(&B)) 2707 return false; 2708 bool Changed = false; 2709 RegisterSet AVB = AVs; 2710 RegisterSet Defs; 2711 2712 for (auto I = B.begin(), E = B.end(); I != E; ++I, AVB.insert(Defs)) { 2713 MachineInstr *MI = &*I; 2714 Defs.clear(); 2715 HBS::getInstrDefs(*MI, Defs); 2716 2717 unsigned Opc = MI->getOpcode(); 2718 if (Opc == TargetOpcode::COPY || Opc == TargetOpcode::REG_SEQUENCE) 2719 continue; 2720 2721 if (MI->mayStore()) { 2722 bool T = genStoreUpperHalf(MI); 2723 T = T || genStoreImmediate(MI); 2724 Changed |= T; 2725 continue; 2726 } 2727 2728 if (Defs.count() != 1) 2729 continue; 2730 const MachineOperand &Op0 = MI->getOperand(0); 2731 if (!Op0.isReg() || !Op0.isDef()) 2732 continue; 2733 BitTracker::RegisterRef RD = Op0; 2734 if (!BT.has(RD.Reg)) 2735 continue; 2736 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); 2737 const BitTracker::RegisterCell &RC = BT.lookup(RD.Reg); 2738 2739 if (FRC->getID() == Hexagon::DoubleRegsRegClassID) { 2740 bool T = genPackhl(MI, RD, RC); 2741 T = T || simplifyExtractLow(MI, RD, RC, AVB); 2742 Changed |= T; 2743 continue; 2744 } 2745 2746 if (FRC->getID() == Hexagon::IntRegsRegClassID) { 2747 bool T = genBitSplit(MI, RD, RC, AVB); 2748 T = T || simplifyExtractLow(MI, RD, RC, AVB); 2749 T = T || genExtractHalf(MI, RD, RC); 2750 T = T || genCombineHalf(MI, RD, RC); 2751 T = T || genExtractLow(MI, RD, RC); 2752 T = T || simplifyRCmp0(MI, RD); 2753 Changed |= T; 2754 continue; 2755 } 2756 2757 if (FRC->getID() == Hexagon::PredRegsRegClassID) { 2758 bool T = simplifyTstbit(MI, RD, RC); 2759 Changed |= T; 2760 continue; 2761 } 2762 } 2763 return Changed; 2764 } 2765 2766 bool HexagonBitSimplify::runOnMachineFunction(MachineFunction &MF) { 2767 if (skipFunction(MF.getFunction())) 2768 return false; 2769 2770 auto &HST = MF.getSubtarget<HexagonSubtarget>(); 2771 auto &HRI = *HST.getRegisterInfo(); 2772 auto &HII = *HST.getInstrInfo(); 2773 2774 MDT = &getAnalysis<MachineDominatorTree>(); 2775 MachineRegisterInfo &MRI = MF.getRegInfo(); 2776 bool Changed; 2777 2778 Changed = DeadCodeElimination(MF, *MDT).run(); 2779 2780 const HexagonEvaluator HE(HRI, MRI, HII, MF); 2781 BitTracker BT(HE, MF); 2782 LLVM_DEBUG(BT.trace(true)); 2783 BT.run(); 2784 2785 MachineBasicBlock &Entry = MF.front(); 2786 2787 RegisterSet AIG; // Available registers for IG. 2788 ConstGeneration ImmG(BT, HII, MRI); 2789 Changed |= visitBlock(Entry, ImmG, AIG); 2790 2791 RegisterSet ARE; // Available registers for RIE. 2792 RedundantInstrElimination RIE(BT, HII, HRI, MRI); 2793 bool Ried = visitBlock(Entry, RIE, ARE); 2794 if (Ried) { 2795 Changed = true; 2796 BT.run(); 2797 } 2798 2799 RegisterSet ACG; // Available registers for CG. 2800 CopyGeneration CopyG(BT, HII, HRI, MRI); 2801 Changed |= visitBlock(Entry, CopyG, ACG); 2802 2803 RegisterSet ACP; // Available registers for CP. 2804 CopyPropagation CopyP(HRI, MRI); 2805 Changed |= visitBlock(Entry, CopyP, ACP); 2806 2807 Changed = DeadCodeElimination(MF, *MDT).run() || Changed; 