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 "PPCTargetMachine.h" 18 #include "llvm/CodeGen/MachineFunction.h" 19 #include "llvm/CodeGen/MachineInstrBuilder.h" 20 #include "llvm/CodeGen/MachineRegisterInfo.h" 21 #include "llvm/CodeGen/SelectionDAG.h" 22 #include "llvm/CodeGen/SelectionDAGISel.h" 23 #include "llvm/IR/Constants.h" 24 #include "llvm/IR/Function.h" 25 #include "llvm/IR/GlobalAlias.h" 26 #include "llvm/IR/GlobalValue.h" 27 #include "llvm/IR/GlobalVariable.h" 28 #include "llvm/IR/Intrinsics.h" 29 #include "llvm/Support/CommandLine.h" 30 #include "llvm/Support/Debug.h" 31 #include "llvm/Support/ErrorHandling.h" 32 #include "llvm/Support/MathExtras.h" 33 #include "llvm/Support/raw_ostream.h" 34 #include "llvm/Target/TargetOptions.h" 35 using namespace llvm; 36 37 #define DEBUG_TYPE "ppc-codegen" 38 39 // FIXME: Remove this once the bug has been fixed! 40 cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug", 41 cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden); 42 43 namespace llvm { 44 void initializePPCDAGToDAGISelPass(PassRegistry&); 45 } 46 47 namespace { 48 //===--------------------------------------------------------------------===// 49 /// PPCDAGToDAGISel - PPC specific code to select PPC machine 50 /// instructions for SelectionDAG operations. 51 /// 52 class PPCDAGToDAGISel : public SelectionDAGISel { 53 const PPCTargetMachine &TM; 54 const PPCTargetLowering *PPCLowering; 55 const PPCSubtarget *PPCSubTarget; 56 unsigned GlobalBaseReg; 57 public: 58 explicit PPCDAGToDAGISel(PPCTargetMachine &tm) 59 : SelectionDAGISel(tm), TM(tm), 60 PPCLowering(TM.getTargetLowering()), 61 PPCSubTarget(TM.getSubtargetImpl()) { 62 initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry()); 63 } 64 65 bool runOnMachineFunction(MachineFunction &MF) override { 66 // Make sure we re-emit a set of the global base reg if necessary 67 GlobalBaseReg = 0; 68 PPCLowering = TM.getTargetLowering(); 69 PPCSubTarget = TM.getSubtargetImpl(); 70 SelectionDAGISel::runOnMachineFunction(MF); 71 72 if (!PPCSubTarget->isSVR4ABI()) 73 InsertVRSaveCode(MF); 74 75 return true; 76 } 77 78 void PostprocessISelDAG() override; 79 80 /// getI32Imm - Return a target constant with the specified value, of type 81 /// i32. 82 inline SDValue getI32Imm(unsigned Imm) { 83 return CurDAG->getTargetConstant(Imm, MVT::i32); 84 } 85 86 /// getI64Imm - Return a target constant with the specified value, of type 87 /// i64. 88 inline SDValue getI64Imm(uint64_t Imm) { 89 return CurDAG->getTargetConstant(Imm, MVT::i64); 90 } 91 92 /// getSmallIPtrImm - Return a target constant of pointer type. 93 inline SDValue getSmallIPtrImm(unsigned Imm) { 94 return CurDAG->getTargetConstant(Imm, PPCLowering->getPointerTy()); 95 } 96 97 /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s 98 /// with any number of 0s on either side. The 1s are allowed to wrap from 99 /// LSB to MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 100 /// 0x0F0F0000 is not, since all 1s are not contiguous. 101 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME); 102 103 104 /// isRotateAndMask - Returns true if Mask and Shift can be folded into a 105 /// rotate and mask opcode and mask operation. 106 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask, 107 unsigned &SH, unsigned &MB, unsigned &ME); 108 109 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC 110 /// base register. Return the virtual register that holds this value. 111 SDNode *getGlobalBaseReg(); 112 113 // Select - Convert the specified operand from a target-independent to a 114 // target-specific node if it hasn't already been changed. 115 SDNode *Select(SDNode *N) override; 116 117 SDNode *SelectBitfieldInsert(SDNode *N); 118 119 /// SelectCC - Select a comparison of the specified values with the 120 /// specified condition code, returning the CR# of the expression. 121 SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, SDLoc dl); 122 123 /// SelectAddrImm - Returns true if the address N can be represented by 124 /// a base register plus a signed 16-bit displacement [r+imm]. 125 bool SelectAddrImm(SDValue N, SDValue &Disp, 126 SDValue &Base) { 127 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false); 128 } 129 130 /// SelectAddrImmOffs - Return true if the operand is valid for a preinc 131 /// immediate field. Note that the operand at this point is already the 132 /// result of a prior SelectAddressRegImm call. 133 bool SelectAddrImmOffs(SDValue N, SDValue &Out) const { 134 if (N.getOpcode() == ISD::TargetConstant || 135 N.getOpcode() == ISD::TargetGlobalAddress) { 136 Out = N; 137 return true; 138 } 139 140 return false; 141 } 142 143 /// SelectAddrIdx - Given the specified addressed, check to see if it can be 144 /// represented as an indexed [r+r] operation. Returns false if it can 145 /// be represented by [r+imm], which are preferred. 146 bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) { 147 return PPCLowering->SelectAddressRegReg(N, Base, Index, *CurDAG); 148 } 149 150 /// SelectAddrIdxOnly - Given the specified addressed, force it to be 151 /// represented as an indexed [r+r] operation. 152 bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) { 153 return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, *CurDAG); 154 } 155 156 /// SelectAddrImmX4 - Returns true if the address N can be represented by 157 /// a base register plus a signed 16-bit displacement that is a multiple of 4. 158 /// Suitable for use by STD and friends. 159 bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) { 160 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true); 161 } 162 163 // Select an address into a single register. 164 bool SelectAddr(SDValue N, SDValue &Base) { 165 Base = N; 166 return true; 167 } 168 169 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for 170 /// inline asm expressions. It is always correct to compute the value into 171 /// a register. The case of adding a (possibly relocatable) constant to a 172 /// register can be improved, but it is wrong to substitute Reg+Reg for 173 /// Reg in an asm, because the load or store opcode would have to change. 174 bool SelectInlineAsmMemoryOperand(const SDValue &Op, 175 char ConstraintCode, 176 std::vector<SDValue> &OutOps) override { 177 OutOps.push_back(Op); 178 return false; 179 } 180 181 void InsertVRSaveCode(MachineFunction &MF); 182 183 const char *getPassName() const override { 184 return "PowerPC DAG->DAG Pattern Instruction Selection"; 185 } 186 187 // Include the pieces autogenerated from the target description. 188 #include "PPCGenDAGISel.inc" 189 190 private: 191 SDNode *SelectSETCC(SDNode *N); 192 193 void PeepholePPC64(); 194 void PeepholeCROps(); 195 196 bool AllUsersSelectZero(SDNode *N); 197 void SwapAllSelectUsers(SDNode *N); 198 }; 199 } 200 201 /// InsertVRSaveCode - Once the entire function has been instruction selected, 202 /// all virtual registers are created and all machine instructions are built, 203 /// check to see if we need to save/restore VRSAVE. If so, do it. 204 void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) { 205 // Check to see if this function uses vector registers, which means we have to 206 // save and restore the VRSAVE register and update it with the regs we use. 207 // 208 // In this case, there will be virtual registers of vector type created 209 // by the scheduler. Detect them now. 210 bool HasVectorVReg = false; 211 for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) { 212 unsigned Reg = TargetRegisterInfo::index2VirtReg(i); 213 if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) { 214 HasVectorVReg = true; 215 break; 216 } 217 } 218 if (!HasVectorVReg) return; // nothing to do. 219 220 // If we have a vector register, we want to emit code into the entry and exit 221 // blocks to save and restore the VRSAVE register. We do this here (instead 222 // of marking all vector instructions as clobbering VRSAVE) for two reasons: 223 // 224 // 1. This (trivially) reduces the load on the register allocator, by not 225 // having to represent the live range of the VRSAVE register. 226 // 2. This (more significantly) allows us to create a temporary virtual 227 // register to hold the saved VRSAVE value, allowing this temporary to be 228 // register allocated, instead of forcing it to be spilled to the stack. 229 230 // Create two vregs - one to hold the VRSAVE register that is live-in to the 231 // function and one for the value after having bits or'd into it. 232 unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 233 unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 234 235 const TargetInstrInfo &TII = *TM.getInstrInfo(); 236 MachineBasicBlock &EntryBB = *Fn.begin(); 237 DebugLoc dl; 238 // Emit the following code into the entry block: 239 // InVRSAVE = MFVRSAVE 240 // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE 241 // MTVRSAVE UpdatedVRSAVE 242 MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point 243 BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE); 244 BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE), 245 UpdatedVRSAVE).addReg(InVRSAVE); 246 BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE); 247 248 // Find all return blocks, outputting a restore in each epilog. 249 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { 250 if (!BB->empty() && BB->back().isReturn()) { 251 IP = BB->end(); --IP; 252 253 // Skip over all terminator instructions, which are part of the return 254 // sequence. 