1 //===-- AMDGPUISelDAGToDAG.cpp - A dag to dag inst selector for AMDGPU ----===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //==-----------------------------------------------------------------------===// 8 // 9 /// \file 10 /// Defines an instruction selector for the AMDGPU target. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "AMDGPU.h" 15 #include "AMDGPUArgumentUsageInfo.h" 16 #include "AMDGPUISelLowering.h" // For AMDGPUISD 17 #include "AMDGPUInstrInfo.h" 18 #include "AMDGPUPerfHintAnalysis.h" 19 #include "AMDGPUSubtarget.h" 20 #include "AMDGPUTargetMachine.h" 21 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 22 #include "SIDefines.h" 23 #include "SIISelLowering.h" 24 #include "SIInstrInfo.h" 25 #include "SIMachineFunctionInfo.h" 26 #include "SIRegisterInfo.h" 27 #include "llvm/ADT/APInt.h" 28 #include "llvm/ADT/SmallVector.h" 29 #include "llvm/ADT/StringRef.h" 30 #include "llvm/Analysis/LegacyDivergenceAnalysis.h" 31 #include "llvm/Analysis/LoopInfo.h" 32 #include "llvm/Analysis/ValueTracking.h" 33 #include "llvm/CodeGen/FunctionLoweringInfo.h" 34 #include "llvm/CodeGen/ISDOpcodes.h" 35 #include "llvm/CodeGen/MachineFunction.h" 36 #include "llvm/CodeGen/MachineRegisterInfo.h" 37 #include "llvm/CodeGen/SelectionDAG.h" 38 #include "llvm/CodeGen/SelectionDAGISel.h" 39 #include "llvm/CodeGen/SelectionDAGNodes.h" 40 #include "llvm/CodeGen/ValueTypes.h" 41 #include "llvm/IR/BasicBlock.h" 42 #include "llvm/InitializePasses.h" 43 #ifdef EXPENSIVE_CHECKS 44 #include "llvm/IR/Dominators.h" 45 #endif 46 #include "llvm/IR/Instruction.h" 47 #include "llvm/MC/MCInstrDesc.h" 48 #include "llvm/Support/Casting.h" 49 #include "llvm/Support/CodeGen.h" 50 #include "llvm/Support/ErrorHandling.h" 51 #include "llvm/Support/MachineValueType.h" 52 #include "llvm/Support/MathExtras.h" 53 #include <cassert> 54 #include <cstdint> 55 #include <new> 56 #include <vector> 57 58 #define DEBUG_TYPE "isel" 59 60 using namespace llvm; 61 62 namespace llvm { 63 64 class R600InstrInfo; 65 66 } // end namespace llvm 67 68 //===----------------------------------------------------------------------===// 69 // Instruction Selector Implementation 70 //===----------------------------------------------------------------------===// 71 72 namespace { 73 74 static bool isNullConstantOrUndef(SDValue V) { 75 if (V.isUndef()) 76 return true; 77 78 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V); 79 return Const != nullptr && Const->isNullValue(); 80 } 81 82 static bool getConstantValue(SDValue N, uint32_t &Out) { 83 // This is only used for packed vectors, where ussing 0 for undef should 84 // always be good. 85 if (N.isUndef()) { 86 Out = 0; 87 return true; 88 } 89 90 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) { 91 Out = C->getAPIntValue().getSExtValue(); 92 return true; 93 } 94 95 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N)) { 96 Out = C->getValueAPF().bitcastToAPInt().getSExtValue(); 97 return true; 98 } 99 100 return false; 101 } 102 103 // TODO: Handle undef as zero 104 static SDNode *packConstantV2I16(const SDNode *N, SelectionDAG &DAG, 105 bool Negate = false) { 106 assert(N->getOpcode() == ISD::BUILD_VECTOR && N->getNumOperands() == 2); 107 uint32_t LHSVal, RHSVal; 108 if (getConstantValue(N->getOperand(0), LHSVal) && 109 getConstantValue(N->getOperand(1), RHSVal)) { 110 SDLoc SL(N); 111 uint32_t K = Negate ? 112 (-LHSVal & 0xffff) | (-RHSVal << 16) : 113 (LHSVal & 0xffff) | (RHSVal << 16); 114 return DAG.getMachineNode(AMDGPU::S_MOV_B32, SL, N->getValueType(0), 115 DAG.getTargetConstant(K, SL, MVT::i32)); 116 } 117 118 return nullptr; 119 } 120 121 static SDNode *packNegConstantV2I16(const SDNode *N, SelectionDAG &DAG) { 122 return packConstantV2I16(N, DAG, true); 123 } 124 125 /// AMDGPU specific code to select AMDGPU machine instructions for 126 /// SelectionDAG operations. 127 class AMDGPUDAGToDAGISel : public SelectionDAGISel { 128 // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can 129 // make the right decision when generating code for different targets. 130 const GCNSubtarget *Subtarget; 131 132 // Default FP mode for the current function. 133 AMDGPU::SIModeRegisterDefaults Mode; 134 135 bool EnableLateStructurizeCFG; 136 137 public: 138 explicit AMDGPUDAGToDAGISel(TargetMachine *TM = nullptr, 139 CodeGenOpt::Level OptLevel = CodeGenOpt::Default) 140 : SelectionDAGISel(*TM, OptLevel) { 141 EnableLateStructurizeCFG = AMDGPUTargetMachine::EnableLateStructurizeCFG; 142 } 143 ~AMDGPUDAGToDAGISel() override = default; 144 145 void getAnalysisUsage(AnalysisUsage &AU) const override { 146 AU.addRequired<AMDGPUArgumentUsageInfo>(); 147 AU.addRequired<LegacyDivergenceAnalysis>(); 148 #ifdef EXPENSIVE_CHECKS 149 AU.addRequired<DominatorTreeWrapperPass>(); 150 AU.addRequired<LoopInfoWrapperPass>(); 151 #endif 152 SelectionDAGISel::getAnalysisUsage(AU); 153 } 154 155 bool matchLoadD16FromBuildVector(SDNode *N) const; 156 157 bool runOnMachineFunction(MachineFunction &MF) override; 158 void PreprocessISelDAG() override; 159 void Select(SDNode *N) override; 160 StringRef getPassName() const override; 161 void PostprocessISelDAG() override; 162 163 protected: 164 void SelectBuildVector(SDNode *N, unsigned RegClassID); 165 166 private: 167 std::pair<SDValue, SDValue> foldFrameIndex(SDValue N) const; 168 bool isNoNanSrc(SDValue N) const; 169 bool isInlineImmediate(const SDNode *N, bool Negated = false) const; 170 bool isNegInlineImmediate(const SDNode *N) const { 171 return isInlineImmediate(N, true); 172 } 173 174 bool isInlineImmediate16(int64_t Imm) const { 175 return AMDGPU::isInlinableLiteral16(Imm, Subtarget->hasInv2PiInlineImm()); 176 } 177 178 bool isInlineImmediate32(int64_t Imm) const { 179 return AMDGPU::isInlinableLiteral32(Imm, Subtarget->hasInv2PiInlineImm()); 180 } 181 182 bool isInlineImmediate64(int64_t Imm) const { 183 return AMDGPU::isInlinableLiteral64(Imm, Subtarget->hasInv2PiInlineImm()); 184 } 185 186 bool isInlineImmediate(const APFloat &Imm) const { 187 return Subtarget->getInstrInfo()->isInlineConstant(Imm); 188 } 189 190 bool isVGPRImm(const SDNode *N) const; 191 bool isUniformLoad(const SDNode *N) const; 192 bool isUniformBr(const SDNode *N) const; 193 194 MachineSDNode *buildSMovImm64(SDLoc &DL, uint64_t Val, EVT VT) const; 195 196 SDNode *glueCopyToOp(SDNode *N, SDValue NewChain, SDValue Glue) const; 197 SDNode *glueCopyToM0(SDNode *N, SDValue Val) const; 198 SDNode *glueCopyToM0LDSInit(SDNode *N) const; 199 200 const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const; 201 virtual bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset); 202 virtual bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset); 203 bool isDSOffsetLegal(SDValue Base, unsigned Offset, 204 unsigned OffsetBits) const; 205 bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const; 206 bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0, 207 SDValue &Offset1) const; 208 bool SelectDS128Bit8ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0, 209 SDValue &Offset1) const; 210 bool SelectDSReadWrite2(SDValue Ptr, SDValue &Base, SDValue &Offset0, 211 SDValue &Offset1, bool IsDS128) const; 212 bool SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr, 213 SDValue &SOffset, SDValue &Offset, SDValue &Offen, 214 SDValue &Idxen, SDValue &Addr64, SDValue &GLC, SDValue &SLC, 215 SDValue &TFE, SDValue &DLC, SDValue &SWZ) const; 216 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr, 217 SDValue &SOffset, SDValue &Offset, SDValue &GLC, 218 SDValue &SLC, SDValue &TFE, SDValue &DLC, 219 SDValue &SWZ) const; 220 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, 221 SDValue &VAddr, SDValue &SOffset, SDValue &Offset, 222 SDValue &SLC) const; 223 bool SelectMUBUFScratchOffen(SDNode *Parent, 224 SDValue Addr, SDValue &RSrc, SDValue &VAddr, 225 SDValue &SOffset, SDValue &ImmOffset) const; 226 bool SelectMUBUFScratchOffset(SDNode *Parent, 227 SDValue Addr, SDValue &SRsrc, SDValue &Soffset, 228 SDValue &Offset) const; 229 230 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset, 231 SDValue &Offset, SDValue &GLC, SDValue &SLC, 232 SDValue &TFE, SDValue &DLC, SDValue &SWZ) const; 233 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset, 234 SDValue &Offset, SDValue &SLC) const; 235 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset, 236 SDValue &Offset) const; 237 238 template <bool IsSigned> 239 bool SelectFlatOffset(SDNode *N, SDValue Addr, SDValue &VAddr, 240 SDValue &Offset) const; 241 bool SelectGlobalSAddr(SDNode *N, SDValue Addr, SDValue &SAddr, 242 SDValue &VOffset, SDValue &Offset) const; 243 bool SelectScratchSAddr(SDNode *N, SDValue Addr, SDValue &SAddr, 244 SDValue &Offset) const; 245 246 bool SelectSMRDOffset(SDValue ByteOffsetNode, SDValue &Offset, 247 bool &Imm) const; 248 SDValue Expand32BitAddress(SDValue Addr) const; 249 bool SelectSMRD(SDValue Addr, SDValue &SBase, SDValue &Offset, 250 bool &Imm) const; 251 bool SelectSMRDImm(SDValue Addr, SDValue &SBase, SDValue &Offset) const; 252 bool SelectSMRDImm32(SDValue Addr, SDValue &SBase, SDValue &Offset) const; 253 bool SelectSMRDSgpr(SDValue Addr, SDValue &SBase, SDValue &Offset) const; 254 bool SelectSMRDBufferImm(SDValue Addr, SDValue &Offset) const; 255 bool SelectSMRDBufferImm32(SDValue Addr, SDValue &Offset) const; 256 bool SelectMOVRELOffset(SDValue Index, SDValue &Base, SDValue &Offset) const; 257 258 bool SelectVOP3Mods_NNaN(SDValue In, SDValue &Src, SDValue &SrcMods) const; 259 bool SelectVOP3ModsImpl(SDValue In, SDValue &Src, unsigned &SrcMods) const; 260 bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const; 261 bool SelectVOP3NoMods(SDValue In, SDValue &Src) const; 262 bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods, 263 SDValue &Clamp, SDValue &Omod) const; 264 bool SelectVOP3NoMods0(SDValue In, SDValue &Src, SDValue &SrcMods, 265 SDValue &Clamp, SDValue &Omod) const; 266 267 bool SelectVOP3OMods(SDValue In, SDValue &Src, 268 SDValue &Clamp, SDValue &Omod) const; 269 270 bool SelectVOP3PMods(SDValue In, SDValue &Src, SDValue &SrcMods) const; 271 272 bool SelectVOP3OpSel(SDValue In, SDValue &Src, SDValue &SrcMods) const; 273 274 bool SelectVOP3OpSelMods(SDValue In, SDValue &Src, SDValue &SrcMods) const; 275 bool SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src, unsigned &Mods) const; 276 bool SelectVOP3PMadMixMods(SDValue In, SDValue &Src, SDValue &SrcMods) const; 277 278 SDValue getHi16Elt(SDValue In) const; 279 280 SDValue getMaterializedScalarImm32(int64_t Val, const SDLoc &DL) const; 281 282 void SelectADD_SUB_I64(SDNode *N); 283 void SelectAddcSubb(SDNode *N); 284 void SelectUADDO_USUBO(SDNode *N); 285 void SelectDIV_SCALE(SDNode *N); 286 void SelectMAD_64_32(SDNode *N); 287 void SelectFMA_W_CHAIN(SDNode *N); 288 void SelectFMUL_W_CHAIN(SDNode *N); 289 290 SDNode *getS_BFE(unsigned Opcode, const SDLoc &DL, SDValue Val, 291 uint32_t Offset, uint32_t Width); 292 void SelectS_BFEFromShifts(SDNode *N); 293 void SelectS_BFE(SDNode *N); 294 bool isCBranchSCC(const SDNode *N) const; 295 void SelectBRCOND(SDNode *N); 296 void SelectFMAD_FMA(SDNode *N); 297 void SelectATOMIC_CMP_SWAP(SDNode *N); 298 void SelectDSAppendConsume(SDNode *N, unsigned IntrID); 299 void SelectDS_GWS(SDNode *N, unsigned IntrID); 300 void SelectInterpP1F16(SDNode *N); 301 void SelectINTRINSIC_W_CHAIN(SDNode *N); 302 void SelectINTRINSIC_WO_CHAIN(SDNode *N); 303 void SelectINTRINSIC_VOID(SDNode *N); 304 305 protected: 306 // Include the pieces autogenerated from the target description. 307 #include "AMDGPUGenDAGISel.inc" 308 }; 309 310 class R600DAGToDAGISel : public AMDGPUDAGToDAGISel { 311 const R600Subtarget *Subtarget; 312 313 bool isConstantLoad(const MemSDNode *N, int cbID) const; 314 bool SelectGlobalValueConstantOffset(SDValue Addr, SDValue& IntPtr); 315 bool SelectGlobalValueVariableOffset(SDValue Addr, SDValue &BaseReg, 316 SDValue& Offset); 317 public: 318 explicit R600DAGToDAGISel(TargetMachine *TM, CodeGenOpt::Level OptLevel) : 319 AMDGPUDAGToDAGISel(TM, OptLevel) {} 320 321 void Select(SDNode *N) override; 322 323 bool SelectADDRIndirect(SDValue Addr, SDValue &Base, 324 SDValue &Offset) override; 325 bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, 326 SDValue &Offset) override; 327 328 bool runOnMachineFunction(MachineFunction &MF) override; 329 330 void PreprocessISelDAG() override {} 331 332 protected: 333 // Include the pieces autogenerated from the target description. 334 #include "R600GenDAGISel.inc" 335 }; 336 337 static SDValue stripBitcast(SDValue Val) { 338 return Val.getOpcode() == ISD::BITCAST ? Val.getOperand(0) : Val; 339 } 340 341 // Figure out if this is really an extract of the high 16-bits of a dword. 342 static bool isExtractHiElt(SDValue In, SDValue &Out) { 343 In = stripBitcast(In); 344 if (In.getOpcode() != ISD::TRUNCATE) 345 return false; 346 347 SDValue Srl = In.getOperand(0); 348 if (Srl.getOpcode() == ISD::SRL) { 349 if (ConstantSDNode *ShiftAmt = dyn_cast<ConstantSDNode>(Srl.getOperand(1))) { 350 if (ShiftAmt->getZExtValue() == 16) { 351 Out = stripBitcast(Srl.getOperand(0)); 352 return true; 353 } 354 } 355 } 356 357 return false; 358 } 359 360 // Look through operations that obscure just looking at the low 16-bits of the 361 // same register. 