1//===-- AMDGPUISelDAGToDAG.cpp - A dag to dag inst selector for AMDGPU ----===// 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 Defines an instruction selector for the AMDGPU target. 12// 13//===----------------------------------------------------------------------===// 14 15#include "AMDGPUInstrInfo.h" 16#include "AMDGPUIntrinsicInfo.h" 17#include "AMDGPUISelLowering.h" // For AMDGPUISD 18#include "AMDGPUSubtarget.h" 19#include "SIISelLowering.h" 20#include "SIMachineFunctionInfo.h" 21#include "llvm/Analysis/ValueTracking.h" 22#include "llvm/CodeGen/FunctionLoweringInfo.h" 23#include "llvm/CodeGen/MachineFrameInfo.h" 24#include "llvm/CodeGen/PseudoSourceValue.h" 25#include "llvm/CodeGen/SelectionDAG.h" 26#include "llvm/CodeGen/SelectionDAGISel.h" 27#include "llvm/IR/DiagnosticInfo.h" 28 29using namespace llvm; 30 31namespace llvm { 32class R600InstrInfo; 33} 34 35//===----------------------------------------------------------------------===// 36// Instruction Selector Implementation 37//===----------------------------------------------------------------------===// 38 39namespace { 40 41static bool isCBranchSCC(const SDNode *N) { 42 assert(N->getOpcode() == ISD::BRCOND); 43 if (!N->hasOneUse()) 44 return false; 45 46 SDValue Cond = N->getOperand(1); 47 if (Cond.getOpcode() == ISD::CopyToReg) 48 Cond = Cond.getOperand(2); 49 return Cond.getOpcode() == ISD::SETCC && 50 Cond.getOperand(0).getValueType() == MVT::i32 && Cond.hasOneUse(); 51} 52 53/// AMDGPU specific code to select AMDGPU machine instructions for 54/// SelectionDAG operations. 55class AMDGPUDAGToDAGISel : public SelectionDAGISel { 56 // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can 57 // make the right decision when generating code for different targets. 58 const AMDGPUSubtarget *Subtarget; 59 60public: 61 AMDGPUDAGToDAGISel(TargetMachine &TM); 62 virtual ~AMDGPUDAGToDAGISel(); 63 bool runOnMachineFunction(MachineFunction &MF) override; 64 void Select(SDNode *N) override; 65 const char *getPassName() const override; 66 void PreprocessISelDAG() override; 67 void PostprocessISelDAG() override; 68 69private: 70 bool isInlineImmediate(SDNode *N) const; 71 bool FoldOperand(SDValue &Src, SDValue &Sel, SDValue &Neg, SDValue &Abs, 72 const R600InstrInfo *TII); 73 bool FoldOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &); 74 bool FoldDotOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &); 75 76 bool isConstantLoad(const MemSDNode *N, int cbID) const; 77 bool isUniformBr(const SDNode *N) const; 78 79 SDNode *glueCopyToM0(SDNode *N) const; 80 81 const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const; 82 bool SelectGlobalValueConstantOffset(SDValue Addr, SDValue& IntPtr); 83 bool SelectGlobalValueVariableOffset(SDValue Addr, SDValue &BaseReg, 84 SDValue& Offset); 85 bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset); 86 bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset); 87 bool isDSOffsetLegal(const SDValue &Base, unsigned Offset, 88 unsigned OffsetBits) const; 89 bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const; 90 bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0, 91 SDValue &Offset1) const; 92 bool SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr, 93 SDValue &SOffset, SDValue &Offset, SDValue &Offen, 94 SDValue &Idxen, SDValue &Addr64, SDValue &GLC, SDValue &SLC, 95 SDValue &TFE) const; 96 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr, 97 SDValue &SOffset, SDValue &Offset, SDValue &GLC, 98 SDValue &SLC, SDValue &TFE) const; 99 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, 100 SDValue &VAddr, SDValue &SOffset, SDValue &Offset, 101 SDValue &SLC) const; 102 bool SelectMUBUFScratch(SDValue Addr, SDValue &RSrc, SDValue &VAddr, 103 SDValue &SOffset, SDValue &ImmOffset) const; 104 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset, 105 SDValue &Offset, SDValue &GLC, SDValue &SLC, 106 SDValue &TFE) const; 107 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset, 108 SDValue &Offset, SDValue &SLC) const; 109 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset, 110 SDValue &Offset) const; 111 bool SelectMUBUFConstant(SDValue Constant, 112 SDValue &SOffset, 113 SDValue &ImmOffset) const; 114 bool SelectMUBUFIntrinsicOffset(SDValue Offset, SDValue &SOffset, 115 SDValue &ImmOffset) const; 116 bool SelectMUBUFIntrinsicVOffset(SDValue Offset, SDValue &SOffset, 117 SDValue &ImmOffset, SDValue &VOffset) const; 118 119 bool SelectFlat(SDValue Addr, SDValue &VAddr, 120 SDValue &SLC, SDValue &TFE) const; 121 122 bool SelectSMRDOffset(SDValue ByteOffsetNode, SDValue &Offset, 123 bool &Imm) const; 124 bool SelectSMRD(SDValue Addr, SDValue &SBase, SDValue &Offset, 125 bool &Imm) const; 126 bool SelectSMRDImm(SDValue Addr, SDValue &SBase, SDValue &Offset) const; 127 bool SelectSMRDImm32(SDValue Addr, SDValue &SBase, SDValue &Offset) const; 128 bool SelectSMRDSgpr(SDValue Addr, SDValue &SBase, SDValue &Offset) const; 129 bool SelectSMRDBufferImm(SDValue Addr, SDValue &Offset) const; 130 bool SelectSMRDBufferImm32(SDValue Addr, SDValue &Offset) const; 131 bool SelectSMRDBufferSgpr(SDValue Addr, SDValue &Offset) const; 132 bool SelectMOVRELOffset(SDValue Index, SDValue &Base, SDValue &Offset) const; 133 bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const; 134 bool SelectVOP3NoMods(SDValue In, SDValue &Src, SDValue &SrcMods) const; 135 bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods, 136 SDValue &Clamp, SDValue &Omod) const; 137 bool SelectVOP3NoMods0(SDValue In, SDValue &Src, SDValue &SrcMods, 138 SDValue &Clamp, SDValue &Omod) const; 139 140 bool SelectVOP3Mods0Clamp(SDValue In, SDValue &Src, SDValue &SrcMods, 141 SDValue &Omod) const; 142 bool SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src, SDValue &SrcMods, 143 SDValue &Clamp, 144 SDValue &Omod) const; 145 146 void SelectADD_SUB_I64(SDNode *N); 147 void SelectDIV_SCALE(SDNode *N); 148 149 SDNode *getS_BFE(unsigned Opcode, const SDLoc &DL, SDValue Val, 150 uint32_t Offset, uint32_t Width); 151 void SelectS_BFEFromShifts(SDNode *N); 152 void SelectS_BFE(SDNode *N); 153 void SelectBRCOND(SDNode *N); 154 void SelectATOMIC_CMP_SWAP(SDNode *N); 155 156 // Include the pieces autogenerated from the target description. 157#include "AMDGPUGenDAGISel.inc" 158}; 159} // end anonymous namespace 160 161/// \brief This pass converts a legalized DAG into a AMDGPU-specific 162// DAG, ready for instruction scheduling. 