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/CodeGen/MachineFunction.h" 20 #include "llvm/CodeGen/MachineInstrBuilder.h" 21 #include "llvm/CodeGen/MachineRegisterInfo.h" 22 #include "llvm/CodeGen/SelectionDAG.h" 23 #include "llvm/CodeGen/SelectionDAGISel.h" 24 #include "llvm/IR/Constants.h" 25 #include "llvm/IR/Function.h" 26 #include "llvm/IR/GlobalAlias.h" 27 #include "llvm/IR/GlobalValue.h" 28 #include "llvm/IR/GlobalVariable.h" 29 #include "llvm/IR/Intrinsics.h" 30 #include "llvm/Support/CommandLine.h" 31 #include "llvm/Support/Debug.h" 32 #include "llvm/Support/ErrorHandling.h" 33 #include "llvm/Support/MathExtras.h" 34 #include "llvm/Support/raw_ostream.h" 35 #include "llvm/Target/TargetOptions.h" 36 using namespace llvm; 37 38 #define DEBUG_TYPE "ppc-codegen" 39 40 // FIXME: Remove this once the bug has been fixed! 41 cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug", 42 cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden); 43 44 namespace llvm { 45 void initializePPCDAGToDAGISelPass(PassRegistry&); 46 } 47 48 namespace { 49 //===--------------------------------------------------------------------===// 50 /// PPCDAGToDAGISel - PPC specific code to select PPC machine 51 /// instructions for SelectionDAG operations. 52 /// 53 class PPCDAGToDAGISel : public SelectionDAGISel { 54 const PPCTargetMachine &TM; 55 const PPCTargetLowering *PPCLowering; 56 const PPCSubtarget *PPCSubTarget; 57 unsigned GlobalBaseReg; 58 public: 59 explicit PPCDAGToDAGISel(PPCTargetMachine &tm) 60 : SelectionDAGISel(tm), TM(tm), 61 PPCLowering(TM.getSubtargetImpl()->getTargetLowering()), 62 PPCSubTarget(TM.getSubtargetImpl()) { 63 initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry()); 64 } 65 66 bool runOnMachineFunction(MachineFunction &MF) override { 67 // Make sure we re-emit a set of the global base reg if necessary 68 GlobalBaseReg = 0; 69 PPCLowering = TM.getSubtargetImpl()->getTargetLowering(); 70 PPCSubTarget = TM.getSubtargetImpl(); 71 SelectionDAGISel::runOnMachineFunction(MF); 72 73 if (!PPCSubTarget->isSVR4ABI()) 74 InsertVRSaveCode(MF); 75 76 return true; 77 } 78 79 void PostprocessISelDAG() override; 80 81 /// getI32Imm - Return a target constant with the specified value, of type 82 /// i32. 83 inline SDValue getI32Imm(unsigned Imm) { 84 return CurDAG->getTargetConstant(Imm, MVT::i32); 85 } 86 87 /// getI64Imm - Return a target constant with the specified value, of type 88 /// i64. 89 inline SDValue getI64Imm(uint64_t Imm) { 90 return CurDAG->getTargetConstant(Imm, MVT::i64); 91 } 92 93 /// getSmallIPtrImm - Return a target constant of pointer type. 94 inline SDValue getSmallIPtrImm(unsigned Imm) { 95 return CurDAG->getTargetConstant(Imm, PPCLowering->getPointerTy()); 96 } 97 98 /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s 99 /// with any number of 0s on either side. The 1s are allowed to wrap from 100 /// LSB to MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 101 /// 0x0F0F0000 is not, since all 1s are not contiguous. 102 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME); 103 104 105 /// isRotateAndMask - Returns true if Mask and Shift can be folded into a 106 /// rotate and mask opcode and mask operation. 107 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask, 108 unsigned &SH, unsigned &MB, unsigned &ME); 109 110 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC 111 /// base register. Return the virtual register that holds this value. 112 SDNode *getGlobalBaseReg(); 113 114 // Select - Convert the specified operand from a target-independent to a 115 // target-specific node if it hasn't already been changed. 116 SDNode *Select(SDNode *N) override; 117 118 SDNode *SelectBitfieldInsert(SDNode *N); 119 120 /// SelectCC - Select a comparison of the specified values with the 121 /// specified condition code, returning the CR# of the expression. 122 SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, SDLoc dl); 123 124 /// SelectAddrImm - Returns true if the address N can be represented by 125 /// a base register plus a signed 16-bit displacement [r+imm]. 126 bool SelectAddrImm(SDValue N, SDValue &Disp, 127 SDValue &Base) { 128 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false); 129 } 130 131 /// SelectAddrImmOffs - Return true if the operand is valid for a preinc 132 /// immediate field. Note that the operand at this point is already the 133 /// result of a prior SelectAddressRegImm call. 134 bool SelectAddrImmOffs(SDValue N, SDValue &Out) const { 135 if (N.getOpcode() == ISD::TargetConstant || 136 N.getOpcode() == ISD::TargetGlobalAddress) { 137 Out = N; 138 return true; 139 } 140 141 return false; 142 } 143 144 /// SelectAddrIdx - Given the specified addressed, check to see if it can be 145 /// represented as an indexed [r+r] operation. Returns false if it can 146 /// be represented by [r+imm], which are preferred. 147 bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) { 148 return PPCLowering->SelectAddressRegReg(N, Base, Index, *CurDAG); 149 } 150 151 /// SelectAddrIdxOnly - Given the specified addressed, force it to be 152 /// represented as an indexed [r+r] operation. 153 bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) { 154 return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, *CurDAG); 155 } 156 157 /// SelectAddrImmX4 - Returns true if the address N can be represented by 158 /// a base register plus a signed 16-bit displacement that is a multiple of 4. 159 /// Suitable for use by STD and friends. 160 bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) { 161 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true); 162 } 163 164 // Select an address into a single register. 165 bool SelectAddr(SDValue N, SDValue &Base) { 166 Base = N; 167 return true; 168 } 169 170 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for 171 /// inline asm expressions. It is always correct to compute the value into 172 /// a register. The case of adding a (possibly relocatable) constant to a 173 /// register can be improved, but it is wrong to substitute Reg+Reg for 174 /// Reg in an asm, because the load or store opcode would have to change. 175 bool SelectInlineAsmMemoryOperand(const SDValue &Op, 176 char ConstraintCode, 177 std::vector<SDValue> &OutOps) override { 178 OutOps.push_back(Op); 179 return false; 180 } 181 182 void InsertVRSaveCode(MachineFunction &MF); 183 184 const char *getPassName() const override { 185 return "PowerPC DAG->DAG Pattern Instruction Selection"; 186 } 187 188 // Include the pieces autogenerated from the target description. 189 #include "PPCGenDAGISel.inc" 190 191 private: 192 SDNode *SelectSETCC(SDNode *N); 193 194 void PeepholePPC64(); 195 void PeepholeCROps(); 196 197 bool AllUsersSelectZero(SDNode *N); 198 void SwapAllSelectUsers(SDNode *N); 199 }; 200 } 201 202 /// InsertVRSaveCode - Once the entire function has been instruction selected, 203 /// all virtual registers are created and all machine instructions are built, 204 /// check to see if we need to save/restore VRSAVE. If so, do it. 205 void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) { 206 // Check to see if this function uses vector registers, which means we have to 207 // save and restore the VRSAVE register and update it with the regs we use. 208 // 209 // In this case, there will be virtual registers of vector type created 210 // by the scheduler. Detect them now. 211 bool HasVectorVReg = false; 212 for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) { 213 unsigned Reg = TargetRegisterInfo::index2VirtReg(i); 214 if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) { 215 HasVectorVReg = true; 216 break; 217 } 218 } 219 if (!HasVectorVReg) return; // nothing to do. 220 221 // If we have a vector register, we want to emit code into the entry and exit 222 // blocks to save and restore the VRSAVE register. We do this here (instead 223 // of marking all vector instructions as clobbering VRSAVE) for two reasons: 224 // 225 // 1. This (trivially) reduces the load on the register allocator, by not 226 // having to represent the live range of the VRSAVE register. 227 // 2. This (more significantly) allows us to create a temporary virtual 228 // register to hold the saved VRSAVE value, allowing this temporary to be 229 // register allocated, instead of forcing it to be spilled to the stack. 230 231 // Create two vregs - one to hold the VRSAVE register that is live-in to the 232 // function and one for the value after having bits or'd into it. 233 unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 234 unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 235 236 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo(); 237 MachineBasicBlock &EntryBB = *Fn.begin(); 238 DebugLoc dl; 239 // Emit the following code into the entry block: 240 // InVRSAVE = MFVRSAVE 241 // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE 242 // MTVRSAVE UpdatedVRSAVE 243 MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point 244 BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE); 245 BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE), 246 UpdatedVRSAVE).addReg(InVRSAVE); 247 BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE); 248 249 // Find all return blocks, outputting a restore in each epilog. 250 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { 251 if (!BB->empty() && BB->back().isReturn()) { 252 IP = BB->end(); --IP; 253 254 // Skip over all terminator instructions, which are part of the return 255 // sequence. 256 MachineBasicBlock::iterator I2 = IP; 257 while (I2 != BB->begin() && (--I2)->isTerminator()) 258 IP = I2; 259 260 // Emit: MTVRSAVE InVRSave 261 BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE); 262 } 263 } 264 } 265 266 267 /// getGlobalBaseReg - Output the instructions required to put the 268 /// base address to use for accessing globals into a register. 