1 //===-- PPCISelDAGToDAG.cpp - PPC --pattern matching inst selector --------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines a pattern matching instruction selector for PowerPC, 11 // converting from a legalized dag to a PPC dag. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "PPC.h" 16 #include "MCTargetDesc/PPCPredicates.h" 17 #include "PPCMachineFunctionInfo.h" 18 #include "PPCTargetMachine.h" 19 #include "llvm/Analysis/BranchProbabilityInfo.h" 20 #include "llvm/CodeGen/FunctionLoweringInfo.h" 21 #include "llvm/CodeGen/MachineFunction.h" 22 #include "llvm/CodeGen/MachineInstrBuilder.h" 23 #include "llvm/CodeGen/MachineRegisterInfo.h" 24 #include "llvm/CodeGen/SelectionDAG.h" 25 #include "llvm/CodeGen/SelectionDAGISel.h" 26 #include "llvm/IR/Constants.h" 27 #include "llvm/IR/Function.h" 28 #include "llvm/IR/GlobalAlias.h" 29 #include "llvm/IR/GlobalValue.h" 30 #include "llvm/IR/GlobalVariable.h" 31 #include "llvm/IR/Intrinsics.h" 32 #include "llvm/IR/Module.h" 33 #include "llvm/Support/CommandLine.h" 34 #include "llvm/Support/Debug.h" 35 #include "llvm/Support/ErrorHandling.h" 36 #include "llvm/Support/MathExtras.h" 37 #include "llvm/Support/raw_ostream.h" 38 #include "llvm/Target/TargetOptions.h" 39 using namespace llvm; 40 41 #define DEBUG_TYPE "ppc-codegen" 42 43 // FIXME: Remove this once the bug has been fixed! 44 cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug", 45 cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden); 46 47 static cl::opt<bool> 48 UseBitPermRewriter("ppc-use-bit-perm-rewriter", cl::init(true), 49 cl::desc("use aggressive ppc isel for bit permutations"), 50 cl::Hidden); 51 static cl::opt<bool> BPermRewriterNoMasking( 52 "ppc-bit-perm-rewriter-stress-rotates", 53 cl::desc("stress rotate selection in aggressive ppc isel for " 54 "bit permutations"), 55 cl::Hidden); 56 57 static cl::opt<bool> EnableBranchHint( 58 "ppc-use-branch-hint", cl::init(true), 59 cl::desc("Enable static hinting of branches on ppc"), 60 cl::Hidden); 61 62 namespace llvm { 63 void initializePPCDAGToDAGISelPass(PassRegistry&); 64 } 65 66 namespace { 67 //===--------------------------------------------------------------------===// 68 /// PPCDAGToDAGISel - PPC specific code to select PPC machine 69 /// instructions for SelectionDAG operations. 70 /// 71 class PPCDAGToDAGISel : public SelectionDAGISel { 72 const PPCTargetMachine &TM; 73 const PPCSubtarget *PPCSubTarget; 74 const PPCTargetLowering *PPCLowering; 75 unsigned GlobalBaseReg; 76 public: 77 explicit PPCDAGToDAGISel(PPCTargetMachine &tm) 78 : SelectionDAGISel(tm), TM(tm) { 79 initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry()); 80 } 81 82 bool runOnMachineFunction(MachineFunction &MF) override { 83 // Make sure we re-emit a set of the global base reg if necessary 84 GlobalBaseReg = 0; 85 PPCSubTarget = &MF.getSubtarget<PPCSubtarget>(); 86 PPCLowering = PPCSubTarget->getTargetLowering(); 87 SelectionDAGISel::runOnMachineFunction(MF); 88 89 if (!PPCSubTarget->isSVR4ABI()) 90 InsertVRSaveCode(MF); 91 92 return true; 93 } 94 95 void PreprocessISelDAG() override; 96 void PostprocessISelDAG() override; 97 98 /// getI32Imm - Return a target constant with the specified value, of type 99 /// i32. 100 inline SDValue getI32Imm(unsigned Imm, SDLoc dl) { 101 return CurDAG->getTargetConstant(Imm, dl, MVT::i32); 102 } 103 104 /// getI64Imm - Return a target constant with the specified value, of type 105 /// i64. 106 inline SDValue getI64Imm(uint64_t Imm, SDLoc dl) { 107 return CurDAG->getTargetConstant(Imm, dl, MVT::i64); 108 } 109 110 /// getSmallIPtrImm - Return a target constant of pointer type. 111 inline SDValue getSmallIPtrImm(unsigned Imm, SDLoc dl) { 112 return CurDAG->getTargetConstant( 113 Imm, dl, PPCLowering->getPointerTy(CurDAG->getDataLayout())); 114 } 115 116 /// isRotateAndMask - Returns true if Mask and Shift can be folded into a 117 /// rotate and mask opcode and mask operation. 118 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask, 119 unsigned &SH, unsigned &MB, unsigned &ME); 120 121 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC 122 /// base register. Return the virtual register that holds this value. 123 SDNode *getGlobalBaseReg(); 124 125 SDNode *getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset = 0); 126 127 // Select - Convert the specified operand from a target-independent to a 128 // target-specific node if it hasn't already been changed. 129 SDNode *Select(SDNode *N) override; 130 131 SDNode *SelectBitfieldInsert(SDNode *N); 132 SDNode *SelectBitPermutation(SDNode *N); 133 134 /// SelectCC - Select a comparison of the specified values with the 135 /// specified condition code, returning the CR# of the expression. 136 SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, SDLoc dl); 137 138 /// SelectAddrImm - Returns true if the address N can be represented by 139 /// a base register plus a signed 16-bit displacement [r+imm]. 140 bool SelectAddrImm(SDValue N, SDValue &Disp, 141 SDValue &Base) { 142 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false); 143 } 144 145 /// SelectAddrImmOffs - Return true if the operand is valid for a preinc 146 /// immediate field. Note that the operand at this point is already the 147 /// result of a prior SelectAddressRegImm call. 148 bool SelectAddrImmOffs(SDValue N, SDValue &Out) const { 149 if (N.getOpcode() == ISD::TargetConstant || 150 N.getOpcode() == ISD::TargetGlobalAddress) { 151 Out = N; 152 return true; 153 } 154 155 return false; 156 } 157 158 /// SelectAddrIdx - Given the specified addressed, check to see if it can be 159 /// represented as an indexed [r+r] operation. Returns false if it can 160 /// be represented by [r+imm], which are preferred. 161 bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) { 162 return PPCLowering->SelectAddressRegReg(N, Base, Index, *CurDAG); 163 } 164 165 /// SelectAddrIdxOnly - Given the specified addressed, force it to be 166 /// represented as an indexed [r+r] operation. 167 bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) { 168 return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, *CurDAG); 169 } 170 171 /// SelectAddrImmX4 - Returns true if the address N can be represented by 172 /// a base register plus a signed 16-bit displacement that is a multiple of 4. 173 /// Suitable for use by STD and friends. 174 bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) { 175 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true); 176 } 177 178 // Select an address into a single register. 179 bool SelectAddr(SDValue N, SDValue &Base) { 180 Base = N; 181 return true; 182 } 183 184 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for 185 /// inline asm expressions. It is always correct to compute the value into 186 /// a register. The case of adding a (possibly relocatable) constant to a 187 /// register can be improved, but it is wrong to substitute Reg+Reg for 188 /// Reg in an asm, because the load or store opcode would have to change. 189 bool SelectInlineAsmMemoryOperand(const SDValue &Op, 190 unsigned ConstraintID, 191 std::vector<SDValue> &OutOps) override { 192 193 switch(ConstraintID) { 194 default: 195 errs() << "ConstraintID: " << ConstraintID << "\n"; 196 llvm_unreachable("Unexpected asm memory constraint"); 197 case InlineAsm::Constraint_es: 198 case InlineAsm::Constraint_i: 199 case InlineAsm::Constraint_m: 200 case InlineAsm::Constraint_o: 201 case InlineAsm::Constraint_Q: 202 case InlineAsm::Constraint_Z: 203 case InlineAsm::Constraint_Zy: 204 // We need to make sure that this one operand does not end up in r0 205 // (because we might end up lowering this as 0(%op)). 206 const TargetRegisterInfo *TRI = PPCSubTarget->getRegisterInfo(); 207 const TargetRegisterClass *TRC = TRI->getPointerRegClass(*MF, /*Kind=*/1); 208 SDLoc dl(Op); 209 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32); 210 SDValue NewOp = 211 SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, 212 dl, Op.getValueType(), 213 Op, RC), 0); 214 215 OutOps.push_back(NewOp); 216 return false; 217 } 218 return true; 219 } 220 221 void InsertVRSaveCode(MachineFunction &MF); 222 223 const char *getPassName() const override { 224 return "PowerPC DAG->DAG Pattern Instruction Selection"; 225 } 226 227 // Include the pieces autogenerated from the target description. 228 #include "PPCGenDAGISel.inc" 229 230 private: 231 SDNode *SelectSETCC(SDNode *N); 232 233 void PeepholePPC64(); 234 void PeepholePPC64ZExt(); 235 void PeepholeCROps(); 236 237 SDValue combineToCMPB(SDNode *N); 238 void foldBoolExts(SDValue &Res, SDNode *&N); 239 240 bool AllUsersSelectZero(SDNode *N); 241 void SwapAllSelectUsers(SDNode *N); 242 243 SDNode *transferMemOperands(SDNode *N, SDNode *Result); 244 }; 245 } 246 247 /// InsertVRSaveCode - Once the entire function has been instruction selected, 248 /// all virtual registers are created and all machine instructions are built, 249 /// check to see if we need to save/restore VRSAVE. If so, do it. 250 void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) { 251 // Check to see if this function uses vector registers, which means we have to 252 // save and restore the VRSAVE register and update it with the regs we use. 253 // 254 // In this case, there will be virtual registers of vector type created 255 // by the scheduler. Detect them now. 256 bool HasVectorVReg = false; 257 for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) { 258 unsigned Reg = TargetRegisterInfo::index2VirtReg(i); 259 if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) { 260 HasVectorVReg = true; 261 break; 262 } 263 } 264 if (!HasVectorVReg) return; // nothing to do. 265 266 // If we have a vector register, we want to emit code into the entry and exit 267 // blocks to save and restore the VRSAVE register. We do this here (instead 268 // of marking all vector instructions as clobbering VRSAVE) for two reasons: 269 // 270 // 1. This (trivially) reduces the load on the register allocator, by not 271 // having to represent the live range of the VRSAVE register. 272 // 2. This (more significantly) allows us to create a temporary virtual 273 // register to hold the saved VRSAVE value, allowing this temporary to be 274 // register allocated, instead of forcing it to be spilled to the stack. 275 276 // Create two vregs - one to hold the VRSAVE register that is live-in to the 277 // function and one for the value after having bits or'd into it. 278 unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 279 unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 280 281 const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo(); 282 MachineBasicBlock &EntryBB = *Fn.begin(); 283 DebugLoc dl; 284 // Emit the following code into the entry block: 285 // InVRSAVE = MFVRSAVE 286 // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE 287 // MTVRSAVE UpdatedVRSAVE 288 MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point 289 BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE); 290 BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE), 291 UpdatedVRSAVE).addReg(InVRSAVE); 292 BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE); 293 294 // Find all return blocks, outputting a restore in each epilog. 295 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { 296 if (BB->isReturnBlock()) { 297 IP = BB->end(); --IP; 298 299 // Skip over all terminator instructions, which are part of the return 300 // sequence. 301 MachineBasicBlock::iterator I2 = IP; 302 while (I2 != BB->begin() && (--I2)->isTerminator()) 303 IP = I2; 304 305 // Emit: MTVRSAVE InVRSave 306 BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE); 307 } 308 } 309 } 310 311 312 /// getGlobalBaseReg - Output the instructions required to put the 313 /// base address to use for accessing globals into a register. 314 /// 315 SDNode *PPCDAGToDAGISel::getGlobalBaseReg() { 316 if (!GlobalBaseReg) { 317 const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo(); 318 // Insert the set of GlobalBaseReg into the first MBB of the function 319 MachineBasicBlock &FirstMBB = MF->front(); 320 MachineBasicBlock::iterator MBBI = FirstMBB.begin(); 321 const Module *M = MF->getFunction()->getParent(); 322 DebugLoc dl; 323 324 if (PPCLowering->getPointerTy(CurDAG->getDataLayout()) == MVT::i32) { 325 if (PPCSubTarget->isTargetELF()) { 326 GlobalBaseReg = PPC::R30; 327 if (M->getPICLevel() == PICLevel::Small) { 328 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MoveGOTtoLR)); 329 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); 330 MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true); 331 } else { 332 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR)); 333 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); 334 unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 335 BuildMI(FirstMBB, MBBI, dl, 336 TII.get(PPC::UpdateGBR), GlobalBaseReg) 337 .addReg(TempReg, RegState::Define).addReg(GlobalBaseReg); 338 MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true); 339 } 340 } else { 341 GlobalBaseReg = 342 RegInfo->createVirtualRegister(&PPC::GPRC_NOR0RegClass); 343 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR)); 344 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); 345 } 346 } else { 347 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_NOX0RegClass); 348 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8)); 349 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg); 350 } 351 } 352 return CurDAG->getRegister(GlobalBaseReg, 353 PPCLowering->getPointerTy(CurDAG->getDataLayout())) 354 .getNode(); 355 } 356 357 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit 358 /// or 64-bit immediate, and if the value can be accurately represented as a 359 /// sign extension from a 16-bit value. If so, this returns true and the 360 /// immediate. 361 static bool isIntS16Immediate(SDNode *N, short &Imm) { 362 if (N->getOpcode() != ISD::Constant) 363 return false; 364 365 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue(); 366 if (N->getValueType(0) == MVT::i32) 367 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue(); 368 else 369 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue(); 370 } 371 372 static bool isIntS16Immediate(SDValue Op, short &Imm) { 373 return isIntS16Immediate(Op.getNode(), Imm); 374 } 375 376 377 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant 378 /// operand. If so Imm will receive the 32-bit value. 379 static bool isInt32Immediate(SDNode *N, unsigned &Imm) { 380 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) { 381 Imm = cast<ConstantSDNode>(N)->getZExtValue(); 382 return true; 383 } 384 return false; 385 } 386 387 /// isInt64Immediate - This method tests to see if the node is a 64-bit constant 388 /// operand. If so Imm will receive the 64-bit value. 389 static bool isInt64Immediate(SDNode *N, uint64_t &Imm) { 390 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) { 391 Imm = cast<ConstantSDNode>(N)->getZExtValue(); 392 return true; 393 } 394 return false; 395 } 396 397 // isInt32Immediate - This method tests to see if a constant operand. 398 // If so Imm will receive the 32 bit value. 399 static bool isInt32Immediate(SDValue N, unsigned &Imm) { 400 return isInt32Immediate(N.getNode(), Imm); 401 } 402 403 static unsigned getBranchHint(unsigned PCC, FunctionLoweringInfo *FuncInfo, 404 const SDValue &DestMBB) { 405 assert(isa<BasicBlockSDNode>(DestMBB)); 406 407 if (!FuncInfo->BPI) return PPC::BR_NO_HINT; 408 409 const BasicBlock *BB = FuncInfo->MBB->getBasicBlock(); 410 const TerminatorInst *BBTerm = BB->getTerminator(); 411 412 if (BBTerm->getNumSuccessors() != 2) return PPC::BR_NO_HINT; 413 414 const BasicBlock *TBB = BBTerm->getSuccessor(0); 415 const BasicBlock *FBB = BBTerm->getSuccessor(1); 416 417 uint32_t TWeight = FuncInfo->BPI->getEdgeWeight(BB, TBB); 418 uint32_t FWeight = FuncInfo->BPI->getEdgeWeight(BB, FBB); 419 420 // We only want to handle cases which are easy to predict at static time, e.g. 421 // C++ throw statement, that is very likely not taken, or calling never 422 // returned function, e.g. stdlib exit(). So we set Threshold to filter 423 // unwanted cases. 424 // 425 // Below is LLVM branch weight table, we only want to handle case 1, 2 426 // 427 // Case Taken:Nontaken Example 428 // 1. Unreachable 1048575:1 C++ throw, stdlib exit(), 429 // 2. Invoke-terminating 1:1048575 430 // 3. Coldblock 4:64 __builtin_expect 431 // 4. Loop Branch 124:4 For loop 432 // 5. PH/ZH/FPH 20:12 433 const uint32_t Threshold = 10000; 434 435 // Minimal weight should be at least 1 436 if (std::max(TWeight, FWeight) / 437 std::max(1u, std::min(TWeight, FWeight)) < Threshold) 438 return PPC::BR_NO_HINT; 439 440 DEBUG(dbgs() << "Use branch hint for '" << FuncInfo->Fn->getName() << "::" 441 << BB->getName() << "'\n" 442 << " -> " << TBB->getName() << ": " << TWeight << "\n" 443 << " -> " << FBB->getName() << ": " << FWeight << "\n"); 444 445 const BasicBlockSDNode *BBDN = cast<BasicBlockSDNode>(DestMBB); 446 447 // If Dest BasicBlock is False-BasicBlock (FBB), swap branch weight, 448 // because we want 'TWeight' stands for 'branch weight' to Dest BasicBlock 449 if (BBDN->getBasicBlock()->getBasicBlock() != TBB) 450 std::swap(TWeight, FWeight); 451 452 return (TWeight > FWeight) ? PPC::BR_TAKEN_HINT : PPC::BR_NONTAKEN_HINT; 453 } 454 455 // isOpcWithIntImmediate - This method tests to see if the node is a specific 456 // opcode and that it has a immediate integer right operand. 457 // If so Imm will receive the 32 bit value. 458 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) { 459 return N->getOpcode() == Opc 460 && isInt32Immediate(N->getOperand(1).getNode(), Imm); 461 } 462 463 SDNode *PPCDAGToDAGISel::getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset) { 464 SDLoc dl(SN); 465 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 466 SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0)); 467 unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8; 468 if (SN->hasOneUse()) 469 return CurDAG->SelectNodeTo(SN, Opc, N->getValueType(0), TFI, 470 getSmallIPtrImm(Offset, dl)); 471 return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI, 472 getSmallIPtrImm(Offset, dl)); 473 } 474 475 bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask, 476 bool isShiftMask, unsigned &SH, 477 unsigned &MB, unsigned &ME) { 478 // Don't even go down this path for i64, since different logic will be 479 // necessary for rldicl/rldicr/rldimi. 480 if (N->getValueType(0) != MVT::i32) 481 return false; 482 483 unsigned Shift = 32; 484 unsigned Indeterminant = ~0; // bit mask marking indeterminant results 485 unsigned Opcode = N->getOpcode(); 486 if (N->getNumOperands() != 2 || 487 !