163FunctionPass *llvm::createAMDGPUISelDag(TargetMachine &TM) { 164 return new AMDGPUDAGToDAGISel(TM); 165} 166 167AMDGPUDAGToDAGISel::AMDGPUDAGToDAGISel(TargetMachine &TM) 168 : SelectionDAGISel(TM) {} 169 170bool AMDGPUDAGToDAGISel::runOnMachineFunction(MachineFunction &MF) { 171 Subtarget = &MF.getSubtarget<AMDGPUSubtarget>(); 172 return SelectionDAGISel::runOnMachineFunction(MF); 173} 174 175AMDGPUDAGToDAGISel::~AMDGPUDAGToDAGISel() { 176} 177 178bool AMDGPUDAGToDAGISel::isInlineImmediate(SDNode *N) const { 179 const SITargetLowering *TL 180 = static_cast<const SITargetLowering *>(getTargetLowering()); 181 return TL->analyzeImmediate(N) == 0; 182} 183 184/// \brief Determine the register class for \p OpNo 185/// \returns The register class of the virtual register that will be used for 186/// the given operand number \OpNo or NULL if the register class cannot be 187/// determined. 188const TargetRegisterClass *AMDGPUDAGToDAGISel::getOperandRegClass(SDNode *N, 189 unsigned OpNo) const { 190 if (!N->isMachineOpcode()) 191 return nullptr; 192 193 switch (N->getMachineOpcode()) { 194 default: { 195 const MCInstrDesc &Desc = 196 Subtarget->getInstrInfo()->get(N->getMachineOpcode()); 197 unsigned OpIdx = Desc.getNumDefs() + OpNo; 198 if (OpIdx >= Desc.getNumOperands()) 199 return nullptr; 200 int RegClass = Desc.OpInfo[OpIdx].RegClass; 201 if (RegClass == -1) 202 return nullptr; 203 204 return Subtarget->getRegisterInfo()->getRegClass(RegClass); 205 } 206 case AMDGPU::REG_SEQUENCE: { 207 unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 208 const TargetRegisterClass *SuperRC = 209 Subtarget->getRegisterInfo()->getRegClass(RCID); 210 211 SDValue SubRegOp = N->getOperand(OpNo + 1); 212 unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue(); 213 return Subtarget->getRegisterInfo()->getSubClassWithSubReg(SuperRC, 214 SubRegIdx); 215 } 216 } 217} 218 219SDNode *AMDGPUDAGToDAGISel::glueCopyToM0(SDNode *N) const { 220 if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS || 221 cast<MemSDNode>(N)->getAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) 222 return N; 223 224 const SITargetLowering& Lowering = 225 *static_cast<const SITargetLowering*>(getTargetLowering()); 226 227 // Write max value to m0 before each load operation 228 229 SDValue M0 = Lowering.copyToM0(*CurDAG, CurDAG->getEntryNode(), SDLoc(N), 230 CurDAG->getTargetConstant(-1, SDLoc(N), MVT::i32)); 231 232 SDValue Glue = M0.getValue(1); 233 234 SmallVector <SDValue, 8> Ops; 235 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 236 Ops.push_back(N->getOperand(i)); 237 } 238 Ops.push_back(Glue); 239 CurDAG->MorphNodeTo(N, N->getOpcode(), N->getVTList(), Ops); 240 241 return N; 242} 243 244static unsigned selectSGPRVectorRegClassID(unsigned NumVectorElts) { 245 switch (NumVectorElts) { 246 case 1: 247 return AMDGPU::SReg_32RegClassID; 248 case 2: 249 return AMDGPU::SReg_64RegClassID; 250 case 4: 251 return AMDGPU::SReg_128RegClassID; 252 case 8: 253 return AMDGPU::SReg_256RegClassID; 254 case 16: 255 return AMDGPU::SReg_512RegClassID; 256 } 257 258 llvm_unreachable("invalid vector size"); 259} 260 261void AMDGPUDAGToDAGISel::Select(SDNode *N) { 262 unsigned int Opc = N->getOpcode(); 263 if (N->isMachineOpcode()) { 264 N->setNodeId(-1); 265 return; // Already selected. 266 } 267 268 if (isa<AtomicSDNode>(N) || 269 (Opc == AMDGPUISD::ATOMIC_INC || Opc == AMDGPUISD::ATOMIC_DEC)) 270 N = glueCopyToM0(N); 271 272 switch (Opc) { 273 default: break; 274 // We are selecting i64 ADD here instead of custom lower it during 275 // DAG legalization, so we can fold some i64 ADDs used for address 276 // calculation into the LOAD and STORE instructions. 277 case ISD::ADD: 278 case ISD::SUB: { 279 if (N->getValueType(0) != MVT::i64 || 280 Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS) 281 break; 282 283 SelectADD_SUB_I64(N); 284 return; 285 } 286 case ISD::SCALAR_TO_VECTOR: 287 case AMDGPUISD::BUILD_VERTICAL_VECTOR: 288 case ISD::BUILD_VECTOR: { 289 unsigned RegClassID; 290 const AMDGPURegisterInfo *TRI = Subtarget->getRegisterInfo(); 291 EVT VT = N->getValueType(0); 292 unsigned NumVectorElts = VT.getVectorNumElements(); 293 EVT EltVT = VT.getVectorElementType(); 294 assert(EltVT.bitsEq(MVT::i32)); 295 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) { 296 RegClassID = selectSGPRVectorRegClassID(NumVectorElts); 297 } else { 298 // BUILD_VECTOR was lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG 299 // that adds a 128 bits reg copy when going through TwoAddressInstructions 300 // pass. We want to avoid 128 bits copies as much as possible because they 301 // can't be bundled by our scheduler. 302 switch(NumVectorElts) { 303 case 2: RegClassID = AMDGPU::R600_Reg64RegClassID; break; 304 case 4: 305 if (Opc == AMDGPUISD::BUILD_VERTICAL_VECTOR) 306 RegClassID = AMDGPU::R600_Reg128VerticalRegClassID; 307 else 308 RegClassID = AMDGPU::R600_Reg128RegClassID; 309 break; 310 default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR"); 311 } 312 } 313 314 SDLoc DL(N); 315 SDValue RegClass = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32); 316 317 if (NumVectorElts == 1) { 318 CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS, EltVT, N->getOperand(0), 319 RegClass); 320 return; 321 } 322 323 assert(NumVectorElts <= 16 && "Vectors with more than 16 elements not " 324 "supported yet"); 325 // 16 = Max Num Vector Elements 326 // 2 = 2 REG_SEQUENCE operands per element (value, subreg index) 327 // 1 = Vector Register Class 328 SmallVector<SDValue, 16 * 2 + 1> RegSeqArgs(NumVectorElts * 2 + 1); 329 330 RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32); 331 bool IsRegSeq = true; 332 unsigned NOps = N->getNumOperands(); 333 for (unsigned i = 0; i < NOps; i++) { 334 // XXX: Why is this here? 335 if (isa<RegisterSDNode>(N->getOperand(i))) { 336 IsRegSeq = false; 337 break; 338 } 339 RegSeqArgs[1 + (2 * i)] = N->getOperand(i); 340 RegSeqArgs[1 + (2 * i) + 1] = 341 CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), DL, 342 MVT::i32); 343 } 344 345 if (NOps != NumVectorElts) { 346 // Fill in the missing undef elements if this was a scalar_to_vector. 