LLVM API Documentation
00001 //===-- PPCISelDAGToDAG.cpp - PPC --pattern matching inst selector --------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This file defines a pattern matching instruction selector for PowerPC, 00011 // converting from a legalized dag to a PPC dag. 00012 // 00013 //===----------------------------------------------------------------------===// 00014 00015 #include "PPC.h" 00016 #include "MCTargetDesc/PPCPredicates.h" 00017 #include "PPCMachineFunctionInfo.h" 00018 #include "PPCTargetMachine.h" 00019 #include "llvm/CodeGen/MachineFunction.h" 00020 #include "llvm/CodeGen/MachineInstrBuilder.h" 00021 #include "llvm/CodeGen/MachineRegisterInfo.h" 00022 #include "llvm/CodeGen/SelectionDAG.h" 00023 #include "llvm/CodeGen/SelectionDAGISel.h" 00024 #include "llvm/IR/Constants.h" 00025 #include "llvm/IR/Function.h" 00026 #include "llvm/IR/GlobalAlias.h" 00027 #include "llvm/IR/GlobalValue.h" 00028 #include "llvm/IR/GlobalVariable.h" 00029 #include "llvm/IR/Intrinsics.h" 00030 #include "llvm/Support/CommandLine.h" 00031 #include "llvm/Support/Debug.h" 00032 #include "llvm/Support/ErrorHandling.h" 00033 #include "llvm/Support/MathExtras.h" 00034 #include "llvm/Support/raw_ostream.h" 00035 #include "llvm/Target/TargetOptions.h" 00036 using namespace llvm; 00037 00038 #define DEBUG_TYPE "ppc-codegen" 00039 00040 // FIXME: Remove this once the bug has been fixed! 00041 cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug", 00042 cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden); 00043 00044 namespace llvm { 00045 void initializePPCDAGToDAGISelPass(PassRegistry&); 00046 } 00047 00048 namespace { 00049 //===--------------------------------------------------------------------===// 00050 /// PPCDAGToDAGISel - PPC specific code to select PPC machine 00051 /// instructions for SelectionDAG operations. 00052 /// 00053 class PPCDAGToDAGISel : public SelectionDAGISel { 00054 const PPCTargetMachine &TM; 00055 const PPCTargetLowering *PPCLowering; 00056 const PPCSubtarget *PPCSubTarget; 00057 unsigned GlobalBaseReg; 00058 public: 00059 explicit PPCDAGToDAGISel(PPCTargetMachine &tm) 00060 : SelectionDAGISel(tm), TM(tm), 00061 PPCLowering(TM.getSubtargetImpl()->getTargetLowering()), 00062 PPCSubTarget(TM.getSubtargetImpl()) { 00063 initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry()); 00064 } 00065 00066 bool runOnMachineFunction(MachineFunction &MF) override { 00067 // Make sure we re-emit a set of the global base reg if necessary 00068 GlobalBaseReg = 0; 00069 PPCLowering = TM.getSubtargetImpl()->getTargetLowering(); 00070 PPCSubTarget = TM.getSubtargetImpl(); 00071 SelectionDAGISel::runOnMachineFunction(MF); 00072 00073 if (!PPCSubTarget->isSVR4ABI()) 00074 InsertVRSaveCode(MF); 00075 00076 return true; 00077 } 00078 00079 void PostprocessISelDAG() override; 00080 00081 /// getI32Imm - Return a target constant with the specified value, of type 00082 /// i32. 00083 inline SDValue getI32Imm(unsigned Imm) { 00084 return CurDAG->getTargetConstant(Imm, MVT::i32); 00085 } 00086 00087 /// getI64Imm - Return a target constant with the specified value, of type 00088 /// i64. 00089 inline SDValue getI64Imm(uint64_t Imm) { 00090 return CurDAG->getTargetConstant(Imm, MVT::i64); 00091 } 00092 00093 /// getSmallIPtrImm - Return a target constant of pointer type. 00094 inline SDValue getSmallIPtrImm(unsigned Imm) { 00095 return CurDAG->getTargetConstant(Imm, PPCLowering->getPointerTy()); 00096 } 00097 00098 /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s 00099 /// with any number of 0s on either side. The 1s are allowed to wrap from 00100 /// LSB to MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 00101 /// 0x0F0F0000 is not, since all 1s are not contiguous. 00102 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME); 00103 00104 00105 /// isRotateAndMask - Returns true if Mask and Shift can be folded into a 00106 /// rotate and mask opcode and mask operation. 00107 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask, 00108 unsigned &SH, unsigned &MB, unsigned &ME); 00109 00110 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC 00111 /// base register. Return the virtual register that holds this value. 00112 SDNode *getGlobalBaseReg(); 00113 00114 // Select - Convert the specified operand from a target-independent to a 00115 // target-specific node if it hasn't already been changed. 