269 /// 270 SDNode *PPCDAGToDAGISel::getGlobalBaseReg() { 271 if (!GlobalBaseReg) { 272 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo(); 273 // Insert the set of GlobalBaseReg into the first MBB of the function 274 MachineBasicBlock &FirstMBB = MF->front(); 275 MachineBasicBlock::iterator MBBI = FirstMBB.begin(); 276 DebugLoc dl; 277 278 if (PPCLowering->getPointerTy() == MVT::i32) { 279 if (PPCSubTarget->isTargetELF()) 280 GlobalBaseReg = PPC::R30; 281 else 282 GlobalBaseReg = 283 RegInfo->createVirtualRegister(&PPC::GPRC_NOR0RegClass); 284 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR)); 285 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); 286 if (PPCSubTarget->isTargetELF()) { 287 unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 288 BuildMI(FirstMBB, MBBI, dl, 289 TII.get(PPC::GetGBRO), TempReg).addReg(GlobalBaseReg); 290 BuildMI(FirstMBB, MBBI, dl, 291 TII.get(PPC::UpdateGBR)).addReg(GlobalBaseReg).addReg(TempReg); 292 MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true); 293 } 294 } else { 295 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_NOX0RegClass); 296 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8)); 297 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg); 298 } 299 } 300 return CurDAG->getRegister(GlobalBaseReg, 301 PPCLowering->getPointerTy()).getNode(); 302 } 303 304 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit 305 /// or 64-bit immediate, and if the value can be accurately represented as a 306 /// sign extension from a 16-bit value. If so, this returns true and the 307 /// immediate. 308 static bool isIntS16Immediate(SDNode *N, short &Imm) { 309 if (N->getOpcode() != ISD::Constant) 310 return false; 311 312 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue(); 313 if (N->getValueType(0) == MVT::i32) 314 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue(); 315 else 316 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue(); 317 } 318 319 static bool isIntS16Immediate(SDValue Op, short &Imm) { 320 return isIntS16Immediate(Op.getNode(), Imm); 321 } 322 323 324 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant 325 /// operand. If so Imm will receive the 32-bit value. 326 static bool isInt32Immediate(SDNode *N, unsigned &Imm) { 327 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) { 328 Imm = cast<ConstantSDNode>(N)->getZExtValue(); 329 return true; 330 } 331 return false; 332 } 333 334 /// isInt64Immediate - This method tests to see if the node is a 64-bit constant 335 /// operand. If so Imm will receive the 64-bit value. 336 static bool isInt64Immediate(SDNode *N, uint64_t &Imm) { 337 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) { 338 Imm = cast<ConstantSDNode>(N)->getZExtValue(); 339 return true; 340 } 341 return false; 342 } 343 344 // isInt32Immediate - This method tests to see if a constant operand. 345 // If so Imm will receive the 32 bit value. 346 static bool isInt32Immediate(SDValue N, unsigned &Imm) { 347 return isInt32Immediate(N.getNode(), Imm); 348 } 349 350 351 // isOpcWithIntImmediate - This method tests to see if the node is a specific 352 // opcode and that it has a immediate integer right operand. 353 // If so Imm will receive the 32 bit value. 354 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) { 355 return N->getOpcode() == Opc 356 && isInt32Immediate(N->getOperand(1).getNode(), Imm); 357 } 358 359 bool PPCDAGToDAGISel::isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) { 360 if (!Val) 361 return false; 362 363 if (isShiftedMask_32(Val)) { 364 // look for the first non-zero bit 365 MB = countLeadingZeros(Val); 366 // look for the first zero bit after the run of ones 367 ME = countLeadingZeros((Val - 1) ^ Val); 368 return true; 369 } else { 370 Val = ~Val; // invert mask 371 if (isShiftedMask_32(Val)) { 372 // effectively look for the first zero bit 373 ME = countLeadingZeros(Val) - 1; 374 // effectively look for the first one bit after the run of zeros 375 MB = countLeadingZeros((Val - 1) ^ Val) + 1; 376 return true; 377 } 378 } 379 // no run present 380 return false; 381 } 382 383 bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask, 384 bool isShiftMask, unsigned &SH, 385 unsigned &MB, unsigned &ME) { 386 // Don't even go down this path for i64, since different logic will be 387 // necessary for rldicl/rldicr/rldimi. 388 if (N->getValueType(0) != MVT::i32) 389 return false; 390 391 unsigned Shift = 32; 392 unsigned Indeterminant = ~0; // bit mask marking indeterminant results 393 unsigned Opcode = N->getOpcode(); 394 if (N->getNumOperands() != 2 || 395 !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31)) 396 return false; 397 398 if (Opcode == ISD::SHL) { 399 // apply shift left to mask if it comes first 400 if (isShiftMask) Mask = Mask << Shift; 401 // determine which bits are made indeterminant by shift 402 Indeterminant = ~(0xFFFFFFFFu << Shift); 403 } else if (Opcode == ISD::SRL) { 404 // apply shift right to mask if it comes first 405 if (isShiftMask) Mask = Mask >> Shift; 406 // determine which bits are made indeterminant by shift 407 Indeterminant = ~(0xFFFFFFFFu >> Shift); 408 // adjust for the left rotate 409 Shift = 32 - Shift; 410 } else if (Opcode == ISD::ROTL) { 411 Indeterminant = 0; 412 } else { 413 return false; 414 } 415 416 // if the mask doesn't intersect any Indeterminant bits 417 if (Mask && !(Mask & Indeterminant)) { 418 SH = Shift & 31; 419 // make sure the mask is still a mask (wrap arounds may not be) 420 return isRunOfOnes(Mask, MB, ME); 421 } 422 return false; 423 } 424 425 /// SelectBitfieldInsert - turn an or of two masked values into 426 /// the rotate left word immediate then mask insert (rlwimi) instruction. 427 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) { 428 SDValue Op0 = N->getOperand(0); 429 SDValue Op1 = N->getOperand(1); 430 SDLoc dl(N); 431 432 APInt LKZ, LKO, RKZ, RKO; 433 CurDAG->computeKnownBits(Op0, LKZ, LKO); 434 CurDAG->computeKnownBits(Op1, RKZ, RKO); 435 436 unsigned TargetMask = LKZ.getZExtValue(); 437 unsigned InsertMask = RKZ.getZExtValue(); 438 439 if ((TargetMask | InsertMask) == 0xFFFFFFFF) { 440 unsigned Op0Opc = Op0.getOpcode(); 441 unsigned Op1Opc = Op1.getOpcode(); 442 unsigned Value, SH = 0; 443 TargetMask = ~TargetMask; 444 InsertMask = ~InsertMask; 445 446 // If the LHS has a foldable shift and the RHS does not, then swap it to the 447 // RHS so that we can fold the shift into the insert. 448 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) { 449 if (Op0.getOperand(0).getOpcode() == ISD::SHL || 450 Op0.getOperand(0).getOpcode() == ISD::SRL) { 451 if (Op1.getOperand(0).getOpcode() != ISD::SHL && 452 Op1.getOperand(0).getOpcode() != ISD::SRL) { 453 std::swap(Op0, Op1); 454 std::swap(Op0Opc, Op1Opc); 455 std::swap(TargetMask, InsertMask); 456 } 457 } 458 } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) { 459 if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL && 460 Op1.getOperand(0).getOpcode() != ISD::SRL) { 461 std::swap(Op0, Op1); 462 std::swap(Op0Opc, Op1Opc); 463 std::swap(TargetMask, InsertMask); 464 } 465 } 466 467 unsigned MB, ME; 468 if (isRunOfOnes(InsertMask, MB, ME)) { 469 SDValue Tmp1, Tmp2; 470 471 if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) && 472 isInt32Immediate(Op1.getOperand(1), Value)) { 473 Op1 = Op1.getOperand(0); 474 SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value; 475 } 476 if (Op1Opc == ISD::AND) { 477 // The AND mask might not be a constant, and we need to make sure that 478 // if we're going to fold the masking with the insert, all bits not 479 // know to be zero in the mask are known to be one. 480 APInt MKZ, MKO; 481 CurDAG->computeKnownBits(Op1.getOperand(1), MKZ, MKO); 482 bool CanFoldMask = InsertMask == MKO.getZExtValue(); 483 484 unsigned SHOpc = Op1.getOperand(0).getOpcode(); 485 if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask && 486 isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) { 487 // Note that Value must be in range here (less than 32) because 488 // otherwise there would not be any bits set in InsertMask. 489 Op1 = Op1.getOperand(0).getOperand(0); 490 SH = (SHOpc == ISD::SHL) ? Value : 32 - Value; 491 } 492 } 493 494 SH &= 31; 495 SDValue Ops[] = { Op0, Op1, getI32Imm(SH), getI32Imm(MB), 496 getI32Imm(ME) }; 497 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); 498 } 499 } 500 return nullptr; 501 } 502 503 /// SelectCC - Select a comparison of the specified values with the specified 504 /// condition code, returning the CR# of the expression. 