347 assert(Opc == ISD::SCALAR_TO_VECTOR && NOps < NumVectorElts); 348 349 MachineSDNode *ImpDef = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, 350 DL, EltVT); 351 for (unsigned i = NOps; i < NumVectorElts; ++i) { 352 RegSeqArgs[1 + (2 * i)] = SDValue(ImpDef, 0); 353 RegSeqArgs[1 + (2 * i) + 1] = 354 CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), DL, MVT::i32); 355 } 356 } 357 358 if (!IsRegSeq) 359 break; 360 CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(), RegSeqArgs); 361 return; 362 } 363 case ISD::BUILD_PAIR: { 364 SDValue RC, SubReg0, SubReg1; 365 if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) { 366 break; 367 } 368 SDLoc DL(N); 369 if (N->getValueType(0) == MVT::i128) { 370 RC = CurDAG->getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32); 371 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32); 372 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32); 373 } else if (N->getValueType(0) == MVT::i64) { 374 RC = CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32); 375 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32); 376 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32); 377 } else { 378 llvm_unreachable("Unhandled value type for BUILD_PAIR"); 379 } 380 const SDValue Ops[] = { RC, N->getOperand(0), SubReg0, 381 N->getOperand(1), SubReg1 }; 382 ReplaceNode(N, CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL, 383 N->getValueType(0), Ops)); 384 return; 385 } 386 387 case ISD::Constant: 388 case ISD::ConstantFP: { 389 if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS || 390 N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N)) 391 break; 392 393 uint64_t Imm; 394 if (ConstantFPSDNode *FP = dyn_cast<ConstantFPSDNode>(N)) 395 Imm = FP->getValueAPF().bitcastToAPInt().getZExtValue(); 396 else { 397 ConstantSDNode *C = cast<ConstantSDNode>(N); 398 Imm = C->getZExtValue(); 399 } 400 401 SDLoc DL(N); 402 SDNode *Lo = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, 403 CurDAG->getConstant(Imm & 0xFFFFFFFF, DL, 404 MVT::i32)); 405 SDNode *Hi = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, 406 CurDAG->getConstant(Imm >> 32, DL, MVT::i32)); 407 const SDValue Ops[] = { 408 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32), 409 SDValue(Lo, 0), CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32), 410 SDValue(Hi, 0), CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32) 411 }; 412 413 ReplaceNode(N, CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL, 414 N->getValueType(0), Ops)); 415 return; 416 } 417 case ISD::LOAD: 418 case ISD::STORE: { 419 N = glueCopyToM0(N); 420 break; 421 } 422 423 case AMDGPUISD::BFE_I32: 424 case AMDGPUISD::BFE_U32: { 425 if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS) 426 break; 427 428 // There is a scalar version available, but unlike the vector version which 429 // has a separate operand for the offset and width, the scalar version packs 430 // the width and offset into a single operand. Try to move to the scalar 431 // version if the offsets are constant, so that we can try to keep extended 432 // loads of kernel arguments in SGPRs. 433 434 // TODO: Technically we could try to pattern match scalar bitshifts of 435 // dynamic values, but it's probably not useful. 436 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); 437 if (!Offset) 438 break; 439 440 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2)); 441 if (!Width) 442 break; 443 444 bool Signed = Opc == AMDGPUISD::BFE_I32; 445 446 uint32_t OffsetVal = Offset->getZExtValue(); 447 uint32_t WidthVal = Width->getZExtValue(); 448 449 ReplaceNode(N, getS_BFE(Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32, 450 SDLoc(N), N->getOperand(0), OffsetVal, WidthVal)); 451 return; 452 } 453 case AMDGPUISD::DIV_SCALE: { 454 SelectDIV_SCALE(N); 455 return; 456 } 457 case ISD::CopyToReg: { 458 const SITargetLowering& Lowering = 459 *static_cast<const SITargetLowering*>(getTargetLowering()); 460 Lowering.legalizeTargetIndependentNode(N, *CurDAG); 461 break; 462 } 463 case ISD::AND: 464 case ISD::SRL: 465 case ISD::SRA: 466 case ISD::SIGN_EXTEND_INREG: 467 if (N->getValueType(0) != MVT::i32 || 468 Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS) 469 break; 470 471 SelectS_BFE(N); 472 return; 473 case ISD::BRCOND: 474 SelectBRCOND(N); 475 return; 476 477 case AMDGPUISD::ATOMIC_CMP_SWAP: 478 SelectATOMIC_CMP_SWAP(N); 479 return; 480 } 481 482 SelectCode(N); 483} 484 485bool AMDGPUDAGToDAGISel::isConstantLoad(const MemSDNode *N, int CbId) const { 486 if (!N->readMem()) 487 return false; 488 if (CbId == -1) 489 return N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS; 490 491 return N->getAddressSpace() == AMDGPUAS::CONSTANT_BUFFER_0 + CbId; 492} 493 494bool AMDGPUDAGToDAGISel::isUniformBr(const SDNode *N) const { 495 const BasicBlock *BB = FuncInfo->MBB->getBasicBlock(); 496 const Instruction *Term = BB->getTerminator(); 497 return Term->getMetadata("amdgpu.uniform") || 498 Term->getMetadata("structurizecfg.uniform"); 499} 500 501const char *AMDGPUDAGToDAGISel::getPassName() const { 502 return "AMDGPU DAG->DAG Pattern Instruction Selection"; 503} 504 505//===----------------------------------------------------------------------===// 506// Complex Patterns 507//===----------------------------------------------------------------------===// 508 509bool AMDGPUDAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr, 510 SDValue& IntPtr) { 511 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) { 512 IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, SDLoc(Addr), 513 true); 514 return true; 515 } 516 return false; 517} 518 519bool AMDGPUDAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr, 520 SDValue& BaseReg, SDValue &Offset) { 521 if (!isa<ConstantSDNode>(Addr)) { 522 BaseReg = Addr; 523 Offset = CurDAG->getIntPtrConstant(0, SDLoc(Addr), true); 524 return true; 525 } 526 return false; 527} 528 529bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base, 530 SDValue &Offset) { 531 ConstantSDNode *IMMOffset; 532 533 if (Addr.getOpcode() == ISD::ADD 534 && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1))) 535 && isInt<16>(IMMOffset->getZExtValue())) { 536 537 Base = Addr.getOperand(0); 538 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr), 539 MVT::i32); 540 return true; 541 // If the pointer address is constant, we can move it to the offset field. 