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31)) 488 return false; 489 490 if (Opcode == ISD::SHL) { 491 // apply shift left to mask if it comes first 492 if (isShiftMask) Mask = Mask << Shift; 493 // determine which bits are made indeterminant by shift 494 Indeterminant = ~(0xFFFFFFFFu << Shift); 495 } else if (Opcode == ISD::SRL) { 496 // apply shift right to mask if it comes first 497 if (isShiftMask) Mask = Mask >> Shift; 498 // determine which bits are made indeterminant by shift 499 Indeterminant = ~(0xFFFFFFFFu >> Shift); 500 // adjust for the left rotate 501 Shift = 32 - Shift; 502 } else if (Opcode == ISD::ROTL) { 503 Indeterminant = 0; 504 } else { 505 return false; 506 } 507 508 // if the mask doesn't intersect any Indeterminant bits 509 if (Mask && !(Mask & Indeterminant)) { 510 SH = Shift & 31; 511 // make sure the mask is still a mask (wrap arounds may not be) 512 return isRunOfOnes(Mask, MB, ME); 513 } 514 return false; 515 } 516 517 /// SelectBitfieldInsert - turn an or of two masked values into 518 /// the rotate left word immediate then mask insert (rlwimi) instruction. 519 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) { 520 SDValue Op0 = N->getOperand(0); 521 SDValue Op1 = N->getOperand(1); 522 SDLoc dl(N); 523 524 APInt LKZ, LKO, RKZ, RKO; 525 CurDAG->computeKnownBits(Op0, LKZ, LKO); 526 CurDAG->computeKnownBits(Op1, RKZ, RKO); 527 528 unsigned TargetMask = LKZ.getZExtValue(); 529 unsigned InsertMask = RKZ.getZExtValue(); 530 531 if ((TargetMask | InsertMask) == 0xFFFFFFFF) { 532 unsigned Op0Opc = Op0.getOpcode(); 533 unsigned Op1Opc = Op1.getOpcode(); 534 unsigned Value, SH = 0; 535 TargetMask = ~TargetMask; 536 InsertMask = ~InsertMask; 537 538 // If the LHS has a foldable shift and the RHS does not, then swap it to the 539 // RHS so that we can fold the shift into the insert. 540 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) { 541 if (Op0.getOperand(0).getOpcode() == ISD::SHL || 542 Op0.getOperand(0).getOpcode() == ISD::SRL) { 543 if (Op1.getOperand(0).getOpcode() != ISD::SHL && 544 Op1.getOperand(0).getOpcode() != ISD::SRL) { 545 std::swap(Op0, Op1); 546 std::swap(Op0Opc, Op1Opc); 547 std::swap(TargetMask, InsertMask); 548 } 549 } 550 } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) { 551 if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL && 552 Op1.getOperand(0).getOpcode() != ISD::SRL) { 553 std::swap(Op0, Op1); 554 std::swap(Op0Opc, Op1Opc); 555 std::swap(TargetMask, InsertMask); 556 } 557 } 558 559 unsigned MB, ME; 560 if (isRunOfOnes(InsertMask, MB, ME)) { 561 SDValue Tmp1, Tmp2; 562 563 if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) && 564 isInt32Immediate(Op1.getOperand(1), Value)) { 565 Op1 = Op1.getOperand(0); 566 SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value; 567 } 568 if (Op1Opc == ISD::AND) { 569 // The AND mask might not be a constant, and we need to make sure that 570 // if we're going to fold the masking with the insert, all bits not 571 // know to be zero in the mask are known to be one. 572 APInt MKZ, MKO; 573 CurDAG->computeKnownBits(Op1.getOperand(1), MKZ, MKO); 574 bool CanFoldMask = InsertMask == MKO.getZExtValue(); 575 576 unsigned SHOpc = Op1.getOperand(0).getOpcode(); 577 if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask && 578 isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) { 579 // Note that Value must be in range here (less than 32) because 580 // otherwise there would not be any bits set in InsertMask. 581 Op1 = Op1.getOperand(0).getOperand(0); 582 SH = (SHOpc == ISD::SHL) ? Value : 32 - Value; 583 } 584 } 585 586 SH &= 31; 587 SDValue Ops[] = { Op0, Op1, getI32Imm(SH, dl), getI32Imm(MB, dl), 588 getI32Imm(ME, dl) }; 589 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); 590 } 591 } 592 return nullptr; 593 } 594 595 // Predict the number of instructions that would be generated by calling 596 // SelectInt64(N). 597 static unsigned SelectInt64CountDirect(int64_t Imm) { 598 // Assume no remaining bits. 599 unsigned Remainder = 0; 600 // Assume no shift required. 601 unsigned Shift = 0; 602 603 // If it can't be represented as a 32 bit value. 604 if (!isInt<32>(Imm)) { 605 Shift = countTrailingZeros<uint64_t>(Imm); 606 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift; 607 608 // If the shifted value fits 32 bits. 609 if (isInt<32>(ImmSh)) { 610 // Go with the shifted value. 611 Imm = ImmSh; 612 } else { 613 // Still stuck with a 64 bit value. 614 Remainder = Imm; 615 Shift = 32; 616 Imm >>= 32; 617 } 618 } 619 620 // Intermediate operand. 621 unsigned Result = 0; 622 623 // Handle first 32 bits. 624 unsigned Lo = Imm & 0xFFFF; 625 626 // Simple value. 627 if (isInt<16>(Imm)) { 628 // Just the Lo bits. 629 ++Result; 630 } else if (Lo) { 631 // Handle the Hi bits and Lo bits. 632 Result += 2; 633 } else { 634 // Just the Hi bits. 635 ++Result; 636 } 637 638 // If no shift, we're done. 639 if (!Shift) return Result; 640 641 // Shift for next step if the upper 32-bits were not zero. 642 if (Imm) 643 ++Result; 644 645 // Add in the last bits as required. 646 if ((Remainder >> 16) & 0xFFFF) 647 ++Result; 648 if (Remainder & 0xFFFF) 649 ++Result; 650 651 return Result; 652 } 653 654 static uint64_t Rot64(uint64_t Imm, unsigned R) { 655 return (Imm << R) | (Imm >> (64 - R)); 656 } 657 658 static unsigned SelectInt64Count(int64_t Imm) { 659 unsigned Count = SelectInt64CountDirect(Imm); 660 if (Count == 1) 661 return Count; 662 663 for (unsigned r = 1; r < 63; ++r) { 664 uint64_t RImm = Rot64(Imm, r); 665 unsigned RCount = SelectInt64CountDirect(RImm) + 1; 666 Count = std::min(Count, RCount); 667 668 // See comments in SelectInt64 for an explanation of the logic below. 669 unsigned LS = findLastSet(RImm); 670 if (LS != r-1) 671 continue; 672 673 uint64_t OnesMask = -(int64_t) (UINT64_C(1) << (LS+1)); 674 uint64_t RImmWithOnes = RImm | OnesMask; 675 676 RCount = SelectInt64CountDirect(RImmWithOnes) + 1; 677 Count = std::min(Count, RCount); 678 } 679 680 return Count; 681 } 682 683 // Select a 64-bit constant. For cost-modeling purposes, SelectInt64Count 684 // (above) needs to be kept in sync with this function. 685 static SDNode *SelectInt64Direct(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) { 686 // Assume no remaining bits. 687 unsigned Remainder = 0; 688 // Assume no shift required. 689 unsigned Shift = 0; 690 691 // If it can't be represented as a 32 bit value. 692 if (!isInt<32>(Imm)) { 693 Shift = countTrailingZeros<uint64_t>(Imm); 694 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift; 695 696 // If the shifted value fits 32 bits. 697 if (isInt<32>(ImmSh)) { 698 // Go with the shifted value. 699 Imm = ImmSh; 700 } else { 701 // Still stuck with a 64 bit value. 702 Remainder = Imm; 703 Shift = 32; 704 Imm >>= 32; 705 } 706 } 707 708 // Intermediate operand. 709 SDNode *Result; 710 711 // Handle first 32 bits. 712 unsigned Lo = Imm & 0xFFFF; 713 unsigned Hi = (Imm >> 16) & 0xFFFF; 714 715 auto getI32Imm = [CurDAG, dl](unsigned Imm) { 716 return CurDAG->getTargetConstant(Imm, dl, MVT::i32); 717 }; 718 719 // Simple value. 720 if (isInt<16>(Imm)) { 721 // Just the Lo bits. 722 Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo)); 723 } else if (Lo) { 724 // Handle the Hi bits. 725 unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8; 726 Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi)); 727 // And Lo bits. 728 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, 729 SDValue(Result, 0), getI32Imm(Lo)); 730 } else { 731 // Just the Hi bits. 732 Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi)); 733 } 734 735 // If no shift, we're done. 736 if (!Shift) return Result; 737 738 // Shift for next step if the upper 32-bits were not zero. 739 if (Imm) { 740 Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, 741 SDValue(Result, 0), 742 getI32Imm(Shift), 743 getI32Imm(63 - Shift)); 744 } 745 746 // Add in the last bits as required. 747 if ((Hi = (Remainder >> 16) & 0xFFFF)) { 748 Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64, 749 SDValue(Result, 0), getI32Imm(Hi)); 750 } 751 if ((Lo = Remainder & 0xFFFF)) { 752 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, 753 SDValue(Result, 0), getI32Imm(Lo)); 754 } 755 756 return Result; 757 } 758 759 static SDNode *SelectInt64(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) { 760 unsigned Count = SelectInt64CountDirect(Imm); 761 if (Count == 1) 762 return SelectInt64Direct(CurDAG, dl, Imm); 763 764 unsigned RMin = 0; 765 766 int64_t MatImm; 767 unsigned MaskEnd; 768 769 for (unsigned r = 1; r < 63; ++r) { 770 uint64_t RImm = Rot64(Imm, r); 771 unsigned RCount = SelectInt64CountDirect(RImm) + 1; 772 if (RCount < Count) { 773 Count = RCount; 774 RMin = r; 775 MatImm = RImm; 776 MaskEnd = 63; 777 } 778 779 // If the immediate to generate has many trailing zeros, it might be 780 // worthwhile to generate a rotated value with too many leading ones 781 // (because that's free with li/lis's sign-extension semantics), and then 782 // mask them off after rotation. 783 784 unsigned LS = findLastSet(RImm); 785 // We're adding (63-LS) higher-order ones, and we expect to mask them off 786 // after performing the inverse rotation by (64-r). So we need that: 787 // 63-LS == 64-r => LS == r-1 788 if (LS != r-1) 789 continue; 790 791 uint64_t OnesMask = -(int64_t) (UINT64_C(1) << (LS+1)); 792 uint64_t RImmWithOnes = RImm | OnesMask; 793 794 RCount = SelectInt64CountDirect(RImmWithOnes) + 1; 795 if (RCount < Count) { 796 Count = RCount; 797 RMin = r; 798 MatImm = RImmWithOnes; 799 MaskEnd = LS; 800 } 801 } 802 803 if (!RMin) 804 return SelectInt64Direct(CurDAG, dl, Imm); 805 806 auto getI32Imm = [CurDAG, dl](unsigned Imm) { 807 return CurDAG->getTargetConstant(Imm, dl, MVT::i32); 808 }; 809 810 SDValue Val = SDValue(SelectInt64Direct(CurDAG, dl, MatImm), 0); 811 return CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Val, 812 getI32Imm(64 - RMin), getI32Imm(MaskEnd)); 813 } 814 815 // Select a 64-bit constant. 816 static SDNode *SelectInt64(SelectionDAG *CurDAG, SDNode *N) { 817 SDLoc dl(N); 818 819 // Get 64 bit value. 820 int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue(); 821 return SelectInt64(CurDAG, dl, Imm); 822 } 823 824 namespace { 825 class BitPermutationSelector { 826 struct ValueBit { 827 SDValue V; 828 829 // The bit number in the value, using a convention where bit 0 is the 830 // lowest-order bit. 831 unsigned Idx; 832 833 enum Kind { 834 ConstZero, 835 Variable 836 } K; 837 838 ValueBit(SDValue V, unsigned I, Kind K = Variable) 839 : V(V), Idx(I), K(K) {} 840 ValueBit(Kind K = Variable) 841 : V(SDValue(nullptr, 0)), Idx(UINT32_MAX), K(K) {} 842 843 bool isZero() const { 844 return K == ConstZero; 845 } 846 847 bool hasValue() const { 848 return K == Variable; 849 } 850 851 SDValue getValue() const { 852 assert(hasValue() && "Cannot get the value of a constant bit"); 853 return V; 854 } 855 856 unsigned getValueBitIndex() const { 857 assert(hasValue() && "Cannot get the value bit index of a constant bit"); 858 return Idx; 859 } 860 }; 861 862 // A bit group has the same underlying value and the same rotate factor. 863 struct BitGroup { 864 SDValue V; 865 unsigned RLAmt; 866 unsigned StartIdx, EndIdx; 867 868 // This rotation amount assumes that the lower 32 bits of the quantity are 869 // replicated in the high 32 bits by the rotation operator (which is done 870 // by rlwinm and friends in 64-bit mode). 871 bool Repl32; 872 // Did converting to Repl32 == true change the rotation factor? If it did, 873 // it decreased it by 32. 874 bool Repl32CR; 875 // Was this group coalesced after setting Repl32 to true? 876 bool Repl32Coalesced; 877 878 BitGroup(SDValue V, unsigned R, unsigned S, unsigned E) 879 : V(V), RLAmt(R), StartIdx(S), EndIdx(E), Repl32(false), Repl32CR(false), 880 Repl32Coalesced(false) { 881 DEBUG(dbgs() << "\tbit group for " << V.getNode() << " RLAmt = " << R << 882 " [" << S << ", " << E << "]\n"); 883 } 884 }; 885 886 // Information on each (Value, RLAmt) pair (like the number of groups 887 // associated with each) used to choose the lowering method. 888 struct ValueRotInfo { 889 SDValue V; 890 unsigned RLAmt; 891 unsigned NumGroups; 892 unsigned FirstGroupStartIdx; 893 bool Repl32; 894 895 ValueRotInfo() 896 : RLAmt(UINT32_MAX), NumGroups(0), FirstGroupStartIdx(UINT32_MAX), 897 Repl32(false) {} 898 899 // For sorting (in reverse order) by NumGroups, and then by 900 // FirstGroupStartIdx. 901 bool operator < (const ValueRotInfo &Other) const { 902 // We need to sort so that the non-Repl32 come first because, when we're 903 // doing masking, the Repl32 bit groups might be subsumed into the 64-bit 904 // masking operation. 905 if (Repl32 < Other.Repl32) 906 return true; 907 else if (Repl32 > Other.Repl32) 908 return false; 909 else if (NumGroups > Other.NumGroups) 910 return true; 911 else if (NumGroups < Other.NumGroups) 912 return false; 913 else if (FirstGroupStartIdx < Other.FirstGroupStartIdx) 914 return true; 915 return false; 916 } 917 }; 918 919 // Return true if something interesting was deduced, return false if we're 920 // providing only a generic representation of V (or something else likewise 921 // uninteresting for instruction selection). 922 bool getValueBits(SDValue V, SmallVector<ValueBit, 64> &Bits) { 923 switch (V.getOpcode()) { 924 default: break; 925 case ISD::ROTL: 926 if (isa<ConstantSDNode>(V.getOperand(1))) { 927 unsigned RotAmt = V.getConstantOperandVal(1); 928 929 SmallVector<ValueBit, 64> LHSBits(Bits.size()); 930 getValueBits(V.getOperand(0), LHSBits); 931 932 for (unsigned i = 0; i < Bits.size(); ++i) 933 Bits[i] = LHSBits[i < RotAmt ? i + (Bits.size() - RotAmt) : i - RotAmt]; 934 935 return true; 936 } 937 break; 938 case ISD::SHL: 939 if (isa<ConstantSDNode>(V.getOperand(1))) { 940 unsigned ShiftAmt = V.getConstantOperandVal(1); 941 942 SmallVector<ValueBit, 64> LHSBits(Bits.size()); 943 getValueBits(V.getOperand(0), LHSBits); 944 945 for (unsigned i = ShiftAmt; i < Bits.size(); ++i) 946 Bits[i] = LHSBits[i - ShiftAmt]; 947 948 for (unsigned i = 0; i < ShiftAmt; ++i) 949 Bits[i] = ValueBit(ValueBit::ConstZero); 950 951 return true; 952 } 953 break; 954 case ISD::SRL: 955 if (isa<ConstantSDNode>(V.getOperand(1))) { 956 unsigned ShiftAmt = V.getConstantOperandVal(1); 957 958 SmallVector<ValueBit, 64> LHSBits(Bits.size()); 959 getValueBits(V.getOperand(0), LHSBits); 960 961 for (unsigned i = 0; i < Bits.size() - ShiftAmt; ++i) 962 Bits[i] = LHSBits[i + ShiftAmt]; 963 964 for (unsigned i = Bits.size() - ShiftAmt; i < Bits.size(); ++i) 965 Bits[i] = ValueBit(ValueBit::ConstZero); 966 967 return true; 968 } 969 break; 970 case ISD::AND: 971 if (isa<ConstantSDNode>(V.getOperand(1))) { 972 uint64_t Mask = V.getConstantOperandVal(1); 973 974 SmallVector<ValueBit, 64> LHSBits(Bits.size()); 975 bool LHSTrivial = getValueBits(V.getOperand(0), LHSBits); 976 977 for (unsigned i = 0; i < Bits.size(); ++i) 978 if (((Mask >> i) & 1) == 1) 979 Bits[i] = LHSBits[i]; 980 else 981 Bits[i] = ValueBit(ValueBit::ConstZero); 982 983 // Mark this as interesting, only if the LHS was also interesting. This 984 // prevents the overall procedure from matching a single immediate 'and' 985 // (which is non-optimal because such an and might be folded with other 986 // things if we don't select it here). 987 return LHSTrivial; 988 } 989 break; 990 case ISD::OR: { 991 SmallVector<ValueBit, 64> LHSBits(Bits.size()), RHSBits(Bits.size()); 992 getValueBits(V.getOperand(0), LHSBits); 993 getValueBits(V.getOperand(1), RHSBits); 994 995 bool AllDisjoint = true; 996 for (unsigned i = 0; i < Bits.size(); ++i) 997 if (LHSBits[i].isZero()) 998 Bits[i] = RHSBits[i]; 999 else if (RHSBits[i].isZero()) 1000 Bits[i] = LHSBits[i]; 1001 else { 1002 AllDisjoint = false; 1003 break; 1004 } 1005 1006 if (!AllDisjoint) 1007 break; 1008 1009 return true; 1010 } 1011 } 1012 1013 for (unsigned i = 0; i < Bits.size(); ++i) 1014 Bits[i] = ValueBit(V, i); 1015 1016 return false; 1017 } 1018 1019 // For each value (except the constant ones), compute the left-rotate amount 1020 // to get it from its original to final position. 1021 void computeRotationAmounts() { 1022 HasZeros = false; 1023 RLAmt.resize(Bits.size()); 1024 for (unsigned i = 0; i < Bits.size(); ++i) 1025 if (Bits[i].hasValue()) { 1026 unsigned VBI = Bits[i].getValueBitIndex(); 1027 if (i >= VBI) 1028 RLAmt[i] = i - VBI; 1029 else 1030 RLAmt[i] = Bits.size() - (VBI - i); 1031 } else if (Bits[i].isZero()) { 1032 HasZeros = true; 1033 RLAmt[i] = UINT32_MAX; 1034 } else { 1035 llvm_unreachable("Unknown value bit type"); 1036 } 1037 } 1038 1039 // Collect groups of consecutive bits with the same underlying value and 1040 // rotation factor. If we're doing late masking, we ignore zeros, otherwise 1041 // they break up groups. 1042 void collectBitGroups(bool LateMask) { 1043 BitGroups.clear(); 1044 1045 unsigned LastRLAmt = RLAmt[0]; 1046 SDValue LastValue = Bits[0].hasValue() ? Bits[0].getValue() : SDValue(); 1047 unsigned LastGroupStartIdx = 0; 1048 for (unsigned i = 1; i < Bits.size(); ++i) { 1049 unsigned ThisRLAmt = RLAmt[i]; 1050 SDValue ThisValue = Bits[i].hasValue() ? Bits[i].getValue() : SDValue(); 1051 if (LateMask && !ThisValue) { 1052 ThisValue = LastValue; 1053 ThisRLAmt = LastRLAmt; 1054 // If we're doing late masking, then the first bit group always starts 1055 // at zero (even if the first bits were zero). 1056 if (BitGroups.empty()) 1057 LastGroupStartIdx = 0; 1058 } 1059 1060 // If this bit has the same underlying value and the same rotate factor as 1061 // the last one, then they're part of the same group. 1062 if (ThisRLAmt == LastRLAmt && ThisValue == LastValue) 1063 continue; 1064 1065 if (LastValue.getNode()) 1066 BitGroups.push_back(BitGroup(LastValue, LastRLAmt, LastGroupStartIdx, 1067 i-1)); 1068 LastRLAmt = ThisRLAmt; 1069 LastValue = ThisValue; 1070 LastGroupStartIdx = i; 1071 } 1072 if (LastValue.getNode()) 1073 BitGroups.push_back(BitGroup(LastValue, LastRLAmt, LastGroupStartIdx, 1074 Bits.size()-1)); 1075 1076 if (BitGroups.empty()) 1077 return; 1078 1079 // We might be able to combine the first and last groups. 1080 if (BitGroups.size() > 1) { 1081 // If the first and last groups are the same, then remove the first group 1082 // in favor of the last group, making the ending index of the last group 1083 // equal to the ending index of the to-be-removed first group. 1084 if (BitGroups[0].StartIdx == 0 && 1085 BitGroups[BitGroups.size()-1].EndIdx == Bits.size()-1 && 1086 BitGroups[0].V == BitGroups[BitGroups.size()-1].V && 1087 BitGroups[0].RLAmt == BitGroups[BitGroups.size()-1].RLAmt) { 1088 DEBUG(dbgs() << "\tcombining final bit group with initial one\n"); 1089 BitGroups[BitGroups.size()-1].EndIdx = BitGroups[0].EndIdx; 1090 BitGroups.erase(BitGroups.begin()); 1091 } 1092 } 1093 } 1094 1095 // Take all (SDValue, RLAmt) pairs and sort them by the number of groups 1096 // associated with each. If there is a degeneracy, pick the one that occurs 1097 // first (in the final value). 1098 void collectValueRotInfo() { 1099 ValueRots.clear(); 1100 1101 for (auto &BG : BitGroups) { 1102 unsigned RLAmtKey = BG.RLAmt + (BG.Repl32 ? 