1 //===-- R600InstrInfo.cpp - R600 Instruction Information ------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 /// \file 11 /// \brief R600 Implementation of TargetInstrInfo. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "R600InstrInfo.h" 16 #include "AMDGPU.h" 17 #include "AMDGPUSubtarget.h" 18 #include "AMDGPUTargetMachine.h" 19 #include "R600Defines.h" 20 #include "R600MachineFunctionInfo.h" 21 #include "R600RegisterInfo.h" 22 #include "llvm/CodeGen/MachineFrameInfo.h" 23 #include "llvm/CodeGen/MachineInstrBuilder.h" 24 #include "llvm/CodeGen/MachineRegisterInfo.h" 25 26 using namespace llvm; 27 28 #define GET_INSTRINFO_CTOR_DTOR 29 #include "AMDGPUGenDFAPacketizer.inc" 30 31 R600InstrInfo::R600InstrInfo(const AMDGPUSubtarget &st) 32 : AMDGPUInstrInfo(st), 33 RI(st) 34 { } 35 36 const R600RegisterInfo &R600InstrInfo::getRegisterInfo() const { 37 return RI; 38 } 39 40 bool R600InstrInfo::isTrig(const MachineInstr &MI) const { 41 return get(MI.getOpcode()).TSFlags & R600_InstFlag::TRIG; 42 } 43 44 bool R600InstrInfo::isVector(const MachineInstr &MI) const { 45 return get(MI.getOpcode()).TSFlags & R600_InstFlag::VECTOR; 46 } 47 48 void 49 R600InstrInfo::copyPhysReg(MachineBasicBlock &MBB, 50 MachineBasicBlock::iterator MI, DebugLoc DL, 51 unsigned DestReg, unsigned SrcReg, 52 bool KillSrc) const { 53 unsigned VectorComponents = 0; 54 if ((AMDGPU::R600_Reg128RegClass.contains(DestReg) || 55 AMDGPU::R600_Reg128VerticalRegClass.contains(DestReg)) && 56 (AMDGPU::R600_Reg128RegClass.contains(SrcReg) || 57 AMDGPU::R600_Reg128VerticalRegClass.contains(SrcReg))) { 58 VectorComponents = 4; 59 } else if((AMDGPU::R600_Reg64RegClass.contains(DestReg) || 60 AMDGPU::R600_Reg64VerticalRegClass.contains(DestReg)) && 61 (AMDGPU::R600_Reg64RegClass.contains(SrcReg) || 62 AMDGPU::R600_Reg64VerticalRegClass.contains(SrcReg))) { 63 VectorComponents = 2; 64 } 65 66 if (VectorComponents > 0) { 67 for (unsigned I = 0; I < VectorComponents; I++) { 68 unsigned SubRegIndex = RI.getSubRegFromChannel(I); 69 buildDefaultInstruction(MBB, MI, AMDGPU::MOV, 70 RI.getSubReg(DestReg, SubRegIndex), 71 RI.getSubReg(SrcReg, SubRegIndex)) 72 .addReg(DestReg, 73 RegState::Define | RegState::Implicit); 74 } 75 } else { 76 MachineInstr *NewMI = buildDefaultInstruction(MBB, MI, AMDGPU::MOV, 77 DestReg, SrcReg); 78 NewMI->getOperand(getOperandIdx(*NewMI, AMDGPU::OpName::src0)) 79 .setIsKill(KillSrc); 80 } 81 } 82 83 /// \returns true if \p MBBI can be moved into a new basic. 84 bool R600InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB, 85 MachineBasicBlock::iterator MBBI) const { 86 for (MachineInstr::const_mop_iterator I = MBBI->operands_begin(), 87 E = MBBI->operands_end(); I != E; ++I) { 88 if (I->isReg() && !TargetRegisterInfo::isVirtualRegister(I->getReg()) && 89 I->isUse() && RI.isPhysRegLiveAcrossClauses(I->getReg())) 90 return false; 91 } 92 return true; 93 } 94 95 unsigned R600InstrInfo::getIEQOpcode() const { 96 return AMDGPU::SETE_INT; 97 } 98 99 bool R600InstrInfo::isMov(unsigned Opcode) const { 100 101 102 switch(Opcode) { 103 default: return false; 104 case AMDGPU::MOV: 105 case AMDGPU::MOV_IMM_F32: 106 case AMDGPU::MOV_IMM_I32: 107 return true; 108 } 109 } 110 111 // Some instructions act as place holders to emulate operations that the GPU 112 // hardware does automatically. This function can be used to check if 113 // an opcode falls into this category. 114 bool R600InstrInfo::isPlaceHolderOpcode(unsigned Opcode) const { 115 switch (Opcode) { 116 default: return false; 117 case AMDGPU::RETURN: 118 return true; 119 } 120 } 121 122 bool R600InstrInfo::isReductionOp(unsigned Opcode) const { 123 return false; 124 } 125 126 bool R600InstrInfo::isCubeOp(unsigned Opcode) const { 127 switch(Opcode) { 128 default: return false; 129 case AMDGPU::CUBE_r600_pseudo: 130 case AMDGPU::CUBE_r600_real: 131 case AMDGPU::CUBE_eg_pseudo: 132 case AMDGPU::CUBE_eg_real: 133 return true; 134 } 135 } 136 137 bool R600InstrInfo::isALUInstr(unsigned Opcode) const { 138 unsigned TargetFlags = get(Opcode).TSFlags; 139 140 return (TargetFlags & R600_InstFlag::ALU_INST); 141 } 142 143 bool R600InstrInfo::hasInstrModifiers(unsigned Opcode) const { 144 unsigned TargetFlags = get(Opcode).TSFlags; 145 146 return ((TargetFlags & R600_InstFlag::OP1) | 147 (TargetFlags & R600_InstFlag::OP2) | 148 (TargetFlags & R600_InstFlag::OP3)); 149 } 150 151 bool R600InstrInfo::isLDSInstr(unsigned Opcode) const { 152 unsigned TargetFlags = get(Opcode).TSFlags; 153 154 return ((TargetFlags & R600_InstFlag::LDS_1A) | 155 (TargetFlags & R600_InstFlag::LDS_1A1D) | 156 (TargetFlags & R600_InstFlag::LDS_1A2D)); 157 } 158 159 bool R600InstrInfo::isLDSNoRetInstr(unsigned Opcode) const { 160 return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) == -1; 161 } 162 163 bool R600InstrInfo::isLDSRetInstr(unsigned Opcode) const { 164 return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) != -1; 165 } 166 167 bool R600InstrInfo::canBeConsideredALU(const MachineInstr *MI) const { 168 if (isALUInstr(MI->getOpcode())) 169 return true; 170 if (isVector(*MI) || isCubeOp(MI->getOpcode())) 171 return true; 172 switch (MI->getOpcode()) { 173 case AMDGPU::PRED_X: 174 case AMDGPU::INTERP_PAIR_XY: 175 case AMDGPU::INTERP_PAIR_ZW: 176 case AMDGPU::INTERP_VEC_LOAD: 177 case AMDGPU::COPY: 178 case AMDGPU::DOT_4: 179 return true; 180 default: 181 return false; 182 } 183 } 184 185 bool R600InstrInfo::isTransOnly(unsigned Opcode) const { 186 if (ST.hasCaymanISA()) 187 return false; 188 return (get(Opcode).getSchedClass() == AMDGPU::Sched::TransALU); 189 } 190 191 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const { 192 return isTransOnly(MI->getOpcode()); 193 } 194 195 bool