2808 2809 BT.run(); 2810 RegisterSet ABS; // Available registers for BS. 2811 BitSimplification BitS(BT, *MDT, HII, HRI, MRI, MF); 2812 Changed |= visitBlock(Entry, BitS, ABS); 2813 2814 Changed = DeadCodeElimination(MF, *MDT).run() || Changed; 2815 2816 if (Changed) { 2817 for (auto &B : MF) 2818 for (auto &I : B) 2819 I.clearKillInfo(); 2820 DeadCodeElimination(MF, *MDT).run(); 2821 } 2822 return Changed; 2823 } 2824 2825 // Recognize loops where the code at the end of the loop matches the code 2826 // before the entry of the loop, and the matching code is such that is can 2827 // be simplified. This pass relies on the bit simplification above and only 2828 // prepares code in a way that can be handled by the bit simplifcation. 2829 // 2830 // This is the motivating testcase (and explanation): 2831 // 2832 // { 2833 // loop0(.LBB0_2, r1) // %for.body.preheader 2834 // r5:4 = memd(r0++#8) 2835 // } 2836 // { 2837 // r3 = lsr(r4, #16) 2838 // r7:6 = combine(r5, r5) 2839 // } 2840 // { 2841 // r3 = insert(r5, #16, #16) 2842 // r7:6 = vlsrw(r7:6, #16) 2843 // } 2844 // .LBB0_2: 2845 // { 2846 // memh(r2+#4) = r5 2847 // memh(r2+#6) = r6 # R6 is really R5.H 2848 // } 2849 // { 2850 // r2 = add(r2, #8) 2851 // memh(r2+#0) = r4 2852 // memh(r2+#2) = r3 # R3 is really R4.H 2853 // } 2854 // { 2855 // r5:4 = memd(r0++#8) 2856 // } 2857 // { # "Shuffling" code that sets up R3 and R6 2858 // r3 = lsr(r4, #16) # so that their halves can be stored in the 2859 // r7:6 = combine(r5, r5) # next iteration. This could be folded into 2860 // } # the stores if the code was at the beginning 2861 // { # of the loop iteration. Since the same code 2862 // r3 = insert(r5, #16, #16) # precedes the loop, it can actually be moved 2863 // r7:6 = vlsrw(r7:6, #16) # there. 2864 // }:endloop0 2865 // 2866 // 2867 // The outcome: 2868 // 2869 // { 2870 // loop0(.LBB0_2, r1) 2871 // r5:4 = memd(r0++#8) 2872 // } 2873 // .LBB0_2: 2874 // { 2875 // memh(r2+#4) = r5 2876 // memh(r2+#6) = r5.h 2877 // } 2878 // { 2879 // r2 = add(r2, #8) 2880 // memh(r2+#0) = r4 2881 // memh(r2+#2) = r4.h 2882 // } 2883 // { 2884 // r5:4 = memd(r0++#8) 2885 // }:endloop0 2886 2887 namespace llvm { 2888 2889 FunctionPass *createHexagonLoopRescheduling(); 2890 void initializeHexagonLoopReschedulingPass(PassRegistry&); 2891 2892 } // end namespace llvm 2893 2894 namespace { 2895 2896 class HexagonLoopRescheduling : public MachineFunctionPass { 2897 public: 2898 static char ID; 2899 2900 HexagonLoopRescheduling() : MachineFunctionPass(ID) { 2901 initializeHexagonLoopReschedulingPass(*PassRegistry::getPassRegistry()); 2902 } 2903 2904 bool runOnMachineFunction(MachineFunction &MF) override; 2905 2906 private: 2907 const HexagonInstrInfo *HII = nullptr; 2908 const HexagonRegisterInfo *HRI = nullptr; 2909 MachineRegisterInfo *MRI = nullptr; 2910 BitTracker *BTP = nullptr; 2911 2912 struct LoopCand { 2913 LoopCand(MachineBasicBlock *lb, MachineBasicBlock *pb, 2914 