64 : 0); 1103 ValueRotInfo &VRI = ValueRots[std::make_pair(BG.V, RLAmtKey)]; 1104 VRI.V = BG.V; 1105 VRI.RLAmt = BG.RLAmt; 1106 VRI.Repl32 = BG.Repl32; 1107 VRI.NumGroups += 1; 1108 VRI.FirstGroupStartIdx = std::min(VRI.FirstGroupStartIdx, BG.StartIdx); 1109 } 1110 1111 // Now that we've collected the various ValueRotInfo instances, we need to 1112 // sort them. 1113 ValueRotsVec.clear(); 1114 for (auto &I : ValueRots) { 1115 ValueRotsVec.push_back(I.second); 1116 } 1117 std::sort(ValueRotsVec.begin(), ValueRotsVec.end()); 1118 } 1119 1120 // In 64-bit mode, rlwinm and friends have a rotation operator that 1121 // replicates the low-order 32 bits into the high-order 32-bits. The mask 1122 // indices of these instructions can only be in the lower 32 bits, so they 1123 // can only represent some 64-bit bit groups. However, when they can be used, 1124 // the 32-bit replication can be used to represent, as a single bit group, 1125 // otherwise separate bit groups. We'll convert to replicated-32-bit bit 1126 // groups when possible. Returns true if any of the bit groups were 1127 // converted. 1128 void assignRepl32BitGroups() { 1129 // If we have bits like this: 1130 // 1131 // Indices: 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 1132 // V bits: ... 7 6 5 4 3 2 1 0 31 30 29 28 27 26 25 24 1133 // Groups: | RLAmt = 8 | RLAmt = 40 | 1134 // 1135 // But, making use of a 32-bit operation that replicates the low-order 32 1136 // bits into the high-order 32 bits, this can be one bit group with a RLAmt 1137 // of 8. 1138 1139 auto IsAllLow32 = [this](BitGroup & BG) { 1140 if (BG.StartIdx <= BG.EndIdx) { 1141 for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i) { 1142 if (!Bits[i].hasValue()) 1143 continue; 1144 if (Bits[i].getValueBitIndex() >= 32) 1145 return false; 1146 } 1147 } else { 1148 for (unsigned i = BG.StartIdx; i < Bits.size(); ++i) { 1149 if (!Bits[i].hasValue()) 1150 continue; 1151 if (Bits[i].getValueBitIndex() >= 32) 1152 return false; 1153 } 1154 for (unsigned i = 0; i <= BG.EndIdx; ++i) { 1155 if (!Bits[i].hasValue()) 1156 continue; 1157 if (Bits[i].getValueBitIndex() >= 32) 1158 return false; 1159 } 1160 } 1161 1162 return true; 1163 }; 1164 1165 for (auto &BG : BitGroups) { 1166 if (BG.StartIdx < 32 && BG.EndIdx < 32) { 1167 if (IsAllLow32(BG)) { 1168 if (BG.RLAmt >= 32) { 1169 BG.RLAmt -= 32; 1170 BG.Repl32CR = true; 1171 } 1172 1173 BG.Repl32 = true; 1174 1175 DEBUG(dbgs() << "\t32-bit replicated bit group for " << 1176 BG.V.getNode() << " RLAmt = " << BG.RLAmt << 1177 " [" << BG.StartIdx << ", " << BG.EndIdx << "]\n"); 1178 } 1179 } 1180 } 1181 1182 // Now walk through the bit groups, consolidating where possible. 1183 for (auto I = BitGroups.begin(); I != BitGroups.end();) { 1184 // We might want to remove this bit group by merging it with the previous 1185 // group (which might be the ending group). 1186 auto IP = (I == BitGroups.begin()) ? 1187 std::prev(BitGroups.end()) : std::prev(I); 1188 if (I->Repl32 && IP->Repl32 && I->V == IP->V && I->RLAmt == IP->RLAmt && 1189 I->StartIdx == (IP->EndIdx + 1) % 64 && I != IP) { 1190 1191 DEBUG(dbgs() << "\tcombining 32-bit replicated bit group for " << 1192 I->V.getNode() << " RLAmt = " << I->RLAmt << 1193 " [" << I->StartIdx << ", " << I->EndIdx << 1194 "] with group with range [" << 1195 IP->StartIdx << ", " << IP->EndIdx << "]\n"); 1196 1197 IP->EndIdx = I->EndIdx; 1198 IP->Repl32CR = IP->Repl32CR || I->Repl32CR; 1199 IP->Repl32Coalesced = true; 1200 I = BitGroups.erase(I); 1201 continue; 1202 } else { 1203 // There is a special case worth handling: If there is a single group 1204 // covering the entire upper 32 bits, and it can be merged with both 1205 // the next and previous groups (which might be the same group), then 1206 // do so. If it is the same group (so there will be only one group in 1207 // total), then we need to reverse the order of the range so that it 1208 // covers the entire 64 bits. 1209 if (I->StartIdx == 32 && I->EndIdx == 63) { 1210 assert(std::next(I) == BitGroups.end() && 1211 "bit group ends at index 63 but there is another?"); 1212 auto IN = BitGroups.begin(); 1213 1214 if (IP->Repl32 && IN->Repl32 && I->V == IP->V && I->V == IN->V && 1215 (I->RLAmt % 32) == IP->RLAmt && (I->RLAmt % 32) == IN->RLAmt && 1216 IP->EndIdx == 31 && IN->StartIdx == 0 && I != IP && 1217 IsAllLow32(*I)) { 1218 1219 DEBUG(dbgs() << "\tcombining bit group for " << 1220 I->V.getNode() << " RLAmt = " << I->RLAmt << 1221 " [" << I->StartIdx << ", " << I->EndIdx << 1222 "] with 32-bit replicated groups with ranges [" << 1223 IP->StartIdx << ", " << IP->EndIdx << "] and [" << 1224 IN->StartIdx << ", " << IN->EndIdx << "]\n"); 1225 1226 if (IP == IN) { 1227 // There is only one other group; change it to cover the whole 1228 // range (backward, so that it can still be Repl32 but cover the 1229 // whole 64-bit range). 1230 IP->StartIdx = 31; 1231 IP->EndIdx = 30; 1232 IP->Repl32CR = IP->Repl32CR || I->RLAmt >= 32; 1233 IP->Repl32Coalesced = true; 1234 I = BitGroups.erase(I); 1235 } else { 1236 // There are two separate groups, one before this group and one 1237 // after us (at the beginning). We're going to remove this group, 1238 // but also the group at the very beginning. 1239 IP->EndIdx = IN->EndIdx; 1240 IP->Repl32CR = IP->Repl32CR || IN->Repl32CR || I->RLAmt >= 32; 1241 IP->Repl32Coalesced = true; 1242 I = BitGroups.erase(I); 1243 BitGroups.erase(BitGroups.begin()); 1244 } 1245 1246 // This must be the last group in the vector (and we might have 1247 // just invalidated the iterator above), so break here. 1248 break; 1249 } 1250 } 1251 } 1252 1253 ++I; 1254 } 1255 } 1256 1257 SDValue getI32Imm(unsigned Imm, SDLoc dl) { 1258 return CurDAG->getTargetConstant(Imm, dl, MVT::i32); 1259 } 1260 1261 uint64_t getZerosMask() { 1262 uint64_t Mask = 0; 1263 for (unsigned i = 0; i < Bits.size(); ++i) { 1264 if (Bits[i].hasValue()) 1265 continue; 1266 Mask |= (UINT64_C(1) << i); 1267 } 1268 1269 return ~Mask; 1270 } 1271 1272 // Depending on the number of groups for a particular value, it might be 1273 // better to rotate, mask explicitly (using andi/andis), and then or the 1274 // result. Select this part of the result first. 1275 void SelectAndParts32(SDLoc dl, SDValue &Res, unsigned *InstCnt) { 1276 if (BPermRewriterNoMasking) 1277 return; 1278 1279 for (ValueRotInfo &VRI : ValueRotsVec) { 1280 unsigned Mask = 0; 1281 for (unsigned i = 0; i < Bits.size(); ++i) { 1282 if (!Bits[i].hasValue() || Bits[i].getValue() != VRI.V) 1283 continue; 1284 if (RLAmt[i] != VRI.RLAmt) 1285 continue; 1286 Mask |= (1u << i); 1287 } 1288 1289 // Compute the masks for andi/andis that would be necessary. 1290 unsigned ANDIMask = (Mask & UINT16_MAX), ANDISMask = Mask >> 16; 1291 assert((ANDIMask != 0 || ANDISMask != 0) && 1292 "No set bits in mask for value bit groups"); 1293 bool NeedsRotate = VRI.RLAmt != 0; 1294 1295 // We're trying to minimize the number of instructions. If we have one 1296 // group, using one of andi/andis can break even. If we have three 1297 // groups, we can use both andi and andis and break even (to use both 1298 // andi and andis we also need to or the results together). We need four 1299 // groups if we also need to rotate. To use andi/andis we need to do more 1300 // than break even because rotate-and-mask instructions tend to be easier 1301 // to schedule. 1302 1303 // FIXME: We've biased here against using andi/andis, which is right for 1304 // POWER cores, but not optimal everywhere. For example, on the A2, 1305 // andi/andis have single-cycle latency whereas the rotate-and-mask 1306 // instructions take two cycles, and it would be better to bias toward 1307 // andi/andis in break-even cases. 1308 1309 unsigned NumAndInsts = (unsigned) NeedsRotate + 1310 (unsigned) (ANDIMask != 0) + 1311 (unsigned) (ANDISMask != 0) + 1312 (unsigned) (ANDIMask != 0 && ANDISMask != 0) + 1313 (unsigned) (bool) Res; 1314 1315 DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() << 1316 " RL: " << VRI.RLAmt << ":" << 1317 "\n\t\t\tisel using masking: " << NumAndInsts << 1318 " using rotates: " << VRI.NumGroups << "\n"); 1319 1320 if (NumAndInsts >= VRI.NumGroups) 1321 continue; 1322 1323 DEBUG(dbgs() << "\t\t\t\tusing masking\n"); 1324 1325 if (InstCnt) *InstCnt += NumAndInsts; 1326 1327 SDValue VRot; 1328 if (VRI.RLAmt) { 1329 SDValue Ops[] = 1330 { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl), 1331 getI32Imm(31, dl) }; 1332 VRot = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, 1333 Ops), 0); 1334 } else { 1335 VRot = VRI.V; 1336 } 1337 1338 SDValue ANDIVal, ANDISVal; 1339 if (ANDIMask != 0) 1340 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32, 1341 VRot, getI32Imm(ANDIMask, dl)), 0); 1342 if (ANDISMask != 0) 1343 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32, 1344 VRot, getI32Imm(ANDISMask, dl)), 0); 1345 1346 SDValue TotalVal; 1347 if (!ANDIVal) 1348 TotalVal = ANDISVal; 1349 else if (!ANDISVal) 1350 TotalVal = ANDIVal; 1351 else 1352 TotalVal = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32, 1353 ANDIVal, ANDISVal), 0); 1354 1355 if (!Res) 1356 Res = TotalVal; 1357 else 1358 Res = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32, 1359 Res, TotalVal), 0); 1360 1361 // Now, remove all groups with this underlying value and rotation 1362 // factor. 1363 eraseMatchingBitGroups([VRI](const BitGroup &BG) { 1364 return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt; 1365 }); 1366 } 1367 } 1368 1369 // Instruction selection for the 32-bit case. 1370 SDNode *Select32(SDNode *N, bool LateMask, unsigned *InstCnt) { 1371 SDLoc dl(N); 1372 SDValue Res; 1373 1374 if (InstCnt) *InstCnt = 0; 1375 1376 // Take care of cases that should use andi/andis first. 1377 SelectAndParts32(dl, Res, InstCnt); 1378 1379 // If we've not yet selected a 'starting' instruction, and we have no zeros 1380 // to fill in, select the (Value, RLAmt) with the highest priority (largest 1381 // number of groups), and start with this rotated value. 1382 if ((!HasZeros || LateMask) && !Res) { 1383 ValueRotInfo &VRI = ValueRotsVec[0]; 1384 if (VRI.RLAmt) { 1385 if (InstCnt) *InstCnt += 1; 1386 SDValue Ops[] = 1387 { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl), 1388 getI32Imm(31, dl) }; 1389 Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 1390 0); 1391 } else { 1392 Res = VRI.V; 1393 } 1394 1395 // Now, remove all groups with this underlying value and rotation factor. 1396 eraseMatchingBitGroups([VRI](const BitGroup &BG) { 1397 return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt; 1398 }); 1399 } 1400 1401 if (InstCnt) *InstCnt += BitGroups.size(); 1402 1403 // Insert the other groups (one at a time). 1404 for (auto &BG : BitGroups) { 1405 if (!Res) { 1406 SDValue Ops[] = 1407 { BG.V, getI32Imm(BG.RLAmt, dl), 1408 getI32Imm(Bits.size() - BG.EndIdx - 1, dl), 1409 getI32Imm(Bits.size() - BG.StartIdx - 1, dl) }; 1410 Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); 1411 } else { 1412 SDValue Ops[] = 1413 { Res, BG.V, getI32Imm(BG.RLAmt, dl), 1414 getI32Imm(Bits.size() - BG.EndIdx - 1, dl), 1415 getI32Imm(Bits.size() - BG.StartIdx - 1, dl) }; 1416 Res = SDValue(CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops), 0); 1417 } 1418 } 1419 1420 if (LateMask) { 1421 unsigned Mask = (unsigned) getZerosMask(); 1422 1423 unsigned ANDIMask = (Mask & UINT16_MAX), ANDISMask = Mask >> 16; 1424 assert((ANDIMask != 0 || ANDISMask != 0) && 1425 "No set bits in zeros mask?"); 1426 1427 if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) + 1428 (unsigned) (ANDISMask != 0) + 1429 (unsigned) (ANDIMask != 0 && ANDISMask != 0); 1430 1431 SDValue ANDIVal, ANDISVal; 1432 if (ANDIMask != 0) 1433 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32, 1434 Res, getI32Imm(ANDIMask, dl)), 0); 1435 if (ANDISMask != 0) 1436 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32, 1437 Res, getI32Imm(ANDISMask, dl)), 0); 1438 1439 if (!ANDIVal) 1440 Res = ANDISVal; 1441 else if (!ANDISVal) 1442 Res = ANDIVal; 1443 else 1444 Res = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32, 1445 ANDIVal, ANDISVal), 0); 1446 } 1447 1448 return Res.getNode(); 1449 } 1450 1451 unsigned SelectRotMask64Count(unsigned RLAmt, bool Repl32, 1452 unsigned MaskStart, unsigned MaskEnd, 1453 bool IsIns) { 1454 // In the notation used by the instructions, 'start' and 'end' are reversed 1455 // because bits are counted from high to low order. 1456 unsigned InstMaskStart = 64 - MaskEnd - 1, 1457 InstMaskEnd = 64 - MaskStart - 1; 1458 1459 if (Repl32) 1460 return 1; 1461 1462 if ((!IsIns && (InstMaskEnd == 63 || InstMaskStart == 0)) || 1463 InstMaskEnd == 63 - RLAmt) 1464 return 1; 1465 1466 return 2; 1467 } 1468 1469 // For 64-bit values, not all combinations of rotates and masks are 1470 // available. Produce one if it is available. 1471 SDValue SelectRotMask64(SDValue V, SDLoc dl, unsigned RLAmt, bool Repl32, 1472 unsigned MaskStart, unsigned MaskEnd, 1473 unsigned *InstCnt = nullptr) { 1474 // In the notation used by the instructions, 'start' and 'end' are reversed 1475 // because bits are counted from high to low order. 1476 unsigned InstMaskStart = 64 - MaskEnd - 1, 1477 InstMaskEnd = 64 - MaskStart - 1; 1478 1479 if (InstCnt) *InstCnt += 1; 1480 1481 if (Repl32) { 1482 // This rotation amount assumes that the lower 32 bits of the quantity 1483 // are replicated in the high 32 bits by the rotation operator (which is 1484 // done by rlwinm and friends). 1485 assert(InstMaskStart >= 32 && "Mask cannot start out of range"); 1486 assert(InstMaskEnd >= 32 && "Mask cannot end out of range"); 1487 SDValue Ops[] = 1488 { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl), 1489 getI32Imm(InstMaskEnd - 32, dl) }; 1490 return SDValue(CurDAG->getMachineNode(PPC::RLWINM8, dl, MVT::i64, 1491 Ops), 0); 1492 } 1493 1494 if (InstMaskEnd == 63) { 1495 SDValue Ops[] = 1496 { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) }; 1497 return SDValue(CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, Ops), 0); 1498 } 1499 1500 if (InstMaskStart == 0) { 1501 SDValue Ops[] = 1502 { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskEnd, dl) }; 1503 return SDValue(CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Ops), 0); 1504 } 1505 1506 if (InstMaskEnd == 63 - RLAmt) { 1507 SDValue Ops[] = 1508 { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) }; 1509 return SDValue(CurDAG->getMachineNode(PPC::RLDIC, dl, MVT::i64, Ops), 0); 1510 } 1511 1512 // We cannot do this with a single instruction, so we'll use two. The 1513 // problem is that we're not free to choose both a rotation amount and mask 1514 // start and end independently. We can choose an arbitrary mask start and 1515 // end, but then the rotation amount is fixed. Rotation, however, can be 1516 // inverted, and so by applying an "inverse" rotation first, we can get the 1517 // desired result. 1518 if (InstCnt) *InstCnt += 1; 1519 1520 // The rotation mask for the second instruction must be MaskStart. 1521 unsigned RLAmt2 = MaskStart; 1522 // The first instruction must rotate V so that the overall rotation amount 1523 // is RLAmt. 1524 unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64; 1525 if (RLAmt1) 1526 V = SelectRotMask64(V, dl, RLAmt1, false, 0, 63); 1527 return SelectRotMask64(V, dl, RLAmt2, false, MaskStart, MaskEnd); 1528 } 1529 1530 // For 64-bit values, not all combinations of rotates and masks are 1531 // available. Produce a rotate-mask-and-insert if one is available. 1532 SDValue SelectRotMaskIns64(SDValue Base, SDValue V, SDLoc dl, unsigned RLAmt, 1533 bool Repl32, unsigned MaskStart, 1534 unsigned MaskEnd, unsigned *InstCnt = nullptr) { 1535 // In the notation used by the instructions, 'start' and 'end' are reversed 1536 // because bits are counted from high to low order. 1537 unsigned InstMaskStart = 64 - MaskEnd - 1, 1538 InstMaskEnd = 64 - MaskStart - 1; 1539 1540 if (InstCnt) *InstCnt += 1; 1541 1542 if (Repl32) { 1543 // This rotation amount assumes that the lower 32 bits of the quantity 1544 // are replicated in the high 32 bits by the rotation operator (which is 1545 // done by rlwinm and friends). 1546 assert(InstMaskStart >= 32 && "Mask cannot start out of range"); 1547 assert(InstMaskEnd >= 32 && "Mask cannot end out of range"); 1548 SDValue Ops[] = 1549 { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl), 1550 getI32Imm(InstMaskEnd - 32, dl) }; 1551 return SDValue(CurDAG->getMachineNode(PPC::RLWIMI8, dl, MVT::i64, 1552 Ops), 0); 1553 } 1554 1555 if (InstMaskEnd == 63 - RLAmt) { 1556 SDValue Ops[] = 1557 { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) }; 1558 return SDValue(CurDAG->getMachineNode(PPC::RLDIMI, dl, MVT::i64, Ops), 0); 1559 } 1560 1561 // We cannot do this with a single instruction, so we'll use two. The 1562 // problem is that we're not free to choose both a rotation amount and mask 1563 // start and end independently. We can choose an arbitrary mask start and 1564 // end, but then the rotation amount is fixed. Rotation, however, can be 1565 // inverted, and so by applying an "inverse" rotation first, we can get the 1566 // desired result. 1567 if (InstCnt) *InstCnt += 1; 1568 1569 // The rotation mask for the second instruction must be MaskStart. 1570 unsigned RLAmt2 = MaskStart; 1571 // The first instruction must rotate V so that the overall rotation amount 1572 // is RLAmt. 1573 unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64; 1574 if (RLAmt1) 1575 V = SelectRotMask64(V, dl, RLAmt1, false, 0, 63); 1576 return SelectRotMaskIns64(Base, V, dl, RLAmt2, false, MaskStart, MaskEnd); 1577 } 1578 1579 void SelectAndParts64(SDLoc dl, SDValue &Res, unsigned *InstCnt) { 1580 if (BPermRewriterNoMasking) 1581 return; 1582 1583 // The idea here is the same as in the 32-bit version, but with additional 1584 // complications from the fact that Repl32 might be true. Because we 1585 // aggressively convert bit groups to Repl32 form (which, for small 1586 // rotation factors, involves no other change), and then coalesce, it might 1587 // be the case that a single 64-bit masking operation could handle both 1588 // some Repl32 groups and some non-Repl32 groups. If converting to Repl32 1589 // form allowed coalescing, then we must use a 32-bit rotaton in order to 1590 // completely capture the new combined bit group. 1591 1592 for (ValueRotInfo &VRI : ValueRotsVec) { 1593 uint64_t Mask = 0; 1594 1595 // We need to add to the mask all bits from the associated bit groups. 1596 // If Repl32 is false, we need to add bits from bit groups that have 1597 // Repl32 true, but are trivially convertable to Repl32 false. Such a 1598 // group is trivially convertable if it overlaps only with the lower 32 1599 // bits, and the group has not been coalesced. 