R600InstrInfo::isVectorOnly(unsigned Opcode) const { 196 return (get(Opcode).getSchedClass() == AMDGPU::Sched::VecALU); 197 } 198 199 bool R600InstrInfo::isVectorOnly(const MachineInstr *MI) const { 200 return isVectorOnly(MI->getOpcode()); 201 } 202 203 bool R600InstrInfo::isExport(unsigned Opcode) const { 204 return (get(Opcode).TSFlags & R600_InstFlag::IS_EXPORT); 205 } 206 207 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const { 208 return ST.hasVertexCache() && IS_VTX(get(Opcode)); 209 } 210 211 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const { 212 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>(); 213 return MFI->ShaderType != ShaderType::COMPUTE && usesVertexCache(MI->getOpcode()); 214 } 215 216 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const { 217 return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode)); 218 } 219 220 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const { 221 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>(); 222 return (MFI->ShaderType == ShaderType::COMPUTE && usesVertexCache(MI->getOpcode())) || 223 usesTextureCache(MI->getOpcode()); 224 } 225 226 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const { 227 switch (Opcode) { 228 case AMDGPU::KILLGT: 229 case AMDGPU::GROUP_BARRIER: 230 return true; 231 default: 232 return false; 233 } 234 } 235 236 bool R600InstrInfo::usesAddressRegister(MachineInstr *MI) const { 237 return MI->findRegisterUseOperandIdx(AMDGPU::AR_X) != -1; 238 } 239 240 bool R600InstrInfo::definesAddressRegister(MachineInstr *MI) const { 241 return MI->findRegisterDefOperandIdx(AMDGPU::AR_X) != -1; 242 } 243 244 bool R600InstrInfo::readsLDSSrcReg(const MachineInstr *MI) const { 245 if (!isALUInstr(MI->getOpcode())) { 246 return false; 247 } 248 for (MachineInstr::const_mop_iterator I = MI->operands_begin(), 249 E = MI->operands_end(); I != E; ++I) { 250 if (!I->isReg() || !I->isUse() || 251 TargetRegisterInfo::isVirtualRegister(I->getReg())) 252 continue; 253 254 if (AMDGPU::R600_LDS_SRC_REGRegClass.contains(I->getReg())) 255 return true; 256 } 257 return false; 258 } 259 260 int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const { 261 static const unsigned OpTable[] = { 262 AMDGPU::OpName::src0, 263 AMDGPU::OpName::src1, 264 AMDGPU::OpName::src2 265 }; 266 267 assert (SrcNum < 3); 268 return getOperandIdx(Opcode, OpTable[SrcNum]); 269 } 270 271 #define SRC_SEL_ROWS 11 272 int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const { 273 static const unsigned SrcSelTable[SRC_SEL_ROWS][2] = { 274 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel}, 275 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel}, 276 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel}, 277 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X}, 278 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y}, 279 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z}, 280 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W}, 281 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X}, 282 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y}, 283 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z}, 284 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W} 285 }; 286 287 for (unsigned i = 0; i < SRC_SEL_ROWS; ++i) { 288 if (getOperandIdx(Opcode, SrcSelTable[i][0]) == (int)SrcIdx) { 289 return getOperandIdx(Opcode, SrcSelTable[i][1]); 290 } 291 } 292 return -1; 293 } 294 #undef SRC_SEL_ROWS 295 296 SmallVector<std::pair<MachineOperand *, int64_t>, 3> 297 R600InstrInfo::getSrcs(MachineInstr *MI) const { 298 SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result; 299 300 if (MI->getOpcode() == AMDGPU::DOT_4) { 301 static const unsigned OpTable[8][2] = { 302 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X}, 303 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y}, 304 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z}, 305 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W}, 306 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X}, 307 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y}, 308 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z}, 309 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}, 310 }; 311 312 for (unsigned j = 0; j < 8; j++) { 313 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(), 314 OpTable[j][0])); 315 unsigned Reg = MO.getReg(); 316 if (Reg == AMDGPU::ALU_CONST) { 317 unsigned Sel = MI->getOperand(getOperandIdx(MI->getOpcode(), 318 OpTable[j][1])).getImm(); 319 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel)); 320 continue; 321 } 322 323 } 324 return Result; 325 } 326 327 static const unsigned OpTable[3][2] = { 328 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel}, 329 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel}, 330 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel}, 331 }; 332 333 for (unsigned j = 0; j < 3; j++) { 334 int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]); 335 if (SrcIdx < 0) 336 break; 337 MachineOperand &MO = MI->getOperand(SrcIdx); 338 unsigned Reg = MI->getOperand(SrcIdx).getReg(); 339 if (Reg == AMDGPU::ALU_CONST) { 340 unsigned Sel = MI->getOperand( 341 getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm(); 342 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel)); 343 continue; 344 } 345 if (Reg == AMDGPU::ALU_LITERAL_X) { 346 unsigned Imm = MI->getOperand( 347 getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal)).getImm(); 348 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Imm)); 349 continue; 350 } 351 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, 0)); 352 } 353 return Result; 354 } 355 356 std::vector<std::pair<int, unsigned> > 357 