00116 SDNode *Select(SDNode *N) override; 00117 00118 SDNode *SelectBitfieldInsert(SDNode *N); 00119 00120 /// SelectCC - Select a comparison of the specified values with the 00121 /// specified condition code, returning the CR# of the expression. 00122 SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, SDLoc dl); 00123 00124 /// SelectAddrImm - Returns true if the address N can be represented by 00125 /// a base register plus a signed 16-bit displacement [r+imm]. 00126 bool SelectAddrImm(SDValue N, SDValue &Disp, 00127 SDValue &Base) { 00128 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false); 00129 } 00130 00131 /// SelectAddrImmOffs - Return true if the operand is valid for a preinc 00132 /// immediate field. Note that the operand at this point is already the 00133 /// result of a prior SelectAddressRegImm call. 00134 bool SelectAddrImmOffs(SDValue N, SDValue &Out) const { 00135 if (N.getOpcode() == ISD::TargetConstant || 00136 N.getOpcode() == ISD::TargetGlobalAddress) { 00137 Out = N; 00138 return true; 00139 } 00140 00141 return false; 00142 } 00143 00144 /// SelectAddrIdx - Given the specified addressed, check to see if it can be 00145 /// represented as an indexed [r+r] operation. Returns false if it can 00146 /// be represented by [r+imm], which are preferred. 00147 bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) { 00148 return PPCLowering->SelectAddressRegReg(N, Base, Index, *CurDAG); 00149 } 00150 00151 /// SelectAddrIdxOnly - Given the specified addressed, force it to be 00152 /// represented as an indexed [r+r] operation. 00153 bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) { 00154 return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, *CurDAG); 00155 } 00156 00157 /// SelectAddrImmX4 - Returns true if the address N can be represented by 00158 /// a base register plus a signed 16-bit displacement that is a multiple of 4. 00159 /// Suitable for use by STD and friends. 00160 bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) { 00161 return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true); 00162 } 00163 00164 // Select an address into a single register. 00165 bool SelectAddr(SDValue N, SDValue &Base) { 00166 Base = N; 00167 return true; 00168 } 00169 00170 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for 00171 /// inline asm expressions. It is always correct to compute the value into 00172 /// a register. The case of adding a (possibly relocatable) constant to a 00173 /// register can be improved, but it is wrong to substitute Reg+Reg for 00174 /// Reg in an asm, because the load or store opcode would have to change. 00175 bool SelectInlineAsmMemoryOperand(const SDValue &Op, 00176 char ConstraintCode, 00177 std::vector<SDValue> &OutOps) override { 00178 OutOps.push_back(Op); 00179 return false; 00180 } 00181 00182 void InsertVRSaveCode(MachineFunction &MF); 00183 00184 const char *getPassName() const override { 00185 return "PowerPC DAG->DAG Pattern Instruction Selection"; 00186 } 00187 00188 // Include the pieces autogenerated from the target description. 00189 #include "PPCGenDAGISel.inc" 00190 00191 private: 00192 SDNode *SelectSETCC(SDNode *N); 00193 00194 void PeepholePPC64(); 00195 void PeepholeCROps(); 00196 00197 bool AllUsersSelectZero(SDNode *N); 00198 void SwapAllSelectUsers(SDNode *N); 00199 }; 00200 } 00201 00202 /// InsertVRSaveCode - Once the entire function has been instruction selected, 00203 /// all virtual registers are created and all machine instructions are built, 00204 /// check to see if we need to save/restore VRSAVE. If so, do it. 00205 void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) { 00206 // Check to see if this function uses vector registers, which means we have to 00207 // save and restore the VRSAVE register and update it with the regs we use. 00208 // 00209 // In this case, there will be virtual registers of vector type created 00210 // by the scheduler. Detect them now. 00211 bool HasVectorVReg = false; 00212 for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) { 00213 unsigned Reg = TargetRegisterInfo::index2VirtReg(i); 00214 if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) { 00215 HasVectorVReg = true; 00216 break; 00217 } 00218 } 00219 if (!HasVectorVReg) return; // nothing to do. 00220 00221 // If we have a vector register, we want to emit code into the entry and exit 00222 // blocks to save and restore the VRSAVE register. We do this here (instead 00223 // of marking all vector instructions as clobbering VRSAVE) for two reasons: 00224 // 00225 // 1. This (trivially) reduces the load on the register allocator, by not 00226 // having to represent the live range of the VRSAVE register. 