1600 auto MatchingBG = [VRI](const BitGroup &BG) { 1601 if (VRI.V != BG.V) 1602 return false; 1603 1604 unsigned EffRLAmt = BG.RLAmt; 1605 if (!VRI.Repl32 && BG.Repl32) { 1606 if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx <= BG.EndIdx && 1607 !BG.Repl32Coalesced) { 1608 if (BG.Repl32CR) 1609 EffRLAmt += 32; 1610 } else { 1611 return false; 1612 } 1613 } else if (VRI.Repl32 != BG.Repl32) { 1614 return false; 1615 } 1616 1617 if (VRI.RLAmt != EffRLAmt) 1618 return false; 1619 1620 return true; 1621 }; 1622 1623 for (auto &BG : BitGroups) { 1624 if (!MatchingBG(BG)) 1625 continue; 1626 1627 if (BG.StartIdx <= BG.EndIdx) { 1628 for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i) 1629 Mask |= (UINT64_C(1) << i); 1630 } else { 1631 for (unsigned i = BG.StartIdx; i < Bits.size(); ++i) 1632 Mask |= (UINT64_C(1) << i); 1633 for (unsigned i = 0; i <= BG.EndIdx; ++i) 1634 Mask |= (UINT64_C(1) << i); 1635 } 1636 } 1637 1638 // We can use the 32-bit andi/andis technique if the mask does not 1639 // require any higher-order bits. This can save an instruction compared 1640 // to always using the general 64-bit technique. 1641 bool Use32BitInsts = isUInt<32>(Mask); 1642 // Compute the masks for andi/andis that would be necessary. 1643 unsigned ANDIMask = (Mask & UINT16_MAX), 1644 ANDISMask = (Mask >> 16) & UINT16_MAX; 1645 1646 bool NeedsRotate = VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(Mask)); 1647 1648 unsigned NumAndInsts = (unsigned) NeedsRotate + 1649 (unsigned) (bool) Res; 1650 if (Use32BitInsts) 1651 NumAndInsts += (unsigned) (ANDIMask != 0) + (unsigned) (ANDISMask != 0) + 1652 (unsigned) (ANDIMask != 0 && ANDISMask != 0); 1653 else 1654 NumAndInsts += SelectInt64Count(Mask) + /* and */ 1; 1655 1656 unsigned NumRLInsts = 0; 1657 bool FirstBG = true; 1658 for (auto &BG : BitGroups) { 1659 if (!MatchingBG(BG)) 1660 continue; 1661 NumRLInsts += 1662 SelectRotMask64Count(BG.RLAmt, BG.Repl32, BG.StartIdx, BG.EndIdx, 1663 !FirstBG); 1664 FirstBG = false; 1665 } 1666 1667 DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() << 1668 " RL: " << VRI.RLAmt << (VRI.Repl32 ? " (32):" : ":") << 1669 "\n\t\t\tisel using masking: " << NumAndInsts << 1670 " using rotates: " << NumRLInsts << "\n"); 1671 1672 // When we'd use andi/andis, we bias toward using the rotates (andi only 1673 // has a record form, and is cracked on POWER cores). However, when using 1674 // general 64-bit constant formation, bias toward the constant form, 1675 // because that exposes more opportunities for CSE. 1676 if (NumAndInsts > NumRLInsts) 1677 continue; 1678 if (Use32BitInsts && NumAndInsts == NumRLInsts) 1679 continue; 1680 1681 DEBUG(dbgs() << "\t\t\t\tusing masking\n"); 1682 1683 if (InstCnt) *InstCnt += NumAndInsts; 1684 1685 SDValue VRot; 1686 // We actually need to generate a rotation if we have a non-zero rotation 1687 // factor or, in the Repl32 case, if we care about any of the 1688 // higher-order replicated bits. In the latter case, we generate a mask 1689 // backward so that it actually includes the entire 64 bits. 1690 if (VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(Mask))) 1691 VRot = SelectRotMask64(VRI.V, dl, VRI.RLAmt, VRI.Repl32, 1692 VRI.Repl32 ? 31 : 0, VRI.Repl32 ? 30 : 63); 1693 else 1694 VRot = VRI.V; 1695 1696 SDValue TotalVal; 1697 if (Use32BitInsts) { 1698 assert((ANDIMask != 0 || ANDISMask != 0) && 1699 "No set bits in mask when using 32-bit ands for 64-bit value"); 1700 1701 SDValue ANDIVal, ANDISVal; 1702 if (ANDIMask != 0) 1703 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64, 1704 VRot, getI32Imm(ANDIMask, dl)), 0); 1705 if (ANDISMask != 0) 1706 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64, 1707 VRot, getI32Imm(ANDISMask, dl)), 0); 1708 1709 if (!ANDIVal) 1710 TotalVal = ANDISVal; 1711 else if (!ANDISVal) 1712 TotalVal = ANDIVal; 1713 else 1714 TotalVal = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64, 1715 ANDIVal, ANDISVal), 0); 1716 } else { 1717 TotalVal = SDValue(SelectInt64(CurDAG, dl, Mask), 0); 1718 TotalVal = 1719 SDValue(CurDAG->getMachineNode(PPC::AND8, dl, MVT::i64, 1720 VRot, TotalVal), 0); 1721 } 1722 1723 if (!Res) 1724 Res = TotalVal; 1725 else 1726 Res = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64, 1727 Res, TotalVal), 0); 1728 1729 // Now, remove all groups with this underlying value and rotation 1730 // factor. 1731 eraseMatchingBitGroups(MatchingBG); 1732 } 1733 } 1734 1735 // Instruction selection for the 64-bit case. 1736 SDNode *Select64(SDNode *N, bool LateMask, unsigned *InstCnt) { 1737 SDLoc dl(N); 1738 SDValue Res; 1739 1740 if (InstCnt) *InstCnt = 0; 1741 1742 // Take care of cases that should use andi/andis first. 1743 SelectAndParts64(dl, Res, InstCnt); 1744 1745 // If we've not yet selected a 'starting' instruction, and we have no zeros 1746 // to fill in, select the (Value, RLAmt) with the highest priority (largest 1747 // number of groups), and start with this rotated value. 1748 if ((!HasZeros || LateMask) && !Res) { 1749 // If we have both Repl32 groups and non-Repl32 groups, the non-Repl32 1750 // groups will come first, and so the VRI representing the largest number 1751 // of groups might not be first (it might be the first Repl32 groups). 1752 unsigned MaxGroupsIdx = 0; 1753 if (!ValueRotsVec[0].Repl32) { 1754 for (unsigned i = 0, ie = ValueRotsVec.size(); i < ie; ++i) 1755 if (ValueRotsVec[i].Repl32) { 1756 if (ValueRotsVec[i].NumGroups > ValueRotsVec[0].NumGroups) 1757 MaxGroupsIdx = i; 1758 break; 1759 } 1760 } 1761 1762 ValueRotInfo &VRI = ValueRotsVec[MaxGroupsIdx]; 1763 bool NeedsRotate = false; 1764 if (VRI.RLAmt) { 1765 NeedsRotate = true; 1766 } else if (VRI.Repl32) { 1767 for (auto &BG : BitGroups) { 1768 if (BG.V != VRI.V || BG.RLAmt != VRI.RLAmt || 1769 BG.Repl32 != VRI.Repl32) 1770 continue; 1771 1772 // We don't need a rotate if the bit group is confined to the lower 1773 // 32 bits. 1774 if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx < BG.EndIdx) 1775 continue; 1776 1777 NeedsRotate = true; 1778 break; 1779 } 1780 } 1781 1782 if (NeedsRotate) 1783 Res = SelectRotMask64(VRI.V, dl, VRI.RLAmt, VRI.Repl32, 1784 VRI.Repl32 ? 31 : 0, VRI.Repl32 ? 30 : 63, 1785 InstCnt); 1786 else 1787 Res = VRI.V; 1788 1789 // Now, remove all groups with this underlying value and rotation factor. 1790 if (Res) 1791 eraseMatchingBitGroups([VRI](const BitGroup &BG) { 1792 return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt && 1793 BG.Repl32 == VRI.Repl32; 1794 }); 1795 } 1796 1797 // Because 64-bit rotates are more flexible than inserts, we might have a 1798 // preference regarding which one we do first (to save one instruction). 1799 if (!Res) 1800 for (auto I = BitGroups.begin(), IE = BitGroups.end(); I != IE; ++I) { 1801 if (SelectRotMask64Count(I->RLAmt, I->Repl32, I->StartIdx, I->EndIdx, 1802 false) < 1803 SelectRotMask64Count(I->RLAmt, I->Repl32, I->StartIdx, I->EndIdx, 1804 true)) { 1805 if (I != BitGroups.begin()) { 1806 BitGroup BG = *I; 1807 BitGroups.erase(I); 1808 BitGroups.insert(BitGroups.begin(), BG); 1809 } 1810 1811 break; 1812 } 1813 } 1814 1815 // Insert the other groups (one at a time). 1816 for (auto &BG : BitGroups) { 1817 if (!Res) 1818 Res = SelectRotMask64(BG.V, dl, BG.RLAmt, BG.Repl32, BG.StartIdx, 1819 BG.EndIdx, InstCnt); 1820 else 1821 Res = SelectRotMaskIns64(Res, BG.V, dl, BG.RLAmt, BG.Repl32, 1822 BG.StartIdx, BG.EndIdx, InstCnt); 1823 } 1824 1825 if (LateMask) { 1826 uint64_t Mask = getZerosMask(); 1827 1828 // We can use the 32-bit andi/andis technique if the mask does not 1829 // require any higher-order bits. This can save an instruction compared 1830 // to always using the general 64-bit technique. 1831 bool Use32BitInsts = isUInt<32>(Mask); 1832 // Compute the masks for andi/andis that would be necessary. 1833 unsigned ANDIMask = (Mask & UINT16_MAX), 1834 ANDISMask = (Mask >> 16) & UINT16_MAX; 1835 1836 if (Use32BitInsts) { 1837 assert((ANDIMask != 0 || ANDISMask != 0) && 1838 "No set bits in mask when using 32-bit ands for 64-bit value"); 1839 1840 if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) + 1841 (unsigned) (ANDISMask != 0) + 1842 (unsigned) (ANDIMask != 0 && ANDISMask != 0); 1843 1844 SDValue ANDIVal, ANDISVal; 1845 if (ANDIMask != 0) 1846 ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64, 1847 Res, getI32Imm(ANDIMask, dl)), 0); 1848 if (ANDISMask != 0) 1849 ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64, 1850 Res, getI32Imm(ANDISMask, dl)), 0); 1851 1852 if (!ANDIVal) 1853 Res = ANDISVal; 1854 else if (!ANDISVal) 1855 Res = ANDIVal; 1856 else 1857 Res = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64, 1858 ANDIVal, ANDISVal), 0); 1859 } else { 1860 if (InstCnt) *InstCnt += SelectInt64Count(Mask) + /* and */ 1; 1861 1862 SDValue MaskVal = SDValue(SelectInt64(CurDAG, dl, Mask), 0); 1863 Res = 1864 SDValue(CurDAG->getMachineNode(PPC::AND8, dl, MVT::i64, 1865 Res, MaskVal), 0); 1866 } 1867 } 1868 1869 return Res.getNode(); 1870 } 1871 1872 SDNode *Select(SDNode *N, bool LateMask, unsigned *InstCnt = nullptr) { 1873 // Fill in BitGroups. 1874 collectBitGroups(LateMask); 1875 if (BitGroups.empty()) 1876 return nullptr; 1877 1878 // For 64-bit values, figure out when we can use 32-bit instructions. 1879 if (Bits.size() == 64) 1880 assignRepl32BitGroups(); 1881 1882 // Fill in ValueRotsVec. 1883 collectValueRotInfo(); 1884 1885 if (Bits.size() == 32) { 1886 return Select32(N, LateMask, InstCnt); 1887 } else { 1888 assert(Bits.size() == 64 && "Not 64 bits here?"); 1889 return Select64(N, LateMask, InstCnt); 1890 } 1891 1892 return nullptr; 1893 } 1894 1895 void eraseMatchingBitGroups(function_ref<bool(const BitGroup &)> F) { 1896 BitGroups.erase(std::remove_if(BitGroups.begin(), BitGroups.end(), F), 1897 BitGroups.end()); 1898 } 1899 1900 SmallVector<ValueBit, 64> Bits; 1901 1902 bool HasZeros; 1903 SmallVector<unsigned, 64> RLAmt; 1904 1905 SmallVector<BitGroup, 16> BitGroups; 1906 1907 DenseMap<std::pair<SDValue, unsigned>, ValueRotInfo> ValueRots; 1908 SmallVector<ValueRotInfo, 16> ValueRotsVec; 1909 1910 SelectionDAG *CurDAG; 1911 1912 public: 1913 BitPermutationSelector(SelectionDAG *DAG) 1914 : CurDAG(DAG) {} 1915 1916 // Here we try to match complex bit permutations into a set of 1917 // rotate-and-shift/shift/and/or instructions, using a set of heuristics 1918 // known to produce optimial code for common cases (like i32 byte swapping). 1919 SDNode *Select(SDNode *N) { 1920 Bits.resize(N->getValueType(0).getSizeInBits()); 1921 if (!getValueBits(SDValue(N, 0), Bits)) 1922 return nullptr; 1923 1924 DEBUG(dbgs() << "Considering bit-permutation-based instruction" 1925 " selection for: "); 1926 DEBUG(N->dump(CurDAG)); 1927 1928 // Fill it RLAmt and set HasZeros. 1929 computeRotationAmounts(); 1930 1931 if (!HasZeros) 1932 return Select(N, false); 1933 1934 // We currently have two techniques for handling results with zeros: early 1935 // masking (the default) and late masking. Late masking is sometimes more 1936 // efficient, but because the structure of the bit groups is different, it 1937 // is hard to tell without generating both and comparing the results. With 1938 // late masking, we ignore zeros in the resulting value when inserting each 1939 // set of bit groups, and then mask in the zeros at the end. With early 1940 // masking, we only insert the non-zero parts of the result at every step. 1941 1942 unsigned InstCnt, InstCntLateMask; 1943 DEBUG(dbgs() << "\tEarly masking:\n"); 1944 SDNode *RN = Select(N, false, &InstCnt); 1945 DEBUG(dbgs() << "\t\tisel would use " << InstCnt << " instructions\n"); 1946 1947 DEBUG(dbgs() << "\tLate masking:\n"); 1948 SDNode *RNLM = Select(N, true, &InstCntLateMask); 1949 DEBUG(dbgs() << "\t\tisel would use " << InstCntLateMask << 1950 " instructions\n"); 1951 1952 if (InstCnt <= InstCntLateMask) { 1953 DEBUG(dbgs() << "\tUsing early-masking for isel\n"); 1954 return RN; 1955 } 1956 1957 DEBUG(dbgs() << "\tUsing late-masking for isel\n"); 1958 return RNLM; 1959 } 1960 }; 1961 } // anonymous namespace 1962 1963 SDNode *PPCDAGToDAGISel::SelectBitPermutation(SDNode *N) { 1964 if (N->getValueType(0) != MVT::i32 && 1965 N->getValueType(0) != MVT::i64) 1966 return nullptr; 1967 1968 if (!UseBitPermRewriter) 1969 return nullptr; 1970 1971 switch (N->getOpcode()) { 1972 default: break; 1973 case ISD::ROTL: 1974 case ISD::SHL: 1975 case ISD::SRL: 1976 case ISD::AND: 1977 case ISD::OR: { 1978 BitPermutationSelector BPS(CurDAG); 1979 return BPS.Select(N); 1980 } 1981 } 1982 1983 return nullptr; 1984 } 1985 1986 /// SelectCC - Select a comparison of the specified values with the specified 1987 /// condition code, returning the CR# of the expression. 1988 SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, 1989 ISD::CondCode CC, SDLoc dl) { 1990 // Always select the LHS. 1991 unsigned Opc; 1992 1993 if (LHS.getValueType() == MVT::i32) { 1994 unsigned Imm; 1995 if (CC == ISD::SETEQ || CC == ISD::SETNE) { 1996 if (isInt32Immediate(RHS, Imm)) { 1997 // SETEQ/SETNE comparison with 16-bit immediate, fold it. 1998 if (isUInt<16>(Imm)) 1999 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, 2000 getI32Imm(Imm & 0xFFFF, dl)), 2001 0); 2002 // If this is a 16-bit signed immediate, fold it. 2003 if (isInt<16>((int)Imm)) 2004 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, 2005 getI32Imm(Imm & 0xFFFF, dl)), 2006 0); 2007 2008 // For non-equality comparisons, the default code would materialize the 2009 // constant, then compare against it, like this: 2010 // lis r2, 4660 2011 // ori r2, r2, 22136 2012 // cmpw cr0, r3, r2 2013 // Since we are just comparing for equality, we can emit this instead: 2014 // xoris r0,r3,0x1234 2015 // cmplwi cr0,r0,0x5678 2016 // beq cr0,L6 2017 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS, 2018 getI32Imm(Imm >> 16, dl)), 0); 2019 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor, 2020 getI32Imm(Imm & 0xFFFF, dl)), 0); 2021 } 2022 Opc = PPC::CMPLW; 2023 } else if (ISD::isUnsignedIntSetCC(CC)) { 2024 if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm)) 2025 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, 2026 getI32Imm(Imm & 0xFFFF, dl)), 0); 2027 Opc = PPC::CMPLW; 2028 } else { 2029 short SImm; 2030 if (isIntS16Immediate(RHS, SImm)) 2031 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, 2032 getI32Imm((int)SImm & 0xFFFF, 2033 dl)), 2034 0); 2035 Opc = PPC::CMPW; 2036 } 2037 } else if (LHS.getValueType() == MVT::i64) { 2038 uint64_t Imm; 2039 if (CC == ISD::SETEQ || CC == ISD::SETNE) { 2040 if (isInt64Immediate(RHS.getNode(), Imm)) { 2041 // SETEQ/SETNE comparison with 16-bit immediate, fold it. 2042 if (isUInt<16>(Imm)) 2043 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, 2044 getI32Imm(Imm & 0xFFFF, dl)), 2045 0); 2046 // If this is a 16-bit signed immediate, fold it. 2047 if (isInt<16>(Imm)) 2048 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, 2049 getI32Imm(Imm & 0xFFFF, dl)), 2050 0); 2051 2052 // For non-equality comparisons, the default code would materialize the 2053 // constant, then compare against it, like this: 2054 // lis r2, 4660 2055 // ori r2, r2, 22136 2056 // cmpd cr0, r3, r2 2057 // Since we are just comparing for equality, we can emit this instead: 2058 // xoris r0,r3,0x1234 2059 // cmpldi cr0,r0,0x5678 2060 // beq cr0,L6 2061 if (isUInt<32>(Imm)) { 2062 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS, 2063 getI64Imm(Imm >> 16, dl)), 0); 2064 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor, 2065 getI64Imm(Imm & 0xFFFF, dl)), 2066 0); 2067 } 2068 } 2069 Opc = PPC::CMPLD; 2070 } else if (ISD::isUnsignedIntSetCC(CC)) { 2071 if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm)) 2072 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, 2073 getI64Imm(Imm & 0xFFFF, dl)), 0); 2074 Opc = PPC::CMPLD; 2075 } else { 2076 short SImm; 2077 if (isIntS16Immediate(RHS, SImm)) 2078 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, 2079 getI64Imm(SImm & 0xFFFF, dl)), 2080 0); 2081 Opc = PPC::CMPD; 2082 } 2083 } else if (LHS.getValueType() == MVT::f32) { 2084 Opc = PPC::FCMPUS; 2085 } else { 2086 assert(LHS.getValueType() == MVT::f64 && "Unknown vt!"); 2087 Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD; 2088 } 2089 return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0); 2090 } 2091 2092 static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) { 2093 switch (CC) { 2094 case ISD::SETUEQ: 2095 case ISD::SETONE: 2096 case ISD::SETOLE: 2097 case ISD::SETOGE: 2098 llvm_unreachable("Should be lowered by legalize!"); 2099 default: llvm_unreachable("Unknown condition!"); 2100 case ISD::SETOEQ: 2101 case ISD::SETEQ: return PPC::PRED_EQ; 2102 case ISD::SETUNE: 2103 case ISD::SETNE: return PPC::PRED_NE; 2104 case ISD::SETOLT: 2105 case ISD::SETLT: return PPC::PRED_LT; 2106 case ISD::SETULE: 2107 case ISD::SETLE: return PPC::PRED_LE; 2108 case ISD::SETOGT: 2109 case ISD::SETGT: return PPC::PRED_GT; 2110 case ISD::SETUGE: 2111 case ISD::SETGE: return PPC::PRED_GE; 2112 case ISD::SETO: return PPC::PRED_NU; 2113 case ISD::SETUO: return PPC::PRED_UN; 2114 // These two are invalid for floating point. Assume we have int. 2115 case ISD::SETULT: return PPC::PRED_LT; 2116 case ISD::SETUGT: return PPC::PRED_GT; 2117 } 2118 } 2119 2120 /// getCRIdxForSetCC - Return the index of the condition register field 2121 /// associated with the SetCC condition, and whether or not the field is 2122 /// treated as inverted. That is, lt = 0; ge = 0 inverted. 2123 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) { 2124 Invert = false; 2125 switch (CC) { 2126 default: llvm_unreachable("Unknown condition!"); 2127 case ISD::SETOLT: 2128 case ISD::SETLT: return 0; // Bit #0 = SETOLT 2129 case ISD::SETOGT: 2130 case ISD::SETGT: return 1; // Bit #1 = SETOGT 2131 case ISD::SETOEQ: 2132 case ISD::SETEQ: return 2; // Bit #2 = SETOEQ 2133 case ISD::SETUO: return 3; // Bit #3 = SETUO 2134 case ISD::SETUGE: 2135 case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE 2136 case ISD::SETULE: 2137 case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE 2138 case ISD::SETUNE: 2139 case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE 2140 case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO 2141 case ISD::SETUEQ: 2142 case ISD::SETOGE: 2143 case ISD::SETOLE: 2144 case ISD::SETONE: 2145 llvm_unreachable("Invalid branch code: should be expanded by legalize"); 2146 // These are invalid for floating point. Assume integer. 2147 case ISD::SETULT: return 0; 2148 case ISD::SETUGT: return 1; 2149 } 2150 } 2151 2152 // getVCmpInst: return the vector compare instruction for the specified 2153 // vector type and condition code. Since this is for altivec specific code, 2154 // only support the altivec types (v16i8, v8i16, v4i32, v2i64, and v4f32). 