255 MachineBasicBlock::iterator I2 = IP; 256 while (I2 != BB->begin() && (--I2)->isTerminator()) 257 IP = I2; 258 259 // Emit: MTVRSAVE InVRSave 260 BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE); 261 } 262 } 263 } 264 265 266 /// getGlobalBaseReg - Output the instructions required to put the 267 /// base address to use for accessing globals into a register. 268 /// 269 SDNode *PPCDAGToDAGISel::getGlobalBaseReg() { 270 if (!GlobalBaseReg) { 271 const TargetInstrInfo &TII = *TM.getInstrInfo(); 272 // Insert the set of GlobalBaseReg into the first MBB of the function 273 MachineBasicBlock &FirstMBB = MF->front(); 274 MachineBasicBlock::iterator MBBI = FirstMBB.begin(); 275 DebugLoc dl; 276 277 if (PPCLowering->getPointerTy() == MVT::i32) { 278 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::GPRC_NOR0RegClass); 279 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR)); 280 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); 281 } else { 282 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_NOX0RegClass); 283 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8)); 284 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg); 285 } 286 } 287 return CurDAG->getRegister(GlobalBaseReg, 288 PPCLowering->getPointerTy()).getNode(); 289 } 290 291 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit 292 /// or 64-bit immediate, and if the value can be accurately represented as a 293 /// sign extension from a 16-bit value. If so, this returns true and the 294 /// immediate. 295 static bool isIntS16Immediate(SDNode *N, short &Imm) { 296 if (N->getOpcode() != ISD::Constant) 297 return false; 298 299 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue(); 300 if (N->getValueType(0) == MVT::i32) 301 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue(); 302 else 303 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue(); 304 } 305 306 static bool isIntS16Immediate(SDValue Op, short &Imm) { 307 return isIntS16Immediate(Op.getNode(), Imm); 308 } 309 310 311 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant 312 /// operand. If so Imm will receive the 32-bit value. 313 static bool isInt32Immediate(SDNode *N, unsigned &Imm) { 314 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) { 315 Imm = cast<ConstantSDNode>(N)->getZExtValue(); 316 return true; 317 } 318 return false; 319 } 320 321 /// isInt64Immediate - This method tests to see if the node is a 64-bit constant 322 /// operand. If so Imm will receive the 64-bit value. 323 static bool isInt64Immediate(SDNode *N, uint64_t &Imm) { 324 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) { 325 Imm = cast<ConstantSDNode>(N)->getZExtValue(); 326 return true; 327 } 328 return false; 329 } 330 331 // isInt32Immediate - This method tests to see if a constant operand. 332 // If so Imm will receive the 32 bit value. 333 static bool isInt32Immediate(SDValue N, unsigned &Imm) { 334 return isInt32Immediate(N.getNode(), Imm); 335 } 336 337 338 // isOpcWithIntImmediate - This method tests to see if the node is a specific 339 // opcode and that it has a immediate integer right operand. 340 // If so Imm will receive the 32 bit value. 341 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) { 342 return N->getOpcode() == Opc 343 && isInt32Immediate(N->getOperand(1).getNode(), Imm); 344 } 345 346 bool PPCDAGToDAGISel::isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) { 347 if (!Val) 348 return false; 349 350 if (isShiftedMask_32(Val)) { 351 // look for the first non-zero bit 352 MB = countLeadingZeros(Val); 353 // look for the first zero bit after the run of ones 354 ME = countLeadingZeros((Val - 1) ^ Val); 355 return true; 356 } else { 357 Val = ~Val; // invert mask 358 if (isShiftedMask_32(Val)) { 359 // effectively look for the first zero bit 360 ME = countLeadingZeros(Val) - 1; 361 // effectively look for the first one bit after the run of zeros 362 MB = countLeadingZeros((Val - 1) ^ Val) + 1; 363 return true; 364 } 365 } 366 // no run present 367 return false; 368 } 369 370 bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask, 371 bool isShiftMask, unsigned &SH, 372 unsigned &MB, unsigned &ME) { 373 // Don't even go down this path for i64, since different logic will be 374 // necessary for rldicl/rldicr/rldimi. 375 if (N->getValueType(0) != MVT::i32) 376 return false; 377 378 unsigned Shift = 32; 379 unsigned Indeterminant = ~0; // bit mask marking indeterminant results 380 unsigned Opcode = N->getOpcode(); 381 if (N->getNumOperands() != 2 || 382 !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31)) 383 return false; 384 385 if (Opcode == ISD::SHL) { 386 // apply shift left to mask if it comes first 387 if (isShiftMask) Mask = Mask << Shift; 388 // determine which bits are made indeterminant by shift 389 Indeterminant = ~(0xFFFFFFFFu << Shift); 390 } else if (Opcode == ISD::SRL) { 391 // apply shift right to mask if it comes first 392 if (isShiftMask) Mask = Mask >> Shift; 393 // determine which bits are made indeterminant by shift 394 Indeterminant = ~(0xFFFFFFFFu >> Shift); 395 // adjust for the left rotate 396 Shift = 32 - Shift; 397 } else if (Opcode == ISD::ROTL) { 398 Indeterminant = 0; 399 } else { 400 return false; 401 } 402 403 // if the mask doesn't intersect any Indeterminant bits 404 if (Mask && !(Mask & Indeterminant)) { 405 SH = Shift & 31; 406 // make sure the mask is still a mask (wrap arounds may not be) 407 return isRunOfOnes(Mask, MB, ME); 408 } 409 return false; 410 } 411 412 /// SelectBitfieldInsert - turn an or of two masked values into 413 /// the rotate left word immediate then mask insert (rlwimi) instruction. 414 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) { 415 SDValue Op0 = N->getOperand(0); 416 SDValue Op1 = N->getOperand(1); 417 SDLoc dl(N); 418 419 APInt LKZ, LKO, RKZ, RKO; 420 CurDAG->computeKnownBits(Op0, LKZ, LKO); 421 CurDAG->computeKnownBits(Op1, RKZ, RKO); 422 423 unsigned TargetMask = LKZ.getZExtValue(); 424 unsigned InsertMask = RKZ.getZExtValue(); 425 426 if ((TargetMask | InsertMask) == 0xFFFFFFFF) { 427 unsigned Op0Opc = Op0.getOpcode(); 428 unsigned Op1Opc = Op1.getOpcode(); 429 unsigned Value, SH = 0; 430 TargetMask = ~TargetMask; 431 InsertMask = ~InsertMask; 432 433 // If the LHS has a foldable shift and the RHS does not, then swap it to the 434 // RHS so that we can fold the shift into the insert. 435 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) { 436 if (Op0.getOperand(0).getOpcode() == ISD::SHL || 437 Op0.getOperand(0).getOpcode() == ISD::SRL) { 438 if (Op1.getOperand(0).getOpcode() != ISD::SHL && 439 Op1.getOperand(0).getOpcode() != ISD::SRL) { 440 std::swap(Op0, Op1); 441 std::swap(Op0Opc, Op1Opc); 442 std::swap(TargetMask, InsertMask); 443 } 444 } 445 } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) { 446 if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL && 447 Op1.getOperand(0).getOpcode() != ISD::SRL) { 448 std::swap(Op0, Op1); 449 std::swap(Op0Opc, Op1Opc); 450 std::swap(TargetMask, InsertMask); 451 } 452 } 453 454 unsigned MB, ME; 455 if (isRunOfOnes(InsertMask, MB, ME)) { 456 SDValue Tmp1, Tmp2; 457 458 if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) && 459 isInt32Immediate(Op1.getOperand(1), Value)) { 460 Op1 = Op1.getOperand(0); 461 SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value; 462 } 463 if (Op1Opc == ISD::AND) { 464 // The AND mask might not be a constant, and we need to make sure that 465 // if we're going to fold the masking with the insert, all bits not 466 // know to be zero in the mask are known to be one. 467 APInt MKZ, MKO; 468 CurDAG->computeKnownBits(Op1.getOperand(1), MKZ, MKO); 469 bool CanFoldMask = InsertMask == MKO.getZExtValue(); 470 471 unsigned SHOpc = Op1.getOperand(0).getOpcode(); 472 if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask && 473 isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) { 474 // Note that Value must be in range here (less than 32) because 475 // otherwise there would not be any bits set in InsertMask. 476 Op1 = Op1.getOperand(0).getOperand(0); 477 SH = (SHOpc == ISD::SHL) ? Value : 32 - Value; 478 } 479 } 480 481 SH &= 31; 482 SDValue Ops[] = { Op0, Op1, getI32Imm(SH), getI32Imm(MB), 483 getI32Imm(ME) }; 484 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); 485 } 486 } 487 return nullptr; 488 } 489 490 /// SelectCC - Select a comparison of the specified values with the specified 491 /// condition code, returning the CR# of the expression. 