R600InstrInfo::ExtractSrcs(MachineInstr *MI, 358 const DenseMap<unsigned, unsigned> &PV, 359 unsigned &ConstCount) const { 360 ConstCount = 0; 361 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> Srcs = getSrcs(MI); 362 const std::pair<int, unsigned> DummyPair(-1, 0); 363 std::vector<std::pair<int, unsigned> > Result; 364 unsigned i = 0; 365 for (unsigned n = Srcs.size(); i < n; ++i) { 366 unsigned Reg = Srcs[i].first->getReg(); 367 unsigned Index = RI.getEncodingValue(Reg) & 0xff; 368 if (Reg == AMDGPU::OQAP) { 369 Result.push_back(std::pair<int, unsigned>(Index, 0)); 370 } 371 if (PV.find(Reg) != PV.end()) { 372 // 255 is used to tells its a PS/PV reg 373 Result.push_back(std::pair<int, unsigned>(255, 0)); 374 continue; 375 } 376 if (Index > 127) { 377 ConstCount++; 378 Result.push_back(DummyPair); 379 continue; 380 } 381 unsigned Chan = RI.getHWRegChan(Reg); 382 Result.push_back(std::pair<int, unsigned>(Index, Chan)); 383 } 384 for (; i < 3; ++i) 385 Result.push_back(DummyPair); 386 return Result; 387 } 388 389 static std::vector<std::pair<int, unsigned> > 390 Swizzle(std::vector<std::pair<int, unsigned> > Src, 391 R600InstrInfo::BankSwizzle Swz) { 392 if (Src[0] == Src[1]) 393 Src[1].first = -1; 394 switch (Swz) { 395 case R600InstrInfo::ALU_VEC_012_SCL_210: 396 break; 397 case R600InstrInfo::ALU_VEC_021_SCL_122: 398 std::swap(Src[1], Src[2]); 399 break; 400 case R600InstrInfo::ALU_VEC_102_SCL_221: 401 std::swap(Src[0], Src[1]); 402 break; 403 case R600InstrInfo::ALU_VEC_120_SCL_212: 404 std::swap(Src[0], Src[1]); 405 std::swap(Src[0], Src[2]); 406 break; 407 case R600InstrInfo::ALU_VEC_201: 408 std::swap(Src[0], Src[2]); 409 std::swap(Src[0], Src[1]); 410 break; 411 case R600InstrInfo::ALU_VEC_210: 412 std::swap(Src[0], Src[2]); 413 break; 414 } 415 return Src; 416 } 417 418 static unsigned 419 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) { 420 switch (Swz) { 421 case R600InstrInfo::ALU_VEC_012_SCL_210: { 422 unsigned Cycles[3] = { 2, 1, 0}; 423 return Cycles[Op]; 424 } 425 case R600InstrInfo::ALU_VEC_021_SCL_122: { 426 unsigned Cycles[3] = { 1, 2, 2}; 427 return Cycles[Op]; 428 } 429 case R600InstrInfo::ALU_VEC_120_SCL_212: { 430 unsigned Cycles[3] = { 2, 1, 2}; 431 return Cycles[Op]; 432 } 433 case R600InstrInfo::ALU_VEC_102_SCL_221: { 434 unsigned Cycles[3] = { 2, 2, 1}; 435 return Cycles[Op]; 436 } 437 default: 438 llvm_unreachable("Wrong Swizzle for Trans Slot"); 439 return 0; 440 } 441 } 442 443 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed 444 /// in the same Instruction Group while meeting read port limitations given a 445 /// Swz swizzle sequence. 446 unsigned R600InstrInfo::isLegalUpTo( 447 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs, 448 const std::vector<R600InstrInfo::BankSwizzle> &Swz, 449 const std::vector<std::pair<int, unsigned> > &TransSrcs, 450 R600InstrInfo::BankSwizzle TransSwz) const { 451 int Vector[4][3]; 452 memset(Vector, -1, sizeof(Vector)); 453 for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) { 454 const std::vector<std::pair<int, unsigned> > &Srcs = 455 Swizzle(IGSrcs[i], Swz[i]); 456 for (unsigned j = 0; j < 3; j++) { 457 const std::pair<int, unsigned> &Src = Srcs[j]; 458 if (Src.first < 0 || Src.first == 255) 459 continue; 460 if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) { 461 if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 && 462 Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) { 463 // The value from output queue A (denoted by register OQAP) can 464 // only be fetched during the first cycle. 465 return false; 466 } 467 // OQAP does not count towards the normal read port restrictions 468 continue; 469 } 470 if (Vector[Src.second][j] < 0) 471 Vector[Src.second][j] = Src.first; 472 if (Vector[Src.second][j] != Src.first) 473 return i; 474 } 475 } 476 // Now check Trans Alu 477 for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) { 478 const std::pair<int, unsigned> &Src = TransSrcs[i]; 479 unsigned Cycle = getTransSwizzle(TransSwz, i); 480 if (Src.first < 0) 481 continue; 482 if (Src.first == 255) 483 continue; 484 if (Vector[Src.second][Cycle] < 0) 485 Vector[Src.second][Cycle] = Src.first; 486 if (Vector[Src.second][Cycle] != Src.first) 487 return IGSrcs.size() - 1; 488 } 489 return IGSrcs.size(); 490 } 491 492 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next 493 /// (in lexicographic term) swizzle sequence assuming that all swizzles after 494 /// Idx can be skipped 495 static bool 496 NextPossibleSolution( 497 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate, 498 unsigned Idx) { 499 assert(Idx < SwzCandidate.size()); 500 int ResetIdx = Idx; 501 while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210) 502 ResetIdx --; 503 for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) { 504 SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210; 505 } 506 if (ResetIdx == -1) 507 return false; 508 int NextSwizzle = SwzCandidate[ResetIdx] + 1; 509 SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle; 510 return true; 511 } 512 513 /// Enumerate all possible Swizzle sequence to find one that can meet all 514 /// read port requirements. 515 bool R600InstrInfo::FindSwizzleForVectorSlot( 516 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs, 517 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate, 518 const std::vector<std::pair<int, unsigned> > &TransSrcs, 519 R600InstrInfo::BankSwizzle TransSwz) const { 520 unsigned ValidUpTo = 0; 521 do { 522 ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz); 523 if (ValidUpTo == IGSrcs.size()) 524 return true; 525 } while (NextPossibleSolution(SwzCandidate, ValidUpTo)); 526 return false; 527 } 528 529 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read 530 /// a const, and can't read a gpr at cycle 1 if they read 2 const. 