00227 // 2. This (more significantly) allows us to create a temporary virtual 00228 // register to hold the saved VRSAVE value, allowing this temporary to be 00229 // register allocated, instead of forcing it to be spilled to the stack. 00230 00231 // Create two vregs - one to hold the VRSAVE register that is live-in to the 00232 // function and one for the value after having bits or'd into it. 00233 unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 00234 unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 00235 00236 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo(); 00237 MachineBasicBlock &EntryBB = *Fn.begin(); 00238 DebugLoc dl; 00239 // Emit the following code into the entry block: 00240 // InVRSAVE = MFVRSAVE 00241 // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE 00242 // MTVRSAVE UpdatedVRSAVE 00243 MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point 00244 BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE); 00245 BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE), 00246 UpdatedVRSAVE).addReg(InVRSAVE); 00247 BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE); 00248 00249 // Find all return blocks, outputting a restore in each epilog. 00250 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { 00251 if (!BB->empty() && BB->back().isReturn()) { 00252 IP = BB->end(); --IP; 00253 00254 // Skip over all terminator instructions, which are part of the return 00255 // sequence. 00256 MachineBasicBlock::iterator I2 = IP; 00257 while (I2 != BB->begin() && (--I2)->isTerminator()) 00258 IP = I2; 00259 00260 // Emit: MTVRSAVE InVRSave 00261 BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE); 00262 } 00263 } 00264 } 00265 00266 00267 /// getGlobalBaseReg - Output the instructions required to put the 00268 /// base address to use for accessing globals into a register. 00269 /// 00270 SDNode *PPCDAGToDAGISel::getGlobalBaseReg() { 00271 if (!GlobalBaseReg) { 00272 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo(); 00273 // Insert the set of GlobalBaseReg into the first MBB of the function 00274 MachineBasicBlock &FirstMBB = MF->front(); 00275 MachineBasicBlock::iterator MBBI = FirstMBB.begin(); 00276 DebugLoc dl; 00277 00278 if (PPCLowering->getPointerTy() == MVT::i32) { 00279 if (PPCSubTarget->isTargetELF()) 00280 GlobalBaseReg = PPC::R30; 00281 else 00282 GlobalBaseReg = 00283 RegInfo->createVirtualRegister(&PPC::GPRC_NOR0RegClass); 00284 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR)); 00285 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); 00286 if (PPCSubTarget->isTargetELF()) { 00287 unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); 00288 BuildMI(FirstMBB, MBBI, dl, 00289 TII.get(PPC::GetGBRO), TempReg).addReg(GlobalBaseReg); 00290 BuildMI(FirstMBB, MBBI, dl, 00291 TII.get(PPC::UpdateGBR)).addReg(GlobalBaseReg).addReg(TempReg); 00292 MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true); 00293 } 00294 } else { 00295 GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_NOX0RegClass); 00296 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8)); 00297 BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg); 00298 } 00299 } 00300 return CurDAG->getRegister(GlobalBaseReg, 00301 PPCLowering->getPointerTy()).getNode(); 00302 } 00303 00304 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit 00305 /// or 64-bit immediate, and if the value can be accurately represented as a 00306 /// sign extension from a 16-bit value. If so, this returns true and the 00307 /// immediate. 00308 static bool isIntS16Immediate(SDNode *N, short &Imm) { 00309 if (N->getOpcode() != ISD::Constant) 00310 return false; 00311 00312 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue(); 00313 if (N->getValueType(0) == MVT::i32) 00314 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue(); 00315 else 00316 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue(); 00317 } 00318 00319 static bool isIntS16Immediate(SDValue Op, short &Imm) { 00320 return isIntS16Immediate(Op.getNode(), Imm); 00321 } 00322 00323 00324 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant 00325 /// operand. If so Imm will receive the 32-bit value. 00326 static bool isInt32Immediate(SDNode *N, unsigned &Imm) { 00327 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) { 00328 Imm = cast<ConstantSDNode>(N)->getZExtValue(); 00329 return true; 00330 } 00331 return false; 00332 } 00333 00334 /// isInt64Immediate - This method tests to see if the node is a 64-bit constant 00335 /// operand. If so Imm will receive the 64-bit value. 