362 static SDValue stripExtractLoElt(SDValue In) { 363 if (In.getOpcode() == ISD::TRUNCATE) { 364 SDValue Src = In.getOperand(0); 365 if (Src.getValueType().getSizeInBits() == 32) 366 return stripBitcast(Src); 367 } 368 369 return In; 370 } 371 372 } // end anonymous namespace 373 374 INITIALIZE_PASS_BEGIN(AMDGPUDAGToDAGISel, "amdgpu-isel", 375 "AMDGPU DAG->DAG Pattern Instruction Selection", false, false) 376 INITIALIZE_PASS_DEPENDENCY(AMDGPUArgumentUsageInfo) 377 INITIALIZE_PASS_DEPENDENCY(AMDGPUPerfHintAnalysis) 378 INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis) 379 #ifdef EXPENSIVE_CHECKS 380 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 381 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 382 #endif 383 INITIALIZE_PASS_END(AMDGPUDAGToDAGISel, "amdgpu-isel", 384 "AMDGPU DAG->DAG Pattern Instruction Selection", false, false) 385 386 /// This pass converts a legalized DAG into a AMDGPU-specific 387 // DAG, ready for instruction scheduling. 388 FunctionPass *llvm::createAMDGPUISelDag(TargetMachine *TM, 389 CodeGenOpt::Level OptLevel) { 390 return new AMDGPUDAGToDAGISel(TM, OptLevel); 391 } 392 393 /// This pass converts a legalized DAG into a R600-specific 394 // DAG, ready for instruction scheduling. 395 FunctionPass *llvm::createR600ISelDag(TargetMachine *TM, 396 CodeGenOpt::Level OptLevel) { 397 return new R600DAGToDAGISel(TM, OptLevel); 398 } 399 400 bool AMDGPUDAGToDAGISel::runOnMachineFunction(MachineFunction &MF) { 401 #ifdef EXPENSIVE_CHECKS 402 DominatorTree & DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 403 LoopInfo * LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 404 for (auto &L : LI->getLoopsInPreorder()) { 405 assert(L->isLCSSAForm(DT)); 406 } 407 #endif 408 Subtarget = &MF.getSubtarget<GCNSubtarget>(); 409 Mode = AMDGPU::SIModeRegisterDefaults(MF.getFunction()); 410 return SelectionDAGISel::runOnMachineFunction(MF); 411 } 412 413 bool AMDGPUDAGToDAGISel::matchLoadD16FromBuildVector(SDNode *N) const { 414 assert(Subtarget->d16PreservesUnusedBits()); 415 MVT VT = N->getValueType(0).getSimpleVT(); 416 if (VT != MVT::v2i16 && VT != MVT::v2f16) 417 return false; 418 419 SDValue Lo = N->getOperand(0); 420 SDValue Hi = N->getOperand(1); 421 422 LoadSDNode *LdHi = dyn_cast<LoadSDNode>(stripBitcast(Hi)); 423 424 // build_vector lo, (load ptr) -> load_d16_hi ptr, lo 425 // build_vector lo, (zextload ptr from i8) -> load_d16_hi_u8 ptr, lo 426 // build_vector lo, (sextload ptr from i8) -> load_d16_hi_i8 ptr, lo 427 428 // Need to check for possible indirect dependencies on the other half of the 429 // vector to avoid introducing a cycle. 430 if (LdHi && Hi.hasOneUse() && !LdHi->isPredecessorOf(Lo.getNode())) { 431 SDVTList VTList = CurDAG->getVTList(VT, MVT::Other); 432 433 SDValue TiedIn = CurDAG->getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), VT, Lo); 434 SDValue Ops[] = { 435 LdHi->getChain(), LdHi->getBasePtr(), TiedIn 436 }; 437 438 unsigned LoadOp = AMDGPUISD::LOAD_D16_HI; 439 if (LdHi->getMemoryVT() == MVT::i8) { 440 LoadOp = LdHi->getExtensionType() == ISD::SEXTLOAD ? 441 AMDGPUISD::LOAD_D16_HI_I8 : AMDGPUISD::LOAD_D16_HI_U8; 442 } else { 443 assert(LdHi->getMemoryVT() == MVT::i16); 444 } 445 446 SDValue NewLoadHi = 447 CurDAG->getMemIntrinsicNode(LoadOp, SDLoc(LdHi), VTList, 448 Ops, LdHi->getMemoryVT(), 449 LdHi->getMemOperand()); 450 451 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), NewLoadHi); 452 CurDAG->ReplaceAllUsesOfValueWith(SDValue(LdHi, 1), NewLoadHi.getValue(1)); 453 return true; 454 } 455 456 // build_vector (load ptr), hi -> load_d16_lo ptr, hi 457 // build_vector (zextload ptr from i8), hi -> load_d16_lo_u8 ptr, hi 458 // build_vector (sextload ptr from i8), hi -> load_d16_lo_i8 ptr, hi 459 LoadSDNode *LdLo = dyn_cast<LoadSDNode>(stripBitcast(Lo)); 460 if (LdLo && Lo.hasOneUse()) { 461 SDValue TiedIn = getHi16Elt(Hi); 462 if (!TiedIn || LdLo->isPredecessorOf(TiedIn.getNode())) 463 return false; 464 465 SDVTList VTList = CurDAG->getVTList(VT, MVT::Other); 466 unsigned LoadOp = AMDGPUISD::LOAD_D16_LO; 467 if (LdLo->getMemoryVT() == MVT::i8) { 468 LoadOp = LdLo->getExtensionType() == ISD::SEXTLOAD ? 469 AMDGPUISD::LOAD_D16_LO_I8 : AMDGPUISD::LOAD_D16_LO_U8; 470 } else { 471 assert(LdLo->getMemoryVT() == MVT::i16); 472 } 473 474 TiedIn = CurDAG->getNode(ISD::BITCAST, SDLoc(N), VT, TiedIn); 475 476 SDValue Ops[] = { 477 LdLo->getChain(), LdLo->getBasePtr(), TiedIn 478 }; 479 480 SDValue NewLoadLo = 481 CurDAG->getMemIntrinsicNode(LoadOp, SDLoc(LdLo), VTList, 482 Ops, LdLo->getMemoryVT(), 483 LdLo->getMemOperand()); 484 485 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), NewLoadLo); 486 CurDAG->ReplaceAllUsesOfValueWith(SDValue(LdLo, 1), NewLoadLo.getValue(1)); 487 return true; 488 } 489 490 return false; 491 } 492 493 void AMDGPUDAGToDAGISel::PreprocessISelDAG() { 494 if (!Subtarget->d16PreservesUnusedBits()) 495 return; 496 497 SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end(); 498 499 bool MadeChange = false; 500 while (Position != CurDAG->allnodes_begin()) { 501 SDNode *N = &*--Position; 502 if (N->use_empty()) 503 continue; 504 505 switch (N->getOpcode()) { 506 case ISD::BUILD_VECTOR: 507 MadeChange |= matchLoadD16FromBuildVector(N); 508 break; 509 default: 510 break; 511 } 512 } 513 514 if (MadeChange) { 515 CurDAG->RemoveDeadNodes(); 516 LLVM_DEBUG(dbgs() << "After PreProcess:\n"; 517 CurDAG->dump();); 518 } 519 } 520 521 bool AMDGPUDAGToDAGISel::isNoNanSrc(SDValue N) const { 522 if (TM.Options.NoNaNsFPMath) 523 return true; 524 525 // TODO: Move into isKnownNeverNaN 526 if (N->getFlags().hasNoNaNs()) 527 return true; 528 529 return CurDAG->isKnownNeverNaN(N); 530 } 531 532 bool AMDGPUDAGToDAGISel::isInlineImmediate(const SDNode *N, 533 bool Negated) const { 534 if (N->isUndef()) 535 return true; 536 537 const SIInstrInfo *TII = Subtarget->getInstrInfo(); 538 if (Negated) { 539 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) 540 return TII->isInlineConstant(-C->getAPIntValue()); 541 542 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N)) 543 return TII->isInlineConstant(-C->getValueAPF().bitcastToAPInt()); 544 545 } else { 546 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) 547 return TII->isInlineConstant(C->getAPIntValue()); 548 549 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N)) 550 return TII->isInlineConstant(C->getValueAPF().bitcastToAPInt()); 551 } 552 553 return false; 554 } 555 556 /// Determine the register class for \p OpNo 557 /// \returns The register class of the virtual register that will be used for 558 /// the given operand number \OpNo or NULL if the register class cannot be 559 /// determined. 560 const TargetRegisterClass *AMDGPUDAGToDAGISel::getOperandRegClass(SDNode *N, 561 unsigned OpNo) const { 562 if (!N->isMachineOpcode()) { 563 if (N->getOpcode() == ISD::CopyToReg) { 564 Register Reg = cast<RegisterSDNode>(N->getOperand(1))->getReg(); 565 if (Reg.isVirtual()) { 566 MachineRegisterInfo &MRI = CurDAG->getMachineFunction().getRegInfo(); 567 return MRI.getRegClass(Reg); 568 } 569 570 const SIRegisterInfo *TRI 571 = static_cast<const GCNSubtarget *>(Subtarget)->getRegisterInfo(); 572 return TRI->getPhysRegClass(Reg); 573 } 574 575 return nullptr; 576 } 577 578 switch (N->getMachineOpcode()) { 579 default: { 580 const MCInstrDesc &Desc = 581 Subtarget->getInstrInfo()->get(N->getMachineOpcode()); 582 unsigned OpIdx = Desc.getNumDefs() + OpNo; 583 if (OpIdx >= Desc.getNumOperands()) 584 return nullptr; 585 int RegClass = Desc.OpInfo[OpIdx].RegClass; 586 if (RegClass == -1) 587 return nullptr; 588 589 return Subtarget->getRegisterInfo()->getRegClass(RegClass); 590 } 591 case AMDGPU::REG_SEQUENCE: { 592 unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 593 const TargetRegisterClass *SuperRC = 594 Subtarget->getRegisterInfo()->getRegClass(RCID); 595 596 SDValue SubRegOp = N->getOperand(OpNo + 1); 597 unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue(); 598 return Subtarget->getRegisterInfo()->getSubClassWithSubReg(SuperRC, 599 SubRegIdx); 600 } 601 } 602 } 603 604 SDNode *AMDGPUDAGToDAGISel::glueCopyToOp(SDNode *N, SDValue NewChain, 605 SDValue Glue) const { 606 SmallVector <SDValue, 8> Ops; 607 Ops.push_back(NewChain); // Replace the chain. 608 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i) 609 Ops.push_back(N->getOperand(i)); 610 611 Ops.push_back(Glue); 612 return CurDAG->MorphNodeTo(N, N->getOpcode(), N->getVTList(), Ops); 613 } 614 615 SDNode *AMDGPUDAGToDAGISel::glueCopyToM0(SDNode *N, SDValue Val) const { 616 const SITargetLowering& Lowering = 617 *static_cast<const SITargetLowering*>(getTargetLowering()); 618 619 assert(N->getOperand(0).getValueType() == MVT::Other && "Expected chain"); 620 621 SDValue M0 = Lowering.copyToM0(*CurDAG, N->getOperand(0), SDLoc(N), Val); 622 return glueCopyToOp(N, M0, M0.getValue(1)); 623 } 624 625 SDNode *AMDGPUDAGToDAGISel::glueCopyToM0LDSInit(SDNode *N) const { 626 unsigned AS = cast<MemSDNode>(N)->getAddressSpace(); 627 if (AS == AMDGPUAS::LOCAL_ADDRESS) { 628 if (Subtarget->ldsRequiresM0Init()) 629 return glueCopyToM0(N, CurDAG->getTargetConstant(-1, SDLoc(N), MVT::i32)); 630 } else if (AS == AMDGPUAS::REGION_ADDRESS) { 631 MachineFunction &MF = CurDAG->getMachineFunction(); 632 unsigned Value = MF.getInfo<SIMachineFunctionInfo>()->getGDSSize(); 633 return 634 glueCopyToM0(N, CurDAG->getTargetConstant(Value, SDLoc(N), MVT::i32)); 635 } 636 return N; 637 } 638 639 MachineSDNode *AMDGPUDAGToDAGISel::buildSMovImm64(SDLoc &DL, uint64_t Imm, 640 EVT VT) const { 641 SDNode *Lo = CurDAG->getMachineNode( 642 AMDGPU::S_MOV_B32, DL, MVT::i32, 643 CurDAG->getTargetConstant(Imm & 0xFFFFFFFF, DL, MVT::i32)); 644 SDNode *Hi = 645 CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, 646 CurDAG->getTargetConstant(Imm >> 32, DL, MVT::i32)); 647 const SDValue Ops[] = { 648 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32), 649 SDValue(Lo, 0), CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32), 650 SDValue(Hi, 0), CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32)}; 651 652 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL, VT, Ops); 653 } 654 655 void AMDGPUDAGToDAGISel::SelectBuildVector(SDNode *N, unsigned RegClassID) { 656 EVT VT = N->getValueType(0); 657 unsigned NumVectorElts = VT.getVectorNumElements(); 658 EVT EltVT = VT.getVectorElementType(); 659 SDLoc DL(N); 660 SDValue RegClass = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32); 661 662 if (NumVectorElts == 1) { 663 CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS, EltVT, N->getOperand(0), 664 RegClass); 665 return; 666 } 667 668 assert(NumVectorElts <= 32 && "Vectors with more than 32 elements not " 669 "supported yet"); 670 // 32 = Max Num Vector Elements 671 // 2 = 2 REG_SEQUENCE operands per element (value, subreg index) 672 // 1 = Vector Register Class 673 SmallVector<SDValue, 32 * 2 + 1> RegSeqArgs(NumVectorElts * 2 + 1); 674 675 bool IsGCN = CurDAG->getSubtarget().getTargetTriple().getArch() == 676 Triple::amdgcn; 677 RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32); 678 bool IsRegSeq = true; 679 unsigned NOps = N->getNumOperands(); 680 for (unsigned i = 0; i < NOps; i++) { 681 // XXX: Why is this here? 682 if (isa<RegisterSDNode>(N->getOperand(i))) { 683 IsRegSeq = false; 684 break; 685 } 686 unsigned Sub = IsGCN ? SIRegisterInfo::getSubRegFromChannel(i) 687 : R600RegisterInfo::getSubRegFromChannel(i); 688 RegSeqArgs[1 + (2 * i)] = N->getOperand(i); 689 RegSeqArgs[1 + (2 * i) + 1] = CurDAG->getTargetConstant(Sub, DL, MVT::i32); 690 } 691 if (NOps != NumVectorElts) { 692 // Fill in the missing undef elements if this was a scalar_to_vector. 693 assert(N->getOpcode() == ISD::SCALAR_TO_VECTOR && NOps < NumVectorElts); 694 MachineSDNode *ImpDef = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, 695 DL, EltVT); 696 for (unsigned i = NOps; i < NumVectorElts; ++i) { 697 unsigned Sub = IsGCN ? SIRegisterInfo::getSubRegFromChannel(i) 698 : R600RegisterInfo::getSubRegFromChannel(i); 699 RegSeqArgs[1 + (2 * i)] = SDValue(ImpDef, 0); 700 RegSeqArgs[1 + (2 * i) + 1] = 701 CurDAG->getTargetConstant(Sub, DL, MVT::i32); 702 } 703 } 704 705 if (!IsRegSeq) 706 SelectCode(N); 707 CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(), RegSeqArgs); 708 } 709 710 void AMDGPUDAGToDAGISel::Select(SDNode *N) { 711 unsigned int Opc = N->getOpcode(); 712 if (N->isMachineOpcode()) { 713 N->setNodeId(-1); 714 return; // Already selected. 715 } 716 717 // isa<MemSDNode> almost works but is slightly too permissive for some DS 718 // intrinsics. 719 if (Opc == ISD::LOAD || Opc == ISD::STORE || isa<AtomicSDNode>(N) || 720 (Opc == AMDGPUISD::ATOMIC_INC || Opc == AMDGPUISD::ATOMIC_DEC || 721 Opc == ISD::ATOMIC_LOAD_FADD || 722 Opc == AMDGPUISD::ATOMIC_LOAD_FMIN || 723 Opc == AMDGPUISD::ATOMIC_LOAD_FMAX)) { 724 N = glueCopyToM0LDSInit(N); 725 SelectCode(N); 726 return; 727 } 728 729 switch (Opc) { 730 default: 731 break; 732 // We are selecting i64 ADD here instead of custom lower it during 733 // DAG legalization, so we can fold some i64 ADDs used for address 734 // calculation into the LOAD and STORE instructions. 