542 } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr)) 543 && isInt<16>(IMMOffset->getZExtValue())) { 544 Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), 545 SDLoc(CurDAG->getEntryNode()), 546 AMDGPU::ZERO, MVT::i32); 547 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr), 548 MVT::i32); 549 return true; 550 } 551 552 // Default case, no offset 553 Base = Addr; 554 Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i32); 555 return true; 556} 557 558bool AMDGPUDAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base, 559 SDValue &Offset) { 560 ConstantSDNode *C; 561 SDLoc DL(Addr); 562 563 if ((C = dyn_cast<ConstantSDNode>(Addr))) { 564 Base = CurDAG->getRegister(AMDGPU::INDIRECT_BASE_ADDR, MVT::i32); 565 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32); 566 } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) && 567 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) { 568 Base = Addr.getOperand(0); 569 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32); 570 } else { 571 Base = Addr; 572 Offset = CurDAG->getTargetConstant(0, DL, MVT::i32); 573 } 574 575 return true; 576} 577 578void AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) { 579 SDLoc DL(N); 580 SDValue LHS = N->getOperand(0); 581 SDValue RHS = N->getOperand(1); 582 583 bool IsAdd = (N->getOpcode() == ISD::ADD); 584 585 SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32); 586 SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32); 587 588 SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 589 DL, MVT::i32, LHS, Sub0); 590 SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 591 DL, MVT::i32, LHS, Sub1); 592 593 SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 594 DL, MVT::i32, RHS, Sub0); 595 SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 596 DL, MVT::i32, RHS, Sub1); 597 598 SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue); 599 SDValue AddLoArgs[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) }; 600 601 unsigned Opc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32; 602 unsigned CarryOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32; 603 604 SDNode *AddLo = CurDAG->getMachineNode( Opc, DL, VTList, AddLoArgs); 605 SDValue Carry(AddLo, 1); 606 SDNode *AddHi 607 = CurDAG->getMachineNode(CarryOpc, DL, MVT::i32, 608 SDValue(Hi0, 0), SDValue(Hi1, 0), Carry); 609 610 SDValue Args[5] = { 611 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32), 612 SDValue(AddLo,0), 613 Sub0, 614 SDValue(AddHi,0), 615 Sub1, 616 }; 617 CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, MVT::i64, Args); 618} 619 620// We need to handle this here because tablegen doesn't support matching 621// instructions with multiple outputs. 622void AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) { 623 SDLoc SL(N); 624 EVT VT = N->getValueType(0); 625 626 assert(VT == MVT::f32 || VT == MVT::f64); 627 628 unsigned Opc 629 = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32; 630 631 // src0_modifiers, src0, src1_modifiers, src1, src2_modifiers, src2, clamp, 632 // omod 633 SDValue Ops[8]; 634 635 SelectVOP3Mods0(N->getOperand(0), Ops[1], Ops[0], Ops[6], Ops[7]); 636 SelectVOP3Mods(N->getOperand(1), Ops[3], Ops[2]); 637 SelectVOP3Mods(N->getOperand(2), Ops[5], Ops[4]); 638 CurDAG->SelectNodeTo(N, Opc, VT, MVT::i1, Ops); 639} 640 641bool AMDGPUDAGToDAGISel::isDSOffsetLegal(const SDValue &Base, unsigned Offset, 642 unsigned OffsetBits) const { 643 if ((OffsetBits == 16 && !isUInt<16>(Offset)) || 644 (OffsetBits == 8 && !isUInt<8>(Offset))) 645 return false; 646 647 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS || 648 Subtarget->unsafeDSOffsetFoldingEnabled()) 649 return true; 650 651 // On Southern Islands instruction with a negative base value and an offset 652 // don't seem to work. 653 return CurDAG->SignBitIsZero(Base); 654} 655 656bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base, 657 SDValue &Offset) const { 658 SDLoc DL(Addr); 659 if (CurDAG->isBaseWithConstantOffset(Addr)) { 660 SDValue N0 = Addr.getOperand(0); 661 SDValue N1 = Addr.getOperand(1); 662 ConstantSDNode *C1 = cast<ConstantSDNode>(N1); 663 if (isDSOffsetLegal(N0, C1->getSExtValue(), 16)) { 664 // (add n0, c0) 665 Base = N0; 666 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16); 667 return true; 668 } 669 } else if (Addr.getOpcode() == ISD::SUB) { 670 // sub C, x -> add (sub 0, x), C 671 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr.getOperand(0))) { 672 int64_t ByteOffset = C->getSExtValue(); 673 if (isUInt<16>(ByteOffset)) { 674 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32); 675 676 // XXX - This is kind of hacky. Create a dummy sub node so we can check 677 // the known bits in isDSOffsetLegal. We need to emit the selected node 678 // here, so this is thrown away. 679 SDValue Sub = CurDAG->getNode(ISD::SUB, DL, MVT::i32, 680 Zero, Addr.getOperand(1)); 681 682 if (isDSOffsetLegal(Sub, ByteOffset, 16)) { 683 MachineSDNode *MachineSub 684 = CurDAG->getMachineNode(AMDGPU::V_SUB_I32_e32, DL, MVT::i32, 685 Zero, Addr.getOperand(1)); 686 687 Base = SDValue(MachineSub, 0); 688 Offset = CurDAG->getTargetConstant(ByteOffset, DL, MVT::i16); 689 return true; 690 } 691 } 692 } 693 } else if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) { 694 // If we have a constant address, prefer to put the constant into the 695 // offset. This can save moves to load the constant address since multiple 696 // operations can share the zero base address register, and enables merging 697 // into read2 / write2 instructions. 698 699 SDLoc DL(Addr); 700 701 if (isUInt<16>(CAddr->getZExtValue())) { 702 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32); 703 MachineSDNode *MovZero = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32, 704 DL, MVT::i32, Zero); 705 Base = SDValue(MovZero, 0); 706 Offset = CurDAG->getTargetConstant(CAddr->getZExtValue(), DL, MVT::i16); 707 return true; 708 } 709 } 710 711 // default case 712 Base = Addr; 713 Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i16); 714 return true; 715} 716 717// TODO: If offset is too big, put low 16-bit into offset. 718bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base, 719 SDValue &Offset0, 720 SDValue &Offset1) const { 721 SDLoc DL(Addr); 722 723 if (CurDAG->isBaseWithConstantOffset(Addr)) { 724 SDValue N0 = Addr.getOperand(0); 725 SDValue N1 = Addr.getOperand(1); 726 ConstantSDNode *C1 = cast<ConstantSDNode>(N1); 727 unsigned DWordOffset0 = C1->getZExtValue() / 4; 728 unsigned DWordOffset1 = DWordOffset0 + 1; 729 // (add n0, c0) 730 if (isDSOffsetLegal(N0, DWordOffset1, 8)) { 731 Base = N0; 732 Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8); 733 Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8); 734 return true; 735 } 736 } else if (Addr.getOpcode() == ISD::SUB) { 737 // sub C, x -> add (sub 0, x), C 738 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr.getOperand(0))) { 739 unsigned DWordOffset0 = C->getZExtValue() / 4; 740 unsigned DWordOffset1 = DWordOffset0 + 1; 741 742 if (isUInt<8>(DWordOffset0)) { 743 SDLoc DL(Addr); 744 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32); 745 746 // XXX - This is kind of hacky. Create a dummy sub node so we can