492 SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, 493 ISD::CondCode CC, SDLoc dl) { 494 // Always select the LHS. 495 unsigned Opc; 496 497 if (LHS.getValueType() == MVT::i32) { 498 unsigned Imm; 499 if (CC == ISD::SETEQ || CC == ISD::SETNE) { 500 if (isInt32Immediate(RHS, Imm)) { 501 // SETEQ/SETNE comparison with 16-bit immediate, fold it. 502 if (isUInt<16>(Imm)) 503 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, 504 getI32Imm(Imm & 0xFFFF)), 0); 505 // If this is a 16-bit signed immediate, fold it. 506 if (isInt<16>((int)Imm)) 507 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, 508 getI32Imm(Imm & 0xFFFF)), 0); 509 510 // For non-equality comparisons, the default code would materialize the 511 // constant, then compare against it, like this: 512 // lis r2, 4660 513 // ori r2, r2, 22136 514 // cmpw cr0, r3, r2 515 // Since we are just comparing for equality, we can emit this instead: 516 // xoris r0,r3,0x1234 517 // cmplwi cr0,r0,0x5678 518 // beq cr0,L6 519 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS, 520 getI32Imm(Imm >> 16)), 0); 521 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor, 522 getI32Imm(Imm & 0xFFFF)), 0); 523 } 524 Opc = PPC::CMPLW; 525 } else if (ISD::isUnsignedIntSetCC(CC)) { 526 if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm)) 527 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, 528 getI32Imm(Imm & 0xFFFF)), 0); 529 Opc = PPC::CMPLW; 530 } else { 531 short SImm; 532 if (isIntS16Immediate(RHS, SImm)) 533 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, 534 getI32Imm((int)SImm & 0xFFFF)), 535 0); 536 Opc = PPC::CMPW; 537 } 538 } else if (LHS.getValueType() == MVT::i64) { 539 uint64_t Imm; 540 if (CC == ISD::SETEQ || CC == ISD::SETNE) { 541 if (isInt64Immediate(RHS.getNode(), Imm)) { 542 // SETEQ/SETNE comparison with 16-bit immediate, fold it. 543 if (isUInt<16>(Imm)) 544 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, 545 getI32Imm(Imm & 0xFFFF)), 0); 546 // If this is a 16-bit signed immediate, fold it. 547 if (isInt<16>(Imm)) 548 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, 549 getI32Imm(Imm & 0xFFFF)), 0); 550 551 // For non-equality comparisons, the default code would materialize the 552 // constant, then compare against it, like this: 553 // lis r2, 4660 554 // ori r2, r2, 22136 555 // cmpd cr0, r3, r2 556 // Since we are just comparing for equality, we can emit this instead: 557 // xoris r0,r3,0x1234 558 // cmpldi cr0,r0,0x5678 559 // beq cr0,L6 560 if (isUInt<32>(Imm)) { 561 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS, 562 getI64Imm(Imm >> 16)), 0); 563 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor, 564 getI64Imm(Imm & 0xFFFF)), 0); 565 } 566 } 567 Opc = PPC::CMPLD; 568 } else if (ISD::isUnsignedIntSetCC(CC)) { 569 if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm)) 570 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, 571 getI64Imm(Imm & 0xFFFF)), 0); 572 Opc = PPC::CMPLD; 573 } else { 574 short SImm; 575 if (isIntS16Immediate(RHS, SImm)) 576 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, 577 getI64Imm(SImm & 0xFFFF)), 578 0); 579 Opc = PPC::CMPD; 580 } 581 } else if (LHS.getValueType() == MVT::f32) { 582 Opc = PPC::FCMPUS; 583 } else { 584 assert(LHS.getValueType() == MVT::f64 && "Unknown vt!"); 585 Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD; 586 } 587 return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0); 588 } 589 590 static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) { 591 switch (CC) { 592 case ISD::SETUEQ: 593 case ISD::SETONE: 594 case ISD::SETOLE: 595 case ISD::SETOGE: 596 llvm_unreachable("Should be lowered by legalize!"); 597 default: llvm_unreachable("Unknown condition!"); 598 case ISD::SETOEQ: 599 case ISD::SETEQ: return PPC::PRED_EQ; 600 case ISD::SETUNE: 601 case ISD::SETNE: return PPC::PRED_NE; 602 case ISD::SETOLT: 603 case ISD::SETLT: return PPC::PRED_LT; 604 case ISD::SETULE: 605 case ISD::SETLE: return PPC::PRED_LE; 606 case ISD::SETOGT: 607 case ISD::SETGT: return PPC::PRED_GT; 608 case ISD::SETUGE: 609 case ISD::SETGE: return PPC::PRED_GE; 610 case ISD::SETO: return PPC::PRED_NU; 611 case ISD::SETUO: return PPC::PRED_UN; 612 // These two are invalid for floating point. Assume we have int. 613 case ISD::SETULT: return PPC::PRED_LT; 614 case ISD::SETUGT: return PPC::PRED_GT; 615 } 616 } 617 618 /// getCRIdxForSetCC - Return the index of the condition register field 619 /// associated with the SetCC condition, and whether or not the field is 620 /// treated as inverted. That is, lt = 0; ge = 0 inverted. 621 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) { 622 Invert = false; 623 switch (CC) { 624 default: llvm_unreachable("Unknown condition!"); 625 case ISD::SETOLT: 626 case ISD::SETLT: return 0; // Bit #0 = SETOLT 627 case ISD::SETOGT: 628 case ISD::SETGT: return 1; // Bit #1 = SETOGT 629 case ISD::SETOEQ: 630 case ISD::SETEQ: return 2; // Bit #2 = SETOEQ 631 case ISD::SETUO: return 3; // Bit #3 = SETUO 632 case ISD::SETUGE: 633 case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE 634 case ISD::SETULE: 635 case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE 636 case ISD::SETUNE: 637 case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE 638 case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO 639 case ISD::SETUEQ: 640 case ISD::SETOGE: 641 case ISD::SETOLE: 642 case ISD::SETONE: 643 llvm_unreachable("Invalid branch code: should be expanded by legalize"); 644 // These are invalid for floating point. Assume integer. 645 case ISD::SETULT: return 0; 646 case ISD::SETUGT: return 1; 647 } 648 } 649 650 // getVCmpInst: return the vector compare instruction for the specified 651 // vector type and condition code. Since this is for altivec specific code, 652 // only support the altivec types (v16i8, v8i16, v4i32, and v4f32). 653 static unsigned int getVCmpInst(MVT::SimpleValueType VecVT, ISD::CondCode CC, 654 bool HasVSX) { 655 switch (CC) { 656 case ISD::SETEQ: 657 case ISD::SETUEQ: 658 case ISD::SETNE: 659 case ISD::SETUNE: 660 if (VecVT == MVT::v16i8) 661 return PPC::VCMPEQUB; 662 else if (VecVT == MVT::v8i16) 663 return PPC::VCMPEQUH; 664 else if (VecVT == MVT::v4i32) 665 return PPC::VCMPEQUW; 666 // v4f32 != v4f32 could be translate to unordered not equal 667 else if (VecVT == MVT::v4f32) 668 return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP; 669 else if (VecVT == MVT::v2f64) 670 return PPC::XVCMPEQDP; 671 break; 672 case ISD::SETLT: 673 case ISD::SETGT: 674 case ISD::SETLE: 675 case ISD::SETGE: 676 if (VecVT == MVT::v16i8) 677 return PPC::VCMPGTSB; 678 else if (VecVT == MVT::v8i16) 679 return PPC::VCMPGTSH; 680 else if (VecVT == MVT::v4i32) 681 return PPC::VCMPGTSW; 682 else if (VecVT == MVT::v4f32) 683 return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP; 684 else if (VecVT == MVT::v2f64) 685 return PPC::XVCMPGTDP; 686 break; 687 case ISD::SETULT: 688 case ISD::SETUGT: 689 case ISD::SETUGE: 690 case ISD::SETULE: 691 if (VecVT == MVT::v16i8) 692 return PPC::VCMPGTUB; 693 else if (VecVT == MVT::v8i16) 694 return PPC::VCMPGTUH; 695 else if (VecVT == MVT::v4i32) 696 return PPC::VCMPGTUW; 697 break; 698 case ISD::SETOEQ: 699 if (VecVT == MVT::v4f32) 700 return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP; 701 else if (VecVT == MVT::v2f64) 702 return PPC::XVCMPEQDP; 703 break; 704 case ISD::SETOLT: 705 case ISD::SETOGT: 706 case ISD::SETOLE: 707 if (VecVT == MVT::v4f32) 708 return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP; 709 else if (VecVT == MVT::v2f64) 710 return PPC::XVCMPGTDP; 711 break; 712 case ISD::SETOGE: 713 if (VecVT == MVT::v4f32) 714 return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP; 715 else if (VecVT == MVT::v2f64) 716 return PPC::XVCMPGEDP; 717 break; 718 default: 719 break; 720 } 721 llvm_unreachable("Invalid integer vector compare condition"); 722 } 723 724 // getVCmpEQInst: return the equal compare instruction for the specified vector 725 // type. Since this is for altivec specific code, only support the altivec 726 // types (v16i8, v8i16, v4i32, and v4f32). 727 static unsigned int getVCmpEQInst(MVT::SimpleValueType VecVT, bool HasVSX) { 728 switch (VecVT) { 729 case MVT::v16i8: 730 return PPC::VCMPEQUB; 731 case MVT::v8i16: 732 return PPC::VCMPEQUH; 733 case MVT::v4i32: 734 return PPC::VCMPEQUW; 735 case MVT::v4f32: 736 return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP; 737 case MVT::v2f64: 738 return PPC::XVCMPEQDP; 739 default: 740 llvm_unreachable("Invalid integer vector compare condition"); 741 } 742 } 743 744 SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) { 745 SDLoc dl(N); 746 unsigned Imm; 747 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); 748 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy(); 749 bool isPPC64 = (PtrVT == MVT::i64); 750 751 if (!PPCSubTarget->useCRBits() && 752 isInt32Immediate(N->getOperand(1), Imm)) { 753 // We can codegen setcc op, imm very efficiently compared to a brcond. 754 // Check for those cases here. 755 // setcc op, 0 756 if (Imm == 0) { 757 SDValue Op = N->getOperand(0); 758 switch (CC) { 759 default: break; 760 case ISD::SETEQ: { 761 Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0); 762 SDValue Ops[] = { Op, getI32Imm(27), getI32Imm(5), getI32Imm(31) }; 763 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 764 } 765 case ISD::SETNE: { 766 if (isPPC64) break; 767 SDValue AD = 768 SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 769 Op, getI32Imm(~0U)), 0); 770 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, 771 AD.getValue(1)); 772 } 773 case ISD::SETLT: { 774 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 775 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 776 } 777 case ISD::SETGT: { 778 SDValue T = 779 SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0); 780 T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0); 781 SDValue Ops[] = { T, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 782 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 783 } 784 } 785 } else if (Imm == ~0U) { // setcc op, -1 786 SDValue Op = N->getOperand(0); 787 switch (CC) { 788 default: break; 789 case ISD::SETEQ: 790 if (isPPC64) break; 791 Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 792 Op, getI32Imm(1)), 0); 793 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, 794 SDValue(CurDAG->getMachineNode(PPC::LI, dl, 795 MVT::i32, 796 getI32Imm(0)), 0), 797 Op.getValue(1)); 798 case ISD::SETNE: { 799 if (isPPC64) break; 800 Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0); 801 SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 802 Op, getI32Imm(~0U)); 803 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0), 804 Op, SDValue(AD, 1)); 805 } 806 case ISD::SETLT: { 807 SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op, 808 getI32Imm(1)), 0); 809 SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD, 810 Op), 0); 811 SDValue Ops[] = { AN, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 812 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 813 } 814 case ISD::SETGT: { 815 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 816 Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 817 0); 818 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, 819 getI32Imm(1)); 820 } 821 } 822 } 823 } 824 825 SDValue LHS = N->getOperand(0); 826 SDValue RHS = N->getOperand(1); 827 828 // Altivec Vector compare instructions do not set any CR register by default and 829 // vector compare operations return the same type as the operands. 