2155 static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC, 2156 bool HasVSX, bool &Swap, bool &Negate) { 2157 Swap = false; 2158 Negate = false; 2159 2160 if (VecVT.isFloatingPoint()) { 2161 /* Handle some cases by swapping input operands. */ 2162 switch (CC) { 2163 case ISD::SETLE: CC = ISD::SETGE; Swap = true; break; 2164 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break; 2165 case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break; 2166 case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break; 2167 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break; 2168 case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break; 2169 default: break; 2170 } 2171 /* Handle some cases by negating the result. */ 2172 switch (CC) { 2173 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break; 2174 case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break; 2175 case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break; 2176 case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break; 2177 default: break; 2178 } 2179 /* We have instructions implementing the remaining cases. */ 2180 switch (CC) { 2181 case ISD::SETEQ: 2182 case ISD::SETOEQ: 2183 if (VecVT == MVT::v4f32) 2184 return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP; 2185 else if (VecVT == MVT::v2f64) 2186 return PPC::XVCMPEQDP; 2187 break; 2188 case ISD::SETGT: 2189 case ISD::SETOGT: 2190 if (VecVT == MVT::v4f32) 2191 return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP; 2192 else if (VecVT == MVT::v2f64) 2193 return PPC::XVCMPGTDP; 2194 break; 2195 case ISD::SETGE: 2196 case ISD::SETOGE: 2197 if (VecVT == MVT::v4f32) 2198 return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP; 2199 else if (VecVT == MVT::v2f64) 2200 return PPC::XVCMPGEDP; 2201 break; 2202 default: 2203 break; 2204 } 2205 llvm_unreachable("Invalid floating-point vector compare condition"); 2206 } else { 2207 /* Handle some cases by swapping input operands. */ 2208 switch (CC) { 2209 case ISD::SETGE: CC = ISD::SETLE; Swap = true; break; 2210 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break; 2211 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break; 2212 case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break; 2213 default: break; 2214 } 2215 /* Handle some cases by negating the result. */ 2216 switch (CC) { 2217 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break; 2218 case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break; 2219 case ISD::SETLE: CC = ISD::SETGT; Negate = true; break; 2220 case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break; 2221 default: break; 2222 } 2223 /* We have instructions implementing the remaining cases. */ 2224 switch (CC) { 2225 case ISD::SETEQ: 2226 case ISD::SETUEQ: 2227 if (VecVT == MVT::v16i8) 2228 return PPC::VCMPEQUB; 2229 else if (VecVT == MVT::v8i16) 2230 return PPC::VCMPEQUH; 2231 else if (VecVT == MVT::v4i32) 2232 return PPC::VCMPEQUW; 2233 else if (VecVT == MVT::v2i64) 2234 return PPC::VCMPEQUD; 2235 break; 2236 case ISD::SETGT: 2237 if (VecVT == MVT::v16i8) 2238 return PPC::VCMPGTSB; 2239 else if (VecVT == MVT::v8i16) 2240 return PPC::VCMPGTSH; 2241 else if (VecVT == MVT::v4i32) 2242 return PPC::VCMPGTSW; 2243 else if (VecVT == MVT::v2i64) 2244 return PPC::VCMPGTSD; 2245 break; 2246 case ISD::SETUGT: 2247 if (VecVT == MVT::v16i8) 2248 return PPC::VCMPGTUB; 2249 else if (VecVT == MVT::v8i16) 2250 return PPC::VCMPGTUH; 2251 else if (VecVT == MVT::v4i32) 2252 return PPC::VCMPGTUW; 2253 else if (VecVT == MVT::v2i64) 2254 return PPC::VCMPGTUD; 2255 break; 2256 default: 2257 break; 2258 } 2259 llvm_unreachable("Invalid integer vector compare condition"); 2260 } 2261 } 2262 2263 SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) { 2264 SDLoc dl(N); 2265 unsigned Imm; 2266 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); 2267 EVT PtrVT = 2268 CurDAG->getTargetLoweringInfo().getPointerTy(CurDAG->getDataLayout()); 2269 bool isPPC64 = (PtrVT == MVT::i64); 2270 2271 if (!PPCSubTarget->useCRBits() && 2272 isInt32Immediate(N->getOperand(1), Imm)) { 2273 // We can codegen setcc op, imm very efficiently compared to a brcond. 2274 // Check for those cases here. 2275 // setcc op, 0 2276 if (Imm == 0) { 2277 SDValue Op = N->getOperand(0); 2278 switch (CC) { 2279 default: break; 2280 case ISD::SETEQ: { 2281 Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0); 2282 SDValue Ops[] = { Op, getI32Imm(27, dl), getI32Imm(5, dl), 2283 getI32Imm(31, dl) }; 2284 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 2285 } 2286 case ISD::SETNE: { 2287 if (isPPC64) break; 2288 SDValue AD = 2289 SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 2290 Op, getI32Imm(~0U, dl)), 0); 2291 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, 2292 AD.getValue(1)); 2293 } 2294 case ISD::SETLT: { 2295 SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl), 2296 getI32Imm(31, dl) }; 2297 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 2298 } 2299 case ISD::SETGT: { 2300 SDValue T = 2301 SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0); 2302 T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0); 2303 SDValue Ops[] = { T, getI32Imm(1, dl), getI32Imm(31, dl), 2304 getI32Imm(31, dl) }; 2305 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 2306 } 2307 } 2308 } else if (Imm == ~0U) { // setcc op, -1 2309 SDValue Op = N->getOperand(0); 2310 switch (CC) { 2311 default: break; 2312 case ISD::SETEQ: 2313 if (isPPC64) break; 2314 Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 2315 Op, getI32Imm(1, dl)), 0); 2316 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, 2317 SDValue(CurDAG->getMachineNode(PPC::LI, dl, 2318 MVT::i32, 2319 getI32Imm(0, dl)), 2320 0), Op.getValue(1)); 2321 case ISD::SETNE: { 2322 if (isPPC64) break; 2323 Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0); 2324 SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 2325 Op, getI32Imm(~0U, dl)); 2326 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0), 2327 Op, SDValue(AD, 1)); 2328 } 2329 case ISD::SETLT: { 2330 SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op, 2331 getI32Imm(1, dl)), 0); 2332 SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD, 2333 Op), 0); 2334 SDValue Ops[] = { AN, getI32Imm(1, dl), getI32Imm(31, dl), 2335 getI32Imm(31, dl) }; 2336 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 2337 } 2338 case ISD::SETGT: { 2339 SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl), 2340 getI32Imm(31, dl) }; 2341 Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); 2342 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, 2343 getI32Imm(1, dl)); 2344 } 2345 } 2346 } 2347 } 2348 2349 SDValue LHS = N->getOperand(0); 2350 SDValue RHS = N->getOperand(1); 2351 2352 // Altivec Vector compare instructions do not set any CR register by default and 2353 // vector compare operations return the same type as the operands. 2354 if (LHS.getValueType().isVector()) { 2355 if (PPCSubTarget->hasQPX()) 2356 return nullptr; 2357 2358 EVT VecVT = LHS.getValueType(); 2359 bool Swap, Negate; 2360 unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC, 2361 PPCSubTarget->hasVSX(), Swap, Negate); 2362 if (Swap) 2363 std::swap(LHS, RHS); 2364 2365 EVT ResVT = VecVT.changeVectorElementTypeToInteger(); 2366 if (Negate) { 2367 SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, ResVT, LHS, RHS), 0); 2368 return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR : 2369 PPC::VNOR, 2370 ResVT, VCmp, VCmp); 2371 } 2372 2373 return CurDAG->SelectNodeTo(N, VCmpInst, ResVT, LHS, RHS); 2374 } 2375 2376 if (PPCSubTarget->useCRBits()) 2377 return nullptr; 2378 2379 bool Inv; 2380 unsigned Idx = getCRIdxForSetCC(CC, Inv); 2381 SDValue CCReg = SelectCC(LHS, RHS, CC, dl); 2382 SDValue IntCR; 2383 2384 // Force the ccreg into CR7. 2385 SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32); 2386 2387 SDValue InFlag(nullptr, 0); // Null incoming flag value. 2388 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg, 2389 InFlag).getValue(1); 2390 2391 IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg, 2392 CCReg), 0); 2393 2394 SDValue Ops[] = { IntCR, getI32Imm((32 - (3 - Idx)) & 31, dl), 2395 getI32Imm(31, dl), getI32Imm(31, dl) }; 2396 if (!Inv) 2397 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 2398 2399 // Get the specified bit. 2400 SDValue Tmp = 2401 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); 2402 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1, dl)); 2403 } 2404 2405 SDNode *PPCDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) { 2406 // Transfer memoperands. 2407 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 2408 MemOp[0] = cast<MemSDNode>(N)->getMemOperand(); 2409 cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1); 2410 return Result; 2411 } 2412 2413 2414 // Select - Convert the specified operand from a target-independent to a 2415 // target-specific node if it hasn't already been changed. 2416 SDNode *PPCDAGToDAGISel::Select(SDNode *N) { 2417 SDLoc dl(N); 2418 if (N->isMachineOpcode()) { 2419 N->setNodeId(-1); 2420 return nullptr; // Already selected. 2421 } 2422 2423 // In case any misguided DAG-level optimizations form an ADD with a 2424 // TargetConstant operand, crash here instead of miscompiling (by selecting 2425 // an r+r add instead of some kind of r+i add). 2426 if (N->getOpcode() == ISD::ADD && 2427 N->getOperand(1).getOpcode() == ISD::TargetConstant) 2428 llvm_unreachable("Invalid ADD with TargetConstant operand"); 2429 2430 // Try matching complex bit permutations before doing anything else. 2431 if (SDNode *NN = SelectBitPermutation(N)) 2432 return NN; 2433 2434 switch (N->getOpcode()) { 2435 default: break; 2436 2437 case ISD::Constant: { 2438 if (N->getValueType(0) == MVT::i64) 2439 return SelectInt64(CurDAG, N); 2440 break; 2441 } 2442 2443 case ISD::SETCC: { 2444 SDNode *SN = SelectSETCC(N); 2445 if (SN) 2446 return SN; 2447 break; 2448 } 2449 case PPCISD::GlobalBaseReg: 2450 return getGlobalBaseReg(); 2451 2452 case ISD::FrameIndex: 2453 return getFrameIndex(N, N); 2454 2455 case PPCISD::MFOCRF: { 2456 SDValue InFlag = N->getOperand(1); 2457 return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, 2458 N->getOperand(0), InFlag); 2459 } 2460 2461 case PPCISD::READ_TIME_BASE: { 2462 return CurDAG->getMachineNode(PPC::ReadTB, dl, MVT::i32, MVT::i32, 2463 MVT::Other, N->getOperand(0)); 2464 } 2465 2466 case PPCISD::SRA_ADDZE: { 2467 SDValue N0 = N->getOperand(0); 2468 SDValue ShiftAmt = 2469 CurDAG->getTargetConstant(*cast<ConstantSDNode>(N->getOperand(1))-> 2470 getConstantIntValue(), dl, 2471 N->getValueType(0)); 2472 if (N->getValueType(0) == MVT::i64) { 2473 SDNode *Op = 2474 CurDAG->getMachineNode(PPC::SRADI, dl, MVT::i64, MVT::Glue, 2475 N0, ShiftAmt); 2476 return CurDAG->SelectNodeTo(N, PPC::ADDZE8, MVT::i64, 2477 SDValue(Op, 0), SDValue(Op, 1)); 2478 } else { 2479 assert(N->getValueType(0) == MVT::i32 && 2480 "Expecting i64 or i32 in PPCISD::SRA_ADDZE"); 2481 SDNode *Op = 2482 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue, 2483 N0, ShiftAmt); 2484 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, 2485 SDValue(Op, 0), SDValue(Op, 1)); 2486 } 2487 } 2488 2489 case ISD::LOAD: { 2490 // Handle preincrement loads. 2491 LoadSDNode *LD = cast<LoadSDNode>(N); 2492 EVT LoadedVT = LD->getMemoryVT(); 2493 2494 // Normal loads are handled by code generated from the .td file. 2495 if (LD->getAddressingMode() != ISD::PRE_INC) 2496 break; 2497 2498 SDValue Offset = LD->getOffset(); 2499 if (Offset.getOpcode() == ISD::TargetConstant || 2500 Offset.getOpcode() == ISD::TargetGlobalAddress) { 2501 2502 unsigned Opcode; 2503 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; 2504 if (LD->getValueType(0) != MVT::i64) { 2505 // Handle PPC32 integer and normal FP loads. 2506 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 2507 switch (LoadedVT.getSimpleVT().SimpleTy) { 2508 default: llvm_unreachable("Invalid PPC load type!"); 2509 case MVT::f64: Opcode = PPC::LFDU; break; 2510 case MVT::f32: Opcode = PPC::LFSU; break; 2511 case MVT::i32: Opcode = PPC::LWZU; break; 2512 case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break; 2513 case MVT::i1: 2514 case MVT::i8: Opcode = PPC::LBZU; break; 2515 } 2516 } else { 2517 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!"); 2518 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 2519 switch (LoadedVT.getSimpleVT().SimpleTy) { 2520 default: llvm_unreachable("Invalid PPC load type!"); 2521 case MVT::i64: Opcode = PPC::LDU; break; 2522 case MVT::i32: Opcode = PPC::LWZU8; break; 2523 case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break; 2524 case MVT::i1: 2525 case MVT::i8: Opcode = PPC::LBZU8; break; 2526 } 2527 } 2528 2529 SDValue Chain = LD->getChain(); 2530 SDValue Base = LD->getBasePtr(); 2531 SDValue Ops[] = { Offset, Base, Chain }; 2532 return transferMemOperands( 2533 N, CurDAG->getMachineNode( 2534 Opcode, dl, LD->getValueType(0), 2535 PPCLowering->getPointerTy(CurDAG->getDataLayout()), MVT::Other, 2536 Ops)); 2537 } else { 2538 unsigned Opcode; 2539 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; 2540 if (LD->getValueType(0) != MVT::i64) { 2541 // Handle PPC32 integer and normal FP loads. 2542 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 2543 switch (LoadedVT.getSimpleVT().SimpleTy) { 2544 default: llvm_unreachable("Invalid PPC load type!"); 2545 case MVT::v4f64: Opcode = PPC::QVLFDUX; break; // QPX 2546 case MVT::v4f32: Opcode = PPC::QVLFSUX; break; // QPX 2547 case MVT::f64: Opcode = PPC::LFDUX; break; 2548 case MVT::f32: Opcode = PPC::LFSUX; break; 2549 case MVT::i32: Opcode = PPC::LWZUX; break; 2550 case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break; 2551 case MVT::i1: 2552 case MVT::i8: Opcode = PPC::LBZUX; break; 2553 } 2554 } else { 2555 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!"); 2556 assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) && 2557 "Invalid sext update load"); 2558 switch (LoadedVT.getSimpleVT().SimpleTy) { 2559 default: llvm_unreachable("Invalid PPC load type!"); 2560 case MVT::i64: Opcode = PPC::LDUX; break; 2561 case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break; 2562 case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break; 2563 case MVT::i1: 2564 case MVT::i8: Opcode = PPC::LBZUX8; break; 2565 } 2566 } 2567 2568 SDValue Chain = LD->getChain(); 2569 SDValue Base = LD->getBasePtr(); 2570 SDValue Ops[] = { Base, Offset, Chain }; 2571 return transferMemOperands( 2572 N, CurDAG->getMachineNode( 2573 Opcode, dl, LD->getValueType(0), 2574 PPCLowering->getPointerTy(CurDAG->getDataLayout()), MVT::Other, 2575 Ops)); 2576 } 2577 } 2578 2579 case ISD::AND: { 2580 unsigned Imm, Imm2, SH, MB, ME; 2581 uint64_t Imm64; 2582 2583 // If this is an and of a value rotated between 0 and 31 bits and then and'd 2584 // with a mask, emit rlwinm 2585 if (isInt32Immediate(N->getOperand(1), Imm) && 2586 isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) { 2587 SDValue Val = N->getOperand(0).getOperand(0); 2588 SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl), 2589 getI32Imm(ME, dl) }; 2590 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 2591 } 2592 // If this is just a masked value where the input is not handled above, and 2593 // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm 2594 if (isInt32Immediate(N->getOperand(1), Imm) && 2595 isRunOfOnes(Imm, MB, ME) && 2596 N->getOperand(0).getOpcode() != ISD::ROTL) { 2597 SDValue Val = N->getOperand(0); 2598 SDValue Ops[] = { Val, getI32Imm(0, dl), getI32Imm(MB, dl), 2599 getI32Imm(ME, dl) }; 2600 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 2601 } 2602 // If this is a 64-bit zero-extension mask, emit rldicl. 2603 if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) && 2604 isMask_64(Imm64)) { 2605 SDValue Val = N->getOperand(0); 2606 MB = 64 - countTrailingOnes(Imm64); 2607 SH = 0; 2608 2609 // If the operand is a logical right shift, we can fold it into this 2610 // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb) 2611 // for n <= mb. The right shift is really a left rotate followed by a 2612 // mask, and this mask is a more-restrictive sub-mask of the mask implied 2613 // by the shift. 2614 if (Val.getOpcode() == ISD::SRL && 2615 isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) { 2616 assert(Imm < 64 && "Illegal shift amount"); 2617 Val = Val.getOperand(0); 2618 SH = 64 - Imm; 2619 } 2620 2621 SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl) }; 2622 return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops); 2623 } 2624 // AND X, 0 -> 0, not "rlwinm 32". 2625 if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) { 2626 ReplaceUses(SDValue(N, 0), N->getOperand(1)); 2627 return nullptr; 2628 } 2629 // ISD::OR doesn't get all the bitfield insertion fun. 2630 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) might be a 2631 // bitfield insert. 2632 if (isInt32Immediate(N->getOperand(1), Imm) && 2633 N->getOperand(0).getOpcode() == ISD::OR && 2634 isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) { 2635 // The idea here is to check whether this is equivalent to: 2636 // (c1 & m) | (x & ~m) 2637 // where m is a run-of-ones mask. The logic here is that, for each bit in 2638 // c1 and c2: 2639 // - if both are 1, then the output will be 1. 2640 // - if both are 0, then the output will be 0. 2641 // - if the bit in c1 is 0, and the bit in c2 is 1, then the output will 2642 // come from x. 2643 // - if the bit in c1 is 1, and the bit in c2 is 0, then the output will 2644 // be 0. 2645 // If that last condition is never the case, then we can form m from the 2646 // bits that are the same between c1 and c2. 2647 unsigned MB, ME; 2648 if (isRunOfOnes(~(Imm^Imm2), MB, ME) && !(~Imm & Imm2)) { 2649 SDValue Ops[] = { N->getOperand(0).getOperand(0), 2650 N->getOperand(0).getOperand(1), 2651 getI32Imm(0, dl), getI32Imm(MB, dl), 2652 getI32Imm(ME, dl) }; 2653 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); 2654 } 2655 } 2656 2657 // Other cases are autogenerated. 2658 break; 2659 } 2660 case ISD::OR: { 2661 if (N->getValueType(0) == MVT::i32) 2662 if (SDNode *I = SelectBitfieldInsert(N)) 2663 return I; 2664 2665 short Imm; 2666 if (N->getOperand(0)->getOpcode() == ISD::FrameIndex && 2667 isIntS16Immediate(N->getOperand(1), Imm)) { 2668 APInt LHSKnownZero, LHSKnownOne; 2669 CurDAG->computeKnownBits(N->getOperand(0), LHSKnownZero, LHSKnownOne); 2670 2671 // If this is equivalent to an add, then we can fold it with the 2672 // FrameIndex calculation. 2673 if ((LHSKnownZero.getZExtValue()|~(uint64_t)Imm) == ~0ULL) 2674 return getFrameIndex(N, N->getOperand(0).getNode(), (int)Imm); 2675 } 2676 2677 // Other cases are autogenerated. 2678 break; 2679 } 2680 case ISD::ADD: { 2681 short Imm; 2682 if (N->getOperand(0)->getOpcode() == ISD::FrameIndex && 2683 isIntS16Immediate(N->getOperand(1), Imm)) 2684 return getFrameIndex(N, N->getOperand(0).getNode(), (int)Imm); 2685 2686 break; 2687 } 2688 case ISD::SHL: { 2689 unsigned Imm, SH, MB, ME; 2690 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && 2691 isRotateAndMask(N, Imm, true, SH, MB, ME)) { 2692 SDValue Ops[] = { N->getOperand(0).getOperand(0), 2693 getI32Imm(SH, dl), getI32Imm(MB, dl), 2694 getI32Imm(ME, dl) }; 2695 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 2696 } 2697 2698 // Other cases are autogenerated. 2699 break; 2700 } 2701 case ISD::SRL: { 2702 unsigned Imm, SH, MB, ME; 2703 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && 2704 isRotateAndMask(N, Imm, true, SH, MB, ME)) { 2705 SDValue Ops[] = { N->getOperand(0).getOperand(0), 2706 getI32Imm(SH, dl), getI32Imm(MB, dl), 2707 getI32Imm(ME, dl) }; 2708 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 2709 } 2710 2711 // Other cases are autogenerated. 