check 747 // the known bits in isDSOffsetLegal. We need to emit the selected node 748 // here, so this is thrown away. 749 SDValue Sub = CurDAG->getNode(ISD::SUB, DL, MVT::i32, 750 Zero, Addr.getOperand(1)); 751 752 if (isDSOffsetLegal(Sub, DWordOffset1, 8)) { 753 MachineSDNode *MachineSub 754 = CurDAG->getMachineNode(AMDGPU::V_SUB_I32_e32, DL, MVT::i32, 755 Zero, Addr.getOperand(1)); 756 757 Base = SDValue(MachineSub, 0); 758 Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8); 759 Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8); 760 return true; 761 } 762 } 763 } 764 } else if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) { 765 unsigned DWordOffset0 = CAddr->getZExtValue() / 4; 766 unsigned DWordOffset1 = DWordOffset0 + 1; 767 assert(4 * DWordOffset0 == CAddr->getZExtValue()); 768 769 if (isUInt<8>(DWordOffset0) && isUInt<8>(DWordOffset1)) { 770 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32); 771 MachineSDNode *MovZero 772 = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32, 773 DL, MVT::i32, Zero); 774 Base = SDValue(MovZero, 0); 775 Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8); 776 Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8); 777 return true; 778 } 779 } 780 781 // default case 782 Base = Addr; 783 Offset0 = CurDAG->getTargetConstant(0, DL, MVT::i8); 784 Offset1 = CurDAG->getTargetConstant(1, DL, MVT::i8); 785 return true; 786} 787 788static bool isLegalMUBUFImmOffset(const ConstantSDNode *Imm) { 789 return isUInt<12>(Imm->getZExtValue()); 790} 791 792bool AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr, 793 SDValue &VAddr, SDValue &SOffset, 794 SDValue &Offset, SDValue &Offen, 795 SDValue &Idxen, SDValue &Addr64, 796 SDValue &GLC, SDValue &SLC, 797 SDValue &TFE) const { 798 // Subtarget prefers to use flat instruction 799 if (Subtarget->useFlatForGlobal()) 800 return false; 801 802 SDLoc DL(Addr); 803 804 if (!GLC.getNode()) 805 GLC = CurDAG->getTargetConstant(0, DL, MVT::i1); 806 if (!SLC.getNode()) 807 SLC = CurDAG->getTargetConstant(0, DL, MVT::i1); 808 TFE = CurDAG->getTargetConstant(0, DL, MVT::i1); 809 810 Idxen = CurDAG->getTargetConstant(0, DL, MVT::i1); 811 Offen = CurDAG->getTargetConstant(0, DL, MVT::i1); 812 Addr64 = CurDAG->getTargetConstant(0, DL, MVT::i1); 813 SOffset = CurDAG->getTargetConstant(0, DL, MVT::i32); 814 815 if (CurDAG->isBaseWithConstantOffset(Addr)) { 816 SDValue N0 = Addr.getOperand(0); 817 SDValue N1 = Addr.getOperand(1); 818 ConstantSDNode *C1 = cast<ConstantSDNode>(N1); 819 820 if (N0.getOpcode() == ISD::ADD) { 821 // (add (add N2, N3), C1) -> addr64 822 SDValue N2 = N0.getOperand(0); 823 SDValue N3 = N0.getOperand(1); 824 Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1); 825 Ptr = N2; 826 VAddr = N3; 827 } else { 828 829 // (add N0, C1) -> offset 830 VAddr = CurDAG->getTargetConstant(0, DL, MVT::i32); 831 Ptr = N0; 832 } 833 834 if (isLegalMUBUFImmOffset(C1)) { 835 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16); 836 return true; 837 } 838 839 if (isUInt<32>(C1->getZExtValue())) { 840 // Illegal offset, store it in soffset. 841 Offset = CurDAG->getTargetConstant(0, DL, MVT::i16); 842 SOffset = SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, 843 CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32)), 844 0); 845 return true; 846 } 847 } 848 849 if (Addr.getOpcode() == ISD::ADD) { 850 // (add N0, N1) -> addr64 851 SDValue N0 = Addr.getOperand(0); 852 SDValue N1 = Addr.getOperand(1); 853 Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1); 854 Ptr = N0; 855 VAddr = N1; 856 Offset = CurDAG->getTargetConstant(0, DL, MVT::i16); 857 return true; 858 } 859 860 // default case -> offset 861 VAddr = CurDAG->getTargetConstant(0, DL, MVT::i32); 862 Ptr = Addr; 863 Offset = CurDAG->getTargetConstant(0, DL, MVT::i16); 864 865 return true; 866} 867 868bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, 869 SDValue &VAddr, SDValue &SOffset, 870 SDValue &Offset, SDValue &GLC, 871 SDValue &SLC, SDValue &TFE) const { 872 SDValue Ptr, Offen, Idxen, Addr64; 873 874 // addr64 bit was removed for volcanic islands. 875 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) 876 return false; 877 878 if (!SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64, 879 GLC, SLC, TFE)) 880 return false; 881 882 ConstantSDNode *C = cast<ConstantSDNode>(Addr64); 883 if (C->getSExtValue()) { 884 SDLoc DL(Addr); 885 886 const SITargetLowering& Lowering = 887 *static_cast<const SITargetLowering*>(getTargetLowering()); 888 889 SRsrc = SDValue(Lowering.wrapAddr64Rsrc(*CurDAG, DL, Ptr), 0); 890 return true; 891 } 892 893 return false; 894} 895 896bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, 897 SDValue &VAddr, SDValue &SOffset, 898 SDValue &Offset, 899 SDValue &SLC) const { 900 SLC = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i1); 901 SDValue GLC, TFE; 902 903 return SelectMUBUFAddr64(Addr, SRsrc, VAddr, SOffset, Offset, GLC, SLC, TFE); 904} 905 906bool AMDGPUDAGToDAGISel::SelectMUBUFScratch(SDValue Addr, SDValue &Rsrc, 907 SDValue &VAddr, SDValue &SOffset, 908 SDValue &ImmOffset) const { 909 910 SDLoc DL(Addr); 911 MachineFunction &MF = CurDAG->getMachineFunction(); 912 const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>(); 913 914 Rsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32); 915 SOffset = CurDAG->getRegister(Info->getScratchWaveOffsetReg(), MVT::i32); 916 917 // (add n0, c1) 918 if (CurDAG->isBaseWithConstantOffset(Addr)) { 919 SDValue N0 = Addr.getOperand(0); 920 SDValue N1 = Addr.getOperand(1); 921 922 // Offsets in vaddr must be positive. 