505 SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, 506 ISD::CondCode CC, SDLoc dl) { 507 // Always select the LHS. 508 unsigned Opc; 509 510 if (LHS.getValueType() == MVT::i32) { 511 unsigned Imm; 512 if (CC == ISD::SETEQ || CC == ISD::SETNE) { 513 if (isInt32Immediate(RHS, Imm)) { 514 // SETEQ/SETNE comparison with 16-bit immediate, fold it. 515 if (isUInt<16>(Imm)) 516 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, 517 getI32Imm(Imm & 0xFFFF)), 0); 518 // If this is a 16-bit signed immediate, fold it. 519 if (isInt<16>((int)Imm)) 520 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, 521 getI32Imm(Imm & 0xFFFF)), 0); 522 523 // For non-equality comparisons, the default code would materialize the 524 // constant, then compare against it, like this: 525 // lis r2, 4660 526 // ori r2, r2, 22136 527 // cmpw cr0, r3, r2 528 // Since we are just comparing for equality, we can emit this instead: 529 // xoris r0,r3,0x1234 530 // cmplwi cr0,r0,0x5678 531 // beq cr0,L6 532 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS, 533 getI32Imm(Imm >> 16)), 0); 534 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor, 535 getI32Imm(Imm & 0xFFFF)), 0); 536 } 537 Opc = PPC::CMPLW; 538 } else if (ISD::isUnsignedIntSetCC(CC)) { 539 if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm)) 540 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, 541 getI32Imm(Imm & 0xFFFF)), 0); 542 Opc = PPC::CMPLW; 543 } else { 544 short SImm; 545 if (isIntS16Immediate(RHS, SImm)) 546 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, 547 getI32Imm((int)SImm & 0xFFFF)), 548 0); 549 Opc = PPC::CMPW; 550 } 551 } else if (LHS.getValueType() == MVT::i64) { 552 uint64_t Imm; 553 if (CC == ISD::SETEQ || CC == ISD::SETNE) { 554 if (isInt64Immediate(RHS.getNode(), Imm)) { 555 // SETEQ/SETNE comparison with 16-bit immediate, fold it. 556 if (isUInt<16>(Imm)) 557 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, 558 getI32Imm(Imm & 0xFFFF)), 0); 559 // If this is a 16-bit signed immediate, fold it. 560 if (isInt<16>(Imm)) 561 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, 562 getI32Imm(Imm & 0xFFFF)), 0); 563 564 // For non-equality comparisons, the default code would materialize the 565 // constant, then compare against it, like this: 566 // lis r2, 4660 567 // ori r2, r2, 22136 568 // cmpd cr0, r3, r2 569 // Since we are just comparing for equality, we can emit this instead: 570 // xoris r0,r3,0x1234 571 // cmpldi cr0,r0,0x5678 572 // beq cr0,L6 573 if (isUInt<32>(Imm)) { 574 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS, 575 getI64Imm(Imm >> 16)), 0); 576 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor, 577 getI64Imm(Imm & 0xFFFF)), 0); 578 } 579 } 580 Opc = PPC::CMPLD; 581 } else if (ISD::isUnsignedIntSetCC(CC)) { 582 if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm)) 583 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, 584 getI64Imm(Imm & 0xFFFF)), 0); 585 Opc = PPC::CMPLD; 586 } else { 587 short SImm; 588 if (isIntS16Immediate(RHS, SImm)) 589 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, 590 getI64Imm(SImm & 0xFFFF)), 591 0); 592 Opc = PPC::CMPD; 593 } 594 } else if (LHS.getValueType() == MVT::f32) { 595 Opc = PPC::FCMPUS; 596 } else { 597 assert(LHS.getValueType() == MVT::f64 && "Unknown vt!"); 598 Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD; 599 } 600 return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0); 601 } 602 603 static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) { 604 switch (CC) { 605 case ISD::SETUEQ: 606 case ISD::SETONE: 607 case ISD::SETOLE: 608 case ISD::SETOGE: 609 llvm_unreachable("Should be lowered by legalize!"); 610 default: llvm_unreachable("Unknown condition!"); 611 case ISD::SETOEQ: 612 case ISD::SETEQ: return PPC::PRED_EQ; 613 case ISD::SETUNE: 614 case ISD::SETNE: return PPC::PRED_NE; 615 case ISD::SETOLT: 616 case ISD::SETLT: return PPC::PRED_LT; 617 case ISD::SETULE: 618 case ISD::SETLE: return PPC::PRED_LE; 619 case ISD::SETOGT: 620 case ISD::SETGT: return PPC::PRED_GT; 621 case ISD::SETUGE: 622 case ISD::SETGE: return PPC::PRED_GE; 623 case ISD::SETO: return PPC::PRED_NU; 624 case ISD::SETUO: return PPC::PRED_UN; 625 // These two are invalid for floating point. Assume we have int. 626 case ISD::SETULT: return PPC::PRED_LT; 627 case ISD::SETUGT: return PPC::PRED_GT; 628 } 629 } 630 631 /// getCRIdxForSetCC - Return the index of the condition register field 632 /// associated with the SetCC condition, and whether or not the field is 633 /// treated as inverted. That is, lt = 0; ge = 0 inverted. 634 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) { 635 Invert = false; 636 switch (CC) { 637 default: llvm_unreachable("Unknown condition!"); 638 case ISD::SETOLT: 639 case ISD::SETLT: return 0; // Bit #0 = SETOLT 640 case ISD::SETOGT: 641 case ISD::SETGT: return 1; // Bit #1 = SETOGT 642 case ISD::SETOEQ: 643 case ISD::SETEQ: return 2; // Bit #2 = SETOEQ 644 case ISD::SETUO: return 3; // Bit #3 = SETUO 645 case ISD::SETUGE: 646 case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE 647 case ISD::SETULE: 648 case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE 649 case ISD::SETUNE: 650 case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE 651 case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO 652 case ISD::SETUEQ: 653 case ISD::SETOGE: 654 case ISD::SETOLE: 655 case ISD::SETONE: 656 llvm_unreachable("Invalid branch code: should be expanded by legalize"); 657 // These are invalid for floating point. Assume integer. 658 case ISD::SETULT: return 0; 659 case ISD::SETUGT: return 1; 660 } 661 } 662 663 // getVCmpInst: return the vector compare instruction for the specified 664 // vector type and condition code. Since this is for altivec specific code, 665 // only support the altivec types (v16i8, v8i16, v4i32, and v4f32). 666 static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC, 667 bool HasVSX, bool &Swap, bool &Negate) { 668 Swap = false; 669 Negate = false; 670 671 if (VecVT.isFloatingPoint()) { 672 /* Handle some cases by swapping input operands. */ 673 switch (CC) { 674 case ISD::SETLE: CC = ISD::SETGE; Swap = true; break; 675 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break; 676 case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break; 677 case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break; 678 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break; 679 case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break; 680 default: break; 681 } 682 /* Handle some cases by negating the result. */ 683 switch (CC) { 684 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break; 685 case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break; 686 case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break; 687 case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break; 688 default: break; 689 } 690 /* We have instructions implementing the remaining cases. */ 691 switch (CC) { 692 case ISD::SETEQ: 693 case ISD::SETOEQ: 694 if (VecVT == MVT::v4f32) 695 return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP; 696 else if (VecVT == MVT::v2f64) 697 return PPC::XVCMPEQDP; 698 break; 699 case ISD::SETGT: 700 case ISD::SETOGT: 701 if (VecVT == MVT::v4f32) 702 return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP; 703 else if (VecVT == MVT::v2f64) 704 return PPC::XVCMPGTDP; 705 break; 706 case ISD::SETGE: 707 case ISD::SETOGE: 708 if (VecVT == MVT::v4f32) 709 return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP; 710 else if (VecVT == MVT::v2f64) 711 return PPC::XVCMPGEDP; 712 break; 713 default: 714 break; 715 } 716 llvm_unreachable("Invalid floating-point vector compare condition"); 717 } else { 718 /* Handle some cases by swapping input operands. */ 719 switch (CC) { 720 case ISD::SETGE: CC = ISD::SETLE; Swap = true; break; 721 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break; 722 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break; 723 case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break; 724 default: break; 725 } 726 /* Handle some cases by negating the result. */ 727 switch (CC) { 728 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break; 729 case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break; 730 case ISD::SETLE: CC = ISD::SETGT; Negate = true; break; 731 case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break; 732 default: break; 733 } 734 /* We have instructions implementing the remaining cases. */ 735 switch (CC) { 736 case ISD::SETEQ: 737 case ISD::SETUEQ: 738 if (VecVT == MVT::v16i8) 739 return PPC::VCMPEQUB; 740 else if (VecVT == MVT::v8i16) 741 return PPC::VCMPEQUH; 742 else if (VecVT == MVT::v4i32) 743 return PPC::VCMPEQUW; 744 break; 745 case ISD::SETGT: 746 if (VecVT == MVT::v16i8) 747 return PPC::VCMPGTSB; 748 else if (VecVT == MVT::v8i16) 749 return PPC::VCMPGTSH; 750 else if (VecVT == MVT::v4i32) 751 return PPC::VCMPGTSW; 752 break; 753 case ISD::SETUGT: 754 if (VecVT == MVT::v16i8) 755 return PPC::VCMPGTUB; 756 else if (VecVT == MVT::v8i16) 757 return PPC::VCMPGTUH; 758 else if (VecVT == MVT::v4i32) 759 return PPC::VCMPGTUW; 760 break; 761 default: 762 break; 763 } 764 llvm_unreachable("Invalid integer vector compare condition"); 765 } 766 } 767 768 SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) { 769 SDLoc dl(N); 770 unsigned Imm; 771 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); 772 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy(); 773 bool isPPC64 = (PtrVT == MVT::i64); 774 775 if (!PPCSubTarget->useCRBits() && 776 isInt32Immediate(N->getOperand(1), Imm)) { 777 // We can codegen setcc op, imm very efficiently compared to a brcond. 