830 if (LHS.getValueType().isVector()) { 831 EVT VecVT = LHS.getValueType(); 832 MVT::SimpleValueType VT = VecVT.getSimpleVT().SimpleTy; 833 unsigned int VCmpInst = getVCmpInst(VT, CC, PPCSubTarget->hasVSX()); 834 835 switch (CC) { 836 case ISD::SETEQ: 837 case ISD::SETOEQ: 838 case ISD::SETUEQ: 839 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS); 840 case ISD::SETNE: 841 case ISD::SETONE: 842 case ISD::SETUNE: { 843 SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0); 844 return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR : 845 PPC::VNOR, 846 VecVT, VCmp, VCmp); 847 } 848 case ISD::SETLT: 849 case ISD::SETOLT: 850 case ISD::SETULT: 851 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, RHS, LHS); 852 case ISD::SETGT: 853 case ISD::SETOGT: 854 case ISD::SETUGT: 855 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS); 856 case ISD::SETGE: 857 case ISD::SETOGE: 858 case ISD::SETUGE: { 859 // Small optimization: Altivec provides a 'Vector Compare Greater Than 860 // or Equal To' instruction (vcmpgefp), so in this case there is no 861 // need for extra logic for the equal compare. 862 if (VecVT.getSimpleVT().isFloatingPoint()) { 863 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS); 864 } else { 865 SDValue VCmpGT(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0); 866 unsigned int VCmpEQInst = getVCmpEQInst(VT, PPCSubTarget->hasVSX()); 867 SDValue VCmpEQ(CurDAG->getMachineNode(VCmpEQInst, dl, VecVT, LHS, RHS), 0); 868 return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLOR : 869 PPC::VOR, 870 VecVT, VCmpGT, VCmpEQ); 871 } 872 } 873 case ISD::SETLE: 874 case ISD::SETOLE: 875 case ISD::SETULE: { 876 SDValue VCmpLE(CurDAG->getMachineNode(VCmpInst, dl, VecVT, RHS, LHS), 0); 877 unsigned int VCmpEQInst = getVCmpEQInst(VT, PPCSubTarget->hasVSX()); 878 SDValue VCmpEQ(CurDAG->getMachineNode(VCmpEQInst, dl, VecVT, LHS, RHS), 0); 879 return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLOR : 880 PPC::VOR, 881 VecVT, VCmpLE, VCmpEQ); 882 } 883 default: 884 llvm_unreachable("Invalid vector compare type: should be expanded by legalize"); 885 } 886 } 887 888 if (PPCSubTarget->useCRBits()) 889 return nullptr; 890 891 bool Inv; 892 unsigned Idx = getCRIdxForSetCC(CC, Inv); 893 SDValue CCReg = SelectCC(LHS, RHS, CC, dl); 894 SDValue IntCR; 895 896 // Force the ccreg into CR7. 897 SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32); 898 899 SDValue InFlag(nullptr, 0); // Null incoming flag value. 900 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg, 901 InFlag).getValue(1); 902 903 IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg, 904 CCReg), 0); 905 906 SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31), 907 getI32Imm(31), getI32Imm(31) }; 908 if (!Inv) 909 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 910 911 // Get the specified bit. 912 SDValue Tmp = 913 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); 914 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1)); 915 } 916 917 918 // Select - Convert the specified operand from a target-independent to a 919 // target-specific node if it hasn't already been changed. 920 SDNode *PPCDAGToDAGISel::Select(SDNode *N) { 921 SDLoc dl(N); 922 if (N->isMachineOpcode()) { 923 N->setNodeId(-1); 924 return nullptr; // Already selected. 925 } 926 927 switch (N->getOpcode()) { 928 default: break; 929 930 case ISD::Constant: { 931 if (N->getValueType(0) == MVT::i64) { 932 // Get 64 bit value. 933 int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue(); 934 // Assume no remaining bits. 935 unsigned Remainder = 0; 936 // Assume no shift required. 937 unsigned Shift = 0; 938 939 // If it can't be represented as a 32 bit value. 940 if (!isInt<32>(Imm)) { 941 Shift = countTrailingZeros<uint64_t>(Imm); 942 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift; 943 944 // If the shifted value fits 32 bits. 945 if (isInt<32>(ImmSh)) { 946 // Go with the shifted value. 947 Imm = ImmSh; 948 } else { 949 // Still stuck with a 64 bit value. 950 Remainder = Imm; 951 Shift = 32; 952 Imm >>= 32; 953 } 954 } 955 956 // Intermediate operand. 957 SDNode *Result; 958 959 // Handle first 32 bits. 960 unsigned Lo = Imm & 0xFFFF; 961 unsigned Hi = (Imm >> 16) & 0xFFFF; 962 963 // Simple value. 964 if (isInt<16>(Imm)) { 965 // Just the Lo bits. 966 Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo)); 967 } else if (Lo) { 968 // Handle the Hi bits. 969 unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8; 970 Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi)); 971 // And Lo bits. 972 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, 973 SDValue(Result, 0), getI32Imm(Lo)); 974 } else { 975 // Just the Hi bits. 976 Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi)); 977 } 978 979 // If no shift, we're done. 980 if (!Shift) return Result; 981 982 // Shift for next step if the upper 32-bits were not zero. 983 if (Imm) { 984 Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, 985 SDValue(Result, 0), 986 getI32Imm(Shift), 987 getI32Imm(63 - Shift)); 988 } 989 990 // Add in the last bits as required. 991 if ((Hi = (Remainder >> 16) & 0xFFFF)) { 992 Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64, 993 SDValue(Result, 0), getI32Imm(Hi)); 994 } 995 if ((Lo = Remainder & 0xFFFF)) { 996 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, 997 SDValue(Result, 0), getI32Imm(Lo)); 998 } 999 1000 return Result; 1001 } 1002 break; 1003 } 1004 1005 case ISD::SETCC: { 1006 SDNode *SN = SelectSETCC(N); 1007 if (SN) 1008 return SN; 1009 break; 1010 } 1011 case PPCISD::GlobalBaseReg: 1012 return getGlobalBaseReg(); 1013 1014 case ISD::FrameIndex: { 1015 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 1016 SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0)); 1017 unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8; 1018 if (N->hasOneUse()) 1019 return CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), TFI, 1020 getSmallIPtrImm(0)); 1021 return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI, 1022 getSmallIPtrImm(0)); 1023 } 1024 1025 case PPCISD::MFOCRF: { 1026 SDValue InFlag = N->getOperand(1); 1027 return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, 1028 N->getOperand(0), InFlag); 1029 } 1030 1031 case ISD::SDIV: { 1032 // FIXME: since this depends on the setting of the carry flag from the srawi 1033 // we should really be making notes about that for the scheduler. 1034 // FIXME: It sure would be nice if we could cheaply recognize the 1035 // srl/add/sra pattern the dag combiner will generate for this as 1036 // sra/addze rather than having to handle sdiv ourselves. oh well. 1037 unsigned Imm; 1038 if (isInt32Immediate(N->getOperand(1), Imm)) { 1039 SDValue N0 = N->getOperand(0); 1040 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) { 1041 SDNode *Op = 1042 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue, 1043 N0, getI32Imm(Log2_32(Imm))); 1044 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, 1045 SDValue(Op, 0), SDValue(Op, 1)); 1046 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) { 1047 SDNode *Op = 1048 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue, 1049 N0, getI32Imm(Log2_32(-Imm))); 1050 SDValue PT = 1051 SDValue(CurDAG->getMachineNode(PPC::ADDZE, dl, MVT::i32, 1052 SDValue(Op, 0), SDValue(Op, 1)), 1053 0); 1054 return CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT); 1055 } 1056 } 1057 1058 // Other cases are autogenerated. 1059 break; 1060 } 1061 1062 case ISD::LOAD: { 1063 // Handle preincrement loads. 1064 LoadSDNode *LD = cast<LoadSDNode>(N); 1065 EVT LoadedVT = LD->getMemoryVT(); 1066 1067 // Normal loads are handled by code generated from the .td file. 1068 if (LD->getAddressingMode() != ISD::PRE_INC) 1069 break; 1070 1071 SDValue Offset = LD->getOffset(); 1072 if (Offset.getOpcode() == ISD::TargetConstant || 1073 Offset.getOpcode() == ISD::TargetGlobalAddress) { 1074 1075 unsigned Opcode; 1076 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; 1077 if (LD->getValueType(0) != MVT::i64) { 1078 // Handle PPC32 integer and normal FP loads. 1079 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 1080 switch (LoadedVT.getSimpleVT().SimpleTy) { 1081 default: llvm_unreachable("Invalid PPC load type!"); 1082 case MVT::f64: Opcode = PPC::LFDU; break; 1083 case MVT::f32: Opcode = PPC::LFSU; break; 1084 case MVT::i32: Opcode = PPC::LWZU; break; 1085 case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break; 1086 case MVT::i1: 1087 case MVT::i8: Opcode = PPC::LBZU; break; 1088 } 1089 } else { 1090 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!"); 1091 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 1092 switch (LoadedVT.getSimpleVT().SimpleTy) { 1093 default: llvm_unreachable("Invalid PPC load type!"); 1094 case MVT::i64: Opcode = PPC::LDU; break; 1095 case MVT::i32: Opcode = PPC::LWZU8; break; 1096 case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break; 1097 case MVT::i1: 1098 case MVT::i8: Opcode = PPC::LBZU8; break; 1099 } 1100 } 1101 1102 SDValue Chain = LD->getChain(); 1103 SDValue Base = LD->getBasePtr(); 1104 SDValue Ops[] = { Offset, Base, Chain }; 1105 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0), 1106 PPCLowering->getPointerTy(), 1107 MVT::Other, Ops); 1108 } else { 1109 unsigned Opcode; 1110 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; 1111 if (LD->getValueType(0) != MVT::i64) { 1112 // Handle PPC32 integer and normal FP loads. 