531 static bool 532 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz, 533 const std::vector<std::pair<int, unsigned> > &TransOps, 534 unsigned ConstCount) { 535 // TransALU can't read 3 constants 536 if (ConstCount > 2) 537 return false; 538 for (unsigned i = 0, e = TransOps.size(); i < e; ++i) { 539 const std::pair<int, unsigned> &Src = TransOps[i]; 540 unsigned Cycle = getTransSwizzle(TransSwz, i); 541 if (Src.first < 0) 542 continue; 543 if (ConstCount > 0 && Cycle == 0) 544 return false; 545 if (ConstCount > 1 && Cycle == 1) 546 return false; 547 } 548 return true; 549 } 550 551 bool 552 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG, 553 const DenseMap<unsigned, unsigned> &PV, 554 std::vector<BankSwizzle> &ValidSwizzle, 555 bool isLastAluTrans) 556 const { 557 //Todo : support shared src0 - src1 operand 558 559 std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs; 560 ValidSwizzle.clear(); 561 unsigned ConstCount; 562 BankSwizzle TransBS = ALU_VEC_012_SCL_210; 563 for (unsigned i = 0, e = IG.size(); i < e; ++i) { 564 IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount)); 565 unsigned Op = getOperandIdx(IG[i]->getOpcode(), 566 AMDGPU::OpName::bank_swizzle); 567 ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle) 568 IG[i]->getOperand(Op).getImm()); 569 } 570 std::vector<std::pair<int, unsigned> > TransOps; 571 if (!isLastAluTrans) 572 return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS); 573 574 TransOps = IGSrcs.back(); 575 IGSrcs.pop_back(); 576 ValidSwizzle.pop_back(); 577 578 static const R600InstrInfo::BankSwizzle TransSwz[] = { 579 ALU_VEC_012_SCL_210, 580 ALU_VEC_021_SCL_122, 581 ALU_VEC_120_SCL_212, 582 ALU_VEC_102_SCL_221 583 }; 584 for (unsigned i = 0; i < 4; i++) { 585 TransBS = TransSwz[i]; 586 if (!isConstCompatible(TransBS, TransOps, ConstCount)) 587 continue; 588 bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, 589 TransBS); 590 if (Result) { 591 ValidSwizzle.push_back(TransBS); 592 return true; 593 } 594 } 595 596 return false; 597 } 598 599 600 bool 601 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts) 602 const { 603 assert (Consts.size() <= 12 && "Too many operands in instructions group"); 604 unsigned Pair1 = 0, Pair2 = 0; 605 for (unsigned i = 0, n = Consts.size(); i < n; ++i) { 606 unsigned ReadConstHalf = Consts[i] & 2; 607 unsigned ReadConstIndex = Consts[i] & (~3); 608 unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf; 609 if (!Pair1) { 610 Pair1 = ReadHalfConst; 611 continue; 612 } 613 if (Pair1 == ReadHalfConst) 614 continue; 615 if (!Pair2) { 616 Pair2 = ReadHalfConst; 617 continue; 618 } 619 if (Pair2 != ReadHalfConst) 620 return false; 621 } 622 return true; 623 } 624 625 bool 626 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs) 627 const { 628 std::vector<unsigned> Consts; 629 SmallSet<int64_t, 4> Literals; 630 for (unsigned i = 0, n = MIs.size(); i < n; i++) { 631 MachineInstr *MI = MIs[i]; 632 if (!isALUInstr(MI->getOpcode())) 633 continue; 634 635 const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Srcs = 636 getSrcs(MI); 637 638 for (unsigned j = 0, e = Srcs.size(); j < e; j++) { 639 std::pair<MachineOperand *, unsigned> Src = Srcs[j]; 640 if (Src.first->getReg() == AMDGPU::ALU_LITERAL_X) 641 Literals.insert(Src.second); 642 if (Literals.size() > 4) 643 return false; 644 if (Src.first->getReg() == AMDGPU::ALU_CONST) 645 Consts.push_back(Src.second); 646 if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) || 647 AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) { 648 unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff; 649 unsigned Chan = RI.getHWRegChan(Src.first->getReg()); 650 Consts.push_back((Index << 2) | Chan); 651 } 652 } 653 } 654 return fitsConstReadLimitations(Consts); 655 } 656 657 DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM, 658 const ScheduleDAG *DAG) const { 659 const InstrItineraryData *II = TM->getInstrItineraryData(); 660 return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II); 661 } 662 663 static bool 664 isPredicateSetter(unsigned Opcode) { 665 switch (Opcode) { 666 case AMDGPU::PRED_X: 667 return true; 668 default: 669 return false; 670 } 671 } 672 673 static MachineInstr * 674 findFirstPredicateSetterFrom(MachineBasicBlock &MBB, 675 MachineBasicBlock::iterator I) { 676 while (I != MBB.begin()) { 677 --I; 678 MachineInstr *MI = I; 679 if (isPredicateSetter(MI->getOpcode())) 680 return MI; 681 } 682 683 return nullptr; 684 } 685 686 static 687 bool isJump(unsigned Opcode) { 688 return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND; 689 } 690 691 static bool isBranch(unsigned Opcode) { 692 return Opcode == AMDGPU::BRANCH || Opcode == AMDGPU::BRANCH_COND_i32 || 693 Opcode == AMDGPU::BRANCH_COND_f32; 694 } 695 696 bool 697 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, 698 MachineBasicBlock *&TBB, 699 MachineBasicBlock *&FBB, 700 SmallVectorImpl<MachineOperand> &Cond, 701 bool AllowModify) const { 702 // Most of the following comes from the ARM implementation of AnalyzeBranch 703 704 // If the block has no terminators, it just falls into the block after it. 705 MachineBasicBlock::iterator I = MBB.end(); 706 if (I == MBB.begin()) 707 return false; 708 --I; 709 while (I->isDebugValue()) { 710 if (I == MBB.begin()) 711 return false; 712 --I; 713 } 714 // AMDGPU::BRANCH* instructions are only available after isel and are not 715 // handled 716 if (isBranch(I->getOpcode())) 717 return true; 718 if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) { 719 return false; 720 } 721 722 // Remove successive JUMP 723 while (I != MBB.begin() && std::prev(I)->getOpcode() == AMDGPU::JUMP) { 724 MachineBasicBlock::iterator PriorI = std::prev(I); 725 if (AllowModify) 726 I->removeFromParent(); 727 I = PriorI; 728 } 729 MachineInstr *LastInst = I; 730 731 // If there is only one terminator instruction, process it. 732 unsigned LastOpc = LastInst->getOpcode(); 733 if (I == MBB.begin() || 734 !