778 // Check for those cases here. 779 // setcc op, 0 780 if (Imm == 0) { 781 SDValue Op = N->getOperand(0); 782 switch (CC) { 783 default: break; 784 case ISD::SETEQ: { 785 Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0); 786 SDValue Ops[] = { Op, getI32Imm(27), getI32Imm(5), getI32Imm(31) }; 787 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 788 } 789 case ISD::SETNE: { 790 if (isPPC64) break; 791 SDValue AD = 792 SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 793 Op, getI32Imm(~0U)), 0); 794 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, 795 AD.getValue(1)); 796 } 797 case ISD::SETLT: { 798 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 799 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 800 } 801 case ISD::SETGT: { 802 SDValue T = 803 SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0); 804 T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0); 805 SDValue Ops[] = { T, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 806 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 807 } 808 } 809 } else if (Imm == ~0U) { // setcc op, -1 810 SDValue Op = N->getOperand(0); 811 switch (CC) { 812 default: break; 813 case ISD::SETEQ: 814 if (isPPC64) break; 815 Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 816 Op, getI32Imm(1)), 0); 817 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, 818 SDValue(CurDAG->getMachineNode(PPC::LI, dl, 819 MVT::i32, 820 getI32Imm(0)), 0), 821 Op.getValue(1)); 822 case ISD::SETNE: { 823 if (isPPC64) break; 824 Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0); 825 SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 826 Op, getI32Imm(~0U)); 827 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0), 828 Op, SDValue(AD, 1)); 829 } 830 case ISD::SETLT: { 831 SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op, 832 getI32Imm(1)), 0); 833 SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD, 834 Op), 0); 835 SDValue Ops[] = { AN, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 836 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 837 } 838 case ISD::SETGT: { 839 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 840 Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 841 0); 842 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, 843 getI32Imm(1)); 844 } 845 } 846 } 847 } 848 849 SDValue LHS = N->getOperand(0); 850 SDValue RHS = N->getOperand(1); 851 852 // Altivec Vector compare instructions do not set any CR register by default and 853 // vector compare operations return the same type as the operands. 854 if (LHS.getValueType().isVector()) { 855 EVT VecVT = LHS.getValueType(); 856 bool Swap, Negate; 857 unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC, 858 PPCSubTarget->hasVSX(), Swap, Negate); 859 if (Swap) 860 std::swap(LHS, RHS); 861 862 if (Negate) { 863 SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0); 864 return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR : 865 PPC::VNOR, 866 VecVT, VCmp, VCmp); 867 } 868 869 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS); 870 } 871 872 if (PPCSubTarget->useCRBits()) 873 return nullptr; 874 875 bool Inv; 876 unsigned Idx = getCRIdxForSetCC(CC, Inv); 877 SDValue CCReg = SelectCC(LHS, RHS, CC, dl); 878 SDValue IntCR; 879 880 // Force the ccreg into CR7. 881 SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32); 882 883 SDValue InFlag(nullptr, 0); // Null incoming flag value. 884 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg, 885 InFlag).getValue(1); 886 887 IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg, 888 CCReg), 0); 889 890 SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31), 891 getI32Imm(31), getI32Imm(31) }; 892 if (!Inv) 893 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 894 895 // Get the specified bit. 896 SDValue Tmp = 897 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); 898 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1)); 899 } 900 901 902 // Select - Convert the specified operand from a target-independent to a 903 // target-specific node if it hasn't already been changed. 904 SDNode *PPCDAGToDAGISel::Select(SDNode *N) { 905 SDLoc dl(N); 906 if (N->isMachineOpcode()) { 907 N->setNodeId(-1); 908 return nullptr; // Already selected. 909 } 910 911 // In case any misguided DAG-level optimizations form an ADD with a 912 // TargetConstant operand, crash here instead of miscompiling (by selecting 913 // an r+r add instead of some kind of r+i add). 914 if (N->getOpcode() == ISD::ADD && 915 N->getOperand(1).getOpcode() == ISD::TargetConstant) 916 llvm_unreachable("Invalid ADD with TargetConstant operand"); 917 918 switch (N->getOpcode()) { 919 default: break; 920 921 case ISD::Constant: { 922 if (N->getValueType(0) == MVT::i64) { 923 // Get 64 bit value. 924 int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue(); 925 // Assume no remaining bits. 926 unsigned Remainder = 0; 927 // Assume no shift required. 928 unsigned Shift = 0; 929 930 // If it can't be represented as a 32 bit value. 931 if (!isInt<32>(Imm)) { 932 Shift = countTrailingZeros<uint64_t>(Imm); 933 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift; 934 935 // If the shifted value fits 32 bits. 936 if (isInt<32>(ImmSh)) { 937 // Go with the shifted value. 938 Imm = ImmSh; 939 } else { 940 // Still stuck with a 64 bit value. 941 Remainder = Imm; 942 Shift = 32; 943 Imm >>= 32; 944 } 945 } 946 947 // Intermediate operand. 948 SDNode *Result; 949 950 // Handle first 32 bits. 951 unsigned Lo = Imm & 0xFFFF; 952 unsigned Hi = (Imm >> 16) & 0xFFFF; 953 954 // Simple value. 955 if (isInt<16>(Imm)) { 956 // Just the Lo bits. 957 Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo)); 958 } else if (Lo) { 959 // Handle the Hi bits. 960 unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8; 961 Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi)); 962 // And Lo bits. 963 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, 964 SDValue(Result, 0), getI32Imm(Lo)); 965 } else { 966 // Just the Hi bits. 967 Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi)); 968 } 969 970 // If no shift, we're done. 971 if (!Shift) return Result; 972 973 // Shift for next step if the upper 32-bits were not zero. 974 if (Imm) { 975 Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, 976 SDValue(Result, 0), 977 getI32Imm(Shift), 978 getI32Imm(63 - Shift)); 979 } 980 981 // Add in the last bits as required. 982 if ((Hi = (Remainder >> 16) & 0xFFFF)) { 983 Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64, 984 SDValue(Result, 0), getI32Imm(Hi)); 985 } 986 if ((Lo = Remainder & 0xFFFF)) { 987 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, 988 SDValue(Result, 0), getI32Imm(Lo)); 989 } 990 991 return Result; 992 } 993 break; 994 } 995 996 case ISD::SETCC: { 997 SDNode *SN = SelectSETCC(N); 998 if (SN) 999 return SN; 1000 break; 1001 } 1002 case PPCISD::GlobalBaseReg: 1003 return getGlobalBaseReg(); 1004 1005 case ISD::FrameIndex: { 1006 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 1007 SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0)); 1008 unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8; 1009 if (N->hasOneUse()) 1010 return CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), TFI, 1011 getSmallIPtrImm(0)); 1012 return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI, 1013 getSmallIPtrImm(0)); 1014 } 1015 1016 case PPCISD::MFOCRF: { 1017 SDValue InFlag = N->getOperand(1); 1018 return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, 1019 N->getOperand(0), InFlag); 1020 } 1021 1022 case ISD::SDIV: { 1023 // FIXME: since this depends on the setting of the carry flag from the srawi 1024 // we should really be making notes about that for the scheduler. 1025 // FIXME: It sure would be nice if we could cheaply recognize the 1026 // srl/add/sra pattern the dag combiner will generate for this as 1027 // sra/addze rather than having to handle sdiv ourselves. oh well. 1028 unsigned Imm; 1029 if (isInt32Immediate(N->getOperand(1), Imm)) { 1030 SDValue N0 = N->getOperand(0); 1031 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) { 1032 SDNode *Op = 1033 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue, 1034 N0, getI32Imm(Log2_32(Imm))); 1035 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, 1036 SDValue(Op, 0), SDValue(Op, 1)); 1037 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) { 1038 SDNode *Op = 1039 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue, 1040 N0, getI32Imm(Log2_32(-Imm))); 1041 SDValue PT = 1042 SDValue(CurDAG->getMachineNode(PPC::ADDZE, dl, MVT::i32, 1043 SDValue(Op, 0), SDValue(Op, 1)), 1044 0); 1045 return CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT); 1046 } 1047 } 1048 1049 // Other cases are autogenerated. 1050 break; 1051 } 1052 1053 case ISD::LOAD: { 1054 // Handle preincrement loads. 1055 LoadSDNode *LD = cast<LoadSDNode>(N); 1056 EVT LoadedVT = LD->getMemoryVT(); 1057 1058 // Normal loads are handled by code generated from the .td file. 1059 if (LD->getAddressingMode() != ISD::PRE_INC) 1060 break; 1061 1062 SDValue Offset = LD->getOffset(); 1063 if (Offset.getOpcode() == ISD::TargetConstant || 1064 Offset.getOpcode() == ISD::TargetGlobalAddress) { 1065 1066 unsigned Opcode; 1067 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; 1068 if (LD->getValueType(0) != MVT::i64) { 1069 // Handle PPC32 integer and normal FP loads. 1070 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 1071 switch (LoadedVT.getSimpleVT().SimpleTy) { 1072 default: llvm_unreachable("Invalid PPC load type!"); 1073 case MVT::f64: Opcode = PPC::LFDU; break; 1074 case MVT::f32: Opcode = PPC::LFSU; break; 1075 case MVT::i32: Opcode = PPC::LWZU; break; 1076 case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break; 1077 case MVT::i1: 1078 case MVT::i8: Opcode = PPC::LBZU; break; 1079 } 1080 } else { 1081 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!"); 1082 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 1083 switch (LoadedVT.getSimpleVT().SimpleTy) { 1084 default: llvm_unreachable("Invalid PPC load type!"); 1085 case MVT::i64: Opcode = PPC::LDU; break; 1086 case MVT::i32: Opcode = PPC::LWZU8; break; 1087 case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break; 1088 case MVT::i1: 1089 case MVT::i8: Opcode = PPC::LBZU8; break; 1090 } 1091 } 1092 1093 SDValue Chain = LD->getChain(); 1094 SDValue Base = LD->getBasePtr(); 1095 SDValue Ops[] = { Offset, Base, Chain }; 1096 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0), 1097 PPCLowering->getPointerTy(), 1098 MVT::Other, Ops); 1099 } else { 1100 unsigned Opcode; 1101 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; 1102 if (LD->getValueType(0) != MVT::i64) { 1103 // Handle PPC32 integer and normal FP loads. 