735 case ISD::ADDC: 736 case ISD::ADDE: 737 case ISD::SUBC: 738 case ISD::SUBE: { 739 if (N->getValueType(0) != MVT::i64) 740 break; 741 742 SelectADD_SUB_I64(N); 743 return; 744 } 745 case ISD::ADDCARRY: 746 case ISD::SUBCARRY: 747 if (N->getValueType(0) != MVT::i32) 748 break; 749 750 SelectAddcSubb(N); 751 return; 752 case ISD::UADDO: 753 case ISD::USUBO: { 754 SelectUADDO_USUBO(N); 755 return; 756 } 757 case AMDGPUISD::FMUL_W_CHAIN: { 758 SelectFMUL_W_CHAIN(N); 759 return; 760 } 761 case AMDGPUISD::FMA_W_CHAIN: { 762 SelectFMA_W_CHAIN(N); 763 return; 764 } 765 766 case ISD::SCALAR_TO_VECTOR: 767 case ISD::BUILD_VECTOR: { 768 EVT VT = N->getValueType(0); 769 unsigned NumVectorElts = VT.getVectorNumElements(); 770 if (VT.getScalarSizeInBits() == 16) { 771 if (Opc == ISD::BUILD_VECTOR && NumVectorElts == 2) { 772 if (SDNode *Packed = packConstantV2I16(N, *CurDAG)) { 773 ReplaceNode(N, Packed); 774 return; 775 } 776 } 777 778 break; 779 } 780 781 assert(VT.getVectorElementType().bitsEq(MVT::i32)); 782 unsigned RegClassID = 783 SIRegisterInfo::getSGPRClassForBitWidth(NumVectorElts * 32)->getID(); 784 SelectBuildVector(N, RegClassID); 785 return; 786 } 787 case ISD::BUILD_PAIR: { 788 SDValue RC, SubReg0, SubReg1; 789 SDLoc DL(N); 790 if (N->getValueType(0) == MVT::i128) { 791 RC = CurDAG->getTargetConstant(AMDGPU::SGPR_128RegClassID, DL, MVT::i32); 792 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32); 793 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32); 794 } else if (N->getValueType(0) == MVT::i64) { 795 RC = CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32); 796 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32); 797 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32); 798 } else { 799 llvm_unreachable("Unhandled value type for BUILD_PAIR"); 800 } 801 const SDValue Ops[] = { RC, N->getOperand(0), SubReg0, 802 N->getOperand(1), SubReg1 }; 803 ReplaceNode(N, CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL, 804 N->getValueType(0), Ops)); 805 return; 806 } 807 808 case ISD::Constant: 809 case ISD::ConstantFP: { 810 if (N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N)) 811 break; 812 813 uint64_t Imm; 814 if (ConstantFPSDNode *FP = dyn_cast<ConstantFPSDNode>(N)) 815 Imm = FP->getValueAPF().bitcastToAPInt().getZExtValue(); 816 else { 817 ConstantSDNode *C = cast<ConstantSDNode>(N); 818 Imm = C->getZExtValue(); 819 } 820 821 SDLoc DL(N); 822 ReplaceNode(N, buildSMovImm64(DL, Imm, N->getValueType(0))); 823 return; 824 } 825 case AMDGPUISD::BFE_I32: 826 case AMDGPUISD::BFE_U32: { 827 // There is a scalar version available, but unlike the vector version which 828 // has a separate operand for the offset and width, the scalar version packs 829 // the width and offset into a single operand. Try to move to the scalar 830 // version if the offsets are constant, so that we can try to keep extended 831 // loads of kernel arguments in SGPRs. 832 833 // TODO: Technically we could try to pattern match scalar bitshifts of 834 // dynamic values, but it's probably not useful. 835 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); 836 if (!Offset) 837 break; 838 839 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2)); 840 if (!Width) 841 break; 842 843 bool Signed = Opc == AMDGPUISD::BFE_I32; 844 845 uint32_t OffsetVal = Offset->getZExtValue(); 846 uint32_t WidthVal = Width->getZExtValue(); 847 848 ReplaceNode(N, getS_BFE(Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32, 849 SDLoc(N), N->getOperand(0), OffsetVal, WidthVal)); 850 return; 851 } 852 case AMDGPUISD::DIV_SCALE: { 853 SelectDIV_SCALE(N); 854 return; 855 } 856 case AMDGPUISD::MAD_I64_I32: 857 case AMDGPUISD::MAD_U64_U32: { 858 SelectMAD_64_32(N); 859 return; 860 } 861 case ISD::CopyToReg: { 862 const SITargetLowering& Lowering = 863 *static_cast<const SITargetLowering*>(getTargetLowering()); 864 N = Lowering.legalizeTargetIndependentNode(N, *CurDAG); 865 break; 866 } 867 case ISD::AND: 868 case ISD::SRL: 869 case ISD::SRA: 870 case ISD::SIGN_EXTEND_INREG: 871 if (N->getValueType(0) != MVT::i32) 872 break; 873 874 SelectS_BFE(N); 875 return; 876 case ISD::BRCOND: 877 SelectBRCOND(N); 878 return; 879 case ISD::FMAD: 880 case ISD::FMA: 881 SelectFMAD_FMA(N); 882 return; 883 case AMDGPUISD::ATOMIC_CMP_SWAP: 884 SelectATOMIC_CMP_SWAP(N); 885 return; 886 case AMDGPUISD::CVT_PKRTZ_F16_F32: 887 case AMDGPUISD::CVT_PKNORM_I16_F32: 888 case AMDGPUISD::CVT_PKNORM_U16_F32: 889 case AMDGPUISD::CVT_PK_U16_U32: 890 case AMDGPUISD::CVT_PK_I16_I32: { 891 // Hack around using a legal type if f16 is illegal. 892 if (N->getValueType(0) == MVT::i32) { 893 MVT NewVT = Opc == AMDGPUISD::CVT_PKRTZ_F16_F32 ? MVT::v2f16 : MVT::v2i16; 894 N = CurDAG->MorphNodeTo(N, N->getOpcode(), CurDAG->getVTList(NewVT), 895 { N->getOperand(0), N->getOperand(1) }); 896 SelectCode(N); 897 return; 898 } 899 900 break; 901 } 902 case ISD::INTRINSIC_W_CHAIN: { 903 SelectINTRINSIC_W_CHAIN(N); 904 return; 905 } 906 case ISD::INTRINSIC_WO_CHAIN: { 907 SelectINTRINSIC_WO_CHAIN(N); 908 return; 909 } 910 case ISD::INTRINSIC_VOID: { 911 SelectINTRINSIC_VOID(N); 912 return; 913 } 914 } 915 916 SelectCode(N); 917 } 918 919 bool AMDGPUDAGToDAGISel::isUniformBr(const SDNode *N) const { 920 const BasicBlock *BB = FuncInfo->MBB->getBasicBlock(); 921 const Instruction *Term = BB->getTerminator(); 922 return Term->getMetadata("amdgpu.uniform") || 923 Term->getMetadata("structurizecfg.uniform"); 924 } 925 926 StringRef AMDGPUDAGToDAGISel::getPassName() const { 927 return "AMDGPU DAG->DAG Pattern Instruction Selection"; 928 } 929 930 //===----------------------------------------------------------------------===// 931 // Complex Patterns 932 //===----------------------------------------------------------------------===// 933 934 bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base, 935 SDValue &Offset) { 936 return false; 937 } 938 939 bool AMDGPUDAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base, 940 SDValue &Offset) { 941 ConstantSDNode *C; 942 SDLoc DL(Addr); 943 944 if ((C = dyn_cast<ConstantSDNode>(Addr))) { 945 Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32); 946 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32); 947 } else if ((Addr.getOpcode() == AMDGPUISD::DWORDADDR) && 948 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(0)))) { 949 Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32); 950 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32); 951 } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) && 952 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) { 953 Base = Addr.getOperand(0); 954 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32); 955 } else { 956 Base = Addr; 957 Offset = CurDAG->getTargetConstant(0, DL, MVT::i32); 958 } 959 960 return true; 961 } 962 963 SDValue AMDGPUDAGToDAGISel::getMaterializedScalarImm32(int64_t Val, 964 const SDLoc &DL) const { 965 SDNode *Mov = CurDAG->getMachineNode( 966 AMDGPU::S_MOV_B32, DL, MVT::i32, 967 CurDAG->getTargetConstant(Val, DL, MVT::i32)); 968 return SDValue(Mov, 0); 969 } 970 971 // FIXME: Should only handle addcarry/subcarry 972 void AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) { 973 SDLoc DL(N); 974 SDValue LHS = N->getOperand(0); 975 SDValue RHS = N->getOperand(1); 976 977 unsigned Opcode = N->getOpcode(); 978 bool ConsumeCarry = (Opcode == ISD::ADDE || Opcode == ISD::SUBE); 979 bool ProduceCarry = 980 ConsumeCarry || Opcode == ISD::ADDC || Opcode == ISD::SUBC; 981 bool IsAdd = Opcode == ISD::ADD || Opcode == ISD::ADDC || Opcode == ISD::ADDE; 982 983 SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32); 984 SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32); 985 986 SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 987 DL, MVT::i32, LHS, Sub0); 988 SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 989 DL, MVT::i32, LHS, Sub1); 990 991 SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 992 DL, MVT::i32, RHS, Sub0); 993 SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 994 DL, MVT::i32, RHS, Sub1); 995 996 SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue); 997 998 static const unsigned OpcMap[2][2][2] = { 999 {{AMDGPU::S_SUB_U32, AMDGPU::S_ADD_U32}, 1000 {AMDGPU::V_SUB_CO_U32_e32, AMDGPU::V_ADD_CO_U32_e32}}, 1001 {{AMDGPU::S_SUBB_U32, AMDGPU::S_ADDC_U32}, 1002 {AMDGPU::V_SUBB_U32_e32, AMDGPU::V_ADDC_U32_e32}}}; 1003 1004 unsigned Opc = OpcMap[0][N->isDivergent()][IsAdd]; 1005 unsigned CarryOpc = OpcMap[1][N->isDivergent()][IsAdd]; 1006 1007 SDNode *AddLo; 1008 if (!ConsumeCarry) { 1009 SDValue Args[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) }; 1010 AddLo = CurDAG->getMachineNode(Opc, DL, VTList, Args); 1011 } else { 1012 SDValue Args[] = { SDValue(Lo0, 0), SDValue(Lo1, 0), N->getOperand(2) }; 1013 AddLo = CurDAG->getMachineNode(CarryOpc, DL, VTList, Args); 1014 } 1015 SDValue AddHiArgs[] = { 1016 SDValue(Hi0, 0), 1017 SDValue(Hi1, 0), 1018 SDValue(AddLo, 1) 1019 }; 1020 SDNode *AddHi = CurDAG->getMachineNode(CarryOpc, DL, VTList, AddHiArgs); 1021 1022 SDValue RegSequenceArgs[] = { 1023 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32), 1024 SDValue(AddLo,0), 1025 Sub0, 1026 SDValue(AddHi,0), 1027 Sub1, 1028 }; 1029 SDNode *RegSequence = CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL, 1030 MVT::i64, RegSequenceArgs); 1031 1032 if (ProduceCarry) { 1033 // Replace the carry-use 1034 ReplaceUses(SDValue(N, 1), SDValue(AddHi, 1)); 1035 } 1036 1037 // Replace the remaining uses. 1038 ReplaceNode(N, RegSequence); 1039 } 1040 1041 void AMDGPUDAGToDAGISel::SelectAddcSubb(SDNode *N) { 1042 SDLoc DL(N); 1043 SDValue LHS = N->getOperand(0); 1044 SDValue RHS = N->getOperand(1); 1045 SDValue CI = N->getOperand(2); 1046 1047 if (N->isDivergent()) { 1048 unsigned Opc = N->getOpcode() == ISD::ADDCARRY ? AMDGPU::V_ADDC_U32_e64 1049 : AMDGPU::V_SUBB_U32_e64; 1050 CurDAG->SelectNodeTo( 1051 N, Opc, N->getVTList(), 1052 {LHS, RHS, CI, 1053 CurDAG->getTargetConstant(0, {}, MVT::i1) /*clamp bit*/}); 1054 } else { 1055 unsigned Opc = N->getOpcode() == ISD::ADDCARRY ? AMDGPU::S_ADD_CO_PSEUDO 1056 : AMDGPU::S_SUB_CO_PSEUDO; 1057 CurDAG->SelectNodeTo(N, Opc, N->getVTList(), {LHS, RHS, CI}); 1058 } 1059 } 1060 1061 void AMDGPUDAGToDAGISel::SelectUADDO_USUBO(SDNode *N) { 1062 // The name of the opcodes are misleading. v_add_i32/v_sub_i32 have unsigned 1063 // carry out despite the _i32 name. These were renamed in VI to _U32. 1064 // FIXME: We should probably rename the opcodes here. 1065 bool IsAdd = N->getOpcode() == ISD::UADDO; 1066 bool IsVALU = N->isDivergent(); 1067 1068 for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end(); UI != E; 1069 ++UI) 1070 if (UI.getUse().getResNo() == 1) { 1071 if ((IsAdd && (UI->getOpcode() != ISD::ADDCARRY)) || 1072 (!IsAdd && (UI->getOpcode() != ISD::SUBCARRY))) { 1073 IsVALU = true; 1074 break; 1075 } 1076 } 1077 1078 if (IsVALU) { 1079 unsigned Opc = IsAdd ? AMDGPU::V_ADD_CO_U32_e64 : AMDGPU::V_SUB_CO_U32_e64; 1080 1081 CurDAG->SelectNodeTo( 1082 N, Opc, N->getVTList(), 1083 {N->getOperand(0), N->getOperand(1), 1084 CurDAG->getTargetConstant(0, {}, MVT::i1) /*clamp bit*/}); 1085 } else { 1086 unsigned Opc = N->getOpcode() == ISD::UADDO ? AMDGPU::S_UADDO_PSEUDO 1087 : AMDGPU::S_USUBO_PSEUDO; 1088 1089 CurDAG->SelectNodeTo(N, Opc, N->getVTList(), 1090 {N->getOperand(0), N->getOperand(1)}); 1091 } 1092 } 1093 1094 void AMDGPUDAGToDAGISel::SelectFMA_W_CHAIN(SDNode *N) { 1095 SDLoc SL(N); 1096 // src0_modifiers, src0, src1_modifiers, src1, src2_modifiers, src2, clamp, omod 1097 SDValue Ops[10]; 1098 1099 SelectVOP3Mods0(N->getOperand(1), Ops[1], Ops[0], Ops[6], Ops[7]); 1100 SelectVOP3Mods(N->getOperand(2), Ops[3], Ops[2]); 1101 SelectVOP3Mods(N->getOperand(3), Ops[5], Ops[4]); 1102 Ops[8] = N->getOperand(0); 1103 Ops[9] = N->getOperand(4); 1104 1105 CurDAG->SelectNodeTo(N, AMDGPU::V_FMA_F32, N->getVTList(), Ops); 1106 } 1107 1108 void AMDGPUDAGToDAGISel::SelectFMUL_W_CHAIN(SDNode *N) { 1109 SDLoc SL(N); 1110 // src0_modifiers, src0, src1_modifiers, src1, clamp, omod 1111 SDValue Ops[8]; 1112 1113 SelectVOP3Mods0(N->getOperand(1), Ops[1], Ops[0], Ops[4], Ops[5]); 1114 SelectVOP3Mods(N->getOperand(2), Ops[3], Ops[2]); 1115 Ops[6] = N->getOperand(0); 1116 Ops[7] = N->getOperand(3); 1117 1118 CurDAG->SelectNodeTo(N, AMDGPU::V_MUL_F32_e64, N->getVTList(), Ops); 1119 } 1120 1121 // We need to handle this here because tablegen doesn't support matching 1122 // instructions with multiple outputs. 1123 void AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) { 1124 SDLoc SL(N); 1125 EVT VT = N->getValueType(0); 1126 1127 assert(VT == MVT::f32 || VT == MVT::f64); 1128 1129 unsigned Opc 1130 = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32; 1131 1132 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2) }; 1133 CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops); 1134 } 1135 1136 // We need to handle this here because tablegen doesn't support matching 1137 // instructions with multiple outputs. 