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) { 735 if (LastOpc == AMDGPU::JUMP) { 736 TBB = LastInst->getOperand(0).getMBB(); 737 return false; 738 } else if (LastOpc == AMDGPU::JUMP_COND) { 739 MachineInstr *predSet = I; 740 while (!isPredicateSetter(predSet->getOpcode())) { 741 predSet = --I; 742 } 743 TBB = LastInst->getOperand(0).getMBB(); 744 Cond.push_back(predSet->getOperand(1)); 745 Cond.push_back(predSet->getOperand(2)); 746 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false)); 747 return false; 748 } 749 return true; // Can't handle indirect branch. 750 } 751 752 // Get the instruction before it if it is a terminator. 753 MachineInstr *SecondLastInst = I; 754 unsigned SecondLastOpc = SecondLastInst->getOpcode(); 755 756 // If the block ends with a B and a Bcc, handle it. 757 if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) { 758 MachineInstr *predSet = --I; 759 while (!isPredicateSetter(predSet->getOpcode())) { 760 predSet = --I; 761 } 762 TBB = SecondLastInst->getOperand(0).getMBB(); 763 FBB = LastInst->getOperand(0).getMBB(); 764 Cond.push_back(predSet->getOperand(1)); 765 Cond.push_back(predSet->getOperand(2)); 766 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false)); 767 return false; 768 } 769 770 // Otherwise, can't handle this. 771 return true; 772 } 773 774 static 775 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) { 776 for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend(); 777 It != E; ++It) { 778 if (It->getOpcode() == AMDGPU::CF_ALU || 779 It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE) 780 return std::prev(It.base()); 781 } 782 return MBB.end(); 783 } 784 785 unsigned 786 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB, 787 MachineBasicBlock *TBB, 788 MachineBasicBlock *FBB, 789 const SmallVectorImpl<MachineOperand> &Cond, 790 DebugLoc DL) const { 791 assert(TBB && "InsertBranch must not be told to insert a fallthrough"); 792 793 if (!FBB) { 794 if (Cond.empty()) { 795 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB); 796 return 1; 797 } else { 798 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end()); 799 assert(PredSet && "No previous predicate !"); 800 addFlag(PredSet, 0, MO_FLAG_PUSH); 801 PredSet->getOperand(2).setImm(Cond[1].getImm()); 802 803 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND)) 804 .addMBB(TBB) 805 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill); 806 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB); 807 if (CfAlu == MBB.end()) 808 return 1; 809 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU); 810 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE)); 811 return 1; 812 } 813 } else { 814 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end()); 815 assert(PredSet && "No previous predicate !"); 816 addFlag(PredSet, 0, MO_FLAG_PUSH); 817 PredSet->getOperand(2).setImm(Cond[1].getImm()); 818 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND)) 819 .addMBB(TBB) 820 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill); 821 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB); 822 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB); 823 if (CfAlu == MBB.end()) 824 return 2; 825 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU); 826 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE)); 827 return 2; 828 } 829 } 830 831 unsigned 832 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { 833 834 // Note : we leave PRED* instructions there. 835 // They may be needed when predicating instructions. 836 837 MachineBasicBlock::iterator I = MBB.end(); 838 839 if (I == MBB.begin()) { 840 return 0; 841 } 842 --I; 843 switch (I->getOpcode()) { 844 default: 845 return 0; 846 case AMDGPU::JUMP_COND: { 847 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I); 848 clearFlag(predSet, 0, MO_FLAG_PUSH); 849 I->eraseFromParent(); 850 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB); 851 if (CfAlu == MBB.end()) 852 break; 853 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE); 854 CfAlu->setDesc(get(AMDGPU::CF_ALU)); 855 break; 856 } 857 case AMDGPU::JUMP: 858 I->eraseFromParent(); 859 break; 860 } 861 I = MBB.end(); 862 863 if (I == MBB.begin()) { 864 return 1; 865 } 866 --I; 867 switch (I->getOpcode()) { 868 // FIXME: only one case?? 869 default: 870 return 1; 871 case AMDGPU::JUMP_COND: { 872 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I); 873 clearFlag(predSet, 0, MO_FLAG_PUSH); 874 I->eraseFromParent(); 875 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB); 876 if (CfAlu == MBB.end()) 877 break; 878 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE); 879 CfAlu->setDesc(get(AMDGPU::CF_ALU)); 880 break; 881 } 882 case AMDGPU::JUMP: 883 I->eraseFromParent(); 884 break; 885 } 886 return 2; 887 } 888 889 bool 890 R600InstrInfo::isPredicated(const MachineInstr *MI) const { 891 int idx = MI->findFirstPredOperandIdx(); 892 if (idx < 0) 893 return false; 894 895 unsigned Reg = MI->getOperand(idx).getReg(); 896 switch (Reg) { 897 default: return false; 898 case AMDGPU::PRED_SEL_ONE: 899 case AMDGPU::PRED_SEL_ZERO: 900 case AMDGPU::PREDICATE_BIT: 901 return true; 902 } 903 } 904 905 bool 906 R600InstrInfo::isPredicable(MachineInstr *MI) const { 907 // XXX: KILL* instructions can be predicated, but they must be the last 908 // instruction in a clause, so this means any instructions after them cannot 909 // be predicated. Until we have proper support for instruction clauses in the 910 // backend, we will mark KILL* instructions as unpredicable. 