1104 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 1105 switch (LoadedVT.getSimpleVT().SimpleTy) { 1106 default: llvm_unreachable("Invalid PPC load type!"); 1107 case MVT::f64: Opcode = PPC::LFDUX; break; 1108 case MVT::f32: Opcode = PPC::LFSUX; break; 1109 case MVT::i32: Opcode = PPC::LWZUX; break; 1110 case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break; 1111 case MVT::i1: 1112 case MVT::i8: Opcode = PPC::LBZUX; break; 1113 } 1114 } else { 1115 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!"); 1116 assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) && 1117 "Invalid sext update load"); 1118 switch (LoadedVT.getSimpleVT().SimpleTy) { 1119 default: llvm_unreachable("Invalid PPC load type!"); 1120 case MVT::i64: Opcode = PPC::LDUX; break; 1121 case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break; 1122 case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break; 1123 case MVT::i1: 1124 case MVT::i8: Opcode = PPC::LBZUX8; break; 1125 } 1126 } 1127 1128 SDValue Chain = LD->getChain(); 1129 SDValue Base = LD->getBasePtr(); 1130 SDValue Ops[] = { Base, Offset, Chain }; 1131 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0), 1132 PPCLowering->getPointerTy(), 1133 MVT::Other, Ops); 1134 } 1135 } 1136 1137 case ISD::AND: { 1138 unsigned Imm, Imm2, SH, MB, ME; 1139 uint64_t Imm64; 1140 1141 // If this is an and of a value rotated between 0 and 31 bits and then and'd 1142 // with a mask, emit rlwinm 1143 if (isInt32Immediate(N->getOperand(1), Imm) && 1144 isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) { 1145 SDValue Val = N->getOperand(0).getOperand(0); 1146 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; 1147 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 1148 } 1149 // If this is just a masked value where the input is not handled above, and 1150 // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm 1151 if (isInt32Immediate(N->getOperand(1), Imm) && 1152 isRunOfOnes(Imm, MB, ME) && 1153 N->getOperand(0).getOpcode() != ISD::ROTL) { 1154 SDValue Val = N->getOperand(0); 1155 SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB), getI32Imm(ME) }; 1156 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 1157 } 1158 // If this is a 64-bit zero-extension mask, emit rldicl. 1159 if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) && 1160 isMask_64(Imm64)) { 1161 SDValue Val = N->getOperand(0); 1162 MB = 64 - CountTrailingOnes_64(Imm64); 1163 SH = 0; 1164 1165 // If the operand is a logical right shift, we can fold it into this 1166 // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb) 1167 // for n <= mb. The right shift is really a left rotate followed by a 1168 // mask, and this mask is a more-restrictive sub-mask of the mask implied 1169 // by the shift. 1170 if (Val.getOpcode() == ISD::SRL && 1171 isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) { 1172 assert(Imm < 64 && "Illegal shift amount"); 1173 Val = Val.getOperand(0); 1174 SH = 64 - Imm; 1175 } 1176 1177 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB) }; 1178 return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops); 1179 } 1180 // AND X, 0 -> 0, not "rlwinm 32". 1181 if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) { 1182 ReplaceUses(SDValue(N, 0), N->getOperand(1)); 1183 return nullptr; 1184 } 1185 // ISD::OR doesn't get all the bitfield insertion fun. 1186 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert 1187 if (isInt32Immediate(N->getOperand(1), Imm) && 1188 N->getOperand(0).getOpcode() == ISD::OR && 1189 isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) { 1190 unsigned MB, ME; 1191 Imm = ~(Imm^Imm2); 1192 if (isRunOfOnes(Imm, MB, ME)) { 1193 SDValue Ops[] = { N->getOperand(0).getOperand(0), 1194 N->getOperand(0).getOperand(1), 1195 getI32Imm(0), getI32Imm(MB),getI32Imm(ME) }; 1196 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); 1197 } 1198 } 1199 1200 // Other cases are autogenerated. 1201 break; 1202 } 1203 case ISD::OR: 1204 if (N->getValueType(0) == MVT::i32) 1205 if (SDNode *I = SelectBitfieldInsert(N)) 1206 return I; 1207 1208 // Other cases are autogenerated. 1209 break; 1210 case ISD::SHL: { 1211 unsigned Imm, SH, MB, ME; 1212 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && 1213 isRotateAndMask(N, Imm, true, SH, MB, ME)) { 1214 SDValue Ops[] = { N->getOperand(0).getOperand(0), 1215 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; 1216 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 1217 } 1218 1219 // Other cases are autogenerated. 1220 break; 1221 } 1222 case ISD::SRL: { 1223 unsigned Imm, SH, MB, ME; 1224 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && 1225 isRotateAndMask(N, Imm, true, SH, MB, ME)) { 1226 SDValue Ops[] = { N->getOperand(0).getOperand(0), 1227 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; 1228 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 1229 } 1230 1231 // Other cases are autogenerated. 1232 break; 1233 } 1234 // FIXME: Remove this once the ANDI glue bug is fixed: 1235 case PPCISD::ANDIo_1_EQ_BIT: 1236 case PPCISD::ANDIo_1_GT_BIT: { 1237 if (!ANDIGlueBug) 1238 break; 1239 1240 EVT InVT = N->getOperand(0).getValueType(); 1241 assert((InVT == MVT::i64 || InVT == MVT::i32) && 1242 "Invalid input type for ANDIo_1_EQ_BIT"); 1243 1244 unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo; 1245 SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue, 1246 N->getOperand(0), 1247 CurDAG->getTargetConstant(1, InVT)), 0); 1248 SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32); 1249 SDValue SRIdxVal = 1250 CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ? 1251 PPC::sub_eq : PPC::sub_gt, MVT::i32); 1252 1253 return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1, 1254 CR0Reg, SRIdxVal, 1255 SDValue(AndI.getNode(), 1) /* glue */); 1256 } 1257 case ISD::SELECT_CC: { 1258 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get(); 1259 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy(); 1260 bool isPPC64 = (PtrVT == MVT::i64); 1261 1262 // If this is a select of i1 operands, we'll pattern match it. 1263 if (PPCSubTarget->useCRBits() && 1264 N->getOperand(0).getValueType() == MVT::i1) 1265 break; 1266 1267 // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc 1268 if (!isPPC64) 1269 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1))) 1270 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2))) 1271 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3))) 1272 if (N1C->isNullValue() && N3C->isNullValue() && 1273 N2C->getZExtValue() == 1ULL && CC == ISD::SETNE && 1274 // FIXME: Implement this optzn for PPC64. 1275 N->getValueType(0) == MVT::i32) { 1276 SDNode *Tmp = 1277 CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 1278 N->getOperand(0), getI32Imm(~0U)); 1279 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, 1280 SDValue(Tmp, 0), N->getOperand(0), 1281 SDValue(Tmp, 1)); 1282 } 1283 1284 SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl); 1285 1286 if (N->getValueType(0) == MVT::i1) { 1287 // An i1 select is: (c & t) | (!c & f). 1288 bool Inv; 1289 unsigned Idx = getCRIdxForSetCC(CC, Inv); 1290 1291 unsigned SRI; 1292 switch (Idx) { 1293 default: llvm_unreachable("Invalid CC index"); 1294 case 0: SRI = PPC::sub_lt; break; 1295 case 1: SRI = PPC::sub_gt; break; 1296 case 2: SRI = PPC::sub_eq; break; 1297 case 3: SRI = PPC::sub_un; break; 1298 } 1299 1300 SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg); 1301 1302 SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1, 1303 CCBit, CCBit), 0); 1304 SDValue C = Inv ? NotCCBit : CCBit, 1305 NotC = Inv ? CCBit : NotCCBit; 1306 1307 SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1, 1308 C, N->getOperand(2)), 0); 1309 SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1, 1310 NotC, N->getOperand(3)), 0); 1311 1312 return CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF); 1313 } 1314 1315 unsigned BROpc = getPredicateForSetCC(CC); 1316 1317 unsigned SelectCCOp; 1318 if (N->getValueType(0) == MVT::i32) 1319 SelectCCOp = PPC::SELECT_CC_I4; 1320 else if (N->getValueType(0) == MVT::i64) 1321 SelectCCOp = PPC::SELECT_CC_I8; 1322 else if (N->getValueType(0) == MVT::f32) 1323 SelectCCOp = PPC::SELECT_CC_F4; 1324 else if (N->getValueType(0) == MVT::f64) 1325 if (PPCSubTarget->hasVSX()) 1326 SelectCCOp = PPC::SELECT_CC_VSFRC; 1327 else 1328 SelectCCOp = PPC::SELECT_CC_F8; 1329 else if (N->getValueType(0) == MVT::v2f64 || 1330 N->getValueType(0) == MVT::v2i64) 1331 SelectCCOp = PPC::SELECT_CC_VSRC; 1332 else 1333 SelectCCOp = PPC::SELECT_CC_VRRC; 1334 1335 SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3), 1336 getI32Imm(BROpc) }; 1337 return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops); 1338 } 1339 case ISD::VSELECT: 1340 if (PPCSubTarget->hasVSX()) { 1341 SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) }; 1342 return CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops); 1343 } 1344 1345 break; 1346 case ISD::VECTOR_SHUFFLE: 1347 if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 || 1348 N->getValueType(0) == MVT::v2i64)) { 1349 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 1350 1351 SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1), 1352 Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1); 1353 unsigned DM[2]; 1354 1355 for (int i = 0; i < 2; ++i) 1356 if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2) 1357 DM[i] = 0; 1358 else 1359 DM[i] = 1; 1360 1361 SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), MVT::i32); 1362 1363 if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 && 1364 Op1.getOpcode() == ISD::SCALAR_TO_VECTOR && 1365 isa<LoadSDNode>(Op1.getOperand(0))) { 1366 LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0)); 1367 SDValue Base, Offset; 1368 1369 if (LD->isUnindexed() && 1370 SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) { 1371 SDValue Chain = LD->getChain(); 1372 SDValue Ops[] = { Base, Offset, Chain }; 1373 return CurDAG->SelectNodeTo(N, PPC::LXVDSX, 1374 N->getValueType(0), Ops); 1375 } 1376 } 1377 1378 SDValue Ops[] = { Op1, Op2, DMV }; 1379 return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops); 1380 } 1381 1382 break; 1383 case PPCISD::BDNZ: 1384 case PPCISD::BDZ: { 1385 bool IsPPC64 = PPCSubTarget->isPPC64(); 1386 SDValue Ops[] = { N->getOperand(1), N->getOperand(0) }; 1387 return CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ ? 