1113 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 1114 switch (LoadedVT.getSimpleVT().SimpleTy) { 1115 default: llvm_unreachable("Invalid PPC load type!"); 1116 case MVT::f64: Opcode = PPC::LFDUX; break; 1117 case MVT::f32: Opcode = PPC::LFSUX; break; 1118 case MVT::i32: Opcode = PPC::LWZUX; break; 1119 case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break; 1120 case MVT::i1: 1121 case MVT::i8: Opcode = PPC::LBZUX; break; 1122 } 1123 } else { 1124 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!"); 1125 assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) && 1126 "Invalid sext update load"); 1127 switch (LoadedVT.getSimpleVT().SimpleTy) { 1128 default: llvm_unreachable("Invalid PPC load type!"); 1129 case MVT::i64: Opcode = PPC::LDUX; break; 1130 case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break; 1131 case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break; 1132 case MVT::i1: 1133 case MVT::i8: Opcode = PPC::LBZUX8; break; 1134 } 1135 } 1136 1137 SDValue Chain = LD->getChain(); 1138 SDValue Base = LD->getBasePtr(); 1139 SDValue Ops[] = { Base, Offset, Chain }; 1140 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0), 1141 PPCLowering->getPointerTy(), 1142 MVT::Other, Ops); 1143 } 1144 } 1145 1146 case ISD::AND: { 1147 unsigned Imm, Imm2, SH, MB, ME; 1148 uint64_t Imm64; 1149 1150 // If this is an and of a value rotated between 0 and 31 bits and then and'd 1151 // with a mask, emit rlwinm 1152 if (isInt32Immediate(N->getOperand(1), Imm) && 1153 isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) { 1154 SDValue Val = N->getOperand(0).getOperand(0); 1155 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; 1156 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 1157 } 1158 // If this is just a masked value where the input is not handled above, and 1159 // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm 1160 if (isInt32Immediate(N->getOperand(1), Imm) && 1161 isRunOfOnes(Imm, MB, ME) && 1162 N->getOperand(0).getOpcode() != ISD::ROTL) { 1163 SDValue Val = N->getOperand(0); 1164 SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB), getI32Imm(ME) }; 1165 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 1166 } 1167 // If this is a 64-bit zero-extension mask, emit rldicl. 1168 if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) && 1169 isMask_64(Imm64)) { 1170 SDValue Val = N->getOperand(0); 1171 MB = 64 - CountTrailingOnes_64(Imm64); 1172 SH = 0; 1173 1174 // If the operand is a logical right shift, we can fold it into this 1175 // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb) 1176 // for n <= mb. The right shift is really a left rotate followed by a 1177 // mask, and this mask is a more-restrictive sub-mask of the mask implied 1178 // by the shift. 1179 if (Val.getOpcode() == ISD::SRL && 1180 isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) { 1181 assert(Imm < 64 && "Illegal shift amount"); 1182 Val = Val.getOperand(0); 1183 SH = 64 - Imm; 1184 } 1185 1186 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB) }; 1187 return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops); 1188 } 1189 // AND X, 0 -> 0, not "rlwinm 32". 1190 if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) { 1191 ReplaceUses(SDValue(N, 0), N->getOperand(1)); 1192 return nullptr; 1193 } 1194 // ISD::OR doesn't get all the bitfield insertion fun. 1195 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert 1196 if (isInt32Immediate(N->getOperand(1), Imm) && 1197 N->getOperand(0).getOpcode() == ISD::OR && 1198 isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) { 1199 unsigned MB, ME; 1200 Imm = ~(Imm^Imm2); 1201 if (isRunOfOnes(Imm, MB, ME)) { 1202 SDValue Ops[] = { N->getOperand(0).getOperand(0), 1203 N->getOperand(0).getOperand(1), 1204 getI32Imm(0), getI32Imm(MB),getI32Imm(ME) }; 1205 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); 1206 } 1207 } 1208 1209 // Other cases are autogenerated. 1210 break; 1211 } 1212 case ISD::OR: 1213 if (N->getValueType(0) == MVT::i32) 1214 if (SDNode *I = SelectBitfieldInsert(N)) 1215 return I; 1216 1217 // Other cases are autogenerated. 1218 break; 1219 case ISD::SHL: { 1220 unsigned Imm, SH, MB, ME; 1221 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && 1222 isRotateAndMask(N, Imm, true, SH, MB, ME)) { 1223 SDValue Ops[] = { N->getOperand(0).getOperand(0), 1224 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; 1225 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 1226 } 1227 1228 // Other cases are autogenerated. 1229 break; 1230 } 1231 case ISD::SRL: { 1232 unsigned Imm, SH, MB, ME; 1233 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && 1234 isRotateAndMask(N, Imm, true, SH, MB, ME)) { 1235 SDValue Ops[] = { N->getOperand(0).getOperand(0), 1236 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; 1237 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 1238 } 1239 1240 // Other cases are autogenerated. 1241 break; 1242 } 1243 // FIXME: Remove this once the ANDI glue bug is fixed: 1244 case PPCISD::ANDIo_1_EQ_BIT: 1245 case PPCISD::ANDIo_1_GT_BIT: { 1246 if (!ANDIGlueBug) 1247 break; 1248 1249 EVT InVT = N->getOperand(0).getValueType(); 1250 assert((InVT == MVT::i64 || InVT == MVT::i32) && 1251 "Invalid input type for ANDIo_1_EQ_BIT"); 1252 1253 unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo; 1254 SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue, 1255 N->getOperand(0), 1256 CurDAG->getTargetConstant(1, InVT)), 0); 1257 SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32); 1258 SDValue SRIdxVal = 1259 CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ? 1260 PPC::sub_eq : PPC::sub_gt, MVT::i32); 1261 1262 return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1, 1263 CR0Reg, SRIdxVal, 1264 SDValue(AndI.getNode(), 1) /* glue */); 1265 } 1266 case ISD::SELECT_CC: { 1267 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get(); 1268 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy(); 1269 bool isPPC64 = (PtrVT == MVT::i64); 1270 1271 // If this is a select of i1 operands, we'll pattern match it. 1272 if (PPCSubTarget->useCRBits() && 1273 N->getOperand(0).getValueType() == MVT::i1) 1274 break; 1275 1276 // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc 1277 if (!isPPC64) 1278 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1))) 1279 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2))) 1280 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3))) 1281 if (N1C->isNullValue() && N3C->isNullValue() && 1282 N2C->getZExtValue() == 1ULL && CC == ISD::SETNE && 1283 // FIXME: Implement this optzn for PPC64. 1284 N->getValueType(0) == MVT::i32) { 1285 SDNode *Tmp = 1286 CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 1287 N->getOperand(0), getI32Imm(~0U)); 1288 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, 1289 SDValue(Tmp, 0), N->getOperand(0), 1290 SDValue(Tmp, 1)); 1291 } 1292 1293 SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl); 1294 1295 if (N->getValueType(0) == MVT::i1) { 1296 // An i1 select is: (c & t) | (!c & f). 1297 bool Inv; 1298 unsigned Idx = getCRIdxForSetCC(CC, Inv); 1299 1300 unsigned SRI; 1301 switch (Idx) { 1302 default: llvm_unreachable("Invalid CC index"); 1303 case 0: SRI = PPC::sub_lt; break; 1304 case 1: SRI = PPC::sub_gt; break; 1305 case 2: SRI = PPC::sub_eq; break; 1306 case 3: SRI = PPC::sub_un; break; 1307 } 1308 1309 SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg); 1310 1311 SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1, 1312 CCBit, CCBit), 0); 1313 SDValue C = Inv ? NotCCBit : CCBit, 1314 NotC = Inv ? CCBit : NotCCBit; 1315 1316 SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1, 1317 C, N->getOperand(2)), 0); 1318 SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1, 1319 NotC, N->getOperand(3)), 0); 1320 1321 return CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF); 1322 } 1323 1324 unsigned BROpc = getPredicateForSetCC(CC); 1325 1326 unsigned SelectCCOp; 1327 if (N->getValueType(0) == MVT::i32) 1328 SelectCCOp = PPC::SELECT_CC_I4; 1329 else if (N->getValueType(0) == MVT::i64) 1330 SelectCCOp = PPC::SELECT_CC_I8; 1331 else if (N->getValueType(0) == MVT::f32) 1332 SelectCCOp = PPC::SELECT_CC_F4; 1333 else if (N->getValueType(0) == MVT::f64) 1334 SelectCCOp = PPC::SELECT_CC_F8; 1335 else 1336 SelectCCOp = PPC::SELECT_CC_VRRC; 1337 1338 SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3), 1339 getI32Imm(BROpc) }; 1340 return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops); 1341 } 1342 case ISD::VSELECT: 1343 if (PPCSubTarget->hasVSX()) { 1344 SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) }; 1345 return CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops); 1346 } 1347 1348 break; 1349 case ISD::VECTOR_SHUFFLE: 1350 if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 || 1351 N->getValueType(0) == MVT::v2i64)) { 1352 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 1353 1354 SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1), 1355 Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1); 1356 unsigned DM[2]; 1357 1358 for (int i = 0; i < 2; ++i) 1359 if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2) 1360 DM[i] = 0; 1361 else 1362 DM[i] = 1; 1363 1364 SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), MVT::i32); 1365 1366 if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 && 1367 Op1.getOpcode() == ISD::SCALAR_TO_VECTOR && 1368 isa<LoadSDNode>(Op1.getOperand(0))) { 1369 LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0)); 1370 SDValue Base, Offset; 1371 1372 if (LD->isUnindexed() && 1373 SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) { 1374 SDValue Chain = LD->getChain(); 1375 SDValue Ops[] = { Base, Offset, Chain }; 1376 return CurDAG->SelectNodeTo(N, PPC::LXVDSX, 1377 N->getValueType(0), Ops); 1378 } 1379 } 1380 1381 SDValue Ops[] = { Op1, Op2, DMV }; 1382 return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops); 1383 } 1384 1385 break; 1386 case PPCISD::BDNZ: 1387 case PPCISD::BDZ: { 1388 bool IsPPC64 = PPCSubTarget->isPPC64(); 1389 SDValue Ops[] = { N->getOperand(1), N->getOperand(0) }; 1390 return CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ ? 