923 ConstantSDNode *C1 = cast<ConstantSDNode>(N1); 924 if (isLegalMUBUFImmOffset(C1)) { 925 VAddr = N0; 926 ImmOffset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16); 927 return true; 928 } 929 } 930 931 // (node) 932 VAddr = Addr; 933 ImmOffset = CurDAG->getTargetConstant(0, DL, MVT::i16); 934 return true; 935} 936 937bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, 938 SDValue &SOffset, SDValue &Offset, 939 SDValue &GLC, SDValue &SLC, 940 SDValue &TFE) const { 941 SDValue Ptr, VAddr, Offen, Idxen, Addr64; 942 const SIInstrInfo *TII = 943 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo()); 944 945 if (!SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64, 946 GLC, SLC, TFE)) 947 return false; 948 949 if (!cast<ConstantSDNode>(Offen)->getSExtValue() && 950 !cast<ConstantSDNode>(Idxen)->getSExtValue() && 951 !cast<ConstantSDNode>(Addr64)->getSExtValue()) { 952 uint64_t Rsrc = TII->getDefaultRsrcDataFormat() | 953 APInt::getAllOnesValue(32).getZExtValue(); // Size 954 SDLoc DL(Addr); 955 956 const SITargetLowering& Lowering = 957 *static_cast<const SITargetLowering*>(getTargetLowering()); 958 959 SRsrc = SDValue(Lowering.buildRSRC(*CurDAG, DL, Ptr, 0, Rsrc), 0); 960 return true; 961 } 962 return false; 963} 964 965bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, 966 SDValue &Soffset, SDValue &Offset 967 ) const { 968 SDValue GLC, SLC, TFE; 969 970 return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE); 971} 972bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, 973 SDValue &Soffset, SDValue &Offset, 974 SDValue &SLC) const { 975 SDValue GLC, TFE; 976 977 return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE); 978} 979 980bool AMDGPUDAGToDAGISel::SelectMUBUFConstant(SDValue Constant, 981 SDValue &SOffset, 982 SDValue &ImmOffset) const { 983 SDLoc DL(Constant); 984 uint32_t Imm = cast<ConstantSDNode>(Constant)->getZExtValue(); 985 uint32_t Overflow = 0; 986 987 if (Imm >= 4096) { 988 if (Imm <= 4095 + 64) { 989 // Use an SOffset inline constant for 1..64 990 Overflow = Imm - 4095; 991 Imm = 4095; 992 } else { 993 // Try to keep the same value in SOffset for adjacent loads, so that 994 // the corresponding register contents can be re-used. 995 // 996 // Load values with all low-bits set into SOffset, so that a larger 997 // range of values can be covered using s_movk_i32 998 uint32_t High = (Imm + 1) & ~4095; 999 uint32_t Low = (Imm + 1) & 4095; 1000 Imm = Low; 1001 Overflow = High - 1; 1002 } 1003 } 1004 1005 // There is a hardware bug in SI and CI which prevents address clamping in 1006 // MUBUF instructions from working correctly with SOffsets. The immediate 1007 // offset is unaffected. 1008 if (Overflow > 0 && 1009 Subtarget->getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS) 1010 return false; 1011 1012 ImmOffset = CurDAG->getTargetConstant(Imm, DL, MVT::i16); 1013 1014 if (Overflow <= 64) 1015 SOffset = CurDAG->getTargetConstant(Overflow, DL, MVT::i32); 1016 else 1017 SOffset = SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, 1018 CurDAG->getTargetConstant(Overflow, DL, MVT::i32)), 1019 0); 1020 1021 return true; 1022} 1023 1024bool AMDGPUDAGToDAGISel::SelectMUBUFIntrinsicOffset(SDValue Offset, 1025 SDValue &SOffset, 1026 SDValue &ImmOffset) const { 1027 SDLoc DL(Offset); 1028 1029 if (!isa<ConstantSDNode>(Offset)) 1030 return false; 1031 1032 return SelectMUBUFConstant(Offset, SOffset, ImmOffset); 1033} 1034 1035bool AMDGPUDAGToDAGISel::SelectMUBUFIntrinsicVOffset(SDValue Offset, 1036 SDValue &SOffset, 1037 SDValue &ImmOffset, 1038 SDValue &VOffset) const { 1039 SDLoc DL(Offset); 1040 1041 // Don't generate an unnecessary voffset for constant offsets. 1042 if (isa<ConstantSDNode>(Offset)) { 1043 SDValue Tmp1, Tmp2; 1044 1045 // When necessary, use a voffset in <= CI anyway to work around a hardware 1046 // bug. 1047 if (Subtarget->getGeneration() > AMDGPUSubtarget::SEA_ISLANDS || 1048 SelectMUBUFConstant(Offset, Tmp1, Tmp2)) 1049 return false; 1050 } 1051 1052 if (CurDAG->isBaseWithConstantOffset(Offset)) { 1053 SDValue N0 = Offset.getOperand(0); 1054 SDValue N1 = Offset.getOperand(1); 1055 if (cast<ConstantSDNode>(N1)->getSExtValue() >= 0 && 1056 SelectMUBUFConstant(N1, SOffset, ImmOffset)) { 1057 VOffset = N0; 1058 return true; 1059 } 1060 } 1061 1062 SOffset = CurDAG->getTargetConstant(0, DL, MVT::i32); 1063 ImmOffset = CurDAG->getTargetConstant(0, DL, MVT::i16); 1064 VOffset = Offset; 1065 1066 return true; 1067} 1068 1069bool AMDGPUDAGToDAGISel::SelectFlat(SDValue Addr, 1070 SDValue &VAddr, 1071 SDValue &SLC, 1072 SDValue &TFE) const { 1073 VAddr = Addr; 1074 TFE = SLC = CurDAG->getTargetConstant(0, SDLoc(), MVT::i1); 1075 return true; 1076} 1077 1078/// 1079/// \param EncodedOffset This is the immediate value that will be encoded 1080/// directly into the instruction. On SI/CI the \p EncodedOffset 1081/// will be in units of dwords and on VI+ it will be units of bytes. 1082static bool isLegalSMRDImmOffset(const AMDGPUSubtarget *ST, 1083 int64_t EncodedOffset) { 1084 return ST->getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS ? 1085 isUInt<8>(EncodedOffset) : isUInt<20>(EncodedOffset); 1086} 1087 1088bool AMDGPUDAGToDAGISel::SelectSMRDOffset(SDValue ByteOffsetNode, 1089 SDValue &Offset, bool &Imm) const { 1090 1091 // FIXME: Handle non-constant offsets. 1092 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ByteOffsetNode); 1093 if (!C) 1094 return false; 1095 1096 SDLoc SL(ByteOffsetNode); 1097 AMDGPUSubtarget::Generation Gen = Subtarget->getGeneration(); 1098 int64_t ByteOffset = C->getSExtValue(); 1099 int64_t EncodedOffset = Gen < AMDGPUSubtarget::VOLCANIC_ISLANDS ? 1100 ByteOffset >> 2 : ByteOffset; 1101 1102 if (isLegalSMRDImmOffset(Subtarget, EncodedOffset)) { 1103 Offset = CurDAG->getTargetConstant(EncodedOffset, SL, MVT::i32); 1104 Imm = true; 1105 return true; 1106 } 1107 1108 if (!isUInt<32>(EncodedOffset) || !isUInt<32>(ByteOffset)) 1109 return false; 1110 1111 if (Gen == AMDGPUSubtarget::SEA_ISLANDS && isUInt<32>(EncodedOffset)) { 1112 // 32-bit Immediates are supported on Sea Islands. 1113 Offset = CurDAG->getTargetConstant(EncodedOffset, SL, MVT::i32); 1114 } else { 1115 SDValue C32Bit = CurDAG->getTargetConstant(ByteOffset, SL, MVT::i32); 1116 Offset = SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32, 1117 C32Bit), 0); 1118 } 1119 Imm = false; 1120 return true; 1121} 1122 1123bool AMDGPUDAGToDAGISel::SelectSMRD(SDValue Addr, SDValue &SBase, 1124 SDValue &Offset, bool &Imm) const { 1125 1126 SDLoc SL(Addr); 1127 if (CurDAG->isBaseWithConstantOffset(Addr)) { 1128 SDValue N0 = Addr.getOperand(0); 1129 SDValue N1 = Addr.getOperand(1); 1130 1131 if (SelectSMRDOffset(N1, Offset, Imm)) { 1132 SBase = N0; 1133 return true; 1134 } 1135 } 1136 SBase = Addr; 1137 Offset = CurDAG->getTargetConstant(0, SL, MVT::i32); 1138 Imm = true; 1139 return true; 1140} 1141 1142bool AMDGPUDAGToDAGISel::SelectSMRDImm(SDValue Addr, SDValue &SBase, 1143 SDValue &Offset) const { 1144 bool Imm; 1145 return SelectSMRD(Addr, SBase, Offset, Imm) && Imm; 1146} 1147 1148bool AMDGPUDAGToDAGISel::SelectSMRDImm32(SDValue Addr, SDValue &SBase, 1149 SDValue &Offset) const { 1150 1151 if (Subtarget->getGeneration() != AMDGPUSubtarget::SEA_ISLANDS) 1152 return false; 1153 1154 bool Imm; 1155 if (!SelectSMRD(Addr, SBase, Offset, Imm)) 1156 return false; 1157 1158 return !Imm && isa<ConstantSDNode>(Offset); 1159} 1160 1161bool AMDGPUDAGToDAGISel::SelectSMRDSgpr(SDValue Addr, SDValue &SBase, 1162 SDValue &Offset) const { 1163 bool Imm; 1164 return SelectSMRD(Addr, SBase, Offset, Imm) && !Imm && 1165 !isa<ConstantSDNode>(Offset); 1166} 1167 1168bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm(SDValue Addr, 1169 SDValue &Offset) const { 1170 bool Imm; 1171 return SelectSMRDOffset(Addr, Offset, Imm) && Imm; 1172} 1173 1174bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm32(SDValue Addr, 1175 SDValue &Offset) const { 1176 if (Subtarget->getGeneration() != AMDGPUSubtarget::SEA_ISLANDS) 1177 return false; 1178 1179 bool Imm; 1180 if (!SelectSMRDOffset(Addr, Offset, Imm)) 1181 return false; 1182 1183 return !Imm && isa<ConstantSDNode>(Offset); 1184} 1185 1186bool AMDGPUDAGToDAGISel::SelectSMRDBufferSgpr(SDValue Addr, 1187 SDValue &Offset) const { 1188 bool Imm; 1189 return SelectSMRDOffset(Addr, Offset, Imm) && !Imm && 1190 !isa<ConstantSDNode>(Offset); 1191} 1192 1193bool AMDGPUDAGToDAGISel::SelectMOVRELOffset(SDValue Index, 1194 SDValue &Base, 1195 SDValue &Offset) const { 1196 SDLoc DL(Index); 1197 1198 if (CurDAG->isBaseWithConstantOffset(Index)) { 1199 SDValue N0 = Index.getOperand(0); 1200 SDValue N1 = Index.getOperand(1); 1201 ConstantSDNode *C1 = cast<ConstantSDNode>(N1); 1202 1203 // (add n0, c0) 1204 Base = N0; 1205 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32); 1206 return true; 1207 } 1208 1209 if (isa<ConstantSDNode>(Index)) 1210 return false; 1211 1212 Base = Index; 1213 Offset = CurDAG->getTargetConstant(0, DL, MVT::i32); 1214 return true; 1215} 1216 1217SDNode *AMDGPUDAGToDAGISel::getS_BFE(unsigned Opcode, const SDLoc &DL, 1218 SDValue Val, uint32_t Offset, 1219 uint32_t Width) { 1220 // Transformation function, pack the offset and width of a BFE into 1221 // the format expected by the S_BFE_I32 / S_BFE_U32. In the second 1222 // source, bits [5:0] contain the offset and bits [22:16] the width. 1223 uint32_t PackedVal = Offset | (Width << 16); 1224 SDValue PackedConst = CurDAG->getTargetConstant(PackedVal, DL, MVT::i32); 1225 1226 return CurDAG->getMachineNode(Opcode, DL, MVT::i32, Val, PackedConst); 1227} 1228 1229void AMDGPUDAGToDAGISel::SelectS_BFEFromShifts(SDNode *N) { 1230 // "(a << b) srl c)" ---> "BFE_U32 a, (c-b), (32-c) 1231 // "(a << b) sra c)" ---> "BFE_I32 a, (c-b), (32-c) 1232 // Predicate: 0 < b <= c < 32 1233 1234 const SDValue &Shl = N->getOperand(0); 1235 ConstantSDNode *B = dyn_cast<ConstantSDNode>(Shl->getOperand(1)); 1236 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); 1237 1238 if (B && C) { 1239 uint32_t BVal = B->getZExtValue(); 1240 uint32_t CVal = C->getZExtValue(); 1241 1242 if (0 < BVal && BVal <= CVal && CVal < 32) { 1243 bool Signed = N->getOpcode() == ISD::SRA; 1244 unsigned Opcode = Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32; 1245 1246 ReplaceNode(N, getS_BFE(Opcode, SDLoc(N), Shl.getOperand(0), CVal - BVal, 1247 32 - CVal)); 1248 return; 1249 } 1250 } 1251 SelectCode(N); 1252} 1253 1254void AMDGPUDAGToDAGISel::SelectS_BFE(SDNode *N) { 1255 switch (N->getOpcode()) { 1256 case ISD::AND: 1257 if (N->getOperand(0).getOpcode() == ISD::SRL) { 1258 // "(a srl b) & mask" ---> "BFE_U32 a, b, popcount(mask)" 1259 // Predicate: isMask(mask) 1260 const SDValue &Srl = N->getOperand(0); 1261 ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(Srl.getOperand(1)); 1262 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(N->getOperand(1)); 1263 1264 if (Shift && Mask) { 1265 uint32_t ShiftVal = Shift->getZExtValue(); 1266 uint32_t MaskVal = Mask->getZExtValue(); 1267 1268 if (isMask_32(MaskVal)) { 1269 uint32_t WidthVal = countPopulation(MaskVal); 1270 1271 ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N), 1272 Srl.getOperand(0), ShiftVal, WidthVal)); 1273 return; 1274 } 1275 } 1276 } 1277 break; 1278 case ISD::SRL: 1279 if (N->getOperand(0).getOpcode() == ISD::AND) { 1280 // "(a & mask) srl b)" ---> "BFE_U32 a, b, popcount(mask >> b)" 1281 // Predicate: isMask(mask >> b) 1282 const SDValue &And = N->getOperand(0); 1283 ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(N->getOperand(1)); 1284 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(And->getOperand(1)); 1285 1286 if (Shift && Mask) { 1287 uint32_t ShiftVal = Shift->getZExtValue(); 1288 uint32_t MaskVal = Mask->getZExtValue() >> ShiftVal; 1289 1290 if (isMask_32(MaskVal)) { 1291 uint32_t WidthVal = countPopulation(MaskVal); 1292 1293 ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N), 1294 And.getOperand(0), ShiftVal, WidthVal)); 1295 return; 1296 } 1297 } 1298 } else if (N->getOperand(0).getOpcode() == ISD::SHL) { 1299 SelectS_BFEFromShifts(N); 1300 return; 1301 } 1302 break; 1303 case ISD::SRA: 1304 if (N->getOperand(0).getOpcode() == ISD::SHL) { 1305 SelectS_BFEFromShifts(N); 1306 return; 1307 } 1308 break; 1309 1310 case ISD::SIGN_EXTEND_INREG: { 1311 // sext_inreg (srl x, 16), i8 -> bfe_i32 x, 16, 8 1312 SDValue Src = N->getOperand(0); 1313 if (Src.getOpcode() != ISD::SRL) 1314 break; 1315 1316 const ConstantSDNode *Amt = dyn_cast<ConstantSDNode>(Src.getOperand(1)); 1317 if (!Amt) 1318 break; 1319 1320 unsigned Width = cast<VTSDNode>(N->getOperand(1))->getVT().getSizeInBits(); 1321 ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_I32, SDLoc(N), Src.getOperand(0), 1322 Amt->getZExtValue(), Width)); 1323 return; 1324 } 1325 } 1326 1327 SelectCode(N); 1328} 1329 1330void AMDGPUDAGToDAGISel::SelectBRCOND(SDNode *N) { 1331 SDValue Cond = N->getOperand(1); 1332 1333 if (isCBranchSCC(N)) { 1334 // This brcond will use S_CBRANCH_SCC*, so let tablegen handle it. 1335 SelectCode(N); 1336 return; 1337 } 1338 1339 // The result of VOPC instructions is or'd against ~EXEC before it is 1340 // written to vcc or another SGPR. This means that the value '1' is always 1341 // written to the corresponding bit for results that are masked. In order 1342 // to correctly check against vccz, we need to and VCC with the EXEC 1343 // register in order to clear the value from the masked bits. 1344 1345 SDLoc SL(N); 1346 1347 SDNode *MaskedCond = 1348 CurDAG->getMachineNode(AMDGPU::S_AND_B64, SL, MVT::i1, 1349 CurDAG->getRegister(AMDGPU::EXEC, MVT::i1), 1350 Cond); 1351 SDValue VCC = CurDAG->getCopyToReg(N->getOperand(0), SL, AMDGPU::VCC, 1352 SDValue(MaskedCond, 0), 1353 SDValue()); // Passing SDValue() adds a 1354 // glue output. 1355 CurDAG->SelectNodeTo(N, AMDGPU::S_CBRANCH_VCCNZ, MVT::Other, 1356 N->getOperand(2), // Basic Block 1357 VCC.getValue(0), // Chain 1358 VCC.getValue(1)); // Glue 1359 return; 1360} 1361 1362// This is here because there isn't a way to use the generated sub0_sub1 as the 1363// subreg index to EXTRACT_SUBREG in tablegen. 