1138 void AMDGPUDAGToDAGISel::SelectMAD_64_32(SDNode *N) { 1139 SDLoc SL(N); 1140 bool Signed = N->getOpcode() == AMDGPUISD::MAD_I64_I32; 1141 unsigned Opc = Signed ? AMDGPU::V_MAD_I64_I32 : AMDGPU::V_MAD_U64_U32; 1142 1143 SDValue Clamp = CurDAG->getTargetConstant(0, SL, MVT::i1); 1144 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2), 1145 Clamp }; 1146 CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops); 1147 } 1148 1149 bool AMDGPUDAGToDAGISel::isDSOffsetLegal(SDValue Base, unsigned Offset, 1150 unsigned OffsetBits) const { 1151 if ((OffsetBits == 16 && !isUInt<16>(Offset)) || 1152 (OffsetBits == 8 && !isUInt<8>(Offset))) 1153 return false; 1154 1155 if (Subtarget->hasUsableDSOffset() || 1156 Subtarget->unsafeDSOffsetFoldingEnabled()) 1157 return true; 1158 1159 // On Southern Islands instruction with a negative base value and an offset 1160 // don't seem to work. 1161 return CurDAG->SignBitIsZero(Base); 1162 } 1163 1164 bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base, 1165 SDValue &Offset) const { 1166 SDLoc DL(Addr); 1167 if (CurDAG->isBaseWithConstantOffset(Addr)) { 1168 SDValue N0 = Addr.getOperand(0); 1169 SDValue N1 = Addr.getOperand(1); 1170 ConstantSDNode *C1 = cast<ConstantSDNode>(N1); 1171 if (isDSOffsetLegal(N0, C1->getSExtValue(), 16)) { 1172 // (add n0, c0) 1173 Base = N0; 1174 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16); 1175 return true; 1176 } 1177 } else if (Addr.getOpcode() == ISD::SUB) { 1178 // sub C, x -> add (sub 0, x), C 1179 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr.getOperand(0))) { 1180 int64_t ByteOffset = C->getSExtValue(); 1181 if (isUInt<16>(ByteOffset)) { 1182 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32); 1183 1184 // XXX - This is kind of hacky. Create a dummy sub node so we can check 1185 // the known bits in isDSOffsetLegal. We need to emit the selected node 1186 // here, so this is thrown away. 1187 SDValue Sub = CurDAG->getNode(ISD::SUB, DL, MVT::i32, 1188 Zero, Addr.getOperand(1)); 1189 1190 if (isDSOffsetLegal(Sub, ByteOffset, 16)) { 1191 SmallVector<SDValue, 3> Opnds; 1192 Opnds.push_back(Zero); 1193 Opnds.push_back(Addr.getOperand(1)); 1194 1195 // FIXME: Select to VOP3 version for with-carry. 1196 unsigned SubOp = AMDGPU::V_SUB_CO_U32_e32; 1197 if (Subtarget->hasAddNoCarry()) { 1198 SubOp = AMDGPU::V_SUB_U32_e64; 1199 Opnds.push_back( 1200 CurDAG->getTargetConstant(0, {}, MVT::i1)); // clamp bit 1201 } 1202 1203 MachineSDNode *MachineSub = 1204 CurDAG->getMachineNode(SubOp, DL, MVT::i32, Opnds); 1205 1206 Base = SDValue(MachineSub, 0); 1207 Offset = CurDAG->getTargetConstant(ByteOffset, DL, MVT::i16); 1208 return true; 1209 } 1210 } 1211 } 1212 } else if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) { 1213 // If we have a constant address, prefer to put the constant into the 1214 // offset. This can save moves to load the constant address since multiple 1215 // operations can share the zero base address register, and enables merging 1216 // into read2 / write2 instructions. 1217 1218 SDLoc DL(Addr); 1219 1220 if (isUInt<16>(CAddr->getZExtValue())) { 1221 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32); 1222 MachineSDNode *MovZero = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32, 1223 DL, MVT::i32, Zero); 1224 Base = SDValue(MovZero, 0); 1225 Offset = CurDAG->getTargetConstant(CAddr->getZExtValue(), DL, MVT::i16); 1226 return true; 1227 } 1228 } 1229 1230 // default case 1231 Base = Addr; 1232 Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i16); 1233 return true; 1234 } 1235 1236 // TODO: If offset is too big, put low 16-bit into offset. 1237 bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base, 1238 SDValue &Offset0, 1239 SDValue &Offset1) const { 1240 return SelectDSReadWrite2(Addr, Base, Offset0, Offset1, false); 1241 } 1242 1243 bool AMDGPUDAGToDAGISel::SelectDS128Bit8ByteAligned(SDValue Addr, SDValue &Base, 1244 SDValue &Offset0, 1245 SDValue &Offset1) const { 1246 return SelectDSReadWrite2(Addr, Base, Offset0, Offset1, true); 1247 } 1248 1249 bool AMDGPUDAGToDAGISel::SelectDSReadWrite2(SDValue Addr, SDValue &Base, 1250 SDValue &Offset0, SDValue &Offset1, 1251 bool IsDS128) const { 1252 SDLoc DL(Addr); 1253 unsigned Align = IsDS128 ? 8 : 4; 1254 1255 if (CurDAG->isBaseWithConstantOffset(Addr)) { 1256 SDValue N0 = Addr.getOperand(0); 1257 SDValue N1 = Addr.getOperand(1); 1258 ConstantSDNode *C1 = cast<ConstantSDNode>(N1); 1259 unsigned OffsetValue0 = C1->getZExtValue() / Align; 1260 unsigned OffsetValue1 = OffsetValue0 + 1; 1261 // (add n0, c0) 1262 if (isDSOffsetLegal(N0, OffsetValue1, 8)) { 1263 Base = N0; 1264 Offset0 = CurDAG->getTargetConstant(OffsetValue0, DL, MVT::i8); 1265 Offset1 = CurDAG->getTargetConstant(OffsetValue1, DL, MVT::i8); 1266 return true; 1267 } 1268 } else if (Addr.getOpcode() == ISD::SUB) { 1269 // sub C, x -> add (sub 0, x), C 1270 if (const ConstantSDNode *C = 1271 dyn_cast<ConstantSDNode>(Addr.getOperand(0))) { 1272 unsigned OffsetValue0 = C->getZExtValue() / Align; 1273 unsigned OffsetValue1 = OffsetValue0 + 1; 1274 1275 if (isUInt<8>(OffsetValue0)) { 1276 SDLoc DL(Addr); 1277 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32); 1278 1279 // XXX - This is kind of hacky. Create a dummy sub node so we can check 1280 // the known bits in isDSOffsetLegal. We need to emit the selected node 1281 // here, so this is thrown away. 1282 SDValue Sub = 1283 CurDAG->getNode(ISD::SUB, DL, MVT::i32, Zero, Addr.getOperand(1)); 1284 1285 if (isDSOffsetLegal(Sub, OffsetValue1, 8)) { 1286 SmallVector<SDValue, 3> Opnds; 1287 Opnds.push_back(Zero); 1288 Opnds.push_back(Addr.getOperand(1)); 1289 unsigned SubOp = AMDGPU::V_SUB_CO_U32_e32; 1290 if (Subtarget->hasAddNoCarry()) { 1291 SubOp = AMDGPU::V_SUB_U32_e64; 1292 Opnds.push_back( 1293 CurDAG->getTargetConstant(0, {}, MVT::i1)); // clamp bit 1294 } 1295 1296 MachineSDNode *MachineSub = CurDAG->getMachineNode( 1297 SubOp, DL, (IsDS128 ? MVT::i64 : MVT::i32), Opnds); 1298 1299 Base = SDValue(MachineSub, 0); 1300 Offset0 = CurDAG->getTargetConstant(OffsetValue0, DL, MVT::i8); 1301 Offset1 = CurDAG->getTargetConstant(OffsetValue1, DL, MVT::i8); 1302 return true; 1303 } 1304 } 1305 } 1306 } else if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) { 1307 unsigned OffsetValue0 = CAddr->getZExtValue() / Align; 1308 unsigned OffsetValue1 = OffsetValue0 + 1; 1309 bool OffsetIsAligned = Align * OffsetValue0 == CAddr->getZExtValue(); 1310 1311 if (isUInt<8>(OffsetValue0) && isUInt<8>(OffsetValue1) && OffsetIsAligned) { 1312 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32); 1313 MachineSDNode *MovZero = 1314 CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32, DL, MVT::i32, Zero); 1315 Base = SDValue(MovZero, 0); 1316 Offset0 = CurDAG->getTargetConstant(OffsetValue0, DL, MVT::i8); 1317 Offset1 = CurDAG->getTargetConstant(OffsetValue1, DL, MVT::i8); 1318 return true; 1319 } 1320 } 1321 1322 // default case 1323 1324 Base = Addr; 1325 Offset0 = CurDAG->getTargetConstant(0, DL, MVT::i8); 1326 Offset1 = CurDAG->getTargetConstant(1, DL, MVT::i8); 1327 return true; 1328 } 1329 1330 bool AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr, 1331 SDValue &VAddr, SDValue &SOffset, 1332 SDValue &Offset, SDValue &Offen, 1333 SDValue &Idxen, SDValue &Addr64, 1334 SDValue &GLC, SDValue &SLC, 1335 SDValue &TFE, SDValue &DLC, 1336 SDValue &SWZ) const { 1337 // Subtarget prefers to use flat instruction 1338 // FIXME: This should be a pattern predicate and not reach here 1339 if (Subtarget->useFlatForGlobal()) 1340 return false; 1341 1342 SDLoc DL(Addr); 1343 1344 if (!GLC.getNode()) 1345 GLC = CurDAG->getTargetConstant(0, DL, MVT::i1); 1346 if (!SLC.getNode()) 1347 SLC = CurDAG->getTargetConstant(0, DL, MVT::i1); 1348 TFE = CurDAG->getTargetConstant(0, DL, MVT::i1); 1349 DLC = CurDAG->getTargetConstant(0, DL, MVT::i1); 1350 SWZ = CurDAG->getTargetConstant(0, DL, MVT::i1); 1351 1352 Idxen = CurDAG->getTargetConstant(0, DL, MVT::i1); 1353 Offen = CurDAG->getTargetConstant(0, DL, MVT::i1); 1354 Addr64 = CurDAG->getTargetConstant(0, DL, MVT::i1); 1355 SOffset = CurDAG->getTargetConstant(0, DL, MVT::i32); 1356 1357 ConstantSDNode *C1 = nullptr; 1358 SDValue N0 = Addr; 1359 if (CurDAG->isBaseWithConstantOffset(Addr)) { 1360 C1 = cast<ConstantSDNode>(Addr.getOperand(1)); 1361 if (isUInt<32>(C1->getZExtValue())) 1362 N0 = Addr.getOperand(0); 1363 else 1364 C1 = nullptr; 1365 } 1366 1367 if (N0.getOpcode() == ISD::ADD) { 1368 // (add N2, N3) -> addr64, or 1369 // (add (add N2, N3), C1) -> addr64 1370 SDValue N2 = N0.getOperand(0); 1371 SDValue N3 = N0.getOperand(1); 1372 Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1); 1373 1374 if (N2->isDivergent()) { 1375 if (N3->isDivergent()) { 1376 // Both N2 and N3 are divergent. Use N0 (the result of the add) as the 1377 // addr64, and construct the resource from a 0 address. 1378 Ptr = SDValue(buildSMovImm64(DL, 0, MVT::v2i32), 0); 1379 VAddr = N0; 1380 } else { 1381 // N2 is divergent, N3 is not. 1382 Ptr = N3; 1383 VAddr = N2; 1384 } 1385 } else { 1386 // N2 is not divergent. 1387 Ptr = N2; 1388 VAddr = N3; 1389 } 1390 Offset = CurDAG->getTargetConstant(0, DL, MVT::i16); 1391 } else if (N0->isDivergent()) { 1392 // N0 is divergent. Use it as the addr64, and construct the resource from a 1393 // 0 address. 1394 Ptr = SDValue(buildSMovImm64(DL, 0, MVT::v2i32), 0); 1395 VAddr = N0; 1396 Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1); 1397 } else { 1398 // N0 -> offset, or 1399 // (N0 + C1) -> offset 1400 VAddr = CurDAG->getTargetConstant(0, DL, MVT::i32); 1401 Ptr = N0; 1402 } 1403 1404 if (!C1) { 1405 // No offset. 1406 Offset = CurDAG->getTargetConstant(0, DL, MVT::i16); 1407 return true; 1408 } 1409 1410 if (SIInstrInfo::isLegalMUBUFImmOffset(C1->getZExtValue())) { 1411 // Legal offset for instruction. 1412 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16); 1413 return true; 1414 } 1415 1416 // Illegal offset, store it in soffset. 1417 Offset = CurDAG->getTargetConstant(0, DL, MVT::i16); 1418 SOffset = 1419 SDValue(CurDAG->getMachineNode( 1420 AMDGPU::S_MOV_B32, DL, MVT::i32, 1421 CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32)), 1422 0); 1423 return true; 1424 } 1425 1426 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, 1427 SDValue &VAddr, SDValue &SOffset, 1428 SDValue &Offset, SDValue &GLC, 1429 SDValue &SLC, SDValue &TFE, 1430 SDValue &DLC, SDValue &SWZ) const { 1431 SDValue Ptr, Offen, Idxen, Addr64; 1432 1433 // addr64 bit was removed for volcanic islands. 1434 // FIXME: This should be a pattern predicate and not reach here 1435 if (!Subtarget->hasAddr64()) 1436 return false; 1437 1438 if (!SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64, 1439 GLC, SLC, TFE, DLC, SWZ)) 1440 return false; 1441 1442 ConstantSDNode *C = cast<ConstantSDNode>(Addr64); 1443 if (C->getSExtValue()) { 1444 SDLoc DL(Addr); 1445 1446 const SITargetLowering& Lowering = 1447 *static_cast<const SITargetLowering*>(getTargetLowering()); 1448 1449 SRsrc = SDValue(Lowering.wrapAddr64Rsrc(*CurDAG, DL, Ptr), 0); 1450 return true; 1451 } 1452 1453 return false; 1454 } 1455 1456 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, 1457 SDValue &VAddr, SDValue &SOffset, 1458 SDValue &Offset, 1459 SDValue &SLC) const { 1460 SLC = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i1); 1461 SDValue GLC, TFE, DLC, SWZ; 1462 1463 return SelectMUBUFAddr64(Addr, SRsrc, VAddr, SOffset, Offset, GLC, SLC, TFE, DLC, SWZ); 1464 } 1465 1466 static bool isStackPtrRelative(const MachinePointerInfo &PtrInfo) { 1467 auto PSV = PtrInfo.V.dyn_cast<const PseudoSourceValue *>(); 1468 return PSV && PSV->isStack(); 1469 } 1470 1471 std::pair<SDValue, SDValue> AMDGPUDAGToDAGISel::foldFrameIndex(SDValue N) const { 1472 SDLoc DL(N); 1473 1474 auto *FI = dyn_cast<FrameIndexSDNode>(N); 1475 SDValue TFI = 1476 FI ? CurDAG->getTargetFrameIndex(FI->getIndex(), FI->getValueType(0)) : N; 1477 1478 // We rebase the base address into an absolute stack address and hence 1479 // use constant 0 for soffset. 1480 return std::make_pair(TFI, CurDAG->getTargetConstant(0, DL, MVT::i32)); 1481 } 1482 1483 bool AMDGPUDAGToDAGISel::SelectMUBUFScratchOffen(SDNode *Parent, 1484 SDValue Addr, SDValue &Rsrc, 1485 SDValue &VAddr, SDValue &SOffset, 1486 SDValue &ImmOffset) const { 1487 1488 SDLoc DL(Addr); 1489 MachineFunction &MF = CurDAG->getMachineFunction(); 1490 const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>(); 1491 1492 Rsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32); 1493 1494 if (ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) { 1495 int64_t Imm = CAddr->getSExtValue(); 1496 const int64_t NullPtr = 1497 AMDGPUTargetMachine::getNullPointerValue(AMDGPUAS::PRIVATE_ADDRESS); 1498 // Don't fold null pointer. 1499 if (Imm != NullPtr) { 1500 SDValue HighBits = CurDAG->getTargetConstant(Imm & ~4095, DL, MVT::i32); 1501 MachineSDNode *MovHighBits = CurDAG->getMachineNode( 1502 AMDGPU::V_MOV_B32_e32, DL, MVT::i32, HighBits); 1503 VAddr = SDValue(MovHighBits, 0); 1504 1505 // In a call sequence, stores to the argument stack area are relative to the 1506 // stack pointer. 