00336 static bool isInt64Immediate(SDNode *N, uint64_t &Imm) { 00337 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) { 00338 Imm = cast<ConstantSDNode>(N)->getZExtValue(); 00339 return true; 00340 } 00341 return false; 00342 } 00343 00344 // isInt32Immediate - This method tests to see if a constant operand. 00345 // If so Imm will receive the 32 bit value. 00346 static bool isInt32Immediate(SDValue N, unsigned &Imm) { 00347 return isInt32Immediate(N.getNode(), Imm); 00348 } 00349 00350 00351 // isOpcWithIntImmediate - This method tests to see if the node is a specific 00352 // opcode and that it has a immediate integer right operand. 00353 // If so Imm will receive the 32 bit value. 00354 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) { 00355 return N->getOpcode() == Opc 00356 && isInt32Immediate(N->getOperand(1).getNode(), Imm); 00357 } 00358 00359 bool PPCDAGToDAGISel::isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) { 00360 if (!Val) 00361 return false; 00362 00363 if (isShiftedMask_32(Val)) { 00364 // look for the first non-zero bit 00365 MB = countLeadingZeros(Val); 00366 // look for the first zero bit after the run of ones 00367 ME = countLeadingZeros((Val - 1) ^ Val); 00368 return true; 00369 } else { 00370 Val = ~Val; // invert mask 00371 if (isShiftedMask_32(Val)) { 00372 // effectively look for the first zero bit 00373 ME = countLeadingZeros(Val) - 1; 00374 // effectively look for the first one bit after the run of zeros 00375 MB = countLeadingZeros((Val - 1) ^ Val) + 1; 00376 return true; 00377 } 00378 } 00379 // no run present 00380 return false; 00381 } 00382 00383 bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask, 00384 bool isShiftMask, unsigned &SH, 00385 unsigned &MB, unsigned &ME) { 00386 // Don't even go down this path for i64, since different logic will be 00387 // necessary for rldicl/rldicr/rldimi. 00388 if (N->getValueType(0) != MVT::i32) 00389 return false; 00390 00391 unsigned Shift = 32; 00392 unsigned Indeterminant = ~0; // bit mask marking indeterminant results 00393 unsigned Opcode = N->getOpcode(); 00394 if (N->getNumOperands() != 2 || 00395 !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31)) 00396 return false; 00397 00398 if (Opcode == ISD::SHL) { 00399 // apply shift left to mask if it comes first 00400 if (isShiftMask) Mask = Mask << Shift; 00401 // determine which bits are made indeterminant by shift 00402 Indeterminant = ~(0xFFFFFFFFu << Shift); 00403 } else if (Opcode == ISD::SRL) { 00404 // apply shift right to mask if it comes first 00405 if (isShiftMask) Mask = Mask >> Shift; 00406 // determine which bits are made indeterminant by shift 00407 Indeterminant = ~(0xFFFFFFFFu >> Shift); 00408 // adjust for the left rotate 00409 Shift = 32 - Shift; 00410 } else if (Opcode == ISD::ROTL) { 00411 Indeterminant = 0; 00412 } else { 00413 return false; 00414 } 00415 00416 // if the mask doesn't intersect any Indeterminant bits 00417 if (Mask && !(Mask & Indeterminant)) { 00418 SH = Shift & 31; 00419 // make sure the mask is still a mask (wrap arounds may not be) 00420 return isRunOfOnes(Mask, MB, ME); 00421 } 00422 return false; 00423 } 00424 00425 /// SelectBitfieldInsert - turn an or of two masked values into 00426 /// the rotate left word immediate then mask insert (rlwimi) instruction. 00427 SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) { 00428 SDValue Op0 = N->getOperand(0); 00429 SDValue Op1 = N->getOperand(1); 00430 SDLoc dl(N); 00431 00432 APInt LKZ, LKO, RKZ, RKO; 00433 CurDAG->computeKnownBits(Op0, LKZ, LKO); 00434 CurDAG->computeKnownBits(Op1, RKZ, RKO); 00435 00436 unsigned TargetMask = LKZ.getZExtValue(); 00437 unsigned InsertMask = RKZ.getZExtValue(); 00438 00439 if ((TargetMask | InsertMask) == 0xFFFFFFFF) { 00440 unsigned Op0Opc = Op0.getOpcode(); 00441 unsigned Op1Opc = Op1.getOpcode(); 00442 unsigned Value, SH = 0; 00443 TargetMask = ~TargetMask; 00444 InsertMask = ~InsertMask; 00445 00446 // If the LHS has a foldable shift and the RHS does not, then swap it to the 00447 // RHS so that we can fold the shift into the insert. 