1391 (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 1392 (IsPPC64 ? PPC::BDZ8 : PPC::BDZ), 1393 MVT::Other, Ops); 1394 } 1395 case PPCISD::COND_BRANCH: { 1396 // Op #0 is the Chain. 1397 // Op #1 is the PPC::PRED_* number. 1398 // Op #2 is the CR# 1399 // Op #3 is the Dest MBB 1400 // Op #4 is the Flag. 1401 // Prevent PPC::PRED_* from being selected into LI. 1402 SDValue Pred = 1403 getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()); 1404 SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3), 1405 N->getOperand(0), N->getOperand(4) }; 1406 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); 1407 } 1408 case ISD::BR_CC: { 1409 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get(); 1410 unsigned PCC = getPredicateForSetCC(CC); 1411 1412 if (N->getOperand(2).getValueType() == MVT::i1) { 1413 unsigned Opc; 1414 bool Swap; 1415 switch (PCC) { 1416 default: llvm_unreachable("Unexpected Boolean-operand predicate"); 1417 case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break; 1418 case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break; 1419 case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break; 1420 case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break; 1421 case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break; 1422 case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break; 1423 } 1424 1425 SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1, 1426 N->getOperand(Swap ? 3 : 2), 1427 N->getOperand(Swap ? 2 : 3)), 0); 1428 return CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other, 1429 BitComp, N->getOperand(4), N->getOperand(0)); 1430 } 1431 1432 SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl); 1433 SDValue Ops[] = { getI32Imm(PCC), CondCode, 1434 N->getOperand(4), N->getOperand(0) }; 1435 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); 1436 } 1437 case ISD::BRIND: { 1438 // FIXME: Should custom lower this. 1439 SDValue Chain = N->getOperand(0); 1440 SDValue Target = N->getOperand(1); 1441 unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8; 1442 unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8; 1443 Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target, 1444 Chain), 0); 1445 return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain); 1446 } 1447 case PPCISD::TOC_ENTRY: { 1448 assert (PPCSubTarget->isPPC64() && "Only supported for 64-bit ABI"); 1449 1450 // For medium and large code model, we generate two instructions as 1451 // described below. Otherwise we allow SelectCodeCommon to handle this, 1452 // selecting one of LDtoc, LDtocJTI, and LDtocCPT. 1453 CodeModel::Model CModel = TM.getCodeModel(); 1454 if (CModel != CodeModel::Medium && CModel != CodeModel::Large) 1455 break; 1456 1457 // The first source operand is a TargetGlobalAddress or a TargetJumpTable. 1458 // If it is an externally defined symbol, a symbol with common linkage, 1459 // a non-local function address, or a jump table address, or if we are 1460 // generating code for large code model, we generate: 1461 // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>)) 1462 // Otherwise we generate: 1463 // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>) 1464 SDValue GA = N->getOperand(0); 1465 SDValue TOCbase = N->getOperand(1); 1466 SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64, 1467 TOCbase, GA); 1468 1469 if (isa<JumpTableSDNode>(GA) || CModel == CodeModel::Large) 1470 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA, 1471 SDValue(Tmp, 0)); 1472 1473 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) { 1474 const GlobalValue *GValue = G->getGlobal(); 1475 if ((GValue->getType()->getElementType()->isFunctionTy() && 1476 (GValue->isDeclaration() || GValue->isWeakForLinker())) || 1477 GValue->isDeclaration() || GValue->hasCommonLinkage() || 1478 GValue->hasAvailableExternallyLinkage()) 1479 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA, 1480 SDValue(Tmp, 0)); 1481 } 1482 1483 return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64, 1484 SDValue(Tmp, 0), GA); 1485 } 1486 case PPCISD::VADD_SPLAT: { 1487 // This expands into one of three sequences, depending on whether 1488 // the first operand is odd or even, positive or negative. 1489 assert(isa<ConstantSDNode>(N->getOperand(0)) && 1490 isa<ConstantSDNode>(N->getOperand(1)) && 1491 "Invalid operand on VADD_SPLAT!"); 1492 1493 int Elt = N->getConstantOperandVal(0); 1494 int EltSize = N->getConstantOperandVal(1); 1495 unsigned Opc1, Opc2, Opc3; 1496 EVT VT; 1497 1498 if (EltSize == 1) { 1499 Opc1 = PPC::VSPLTISB; 1500 Opc2 = PPC::VADDUBM; 1501 Opc3 = PPC::VSUBUBM; 1502 VT = MVT::v16i8; 1503 } else if (EltSize == 2) { 1504 Opc1 = PPC::VSPLTISH; 1505 Opc2 = PPC::VADDUHM; 1506 Opc3 = PPC::VSUBUHM; 1507 VT = MVT::v8i16; 1508 } else { 1509 assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!"); 1510 Opc1 = PPC::VSPLTISW; 1511 Opc2 = PPC::VADDUWM; 1512 Opc3 = PPC::VSUBUWM; 1513 VT = MVT::v4i32; 1514 } 1515 1516 if ((Elt & 1) == 0) { 1517 // Elt is even, in the range [-32,-18] + [16,30]. 1518 // 1519 // Convert: VADD_SPLAT elt, size 1520 // Into: tmp = VSPLTIS[BHW] elt 1521 // VADDU[BHW]M tmp, tmp 1522 // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4 1523 SDValue EltVal = getI32Imm(Elt >> 1); 1524 SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 1525 SDValue TmpVal = SDValue(Tmp, 0); 1526 return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal); 1527 1528 } else if (Elt > 0) { 1529 // Elt is odd and positive, in the range [17,31]. 1530 // 1531 // Convert: VADD_SPLAT elt, size 1532 // Into: tmp1 = VSPLTIS[BHW] elt-16 1533 // tmp2 = VSPLTIS[BHW] -16 1534 // VSUBU[BHW]M tmp1, tmp2 1535 SDValue EltVal = getI32Imm(Elt - 16); 1536 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 1537 EltVal = getI32Imm(-16); 1538 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 1539 return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0), 1540 SDValue(Tmp2, 0)); 1541 1542 } else { 1543 // Elt is odd and negative, in the range [-31,-17]. 1544 // 1545 // Convert: VADD_SPLAT elt, size 1546 // Into: tmp1 = VSPLTIS[BHW] elt+16 1547 // tmp2 = VSPLTIS[BHW] -16 1548 // VADDU[BHW]M tmp1, tmp2 1549 SDValue EltVal = getI32Imm(Elt + 16); 1550 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 1551 EltVal = getI32Imm(-16); 1552 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 1553 return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0), 1554 SDValue(Tmp2, 0)); 1555 } 1556 } 1557 } 1558 1559 return SelectCode(N); 1560 } 1561 1562 /// PostprocessISelDAG - Perform some late peephole optimizations 1563 /// on the DAG representation. 1564 void PPCDAGToDAGISel::PostprocessISelDAG() { 1565 1566 // Skip peepholes at -O0. 1567 if (TM.getOptLevel() == CodeGenOpt::None) 1568 return; 1569 1570 PeepholePPC64(); 1571 PeepholeCROps(); 1572 } 1573 1574 // Check if all users of this node will become isel where the second operand 1575 // is the constant zero. If this is so, and if we can negate the condition, 1576 // then we can flip the true and false operands. This will allow the zero to 1577 // be folded with the isel so that we don't need to materialize a register 1578 // containing zero. 1579 bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) { 1580 // If we're not using isel, then this does not matter. 