1507 const MachinePointerInfo &PtrInfo 1508 = cast<MemSDNode>(Parent)->getPointerInfo(); 1509 SOffset = isStackPtrRelative(PtrInfo) 1510 ? CurDAG->getRegister(Info->getStackPtrOffsetReg(), MVT::i32) 1511 : CurDAG->getTargetConstant(0, DL, MVT::i32); 1512 ImmOffset = CurDAG->getTargetConstant(Imm & 4095, DL, MVT::i16); 1513 return true; 1514 } 1515 } 1516 1517 if (CurDAG->isBaseWithConstantOffset(Addr)) { 1518 // (add n0, c1) 1519 1520 SDValue N0 = Addr.getOperand(0); 1521 SDValue N1 = Addr.getOperand(1); 1522 1523 // Offsets in vaddr must be positive if range checking is enabled. 1524 // 1525 // The total computation of vaddr + soffset + offset must not overflow. If 1526 // vaddr is negative, even if offset is 0 the sgpr offset add will end up 1527 // overflowing. 1528 // 1529 // Prior to gfx9, MUBUF instructions with the vaddr offset enabled would 1530 // always perform a range check. If a negative vaddr base index was used, 1531 // this would fail the range check. The overall address computation would 1532 // compute a valid address, but this doesn't happen due to the range 1533 // check. For out-of-bounds MUBUF loads, a 0 is returned. 1534 // 1535 // Therefore it should be safe to fold any VGPR offset on gfx9 into the 1536 // MUBUF vaddr, but not on older subtargets which can only do this if the 1537 // sign bit is known 0. 1538 ConstantSDNode *C1 = cast<ConstantSDNode>(N1); 1539 if (SIInstrInfo::isLegalMUBUFImmOffset(C1->getZExtValue()) && 1540 (!Subtarget->privateMemoryResourceIsRangeChecked() || 1541 CurDAG->SignBitIsZero(N0))) { 1542 std::tie(VAddr, SOffset) = foldFrameIndex(N0); 1543 ImmOffset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16); 1544 return true; 1545 } 1546 } 1547 1548 // (node) 1549 std::tie(VAddr, SOffset) = foldFrameIndex(Addr); 1550 ImmOffset = CurDAG->getTargetConstant(0, DL, MVT::i16); 1551 return true; 1552 } 1553 1554 bool AMDGPUDAGToDAGISel::SelectMUBUFScratchOffset(SDNode *Parent, 1555 SDValue Addr, 1556 SDValue &SRsrc, 1557 SDValue &SOffset, 1558 SDValue &Offset) const { 1559 ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr); 1560 if (!CAddr || !SIInstrInfo::isLegalMUBUFImmOffset(CAddr->getZExtValue())) 1561 return false; 1562 1563 SDLoc DL(Addr); 1564 MachineFunction &MF = CurDAG->getMachineFunction(); 1565 const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>(); 1566 1567 SRsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32); 1568 1569 const MachinePointerInfo &PtrInfo = cast<MemSDNode>(Parent)->getPointerInfo(); 1570 1571 // FIXME: Get from MachinePointerInfo? We should only be using the frame 1572 // offset if we know this is in a call sequence. 1573 SOffset = isStackPtrRelative(PtrInfo) 1574 ? CurDAG->getRegister(Info->getStackPtrOffsetReg(), MVT::i32) 1575 : CurDAG->getTargetConstant(0, DL, MVT::i32); 1576 1577 Offset = CurDAG->getTargetConstant(CAddr->getZExtValue(), DL, MVT::i16); 1578 return true; 1579 } 1580 1581 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, 1582 SDValue &SOffset, SDValue &Offset, 1583 SDValue &GLC, SDValue &SLC, 1584 SDValue &TFE, SDValue &DLC, 1585 SDValue &SWZ) const { 1586 SDValue Ptr, VAddr, Offen, Idxen, Addr64; 1587 const SIInstrInfo *TII = 1588 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo()); 1589 1590 if (!SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64, 1591 GLC, SLC, TFE, DLC, SWZ)) 1592 return false; 1593 1594 if (!cast<ConstantSDNode>(Offen)->getSExtValue() && 1595 !cast<ConstantSDNode>(Idxen)->getSExtValue() && 1596 !cast<ConstantSDNode>(Addr64)->getSExtValue()) { 1597 uint64_t Rsrc = TII->getDefaultRsrcDataFormat() | 1598 APInt::getAllOnesValue(32).getZExtValue(); // Size 1599 SDLoc DL(Addr); 1600 1601 const SITargetLowering& Lowering = 1602 *static_cast<const SITargetLowering*>(getTargetLowering()); 1603 1604 SRsrc = SDValue(Lowering.buildRSRC(*CurDAG, DL, Ptr, 0, Rsrc), 0); 1605 return true; 1606 } 1607 return false; 1608 } 1609 1610 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, 1611 SDValue &Soffset, SDValue &Offset 1612 ) const { 1613 SDValue GLC, SLC, TFE, DLC, SWZ; 1614 1615 return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE, DLC, SWZ); 1616 } 1617 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, 1618 SDValue &Soffset, SDValue &Offset, 1619 SDValue &SLC) const { 1620 SDValue GLC, TFE, DLC, SWZ; 1621 1622 return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE, DLC, SWZ); 1623 } 1624 1625 // Find a load or store from corresponding pattern root. 1626 // Roots may be build_vector, bitconvert or their combinations. 1627 static MemSDNode* findMemSDNode(SDNode *N) { 1628 N = AMDGPUTargetLowering::stripBitcast(SDValue(N,0)).getNode(); 1629 if (MemSDNode *MN = dyn_cast<MemSDNode>(N)) 1630 return MN; 1631 assert(isa<BuildVectorSDNode>(N)); 1632 for (SDValue V : N->op_values()) 1633 if (MemSDNode *MN = 1634 dyn_cast<MemSDNode>(AMDGPUTargetLowering::stripBitcast(V))) 1635 return MN; 1636 llvm_unreachable("cannot find MemSDNode in the pattern!"); 1637 } 1638 1639 static bool getBaseWithOffsetUsingSplitOR(SelectionDAG &DAG, SDValue Addr, 1640 SDValue &N0, SDValue &N1) { 1641 if (Addr.getValueType() == MVT::i64 && Addr.getOpcode() == ISD::BITCAST && 1642 Addr.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) { 1643 // As we split 64-bit `or` earlier, it's complicated pattern to match, i.e. 1644 // (i64 (bitcast (v2i32 (build_vector 1645 // (or (extract_vector_elt V, 0), OFFSET), 1646 // (extract_vector_elt V, 1))))) 1647 SDValue Lo = Addr.getOperand(0).getOperand(0); 1648 if (Lo.getOpcode() == ISD::OR && DAG.isBaseWithConstantOffset(Lo)) { 1649 SDValue BaseLo = Lo.getOperand(0); 1650 SDValue BaseHi = Addr.getOperand(0).getOperand(1); 1651 // Check that split base (Lo and Hi) are extracted from the same one. 1652 if (BaseLo.getOpcode() == ISD::EXTRACT_VECTOR_ELT && 1653 BaseHi.getOpcode() == ISD::EXTRACT_VECTOR_ELT && 1654 BaseLo.getOperand(0) == BaseHi.getOperand(0) && 1655 // Lo is statically extracted from index 0. 1656 isa<ConstantSDNode>(BaseLo.getOperand(1)) && 1657 BaseLo.getConstantOperandVal(1) == 0 && 1658 // Hi is statically extracted from index 0. 1659 isa<ConstantSDNode>(BaseHi.getOperand(1)) && 1660 BaseHi.getConstantOperandVal(1) == 1) { 1661 N0 = BaseLo.getOperand(0).getOperand(0); 1662 N1 = Lo.getOperand(1); 1663 return true; 1664 } 1665 } 1666 } 1667 return false; 1668 } 1669 1670 template <bool IsSigned> 1671 bool AMDGPUDAGToDAGISel::SelectFlatOffset(SDNode *N, 1672 SDValue Addr, 1673 SDValue &VAddr, 1674 SDValue &Offset) const { 1675 int64_t OffsetVal = 0; 1676 1677 unsigned AS = findMemSDNode(N)->getAddressSpace(); 1678 1679 if (Subtarget->hasFlatInstOffsets() && 1680 (!Subtarget->hasFlatSegmentOffsetBug() || 1681 AS != AMDGPUAS::FLAT_ADDRESS)) { 1682 SDValue N0, N1; 1683 if (CurDAG->isBaseWithConstantOffset(Addr)) { 1684 N0 = Addr.getOperand(0); 1685 N1 = Addr.getOperand(1); 1686 } else if (getBaseWithOffsetUsingSplitOR(*CurDAG, Addr, N0, N1)) { 1687 assert(N0 && N1 && isa<ConstantSDNode>(N1)); 1688 } 1689 if (N0 && N1) { 1690 uint64_t COffsetVal = cast<ConstantSDNode>(N1)->getSExtValue(); 1691 1692 const SIInstrInfo *TII = Subtarget->getInstrInfo(); 1693 if (TII->isLegalFLATOffset(COffsetVal, AS, IsSigned)) { 1694 Addr = N0; 1695 OffsetVal = COffsetVal; 1696 } else { 1697 // If the offset doesn't fit, put the low bits into the offset field and 1698 // add the rest. 1699 // 1700 // For a FLAT instruction the hardware decides whether to access 1701 // global/scratch/shared memory based on the high bits of vaddr, 1702 // ignoring the offset field, so we have to ensure that when we add 1703 // remainder to vaddr it still points into the same underlying object. 1704 // The easiest way to do that is to make sure that we split the offset 1705 // into two pieces that are both >= 0 or both <= 0. 1706 1707 SDLoc DL(N); 1708 uint64_t RemainderOffset = COffsetVal; 1709 uint64_t ImmField = 0; 1710 const unsigned NumBits = TII->getNumFlatOffsetBits(IsSigned); 1711 if (IsSigned) { 1712 // Use signed division by a power of two to truncate towards 0. 1713 int64_t D = 1LL << (NumBits - 1); 1714 RemainderOffset = (static_cast<int64_t>(COffsetVal) / D) * D; 1715 ImmField = COffsetVal - RemainderOffset; 1716 } else if (static_cast<int64_t>(COffsetVal) >= 0) { 1717 ImmField = COffsetVal & maskTrailingOnes<uint64_t>(NumBits); 1718 RemainderOffset = COffsetVal - ImmField; 1719 } 1720 assert(TII->isLegalFLATOffset(ImmField, AS, IsSigned)); 1721 assert(RemainderOffset + ImmField == COffsetVal); 1722 1723 OffsetVal = ImmField; 1724 1725 SDValue AddOffsetLo = 1726 getMaterializedScalarImm32(Lo_32(RemainderOffset), DL); 1727 SDValue Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1); 1728 1729 if (Addr.getValueType().getSizeInBits() == 32) { 1730 SmallVector<SDValue, 3> Opnds; 1731 Opnds.push_back(N0); 1732 Opnds.push_back(AddOffsetLo); 1733 unsigned AddOp = AMDGPU::V_ADD_CO_U32_e32; 1734 if (Subtarget->hasAddNoCarry()) { 1735 AddOp = AMDGPU::V_ADD_U32_e64; 1736 Opnds.push_back(Clamp); 1737 } 1738 Addr = SDValue(CurDAG->getMachineNode(AddOp, DL, MVT::i32, Opnds), 0); 1739 } else { 1740 // TODO: Should this try to use a scalar add pseudo if the base address 1741 // is uniform and saddr is usable? 1742 SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32); 1743 SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32); 1744 1745 SDNode *N0Lo = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 1746 DL, MVT::i32, N0, Sub0); 1747 SDNode *N0Hi = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 1748 DL, MVT::i32, N0, Sub1); 1749 1750 SDValue AddOffsetHi = 1751 getMaterializedScalarImm32(Hi_32(RemainderOffset), DL); 1752 1753 SDVTList VTs = CurDAG->getVTList(MVT::i32, MVT::i1); 1754 1755 SDNode *Add = 1756 CurDAG->getMachineNode(AMDGPU::V_ADD_CO_U32_e64, DL, VTs, 1757 {AddOffsetLo, SDValue(N0Lo, 0), Clamp}); 1758 1759 SDNode *Addc = CurDAG->getMachineNode( 1760 AMDGPU::V_ADDC_U32_e64, DL, VTs, 1761 {AddOffsetHi, SDValue(N0Hi, 0), SDValue(Add, 1), Clamp}); 1762 1763 SDValue RegSequenceArgs[] = { 1764 CurDAG->getTargetConstant(AMDGPU::VReg_64RegClassID, DL, MVT::i32), 1765 SDValue(Add, 0), Sub0, SDValue(Addc, 0), Sub1}; 1766 1767 Addr = SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL, 1768 MVT::i64, RegSequenceArgs), 1769 0); 1770 } 1771 } 1772 } 1773 } 1774 1775 VAddr = Addr; 1776 Offset = CurDAG->getTargetConstant(OffsetVal, SDLoc(), MVT::i16); 1777 return true; 1778 } 1779 1780 // If this matches zero_extend i32:x, return x 1781 static SDValue matchZExtFromI32(SDValue Op) { 1782 if (Op.getOpcode() != ISD::ZERO_EXTEND) 1783 return SDValue(); 1784 1785 SDValue ExtSrc = Op.getOperand(0); 1786 return (ExtSrc.getValueType() == MVT::i32) ? ExtSrc : SDValue(); 1787 } 1788 1789 // Match (64-bit SGPR base) + (zext vgpr offset) + sext(imm offset) 1790 bool AMDGPUDAGToDAGISel::SelectGlobalSAddr(SDNode *N, 1791 SDValue Addr, 1792 SDValue &SAddr, 1793 SDValue &VOffset, 1794 SDValue &Offset) const { 1795 int64_t ImmOffset = 0; 1796 1797 // Match the immediate offset first, which canonically is moved as low as 1798 // possible. 1799 if (CurDAG->isBaseWithConstantOffset(Addr)) { 1800 SDValue LHS = Addr.getOperand(0); 1801 SDValue RHS = Addr.getOperand(1); 1802 1803 int64_t COffsetVal = cast<ConstantSDNode>(RHS)->getSExtValue(); 1804 const SIInstrInfo *TII = Subtarget->getInstrInfo(); 1805 1806 // TODO: Could split larger constant into VGPR offset. 1807 if (TII->isLegalFLATOffset(COffsetVal, AMDGPUAS::GLOBAL_ADDRESS, true)) { 1808 Addr = LHS; 1809 ImmOffset = COffsetVal; 1810 } 1811 } 1812 1813 // Match the variable offset. 1814 if (Addr.getOpcode() != ISD::ADD) 1815 return false; 1816 1817 SDValue LHS = Addr.getOperand(0); 1818 SDValue RHS = Addr.getOperand(1); 1819 1820 if (!LHS->isDivergent()) { 1821 // add (i64 sgpr), (zero_extend (i32 vgpr)) 1822 if (SDValue ZextRHS = matchZExtFromI32(RHS)) { 1823 SAddr = LHS; 1824 VOffset = ZextRHS; 1825 } 1826 } 1827 1828 if (!SAddr && !RHS->isDivergent()) { 1829 // add (zero_extend (i32 vgpr)), (i64 sgpr) 1830 if (SDValue ZextLHS = matchZExtFromI32(LHS)) { 1831 SAddr = RHS; 1832 VOffset = ZextLHS; 1833 } 1834 } 1835 1836 if (!SAddr) 1837 return false; 1838 1839 Offset = CurDAG->getTargetConstant(ImmOffset, SDLoc(), MVT::i16); 1840 return true; 1841 } 1842 1843 // Match (32-bit SGPR base) + sext(imm offset) 1844 bool AMDGPUDAGToDAGISel::SelectScratchSAddr(SDNode *N, 1845 SDValue Addr, 1846 SDValue &SAddr, 1847 SDValue &Offset) const { 1848 if (Addr->isDivergent()) 1849 return false; 1850 1851 SAddr = Addr; 1852 int64_t COffsetVal = 0; 1853 1854 if (CurDAG->isBaseWithConstantOffset(Addr)) { 1855 COffsetVal = cast<ConstantSDNode>(Addr.getOperand(1))->getSExtValue(); 1856 SAddr = Addr.getOperand(0); 1857 } 1858 1859 if (auto FI = dyn_cast<FrameIndexSDNode>(SAddr)) { 1860 SAddr = CurDAG->getTargetFrameIndex(FI->getIndex(), FI->getValueType(0)); 1861 } else if (SAddr.getOpcode() == ISD::ADD && 1862 isa<FrameIndexSDNode>(SAddr.getOperand(0))) { 1863 // Materialize this into a scalar move for scalar address to avoid 1864 // readfirstlane. 