00448 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) { 00449 if (Op0.getOperand(0).getOpcode() == ISD::SHL || 00450 Op0.getOperand(0).getOpcode() == ISD::SRL) { 00451 if (Op1.getOperand(0).getOpcode() != ISD::SHL && 00452 Op1.getOperand(0).getOpcode() != ISD::SRL) { 00453 std::swap(Op0, Op1); 00454 std::swap(Op0Opc, Op1Opc); 00455 std::swap(TargetMask, InsertMask); 00456 } 00457 } 00458 } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) { 00459 if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL && 00460 Op1.getOperand(0).getOpcode() != ISD::SRL) { 00461 std::swap(Op0, Op1); 00462 std::swap(Op0Opc, Op1Opc); 00463 std::swap(TargetMask, InsertMask); 00464 } 00465 } 00466 00467 unsigned MB, ME; 00468 if (isRunOfOnes(InsertMask, MB, ME)) { 00469 SDValue Tmp1, Tmp2; 00470 00471 if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) && 00472 isInt32Immediate(Op1.getOperand(1), Value)) { 00473 Op1 = Op1.getOperand(0); 00474 SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value; 00475 } 00476 if (Op1Opc == ISD::AND) { 00477 // The AND mask might not be a constant, and we need to make sure that 00478 // if we're going to fold the masking with the insert, all bits not 00479 // know to be zero in the mask are known to be one. 00480 APInt MKZ, MKO; 00481 CurDAG->computeKnownBits(Op1.getOperand(1), MKZ, MKO); 00482 bool CanFoldMask = InsertMask == MKO.getZExtValue(); 00483 00484 unsigned SHOpc = Op1.getOperand(0).getOpcode(); 00485 if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask && 00486 isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) { 00487 // Note that Value must be in range here (less than 32) because 00488 // otherwise there would not be any bits set in InsertMask. 00489 Op1 = Op1.getOperand(0).getOperand(0); 00490 SH = (SHOpc == ISD::SHL) ? Value : 32 - Value; 00491 } 00492 } 00493 00494 SH &= 31; 00495 SDValue Ops[] = { Op0, Op1, getI32Imm(SH), getI32Imm(MB), 00496 getI32Imm(ME) }; 00497 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); 00498 } 00499 } 00500 return nullptr; 00501 } 00502 00503 /// SelectCC - Select a comparison of the specified values with the specified 00504 /// condition code, returning the CR# of the expression. 00505 SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, 00506 ISD::CondCode CC, SDLoc dl) { 00507 // Always select the LHS. 00508 unsigned Opc; 00509 00510 if (LHS.getValueType() == MVT::i32) { 00511 unsigned Imm; 00512 if (CC == ISD::SETEQ || CC == ISD::SETNE) { 00513 if (isInt32Immediate(RHS, Imm)) { 00514 // SETEQ/SETNE comparison with 16-bit immediate, fold it. 00515 if (isUInt<16>(Imm)) 00516 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, 00517 getI32Imm(Imm & 0xFFFF)), 0); 00518 // If this is a 16-bit signed immediate, fold it. 00519 if (isInt<16>((int)Imm)) 00520 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, 00521 getI32Imm(Imm & 0xFFFF)), 0); 00522 00523 // For non-equality comparisons, the default code would materialize the 00524 // constant, then compare against it, like this: 00525 // lis r2, 4660 00526 // ori r2, r2, 22136 00527 // cmpw cr0, r3, r2 00528 // Since we are just comparing for equality, we can emit this instead: 00529 // xoris r0,r3,0x1234 00530 // cmplwi cr0,r0,0x5678 00531 // beq cr0,L6 00532 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS, 00533 getI32Imm(Imm >> 16)), 0); 00534 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor, 00535 getI32Imm(Imm & 0xFFFF)), 0); 00536 } 00537 Opc = PPC::CMPLW; 00538 } else if (ISD::isUnsignedIntSetCC(CC)) { 00539 if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm)) 00540 return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, 00541 getI32Imm(Imm & 0xFFFF)), 0); 00542 Opc = PPC::CMPLW; 00543 } else { 00544 short SImm; 00545 if (isIntS16Immediate(RHS, SImm)) 00546 return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, 00547 getI32Imm((int)SImm & 0xFFFF)), 00548 0); 00549 Opc = PPC::CMPW; 00550 } 00551 } else if (LHS.getValueType() == MVT::i64) { 00552 uint64_t Imm; 00553 if (CC == ISD::SETEQ || CC == ISD::SETNE) { 00554 if (isInt64Immediate(RHS.getNode(), Imm)) { 00555 // SETEQ/SETNE comparison with 16-bit immediate, fold it. 00556 if (isUInt<16>(Imm)) 00557 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, 00558 getI32Imm(Imm & 0xFFFF)), 0); 00559 // If this is a 16-bit signed immediate, fold it. 00560 if (isInt<16>(Imm)) 00561 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, 00562 getI32Imm(Imm & 0xFFFF)), 0); 00563 00564 // For non-equality comparisons, the default code would materialize the 00565 // constant, then compare against it, like this: 00566 // lis r2, 4660 00567 // ori r2, r2, 22136 00568 // cmpd cr0, r3, r2 00569 // Since we are just