MachineBasicBlock *eb) : LB(lb), PB(pb), EB(eb) {} 2915 2916 MachineBasicBlock *LB, *PB, *EB; 2917 }; 2918 using InstrList = std::vector<MachineInstr *>; 2919 struct InstrGroup { 2920 BitTracker::RegisterRef Inp, Out; 2921 InstrList Ins; 2922 }; 2923 struct PhiInfo { 2924 PhiInfo(MachineInstr &P, MachineBasicBlock &B); 2925 2926 unsigned DefR; 2927 BitTracker::RegisterRef LR, PR; // Loop Register, Preheader Register 2928 MachineBasicBlock *LB, *PB; // Loop Block, Preheader Block 2929 }; 2930 2931 static unsigned getDefReg(const MachineInstr *MI); 2932 bool isConst(unsigned Reg) const; 2933 bool isBitShuffle(const MachineInstr *MI, unsigned DefR) const; 2934 bool isStoreInput(const MachineInstr *MI, unsigned DefR) const; 2935 bool isShuffleOf(unsigned OutR, unsigned InpR) const; 2936 bool isSameShuffle(unsigned OutR1, unsigned InpR1, unsigned OutR2, 2937 unsigned &InpR2) const; 2938 void moveGroup(InstrGroup &G, MachineBasicBlock &LB, MachineBasicBlock &PB, 2939 MachineBasicBlock::iterator At, unsigned OldPhiR, unsigned NewPredR); 2940 bool processLoop(LoopCand &C); 2941 }; 2942 2943 } // end anonymous namespace 2944 2945 char HexagonLoopRescheduling::ID = 0; 2946 2947 INITIALIZE_PASS(HexagonLoopRescheduling, "hexagon-loop-resched", 2948 "Hexagon Loop Rescheduling", false, false) 2949 2950 HexagonLoopRescheduling::PhiInfo::PhiInfo(MachineInstr &P, 2951 MachineBasicBlock &B) { 2952 DefR = HexagonLoopRescheduling::getDefReg(&P); 2953 LB = &B; 2954 PB = nullptr; 2955 for (unsigned i = 1, n = P.getNumOperands(); i < n; i += 2) { 2956 const MachineOperand &OpB = P.getOperand(i+1); 2957 if (OpB.getMBB() == &B) { 2958 LR = P.getOperand(i); 2959 continue; 2960 } 2961 PB = OpB.getMBB(); 2962 PR = P.getOperand(i); 2963 } 2964 } 2965 2966 unsigned HexagonLoopRescheduling::getDefReg(const MachineInstr *MI) { 2967 RegisterSet Defs; 2968 HBS::getInstrDefs(*MI, Defs); 2969 if (Defs.count() != 1) 2970 return 0; 2971 return Defs.find_first(); 2972 } 2973 2974 bool HexagonLoopRescheduling::isConst(unsigned Reg) const { 2975 if (!BTP->has(Reg)) 2976 return false; 2977 const BitTracker::RegisterCell &RC = BTP->lookup(Reg); 2978 for (unsigned i = 0, w = RC.width(); i < w; ++i) { 2979 const BitTracker::BitValue &V = RC[i]; 2980 if (!V.is(0) && !V.is(1)) 2981 return false; 2982 } 2983 return true; 2984 } 2985 2986 bool HexagonLoopRescheduling::isBitShuffle(const MachineInstr *MI, 2987 unsigned DefR) const { 2988 unsigned Opc = MI->getOpcode(); 2989 switch (Opc) { 2990 case TargetOpcode::COPY: 2991 case Hexagon::S2_lsr_i_r: 2992 case Hexagon::S2_asr_i_r: 2993 case Hexagon::S2_asl_i_r: 2994 case Hexagon::S2_lsr_i_p: 2995 case Hexagon::S2_asr_i_p: 2996 case Hexagon::S2_asl_i_p: 2997 case Hexagon::S2_insert: 2998 case Hexagon::A2_or: 2999 case Hexagon::A2_orp: 3000 case Hexagon::A2_and: 3001 case Hexagon::A2_andp: 3002 case Hexagon::A2_combinew: 3003 case Hexagon::A4_combineri: 3004 case Hexagon::A4_combineir: 3005 case Hexagon::A2_combineii: 3006 case