2712 break; 2713 } 2714 // FIXME: Remove this once the ANDI glue bug is fixed: 2715 case PPCISD::ANDIo_1_EQ_BIT: 2716 case PPCISD::ANDIo_1_GT_BIT: { 2717 if (!ANDIGlueBug) 2718 break; 2719 2720 EVT InVT = N->getOperand(0).getValueType(); 2721 assert((InVT == MVT::i64 || InVT == MVT::i32) && 2722 "Invalid input type for ANDIo_1_EQ_BIT"); 2723 2724 unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo; 2725 SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue, 2726 N->getOperand(0), 2727 CurDAG->getTargetConstant(1, dl, InVT)), 2728 0); 2729 SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32); 2730 SDValue SRIdxVal = 2731 CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ? 2732 PPC::sub_eq : PPC::sub_gt, dl, MVT::i32); 2733 2734 return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1, 2735 CR0Reg, SRIdxVal, 2736 SDValue(AndI.getNode(), 1) /* glue */); 2737 } 2738 case ISD::SELECT_CC: { 2739 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get(); 2740 EVT PtrVT = 2741 CurDAG->getTargetLoweringInfo().getPointerTy(CurDAG->getDataLayout()); 2742 bool isPPC64 = (PtrVT == MVT::i64); 2743 2744 // If this is a select of i1 operands, we'll pattern match it. 2745 if (PPCSubTarget->useCRBits() && 2746 N->getOperand(0).getValueType() == MVT::i1) 2747 break; 2748 2749 // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc 2750 if (!isPPC64) 2751 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1))) 2752 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2))) 2753 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3))) 2754 if (N1C->isNullValue() && N3C->isNullValue() && 2755 N2C->getZExtValue() == 1ULL && CC == ISD::SETNE && 2756 // FIXME: Implement this optzn for PPC64. 2757 N->getValueType(0) == MVT::i32) { 2758 SDNode *Tmp = 2759 CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 2760 N->getOperand(0), getI32Imm(~0U, dl)); 2761 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, 2762 SDValue(Tmp, 0), N->getOperand(0), 2763 SDValue(Tmp, 1)); 2764 } 2765 2766 SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl); 2767 2768 if (N->getValueType(0) == MVT::i1) { 2769 // An i1 select is: (c & t) | (!c & f). 2770 bool Inv; 2771 unsigned Idx = getCRIdxForSetCC(CC, Inv); 2772 2773 unsigned SRI; 2774 switch (Idx) { 2775 default: llvm_unreachable("Invalid CC index"); 2776 case 0: SRI = PPC::sub_lt; break; 2777 case 1: SRI = PPC::sub_gt; break; 2778 case 2: SRI = PPC::sub_eq; break; 2779 case 3: SRI = PPC::sub_un; break; 2780 } 2781 2782 SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg); 2783 2784 SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1, 2785 CCBit, CCBit), 0); 2786 SDValue C = Inv ? NotCCBit : CCBit, 2787 NotC = Inv ? CCBit : NotCCBit; 2788 2789 SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1, 2790 C, N->getOperand(2)), 0); 2791 SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1, 2792 NotC, N->getOperand(3)), 0); 2793 2794 return CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF); 2795 } 2796 2797 unsigned BROpc = getPredicateForSetCC(CC); 2798 2799 unsigned SelectCCOp; 2800 if (N->getValueType(0) == MVT::i32) 2801 SelectCCOp = PPC::SELECT_CC_I4; 2802 else if (N->getValueType(0) == MVT::i64) 2803 SelectCCOp = PPC::SELECT_CC_I8; 2804 else if (N->getValueType(0) == MVT::f32) 2805 if (PPCSubTarget->hasP8Vector()) 2806 SelectCCOp = PPC::SELECT_CC_VSSRC; 2807 else 2808 SelectCCOp = PPC::SELECT_CC_F4; 2809 else if (N->getValueType(0) == MVT::f64) 2810 if (PPCSubTarget->hasVSX()) 2811 SelectCCOp = PPC::SELECT_CC_VSFRC; 2812 else 2813 SelectCCOp = PPC::SELECT_CC_F8; 2814 else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f64) 2815 SelectCCOp = PPC::SELECT_CC_QFRC; 2816 else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f32) 2817 SelectCCOp = PPC::SELECT_CC_QSRC; 2818 else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4i1) 2819 SelectCCOp = PPC::SELECT_CC_QBRC; 2820 else if (N->getValueType(0) == MVT::v2f64 || 2821 N->getValueType(0) == MVT::v2i64) 2822 SelectCCOp = PPC::SELECT_CC_VSRC; 2823 else 2824 SelectCCOp = PPC::SELECT_CC_VRRC; 2825 2826 SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3), 2827 getI32Imm(BROpc, dl) }; 2828 return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops); 2829 } 2830 case ISD::VSELECT: 2831 if (PPCSubTarget->hasVSX()) { 2832 SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) }; 2833 return CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops); 2834 } 2835 2836 break; 2837 case ISD::VECTOR_SHUFFLE: 2838 if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 || 2839 N->getValueType(0) == MVT::v2i64)) { 2840 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 2841 2842 SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1), 2843 Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1); 2844 unsigned DM[2]; 2845 2846 for (int i = 0; i < 2; ++i) 2847 if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2) 2848 DM[i] = 0; 2849 else 2850 DM[i] = 1; 2851 2852 if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 && 2853 Op1.getOpcode() == ISD::SCALAR_TO_VECTOR && 2854 isa<LoadSDNode>(Op1.getOperand(0))) { 2855 LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0)); 2856 SDValue Base, Offset; 2857 2858 if (LD->isUnindexed() && LD->hasOneUse() && Op1.hasOneUse() && 2859 (LD->getMemoryVT() == MVT::f64 || 2860 LD->getMemoryVT() == MVT::i64) && 2861 SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) { 2862 SDValue Chain = LD->getChain(); 2863 SDValue Ops[] = { Base, Offset, Chain }; 2864 return CurDAG->SelectNodeTo(N, PPC::LXVDSX, 2865 N->getValueType(0), Ops); 2866 } 2867 } 2868 2869 // For little endian, we must swap the input operands and adjust 2870 // the mask elements (reverse and invert them). 2871 if (PPCSubTarget->isLittleEndian()) { 2872 std::swap(Op1, Op2); 2873 unsigned tmp = DM[0]; 2874 DM[0] = 1 - DM[1]; 2875 DM[1] = 1 - tmp; 2876 } 2877 2878 SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), dl, 2879 MVT::i32); 2880 SDValue Ops[] = { Op1, Op2, DMV }; 2881 return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops); 2882 } 2883 2884 break; 2885 case PPCISD::BDNZ: 2886 case PPCISD::BDZ: { 2887 bool IsPPC64 = PPCSubTarget->isPPC64(); 2888 SDValue Ops[] = { N->getOperand(1), N->getOperand(0) }; 2889 return CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ ? 2890 (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 2891 (IsPPC64 ? PPC::BDZ8 : PPC::BDZ), 2892 MVT::Other, Ops); 2893 } 2894 case PPCISD::COND_BRANCH: { 2895 // Op #0 is the Chain. 2896 // Op #1 is the PPC::PRED_* number. 2897 // Op #2 is the CR# 2898 // Op #3 is the Dest MBB 2899 // Op #4 is the Flag. 2900 // Prevent PPC::PRED_* from being selected into LI. 2901 unsigned PCC = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 2902 if (EnableBranchHint) 2903 PCC |= getBranchHint(PCC, FuncInfo, N->getOperand(3)); 2904 2905 SDValue Pred = getI32Imm(PCC, dl); 2906 SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3), 2907 N->getOperand(0), N->getOperand(4) }; 2908 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); 2909 } 2910 case ISD::BR_CC: { 2911 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get(); 2912 unsigned PCC = getPredicateForSetCC(CC); 2913 2914 if (N->getOperand(2).getValueType() == MVT::i1) { 2915 unsigned Opc; 2916 bool Swap; 2917 switch (PCC) { 2918 default: llvm_unreachable("Unexpected Boolean-operand predicate"); 2919 case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break; 2920 case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break; 2921 case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break; 2922 case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break; 2923 case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break; 2924 case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break; 2925 } 2926 2927 SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1, 2928 N->getOperand(Swap ? 3 : 2), 2929 N->getOperand(Swap ? 2 : 3)), 0); 2930 return CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other, 2931 BitComp, N->getOperand(4), N->getOperand(0)); 2932 } 2933 2934 if (EnableBranchHint) 2935 PCC |= getBranchHint(PCC, FuncInfo, N->getOperand(4)); 2936 2937 SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl); 2938 SDValue Ops[] = { getI32Imm(PCC, dl), CondCode, 2939 N->getOperand(4), N->getOperand(0) }; 2940 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); 2941 } 2942 case ISD::BRIND: { 2943 // FIXME: Should custom lower this. 2944 SDValue Chain = N->getOperand(0); 2945 SDValue Target = N->getOperand(1); 2946 unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8; 2947 unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8; 2948 Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target, 2949 Chain), 0); 2950 return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain); 2951 } 2952 case PPCISD::TOC_ENTRY: { 2953 assert ((PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI()) && 2954 "Only supported for 64-bit ABI and 32-bit SVR4"); 2955 if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) { 2956 SDValue GA = N->getOperand(0); 2957 return transferMemOperands(N, CurDAG->getMachineNode(PPC::LWZtoc, dl, 2958 MVT::i32, GA, N->getOperand(1))); 2959 } 2960 2961 // For medium and large code model, we generate two instructions as 2962 // described below. Otherwise we allow SelectCodeCommon to handle this, 2963 // selecting one of LDtoc, LDtocJTI, LDtocCPT, and LDtocBA. 2964 CodeModel::Model CModel = TM.getCodeModel(); 2965 if (CModel != CodeModel::Medium && CModel != CodeModel::Large) 2966 break; 2967 2968 // The first source operand is a TargetGlobalAddress or a TargetJumpTable. 2969 // If it must be toc-referenced according to PPCSubTarget, we generate: 2970 // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>)) 2971 // Otherwise we generate: 2972 // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>) 2973 SDValue GA = N->getOperand(0); 2974 SDValue TOCbase = N->getOperand(1); 2975 SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64, 2976 TOCbase, GA); 2977 2978 if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(GA) || 2979 CModel == CodeModel::Large) 2980 return transferMemOperands(N, CurDAG->getMachineNode(PPC::LDtocL, dl, 2981 MVT::i64, GA, SDValue(Tmp, 0))); 2982 2983 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) { 2984 const GlobalValue *GV = G->getGlobal(); 2985 unsigned char GVFlags = PPCSubTarget->classifyGlobalReference(GV); 2986 if (GVFlags & PPCII::MO_NLP_FLAG) { 2987 return transferMemOperands(N, CurDAG->getMachineNode(PPC::LDtocL, dl, 2988 MVT::i64, GA, SDValue(Tmp, 0))); 2989 } 2990 } 2991 2992 return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64, 2993 SDValue(Tmp, 0), GA); 2994 } 2995 case PPCISD::PPC32_PICGOT: { 2996 // Generate a PIC-safe GOT reference. 2997 assert(!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() && 2998 "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4"); 2999 return CurDAG->SelectNodeTo( 3000 N, PPC::PPC32PICGOT, PPCLowering->getPointerTy(CurDAG->getDataLayout()), 3001 MVT::i32); 3002 } 3003 case PPCISD::VADD_SPLAT: { 3004 // This expands into one of three sequences, depending on whether 3005 // the first operand is odd or even, positive or negative. 3006 assert(isa<ConstantSDNode>(N->getOperand(0)) && 3007 isa<ConstantSDNode>(N->getOperand(1)) && 3008 "Invalid operand on VADD_SPLAT!"); 3009 3010 int Elt = N->getConstantOperandVal(0); 3011 int EltSize = N->getConstantOperandVal(1); 3012 unsigned Opc1, Opc2, Opc3; 3013 EVT VT; 3014 3015 if (EltSize == 1) { 3016 Opc1 = PPC::VSPLTISB; 3017 Opc2 = PPC::VADDUBM; 3018 Opc3 = PPC::VSUBUBM; 3019 VT = MVT::v16i8; 3020 } else if (EltSize == 2) { 3021 Opc1 = PPC::VSPLTISH; 3022 Opc2 = PPC::VADDUHM; 3023 Opc3 = PPC::VSUBUHM; 3024 VT = MVT::v8i16; 3025 } else { 3026 assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!"); 3027 Opc1 = PPC::VSPLTISW; 3028 Opc2 = PPC::VADDUWM; 3029 Opc3 = PPC::VSUBUWM; 3030 VT = MVT::v4i32; 3031 } 3032 3033 if ((Elt & 1) == 0) { 3034 // Elt is even, in the range [-32,-18] + [16,30]. 3035 // 3036 // Convert: VADD_SPLAT elt, size 3037 // Into: tmp = VSPLTIS[BHW] elt 3038 // VADDU[BHW]M tmp, tmp 3039 // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4 3040 SDValue EltVal = getI32Imm(Elt >> 1, dl); 3041 SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 3042 SDValue TmpVal = SDValue(Tmp, 0); 3043 return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal); 3044 3045 } else if (Elt > 0) { 3046 // Elt is odd and positive, in the range [17,31]. 3047 // 3048 // Convert: VADD_SPLAT elt, size 3049 // Into: tmp1 = VSPLTIS[BHW] elt-16 3050 // tmp2 = VSPLTIS[BHW] -16 3051 // VSUBU[BHW]M tmp1, tmp2 3052 SDValue EltVal = getI32Imm(Elt - 16, dl); 3053 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 3054 EltVal = getI32Imm(-16, dl); 3055 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 3056 return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0), 3057 SDValue(Tmp2, 0)); 3058 3059 } else { 3060 // Elt is odd and negative, in the range [-31,-17]. 3061 // 3062 // Convert: VADD_SPLAT elt, size 3063 // Into: tmp1 = VSPLTIS[BHW] elt+16 3064 // tmp2 = VSPLTIS[BHW] -16 3065 // VADDU[BHW]M tmp1, tmp2 3066 SDValue EltVal = getI32Imm(Elt + 16, dl); 3067 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 3068 EltVal = getI32Imm(-16, dl); 3069 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 3070 return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0), 3071 SDValue(Tmp2, 0)); 3072 } 3073 } 3074 } 3075 3076 return SelectCode(N); 3077 } 3078 3079 // If the target supports the cmpb instruction, do the idiom recognition here. 3080 // We don't do this as a DAG combine because we don't want to do it as nodes 3081 // are being combined (because we might miss part of the eventual idiom). We 3082 // don't want to do it during instruction selection because we want to reuse 3083 // the logic for lowering the masking operations already part of the 3084 // instruction selector. 3085 SDValue PPCDAGToDAGISel::combineToCMPB(SDNode *N) { 3086 SDLoc dl(N); 3087 3088 assert(N->getOpcode() == ISD::OR && 3089 "Only OR nodes are supported for CMPB"); 3090 3091 SDValue Res; 3092 if (!PPCSubTarget->hasCMPB()) 3093 return Res; 3094 3095 if (N->getValueType(0) != MVT::i32 && 3096 N->getValueType(0) != MVT::i64) 3097 return Res; 3098 3099 EVT VT = N->getValueType(0); 3100 3101 SDValue RHS, LHS; 3102 bool BytesFound[8] = { 0, 0, 0, 0, 0, 0, 0, 0 }; 3103 uint64_t Mask = 0, Alt = 0; 3104 3105 auto IsByteSelectCC = [this](SDValue O, unsigned &b, 3106 uint64_t &Mask, uint64_t &Alt, 3107 SDValue &LHS, SDValue &RHS) { 3108 if (O.getOpcode() != ISD::SELECT_CC) 3109 return false; 3110 ISD::CondCode CC = cast<CondCodeSDNode>(O.getOperand(4))->get(); 3111 3112 if (!isa<ConstantSDNode>(O.getOperand(2)) || 3113 !isa<ConstantSDNode>(O.getOperand(3))) 3114 return false; 3115 3116 uint64_t PM = O.getConstantOperandVal(2); 3117 uint64_t PAlt = O.getConstantOperandVal(3); 3118 for (b = 0; b < 8; ++b) { 3119 uint64_t Mask = UINT64_C(0xFF) << (8*b); 3120 if (PM && (PM & Mask) == PM && (PAlt & Mask) == PAlt) 3121 break; 3122 } 3123 3124 if (b == 8) 3125 return false; 3126 Mask |= PM; 3127 Alt |= PAlt; 3128 3129 if (!isa<ConstantSDNode>(O.getOperand(1)) || 3130 O.getConstantOperandVal(1) != 0) { 3131 SDValue Op0 = O.getOperand(0), Op1 = O.getOperand(1); 3132 if (Op0.getOpcode() == ISD::TRUNCATE) 3133 Op0 = Op0.getOperand(0); 3134 if (Op1.getOpcode() == ISD::TRUNCATE) 3135 Op1 = Op1.getOperand(0); 3136 3137 if (Op0.getOpcode() == ISD::SRL && Op1.getOpcode() == ISD::SRL && 3138 Op0.getOperand(1) == Op1.getOperand(1) && CC == ISD::SETEQ && 3139 isa<ConstantSDNode>(Op0.getOperand(1))) { 3140 3141 unsigned Bits = Op0.getValueType().getSizeInBits(); 3142 if (b != Bits/8-1) 3143 return false; 3144 if (Op0.getConstantOperandVal(1) != Bits-8) 3145 return false; 3146 3147 LHS = Op0.getOperand(0); 3148 RHS = Op1.getOperand(0); 3149 return true; 3150 } 3151 3152 // When we have small integers (i16 to be specific), the form present 3153 // post-legalization uses SETULT in the SELECT_CC for the 3154 // higher-order byte, depending on the fact that the 3155 // even-higher-order bytes are known to all be zero, for example: 3156 // select_cc (xor $lhs, $rhs), 256, 65280, 0, setult 3157 // (so when the second byte is the same, because all higher-order 3158 // bits from bytes 3 and 4 are known to be zero, the result of the 3159 // xor can be at most 255) 3160 if (Op0.getOpcode() == ISD::XOR && CC == ISD::SETULT && 3161 isa<ConstantSDNode>(O.getOperand(1))) { 3162 3163 uint64_t ULim = O.getConstantOperandVal(1); 3164 if (ULim != (UINT64_C(1) << b*8)) 3165 return false; 3166 3167 // Now we need to make sure that the upper bytes are known to be 3168 // zero. 3169 unsigned Bits = Op0.getValueType().getSizeInBits(); 3170 if (!CurDAG->MaskedValueIsZero(Op0, 3171 APInt::getHighBitsSet(Bits, Bits - (b+1)*8))) 3172 return false; 3173 3174 LHS = Op0.getOperand(0); 3175 RHS = Op0.getOperand(1); 3176 return true; 3177 } 3178 3179 return false; 3180 } 3181 3182 if (CC != ISD::SETEQ) 3183 return false; 3184 3185 SDValue Op = O.getOperand(0); 3186 if (Op.getOpcode() == ISD::AND) { 3187 if (!isa<ConstantSDNode>(Op.getOperand(1))) 3188 return false; 3189 if (Op.getConstantOperandVal(1) != (UINT64_C(0xFF) << (8*b))) 3190 return false; 3191 3192 SDValue XOR = Op.getOperand(0); 3193 if (XOR.getOpcode() == ISD::TRUNCATE) 3194 XOR = XOR.getOperand(0); 3195 if (XOR.getOpcode() != ISD::XOR) 3196 return false; 3197 3198 LHS = XOR.getOperand(0); 3199 RHS = XOR.getOperand(1); 3200 return true; 3201 } else if (Op.getOpcode() == ISD::SRL) { 3202 if (!isa<ConstantSDNode>(Op.getOperand(1))) 3203 return false; 3204 unsigned Bits = Op.getValueType().getSizeInBits(); 3205 if (b != Bits/8-1) 3206 return false; 3207 if (Op.getConstantOperandVal(1) != Bits-8) 3208 return false; 3209 3210 SDValue XOR = Op.getOperand(0); 3211 if (XOR.getOpcode() == ISD::TRUNCATE) 3212 XOR = XOR.getOperand(0); 3213 if (XOR.getOpcode() != ISD::XOR) 3214 return false; 3215 3216 LHS = XOR.getOperand(0); 3217 RHS = XOR.getOperand(1); 3218 return true; 3219 } 3220 3221 return false; 3222 }; 3223 3224 SmallVector<SDValue, 8> Queue(1, SDValue(N, 0)); 3225 while (!Queue.empty()) { 3226 SDValue V = Queue.pop_back_val(); 3227 3228 for (const SDValue &O : V.getNode()->ops()) { 3229 unsigned b; 3230 uint64_t M = 0, A = 0; 3231 SDValue OLHS, ORHS; 3232 if (O.getOpcode() == ISD::OR) { 3233 Queue.push_back(O); 3234 } else if (IsByteSelectCC(O, b, M, A, OLHS, ORHS)) { 3235 if (!LHS) { 3236 LHS = OLHS; 3237 RHS = ORHS; 3238 BytesFound[b] = true; 3239 Mask |= M; 3240 Alt |= A; 3241 } else if ((LHS == ORHS && RHS == OLHS) || 3242 (RHS == ORHS && LHS == OLHS)) { 3243 BytesFound[b] = true; 3244 Mask |= M; 3245 Alt |= A; 3246 } else { 3247 return Res; 3248 } 3249 } else { 3250 return Res; 3251 } 3252 } 3253 } 3254 3255 unsigned LastB = 0, BCnt = 0; 3256 for (unsigned i = 0; i < 8; ++i) 3257 if (BytesFound[LastB]) { 3258 ++BCnt; 3259 LastB = i; 3260 } 3261 3262 if (!LastB || BCnt < 2) 3263 return Res; 3264 3265 // Because we'll be zero-extending the output anyway if don't have a specific 3266 // value for each input byte (via the Mask), we can 'anyext' the inputs. 