1581 if (!PPCSubTarget->hasISEL()) 1582 return false; 1583 1584 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 1585 UI != UE; ++UI) { 1586 SDNode *User = *UI; 1587 if (!User->isMachineOpcode()) 1588 return false; 1589 if (User->getMachineOpcode() != PPC::SELECT_I4 && 1590 User->getMachineOpcode() != PPC::SELECT_I8) 1591 return false; 1592 1593 SDNode *Op2 = User->getOperand(2).getNode(); 1594 if (!Op2->isMachineOpcode()) 1595 return false; 1596 1597 if (Op2->getMachineOpcode() != PPC::LI && 1598 Op2->getMachineOpcode() != PPC::LI8) 1599 return false; 1600 1601 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0)); 1602 if (!C) 1603 return false; 1604 1605 if (!C->isNullValue()) 1606 return false; 1607 } 1608 1609 return true; 1610 } 1611 1612 void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) { 1613 SmallVector<SDNode *, 4> ToReplace; 1614 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 1615 UI != UE; ++UI) { 1616 SDNode *User = *UI; 1617 assert((User->getMachineOpcode() == PPC::SELECT_I4 || 1618 User->getMachineOpcode() == PPC::SELECT_I8) && 1619 "Must have all select users"); 1620 ToReplace.push_back(User); 1621 } 1622 1623 for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(), 1624 UE = ToReplace.end(); UI != UE; ++UI) { 1625 SDNode *User = *UI; 1626 SDNode *ResNode = 1627 CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User), 1628 User->getValueType(0), User->getOperand(0), 1629 User->getOperand(2), 1630 User->getOperand(1)); 1631 1632 DEBUG(dbgs() << "CR Peephole replacing:\nOld: "); 1633 DEBUG(User->dump(CurDAG)); 1634 DEBUG(dbgs() << "\nNew: "); 1635 DEBUG(ResNode->dump(CurDAG)); 1636 DEBUG(dbgs() << "\n"); 1637 1638 ReplaceUses(User, ResNode); 1639 } 1640 } 1641 1642 void PPCDAGToDAGISel::PeepholeCROps() { 1643 bool IsModified; 1644 do { 1645 IsModified = false; 1646 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(), 1647 E = CurDAG->allnodes_end(); I != E; ++I) { 1648 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I); 1649 if (!MachineNode || MachineNode->use_empty()) 1650 continue; 1651 SDNode *ResNode = MachineNode; 1652 1653 bool Op1Set = false, Op1Unset = false, 1654 Op1Not = false, 1655 Op2Set = false, Op2Unset = false, 1656 Op2Not = false; 1657 1658 unsigned Opcode = MachineNode->getMachineOpcode(); 1659 switch (Opcode) { 1660 default: break; 1661 case PPC::CRAND: 1662 case PPC::CRNAND: 1663 case PPC::CROR: 1664 case PPC::CRXOR: 1665 case PPC::CRNOR: 1666 case PPC::CREQV: 1667 case PPC::CRANDC: 1668 case PPC::CRORC: { 1669 SDValue Op = MachineNode->getOperand(1); 1670 if (Op.isMachineOpcode()) { 1671 if (Op.getMachineOpcode() == PPC::CRSET) 1672 Op2Set = true; 1673 else if (Op.getMachineOpcode() == PPC::CRUNSET) 1674 Op2Unset = true; 1675 else if (Op.getMachineOpcode() == PPC::CRNOR && 1676 Op.getOperand(0) == Op.getOperand(1)) 1677 Op2Not = true; 1678 } 1679 } // fallthrough 1680 case PPC::BC: 1681 case PPC::BCn: 1682 case PPC::SELECT_I4: 1683 case PPC::SELECT_I8: 1684 case PPC::SELECT_F4: 1685 case PPC::SELECT_F8: 1686 case PPC::SELECT_VRRC: { 1687 SDValue Op = MachineNode->getOperand(0); 1688 if (Op.isMachineOpcode()) { 1689 if (Op.getMachineOpcode() == PPC::CRSET) 1690 Op1Set = true; 1691 else if (Op.getMachineOpcode() == PPC::CRUNSET) 1692 Op1Unset = true; 1693 else if (Op.getMachineOpcode() == PPC::CRNOR && 1694 Op.getOperand(0) == Op.getOperand(1)) 1695 Op1Not = true; 1696 } 1697 } 1698 break; 1699 } 1700 1701 bool SelectSwap = false; 1702 switch (Opcode) { 1703 default: break; 1704 case PPC::CRAND: 1705 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1706 // x & x = x 1707 ResNode = MachineNode->getOperand(0).getNode(); 1708 else if (Op1Set) 1709 // 1 & y = y 1710 ResNode = MachineNode->getOperand(1).getNode(); 1711 else if (Op2Set) 1712 // x & 1 = x 1713 ResNode = MachineNode->getOperand(0).getNode(); 1714 else if (Op1Unset || Op2Unset) 1715 // x & 0 = 0 & y = 0 1716 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 1717 MVT::i1); 1718 else if (Op1Not) 1719 // ~x & y = andc(y, x) 1720 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 1721 MVT::i1, MachineNode->getOperand(1), 1722 MachineNode->getOperand(0). 1723 getOperand(0)); 1724 else if (Op2Not) 1725 // x & ~y = andc(x, y) 1726 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 1727 MVT::i1, MachineNode->getOperand(0), 1728 MachineNode->getOperand(1). 1729 getOperand(0)); 1730 else if (AllUsersSelectZero(MachineNode)) 1731 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode), 1732 MVT::i1, MachineNode->getOperand(0), 1733 MachineNode->getOperand(1)), 1734 SelectSwap = true; 1735 break; 1736 case PPC::CRNAND: 1737 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1738 // nand(x, x) -> nor(x, x) 1739 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1740 MVT::i1, MachineNode->getOperand(0), 1741 MachineNode->getOperand(0)); 1742 else if (Op1Set) 1743 // nand(1, y) -> nor(y, y) 1744 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1745 MVT::i1, MachineNode->getOperand(1), 1746 MachineNode->getOperand(1)); 1747 else if (Op2Set) 1748 // nand(x, 1) -> nor(x, x) 1749 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1750 MVT::i1, MachineNode->getOperand(0), 1751 MachineNode->getOperand(0)); 1752 else if (Op1Unset || Op2Unset) 1753 // nand(x, 0) = nand(0, y) = 1 1754 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 1755 MVT::i1); 1756 else if (Op1Not) 1757 // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y) 1758 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 1759 MVT::i1, MachineNode->getOperand(0). 1760 getOperand(0), 1761 MachineNode->getOperand(1)); 1762 else if (Op2Not) 1763 // nand(x, ~y) = ~x | y = orc(y, x) 1764 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 1765 MVT::i1, MachineNode->getOperand(1). 1766 getOperand(0), 1767 MachineNode->getOperand(0)); 1768 else if (AllUsersSelectZero(MachineNode)) 1769 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode), 1770 MVT::i1, MachineNode->getOperand(0), 1771 MachineNode->getOperand(1)), 1772 SelectSwap = true; 1773 break; 1774 case PPC::CROR: 1775 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1776 // x | x = x 1777 ResNode = MachineNode->getOperand(0).getNode(); 1778 else if (Op1Set || Op2Set) 1779 // x | 1 = 1 | y = 1 1780 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 1781 MVT::i1); 1782 else if (Op1Unset) 1783 // 0 | y = y 1784 ResNode = MachineNode->getOperand(1).getNode(); 1785 else if (Op2Unset) 1786 // x | 0 = x 1787 ResNode = MachineNode->getOperand(0).getNode(); 1788 else if (Op1Not) 1789 // ~x | y = orc(y, x) 1790 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 1791 MVT::i1, MachineNode->getOperand(1), 1792 MachineNode->getOperand(0). 1793 getOperand(0)); 1794 else if (Op2Not) 1795 // x | ~y = orc(x, y) 1796 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 1797 MVT::i1, MachineNode->getOperand(0), 1798 MachineNode->getOperand(1). 1799 getOperand(0)); 1800 else if (AllUsersSelectZero(MachineNode)) 1801 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1802 MVT::i1, MachineNode->getOperand(0), 1803 MachineNode->getOperand(1)), 1804 SelectSwap = true; 1805 break; 1806 case PPC::CRXOR: 1807 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1808 // xor(x, x) = 0 1809 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 1810 MVT::i1); 1811 else if (Op1Set) 1812 // xor(1, y) -> nor(y, y) 1813 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1814 MVT::i1, MachineNode->getOperand(1), 1815 MachineNode->getOperand(1)); 1816 else if (Op2Set) 1817 // xor(x, 1) -> nor(x, x) 1818 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1819 MVT::i1, MachineNode->getOperand(0), 1820 MachineNode->getOperand(0)); 1821 else if (Op1Unset) 1822 // xor(0, y) = y 1823 ResNode = MachineNode->getOperand(1).getNode(); 1824 else if (Op2Unset) 1825 // xor(x, 0) = x 1826 ResNode = MachineNode->getOperand(0).getNode(); 1827 else if (Op1Not) 1828 // xor(~x, y) = eqv(x, y) 1829 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 1830 MVT::i1, MachineNode->getOperand(0). 1831 getOperand(0), 1832 MachineNode->getOperand(1)); 1833 else if (Op2Not) 1834 // xor(x, ~y) = eqv(x, y) 1835 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 1836 MVT::i1, MachineNode->getOperand(0), 1837 MachineNode->getOperand(1). 1838 getOperand(0)); 1839 else if (AllUsersSelectZero(MachineNode)) 1840 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 1841 MVT::i1, MachineNode->getOperand(0), 1842 MachineNode->getOperand(1)), 1843 SelectSwap = true; 1844 break; 1845 case PPC::CRNOR: 1846 if (Op1Set || Op2Set) 1847 // nor(1, y) -> 0 1848 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 1849 MVT::i1); 1850 else if (Op1Unset) 1851 // nor(0, y) = ~y -> nor(y, y) 1852 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1853 MVT::i1, MachineNode->getOperand(1), 1854 MachineNode->getOperand(1)); 1855 else if (Op2Unset) 1856 // nor(x, 0) = ~x 1857 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1858 MVT::i1, MachineNode->getOperand(0), 1859 MachineNode->getOperand(0)); 1860 else if (Op1Not) 1861 // nor(~x, y) = andc(x, y) 1862 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 1863 MVT::i1, MachineNode->getOperand(0). 1864 getOperand(0), 1865 MachineNode->getOperand(1)); 1866 else if (Op2Not) 1867 // nor(x, ~y) = andc(y, x) 1868 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 1869 MVT::i1, MachineNode->getOperand(1). 1870 getOperand(0), 1871 MachineNode->getOperand(0)); 1872 else if (AllUsersSelectZero(MachineNode)) 1873 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode), 1874 MVT::i1, MachineNode->getOperand(0), 1875 MachineNode->getOperand(1)), 1876 SelectSwap = true; 1877 break; 1878 case PPC::CREQV: 1879 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1880 // eqv(x, x) = 1 1881 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 1882 MVT::i1); 1883 else if (Op1Set) 1884 // eqv(1, y) = y 1885 ResNode = MachineNode->getOperand(1).getNode(); 1886 else if (Op2Set) 1887 // eqv(x, 1) = x 1888 ResNode = MachineNode->getOperand(0).getNode(); 1889 else if (Op1Unset) 1890 // eqv(0, y) = ~y -> nor(y, y) 1891 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1892 MVT::i1, MachineNode->getOperand(1), 1893 MachineNode->getOperand(1)); 1894 else if (Op2Unset) 1895 // eqv(x, 0) = ~x 1896 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1897 MVT::i1, MachineNode->getOperand(0), 1898 MachineNode->getOperand(0)); 1899 else if (Op1Not) 1900 // eqv(~x, y) = xor(x, y) 1901 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 1902 MVT::i1, MachineNode->getOperand(0). 1903 getOperand(0), 1904 MachineNode->getOperand(1)); 1905 else if (Op2Not) 1906 // eqv(x, ~y) = xor(x, y) 1907 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 1908 MVT::i1, MachineNode->getOperand(0), 1909 MachineNode->getOperand(1). 