comparing for equality, we can emit this instead: 00570 // xoris r0,r3,0x1234 00571 // cmpldi cr0,r0,0x5678 00572 // beq cr0,L6 00573 if (isUInt<32>(Imm)) { 00574 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS, 00575 getI64Imm(Imm >> 16)), 0); 00576 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor, 00577 getI64Imm(Imm & 0xFFFF)), 0); 00578 } 00579 } 00580 Opc = PPC::CMPLD; 00581 } else if (ISD::isUnsignedIntSetCC(CC)) { 00582 if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm)) 00583 return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, 00584 getI64Imm(Imm & 0xFFFF)), 0); 00585 Opc = PPC::CMPLD; 00586 } else { 00587 short SImm; 00588 if (isIntS16Immediate(RHS, SImm)) 00589 return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, 00590 getI64Imm(SImm & 0xFFFF)), 00591 0); 00592 Opc = PPC::CMPD; 00593 } 00594 } else if (LHS.getValueType() == MVT::f32) { 00595 Opc = PPC::FCMPUS; 00596 } else { 00597 assert(LHS.getValueType() == MVT::f64 && "Unknown vt!"); 00598 Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD; 00599 } 00600 return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0); 00601 } 00602 00603 static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) { 00604 switch (CC) { 00605 case ISD::SETUEQ: 00606 case ISD::SETONE: 00607 case ISD::SETOLE: 00608 case ISD::SETOGE: 00609 llvm_unreachable("Should be lowered by legalize!"); 00610 default: llvm_unreachable("Unknown condition!"); 00611 case ISD::SETOEQ: 00612 case ISD::SETEQ: return PPC::PRED_EQ; 00613 case ISD::SETUNE: 00614 case ISD::SETNE: return PPC::PRED_NE; 00615 case ISD::SETOLT: 00616 case ISD::SETLT: return PPC::PRED_LT; 00617 case ISD::SETULE: 00618 case ISD::SETLE: return PPC::PRED_LE; 00619 case ISD::SETOGT: 00620 case ISD::SETGT: return PPC::PRED_GT; 00621 case ISD::SETUGE: 00622 case ISD::SETGE: return PPC::PRED_GE; 00623 case ISD::SETO: return PPC::PRED_NU; 00624 case ISD::SETUO: return PPC::PRED_UN; 00625 // These two are invalid for floating point. Assume we have int. 00626 case ISD::SETULT: return PPC::PRED_LT; 00627 case ISD::SETUGT: return PPC::PRED_GT; 00628 } 00629 } 00630 00631 /// getCRIdxForSetCC - Return the index of the condition register field 00632 /// associated with the SetCC condition, and whether or not the field is 00633 /// treated as inverted. That is, lt = 0; ge = 0 inverted. 00634 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) { 00635 Invert = false; 00636 switch (CC) { 00637 default: llvm_unreachable("Unknown condition!"); 00638 case ISD::SETOLT: 00639 case ISD::SETLT: return 0; // Bit #0 = SETOLT 00640 case ISD::SETOGT: 00641 case ISD::SETGT: return 1; // Bit #1 = SETOGT 00642 case ISD::SETOEQ: 00643 case ISD::SETEQ: return 2; // Bit #2 = SETOEQ 00644 case ISD::SETUO: return 3; // Bit #3 = SETUO 00645 case ISD::SETUGE: 00646 case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE 00647 case ISD::SETULE: 00648 case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE 00649 case ISD::SETUNE: 00650 case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE 00651 case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO 00652 case ISD::SETUEQ: 00653 case ISD::SETOGE: 00654 case ISD::SETOLE: 00655 case ISD::SETONE: 00656 llvm_unreachable("Invalid branch code: should be expanded by legalize"); 00657 // These are invalid for floating point. Assume integer. 00658 case ISD::SETULT: return 0; 00659 case ISD::SETUGT: return 1; 00660 } 00661 } 00662 00663 // getVCmpInst: return the vector compare instruction for the specified 00664 // vector type and condition code. Since this is for altivec specific code, 00665 // only support the altivec types (v16i8, v8i16, v4i32, and v4f32). 00666 static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC, 00667 bool HasVSX, bool &Swap, bool &Negate) { 00668 Swap = false; 00669 Negate = false; 00670 00671 if (VecVT.isFloatingPoint()) { 00672 /* Handle some cases by swapping input operands. */ 00673 switch (CC) { 00674 case ISD::SETLE: CC = ISD::SETGE; Swap = true; break; 00675 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break; 00676 case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break; 00677 case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break; 00678 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break; 00679 case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break; 00680 default: break; 00681 } 00682 /* Handle some cases by negating the result. */ 00683 switch (CC) { 00684 