1865 auto FI = cast<FrameIndexSDNode>(SAddr.getOperand(0)); 1866 SDValue TFI = CurDAG->getTargetFrameIndex(FI->getIndex(), 1867 FI->getValueType(0)); 1868 SAddr = SDValue(CurDAG->getMachineNode(AMDGPU::S_ADD_U32, SDLoc(SAddr), 1869 MVT::i32, TFI, SAddr.getOperand(1)), 1870 0); 1871 } 1872 1873 const SIInstrInfo *TII = Subtarget->getInstrInfo(); 1874 1875 if (!TII->isLegalFLATOffset(COffsetVal, AMDGPUAS::PRIVATE_ADDRESS, true)) { 1876 int64_t RemainderOffset = COffsetVal; 1877 int64_t ImmField = 0; 1878 const unsigned NumBits = TII->getNumFlatOffsetBits(true); 1879 // Use signed division by a power of two to truncate towards 0. 1880 int64_t D = 1LL << (NumBits - 1); 1881 RemainderOffset = (COffsetVal / D) * D; 1882 ImmField = COffsetVal - RemainderOffset; 1883 1884 assert(TII->isLegalFLATOffset(ImmField, AMDGPUAS::PRIVATE_ADDRESS, true)); 1885 assert(RemainderOffset + ImmField == COffsetVal); 1886 1887 COffsetVal = ImmField; 1888 1889 SDLoc DL(N); 1890 SDValue AddOffset = 1891 getMaterializedScalarImm32(Lo_32(RemainderOffset), DL); 1892 SAddr = SDValue(CurDAG->getMachineNode(AMDGPU::S_ADD_U32, DL, MVT::i32, 1893 SAddr, AddOffset), 0); 1894 } 1895 1896 Offset = CurDAG->getTargetConstant(COffsetVal, SDLoc(), MVT::i16); 1897 1898 return true; 1899 } 1900 1901 bool AMDGPUDAGToDAGISel::SelectSMRDOffset(SDValue ByteOffsetNode, 1902 SDValue &Offset, bool &Imm) const { 1903 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ByteOffsetNode); 1904 if (!C) { 1905 if (ByteOffsetNode.getValueType().isScalarInteger() && 1906 ByteOffsetNode.getValueType().getSizeInBits() == 32) { 1907 Offset = ByteOffsetNode; 1908 Imm = false; 1909 return true; 1910 } 1911 if (ByteOffsetNode.getOpcode() == ISD::ZERO_EXTEND) { 1912 if (ByteOffsetNode.getOperand(0).getValueType().getSizeInBits() == 32) { 1913 Offset = ByteOffsetNode.getOperand(0); 1914 Imm = false; 1915 return true; 1916 } 1917 } 1918 return false; 1919 } 1920 1921 SDLoc SL(ByteOffsetNode); 1922 // GFX9 and GFX10 have signed byte immediate offsets. 1923 int64_t ByteOffset = C->getSExtValue(); 1924 Optional<int64_t> EncodedOffset = 1925 AMDGPU::getSMRDEncodedOffset(*Subtarget, ByteOffset, false); 1926 if (EncodedOffset) { 1927 Offset = CurDAG->getTargetConstant(*EncodedOffset, SL, MVT::i32); 1928 Imm = true; 1929 return true; 1930 } 1931 1932 // SGPR and literal offsets are unsigned. 1933 if (ByteOffset < 0) 1934 return false; 1935 1936 EncodedOffset = AMDGPU::getSMRDEncodedLiteralOffset32(*Subtarget, ByteOffset); 1937 if (EncodedOffset) { 1938 Offset = CurDAG->getTargetConstant(*EncodedOffset, SL, MVT::i32); 1939 return true; 1940 } 1941 1942 if (!isUInt<32>(ByteOffset) && !isInt<32>(ByteOffset)) 1943 return false; 1944 1945 SDValue C32Bit = CurDAG->getTargetConstant(ByteOffset, SL, MVT::i32); 1946 Offset = SDValue( 1947 CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32, C32Bit), 0); 1948 1949 return true; 1950 } 1951 1952 SDValue AMDGPUDAGToDAGISel::Expand32BitAddress(SDValue Addr) const { 1953 if (Addr.getValueType() != MVT::i32) 1954 return Addr; 1955 1956 // Zero-extend a 32-bit address. 1957 SDLoc SL(Addr); 1958 1959 const MachineFunction &MF = CurDAG->getMachineFunction(); 1960 const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>(); 1961 unsigned AddrHiVal = Info->get32BitAddressHighBits(); 1962 SDValue AddrHi = CurDAG->getTargetConstant(AddrHiVal, SL, MVT::i32); 1963 1964 const SDValue Ops[] = { 1965 CurDAG->getTargetConstant(AMDGPU::SReg_64_XEXECRegClassID, SL, MVT::i32), 1966 Addr, 1967 CurDAG->getTargetConstant(AMDGPU::sub0, SL, MVT::i32), 1968 SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32, AddrHi), 1969 0), 1970 CurDAG->getTargetConstant(AMDGPU::sub1, SL, MVT::i32), 1971 }; 1972 1973 return SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, SL, MVT::i64, 1974 Ops), 0); 1975 } 1976 1977 bool AMDGPUDAGToDAGISel::SelectSMRD(SDValue Addr, SDValue &SBase, 1978 SDValue &Offset, bool &Imm) const { 1979 SDLoc SL(Addr); 1980 1981 // A 32-bit (address + offset) should not cause unsigned 32-bit integer 1982 // wraparound, because s_load instructions perform the addition in 64 bits. 1983 if ((Addr.getValueType() != MVT::i32 || 1984 Addr->getFlags().hasNoUnsignedWrap())) { 1985 SDValue N0, N1; 1986 // Extract the base and offset if possible. 1987 if (CurDAG->isBaseWithConstantOffset(Addr) || 1988 Addr.getOpcode() == ISD::ADD) { 1989 N0 = Addr.getOperand(0); 1990 N1 = Addr.getOperand(1); 1991 } else if (getBaseWithOffsetUsingSplitOR(*CurDAG, Addr, N0, N1)) { 1992 assert(N0 && N1 && isa<ConstantSDNode>(N1)); 1993 } 1994 if (N0 && N1) { 1995 if (SelectSMRDOffset(N1, Offset, Imm)) { 1996 SBase = Expand32BitAddress(N0); 1997 return true; 1998 } 1999 } 2000 } 2001 SBase = Expand32BitAddress(Addr); 2002 Offset = CurDAG->getTargetConstant(0, SL, MVT::i32); 2003 Imm = true; 2004 return true; 2005 } 2006 2007 bool AMDGPUDAGToDAGISel::SelectSMRDImm(SDValue Addr, SDValue &SBase, 2008 SDValue &Offset) const { 2009 bool Imm = false; 2010 return SelectSMRD(Addr, SBase, Offset, Imm) && Imm; 2011 } 2012 2013 bool AMDGPUDAGToDAGISel::SelectSMRDImm32(SDValue Addr, SDValue &SBase, 2014 SDValue &Offset) const { 2015 2016 assert(Subtarget->getGeneration() == AMDGPUSubtarget::SEA_ISLANDS); 2017 2018 bool Imm = false; 2019 if (!SelectSMRD(Addr, SBase, Offset, Imm)) 2020 return false; 2021 2022 return !Imm && isa<ConstantSDNode>(Offset); 2023 } 2024 2025 bool AMDGPUDAGToDAGISel::SelectSMRDSgpr(SDValue Addr, SDValue &SBase, 2026 SDValue &Offset) const { 2027 bool Imm = false; 2028 return SelectSMRD(Addr, SBase, Offset, Imm) && !Imm && 2029 !isa<ConstantSDNode>(Offset); 2030 } 2031 2032 bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm(SDValue Addr, 2033 SDValue &Offset) const { 2034 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr)) { 2035 // The immediate offset for S_BUFFER instructions is unsigned. 2036 if (auto Imm = 2037 AMDGPU::getSMRDEncodedOffset(*Subtarget, C->getZExtValue(), true)) { 2038 Offset = CurDAG->getTargetConstant(*Imm, SDLoc(Addr), MVT::i32); 2039 return true; 2040 } 2041 } 2042 2043 return false; 2044 } 2045 2046 bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm32(SDValue Addr, 2047 SDValue &Offset) const { 2048 assert(Subtarget->getGeneration() == AMDGPUSubtarget::SEA_ISLANDS); 2049 2050 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr)) { 2051 if (auto Imm = AMDGPU::getSMRDEncodedLiteralOffset32(*Subtarget, 2052 C->getZExtValue())) { 2053 Offset = CurDAG->getTargetConstant(*Imm, SDLoc(Addr), MVT::i32); 2054 return true; 2055 } 2056 } 2057 2058 return false; 2059 } 2060 2061 bool AMDGPUDAGToDAGISel::SelectMOVRELOffset(SDValue Index, 2062 SDValue &Base, 2063 SDValue &Offset) const { 2064 SDLoc DL(Index); 2065 2066 if (CurDAG->isBaseWithConstantOffset(Index)) { 2067 SDValue N0 = Index.getOperand(0); 2068 SDValue N1 = Index.getOperand(1); 2069 ConstantSDNode *C1 = cast<ConstantSDNode>(N1); 2070 2071 // (add n0, c0) 2072 // Don't peel off the offset (c0) if doing so could possibly lead 2073 // the base (n0) to be negative. 2074 // (or n0, |c0|) can never change a sign given isBaseWithConstantOffset. 2075 if (C1->getSExtValue() <= 0 || CurDAG->SignBitIsZero(N0) || 2076 (Index->getOpcode() == ISD::OR && C1->getSExtValue() >= 0)) { 2077 Base = N0; 2078 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32); 2079 return true; 2080 } 2081 } 2082 2083 if (isa<ConstantSDNode>(Index)) 2084 return false; 2085 2086 Base = Index; 2087 Offset = CurDAG->getTargetConstant(0, DL, MVT::i32); 2088 return true; 2089 } 2090 2091 SDNode *AMDGPUDAGToDAGISel::getS_BFE(unsigned Opcode, const SDLoc &DL, 2092 SDValue Val, uint32_t Offset, 2093 uint32_t Width) { 2094 // Transformation function, pack the offset and width of a BFE into 2095 // the format expected by the S_BFE_I32 / S_BFE_U32. In the second 2096 // source, bits [5:0] contain the offset and bits [22:16] the width. 2097 uint32_t PackedVal = Offset | (Width << 16); 2098 SDValue PackedConst = CurDAG->getTargetConstant(PackedVal, DL, MVT::i32); 2099 2100 return CurDAG->getMachineNode(Opcode, DL, MVT::i32, Val, PackedConst); 2101 } 2102 2103 void AMDGPUDAGToDAGISel::SelectS_BFEFromShifts(SDNode *N) { 2104 // "(a << b) srl c)" ---> "BFE_U32 a, (c-b), (32-c) 2105 // "(a << b) sra c)" ---> "BFE_I32 a, (c-b), (32-c) 2106 // Predicate: 0 < b <= c < 32 2107 2108 const SDValue &Shl = N->getOperand(0); 2109 ConstantSDNode *B = dyn_cast<ConstantSDNode>(Shl->getOperand(1)); 2110 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); 2111 2112 if (B && C) { 2113 uint32_t BVal = B->getZExtValue(); 2114 uint32_t CVal = C->getZExtValue(); 2115 2116 if (0 < BVal && BVal <= CVal && CVal < 32) { 2117 bool Signed = N->getOpcode() == ISD::SRA; 2118 unsigned Opcode = Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32; 2119 2120 ReplaceNode(N, getS_BFE(Opcode, SDLoc(N), Shl.getOperand(0), CVal - BVal, 2121 32 - CVal)); 2122 return; 2123 } 2124 } 2125 SelectCode(N); 2126 } 2127 2128 void AMDGPUDAGToDAGISel::SelectS_BFE(SDNode *N) { 2129 switch (N->getOpcode()) { 2130 case ISD::AND: 2131 if (N->getOperand(0).getOpcode() == ISD::SRL) { 2132 // "(a srl b) & mask" ---> "BFE_U32 a, b, popcount(mask)" 2133 // Predicate: isMask(mask) 2134 const SDValue &Srl = N->getOperand(0); 2135 ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(Srl.getOperand(1)); 2136 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(N->getOperand(1)); 2137 2138 if (Shift && Mask) { 2139 uint32_t ShiftVal = Shift->getZExtValue(); 2140 uint32_t MaskVal = Mask->getZExtValue(); 2141 2142 if (isMask_32(MaskVal)) { 2143 uint32_t WidthVal = countPopulation(MaskVal); 2144 2145 ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N), 2146 Srl.getOperand(0), ShiftVal, WidthVal)); 2147 return; 2148 } 2149 } 2150 } 2151 break; 2152 case ISD::SRL: 2153 if (N->getOperand(0).getOpcode() == ISD::AND) { 2154 // "(a & mask) srl b)" ---> "BFE_U32 a, b, popcount(mask >> b)" 2155 // Predicate: isMask(mask >> b) 2156 const SDValue &And = N->getOperand(0); 2157 ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(N->getOperand(1)); 2158 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(And->getOperand(1)); 2159 2160 if (Shift && Mask) { 2161 uint32_t ShiftVal = Shift->getZExtValue(); 2162 uint32_t MaskVal = Mask->getZExtValue() >> ShiftVal; 2163 2164 if (isMask_32(MaskVal)) { 2165 uint32_t WidthVal = countPopulation(MaskVal); 2166 2167 ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N), 2168 And.getOperand(0), ShiftVal, WidthVal)); 2169 return; 2170 } 2171 } 2172 } else if (N->getOperand(0).getOpcode() == ISD::SHL) { 2173 SelectS_BFEFromShifts(N); 2174 return; 2175 } 2176 break; 2177 case ISD::SRA: 2178 if (N->getOperand(0).getOpcode() == ISD::SHL) { 2179 SelectS_BFEFromShifts(N); 2180 return; 2181 } 2182 break; 2183 2184 case ISD::SIGN_EXTEND_INREG: { 2185 // sext_inreg (srl x, 16), i8 -> bfe_i32 x, 16, 8 2186 SDValue Src = N->getOperand(0); 2187 if (Src.getOpcode() != ISD::SRL) 2188 break; 2189 2190 const ConstantSDNode *Amt = dyn_cast<ConstantSDNode>(Src.getOperand(1)); 2191 if (!Amt) 2192 break; 2193 2194 unsigned Width = cast<VTSDNode>(N->getOperand(1))->getVT().getSizeInBits(); 2195 ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_I32, SDLoc(N), Src.getOperand(0), 2196 Amt->getZExtValue(), Width)); 2197 return; 2198 } 2199 } 2200 2201 SelectCode(N); 2202 } 2203 2204 bool AMDGPUDAGToDAGISel::isCBranchSCC(const SDNode *N) const { 2205 assert(N->getOpcode() == ISD::BRCOND); 2206 if (!N->hasOneUse()) 2207 return false; 2208 2209 SDValue Cond = N->getOperand(1); 2210 if (Cond.getOpcode() == ISD::CopyToReg) 2211 Cond = Cond.getOperand(2); 2212 2213 if (Cond.getOpcode() != ISD::SETCC || !Cond.hasOneUse()) 2214 return false; 2215 2216 MVT VT = Cond.getOperand(0).getSimpleValueType(); 2217 if (VT == MVT::i32) 2218 return true; 2219 2220 if (VT == MVT::i64) { 2221 auto ST = static_cast<const GCNSubtarget *>(Subtarget); 2222 2223 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); 2224 return (CC == ISD::SETEQ || CC == ISD::SETNE) && ST->hasScalarCompareEq64(); 2225 } 2226 2227 return false; 2228 } 2229 2230 void AMDGPUDAGToDAGISel::SelectBRCOND(SDNode *N) { 2231 SDValue Cond = N->getOperand(1); 2232 2233 if (Cond.isUndef()) { 2234 CurDAG->SelectNodeTo(N, AMDGPU::SI_BR_UNDEF, MVT::Other, 2235 N->getOperand(2), N->getOperand(0)); 2236 return; 2237 } 2238 2239 const GCNSubtarget *ST = static_cast<const GCNSubtarget *>(Subtarget); 2240 const SIRegisterInfo *TRI = ST->getRegisterInfo(); 2241 2242 bool UseSCCBr = isCBranchSCC(N) && isUniformBr(N); 2243 unsigned BrOp = UseSCCBr ? AMDGPU::S_CBRANCH_SCC1 : AMDGPU::S_CBRANCH_VCCNZ; 2244 Register CondReg = UseSCCBr ? AMDGPU::SCC : TRI->getVCC(); 2245 SDLoc SL(N); 2246 2247 if (!UseSCCBr) { 2248 // This is the case that we are selecting to S_CBRANCH_VCCNZ. We have not 2249 // analyzed what generates the vcc value, so we do not know whether vcc 2250 // bits for disabled lanes are 0. Thus we need to mask out bits for 2251 // disabled lanes. 2252 // 2253 // For the case that we select S_CBRANCH_SCC1 and it gets 2254 // changed to S_CBRANCH_VCCNZ in SIFixSGPRCopies, SIFixSGPRCopies calls 2255 // SIInstrInfo::moveToVALU which inserts the S_AND). 2256 // 2257 // We could add an analysis of what generates the vcc value here and omit 2258 // the S_AND when is unnecessary. But it would be better to add a separate 2259 // pass after SIFixSGPRCopies to do the unnecessary S_AND removal, so it 2260 // catches both cases. 