1388 (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 1389 (IsPPC64 ? PPC::BDZ8 : PPC::BDZ), 1390 MVT::Other, Ops); 1391 } 1392 case PPCISD::COND_BRANCH: { 1393 // Op #0 is the Chain. 1394 // Op #1 is the PPC::PRED_* number. 1395 // Op #2 is the CR# 1396 // Op #3 is the Dest MBB 1397 // Op #4 is the Flag. 1398 // Prevent PPC::PRED_* from being selected into LI. 1399 SDValue Pred = 1400 getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()); 1401 SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3), 1402 N->getOperand(0), N->getOperand(4) }; 1403 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); 1404 } 1405 case ISD::BR_CC: { 1406 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get(); 1407 unsigned PCC = getPredicateForSetCC(CC); 1408 1409 if (N->getOperand(2).getValueType() == MVT::i1) { 1410 unsigned Opc; 1411 bool Swap; 1412 switch (PCC) { 1413 default: llvm_unreachable("Unexpected Boolean-operand predicate"); 1414 case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break; 1415 case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break; 1416 case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break; 1417 case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break; 1418 case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break; 1419 case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break; 1420 } 1421 1422 SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1, 1423 N->getOperand(Swap ? 3 : 2), 1424 N->getOperand(Swap ? 2 : 3)), 0); 1425 return CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other, 1426 BitComp, N->getOperand(4), N->getOperand(0)); 1427 } 1428 1429 SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl); 1430 SDValue Ops[] = { getI32Imm(PCC), CondCode, 1431 N->getOperand(4), N->getOperand(0) }; 1432 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); 1433 } 1434 case ISD::BRIND: { 1435 // FIXME: Should custom lower this. 1436 SDValue Chain = N->getOperand(0); 1437 SDValue Target = N->getOperand(1); 1438 unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8; 1439 unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8; 1440 Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target, 1441 Chain), 0); 1442 return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain); 1443 } 1444 case PPCISD::TOC_ENTRY: { 1445 if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) { 1446 SDValue GA = N->getOperand(0); 1447 return CurDAG->getMachineNode(PPC::LWZtoc, dl, MVT::i32, GA, 1448 N->getOperand(1)); 1449 } 1450 assert (PPCSubTarget->isPPC64() && 1451 "Only supported for 64-bit ABI and 32-bit SVR4"); 1452 1453 // For medium and large code model, we generate two instructions as 1454 // described below. Otherwise we allow SelectCodeCommon to handle this, 1455 // selecting one of LDtoc, LDtocJTI, and LDtocCPT. 1456 CodeModel::Model CModel = TM.getCodeModel(); 1457 if (CModel != CodeModel::Medium && CModel != CodeModel::Large) 1458 break; 1459 1460 // The first source operand is a TargetGlobalAddress or a TargetJumpTable. 1461 // If it is an externally defined symbol, a symbol with common linkage, 1462 // a non-local function address, or a jump table address, or if we are 1463 // generating code for large code model, we generate: 1464 // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>)) 1465 // Otherwise we generate: 1466 // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>) 1467 SDValue GA = N->getOperand(0); 1468 SDValue TOCbase = N->getOperand(1); 1469 SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64, 1470 TOCbase, GA); 1471 1472 if (isa<JumpTableSDNode>(GA) || CModel == CodeModel::Large) 1473 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA, 1474 SDValue(Tmp, 0)); 1475 1476 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) { 1477 const GlobalValue *GValue = G->getGlobal(); 1478 if ((GValue->getType()->getElementType()->isFunctionTy() && 1479 (GValue->isDeclaration() || GValue->isWeakForLinker())) || 1480 GValue->isDeclaration() || GValue->hasCommonLinkage() || 1481 GValue->hasAvailableExternallyLinkage()) 1482 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA, 1483 SDValue(Tmp, 0)); 1484 } 1485 1486 return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64, 1487 SDValue(Tmp, 0), GA); 1488 } 1489 case PPCISD::PPC32_PICGOT: { 1490 // Generate a PIC-safe GOT reference. 1491 assert(!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() && 1492 "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4"); 1493 return CurDAG->SelectNodeTo(N, PPC::PPC32PICGOT, PPCLowering->getPointerTy(), MVT::i32); 1494 } 1495 case PPCISD::VADD_SPLAT: { 1496 // This expands into one of three sequences, depending on whether 1497 // the first operand is odd or even, positive or negative. 1498 assert(isa<ConstantSDNode>(N->getOperand(0)) && 1499 isa<ConstantSDNode>(N->getOperand(1)) && 1500 "Invalid operand on VADD_SPLAT!"); 1501 1502 int Elt = N->getConstantOperandVal(0); 1503 int EltSize = N->getConstantOperandVal(1); 1504 unsigned Opc1, Opc2, Opc3; 1505 EVT VT; 1506 1507 if (EltSize == 1) { 1508 Opc1 = PPC::VSPLTISB; 1509 Opc2 = PPC::VADDUBM; 1510 Opc3 = PPC::VSUBUBM; 1511 VT = MVT::v16i8; 1512 } else if (EltSize == 2) { 1513 Opc1 = PPC::VSPLTISH; 1514 Opc2 = PPC::VADDUHM; 1515 Opc3 = PPC::VSUBUHM; 1516 VT = MVT::v8i16; 1517 } else { 1518 assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!"); 1519 Opc1 = PPC::VSPLTISW; 1520 Opc2 = PPC::VADDUWM; 1521 Opc3 = PPC::VSUBUWM; 1522 VT = MVT::v4i32; 1523 } 1524 1525 if ((Elt & 1) == 0) { 1526 // Elt is even, in the range [-32,-18] + [16,30]. 1527 // 1528 // Convert: VADD_SPLAT elt, size 1529 // Into: tmp = VSPLTIS[BHW] elt 1530 // VADDU[BHW]M tmp, tmp 1531 // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4 1532 SDValue EltVal = getI32Imm(Elt >> 1); 1533 SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 1534 SDValue TmpVal = SDValue(Tmp, 0); 1535 return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal); 1536 1537 } else if (Elt > 0) { 1538 // Elt is odd and positive, in the range [17,31]. 1539 // 1540 // Convert: VADD_SPLAT elt, size 1541 // Into: tmp1 = VSPLTIS[BHW] elt-16 1542 // tmp2 = VSPLTIS[BHW] -16 1543 // VSUBU[BHW]M tmp1, tmp2 1544 SDValue EltVal = getI32Imm(Elt - 16); 1545 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 1546 EltVal = getI32Imm(-16); 1547 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 1548 return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0), 1549 SDValue(Tmp2, 0)); 1550 1551 } else { 1552 // Elt is odd and negative, in the range [-31,-17]. 1553 // 1554 // Convert: VADD_SPLAT elt, size 1555 // Into: tmp1 = VSPLTIS[BHW] elt+16 1556 // tmp2 = VSPLTIS[BHW] -16 1557 // VADDU[BHW]M tmp1, tmp2 1558 SDValue EltVal = getI32Imm(Elt + 16); 1559 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 1560 EltVal = getI32Imm(-16); 1561 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 1562 return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0), 1563 SDValue(Tmp2, 0)); 1564 } 1565 } 1566 } 1567 1568 return SelectCode(N); 1569 } 1570 1571 /// PostprocessISelDAG - Perform some late peephole optimizations 1572 /// on the DAG representation. 1573 void PPCDAGToDAGISel::PostprocessISelDAG() { 1574 1575 // Skip peepholes at -O0. 1576 if (TM.getOptLevel() == CodeGenOpt::None) 1577 return; 1578 1579 PeepholePPC64(); 1580 PeepholeCROps(); 1581 } 1582 1583 // Check if all users of this node will become isel where the second operand 1584 // is the constant zero. If this is so, and if we can negate the condition, 1585 // then we can flip the true and false operands. This will allow the zero to 1586 // be folded with the isel so that we don't need to materialize a register 1587 // containing zero. 1588 bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) { 1589 // If we're not using isel, then this does not matter. 