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break; 00685 case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break; 00686 case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break; 00687 case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break; 00688 default: break; 00689 } 00690 /* We have instructions implementing the remaining cases. */ 00691 switch (CC) { 00692 case ISD::SETEQ: 00693 case ISD::SETOEQ: 00694 if (VecVT == MVT::v4f32) 00695 return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP; 00696 else if (VecVT == MVT::v2f64) 00697 return PPC::XVCMPEQDP; 00698 break; 00699 case ISD::SETGT: 00700 case ISD::SETOGT: 00701 if (VecVT == MVT::v4f32) 00702 return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP; 00703 else if (VecVT == MVT::v2f64) 00704 return PPC::XVCMPGTDP; 00705 break; 00706 case ISD::SETGE: 00707 case ISD::SETOGE: 00708 if (VecVT == MVT::v4f32) 00709 return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP; 00710 else if (VecVT == MVT::v2f64) 00711 return PPC::XVCMPGEDP; 00712 break; 00713 default: 00714 break; 00715 } 00716 llvm_unreachable("Invalid floating-point vector compare condition"); 00717 } else { 00718 /* Handle some cases by swapping input operands. */ 00719 switch (CC) { 00720 case ISD::SETGE: CC = ISD::SETLE; Swap = true; break; 00721 case ISD::SETLT: CC = ISD::SETGT; Swap = true; break; 00722 case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break; 00723 case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break; 00724 default: break; 00725 } 00726 /* Handle some cases by negating the result. */ 00727 switch (CC) { 00728 case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break; 00729 case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break; 00730 case ISD::SETLE: CC = ISD::SETGT; Negate = true; break; 00731 case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break; 00732 default: break; 00733 } 00734 /* We have instructions implementing the remaining cases. */ 00735 switch (CC) { 00736 case ISD::SETEQ: 00737 case ISD::SETUEQ: 00738 if (VecVT == MVT::v16i8) 00739 return PPC::VCMPEQUB; 00740 else if (VecVT == MVT::v8i16) 00741 return PPC::VCMPEQUH; 00742 else if (VecVT == MVT::v4i32) 00743 return PPC::VCMPEQUW; 00744 break; 00745 case ISD::SETGT: 00746 if (VecVT == MVT::v16i8) 00747 return PPC::VCMPGTSB; 00748 else if (VecVT == MVT::v8i16) 00749 return PPC::VCMPGTSH; 00750 else if (VecVT == MVT::v4i32) 00751 return PPC::VCMPGTSW; 00752 break; 00753 case ISD::SETUGT: 00754 if (VecVT == MVT::v16i8) 00755 return PPC::VCMPGTUB; 00756 else if (VecVT == MVT::v8i16) 00757 return PPC::VCMPGTUH; 00758 else if (VecVT == MVT::v4i32) 00759 return PPC::VCMPGTUW; 00760 break; 00761 default: 00762 break; 00763 } 00764 llvm_unreachable("Invalid integer vector compare condition"); 00765 } 00766 } 00767 00768 SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) { 00769 SDLoc dl(N); 00770 unsigned Imm; 00771 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); 00772 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy(); 00773 bool isPPC64 = (PtrVT == MVT::i64); 00774 00775 if (!PPCSubTarget->useCRBits() && 00776 isInt32Immediate(N->getOperand(1), Imm)) { 00777 // We can codegen setcc op, imm very efficiently compared to a brcond. 00778 // Check for those cases here. 00779 // setcc op, 0 00780 if (Imm == 0) { 00781 SDValue Op = N->getOperand(0); 00782 switch (CC) { 00783 default: break; 00784 case ISD::SETEQ: { 00785 Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0); 00786 SDValue Ops[] = { Op, getI32Imm(27), getI32Imm(5), getI32Imm(31) }; 00787 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 00788 } 00789 case ISD::SETNE: { 00790 if (isPPC64) break; 00791 SDValue AD = 00792 SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 00793 Op, getI32Imm(~0U)), 0); 00794 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, 00795 AD.getValue(1)); 00796 } 00797 case ISD::SETLT: { 00798 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 00799 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 00800 } 00801 case ISD::SETGT: { 00802 SDValue T = 00803 SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0); 00804 T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0); 00805 SDValue Ops[] = { T, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 00806 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 00807 } 00808 } 00809 } else if (Imm == ~0U) { // setcc op, -1 00810 SDValue Op = N->getOperand(0); 00811 switch (CC) { 00812 default: break; 