2261 Cond = SDValue(CurDAG->getMachineNode(ST->isWave32() ? AMDGPU::S_AND_B32 2262 : AMDGPU::S_AND_B64, 2263 SL, MVT::i1, 2264 CurDAG->getRegister(ST->isWave32() ? AMDGPU::EXEC_LO 2265 : AMDGPU::EXEC, 2266 MVT::i1), 2267 Cond), 2268 0); 2269 } 2270 2271 SDValue VCC = CurDAG->getCopyToReg(N->getOperand(0), SL, CondReg, Cond); 2272 CurDAG->SelectNodeTo(N, BrOp, MVT::Other, 2273 N->getOperand(2), // Basic Block 2274 VCC.getValue(0)); 2275 } 2276 2277 void AMDGPUDAGToDAGISel::SelectFMAD_FMA(SDNode *N) { 2278 MVT VT = N->getSimpleValueType(0); 2279 bool IsFMA = N->getOpcode() == ISD::FMA; 2280 if (VT != MVT::f32 || (!Subtarget->hasMadMixInsts() && 2281 !Subtarget->hasFmaMixInsts()) || 2282 ((IsFMA && Subtarget->hasMadMixInsts()) || 2283 (!IsFMA && Subtarget->hasFmaMixInsts()))) { 2284 SelectCode(N); 2285 return; 2286 } 2287 2288 SDValue Src0 = N->getOperand(0); 2289 SDValue Src1 = N->getOperand(1); 2290 SDValue Src2 = N->getOperand(2); 2291 unsigned Src0Mods, Src1Mods, Src2Mods; 2292 2293 // Avoid using v_mad_mix_f32/v_fma_mix_f32 unless there is actually an operand 2294 // using the conversion from f16. 2295 bool Sel0 = SelectVOP3PMadMixModsImpl(Src0, Src0, Src0Mods); 2296 bool Sel1 = SelectVOP3PMadMixModsImpl(Src1, Src1, Src1Mods); 2297 bool Sel2 = SelectVOP3PMadMixModsImpl(Src2, Src2, Src2Mods); 2298 2299 assert((IsFMA || !Mode.allFP32Denormals()) && 2300 "fmad selected with denormals enabled"); 2301 // TODO: We can select this with f32 denormals enabled if all the sources are 2302 // converted from f16 (in which case fmad isn't legal). 2303 2304 if (Sel0 || Sel1 || Sel2) { 2305 // For dummy operands. 2306 SDValue Zero = CurDAG->getTargetConstant(0, SDLoc(), MVT::i32); 2307 SDValue Ops[] = { 2308 CurDAG->getTargetConstant(Src0Mods, SDLoc(), MVT::i32), Src0, 2309 CurDAG->getTargetConstant(Src1Mods, SDLoc(), MVT::i32), Src1, 2310 CurDAG->getTargetConstant(Src2Mods, SDLoc(), MVT::i32), Src2, 2311 CurDAG->getTargetConstant(0, SDLoc(), MVT::i1), 2312 Zero, Zero 2313 }; 2314 2315 CurDAG->SelectNodeTo(N, 2316 IsFMA ? AMDGPU::V_FMA_MIX_F32 : AMDGPU::V_MAD_MIX_F32, 2317 MVT::f32, Ops); 2318 } else { 2319 SelectCode(N); 2320 } 2321 } 2322 2323 // This is here because there isn't a way to use the generated sub0_sub1 as the 2324 // subreg index to EXTRACT_SUBREG in tablegen. 2325 void AMDGPUDAGToDAGISel::SelectATOMIC_CMP_SWAP(SDNode *N) { 2326 MemSDNode *Mem = cast<MemSDNode>(N); 2327 unsigned AS = Mem->getAddressSpace(); 2328 if (AS == AMDGPUAS::FLAT_ADDRESS) { 2329 SelectCode(N); 2330 return; 2331 } 2332 2333 MVT VT = N->getSimpleValueType(0); 2334 bool Is32 = (VT == MVT::i32); 2335 SDLoc SL(N); 2336 2337 MachineSDNode *CmpSwap = nullptr; 2338 if (Subtarget->hasAddr64()) { 2339 SDValue SRsrc, VAddr, SOffset, Offset, SLC; 2340 2341 if (SelectMUBUFAddr64(Mem->getBasePtr(), SRsrc, VAddr, SOffset, Offset, SLC)) { 2342 unsigned Opcode = Is32 ? AMDGPU::BUFFER_ATOMIC_CMPSWAP_ADDR64_RTN : 2343 AMDGPU::BUFFER_ATOMIC_CMPSWAP_X2_ADDR64_RTN; 2344 SDValue CmpVal = Mem->getOperand(2); 2345 2346 // XXX - Do we care about glue operands? 2347 2348 SDValue Ops[] = { 2349 CmpVal, VAddr, SRsrc, SOffset, Offset, SLC, Mem->getChain() 2350 }; 2351 2352 CmpSwap = CurDAG->getMachineNode(Opcode, SL, Mem->getVTList(), Ops); 2353 } 2354 } 2355 2356 if (!CmpSwap) { 2357 SDValue SRsrc, SOffset, Offset, SLC; 2358 if (SelectMUBUFOffset(Mem->getBasePtr(), SRsrc, SOffset, Offset, SLC)) { 2359 unsigned Opcode = Is32 ? AMDGPU::BUFFER_ATOMIC_CMPSWAP_OFFSET_RTN : 2360 AMDGPU::BUFFER_ATOMIC_CMPSWAP_X2_OFFSET_RTN; 2361 2362 SDValue CmpVal = Mem->getOperand(2); 2363 SDValue Ops[] = { 2364 CmpVal, SRsrc, SOffset, Offset, SLC, Mem->getChain() 2365 }; 2366 2367 CmpSwap = CurDAG->getMachineNode(Opcode, SL, Mem->getVTList(), Ops); 2368 } 2369 } 2370 2371 if (!CmpSwap) { 2372 SelectCode(N); 2373 return; 2374 } 2375 2376 MachineMemOperand *MMO = Mem->getMemOperand(); 2377 CurDAG->setNodeMemRefs(CmpSwap, {MMO}); 2378 2379 unsigned SubReg = Is32 ? AMDGPU::sub0 : AMDGPU::sub0_sub1; 2380 SDValue Extract 2381 = CurDAG->getTargetExtractSubreg(SubReg, SL, VT, SDValue(CmpSwap, 0)); 2382 2383 ReplaceUses(SDValue(N, 0), Extract); 2384 ReplaceUses(SDValue(N, 1), SDValue(CmpSwap, 1)); 2385 CurDAG->RemoveDeadNode(N); 2386 } 2387 2388 void AMDGPUDAGToDAGISel::SelectDSAppendConsume(SDNode *N, unsigned IntrID) { 2389 // The address is assumed to be uniform, so if it ends up in a VGPR, it will 2390 // be copied to an SGPR with readfirstlane. 2391 unsigned Opc = IntrID == Intrinsic::amdgcn_ds_append ? 2392 AMDGPU::DS_APPEND : AMDGPU::DS_CONSUME; 2393 2394 SDValue Chain = N->getOperand(0); 2395 SDValue Ptr = N->getOperand(2); 2396 MemIntrinsicSDNode *M = cast<MemIntrinsicSDNode>(N); 2397 MachineMemOperand *MMO = M->getMemOperand(); 2398 bool IsGDS = M->getAddressSpace() == AMDGPUAS::REGION_ADDRESS; 2399 2400 SDValue Offset; 2401 if (CurDAG->isBaseWithConstantOffset(Ptr)) { 2402 SDValue PtrBase = Ptr.getOperand(0); 2403 SDValue PtrOffset = Ptr.getOperand(1); 2404 2405 const APInt &OffsetVal = cast<ConstantSDNode>(PtrOffset)->getAPIntValue(); 2406 if (isDSOffsetLegal(PtrBase, OffsetVal.getZExtValue(), 16)) { 2407 N = glueCopyToM0(N, PtrBase); 2408 Offset = CurDAG->getTargetConstant(OffsetVal, SDLoc(), MVT::i32); 2409 } 2410 } 2411 2412 if (!Offset) { 2413 N = glueCopyToM0(N, Ptr); 2414 Offset = CurDAG->getTargetConstant(0, SDLoc(), MVT::i32); 2415 } 2416 2417 SDValue Ops[] = { 2418 Offset, 2419 CurDAG->getTargetConstant(IsGDS, SDLoc(), MVT::i32), 2420 Chain, 2421 N->getOperand(N->getNumOperands() - 1) // New glue 2422 }; 2423 2424 SDNode *Selected = CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops); 2425 CurDAG->setNodeMemRefs(cast<MachineSDNode>(Selected), {MMO}); 2426 } 2427 2428 static unsigned gwsIntrinToOpcode(unsigned IntrID) { 2429 switch (IntrID) { 2430 case Intrinsic::amdgcn_ds_gws_init: 2431 return AMDGPU::DS_GWS_INIT; 2432 case Intrinsic::amdgcn_ds_gws_barrier: 2433 return AMDGPU::DS_GWS_BARRIER; 2434 case Intrinsic::amdgcn_ds_gws_sema_v: 2435 return AMDGPU::DS_GWS_SEMA_V; 2436 case Intrinsic::amdgcn_ds_gws_sema_br: 2437 return AMDGPU::DS_GWS_SEMA_BR; 2438 case Intrinsic::amdgcn_ds_gws_sema_p: 2439 return AMDGPU::DS_GWS_SEMA_P; 2440 case Intrinsic::amdgcn_ds_gws_sema_release_all: 2441 return AMDGPU::DS_GWS_SEMA_RELEASE_ALL; 2442 default: 2443 llvm_unreachable("not a gws intrinsic"); 2444 } 2445 } 2446 2447 void AMDGPUDAGToDAGISel::SelectDS_GWS(SDNode *N, unsigned IntrID) { 2448 if (IntrID == Intrinsic::amdgcn_ds_gws_sema_release_all && 2449 !Subtarget->hasGWSSemaReleaseAll()) { 2450 // Let this error. 2451 SelectCode(N); 2452 return; 2453 } 2454 2455 // Chain, intrinsic ID, vsrc, offset 2456 const bool HasVSrc = N->getNumOperands() == 4; 2457 assert(HasVSrc || N->getNumOperands() == 3); 2458 2459 SDLoc SL(N); 2460 SDValue BaseOffset = N->getOperand(HasVSrc ? 3 : 2); 2461 int ImmOffset = 0; 2462 MemIntrinsicSDNode *M = cast<MemIntrinsicSDNode>(N); 2463 MachineMemOperand *MMO = M->getMemOperand(); 2464 2465 // Don't worry if the offset ends up in a VGPR. Only one lane will have 2466 // effect, so SIFixSGPRCopies will validly insert readfirstlane. 2467 2468 // The resource id offset is computed as (<isa opaque base> + M0[21:16] + 2469 // offset field) % 64. Some versions of the programming guide omit the m0 2470 // part, or claim it's from offset 0. 2471 if (ConstantSDNode *ConstOffset = dyn_cast<ConstantSDNode>(BaseOffset)) { 2472 // If we have a constant offset, try to use the 0 in m0 as the base. 2473 // TODO: Look into changing the default m0 initialization value. If the 2474 // default -1 only set the low 16-bits, we could leave it as-is and add 1 to 2475 // the immediate offset. 2476 glueCopyToM0(N, CurDAG->getTargetConstant(0, SL, MVT::i32)); 2477 ImmOffset = ConstOffset->getZExtValue(); 2478 } else { 2479 if (CurDAG->isBaseWithConstantOffset(BaseOffset)) { 2480 ImmOffset = BaseOffset.getConstantOperandVal(1); 2481 BaseOffset = BaseOffset.getOperand(0); 2482 } 2483 2484 // Prefer to do the shift in an SGPR since it should be possible to use m0 2485 // as the result directly. If it's already an SGPR, it will be eliminated 2486 // later. 2487 SDNode *SGPROffset 2488 = CurDAG->getMachineNode(AMDGPU::V_READFIRSTLANE_B32, SL, MVT::i32, 2489 BaseOffset); 2490 // Shift to offset in m0 2491 SDNode *M0Base 2492 = CurDAG->getMachineNode(AMDGPU::S_LSHL_B32, SL, MVT::i32, 2493 SDValue(SGPROffset, 0), 2494 CurDAG->getTargetConstant(16, SL, MVT::i32)); 2495 glueCopyToM0(N, SDValue(M0Base, 0)); 2496 } 2497 2498 SDValue Chain = N->getOperand(0); 2499 SDValue OffsetField = CurDAG->getTargetConstant(ImmOffset, SL, MVT::i32); 2500 2501 // TODO: Can this just be removed from the instruction? 2502 SDValue GDS = CurDAG->getTargetConstant(1, SL, MVT::i1); 2503 2504 const unsigned Opc = gwsIntrinToOpcode(IntrID); 2505 SmallVector<SDValue, 5> Ops; 2506 if (HasVSrc) 2507 Ops.push_back(N->getOperand(2)); 2508 Ops.push_back(OffsetField); 2509 Ops.push_back(GDS); 2510 Ops.push_back(Chain); 2511 2512 SDNode *Selected = CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops); 2513 CurDAG->setNodeMemRefs(cast<MachineSDNode>(Selected), {MMO}); 2514 } 2515 2516 void AMDGPUDAGToDAGISel::SelectInterpP1F16(SDNode *N) { 2517 if (Subtarget->getLDSBankCount() != 16) { 2518 // This is a single instruction with a pattern. 2519 SelectCode(N); 2520 return; 2521 } 2522 2523 SDLoc DL(N); 2524 2525 // This requires 2 instructions. It is possible to write a pattern to support 2526 // this, but the generated isel emitter doesn't correctly deal with multiple 2527 // output instructions using the same physical register input. The copy to m0 2528 // is incorrectly placed before the second instruction. 2529 // 2530 // TODO: Match source modifiers. 2531 // 2532 // def : Pat < 2533 // (int_amdgcn_interp_p1_f16 2534 // (VOP3Mods f32:$src0, i32:$src0_modifiers), 2535 // (i32 timm:$attrchan), (i32 timm:$attr), 2536 // (i1 timm:$high), M0), 2537 // (V_INTERP_P1LV_F16 $src0_modifiers, VGPR_32:$src0, timm:$attr, 2538 // timm:$attrchan, 0, 2539 // (V_INTERP_MOV_F32 2, timm:$attr, timm:$attrchan), timm:$high)> { 2540 // let Predicates = [has16BankLDS]; 2541 // } 2542 2543 // 16 bank LDS 2544 SDValue ToM0 = CurDAG->getCopyToReg(CurDAG->getEntryNode(), DL, AMDGPU::M0, 2545 N->getOperand(5), SDValue()); 2546 2547 SDVTList VTs = CurDAG->getVTList(MVT::f32, MVT::Other); 2548 2549 SDNode *InterpMov = 2550 CurDAG->getMachineNode(AMDGPU::V_INTERP_MOV_F32, DL, VTs, { 2551 CurDAG->getTargetConstant(2, DL, MVT::i32), // P0 2552 N->getOperand(3), // Attr 2553 N->getOperand(2), // Attrchan 2554 ToM0.getValue(1) // In glue 2555 }); 2556 2557 SDNode *InterpP1LV = 2558 CurDAG->getMachineNode(AMDGPU::V_INTERP_P1LV_F16, DL, MVT::f32, { 2559 CurDAG->getTargetConstant(0, DL, MVT::i32), // $src0_modifiers 2560 N->getOperand(1), // Src0 2561 N->getOperand(3), // Attr 2562 N->getOperand(2), // Attrchan 2563 CurDAG->getTargetConstant(0, DL, MVT::i32), // $src2_modifiers 2564 SDValue(InterpMov, 0), // Src2 - holds two f16 values selected by high 2565 N->getOperand(4), // high 2566 CurDAG->getTargetConstant(0, DL, MVT::i1), // $clamp 2567 CurDAG->getTargetConstant(0, DL, MVT::i32), // $omod 2568 SDValue(InterpMov, 1) 2569 }); 2570 2571 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), SDValue(InterpP1LV, 0)); 2572 } 2573 2574 void AMDGPUDAGToDAGISel::SelectINTRINSIC_W_CHAIN(SDNode *N) { 2575 unsigned IntrID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 2576 switch (IntrID) { 2577 case Intrinsic::amdgcn_ds_append: 2578 case Intrinsic::amdgcn_ds_consume: { 2579 if (N->getValueType(0) != MVT::i32) 2580 break; 2581 SelectDSAppendConsume(N, IntrID); 2582 return; 2583 } 2584 } 2585 2586 SelectCode(N); 2587 } 2588 2589 void AMDGPUDAGToDAGISel::SelectINTRINSIC_WO_CHAIN(SDNode *N) { 2590 unsigned IntrID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 2591 unsigned Opcode; 2592 switch (IntrID) { 2593 case Intrinsic::amdgcn_wqm: 2594 Opcode = AMDGPU::WQM; 2595 break; 2596 case Intrinsic::amdgcn_softwqm: 2597 Opcode = AMDGPU::SOFT_WQM; 2598 break; 2599 case Intrinsic::amdgcn_wwm: 2600 Opcode = AMDGPU::WWM; 2601 break; 2602 case Intrinsic::amdgcn_interp_p1_f16: 2603 SelectInterpP1F16(N); 2604 return; 2605 default: 2606 SelectCode(N); 2607 return; 2608 } 2609 2610 SDValue Src = N->getOperand(1); 2611 CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), {Src}); 2612 } 2613 2614 void AMDGPUDAGToDAGISel::SelectINTRINSIC_VOID(SDNode *N) { 2615 unsigned IntrID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 2616 switch (IntrID) { 2617 case Intrinsic::amdgcn_ds_gws_init: 2618 case Intrinsic::amdgcn_ds_gws_barrier: 2619 case Intrinsic::amdgcn_ds_gws_sema_v: 2620 case Intrinsic::amdgcn_ds_gws_sema_br: 2621 case Intrinsic::amdgcn_ds_gws_sema_p: 2622 case Intrinsic::amdgcn_ds_gws_sema_release_all: 2623 SelectDS_GWS(N, IntrID); 2624 return; 2625 default: 2626 break; 2627 } 2628 2629 SelectCode(N); 2630 } 2631 2632 bool AMDGPUDAGToDAGISel::SelectVOP3ModsImpl(SDValue In, SDValue &Src, 2633 unsigned &Mods) const { 2634 Mods = 0; 2635 Src = In; 2636 2637 if (Src.getOpcode() == ISD::FNEG) { 2638 Mods |= SISrcMods::NEG; 2639 Src = Src.getOperand(0); 2640 } 2641 2642 if (Src.getOpcode() == ISD::FABS) { 2643 Mods |= SISrcMods::ABS; 2644 Src = Src.getOperand(0); 2645 } 2646 2647 return true; 2648 } 2649 2650 bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src, 2651 SDValue &SrcMods) const { 2652 unsigned Mods; 2653 if (SelectVOP3ModsImpl(In, Src, Mods)) { 2654 SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32); 2655 return true; 2656 } 2657 2658 return false; 2659 } 2660 2661 bool AMDGPUDAGToDAGISel::SelectVOP3Mods_NNaN(SDValue In, SDValue &Src, 2662 SDValue &SrcMods) const { 2663 SelectVOP3Mods(In, Src, SrcMods); 2664 return isNoNanSrc(Src); 2665 } 2666 2667 bool AMDGPUDAGToDAGISel::SelectVOP3NoMods(SDValue In, SDValue &Src) const { 2668 if (In.getOpcode() == ISD::FABS || In.getOpcode() == ISD::FNEG) 2669 return false; 2670 2671 Src = In; 2672 return true; 2673 } 2674 2675 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0(SDValue In, SDValue &Src, 2676 SDValue &SrcMods, SDValue &Clamp, 2677 SDValue &Omod) const { 2678 SDLoc DL(In); 2679 Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1); 2680 Omod = CurDAG->getTargetConstant(0, DL, MVT::i1); 2681 2682 return SelectVOP3Mods(In, Src, SrcMods); 2683 } 2684 2685 bool AMDGPUDAGToDAGISel::SelectVOP3OMods(SDValue In, SDValue &Src, 2686 SDValue &Clamp, SDValue &Omod) const { 2687 Src = In; 2688 2689 SDLoc DL(In); 2690 Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1); 2691 Omod = CurDAG->getTargetConstant(0, DL, MVT::i1); 2692 2693 return true; 2694 } 2695 2696 bool AMDGPUDAGToDAGISel::SelectVOP3PMods(SDValue In, SDValue &Src, 2697 SDValue &SrcMods) const { 2698 unsigned Mods = 0; 2699 Src = In; 2700 2701 if (Src.getOpcode() == ISD::FNEG) { 2702 Mods ^= (SISrcMods::NEG | SISrcMods::NEG_HI); 2703 Src = Src.getOperand(0); 2704 } 2705 2706 if (Src.getOpcode() == ISD::BUILD_VECTOR) { 2707 unsigned VecMods = Mods; 2708 2709 SDValue Lo = stripBitcast(Src.getOperand(0)); 2710 SDValue Hi = stripBitcast(Src.getOperand(1)); 2711 2712 if (Lo.getOpcode() == ISD::FNEG) { 2713 Lo = stripBitcast(Lo.getOperand(0)); 2714 Mods ^= SISrcMods::NEG; 2715 } 2716 2717 if (Hi.getOpcode() == ISD::FNEG) { 2718 Hi = stripBitcast(Hi.getOperand(0)); 2719 Mods ^= SISrcMods::NEG_HI; 2720 } 2721 2722 if (isExtractHiElt(Lo, Lo)) 2723 Mods |= SISrcMods::OP_SEL_0; 2724 2725 if (isExtractHiElt(Hi, Hi)) 2726 Mods |= SISrcMods::OP_SEL_1; 2727 2728 Lo = stripExtractLoElt(Lo); 2729 Hi = stripExtractLoElt(Hi); 2730 2731 if (Lo == Hi && !isInlineImmediate(Lo.getNode())) { 2732 // Really a scalar input. Just select from the low half of the register to 2733 // avoid packing. 2734 2735 Src = Lo; 2736 SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32); 2737 return true; 2738 } 2739 2740 Mods = VecMods; 2741 } 2742 2743 // Packed instructions do not have abs modifiers. 2744 Mods |= SISrcMods::OP_SEL_1; 2745 2746 SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32); 2747 return true; 2748 } 2749 2750 bool AMDGPUDAGToDAGISel::SelectVOP3OpSel(SDValue In, SDValue &Src, 2751 SDValue &SrcMods) const { 2752 Src = In; 2753 // FIXME: Handle op_sel 2754 SrcMods = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32); 2755 return true; 2756 } 2757 2758 bool AMDGPUDAGToDAGISel::SelectVOP3OpSelMods(SDValue In, SDValue &Src, 2759 SDValue &SrcMods) const { 2760 // FIXME: Handle op_sel 2761 return SelectVOP3Mods(In, Src, SrcMods); 2762 } 2763 2764 // The return value is not whether the match is possible (which it always is), 2765 // but whether or not it a conversion is really used. 2766 bool AMDGPUDAGToDAGISel::SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src, 2767 unsigned &Mods) const { 2768 Mods = 0; 2769 SelectVOP3ModsImpl(In, Src, Mods); 2770 2771 if (Src.getOpcode() == ISD::FP_EXTEND) { 2772 Src = Src.getOperand(0); 2773 assert(Src.getValueType() == MVT::f16); 2774 Src = stripBitcast(Src); 2775 2776 // Be careful about folding modifiers if we already have an abs. fneg is 2777 // applied last, so we don't want to apply an earlier fneg. 2778 if ((Mods & SISrcMods::ABS) == 0) { 2779 unsigned ModsTmp; 2780 SelectVOP3ModsImpl(Src, Src, ModsTmp); 2781 2782 if ((ModsTmp & SISrcMods::NEG) != 0) 2783 Mods ^= SISrcMods::NEG; 2784 2785 if ((ModsTmp & SISrcMods::ABS) != 0) 2786 Mods |= SISrcMods::ABS; 2787 } 2788 2789 // op_sel/op_sel_hi decide the source type and source. 2790 // If the source's op_sel_hi is set, it indicates to do a conversion from fp16. 2791 // If the sources's op_sel is set, it picks the high half of the source 2792 // register. 2793 2794 Mods |= SISrcMods::OP_SEL_1; 2795 if (isExtractHiElt(Src, Src)) { 2796 Mods |= SISrcMods::OP_SEL_0; 2797 2798 // TODO: Should we try to look for neg/abs here? 2799 } 2800 2801 return true; 2802 } 2803 2804 return false; 2805 } 2806 2807 bool AMDGPUDAGToDAGISel::SelectVOP3PMadMixMods(SDValue In, SDValue &Src, 2808 SDValue &SrcMods) const { 2809 unsigned Mods = 0; 2810 SelectVOP3PMadMixModsImpl(In, Src, Mods); 2811 SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32); 2812 return true; 2813 } 2814 2815 SDValue AMDGPUDAGToDAGISel::getHi16Elt(SDValue In) const { 2816 if (In.isUndef()) 2817 return CurDAG->getUNDEF(MVT::i32); 2818 2819 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(In)) { 2820 SDLoc SL(In); 2821 return CurDAG->getConstant(C->getZExtValue() << 16, SL, MVT::i32); 2822 } 2823 2824 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(In)) { 2825 SDLoc SL(In); 2826 return CurDAG->getConstant( 2827 C->getValueAPF().bitcastToAPInt().getZExtValue() << 16, SL, MVT::i32); 2828 } 2829 2830 SDValue Src; 2831 if (isExtractHiElt(In, Src)) 2832 return Src; 2833 2834 return SDValue(); 2835 } 2836 2837 bool AMDGPUDAGToDAGISel::isVGPRImm(const SDNode * N) const { 2838 assert(CurDAG->getTarget().getTargetTriple().getArch() == Triple::amdgcn); 2839 2840 const SIRegisterInfo *SIRI = 2841 static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo()); 2842 const SIInstrInfo * SII = 2843 static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo()); 2844 2845 unsigned Limit = 0; 2846 bool AllUsesAcceptSReg = true; 2847 for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end(); 2848 Limit < 10 && U != E; ++U, ++Limit) { 2849 const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo()); 2850 2851 // If the register class is unknown, it could be an unknown 2852 // register class that needs to be an SGPR, e.g. an inline asm 2853 // constraint 2854 if (!RC || SIRI->isSGPRClass(RC)) 2855 return false; 2856 2857 if (RC != &AMDGPU::VS_32RegClass) { 2858 AllUsesAcceptSReg = false; 2859 SDNode * User = *U; 2860 if (User->isMachineOpcode()) { 2861 unsigned Opc = User->getMachineOpcode(); 2862 MCInstrDesc Desc = SII->get(Opc); 2863 if (Desc.isCommutable()) { 2864 unsigned OpIdx = Desc.getNumDefs() + U.getOperandNo(); 2865 unsigned CommuteIdx1 = TargetInstrInfo::CommuteAnyOperandIndex; 2866 if (SII->findCommutedOpIndices(Desc, OpIdx, CommuteIdx1)) { 2867 unsigned CommutedOpNo = CommuteIdx1 - Desc.getNumDefs(); 2868 const TargetRegisterClass *CommutedRC = getOperandRegClass(*U, CommutedOpNo); 2869 if (CommutedRC == &AMDGPU::VS_32RegClass) 2870 AllUsesAcceptSReg = true; 2871 } 2872 } 2873 } 2874 // If "AllUsesAcceptSReg == false" so far we haven't suceeded 2875 // commuting current user. This means have at least one use 2876 // that strictly require VGPR. Thus, we will not attempt to commute 2877 // other user instructions. 2878 if (!AllUsesAcceptSReg) 2879 break; 2880 } 2881 } 2882 return !AllUsesAcceptSReg && (Limit < 10); 2883 } 2884 2885 bool AMDGPUDAGToDAGISel::isUniformLoad(const SDNode * N) const { 2886 auto Ld = cast<LoadSDNode>(N); 2887 2888 return Ld->getAlignment() >= 4 && 2889 ( 2890 ( 2891 ( 2892 Ld->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS || 2893 Ld->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT 2894 ) 2895 && 2896 !N->isDivergent() 2897 ) 2898 || 2899 ( 2900 Subtarget->getScalarizeGlobalBehavior() && 2901 Ld->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS && 2902 Ld->isSimple() && 2903 !N->isDivergent() && 2904 static_cast<const SITargetLowering *>( 2905 getTargetLowering())->isMemOpHasNoClobberedMemOperand(N) 2906 ) 2907 ); 2908 } 2909 2910 void AMDGPUDAGToDAGISel::PostprocessISelDAG() { 2911 const AMDGPUTargetLowering& Lowering = 2912 *static_cast<const AMDGPUTargetLowering*>(getTargetLowering()); 2913 bool IsModified = false; 2914 do { 2915 IsModified = false; 2916 2917 // Go over all selected nodes and try to fold them a bit more 2918 SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_begin(); 2919 while (Position != CurDAG->allnodes_end()) { 2920 SDNode *Node = &*Position++; 2921 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(Node); 2922 if (!MachineNode) 2923 continue; 2924 2925 SDNode *ResNode = Lowering.PostISelFolding(MachineNode, *CurDAG); 2926 if (ResNode != Node) { 2927 if (ResNode) 2928 ReplaceUses(Node, ResNode); 2929 IsModified = true; 2930 } 2931 } 2932 CurDAG->RemoveDeadNodes(); 2933 } while (IsModified); 2934 } 2935 2936 bool R600DAGToDAGISel::runOnMachineFunction(MachineFunction &MF) { 2937 Subtarget = &MF.getSubtarget<R600Subtarget>(); 2938 return SelectionDAGISel::runOnMachineFunction(MF); 2939 } 2940 2941 bool R600DAGToDAGISel::isConstantLoad(const MemSDNode *N, int CbId) const { 2942 if (!N->readMem()) 2943 return false; 2944 if (CbId == -1) 2945 return N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS || 2946 N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT; 2947 2948 return N->getAddressSpace() == AMDGPUAS::CONSTANT_BUFFER_0 + CbId; 2949 } 2950 2951 bool R600DAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr, 2952 SDValue& IntPtr) { 2953 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) { 2954 IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, SDLoc(Addr), 2955 true); 2956 return true; 2957 } 2958 return false; 2959 } 2960 2961 bool R600DAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr, 2962 SDValue& BaseReg, SDValue &Offset) { 2963 if (!isa<ConstantSDNode>(Addr)) { 2964 BaseReg = Addr; 2965 Offset = CurDAG->getIntPtrConstant(0, SDLoc(Addr), true); 2966 return true; 2967 } 2968 return false; 2969 } 2970 2971 void R600DAGToDAGISel::Select(SDNode *N) { 2972 unsigned int Opc = N->getOpcode(); 2973 if (N->isMachineOpcode()) { 2974 N->setNodeId(-1); 2975 return; // Already selected. 2976 } 2977 2978 switch (Opc) { 2979 default: break; 2980 case AMDGPUISD::BUILD_VERTICAL_VECTOR: 2981 case ISD::SCALAR_TO_VECTOR: 2982 case ISD::BUILD_VECTOR: { 2983 EVT VT = N->getValueType(0); 2984 unsigned NumVectorElts = VT.getVectorNumElements(); 2985 unsigned RegClassID; 2986 // BUILD_VECTOR was lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG 2987 // that adds a 128 bits reg copy when going through TwoAddressInstructions 2988 // pass. We want to avoid 128 bits copies as much as possible because they 2989 // can't be bundled by our scheduler. 2990 switch(NumVectorElts) { 2991 case 2: RegClassID = R600::R600_Reg64RegClassID; break; 2992 case 4: 2993 if (Opc == AMDGPUISD::BUILD_VERTICAL_VECTOR) 2994 RegClassID = R600::R600_Reg128VerticalRegClassID; 2995 else 2996 RegClassID = R600::R600_Reg128RegClassID; 2997 break; 2998 default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR"); 2999 } 3000 SelectBuildVector(N, RegClassID); 3001 return; 3002 } 3003 } 3004 3005 SelectCode(N); 3006 } 3007 3008 bool R600DAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base, 3009 SDValue &Offset) { 3010 ConstantSDNode *C; 3011 SDLoc DL(Addr); 3012 3013 if ((C = dyn_cast<ConstantSDNode>(Addr))) { 3014 Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32); 3015 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32); 3016 } else if ((Addr.getOpcode() == AMDGPUISD::DWORDADDR) && 3017 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(0)))) { 3018 Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32); 3019 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32); 3020 } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) && 3021 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) { 3022 Base = Addr.getOperand(0); 3023 Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32); 3024 } else { 3025 Base = Addr; 3026 Offset = CurDAG->getTargetConstant(0, DL, MVT::i32); 3027 } 3028 3029 return true; 3030 } 3031 3032 bool R600DAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base, 3033 SDValue &Offset) { 3034 ConstantSDNode *IMMOffset; 3035 3036 if (Addr.getOpcode() == ISD::ADD 3037 && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1))) 3038 && isInt<16>(IMMOffset->getZExtValue())) { 3039 3040 Base = Addr.getOperand(0); 3041 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr), 3042 MVT::i32); 3043 return true; 3044 // If the pointer address is constant, we can move it to the offset field. 3045 } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr)) 3046 && isInt<16>(IMMOffset->getZExtValue())) { 3047 Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), 3048 SDLoc(CurDAG->getEntryNode()), 3049 R600::ZERO, MVT::i32); 3050 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr), 3051 MVT::i32); 3052 return true; 3053 } 3054 3055 // Default case, no offset 3056 Base = Addr; 3057 Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i32); 3058 return true; 3059 } 3060