Hexagon::A4_combineii: 3007 case Hexagon::A2_combine_ll: 3008 case Hexagon::A2_combine_lh: 3009 case Hexagon::A2_combine_hl: 3010 case Hexagon::A2_combine_hh: 3011 return true; 3012 } 3013 return false; 3014 } 3015 3016 bool HexagonLoopRescheduling::isStoreInput(const MachineInstr *MI, 3017 unsigned InpR) const { 3018 for (unsigned i = 0, n = MI->getNumOperands(); i < n; ++i) { 3019 const MachineOperand &Op = MI->getOperand(i); 3020 if (!Op.isReg()) 3021 continue; 3022 if (Op.getReg() == InpR) 3023 return i == n-1; 3024 } 3025 return false; 3026 } 3027 3028 bool HexagonLoopRescheduling::isShuffleOf(unsigned OutR, unsigned InpR) const { 3029 if (!BTP->has(OutR) || !BTP->has(InpR)) 3030 return false; 3031 const BitTracker::RegisterCell &OutC = BTP->lookup(OutR); 3032 for (unsigned i = 0, w = OutC.width(); i < w; ++i) { 3033 const BitTracker::BitValue &V = OutC[i]; 3034 if (V.Type != BitTracker::BitValue::Ref) 3035 continue; 3036 if (V.RefI.Reg != InpR) 3037 return false; 3038 } 3039 return true; 3040 } 3041 3042 bool HexagonLoopRescheduling::isSameShuffle(unsigned OutR1, unsigned InpR1, 3043 unsigned OutR2, unsigned &InpR2) const { 3044 if (!BTP->has(OutR1) || !BTP->has(InpR1) || !BTP->has(OutR2)) 3045 return false; 3046 const BitTracker::RegisterCell &OutC1 = BTP->lookup(OutR1); 3047 const BitTracker::RegisterCell &OutC2 = BTP->lookup(OutR2); 3048 unsigned W = OutC1.width(); 3049 unsigned MatchR = 0; 3050 if (W != OutC2.width()) 3051 return false; 3052 for (unsigned i = 0; i < W; ++i) { 3053 const BitTracker::BitValue &V1 = OutC1[i], &V2 = OutC2[i]; 3054 if (V1.Type != V2.Type || V1.Type == BitTracker::BitValue::One) 3055 return false; 3056 if (V1.Type != BitTracker::BitValue::Ref) 3057 continue; 3058 if (V1.RefI.Pos != V2.RefI.Pos) 3059 return false; 3060 if (V1.RefI.Reg != InpR1) 3061 return false; 3062 if (V2.RefI.Reg == 0 || V2.RefI.Reg == OutR2) 3063 return false; 3064 if (!MatchR) 3065 MatchR = V2.RefI.Reg; 3066 else if (V2.RefI.Reg != MatchR) 3067 return false; 3068 } 3069 InpR2 = MatchR; 3070 return true; 3071 } 3072 3073 void HexagonLoopRescheduling::moveGroup(InstrGroup &G, MachineBasicBlock &LB, 3074 MachineBasicBlock &PB, MachineBasicBlock::iterator At, unsigned OldPhiR, 3075 unsigned NewPredR) { 3076 DenseMap<unsigned,unsigned> RegMap; 3077 3078 const TargetRegisterClass *PhiRC = MRI->getRegClass(NewPredR); 3079 Register PhiR = MRI->createVirtualRegister(PhiRC); 3080 BuildMI(LB, At, At->getDebugLoc(), HII->get(TargetOpcode::PHI), PhiR) 3081 .addReg(NewPredR) 3082 .addMBB(&PB) 3083 .addReg(G.Inp.Reg) 3084 .addMBB(&LB); 3085 RegMap.insert(std::make_pair(G.Inp.Reg, PhiR)); 3086 3087 for (unsigned i = G.Ins.size(); i > 0; --i) { 3088 const MachineInstr *SI = G.Ins[i-1]; 3089 unsigned DR = getDefReg(SI); 3090 const TargetRegisterClass *RC = MRI->getRegClass(DR); 3091 Register NewDR = MRI->createVirtualRegister(RC); 3092 DebugLoc DL = SI->getDebugLoc(); 3093 3094 auto MIB = BuildMI(LB, At, DL, HII->get(SI->getOpcode()), NewDR); 