911 912 if (MI->getOpcode() == AMDGPU::KILLGT) { 913 return false; 914 } else if (MI->getOpcode() == AMDGPU::CF_ALU) { 915 // If the clause start in the middle of MBB then the MBB has more 916 // than a single clause, unable to predicate several clauses. 917 if (MI->getParent()->begin() != MachineBasicBlock::iterator(MI)) 918 return false; 919 // TODO: We don't support KC merging atm 920 if (MI->getOperand(3).getImm() != 0 || MI->getOperand(4).getImm() != 0) 921 return false; 922 return true; 923 } else if (isVector(*MI)) { 924 return false; 925 } else { 926 return AMDGPUInstrInfo::isPredicable(MI); 927 } 928 } 929 930 931 bool 932 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB, 933 unsigned NumCyles, 934 unsigned ExtraPredCycles, 935 const BranchProbability &Probability) const{ 936 return true; 937 } 938 939 bool 940 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB, 941 unsigned NumTCycles, 942 unsigned ExtraTCycles, 943 MachineBasicBlock &FMBB, 944 unsigned NumFCycles, 945 unsigned ExtraFCycles, 946 const BranchProbability &Probability) const { 947 return true; 948 } 949 950 bool 951 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB, 952 unsigned NumCyles, 953 const BranchProbability &Probability) 954 const { 955 return true; 956 } 957 958 bool 959 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB, 960 MachineBasicBlock &FMBB) const { 961 return false; 962 } 963 964 965 bool 966 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 967 MachineOperand &MO = Cond[1]; 968 switch (MO.getImm()) { 969 case OPCODE_IS_ZERO_INT: 970 MO.setImm(OPCODE_IS_NOT_ZERO_INT); 971 break; 972 case OPCODE_IS_NOT_ZERO_INT: 973 MO.setImm(OPCODE_IS_ZERO_INT); 974 break; 975 case OPCODE_IS_ZERO: 976 MO.setImm(OPCODE_IS_NOT_ZERO); 977 break; 978 case OPCODE_IS_NOT_ZERO: 979 MO.setImm(OPCODE_IS_ZERO); 980 break; 981 default: 982 return true; 983 } 984 985 MachineOperand &MO2 = Cond[2]; 986 switch (MO2.getReg()) { 987 case AMDGPU::PRED_SEL_ZERO: 988 MO2.setReg(AMDGPU::PRED_SEL_ONE); 989 break; 990 case AMDGPU::PRED_SEL_ONE: 991 MO2.setReg(AMDGPU::PRED_SEL_ZERO); 992 break; 993 default: 994 return true; 995 } 996 return false; 997 } 998 999 bool 1000 R600InstrInfo::DefinesPredicate(MachineInstr *MI, 1001 std::vector<MachineOperand> &Pred) const { 1002 return isPredicateSetter(MI->getOpcode()); 1003 } 1004 1005 1006 bool 1007 R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1, 1008 const SmallVectorImpl<MachineOperand> &Pred2) const { 1009 return false; 1010 } 1011 1012 1013 bool 1014 R600InstrInfo::PredicateInstruction(MachineInstr *MI, 1015 const SmallVectorImpl<MachineOperand> &Pred) const { 1016 int PIdx = MI->findFirstPredOperandIdx(); 1017 1018 if (MI->getOpcode() == AMDGPU::CF_ALU) { 1019 MI->getOperand(8).setImm(0); 1020 return true; 1021 } 1022 1023 if (MI->getOpcode() == AMDGPU::DOT_4) { 1024 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_X)) 1025 .setReg(Pred[2].getReg()); 1026 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Y)) 1027 .setReg(Pred[2].getReg()); 1028 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Z)) 1029 .setReg(Pred[2].getReg()); 1030 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_W)) 1031 .setReg(Pred[2].getReg()); 1032 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI); 1033 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit); 1034 return true; 1035 } 1036 1037 if (PIdx != -1) { 1038 MachineOperand &PMO = MI->getOperand(PIdx); 1039 PMO.setReg(Pred[2].getReg()); 1040 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI); 1041 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit); 1042 return true; 1043 } 1044 1045 return false; 1046 } 1047 1048 unsigned int R600InstrInfo::getPredicationCost(const MachineInstr *) const { 1049 return 2; 1050 } 1051 1052 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 1053 const MachineInstr *MI, 1054 unsigned *PredCost) const { 1055 if (PredCost) 1056 *PredCost = 2; 1057 return 2; 1058 } 1059 1060 bool R600InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const { 1061 1062 switch(MI->getOpcode()) { 1063 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI); 1064 case AMDGPU::R600_EXTRACT_ELT_V2: 1065 case AMDGPU::R600_EXTRACT_ELT_V4: 1066 buildIndirectRead(MI->getParent(), MI, MI->getOperand(0).getReg(), 1067 RI.getHWRegIndex(MI->getOperand(1).getReg()), // Address 1068 MI->getOperand(2).getReg(), 1069 RI.getHWRegChan(MI->getOperand(1).getReg())); 1070 break; 1071 case AMDGPU::R600_INSERT_ELT_V2: 1072 case AMDGPU::R600_INSERT_ELT_V4: 1073 buildIndirectWrite(MI->getParent(), MI, MI->getOperand(2).getReg(), // Value 1074 RI.getHWRegIndex(MI->getOperand(1).getReg()), // Address 1075 MI->getOperand(3).getReg(), // Offset 1076 RI.getHWRegChan(MI->getOperand(1).getReg())); // Channel 1077 break; 1078 } 1079 MI->eraseFromParent(); 1080 return true; 1081 } 1082 1083 void R600InstrInfo::reserveIndirectRegisters(BitVector &Reserved, 1084 const MachineFunction &MF) const { 1085 const AMDGPUFrameLowering *TFL = 1086 static_cast<const AMDGPUFrameLowering*>( 1087 MF.getTarget().getFrameLowering()); 1088 1089 unsigned StackWidth = TFL->getStackWidth(MF); 1090 int End = getIndirectIndexEnd(MF); 1091 1092 if (End == -1) 1093 return; 1094 1095 for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) { 1096 unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index); 1097 Reserved.set(SuperReg); 1098 for (unsigned Chan = 0; Chan < StackWidth; ++Chan) { 1099 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan); 1100 Reserved.set(Reg); 1101 } 1102 } 1103 } 1104 