1590 if (!PPCSubTarget->hasISEL()) 1591 return false; 1592 1593 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 1594 UI != UE; ++UI) { 1595 SDNode *User = *UI; 1596 if (!User->isMachineOpcode()) 1597 return false; 1598 if (User->getMachineOpcode() != PPC::SELECT_I4 && 1599 User->getMachineOpcode() != PPC::SELECT_I8) 1600 return false; 1601 1602 SDNode *Op2 = User->getOperand(2).getNode(); 1603 if (!Op2->isMachineOpcode()) 1604 return false; 1605 1606 if (Op2->getMachineOpcode() != PPC::LI && 1607 Op2->getMachineOpcode() != PPC::LI8) 1608 return false; 1609 1610 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0)); 1611 if (!C) 1612 return false; 1613 1614 if (!C->isNullValue()) 1615 return false; 1616 } 1617 1618 return true; 1619 } 1620 1621 void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) { 1622 SmallVector<SDNode *, 4> ToReplace; 1623 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 1624 UI != UE; ++UI) { 1625 SDNode *User = *UI; 1626 assert((User->getMachineOpcode() == PPC::SELECT_I4 || 1627 User->getMachineOpcode() == PPC::SELECT_I8) && 1628 "Must have all select users"); 1629 ToReplace.push_back(User); 1630 } 1631 1632 for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(), 1633 UE = ToReplace.end(); UI != UE; ++UI) { 1634 SDNode *User = *UI; 1635 SDNode *ResNode = 1636 CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User), 1637 User->getValueType(0), User->getOperand(0), 1638 User->getOperand(2), 1639 User->getOperand(1)); 1640 1641 DEBUG(dbgs() << "CR Peephole replacing:\nOld: "); 1642 DEBUG(User->dump(CurDAG)); 1643 DEBUG(dbgs() << "\nNew: "); 1644 DEBUG(ResNode->dump(CurDAG)); 1645 DEBUG(dbgs() << "\n"); 1646 1647 ReplaceUses(User, ResNode); 1648 } 1649 } 1650 1651 void PPCDAGToDAGISel::PeepholeCROps() { 1652 bool IsModified; 1653 do { 1654 IsModified = false; 1655 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(), 1656 E = CurDAG->allnodes_end(); I != E; ++I) { 1657 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I); 1658 if (!MachineNode || MachineNode->use_empty()) 1659 continue; 1660 SDNode *ResNode = MachineNode; 1661 1662 bool Op1Set = false, Op1Unset = false, 1663 Op1Not = false, 1664 Op2Set = false, Op2Unset = false, 1665 Op2Not = false; 1666 1667 unsigned Opcode = MachineNode->getMachineOpcode(); 1668 switch (Opcode) { 1669 default: break; 1670 case PPC::CRAND: 1671 case PPC::CRNAND: 1672 case PPC::CROR: 1673 case PPC::CRXOR: 1674 case PPC::CRNOR: 1675 case PPC::CREQV: 1676 case PPC::CRANDC: 1677 case PPC::CRORC: { 1678 SDValue Op = MachineNode->getOperand(1); 1679 if (Op.isMachineOpcode()) { 1680 if (Op.getMachineOpcode() == PPC::CRSET) 1681 Op2Set = true; 1682 else if (Op.getMachineOpcode() == PPC::CRUNSET) 1683 Op2Unset = true; 1684 else if (Op.getMachineOpcode() == PPC::CRNOR && 1685 Op.getOperand(0) == Op.getOperand(1)) 1686 Op2Not = true; 1687 } 1688 } // fallthrough 1689 case PPC::BC: 1690 case PPC::BCn: 1691 case PPC::SELECT_I4: 1692 case PPC::SELECT_I8: 1693 case PPC::SELECT_F4: 1694 case PPC::SELECT_F8: 1695 case PPC::SELECT_VRRC: 1696 case PPC::SELECT_VSFRC: 1697 case PPC::SELECT_VSRC: { 1698 SDValue Op = MachineNode->getOperand(0); 1699 if (Op.isMachineOpcode()) { 1700 if (Op.getMachineOpcode() == PPC::CRSET) 1701 Op1Set = true; 1702 else if (Op.getMachineOpcode() == PPC::CRUNSET) 1703 Op1Unset = true; 1704 else if (Op.getMachineOpcode() == PPC::CRNOR && 1705 Op.getOperand(0) == Op.getOperand(1)) 1706 Op1Not = true; 1707 } 1708 } 1709 break; 1710 } 1711 1712 bool SelectSwap = false; 1713 switch (Opcode) { 1714 default: break; 1715 case PPC::CRAND: 1716 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1717 // x & x = x 1718 ResNode = MachineNode->getOperand(0).getNode(); 1719 else if (Op1Set) 1720 // 1 & y = y 1721 ResNode = MachineNode->getOperand(1).getNode(); 1722 else if (Op2Set) 1723 // x & 1 = x 1724 ResNode = MachineNode->getOperand(0).getNode(); 1725 else if (Op1Unset || Op2Unset) 1726 // x & 0 = 0 & y = 0 1727 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 1728 MVT::i1); 1729 else if (Op1Not) 1730 // ~x & y = andc(y, x) 1731 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 1732 MVT::i1, MachineNode->getOperand(1), 1733 MachineNode->getOperand(0). 1734 getOperand(0)); 1735 else if (Op2Not) 1736 // x & ~y = andc(x, y) 1737 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 1738 MVT::i1, MachineNode->getOperand(0), 1739 MachineNode->getOperand(1). 1740 getOperand(0)); 1741 else if (AllUsersSelectZero(MachineNode)) 1742 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode), 1743 MVT::i1, MachineNode->getOperand(0), 1744 MachineNode->getOperand(1)), 1745 SelectSwap = true; 1746 break; 1747 case PPC::CRNAND: 1748 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1749 // nand(x, x) -> nor(x, x) 1750 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1751 MVT::i1, MachineNode->getOperand(0), 1752 MachineNode->getOperand(0)); 1753 else if (Op1Set) 1754 // nand(1, y) -> nor(y, y) 1755 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1756 MVT::i1, MachineNode->getOperand(1), 1757 MachineNode->getOperand(1)); 1758 else if (Op2Set) 1759 // nand(x, 1) -> nor(x, x) 1760 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1761 MVT::i1, MachineNode->getOperand(0), 1762 MachineNode->getOperand(0)); 1763 else if (Op1Unset || Op2Unset) 1764 // nand(x, 0) = nand(0, y) = 1 1765 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 1766 MVT::i1); 1767 else if (Op1Not) 1768 // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y) 1769 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 1770 MVT::i1, MachineNode->getOperand(0). 1771 getOperand(0), 1772 MachineNode->getOperand(1)); 1773 else if (Op2Not) 1774 // nand(x, ~y) = ~x | y = orc(y, x) 1775 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 1776 MVT::i1, MachineNode->getOperand(1). 1777 getOperand(0), 1778 MachineNode->getOperand(0)); 1779 else if (AllUsersSelectZero(MachineNode)) 1780 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode), 1781 MVT::i1, MachineNode->getOperand(0), 1782 MachineNode->getOperand(1)), 1783 SelectSwap = true; 1784 break; 1785 case PPC::CROR: 1786 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1787 // x | x = x 1788 ResNode = MachineNode->getOperand(0).getNode(); 1789 else if (Op1Set || Op2Set) 1790 // x | 1 = 1 | y = 1 1791 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 1792 MVT::i1); 1793 else if (Op1Unset) 1794 // 0 | y = y 1795 ResNode = MachineNode->getOperand(1).getNode(); 1796 else if (Op2Unset) 1797 // x | 0 = x 1798 ResNode = MachineNode->getOperand(0).getNode(); 1799 else if (Op1Not) 1800 // ~x | y = orc(y, x) 1801 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 1802 MVT::i1, MachineNode->getOperand(1), 1803 MachineNode->getOperand(0). 1804 getOperand(0)); 1805 else if (Op2Not) 1806 // x | ~y = orc(x, y) 1807 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 1808 MVT::i1, MachineNode->getOperand(0), 1809 MachineNode->getOperand(1). 1810 getOperand(0)); 1811 else if (AllUsersSelectZero(MachineNode)) 1812 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1813 MVT::i1, MachineNode->getOperand(0), 1814 MachineNode->getOperand(1)), 1815 SelectSwap = true; 1816 break; 1817 case PPC::CRXOR: 1818 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1819 // xor(x, x) = 0 1820 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 1821 MVT::i1); 1822 else if (Op1Set) 1823 // xor(1, y) -> nor(y, y) 1824 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1825 MVT::i1, MachineNode->getOperand(1), 1826 MachineNode->getOperand(1)); 1827 else if (Op2Set) 1828 // xor(x, 1) -> nor(x, x) 1829 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1830 MVT::i1, MachineNode->getOperand(0), 1831 MachineNode->getOperand(0)); 1832 else if (Op1Unset) 1833 // xor(0, y) = y 1834 ResNode = MachineNode->getOperand(1).getNode(); 1835 else if (Op2Unset) 1836 // xor(x, 0) = x 1837 ResNode = MachineNode->getOperand(0).getNode(); 1838 else if (Op1Not) 1839 // xor(~x, y) = eqv(x, y) 1840 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 1841 MVT::i1, MachineNode->getOperand(0). 1842 getOperand(0), 1843 MachineNode->getOperand(1)); 1844 else if (Op2Not) 1845 // xor(x, ~y) = eqv(x, y) 1846 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 1847 MVT::i1, MachineNode->getOperand(0), 1848 MachineNode->getOperand(1). 1849 getOperand(0)); 1850 else if (AllUsersSelectZero(MachineNode)) 1851 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 1852 MVT::i1, MachineNode->getOperand(0), 1853 MachineNode->getOperand(1)), 1854 SelectSwap = true; 1855 break; 1856 case PPC::CRNOR: 1857 if (Op1Set || Op2Set) 1858 // nor(1, y) -> 0 1859 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 1860 MVT::i1); 1861 else if (Op1Unset) 1862 // nor(0, y) = ~y -> nor(y, y) 1863 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1864 MVT::i1, MachineNode->getOperand(1), 1865 MachineNode->getOperand(1)); 1866 else if (Op2Unset) 1867 // nor(x, 0) = ~x 1868 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1869 MVT::i1, MachineNode->getOperand(0), 1870 MachineNode->getOperand(0)); 1871 else if (Op1Not) 1872 // nor(~x, y) = andc(x, y) 1873 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 1874 MVT::i1, MachineNode->getOperand(0). 1875 getOperand(0), 1876 MachineNode->getOperand(1)); 1877 else if (Op2Not) 1878 // nor(x, ~y) = andc(y, x) 1879 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 1880 MVT::i1, MachineNode->getOperand(1). 1881 getOperand(0), 1882 MachineNode->getOperand(0)); 1883 else if (AllUsersSelectZero(MachineNode)) 1884 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode), 1885 MVT::i1, MachineNode->getOperand(0), 1886 MachineNode->getOperand(1)), 1887 SelectSwap = true; 1888 break; 1889 case PPC::CREQV: 1890 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1891 // eqv(x, x) = 1 1892 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 1893 MVT::i1); 1894 else if (Op1Set) 1895 // eqv(1, y) = y 1896 ResNode = MachineNode->getOperand(1).getNode(); 1897 else if (Op2Set) 1898 // eqv(x, 1) = x 1899 ResNode = MachineNode->getOperand(0).getNode(); 1900 else if (Op1Unset) 1901 // eqv(0, y) = ~y -> nor(y, y) 1902 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1903 MVT::i1, MachineNode->getOperand(1), 1904 MachineNode->getOperand(1)); 1905 else if (Op2Unset) 1906 // eqv(x, 0) = ~x 1907 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1908 MVT::i1, MachineNode->getOperand(0), 1909 MachineNode->getOperand(0)); 1910 else if (Op1Not) 1911 // eqv(~x, y) = xor(x, y) 1912 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 1913 MVT::i1, MachineNode->getOperand(0). 1914 getOperand(0), 1915 MachineNode->getOperand(1)); 1916 else if (Op2Not) 1917 // eqv(x, ~y) = xor(x, y) 1918 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 1919 MVT::i1, MachineNode->getOperand(0), 1920 MachineNode->getOperand(1). 