1910 getOperand(0)); 1911 else if (AllUsersSelectZero(MachineNode)) 1912 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 1913 MVT::i1, MachineNode->getOperand(0), 1914 MachineNode->getOperand(1)), 1915 SelectSwap = true; 1916 break; 1917 case PPC::CRANDC: 1918 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1919 // andc(x, x) = 0 1920 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 1921 MVT::i1); 1922 else if (Op1Set) 1923 // andc(1, y) = ~y 1924 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1925 MVT::i1, MachineNode->getOperand(1), 1926 MachineNode->getOperand(1)); 1927 else if (Op1Unset || Op2Set) 1928 // andc(0, y) = andc(x, 1) = 0 1929 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 1930 MVT::i1); 1931 else if (Op2Unset) 1932 // andc(x, 0) = x 1933 ResNode = MachineNode->getOperand(0).getNode(); 1934 else if (Op1Not) 1935 // andc(~x, y) = ~(x | y) = nor(x, y) 1936 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1937 MVT::i1, MachineNode->getOperand(0). 1938 getOperand(0), 1939 MachineNode->getOperand(1)); 1940 else if (Op2Not) 1941 // andc(x, ~y) = x & y 1942 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode), 1943 MVT::i1, MachineNode->getOperand(0), 1944 MachineNode->getOperand(1). 1945 getOperand(0)); 1946 else if (AllUsersSelectZero(MachineNode)) 1947 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 1948 MVT::i1, MachineNode->getOperand(1), 1949 MachineNode->getOperand(0)), 1950 SelectSwap = true; 1951 break; 1952 case PPC::CRORC: 1953 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 1954 // orc(x, x) = 1 1955 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 1956 MVT::i1); 1957 else if (Op1Set || Op2Unset) 1958 // orc(1, y) = orc(x, 0) = 1 1959 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 1960 MVT::i1); 1961 else if (Op2Set) 1962 // orc(x, 1) = x 1963 ResNode = MachineNode->getOperand(0).getNode(); 1964 else if (Op1Unset) 1965 // orc(0, y) = ~y 1966 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 1967 MVT::i1, MachineNode->getOperand(1), 1968 MachineNode->getOperand(1)); 1969 else if (Op1Not) 1970 // orc(~x, y) = ~(x & y) = nand(x, y) 1971 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode), 1972 MVT::i1, MachineNode->getOperand(0). 1973 getOperand(0), 1974 MachineNode->getOperand(1)); 1975 else if (Op2Not) 1976 // orc(x, ~y) = x | y 1977 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode), 1978 MVT::i1, MachineNode->getOperand(0), 1979 MachineNode->getOperand(1). 1980 getOperand(0)); 1981 else if (AllUsersSelectZero(MachineNode)) 1982 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 1983 MVT::i1, MachineNode->getOperand(1), 1984 MachineNode->getOperand(0)), 1985 SelectSwap = true; 1986 break; 1987 case PPC::SELECT_I4: 1988 case PPC::SELECT_I8: 1989 case PPC::SELECT_F4: 1990 case PPC::SELECT_F8: 1991 case PPC::SELECT_VRRC: 1992 if (Op1Set) 1993 ResNode = MachineNode->getOperand(1).getNode(); 1994 else if (Op1Unset) 1995 ResNode = MachineNode->getOperand(2).getNode(); 1996 else if (Op1Not) 1997 ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(), 1998 SDLoc(MachineNode), 1999 MachineNode->getValueType(0), 2000 MachineNode->getOperand(0). 2001 getOperand(0), 2002 MachineNode->getOperand(2), 2003 MachineNode->getOperand(1)); 2004 break; 2005 case PPC::BC: 2006 case PPC::BCn: 2007 if (Op1Not) 2008 ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn : 2009 PPC::BC, 2010 SDLoc(MachineNode), 2011 MVT::Other, 2012 MachineNode->getOperand(0). 2013 getOperand(0), 2014 MachineNode->getOperand(1), 2015 MachineNode->getOperand(2)); 2016 // FIXME: Handle Op1Set, Op1Unset here too. 2017 break; 2018 } 2019 2020 // If we're inverting this node because it is used only by selects that 2021 // we'd like to swap, then swap the selects before the node replacement. 2022 if (SelectSwap) 2023 SwapAllSelectUsers(MachineNode); 2024 2025 if (ResNode != MachineNode) { 2026 DEBUG(dbgs() << "CR Peephole replacing:\nOld: "); 2027 DEBUG(MachineNode->dump(CurDAG)); 2028 DEBUG(dbgs() << "\nNew: "); 2029 DEBUG(ResNode->dump(CurDAG)); 2030 DEBUG(dbgs() << "\n"); 2031 2032 ReplaceUses(MachineNode, ResNode); 2033 IsModified = true; 2034 } 2035 } 2036 if (IsModified) 2037 CurDAG->RemoveDeadNodes(); 2038 } while (IsModified); 2039 } 2040 2041 void PPCDAGToDAGISel::PeepholePPC64() { 2042 // These optimizations are currently supported only for 64-bit SVR4. 2043 if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64()) 2044 return; 2045 2046 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode()); 2047 ++Position; 2048 2049 while (Position != CurDAG->allnodes_begin()) { 2050 SDNode *N = --Position; 2051 // Skip dead nodes and any non-machine opcodes. 2052 if (N->use_empty() || !N->isMachineOpcode()) 2053 continue; 2054 2055 unsigned FirstOp; 2056 unsigned StorageOpcode = N->getMachineOpcode(); 2057 2058 switch (StorageOpcode) { 2059 default: continue; 2060 2061 case PPC::LBZ: 2062 case PPC::LBZ8: 2063 case PPC::LD: 2064 case PPC::LFD: 2065 case PPC::LFS: 2066 case PPC::LHA: 2067 case PPC::LHA8: 2068 case PPC::LHZ: 2069 case PPC::LHZ8: 2070 case PPC::LWA: 2071 case PPC::LWZ: 2072 case PPC::LWZ8: 2073 FirstOp = 0; 2074 break; 2075 2076 case PPC::STB: 2077 case PPC::STB8: 2078 case PPC::STD: 2079 case PPC::STFD: 2080 case PPC::STFS: 2081 case PPC::STH: 2082 case PPC::STH8: 2083 case PPC::STW: 2084 case PPC::STW8: 2085 FirstOp = 1; 2086 break; 2087 } 2088 2089 // If this is a load or store with a zero offset, we may be able to 2090 // fold an add-immediate into the memory operation. 2091 if (!isa<ConstantSDNode>(N->getOperand(FirstOp)) || 2092 N->getConstantOperandVal(FirstOp) != 0) 2093 continue; 2094 2095 SDValue Base = N->getOperand(FirstOp + 1); 2096 if (!Base.isMachineOpcode()) 2097 continue; 2098 2099 unsigned Flags = 0; 2100 bool ReplaceFlags = true; 2101 2102 // When the feeding operation is an add-immediate of some sort, 2103 // determine whether we need to add relocation information to the 2104 // target flags on the immediate operand when we fold it into the 2105 // load instruction. 2106 // 2107 // For something like ADDItocL, the relocation information is 2108 // inferred from the opcode; when we process it in the AsmPrinter, 2109 // we add the necessary relocation there. A load, though, can receive 2110 // relocation from various flavors of ADDIxxx, so we need to carry 2111 // the relocation information in the target flags. 2112 switch (Base.getMachineOpcode()) { 2113 default: continue; 2114 2115 case PPC::ADDI8: 2116 case PPC::ADDI: 2117 // In some cases (such as TLS) the relocation information 2118 // is already in place on the operand, so copying the operand 2119 // is sufficient. 2120 ReplaceFlags = false; 2121 // For these cases, the immediate may not be divisible by 4, in 2122 // which case the fold is illegal for DS-form instructions. (The 2123 // other cases provide aligned addresses and are always safe.) 2124 if ((StorageOpcode == PPC::LWA || 2125 StorageOpcode == PPC::LD || 2126 StorageOpcode == PPC::STD) && 2127 (!isa<ConstantSDNode>(Base.getOperand(1)) || 2128 Base.getConstantOperandVal(1) % 4 != 0)) 2129 continue; 2130 break; 2131 case PPC::ADDIdtprelL: 2132 Flags = PPCII::MO_DTPREL_LO; 2133 break; 2134 case PPC::ADDItlsldL: 2135 Flags = PPCII::MO_TLSLD_LO; 2136 break; 2137 case PPC::ADDItocL: 2138 Flags = PPCII::MO_TOC_LO; 2139 break; 2140 } 2141 2142 // We found an opportunity. Reverse the operands from the add 2143 // immediate and substitute them into the load or store. If 2144 // needed, update the target flags for the immediate operand to 2145 // reflect the necessary relocation information. 2146 DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: "); 2147 DEBUG(Base->dump(CurDAG)); 2148 DEBUG(dbgs() << "\nN: "); 2149 DEBUG(N->dump(CurDAG)); 2150 DEBUG(dbgs() << "\n"); 2151 2152 SDValue ImmOpnd = Base.getOperand(1); 2153 2154 // If the relocation information isn't already present on the 2155 // immediate operand, add it now. 2156 if (ReplaceFlags) { 2157 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) { 2158 SDLoc dl(GA); 2159 const GlobalValue *GV = GA->getGlobal(); 2160 // We can't perform this optimization for data whose alignment 2161 // is insufficient for the instruction encoding. 2162 if (GV->getAlignment() < 4 && 2163 (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD || 2164 StorageOpcode == PPC::LWA)) { 2165 DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n"); 2166 continue; 2167 } 2168 ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, 0, Flags); 2169 } else if (ConstantPoolSDNode *CP = 2170 dyn_cast<ConstantPoolSDNode>(ImmOpnd)) { 2171 const Constant *C = CP->getConstVal(); 2172 ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64, 2173 CP->getAlignment(), 2174 0, Flags); 2175 } 2176 } 2177 2178 if (FirstOp == 1) // Store 2179 (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd, 2180 Base.getOperand(0), N->getOperand(3)); 2181 else // Load 2182 (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0), 2183 N->getOperand(2)); 2184 2185 // The add-immediate may now be dead, in which case remove it. 2186 if (Base.getNode()->use_empty()) 2187 CurDAG->RemoveDeadNode(Base.getNode()); 2188 } 2189 } 2190 2191 2192 /// createPPCISelDag - This pass converts a legalized DAG into a 2193 /// PowerPC-specific DAG, ready for instruction scheduling. 2194 /// 2195 FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) { 2196 return new PPCDAGToDAGISel(TM); 2197 } 2198 2199 static void initializePassOnce(PassRegistry &Registry) { 2200 const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection"; 2201 PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID, 2202 nullptr, false, false); 2203 Registry.registerPass(*PI, true); 2204 } 2205 2206 void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) { 2207 CALL_ONCE_INITIALIZATION(initializePassOnce); 2208 } 2209 2210