1364void AMDGPUDAGToDAGISel::SelectATOMIC_CMP_SWAP(SDNode *N) { 1365 MemSDNode *Mem = cast<MemSDNode>(N); 1366 unsigned AS = Mem->getAddressSpace(); 1367 if (AS == AMDGPUAS::FLAT_ADDRESS) { 1368 SelectCode(N); 1369 return; 1370 } 1371 1372 MVT VT = N->getSimpleValueType(0); 1373 bool Is32 = (VT == MVT::i32); 1374 SDLoc SL(N); 1375 1376 MachineSDNode *CmpSwap = nullptr; 1377 if (Subtarget->hasAddr64()) { 1378 SDValue SRsrc, VAddr, SOffset, Offset, GLC, SLC; 1379 1380 if (SelectMUBUFAddr64(Mem->getBasePtr(), SRsrc, VAddr, SOffset, Offset, SLC)) { 1381 unsigned Opcode = Is32 ? AMDGPU::BUFFER_ATOMIC_CMPSWAP_RTN_ADDR64 : 1382 AMDGPU::BUFFER_ATOMIC_CMPSWAP_X2_RTN_ADDR64; 1383 SDValue CmpVal = Mem->getOperand(2); 1384 1385 // XXX - Do we care about glue operands? 1386 1387 SDValue Ops[] = { 1388 CmpVal, VAddr, SRsrc, SOffset, Offset, SLC, Mem->getChain() 1389 }; 1390 1391 CmpSwap = CurDAG->getMachineNode(Opcode, SL, Mem->getVTList(), Ops); 1392 } 1393 } 1394 1395 if (!CmpSwap) { 1396 SDValue SRsrc, SOffset, Offset, SLC; 1397 if (SelectMUBUFOffset(Mem->getBasePtr(), SRsrc, SOffset, Offset, SLC)) { 1398 unsigned Opcode = Is32 ? AMDGPU::BUFFER_ATOMIC_CMPSWAP_RTN_OFFSET : 1399 AMDGPU::BUFFER_ATOMIC_CMPSWAP_X2_RTN_OFFSET; 1400 1401 SDValue CmpVal = Mem->getOperand(2); 1402 SDValue Ops[] = { 1403 CmpVal, SRsrc, SOffset, Offset, SLC, Mem->getChain() 1404 }; 1405 1406 CmpSwap = CurDAG->getMachineNode(Opcode, SL, Mem->getVTList(), Ops); 1407 } 1408 } 1409 1410 if (!CmpSwap) { 1411 SelectCode(N); 1412 return; 1413 } 1414 1415 MachineSDNode::mmo_iterator MMOs = MF->allocateMemRefsArray(1); 1416 *MMOs = Mem->getMemOperand(); 1417 CmpSwap->setMemRefs(MMOs, MMOs + 1); 1418 1419 unsigned SubReg = Is32 ? AMDGPU::sub0 : AMDGPU::sub0_sub1; 1420 SDValue Extract 1421 = CurDAG->getTargetExtractSubreg(SubReg, SL, VT, SDValue(CmpSwap, 0)); 1422 1423 ReplaceUses(SDValue(N, 0), Extract); 1424 ReplaceUses(SDValue(N, 1), SDValue(CmpSwap, 1)); 1425 CurDAG->RemoveDeadNode(N); 1426} 1427 1428bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src, 1429 SDValue &SrcMods) const { 1430 1431 unsigned Mods = 0; 1432 1433 Src = In; 1434 1435 if (Src.getOpcode() == ISD::FNEG) { 1436 Mods |= SISrcMods::NEG; 1437 Src = Src.getOperand(0); 1438 } 1439 1440 if (Src.getOpcode() == ISD::FABS) { 1441 Mods |= SISrcMods::ABS; 1442 Src = Src.getOperand(0); 1443 } 1444 1445 SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32); 1446 1447 return true; 1448} 1449 1450bool AMDGPUDAGToDAGISel::SelectVOP3NoMods(SDValue In, SDValue &Src, 1451 SDValue &SrcMods) const { 1452 bool Res = SelectVOP3Mods(In, Src, SrcMods); 1453 return Res && cast<ConstantSDNode>(SrcMods)->isNullValue(); 1454} 1455 1456bool AMDGPUDAGToDAGISel::SelectVOP3Mods0(SDValue In, SDValue &Src, 1457 SDValue &SrcMods, SDValue &Clamp, 1458 SDValue &Omod) const { 1459 SDLoc DL(In); 1460 // FIXME: Handle Clamp and Omod 1461 Clamp = CurDAG->getTargetConstant(0, DL, MVT::i32); 1462 Omod = CurDAG->getTargetConstant(0, DL, MVT::i32); 1463 1464 return SelectVOP3Mods(In, Src, SrcMods); 1465} 1466 1467bool AMDGPUDAGToDAGISel::SelectVOP3NoMods0(SDValue In, SDValue &Src, 1468 SDValue &SrcMods, SDValue &Clamp, 1469 SDValue &Omod) const { 1470 bool Res = SelectVOP3Mods0(In, Src, SrcMods, Clamp, Omod); 1471 1472 return Res && cast<ConstantSDNode>(SrcMods)->isNullValue() && 1473 cast<ConstantSDNode>(Clamp)->isNullValue() && 1474 cast<ConstantSDNode>(Omod)->isNullValue(); 1475} 1476 1477bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp(SDValue In, SDValue &Src, 1478 SDValue &SrcMods, 1479 SDValue &Omod) const { 1480 // FIXME: Handle Omod 1481 Omod = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32); 1482 1483 return SelectVOP3Mods(In, Src, SrcMods); 1484} 1485 1486bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src, 1487 SDValue &SrcMods, 1488 SDValue &Clamp, 1489 SDValue &Omod) const { 1490 Clamp = Omod = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32); 1491 return SelectVOP3Mods(In, Src, SrcMods); 1492} 1493 1494void AMDGPUDAGToDAGISel::PreprocessISelDAG() { 1495 MachineFrameInfo *MFI = CurDAG->getMachineFunction().getFrameInfo(); 1496 1497 // Handle the perverse case where a frame index is being stored. We don't 1498 // want to see multiple frame index operands on the same instruction since 1499 // it complicates things and violates some assumptions about frame index 1500 // lowering. 1501 for (int I = MFI->getObjectIndexBegin(), E = MFI->getObjectIndexEnd(); 1502 I != E; ++I) { 1503 SDValue FI = CurDAG->getTargetFrameIndex(I, MVT::i32); 1504 1505 // It's possible that we have a frame index defined in the function that 1506 // isn't used in this block. 1507 if (FI.use_empty()) 1508 continue; 1509 1510 // Skip over the AssertZext inserted during lowering. 1511 SDValue EffectiveFI = FI; 1512 auto It = FI->use_begin(); 1513 if (It->getOpcode() == ISD::AssertZext && FI->hasOneUse()) { 1514 EffectiveFI = SDValue(*It, 0); 1515 It = EffectiveFI->use_begin(); 1516 } 1517 1518 for (auto It = EffectiveFI->use_begin(); !It.atEnd(); ) { 1519 SDUse &Use = It.getUse(); 1520 SDNode *User = Use.getUser(); 1521 unsigned OpIdx = It.getOperandNo(); 1522 ++It; 1523 1524 if (MemSDNode *M = dyn_cast<MemSDNode>(User)) { 1525 unsigned PtrIdx = M->getOpcode() == ISD::STORE ? 2 : 1; 1526 if (OpIdx == PtrIdx) 1527 continue; 1528 1529 unsigned OpN = M->getNumOperands(); 1530 SDValue NewOps[8]; 1531 1532 assert(OpN < array_lengthof(NewOps)); 1533 for (unsigned Op = 0; Op != OpN; ++Op) { 1534 if (Op != OpIdx) { 1535 NewOps[Op] = M->getOperand(Op); 1536 continue; 1537 } 1538 1539 MachineSDNode *Mov = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32, 1540 SDLoc(M), MVT::i32, FI); 1541 NewOps[Op] = SDValue(Mov, 0); 1542 } 1543 1544 CurDAG->UpdateNodeOperands(M, makeArrayRef(NewOps, OpN)); 1545 } 1546 } 1547 } 1548} 1549 1550void AMDGPUDAGToDAGISel::PostprocessISelDAG() { 1551 const AMDGPUTargetLowering& Lowering = 1552 *static_cast<const AMDGPUTargetLowering*>(getTargetLowering()); 1553 bool IsModified = false; 1554 do { 1555 IsModified = false; 1556 // Go over all selected nodes and try to fold them a bit more 1557 for (SDNode &Node : CurDAG->allnodes()) { 1558 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(&Node); 1559 if (!MachineNode) 1560 continue; 1561 1562 SDNode *ResNode = Lowering.PostISelFolding(MachineNode, *CurDAG); 1563 if (ResNode != &Node) { 1564 ReplaceUses(&Node, ResNode); 1565 IsModified = true; 1566 } 1567 } 1568 CurDAG->RemoveDeadNodes(); 1569 } while (IsModified); 1570} 1571