1105 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex, 1106 unsigned Channel) const { 1107 // XXX: Remove when we support a stack width > 2 1108 assert(Channel == 0); 1109 return RegIndex; 1110 } 1111 1112 const TargetRegisterClass *R600InstrInfo::getIndirectAddrRegClass() const { 1113 return &AMDGPU::R600_TReg32_XRegClass; 1114 } 1115 1116 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB, 1117 MachineBasicBlock::iterator I, 1118 unsigned ValueReg, unsigned Address, 1119 unsigned OffsetReg) const { 1120 return buildIndirectWrite(MBB, I, ValueReg, Address, OffsetReg, 0); 1121 } 1122 1123 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB, 1124 MachineBasicBlock::iterator I, 1125 unsigned ValueReg, unsigned Address, 1126 unsigned OffsetReg, 1127 unsigned AddrChan) const { 1128 unsigned AddrReg; 1129 switch (AddrChan) { 1130 default: llvm_unreachable("Invalid Channel"); 1131 case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break; 1132 case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break; 1133 case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break; 1134 case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break; 1135 } 1136 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg, 1137 AMDGPU::AR_X, OffsetReg); 1138 setImmOperand(MOVA, AMDGPU::OpName::write, 0); 1139 1140 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV, 1141 AddrReg, ValueReg) 1142 .addReg(AMDGPU::AR_X, 1143 RegState::Implicit | RegState::Kill); 1144 setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1); 1145 return Mov; 1146 } 1147 1148 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB, 1149 MachineBasicBlock::iterator I, 1150 unsigned ValueReg, unsigned Address, 1151 unsigned OffsetReg) const { 1152 return buildIndirectRead(MBB, I, ValueReg, Address, OffsetReg, 0); 1153 } 1154 1155 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB, 1156 MachineBasicBlock::iterator I, 1157 unsigned ValueReg, unsigned Address, 1158 unsigned OffsetReg, 1159 unsigned AddrChan) const { 1160 unsigned AddrReg; 1161 switch (AddrChan) { 1162 default: llvm_unreachable("Invalid Channel"); 1163 case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break; 1164 case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break; 1165 case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break; 1166 case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break; 1167 } 1168 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg, 1169 AMDGPU::AR_X, 1170 OffsetReg); 1171 setImmOperand(MOVA, AMDGPU::OpName::write, 0); 1172 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV, 1173 ValueReg, 1174 AddrReg) 1175 .addReg(AMDGPU::AR_X, 1176 RegState::Implicit | RegState::Kill); 1177 setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1); 1178 1179 return Mov; 1180 } 1181 1182 unsigned R600InstrInfo::getMaxAlusPerClause() const { 1183 return 115; 1184 } 1185 1186 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB, 1187 MachineBasicBlock::iterator I, 1188 unsigned Opcode, 1189 unsigned DstReg, 1190 unsigned Src0Reg, 1191 unsigned Src1Reg) const { 1192 MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode), 1193 DstReg); // $dst 1194 1195 if (Src1Reg) { 1196 MIB.addImm(0) // $update_exec_mask 1197 .addImm(0); // $update_predicate 1198 } 1199 MIB.addImm(1) // $write 1200 .addImm(0) // $omod 1201 .addImm(0) // $dst_rel 1202 .addImm(0) // $dst_clamp 1203 .addReg(Src0Reg) // $src0 1204 .addImm(0) // $src0_neg 1205 .addImm(0) // $src0_rel 1206 .addImm(0) // $src0_abs 1207 .addImm(-1); // $src0_sel 1208 1209 if (Src1Reg) { 1210 MIB.addReg(Src1Reg) // $src1 1211 .addImm(0) // $src1_neg 1212 .addImm(0) // $src1_rel 1213 .addImm(0) // $src1_abs 1214 .addImm(-1); // $src1_sel 1215 } 1216 1217 //XXX: The r600g finalizer expects this to be 1, once we've moved the 1218 //scheduling to the backend, we can change the default to 0. 1219 MIB.addImm(1) // $last 1220 .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel 1221 .addImm(0) // $literal 1222 .addImm(0); // $bank_swizzle 1223 1224 return MIB; 1225 } 1226 1227 #define OPERAND_CASE(Label) \ 1228 case Label: { \ 1229 static const unsigned Ops[] = \ 1230 { \ 1231 Label##_X, \ 1232 Label##_Y, \ 1233 Label##_Z, \ 1234 Label##_W \ 1235 }; \ 1236 return Ops[Slot]; \ 1237 } 1238 1239 static unsigned getSlotedOps(unsigned Op, unsigned Slot) { 1240 switch (Op) { 1241 OPERAND_CASE(AMDGPU::OpName::update_exec_mask) 1242 OPERAND_CASE(AMDGPU::OpName::update_pred) 1243 OPERAND_CASE(AMDGPU::OpName::write) 1244 OPERAND_CASE(AMDGPU::OpName::omod) 1245 OPERAND_CASE(AMDGPU::OpName::dst_rel) 1246 OPERAND_CASE(AMDGPU::OpName::clamp) 1247 OPERAND_CASE(AMDGPU::OpName::src0) 1248 OPERAND_CASE(AMDGPU::OpName::src0_neg) 1249 OPERAND_CASE(AMDGPU::OpName::src0_rel) 1250 OPERAND_CASE(AMDGPU::OpName::src0_abs) 1251 OPERAND_CASE(AMDGPU::OpName::src0_sel) 1252 OPERAND_CASE(AMDGPU::OpName::src1) 1253 OPERAND_CASE(AMDGPU::OpName::src1_neg) 1254 OPERAND_CASE(AMDGPU::OpName::src1_rel) 1255 OPERAND_CASE(AMDGPU::OpName::src1_abs) 1256 OPERAND_CASE(AMDGPU::OpName::src1_sel) 1257 OPERAND_CASE(AMDGPU::OpName::pred_sel) 1258 default: 1259 llvm_unreachable("Wrong Operand"); 1260 } 1261 } 1262 1263 #undef OPERAND_CASE 1264 1265 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction( 1266 MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg) 1267 const { 1268 assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented"); 1269 unsigned Opcode; 1270 if (ST.getGeneration() <= AMDGPUSubtarget::R700) 1271 Opcode = AMDGPU::DOT4_r600; 1272 else 1273 Opcode = AMDGPU::DOT4_eg; 1274 MachineBasicBlock::iterator I = MI; 1275 MachineOperand &Src0 = MI->getOperand( 1276 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot))); 