3267 if (LHS.getValueType() != VT) { 3268 LHS = CurDAG->getAnyExtOrTrunc(LHS, dl, VT); 3269 RHS = CurDAG->getAnyExtOrTrunc(RHS, dl, VT); 3270 } 3271 3272 Res = CurDAG->getNode(PPCISD::CMPB, dl, VT, LHS, RHS); 3273 3274 bool NonTrivialMask = ((int64_t) Mask) != INT64_C(-1); 3275 if (NonTrivialMask && !Alt) { 3276 // Res = Mask & CMPB 3277 Res = CurDAG->getNode(ISD::AND, dl, VT, Res, 3278 CurDAG->getConstant(Mask, dl, VT)); 3279 } else if (Alt) { 3280 // Res = (CMPB & Mask) | (~CMPB & Alt) 3281 // Which, as suggested here: 3282 // https://graphics.stanford.edu/~seander/bithacks.html#MaskedMerge 3283 // can be written as: 3284 // Res = Alt ^ ((Alt ^ Mask) & CMPB) 3285 // useful because the (Alt ^ Mask) can be pre-computed. 3286 Res = CurDAG->getNode(ISD::AND, dl, VT, Res, 3287 CurDAG->getConstant(Mask ^ Alt, dl, VT)); 3288 Res = CurDAG->getNode(ISD::XOR, dl, VT, Res, 3289 CurDAG->getConstant(Alt, dl, VT)); 3290 } 3291 3292 return Res; 3293 } 3294 3295 // When CR bit registers are enabled, an extension of an i1 variable to a i32 3296 // or i64 value is lowered in terms of a SELECT_I[48] operation, and thus 3297 // involves constant materialization of a 0 or a 1 or both. If the result of 3298 // the extension is then operated upon by some operator that can be constant 3299 // folded with a constant 0 or 1, and that constant can be materialized using 3300 // only one instruction (like a zero or one), then we should fold in those 3301 // operations with the select. 3302 void PPCDAGToDAGISel::foldBoolExts(SDValue &Res, SDNode *&N) { 3303 if (!PPCSubTarget->useCRBits()) 3304 return; 3305 3306 if (N->getOpcode() != ISD::ZERO_EXTEND && 3307 N->getOpcode() != ISD::SIGN_EXTEND && 3308 N->getOpcode() != ISD::ANY_EXTEND) 3309 return; 3310 3311 if (N->getOperand(0).getValueType() != MVT::i1) 3312 return; 3313 3314 if (!N->hasOneUse()) 3315 return; 3316 3317 SDLoc dl(N); 3318 EVT VT = N->getValueType(0); 3319 SDValue Cond = N->getOperand(0); 3320 SDValue ConstTrue = 3321 CurDAG->getConstant(N->getOpcode() == ISD::SIGN_EXTEND ? -1 : 1, dl, VT); 3322 SDValue ConstFalse = CurDAG->getConstant(0, dl, VT); 3323 3324 do { 3325 SDNode *User = *N->use_begin(); 3326 if (User->getNumOperands() != 2) 3327 break; 3328 3329 auto TryFold = [this, N, User, dl](SDValue Val) { 3330 SDValue UserO0 = User->getOperand(0), UserO1 = User->getOperand(1); 3331 SDValue O0 = UserO0.getNode() == N ? Val : UserO0; 3332 SDValue O1 = UserO1.getNode() == N ? Val : UserO1; 3333 3334 return CurDAG->FoldConstantArithmetic(User->getOpcode(), dl, 3335 User->getValueType(0), 3336 O0.getNode(), O1.getNode()); 3337 }; 3338 3339 SDValue TrueRes = TryFold(ConstTrue); 3340 if (!TrueRes) 3341 break; 3342 SDValue FalseRes = TryFold(ConstFalse); 3343 if (!FalseRes) 3344 break; 3345 3346 // For us to materialize these using one instruction, we must be able to 3347 // represent them as signed 16-bit integers. 3348 uint64_t True = cast<ConstantSDNode>(TrueRes)->getZExtValue(), 3349 False = cast<ConstantSDNode>(FalseRes)->getZExtValue(); 3350 if (!isInt<16>(True) || !isInt<16>(False)) 3351 break; 3352 3353 // We can replace User with a new SELECT node, and try again to see if we 3354 // can fold the select with its user. 3355 Res = CurDAG->getSelect(dl, User->getValueType(0), Cond, TrueRes, FalseRes); 3356 N = User; 3357 ConstTrue = TrueRes; 3358 ConstFalse = FalseRes; 3359 } while (N->hasOneUse()); 3360 } 3361 3362 void PPCDAGToDAGISel::PreprocessISelDAG() { 3363 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode()); 3364 ++Position; 3365 3366 bool MadeChange = false; 3367 while (Position != CurDAG->allnodes_begin()) { 3368 SDNode *N = &*--Position; 3369 if (N->use_empty()) 3370 continue; 3371 3372 SDValue Res; 3373 switch (N->getOpcode()) { 3374 default: break; 3375 case ISD::OR: 3376 Res = combineToCMPB(N); 3377 break; 3378 } 3379 3380 if (!Res) 3381 foldBoolExts(Res, N); 3382 3383 if (Res) { 3384 DEBUG(dbgs() << "PPC DAG preprocessing replacing:\nOld: "); 3385 DEBUG(N->dump(CurDAG)); 3386 DEBUG(dbgs() << "\nNew: "); 3387 DEBUG(Res.getNode()->dump(CurDAG)); 3388 DEBUG(dbgs() << "\n"); 3389 3390 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Res); 3391 MadeChange = true; 3392 } 3393 } 3394 3395 if (MadeChange) 3396 CurDAG->RemoveDeadNodes(); 3397 } 3398 3399 /// PostprocessISelDAG - Perform some late peephole optimizations 3400 /// on the DAG representation. 3401 void PPCDAGToDAGISel::PostprocessISelDAG() { 3402 3403 // Skip peepholes at -O0. 3404 if (TM.getOptLevel() == CodeGenOpt::None) 3405 return; 3406 3407 PeepholePPC64(); 3408 PeepholeCROps(); 3409 PeepholePPC64ZExt(); 3410 } 3411 3412 // Check if all users of this node will become isel where the second operand 3413 // is the constant zero. If this is so, and if we can negate the condition, 3414 // then we can flip the true and false operands. This will allow the zero to 3415 // be folded with the isel so that we don't need to materialize a register 3416 // containing zero. 3417 bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) { 3418 // If we're not using isel, then this does not matter. 3419 if (!PPCSubTarget->hasISEL()) 3420 return false; 3421 3422 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 3423 UI != UE; ++UI) { 3424 SDNode *User = *UI; 3425 if (!User->isMachineOpcode()) 3426 return false; 3427 if (User->getMachineOpcode() != PPC::SELECT_I4 && 3428 User->getMachineOpcode() != PPC::SELECT_I8) 3429 return false; 3430 3431 SDNode *Op2 = User->getOperand(2).getNode(); 3432 if (!Op2->isMachineOpcode()) 3433 return false; 3434 3435 if (Op2->getMachineOpcode() != PPC::LI && 3436 Op2->getMachineOpcode() != PPC::LI8) 3437 return false; 3438 3439 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0)); 3440 if (!C) 3441 return false; 3442 3443 if (!C->isNullValue()) 3444 return false; 3445 } 3446 3447 return true; 3448 } 3449 3450 void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) { 3451 SmallVector<SDNode *, 4> ToReplace; 3452 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 3453 UI != UE; ++UI) { 3454 SDNode *User = *UI; 3455 assert((User->getMachineOpcode() == PPC::SELECT_I4 || 3456 User->getMachineOpcode() == PPC::SELECT_I8) && 3457 "Must have all select users"); 3458 ToReplace.push_back(User); 3459 } 3460 3461 for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(), 3462 UE = ToReplace.end(); UI != UE; ++UI) { 3463 SDNode *User = *UI; 3464 SDNode *ResNode = 3465 CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User), 3466 User->getValueType(0), User->getOperand(0), 3467 User->getOperand(2), 3468 User->getOperand(1)); 3469 3470 DEBUG(dbgs() << "CR Peephole replacing:\nOld: "); 3471 DEBUG(User->dump(CurDAG)); 3472 DEBUG(dbgs() << "\nNew: "); 3473 DEBUG(ResNode->dump(CurDAG)); 3474 DEBUG(dbgs() << "\n"); 3475 3476 ReplaceUses(User, ResNode); 3477 } 3478 } 3479 3480 void PPCDAGToDAGISel::PeepholeCROps() { 3481 bool IsModified; 3482 do { 3483 IsModified = false; 3484 for (SDNode &Node : CurDAG->allnodes()) { 3485 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(&Node); 3486 if (!MachineNode || MachineNode->use_empty()) 3487 continue; 3488 SDNode *ResNode = MachineNode; 3489 3490 bool Op1Set = false, Op1Unset = false, 3491 Op1Not = false, 3492 Op2Set = false, Op2Unset = false, 3493 Op2Not = false; 3494 3495 unsigned Opcode = MachineNode->getMachineOpcode(); 3496 switch (Opcode) { 3497 default: break; 3498 case PPC::CRAND: 3499 case PPC::CRNAND: 3500 case PPC::CROR: 3501 case PPC::CRXOR: 3502 case PPC::CRNOR: 3503 case PPC::CREQV: 3504 case PPC::CRANDC: 3505 case PPC::CRORC: { 3506 SDValue Op = MachineNode->getOperand(1); 3507 if (Op.isMachineOpcode()) { 3508 if (Op.getMachineOpcode() == PPC::CRSET) 3509 Op2Set = true; 3510 else if (Op.getMachineOpcode() == PPC::CRUNSET) 3511 Op2Unset = true; 3512 else if (Op.getMachineOpcode() == PPC::CRNOR && 3513 Op.getOperand(0) == Op.getOperand(1)) 3514 Op2Not = true; 3515 } 3516 } // fallthrough 3517 case PPC::BC: 3518 case PPC::BCn: 3519 case PPC::SELECT_I4: 3520 case PPC::SELECT_I8: 3521 case PPC::SELECT_F4: 3522 case PPC::SELECT_F8: 3523 case PPC::SELECT_QFRC: 3524 case PPC::SELECT_QSRC: 3525 case PPC::SELECT_QBRC: 3526 case PPC::SELECT_VRRC: 3527 case PPC::SELECT_VSFRC: 3528 case PPC::SELECT_VSSRC: 3529 case PPC::SELECT_VSRC: { 3530 SDValue Op = MachineNode->getOperand(0); 3531 if (Op.isMachineOpcode()) { 3532 if (Op.getMachineOpcode() == PPC::CRSET) 3533 Op1Set = true; 3534 else if (Op.getMachineOpcode() == PPC::CRUNSET) 3535 Op1Unset = true; 3536 else if (Op.getMachineOpcode() == PPC::CRNOR && 3537 Op.getOperand(0) == Op.getOperand(1)) 3538 Op1Not = true; 3539 } 3540 } 3541 break; 3542 } 3543 3544 bool SelectSwap = false; 3545 switch (Opcode) { 3546 default: break; 3547 case PPC::CRAND: 3548 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 3549 // x & x = x 3550 ResNode = MachineNode->getOperand(0).getNode(); 3551 else if (Op1Set) 3552 // 1 & y = y 3553 ResNode = MachineNode->getOperand(1).getNode(); 3554 else if (Op2Set) 3555 // x & 1 = x 3556 ResNode = MachineNode->getOperand(0).getNode(); 3557 else if (Op1Unset || Op2Unset) 3558 // x & 0 = 0 & y = 0 3559 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 3560 MVT::i1); 3561 else if (Op1Not) 3562 // ~x & y = andc(y, x) 3563 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 3564 MVT::i1, MachineNode->getOperand(1), 3565 MachineNode->getOperand(0). 3566 getOperand(0)); 3567 else if (Op2Not) 3568 // x & ~y = andc(x, y) 3569 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 3570 MVT::i1, MachineNode->getOperand(0), 3571 MachineNode->getOperand(1). 3572 getOperand(0)); 3573 else if (AllUsersSelectZero(MachineNode)) 3574 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode), 3575 MVT::i1, MachineNode->getOperand(0), 3576 MachineNode->getOperand(1)), 3577 SelectSwap = true; 3578 break; 3579 case PPC::CRNAND: 3580 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 3581 // nand(x, x) -> nor(x, x) 3582 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3583 MVT::i1, MachineNode->getOperand(0), 3584 MachineNode->getOperand(0)); 3585 else if (Op1Set) 3586 // nand(1, y) -> nor(y, y) 3587 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3588 MVT::i1, MachineNode->getOperand(1), 3589 MachineNode->getOperand(1)); 3590 else if (Op2Set) 3591 // nand(x, 1) -> nor(x, x) 3592 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3593 MVT::i1, MachineNode->getOperand(0), 3594 MachineNode->getOperand(0)); 3595 else if (Op1Unset || Op2Unset) 3596 // nand(x, 0) = nand(0, y) = 1 3597 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 3598 MVT::i1); 3599 else if (Op1Not) 3600 // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y) 3601 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 3602 MVT::i1, MachineNode->getOperand(0). 3603 getOperand(0), 3604 MachineNode->getOperand(1)); 3605 else if (Op2Not) 3606 // nand(x, ~y) = ~x | y = orc(y, x) 3607 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 3608 MVT::i1, MachineNode->getOperand(1). 3609 getOperand(0), 3610 MachineNode->getOperand(0)); 3611 else if (AllUsersSelectZero(MachineNode)) 3612 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode), 3613 MVT::i1, MachineNode->getOperand(0), 3614 MachineNode->getOperand(1)), 3615 SelectSwap = true; 3616 break; 3617 case PPC::CROR: 3618 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 3619 // x | x = x 3620 ResNode = MachineNode->getOperand(0).getNode(); 3621 else if (Op1Set || Op2Set) 3622 // x | 1 = 1 | y = 1 3623 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 3624 MVT::i1); 3625 else if (Op1Unset) 3626 // 0 | y = y 3627 ResNode = MachineNode->getOperand(1).getNode(); 3628 else if (Op2Unset) 3629 // x | 0 = x 3630 ResNode = MachineNode->getOperand(0).getNode(); 3631 else if (Op1Not) 3632 // ~x | y = orc(y, x) 3633 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 3634 MVT::i1, MachineNode->getOperand(1), 3635 MachineNode->getOperand(0). 3636 getOperand(0)); 3637 else if (Op2Not) 3638 // x | ~y = orc(x, y) 3639 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 3640 MVT::i1, MachineNode->getOperand(0), 3641 MachineNode->getOperand(1). 3642 getOperand(0)); 3643 else if (AllUsersSelectZero(MachineNode)) 3644 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3645 MVT::i1, MachineNode->getOperand(0), 3646 MachineNode->getOperand(1)), 3647 SelectSwap = true; 3648 break; 3649 case PPC::CRXOR: 3650 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 3651 // xor(x, x) = 0 3652 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 3653 MVT::i1); 3654 else if (Op1Set) 3655 // xor(1, y) -> nor(y, y) 3656 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3657 MVT::i1, MachineNode->getOperand(1), 3658 MachineNode->getOperand(1)); 3659 else if (Op2Set) 3660 // xor(x, 1) -> nor(x, x) 3661 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3662 MVT::i1, MachineNode->getOperand(0), 3663 MachineNode->getOperand(0)); 3664 else if (Op1Unset) 3665 // xor(0, y) = y 3666 ResNode = MachineNode->getOperand(1).getNode(); 3667 else if (Op2Unset) 3668 // xor(x, 0) = x 3669 ResNode = MachineNode->getOperand(0).getNode(); 3670 else if (Op1Not) 3671 // xor(~x, y) = eqv(x, y) 3672 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 3673 MVT::i1, MachineNode->getOperand(0). 3674 getOperand(0), 3675 MachineNode->getOperand(1)); 3676 else if (Op2Not) 3677 // xor(x, ~y) = eqv(x, y) 3678 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 3679 MVT::i1, MachineNode->getOperand(0), 3680 MachineNode->getOperand(1). 3681 getOperand(0)); 3682 else if (AllUsersSelectZero(MachineNode)) 3683 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 3684 MVT::i1, MachineNode->getOperand(0), 3685 MachineNode->getOperand(1)), 3686 SelectSwap = true; 3687 break; 3688 case PPC::CRNOR: 3689 if (Op1Set || Op2Set) 3690 // nor(1, y) -> 0 3691 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 3692 MVT::i1); 3693 else if (Op1Unset) 3694 // nor(0, y) = ~y -> nor(y, y) 3695 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3696 MVT::i1, MachineNode->getOperand(1), 3697 MachineNode->getOperand(1)); 3698 else if (Op2Unset) 3699 // nor(x, 0) = ~x 3700 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3701 MVT::i1, MachineNode->getOperand(0), 3702 MachineNode->getOperand(0)); 3703 else if (Op1Not) 3704 // nor(~x, y) = andc(x, y) 3705 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 3706 MVT::i1, MachineNode->getOperand(0). 3707 getOperand(0), 3708 MachineNode->getOperand(1)); 3709 else if (Op2Not) 3710 // nor(x, ~y) = andc(y, x) 3711 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 3712 MVT::i1, MachineNode->getOperand(1). 3713 getOperand(0), 3714 MachineNode->getOperand(0)); 3715 else if (AllUsersSelectZero(MachineNode)) 3716 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode), 3717 MVT::i1, MachineNode->getOperand(0), 3718 MachineNode->getOperand(1)), 3719 SelectSwap = true; 3720 break; 3721 case PPC::CREQV: 3722 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 3723 // eqv(x, x) = 1 3724 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 3725 MVT::i1); 3726 else if (Op1Set) 3727 // eqv(1, y) = y 3728 ResNode = MachineNode->getOperand(1).getNode(); 3729 else if (Op2Set) 3730 // eqv(x, 1) = x 3731 ResNode = MachineNode->getOperand(0).getNode(); 3732 else if (Op1Unset) 3733 // eqv(0, y) = ~y -> nor(y, y) 3734 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3735 MVT::i1, MachineNode->getOperand(1), 3736 MachineNode->getOperand(1)); 3737 else if (Op2Unset) 3738 // eqv(x, 0) = ~x 3739 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3740 MVT::i1, MachineNode->getOperand(0), 3741 MachineNode->getOperand(0)); 3742 else if (Op1Not) 3743 // eqv(~x, y) = xor(x, y) 3744 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 3745 MVT::i1, MachineNode->getOperand(0). 3746 getOperand(0), 3747 MachineNode->getOperand(1)); 3748 else if (Op2Not) 3749 // eqv(x, ~y) = xor(x, y) 3750 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 3751 MVT::i1, MachineNode->getOperand(0), 3752 MachineNode->getOperand(1). 3753 getOperand(0)); 3754 else if (AllUsersSelectZero(MachineNode)) 3755 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 3756 MVT::i1, MachineNode->getOperand(0), 3757 MachineNode->getOperand(1)), 3758 SelectSwap = true; 3759 break; 3760 case PPC::CRANDC: 3761 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 3762 // andc(x, x) = 0 3763 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 3764 MVT::i1); 3765 else if (Op1Set) 3766 // andc(1, y) = ~y 3767 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3768 MVT::i1, MachineNode->getOperand(1), 3769 MachineNode->getOperand(1)); 3770 else if (Op1Unset || Op2Set) 3771 // andc(0, y) = andc(x, 1) = 0 3772 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 3773 MVT::i1); 3774 else if (Op2Unset) 3775 // andc(x, 0) = x 3776 ResNode = MachineNode->getOperand(0).getNode(); 3777 else if (Op1Not) 3778 // andc(~x, y) = ~(x | y) = nor(x, y) 3779 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3780 MVT::i1, MachineNode->getOperand(0). 3781 getOperand(0), 3782 MachineNode->getOperand(1)); 3783 else if (Op2Not) 3784 // andc(x, ~y) = x & y 3785 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode), 3786 MVT::i1, MachineNode->getOperand(0), 3787 MachineNode->getOperand(1). 3788 getOperand(0)); 3789 else if (AllUsersSelectZero(MachineNode)) 3790 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 3791 MVT::i1, MachineNode->getOperand(1), 3792 MachineNode->getOperand(0)), 3793 SelectSwap = true; 3794 break; 3795 case PPC::CRORC: 3796 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 3797 // orc(x, x) = 1 3798 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 3799 MVT::i1); 3800 else if (Op1Set || Op2Unset) 3801 // orc(1, y) = orc(x, 0) = 1 3802 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 3803 MVT::i1); 3804 else if (Op2Set) 3805 // orc(x, 1) = x 3806 ResNode = MachineNode->getOperand(0).getNode(); 3807 else if (Op1Unset) 3808 // orc(0, y) = ~y 3809 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 3810 MVT::i1, MachineNode->getOperand(1), 3811 MachineNode->getOperand(1)); 3812 else if (Op1Not) 3813 // orc(~x, y) = ~(x & y) = nand(x, y) 3814 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode), 3815 MVT::i1, MachineNode->getOperand(0). 3816 getOperand(0), 3817 MachineNode->getOperand(1)); 3818 else if (Op2Not) 3819 // orc(x, ~y) = x | y 3820 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode), 3821 MVT::i1, MachineNode->getOperand(0), 3822 MachineNode->getOperand(1). 