1921 getOperand(0)); 1922 else if (AllUsersSelectZero(MachineNode)) 1923 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 1924 MVT::i1, MachineNode->getOperand(0), 1925 MachineNode->getOperand(1)), 1926 SelectSwap = true; 1927 break; 1928 case PPC::CRANDC: 1929 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1930 // andc(x, x) = 0 1931 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 1932 MVT::i1); 1933 else if (Op1Set) 1934 // andc(1, y) = ~y 1935 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1936 MVT::i1, MachineNode->getOperand(1), 1937 MachineNode->getOperand(1)); 1938 else if (Op1Unset || Op2Set) 1939 // andc(0, y) = andc(x, 1) = 0 1940 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 1941 MVT::i1); 1942 else if (Op2Unset) 1943 // andc(x, 0) = x 1944 ResNode = MachineNode->getOperand(0).getNode(); 1945 else if (Op1Not) 1946 // andc(~x, y) = ~(x | y) = nor(x, y) 1947 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1948 MVT::i1, MachineNode->getOperand(0). 1949 getOperand(0), 1950 MachineNode->getOperand(1)); 1951 else if (Op2Not) 1952 // andc(x, ~y) = x & y 1953 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode), 1954 MVT::i1, MachineNode->getOperand(0), 1955 MachineNode->getOperand(1). 1956 getOperand(0)); 1957 else if (AllUsersSelectZero(MachineNode)) 1958 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 1959 MVT::i1, MachineNode->getOperand(1), 1960 MachineNode->getOperand(0)), 1961 SelectSwap = true; 1962 break; 1963 case PPC::CRORC: 1964 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1965 // orc(x, x) = 1 1966 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 1967 MVT::i1); 1968 else if (Op1Set || Op2Unset) 1969 // orc(1, y) = orc(x, 0) = 1 1970 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 1971 MVT::i1); 1972 else if (Op2Set) 1973 // orc(x, 1) = x 1974 ResNode = MachineNode->getOperand(0).getNode(); 1975 else if (Op1Unset) 1976 // orc(0, y) = ~y 1977 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1978 MVT::i1, MachineNode->getOperand(1), 1979 MachineNode->getOperand(1)); 1980 else if (Op1Not) 1981 // orc(~x, y) = ~(x & y) = nand(x, y) 1982 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode), 1983 MVT::i1, MachineNode->getOperand(0). 1984 getOperand(0), 1985 MachineNode->getOperand(1)); 1986 else if (Op2Not) 1987 // orc(x, ~y) = x | y 1988 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode), 1989 MVT::i1, MachineNode->getOperand(0), 1990 MachineNode->getOperand(1). 1991 getOperand(0)); 1992 else if (AllUsersSelectZero(MachineNode)) 1993 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 1994 MVT::i1, MachineNode->getOperand(1), 1995 MachineNode->getOperand(0)), 1996 SelectSwap = true; 1997 break; 1998 case PPC::SELECT_I4: 1999 case PPC::SELECT_I8: 2000 case PPC::SELECT_F4: 2001 case PPC::SELECT_F8: 2002 case PPC::SELECT_VRRC: 2003 case PPC::SELECT_VSFRC: 2004 case PPC::SELECT_VSRC: 2005 if (Op1Set) 2006 ResNode = MachineNode->getOperand(1).getNode(); 2007 else if (Op1Unset) 2008 ResNode = MachineNode->getOperand(2).getNode(); 2009 else if (Op1Not) 2010 ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(), 2011 SDLoc(MachineNode), 2012 MachineNode->getValueType(0), 2013 MachineNode->getOperand(0). 2014 getOperand(0), 2015 MachineNode->getOperand(2), 2016 MachineNode->getOperand(1)); 2017 break; 2018 case PPC::BC: 2019 case PPC::BCn: 2020 if (Op1Not) 2021 ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn : 2022 PPC::BC, 2023 SDLoc(MachineNode), 2024 MVT::Other, 2025 MachineNode->getOperand(0). 2026 getOperand(0), 2027 MachineNode->getOperand(1), 2028 MachineNode->getOperand(2)); 2029 // FIXME: Handle Op1Set, Op1Unset here too. 2030 break; 2031 } 2032 2033 // If we're inverting this node because it is used only by selects that 2034 // we'd like to swap, then swap the selects before the node replacement. 2035 if (SelectSwap) 2036 SwapAllSelectUsers(MachineNode); 2037 2038 if (ResNode != MachineNode) { 2039 DEBUG(dbgs() << "CR Peephole replacing:\nOld: "); 2040 DEBUG(MachineNode->dump(CurDAG)); 2041 DEBUG(dbgs() << "\nNew: "); 2042 DEBUG(ResNode->dump(CurDAG)); 2043 DEBUG(dbgs() << "\n"); 2044 2045 ReplaceUses(MachineNode, ResNode); 2046 IsModified = true; 2047 } 2048 } 2049 if (IsModified) 2050 CurDAG->RemoveDeadNodes(); 2051 } while (IsModified); 2052 } 2053 2054 void PPCDAGToDAGISel::PeepholePPC64() { 2055 // These optimizations are currently supported only for 64-bit SVR4. 2056 if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64()) 2057 return; 2058 2059 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode()); 2060 ++Position; 2061 2062 while (Position != CurDAG->allnodes_begin()) { 2063 SDNode *N = --Position; 2064 // Skip dead nodes and any non-machine opcodes. 2065 if (N->use_empty() || !N->isMachineOpcode()) 2066 continue; 2067 2068 unsigned FirstOp; 2069 unsigned StorageOpcode = N->getMachineOpcode(); 2070 2071 switch (StorageOpcode) { 2072 default: continue; 2073 2074 case PPC::LBZ: 2075 case PPC::LBZ8: 2076 case PPC::LD: 2077 case PPC::LFD: 2078 case PPC::LFS: 2079 case PPC::LHA: 2080 case PPC::LHA8: 2081 case PPC::LHZ: 2082 case PPC::LHZ8: 2083 case PPC::LWA: 2084 case PPC::LWZ: 2085 case PPC::LWZ8: 2086 FirstOp = 0; 2087 break; 2088 2089 case PPC::STB: 2090 case PPC::STB8: 2091 case PPC::STD: 2092 case PPC::STFD: 2093 case PPC::STFS: 2094 case PPC::STH: 2095 case PPC::STH8: 2096 case PPC::STW: 2097 case PPC::STW8: 2098 FirstOp = 1; 2099 break; 2100 } 2101 2102 // If this is a load or store with a zero offset, we may be able to 2103 // fold an add-immediate into the memory operation. 2104 if (!isa<ConstantSDNode>(N->getOperand(FirstOp)) || 2105 N->getConstantOperandVal(FirstOp) != 0) 2106 continue; 2107 2108 SDValue Base = N->getOperand(FirstOp + 1); 2109 if (!Base.isMachineOpcode()) 2110 continue; 2111 2112 unsigned Flags = 0; 2113 bool ReplaceFlags = true; 2114 2115 // When the feeding operation is an add-immediate of some sort, 2116 // determine whether we need to add relocation information to the 2117 // target flags on the immediate operand when we fold it into the 2118 // load instruction. 2119 // 2120 // For something like ADDItocL, the relocation information is 2121 // inferred from the opcode; when we process it in the AsmPrinter, 2122 // we add the necessary relocation there. A load, though, can receive 2123 // relocation from various flavors of ADDIxxx, so we need to carry 2124 // the relocation information in the target flags. 2125 switch (Base.getMachineOpcode()) { 2126 default: continue; 2127 2128 case PPC::ADDI8: 2129 case PPC::ADDI: 2130 // In some cases (such as TLS) the relocation information 2131 // is already in place on the operand, so copying the operand 2132 // is sufficient. 2133 ReplaceFlags = false; 2134 // For these cases, the immediate may not be divisible by 4, in 2135 // which case the fold is illegal for DS-form instructions. (The 2136 // other cases provide aligned addresses and are always safe.) 2137 if ((StorageOpcode == PPC::LWA || 2138 StorageOpcode == PPC::LD || 2139 StorageOpcode == PPC::STD) && 2140 (!isa<ConstantSDNode>(Base.getOperand(1)) || 2141 Base.getConstantOperandVal(1) % 4 != 0)) 2142 continue; 2143 break; 2144 case PPC::ADDIdtprelL: 2145 Flags = PPCII::MO_DTPREL_LO; 2146 break; 2147 case PPC::ADDItlsldL: 2148 Flags = PPCII::MO_TLSLD_LO; 2149 break; 2150 case PPC::ADDItocL: 2151 Flags = PPCII::MO_TOC_LO; 2152 break; 2153 } 2154 2155 // We found an opportunity. Reverse the operands from the add 2156 // immediate and substitute them into the load or store. If 2157 // needed, update the target flags for the immediate operand to 2158 // reflect the necessary relocation information. 2159 DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: "); 2160 DEBUG(Base->dump(CurDAG)); 2161 DEBUG(dbgs() << "\nN: "); 2162 DEBUG(N->dump(CurDAG)); 2163 DEBUG(dbgs() << "\n"); 2164 2165 SDValue ImmOpnd = Base.getOperand(1); 2166 2167 // If the relocation information isn't already present on the 2168 // immediate operand, add it now. 2169 if (ReplaceFlags) { 2170 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) { 2171 SDLoc dl(GA); 2172 const GlobalValue *GV = GA->getGlobal(); 2173 // We can't perform this optimization for data whose alignment 2174 // is insufficient for the instruction encoding. 2175 if (GV->getAlignment() < 4 && 2176 (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD || 2177 StorageOpcode == PPC::LWA)) { 2178 DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n"); 2179 continue; 2180 } 2181 ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, 0, Flags); 2182 } else if (ConstantPoolSDNode *CP = 2183 dyn_cast<ConstantPoolSDNode>(ImmOpnd)) { 2184 const Constant *C = CP->getConstVal(); 2185 ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64, 2186 CP->getAlignment(), 2187 0, Flags); 2188 } 2189 } 2190 2191 if (FirstOp == 1) // Store 2192 (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd, 2193 Base.getOperand(0), N->getOperand(3)); 2194 else // Load 2195 (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0), 2196 N->getOperand(2)); 2197 2198 // The add-immediate may now be dead, in which case remove it. 2199 if (Base.getNode()->use_empty()) 2200 CurDAG->RemoveDeadNode(Base.getNode()); 2201 } 2202 } 2203 2204 2205 /// createPPCISelDag - This pass converts a legalized DAG into a 2206 /// PowerPC-specific DAG, ready for instruction scheduling. 2207 /// 2208 FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) { 2209 return new PPCDAGToDAGISel(TM); 2210 } 2211 2212 static void initializePassOnce(PassRegistry &Registry) { 2213 const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection"; 2214 PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID, 2215 nullptr, false, false); 2216 Registry.registerPass(*PI, true); 2217 } 2218 2219 void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) { 2220 CALL_ONCE_INITIALIZATION(initializePassOnce); 2221 } 2222 2223