1277 MachineOperand &Src1 = MI->getOperand( 1278 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot))); 1279 MachineInstr *MIB = buildDefaultInstruction( 1280 MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg()); 1281 static const unsigned Operands[14] = { 1282 AMDGPU::OpName::update_exec_mask, 1283 AMDGPU::OpName::update_pred, 1284 AMDGPU::OpName::write, 1285 AMDGPU::OpName::omod, 1286 AMDGPU::OpName::dst_rel, 1287 AMDGPU::OpName::clamp, 1288 AMDGPU::OpName::src0_neg, 1289 AMDGPU::OpName::src0_rel, 1290 AMDGPU::OpName::src0_abs, 1291 AMDGPU::OpName::src0_sel, 1292 AMDGPU::OpName::src1_neg, 1293 AMDGPU::OpName::src1_rel, 1294 AMDGPU::OpName::src1_abs, 1295 AMDGPU::OpName::src1_sel, 1296 }; 1297 1298 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(), 1299 getSlotedOps(AMDGPU::OpName::pred_sel, Slot))); 1300 MIB->getOperand(getOperandIdx(Opcode, AMDGPU::OpName::pred_sel)) 1301 .setReg(MO.getReg()); 1302 1303 for (unsigned i = 0; i < 14; i++) { 1304 MachineOperand &MO = MI->getOperand( 1305 getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot))); 1306 assert (MO.isImm()); 1307 setImmOperand(MIB, Operands[i], MO.getImm()); 1308 } 1309 MIB->getOperand(20).setImm(0); 1310 return MIB; 1311 } 1312 1313 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB, 1314 MachineBasicBlock::iterator I, 1315 unsigned DstReg, 1316 uint64_t Imm) const { 1317 MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg, 1318 AMDGPU::ALU_LITERAL_X); 1319 setImmOperand(MovImm, AMDGPU::OpName::literal, Imm); 1320 return MovImm; 1321 } 1322 1323 MachineInstr *R600InstrInfo::buildMovInstr(MachineBasicBlock *MBB, 1324 MachineBasicBlock::iterator I, 1325 unsigned DstReg, unsigned SrcReg) const { 1326 return buildDefaultInstruction(*MBB, I, AMDGPU::MOV, DstReg, SrcReg); 1327 } 1328 1329 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const { 1330 return getOperandIdx(MI.getOpcode(), Op); 1331 } 1332 1333 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const { 1334 return AMDGPU::getNamedOperandIdx(Opcode, Op); 1335 } 1336 1337 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op, 1338 int64_t Imm) const { 1339 int Idx = getOperandIdx(*MI, Op); 1340 assert(Idx != -1 && "Operand not supported for this instruction."); 1341 assert(MI->getOperand(Idx).isImm()); 1342 MI->getOperand(Idx).setImm(Imm); 1343 } 1344 1345 //===----------------------------------------------------------------------===// 1346 // Instruction flag getters/setters 1347 //===----------------------------------------------------------------------===// 1348 1349 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const { 1350 return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0; 1351 } 1352 1353 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx, 1354 unsigned Flag) const { 1355 unsigned TargetFlags = get(MI->getOpcode()).TSFlags; 1356 int FlagIndex = 0; 1357 if (Flag != 0) { 1358 // If we pass something other than the default value of Flag to this 1359 // function, it means we are want to set a flag on an instruction 1360 // that uses native encoding. 1361 assert(HAS_NATIVE_OPERANDS(TargetFlags)); 1362 bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3; 1363 switch (Flag) { 1364 case MO_FLAG_CLAMP: 1365 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp); 1366 break; 1367 case MO_FLAG_MASK: 1368 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write); 1369 break; 1370 case MO_FLAG_NOT_LAST: 1371 case MO_FLAG_LAST: 1372 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last); 1373 break; 1374 case MO_FLAG_NEG: 1375 switch (SrcIdx) { 1376 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break; 1377 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break; 1378 case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break; 1379 } 1380 break; 1381 1382 case MO_FLAG_ABS: 1383 assert(!IsOP3 && "Cannot set absolute value modifier for OP3 " 1384 "instructions."); 1385 (void)IsOP3; 1386 switch (SrcIdx) { 1387 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break; 1388 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break; 1389 } 1390 break; 1391 1392 default: 1393 FlagIndex = -1; 1394 break; 1395 } 1396 assert(FlagIndex != -1 && "Flag not supported for this instruction"); 1397 } else { 1398 FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags); 1399 assert(FlagIndex != 0 && 1400 "Instruction flags not supported for this instruction"); 1401 } 1402 1403 MachineOperand &FlagOp = MI->getOperand(FlagIndex); 1404 assert(FlagOp.isImm()); 1405 return FlagOp; 1406 } 1407 1408 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand, 1409 unsigned Flag) const { 1410 unsigned TargetFlags = get(MI->getOpcode()).TSFlags; 1411 if (Flag == 0) { 1412 return; 1413 } 1414 if (HAS_NATIVE_OPERANDS(TargetFlags)) { 1415 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag); 1416 if (Flag == MO_FLAG_NOT_LAST) { 1417 clearFlag(MI, Operand, MO_FLAG_LAST); 1418 } else if (Flag == MO_FLAG_MASK) { 1419 clearFlag(MI, Operand, Flag); 1420 } else { 1421 FlagOp.setImm(1); 1422 } 1423 } else { 1424 MachineOperand &FlagOp = getFlagOp(MI, Operand); 1425 FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand))); 1426 } 1427 } 1428 1429 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand, 1430 unsigned Flag) const { 1431 unsigned TargetFlags = get(MI->getOpcode()).TSFlags; 1432 if (HAS_NATIVE_OPERANDS(TargetFlags)) { 1433 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag); 1434 FlagOp.setImm(0); 1435 } else { 1436 MachineOperand &FlagOp = getFlagOp(MI); 1437 unsigned InstFlags = FlagOp.getImm(); 1438 InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand)); 1439 FlagOp.setImm(InstFlags); 1440 } 1441 } 1442