3823 getOperand(0)); 3824 else if (AllUsersSelectZero(MachineNode)) 3825 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 3826 MVT::i1, MachineNode->getOperand(1), 3827 MachineNode->getOperand(0)), 3828 SelectSwap = true; 3829 break; 3830 case PPC::SELECT_I4: 3831 case PPC::SELECT_I8: 3832 case PPC::SELECT_F4: 3833 case PPC::SELECT_F8: 3834 case PPC::SELECT_QFRC: 3835 case PPC::SELECT_QSRC: 3836 case PPC::SELECT_QBRC: 3837 case PPC::SELECT_VRRC: 3838 case PPC::SELECT_VSFRC: 3839 case PPC::SELECT_VSSRC: 3840 case PPC::SELECT_VSRC: 3841 if (Op1Set) 3842 ResNode = MachineNode->getOperand(1).getNode(); 3843 else if (Op1Unset) 3844 ResNode = MachineNode->getOperand(2).getNode(); 3845 else if (Op1Not) 3846 ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(), 3847 SDLoc(MachineNode), 3848 MachineNode->getValueType(0), 3849 MachineNode->getOperand(0). 3850 getOperand(0), 3851 MachineNode->getOperand(2), 3852 MachineNode->getOperand(1)); 3853 break; 3854 case PPC::BC: 3855 case PPC::BCn: 3856 if (Op1Not) 3857 ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn : 3858 PPC::BC, 3859 SDLoc(MachineNode), 3860 MVT::Other, 3861 MachineNode->getOperand(0). 3862 getOperand(0), 3863 MachineNode->getOperand(1), 3864 MachineNode->getOperand(2)); 3865 // FIXME: Handle Op1Set, Op1Unset here too. 3866 break; 3867 } 3868 3869 // If we're inverting this node because it is used only by selects that 3870 // we'd like to swap, then swap the selects before the node replacement. 3871 if (SelectSwap) 3872 SwapAllSelectUsers(MachineNode); 3873 3874 if (ResNode != MachineNode) { 3875 DEBUG(dbgs() << "CR Peephole replacing:\nOld: "); 3876 DEBUG(MachineNode->dump(CurDAG)); 3877 DEBUG(dbgs() << "\nNew: "); 3878 DEBUG(ResNode->dump(CurDAG)); 3879 DEBUG(dbgs() << "\n"); 3880 3881 ReplaceUses(MachineNode, ResNode); 3882 IsModified = true; 3883 } 3884 } 3885 if (IsModified) 3886 CurDAG->RemoveDeadNodes(); 3887 } while (IsModified); 3888 } 3889 3890 // Gather the set of 32-bit operations that are known to have their 3891 // higher-order 32 bits zero, where ToPromote contains all such operations. 3892 static bool PeepholePPC64ZExtGather(SDValue Op32, 3893 SmallPtrSetImpl<SDNode *> &ToPromote) { 3894 if (!Op32.isMachineOpcode()) 3895 return false; 3896 3897 // First, check for the "frontier" instructions (those that will clear the 3898 // higher-order 32 bits. 3899 3900 // For RLWINM and RLWNM, we need to make sure that the mask does not wrap 3901 // around. If it does not, then these instructions will clear the 3902 // higher-order bits. 3903 if ((Op32.getMachineOpcode() == PPC::RLWINM || 3904 Op32.getMachineOpcode() == PPC::RLWNM) && 3905 Op32.getConstantOperandVal(2) <= Op32.getConstantOperandVal(3)) { 3906 ToPromote.insert(Op32.getNode()); 3907 return true; 3908 } 3909 3910 // SLW and SRW always clear the higher-order bits. 3911 if (Op32.getMachineOpcode() == PPC::SLW || 3912 Op32.getMachineOpcode() == PPC::SRW) { 3913 ToPromote.insert(Op32.getNode()); 3914 return true; 3915 } 3916 3917 // For LI and LIS, we need the immediate to be positive (so that it is not 3918 // sign extended). 3919 if (Op32.getMachineOpcode() == PPC::LI || 3920 Op32.getMachineOpcode() == PPC::LIS) { 3921 if (!isUInt<15>(Op32.getConstantOperandVal(0))) 3922 return false; 3923 3924 ToPromote.insert(Op32.getNode()); 3925 return true; 3926 } 3927 3928 // LHBRX and LWBRX always clear the higher-order bits. 3929 if (Op32.getMachineOpcode() == PPC::LHBRX || 3930 Op32.getMachineOpcode() == PPC::LWBRX) { 3931 ToPromote.insert(Op32.getNode()); 3932 return true; 3933 } 3934 3935 // CNTLZW always produces a 64-bit value in [0,32], and so is zero extended. 3936 if (Op32.getMachineOpcode() == PPC::CNTLZW) { 3937 ToPromote.insert(Op32.getNode()); 3938 return true; 3939 } 3940 3941 // Next, check for those instructions we can look through. 3942 3943 // Assuming the mask does not wrap around, then the higher-order bits are 3944 // taken directly from the first operand. 3945 if (Op32.getMachineOpcode() == PPC::RLWIMI && 3946 Op32.getConstantOperandVal(3) <= Op32.getConstantOperandVal(4)) { 3947 SmallPtrSet<SDNode *, 16> ToPromote1; 3948 if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1)) 3949 return false; 3950 3951 ToPromote.insert(Op32.getNode()); 3952 ToPromote.insert(ToPromote1.begin(), ToPromote1.end()); 3953 return true; 3954 } 3955 3956 // For OR, the higher-order bits are zero if that is true for both operands. 3957 // For SELECT_I4, the same is true (but the relevant operand numbers are 3958 // shifted by 1). 3959 if (Op32.getMachineOpcode() == PPC::OR || 3960 Op32.getMachineOpcode() == PPC::SELECT_I4) { 3961 unsigned B = Op32.getMachineOpcode() == PPC::SELECT_I4 ? 1 : 0; 3962 SmallPtrSet<SDNode *, 16> ToPromote1; 3963 if (!PeepholePPC64ZExtGather(Op32.getOperand(B+0), ToPromote1)) 3964 return false; 3965 if (!PeepholePPC64ZExtGather(Op32.getOperand(B+1), ToPromote1)) 3966 return false; 3967 3968 ToPromote.insert(Op32.getNode()); 3969 ToPromote.insert(ToPromote1.begin(), ToPromote1.end()); 3970 return true; 3971 } 3972 3973 // For ORI and ORIS, we need the higher-order bits of the first operand to be 3974 // zero, and also for the constant to be positive (so that it is not sign 3975 // extended). 3976 if (Op32.getMachineOpcode() == PPC::ORI || 3977 Op32.getMachineOpcode() == PPC::ORIS) { 3978 SmallPtrSet<SDNode *, 16> ToPromote1; 3979 if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1)) 3980 return false; 3981 if (!isUInt<15>(Op32.getConstantOperandVal(1))) 3982 return false; 3983 3984 ToPromote.insert(Op32.getNode()); 3985 ToPromote.insert(ToPromote1.begin(), ToPromote1.end()); 3986 return true; 3987 } 3988 3989 // The higher-order bits of AND are zero if that is true for at least one of 3990 // the operands. 3991 if (Op32.getMachineOpcode() == PPC::AND) { 3992 SmallPtrSet<SDNode *, 16> ToPromote1, ToPromote2; 3993 bool Op0OK = 3994 PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1); 3995 bool Op1OK = 3996 PeepholePPC64ZExtGather(Op32.getOperand(1), ToPromote2); 3997 if (!Op0OK && !Op1OK) 3998 return false; 3999 4000 ToPromote.insert(Op32.getNode()); 4001 4002 if (Op0OK) 4003 ToPromote.insert(ToPromote1.begin(), ToPromote1.end()); 4004 4005 if (Op1OK) 4006 ToPromote.insert(ToPromote2.begin(), ToPromote2.end()); 4007 4008 return true; 4009 } 4010 4011 // For ANDI and ANDIS, the higher-order bits are zero if either that is true 4012 // of the first operand, or if the second operand is positive (so that it is 4013 // not sign extended). 4014 if (Op32.getMachineOpcode() == PPC::ANDIo || 4015 Op32.getMachineOpcode() == PPC::ANDISo) { 4016 SmallPtrSet<SDNode *, 16> ToPromote1; 4017 bool Op0OK = 4018 PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1); 4019 bool Op1OK = isUInt<15>(Op32.getConstantOperandVal(1)); 4020 if (!Op0OK && !Op1OK) 4021 return false; 4022 4023 ToPromote.insert(Op32.getNode()); 4024 4025 if (Op0OK) 4026 ToPromote.insert(ToPromote1.begin(), ToPromote1.end()); 4027 4028 return true; 4029 } 4030 4031 return false; 4032 } 4033 4034 void PPCDAGToDAGISel::PeepholePPC64ZExt() { 4035 if (!PPCSubTarget->isPPC64()) 4036 return; 4037 4038 // When we zero-extend from i32 to i64, we use a pattern like this: 4039 // def : Pat<(i64 (zext i32:$in)), 4040 // (RLDICL (INSERT_SUBREG (i64 (IMPLICIT_DEF)), $in, sub_32), 4041 // 0, 32)>; 4042 // There are several 32-bit shift/rotate instructions, however, that will 4043 // clear the higher-order bits of their output, rendering the RLDICL 4044 // unnecessary. When that happens, we remove it here, and redefine the 4045 // relevant 32-bit operation to be a 64-bit operation. 4046 4047 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode()); 4048 ++Position; 4049 4050 bool MadeChange = false; 4051 while (Position != CurDAG->allnodes_begin()) { 4052 SDNode *N = &*--Position; 4053 // Skip dead nodes and any non-machine opcodes. 4054 if (N->use_empty() || !N->isMachineOpcode()) 4055 continue; 4056 4057 if (N->getMachineOpcode() != PPC::RLDICL) 4058 continue; 4059 4060 if (N->getConstantOperandVal(1) != 0 || 4061 N->getConstantOperandVal(2) != 32) 4062 continue; 4063 4064 SDValue ISR = N->getOperand(0); 4065 if (!ISR.isMachineOpcode() || 4066 ISR.getMachineOpcode() != TargetOpcode::INSERT_SUBREG) 4067 continue; 4068 4069 if (!ISR.hasOneUse()) 4070 continue; 4071 4072 if (ISR.getConstantOperandVal(2) != PPC::sub_32) 4073 continue; 4074 4075 SDValue IDef = ISR.getOperand(0); 4076 if (!IDef.isMachineOpcode() || 4077 IDef.getMachineOpcode() != TargetOpcode::IMPLICIT_DEF) 4078 continue; 4079 4080 // We now know that we're looking at a canonical i32 -> i64 zext. See if we 4081 // can get rid of it. 4082 4083 SDValue Op32 = ISR->getOperand(1); 4084 if (!Op32.isMachineOpcode()) 4085 continue; 4086 4087 // There are some 32-bit instructions that always clear the high-order 32 4088 // bits, there are also some instructions (like AND) that we can look 4089 // through. 4090 SmallPtrSet<SDNode *, 16> ToPromote; 4091 if (!PeepholePPC64ZExtGather(Op32, ToPromote)) 4092 continue; 4093 4094 // If the ToPromote set contains nodes that have uses outside of the set 4095 // (except for the original INSERT_SUBREG), then abort the transformation. 4096 bool OutsideUse = false; 4097 for (SDNode *PN : ToPromote) { 4098 for (SDNode *UN : PN->uses()) { 4099 if (!ToPromote.count(UN) && UN != ISR.getNode()) { 4100 OutsideUse = true; 4101 break; 4102 } 4103 } 4104 4105 if (OutsideUse) 4106 break; 4107 } 4108 if (OutsideUse) 4109 continue; 4110 4111 MadeChange = true; 4112 4113 // We now know that this zero extension can be removed by promoting to 4114 // nodes in ToPromote to 64-bit operations, where for operations in the 4115 // frontier of the set, we need to insert INSERT_SUBREGs for their 4116 // operands. 4117 for (SDNode *PN : ToPromote) { 4118 unsigned NewOpcode; 4119 switch (PN->getMachineOpcode()) { 4120 default: 4121 llvm_unreachable("Don't know the 64-bit variant of this instruction"); 4122 case PPC::RLWINM: NewOpcode = PPC::RLWINM8; break; 4123 case PPC::RLWNM: NewOpcode = PPC::RLWNM8; break; 4124 case PPC::SLW: NewOpcode = PPC::SLW8; break; 4125 case PPC::SRW: NewOpcode = PPC::SRW8; break; 4126 case PPC::LI: NewOpcode = PPC::LI8; break; 4127 case PPC::LIS: NewOpcode = PPC::LIS8; break; 4128 case PPC::LHBRX: NewOpcode = PPC::LHBRX8; break; 4129 case PPC::LWBRX: NewOpcode = PPC::LWBRX8; break; 4130 case PPC::CNTLZW: NewOpcode = PPC::CNTLZW8; break; 4131 case PPC::RLWIMI: NewOpcode = PPC::RLWIMI8; break; 4132 case PPC::OR: NewOpcode = PPC::OR8; break; 4133 case PPC::SELECT_I4: NewOpcode = PPC::SELECT_I8; break; 4134 case PPC::ORI: NewOpcode = PPC::ORI8; break; 4135 case PPC::ORIS: NewOpcode = PPC::ORIS8; break; 4136 case PPC::AND: NewOpcode = PPC::AND8; break; 4137 case PPC::ANDIo: NewOpcode = PPC::ANDIo8; break; 4138 case PPC::ANDISo: NewOpcode = PPC::ANDISo8; break; 4139 } 4140 4141 // Note: During the replacement process, the nodes will be in an 4142 // inconsistent state (some instructions will have operands with values 4143 // of the wrong type). Once done, however, everything should be right 4144 // again. 4145 4146 SmallVector<SDValue, 4> Ops; 4147 for (const SDValue &V : PN->ops()) { 4148 if (!ToPromote.count(V.getNode()) && V.getValueType() == MVT::i32 && 4149 !isa<ConstantSDNode>(V)) { 4150 SDValue ReplOpOps[] = { ISR.getOperand(0), V, ISR.getOperand(2) }; 4151 SDNode *ReplOp = 4152 CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, SDLoc(V), 4153 ISR.getNode()->getVTList(), ReplOpOps); 4154 Ops.push_back(SDValue(ReplOp, 0)); 4155 } else { 4156 Ops.push_back(V); 4157 } 4158 } 4159 4160 // Because all to-be-promoted nodes only have users that are other 4161 // promoted nodes (or the original INSERT_SUBREG), we can safely replace 4162 // the i32 result value type with i64. 4163 4164 SmallVector<EVT, 2> NewVTs; 4165 SDVTList VTs = PN->getVTList(); 4166 for (unsigned i = 0, ie = VTs.NumVTs; i != ie; ++i) 4167 if (VTs.VTs[i] == MVT::i32) 4168 NewVTs.push_back(MVT::i64); 4169 else 4170 NewVTs.push_back(VTs.VTs[i]); 4171 4172 DEBUG(dbgs() << "PPC64 ZExt Peephole morphing:\nOld: "); 4173 DEBUG(PN->dump(CurDAG)); 4174 4175 CurDAG->SelectNodeTo(PN, NewOpcode, CurDAG->getVTList(NewVTs), Ops); 4176 4177 DEBUG(dbgs() << "\nNew: "); 4178 DEBUG(PN->dump(CurDAG)); 4179 DEBUG(dbgs() << "\n"); 4180 } 4181 4182 // Now we replace the original zero extend and its associated INSERT_SUBREG 4183 // with the value feeding the INSERT_SUBREG (which has now been promoted to 4184 // return an i64). 4185 4186 DEBUG(dbgs() << "PPC64 ZExt Peephole replacing:\nOld: "); 4187 DEBUG(N->dump(CurDAG)); 4188 DEBUG(dbgs() << "\nNew: "); 4189 DEBUG(Op32.getNode()->dump(CurDAG)); 4190 DEBUG(dbgs() << "\n"); 4191 4192 ReplaceUses(N, Op32.getNode()); 4193 } 4194 4195 if (MadeChange) 4196 CurDAG->RemoveDeadNodes(); 4197 } 4198 4199 void PPCDAGToDAGISel::PeepholePPC64() { 4200 // These optimizations are currently supported only for 64-bit SVR4. 4201 if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64()) 4202 return; 4203 4204 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode()); 4205 ++Position; 4206 4207 while (Position != CurDAG->allnodes_begin()) { 4208 SDNode *N = &*--Position; 4209 // Skip dead nodes and any non-machine opcodes. 4210 if (N->use_empty() || !N->isMachineOpcode()) 4211 continue; 4212 4213 unsigned FirstOp; 4214 unsigned StorageOpcode = N->getMachineOpcode(); 4215 4216 switch (StorageOpcode) { 4217 default: continue; 4218 4219 case PPC::LBZ: 4220 case PPC::LBZ8: 4221 case PPC::LD: 4222 case PPC::LFD: 4223 case PPC::LFS: 4224 case PPC::LHA: 4225 case PPC::LHA8: 4226 case PPC::LHZ: 4227 case PPC::LHZ8: 4228 case PPC::LWA: 4229 case PPC::LWZ: 4230 case PPC::LWZ8: 4231 FirstOp = 0; 4232 break; 4233 4234 case PPC::STB: 4235 case PPC::STB8: 4236 case PPC::STD: 4237 case PPC::STFD: 4238 case PPC::STFS: 4239 case PPC::STH: 4240 case PPC::STH8: 4241 case PPC::STW: 4242 case PPC::STW8: 4243 FirstOp = 1; 4244 break; 4245 } 4246 4247 // If this is a load or store with a zero offset, or within the alignment, 4248 // we may be able to fold an add-immediate into the memory operation. 4249 // The check against alignment is below, as it can't occur until we check 4250 // the arguments to N 4251 if (!isa<ConstantSDNode>(N->getOperand(FirstOp))) 4252 continue; 4253 4254 SDValue Base = N->getOperand(FirstOp + 1); 4255 if (!Base.isMachineOpcode()) 4256 continue; 4257 4258 // On targets with fusion, we don't want this to fire and remove a fusion 4259 // opportunity, unless a) it results in another fusion opportunity or 4260 // b) optimizing for size. 4261 if (PPCSubTarget->hasFusion() && 4262 (!MF->getFunction()->optForSize() && !Base.hasOneUse())) 4263 continue; 4264 4265 unsigned Flags = 0; 4266 bool ReplaceFlags = true; 4267 4268 // When the feeding operation is an add-immediate of some sort, 4269 // determine whether we need to add relocation information to the 4270 // target flags on the immediate operand when we fold it into the 4271 // load instruction. 4272 // 4273 // For something like ADDItocL, the relocation information is 4274 // inferred from the opcode; when we process it in the AsmPrinter, 4275 // we add the necessary relocation there. A load, though, can receive 4276 // relocation from various flavors of ADDIxxx, so we need to carry 4277 // the relocation information in the target flags. 4278 switch (Base.getMachineOpcode()) { 4279 default: continue; 4280 4281 case PPC::ADDI8: 4282 case PPC::ADDI: 4283 // In some cases (such as TLS) the relocation information 4284 // is already in place on the operand, so copying the operand 4285 // is sufficient. 4286 ReplaceFlags = false; 4287 // For these cases, the immediate may not be divisible by 4, in 4288 // which case the fold is illegal for DS-form instructions. (The 4289 // other cases provide aligned addresses and are always safe.) 4290 if ((StorageOpcode == PPC::LWA || 4291 StorageOpcode == PPC::LD || 4292 StorageOpcode == PPC::STD) && 4293 (!isa<ConstantSDNode>(Base.getOperand(1)) || 4294 Base.getConstantOperandVal(1) % 4 != 0)) 4295 continue; 4296 break; 4297 case PPC::ADDIdtprelL: 4298 Flags = PPCII::MO_DTPREL_LO; 4299 break; 4300 case PPC::ADDItlsldL: 4301 Flags = PPCII::MO_TLSLD_LO; 4302 break; 4303 case PPC::ADDItocL: 4304 Flags = PPCII::MO_TOC_LO; 4305 break; 4306 } 4307 4308 SDValue ImmOpnd = Base.getOperand(1); 4309 int MaxDisplacement = 0; 4310 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) { 4311 const GlobalValue *GV = GA->getGlobal(); 4312 MaxDisplacement = GV->getAlignment() - 1; 4313 } 4314 4315 int Offset = N->getConstantOperandVal(FirstOp); 4316 if (Offset < 0 || Offset > MaxDisplacement) 4317 continue; 4318 4319 // We found an opportunity. Reverse the operands from the add 4320 // immediate and substitute them into the load or store. If 4321 // needed, update the target flags for the immediate operand to 4322 // reflect the necessary relocation information. 4323 DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: "); 4324 DEBUG(Base->dump(CurDAG)); 4325 DEBUG(dbgs() << "\nN: "); 4326 DEBUG(N->dump(CurDAG)); 4327 DEBUG(dbgs() << "\n"); 4328 4329 // If the relocation information isn't already present on the 4330 // immediate operand, add it now. 4331 if (ReplaceFlags) { 4332 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) { 4333 SDLoc dl(GA); 4334 const GlobalValue *GV = GA->getGlobal(); 4335 // We can't perform this optimization for data whose alignment 4336 // is insufficient for the instruction encoding. 4337 if (GV->getAlignment() < 4 && 4338 (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD || 4339 StorageOpcode == PPC::LWA || (Offset % 4) != 0)) { 4340 DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n"); 4341 continue; 4342 } 4343 ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, Offset, Flags); 4344 } else if (ConstantPoolSDNode *CP = 4345 dyn_cast<ConstantPoolSDNode>(ImmOpnd)) { 4346 const Constant *C = CP->getConstVal(); 4347 ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64, 4348 CP->getAlignment(), 4349 Offset, Flags); 4350 } 4351 } 4352 4353 if (FirstOp == 1) // Store 4354 (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd, 4355 Base.getOperand(0), N->getOperand(3)); 4356 else // Load 4357 (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0), 4358 N->getOperand(2)); 4359 4360 // The add-immediate may now be dead, in which case remove it. 4361 if (Base.getNode()->use_empty()) 4362 CurDAG->RemoveDeadNode(Base.getNode()); 4363 } 4364 } 4365 4366 4367 /// createPPCISelDag - This pass converts a legalized DAG into a 4368 /// PowerPC-specific DAG, ready for instruction scheduling. 4369 /// 4370 FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) { 4371 return new PPCDAGToDAGISel(TM); 4372 } 4373 4374 static void initializePassOnce(PassRegistry &Registry) { 4375 const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection"; 4376 PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID, 4377 nullptr, false, false); 4378 Registry.registerPass(*PI, true); 4379 } 4380 4381 void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) { 4382 CALL_ONCE_INITIALIZATION(initializePassOnce); 4383 } 4384 4385