3095 for (unsigned j = 0, m = SI->getNumOperands(); j < m; ++j) { 3096 const MachineOperand &Op = SI->getOperand(j); 3097 if (!Op.isReg()) { 3098 MIB.add(Op); 3099 continue; 3100 } 3101 if (!Op.isUse()) 3102 continue; 3103 unsigned UseR = RegMap[Op.getReg()]; 3104 MIB.addReg(UseR, 0, Op.getSubReg()); 3105 } 3106 RegMap.insert(std::make_pair(DR, NewDR)); 3107 } 3108 3109 HBS::replaceReg(OldPhiR, RegMap[G.Out.Reg], *MRI); 3110 } 3111 3112 bool HexagonLoopRescheduling::processLoop(LoopCand &C) { 3113 LLVM_DEBUG(dbgs() << "Processing loop in " << printMBBReference(*C.LB) 3114 << "\n"); 3115 std::vector<PhiInfo> Phis; 3116 for (auto &I : *C.LB) { 3117 if (!I.isPHI()) 3118 break; 3119 unsigned PR = getDefReg(&I); 3120 if (isConst(PR)) 3121 continue; 3122 bool BadUse = false, GoodUse = false; 3123 for (const MachineOperand &MO : MRI->use_operands(PR)) { 3124 const MachineInstr *UseI = MO.getParent(); 3125 if (UseI->getParent() != C.LB) { 3126 BadUse = true; 3127 break; 3128 } 3129 if (isBitShuffle(UseI, PR) || isStoreInput(UseI, PR)) 3130 GoodUse = true; 3131 } 3132 if (BadUse || !GoodUse) 3133 continue; 3134 3135 Phis.push_back(PhiInfo(I, *C.LB)); 3136 } 3137 3138 LLVM_DEBUG({ 3139 dbgs() << "Phis: {"; 3140 for (auto &I : Phis) { 3141 dbgs() << ' ' << printReg(I.DefR, HRI) << "=phi(" 3142 << printReg(I.PR.Reg, HRI, I.PR.Sub) << ":b" << I.PB->getNumber() 3143 << ',' << printReg(I.LR.Reg, HRI, I.LR.Sub) << ":b" 3144 << I.LB->getNumber() << ')'; 3145 } 3146 dbgs() << " }\n"; 3147 }); 3148 3149 if (Phis.empty()) 3150 return false; 3151 3152 bool Changed = false; 3153 InstrList ShufIns; 3154 3155 // Go backwards in the block: for each bit shuffling instruction, check 3156 // if that instruction could potentially be moved to the front of the loop: 3157 // the output of the loop cannot be used in a non-shuffling instruction 3158 // in this loop. 3159 for (auto I = C.LB->rbegin(), E = C.LB->rend(); I != E; ++I) { 3160 if (I->isTerminator()) 3161 continue; 3162 if (I->isPHI()) 3163 break; 3164 3165 RegisterSet Defs; 3166 HBS::getInstrDefs(*I, Defs); 3167 if (Defs.count() != 1) 3168 continue; 3169 Register DefR = Defs.find_first(); 3170 if (!DefR.isVirtual()) 3171 continue; 3172 if (!isBitShuffle(&*I, DefR)) 3173 continue; 3174 3175 bool BadUse = false; 3176 for (auto UI = MRI->use_begin(DefR), UE = MRI->use_end(); UI != UE; ++UI) { 3177 MachineInstr *UseI = UI->getParent(); 3178 if (UseI->getParent() == C.LB) { 3179 if (UseI->isPHI()) { 3180 // If the use is in a phi node in this loop, then it should be 3181 // the value corresponding to the back edge. 3182 unsigned Idx = UI.getOperandNo(); 3183 if (UseI->getOperand(Idx+1).getMBB() != C.LB) 3184 BadUse = true; 3185 } else { 3186 auto F = find(ShufIns, UseI); 3187 if (F == ShufIns.end()) 3188 BadUse = true; 3189 } 3190 } else { 3191 // There is a use outside of the loop, but there is no epilog block 3192 // suitable for a copy-out. 3193 if (C.EB == nullptr) 3194 BadUse = true; 3195 } 3196 if (BadUse) 3197 break; 3198 } 3199 3200 if (BadUse) 3201 continue; 3202 ShufIns.push_back(&*I); 3203 } 3204 3205 // Partition the list of shuffling instructions into instruction groups, 3206 // where each group has to be moved as a whole (i.e. a group is a chain of 3207 // dependent instructions). A group produces a single live output register, 3208 // which is meant to be the input of the loop phi node (although this is 3209 // not checked here yet). It also uses a single register as its input, 3210 // which is some value produced in the loop body. After moving the group 3211 // to the beginning of the loop, that input register would need to be 3212 // the loop-carried register (through a phi node) instead of the (currently 3213 // loop-carried) output register. 3214 using InstrGroupList = std::vector<InstrGroup>; 3215 InstrGroupList Groups; 3216 3217 for (unsigned i = 0, n = ShufIns.size(); i < n; ++i) { 3218 MachineInstr *SI = ShufIns[i]; 3219 if (SI == nullptr) 3220 continue; 3221 3222 InstrGroup G; 3223 G.Ins.push_back(SI); 3224 G.Out.Reg = getDefReg(SI); 3225 RegisterSet Inputs; 3226 HBS::getInstrUses(*SI, Inputs); 3227 3228 for (unsigned j = i+1; j < n; ++j) { 3229 MachineInstr *MI = ShufIns[j]; 3230 if (MI == nullptr) 3231 continue; 3232 RegisterSet Defs; 3233 HBS::getInstrDefs(*MI, Defs); 3234 // If this instruction does not define any pending inputs, skip it. 3235 if (!Defs.intersects(Inputs)) 3236 continue; 3237 // Otherwise, add it to the current group and remove the inputs that 3238 // are defined by MI. 3239 G.Ins.push_back(MI); 3240 Inputs.remove(Defs); 3241 // Then add all registers used by MI. 3242 HBS::getInstrUses(*MI, Inputs); 3243 ShufIns[j] = nullptr; 3244 } 3245 3246 // Only add a group if it requires at most one register. 3247 if (Inputs.count() > 1) 3248 continue; 3249 auto LoopInpEq = [G] (const PhiInfo &P) -> bool { 3250 return G.Out.Reg == P.LR.Reg; 3251 }; 3252 if (llvm::none_of(Phis, LoopInpEq)) 3253 continue; 3254 3255 G.Inp.Reg = Inputs.find_first(); 3256 Groups.push_back(G); 3257 } 3258 3259 LLVM_DEBUG({ 3260 for (unsigned i = 0, n = Groups.size(); i < n; ++i) { 3261 InstrGroup &G = Groups[i]; 3262 dbgs() << "Group[" << i << "] inp: " 3263 << printReg(G.Inp.Reg, HRI, G.Inp.Sub) 3264 << " out: " << printReg(G.Out.Reg, HRI, G.Out.Sub) << "\n"; 3265 for (unsigned j = 0, m = G.Ins.size(); j < m; ++j) 3266 dbgs() << " " << *G.Ins[j]; 3267 } 3268 }); 3269 3270 for (unsigned i = 0, n = Groups.size(); i < n; ++i) { 3271 InstrGroup &G = Groups[i]; 3272 if (!isShuffleOf(G.Out.Reg, G.Inp.Reg)) 3273 continue; 3274 auto LoopInpEq = [G] (const PhiInfo &P) -> bool { 3275 return G.Out.Reg == P.LR.Reg; 3276 }; 3277 auto F = llvm::find_if(Phis, LoopInpEq); 3278 if (F == Phis.end()) 3279 continue; 3280 unsigned PrehR = 0; 3281 if (!isSameShuffle(G.Out.Reg, G.Inp.Reg, F->PR.Reg, PrehR)) { 3282 const MachineInstr *DefPrehR = MRI->getVRegDef(F->PR.Reg); 3283 unsigned Opc = DefPrehR->getOpcode(); 3284 if (Opc != Hexagon::A2_tfrsi && Opc != Hexagon::A2_tfrpi) 3285 continue; 3286 if (!DefPrehR->getOperand(1).isImm()) 3287 continue; 3288 if (DefPrehR->getOperand(1).getImm() != 0) 3289 continue; 3290 const TargetRegisterClass *RC = MRI->getRegClass(G.Inp.Reg); 3291 if (RC != MRI->getRegClass(F->PR.Reg)) { 3292 PrehR = MRI->createVirtualRegister(RC); 3293 unsigned TfrI = (RC == &Hexagon::IntRegsRegClass) ? Hexagon::A2_tfrsi 3294 : Hexagon::A2_tfrpi; 3295 auto T = C.PB->getFirstTerminator(); 3296 DebugLoc DL = (T != C.PB->end()) ? T->getDebugLoc() : DebugLoc(); 3297 BuildMI(*C.PB, T, DL, HII->get(TfrI), PrehR) 3298 .addImm(0); 3299 } else { 3300 PrehR = F->PR.Reg; 3301 } 3302 } 3303 // isSameShuffle could match with PrehR being of a wider class than 3304 // G.Inp.Reg, for example if G shuffles the low 32 bits of its input, 3305 // it would match for the input being a 32-bit register, and PrehR 3306 // being a 64-bit register (where the low 32 bits match). This could 3307 // be handled, but for now skip these cases. 3308 if (MRI->getRegClass(PrehR) != MRI->getRegClass(G.Inp.Reg)) 3309 continue; 3310 moveGroup(G, *F->LB, *F->PB, F->LB->getFirstNonPHI(), F->DefR, PrehR); 3311 Changed = true; 3312 } 3313 3314 return Changed; 3315 } 3316 3317 bool HexagonLoopRescheduling::runOnMachineFunction(MachineFunction &MF) { 3318 if (skipFunction(MF.getFunction())) 3319 return false; 3320 3321 auto &HST = MF.getSubtarget<HexagonSubtarget>(); 3322 HII = HST.getInstrInfo(); 3323 HRI = HST.getRegisterInfo(); 3324 MRI = &MF.getRegInfo(); 3325 const HexagonEvaluator HE(*HRI, *MRI, *HII, MF); 3326 BitTracker BT(HE, MF); 3327 LLVM_DEBUG(BT.trace(true)); 3328 BT.run(); 3329 BTP = &BT; 3330 3331 std::vector<LoopCand> Cand; 3332 3333 for (auto &B : MF) { 3334 if (B.pred_size() != 2 || B.succ_size() != 2) 3335 continue; 3336 MachineBasicBlock *PB = nullptr; 3337 bool IsLoop = false; 3338 for (MachineBasicBlock *Pred : B.predecessors()) { 3339 if (Pred != &B) 3340 PB = Pred; 3341 else 3342 IsLoop = true; 3343 } 3344 if (!IsLoop) 3345 continue; 3346 3347 MachineBasicBlock *EB = nullptr; 3348 for (MachineBasicBlock *Succ : B.successors()) { 3349 if (Succ == &B) 3350 continue; 3351 // Set EP to the epilog block, if it has only 1 predecessor (i.e. the 3352 // edge from B to EP is non-critical. 3353 if (Succ->pred_size() == 1) 3354 EB = Succ; 3355 break; 3356 } 3357 3358 Cand.push_back(LoopCand(&B, PB, EB)); 3359 } 3360 3361 bool Changed = false; 3362 for (auto &C : Cand) 3363 Changed |= processLoop(C); 3364 3365 return Changed; 3366 } 3367 3368 //===----------------------------------------------------------------------===// 3369 // Public Constructor Functions 3370 //===----------------------------------------------------------------------===// 3371 3372 FunctionPass *llvm::createHexagonLoopRescheduling() { 3373 return new HexagonLoopRescheduling(); 3374 } 3375 3376 FunctionPass *llvm::createHexagonBitSimplify() { 3377 return new HexagonBitSimplify(); 3378 } 3379