00813 case ISD::SETEQ: 00814 if (isPPC64) break; 00815 Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 00816 Op, getI32Imm(1)), 0); 00817 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, 00818 SDValue(CurDAG->getMachineNode(PPC::LI, dl, 00819 MVT::i32, 00820 getI32Imm(0)), 0), 00821 Op.getValue(1)); 00822 case ISD::SETNE: { 00823 if (isPPC64) break; 00824 Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0); 00825 SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 00826 Op, getI32Imm(~0U)); 00827 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0), 00828 Op, SDValue(AD, 1)); 00829 } 00830 case ISD::SETLT: { 00831 SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op, 00832 getI32Imm(1)), 0); 00833 SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD, 00834 Op), 0); 00835 SDValue Ops[] = { AN, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 00836 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 00837 } 00838 case ISD::SETGT: { 00839 SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; 00840 Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 00841 0); 00842 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, 00843 getI32Imm(1)); 00844 } 00845 } 00846 } 00847 } 00848 00849 SDValue LHS = N->getOperand(0); 00850 SDValue RHS = N->getOperand(1); 00851 00852 // Altivec Vector compare instructions do not set any CR register by default and 00853 // vector compare operations return the same type as the operands. 00854 if (LHS.getValueType().isVector()) { 00855 EVT VecVT = LHS.getValueType(); 00856 bool Swap, Negate; 00857 unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC, 00858 PPCSubTarget->hasVSX(), Swap, Negate); 00859 if (Swap) 00860 std::swap(LHS, RHS); 00861 00862 if (Negate) { 00863 SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0); 00864 return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR : 00865 PPC::VNOR, 00866 VecVT, VCmp, VCmp); 00867 } 00868 00869 return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS); 00870 } 00871 00872 if (PPCSubTarget->useCRBits()) 00873 return nullptr; 00874 00875 bool Inv; 00876 unsigned Idx = getCRIdxForSetCC(CC, Inv); 00877 SDValue CCReg = SelectCC(LHS, RHS, CC, dl); 00878 SDValue IntCR; 00879 00880 // Force the ccreg into CR7. 00881 SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32); 00882 00883 SDValue InFlag(nullptr, 0); // Null incoming flag value. 00884 CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg, 00885 InFlag).getValue(1); 00886 00887 IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg, 00888 CCReg), 0); 00889 00890 SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31), 00891 getI32Imm(31), getI32Imm(31) }; 00892 if (!Inv) 00893 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 00894 00895 // Get the specified bit. 00896 SDValue Tmp = 00897 SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); 00898 return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1)); 00899 } 00900 00901 00902 // Select - Convert the specified operand from a target-independent to a 00903 // target-specific node if it hasn't already been changed. 00904 SDNode *PPCDAGToDAGISel::Select(SDNode *N) { 00905 SDLoc dl(N); 00906 if (N->isMachineOpcode()) { 00907 N->setNodeId(-1); 00908 return nullptr; // Already selected. 00909 } 00910 00911 // In case any misguided DAG-level optimizations form an ADD with a 00912 // TargetConstant operand, crash here instead of miscompiling (by selecting 00913 // an r+r add instead of some kind of r+i add). 00914 if (N->getOpcode() == ISD::ADD && 00915 N->getOperand(1).getOpcode() == ISD::TargetConstant) 00916 llvm_unreachable("Invalid ADD with TargetConstant operand"); 00917 00918 switch (N->getOpcode()) { 00919 default: break; 00920 00921 case ISD::Constant: { 00922 if (N->getValueType(0) == MVT::i64) { 00923 // Get 64 bit value. 00924 int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue(); 00925 // Assume no remaining bits. 00926 unsigned Remainder = 0; 00927 // Assume no shift required. 00928 unsigned Shift = 0; 00929 00930 // If it can't be represented as a 32 bit value. 00931 if (!isInt<32>(Imm)) { 00932 Shift = countTrailingZeros<uint64_t>(Imm); 00933 int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift; 00934 00935 // If the shifted value fits 32 bits. 00936 if (isInt<32>(ImmSh)) { 00937 // Go with the shifted value. 00938 Imm = ImmSh; 00939 } else { 00940 // Still stuck with a 64 bit value. 00941 Remainder = Imm; 00942 Shift = 32; 00943 Imm >>= 32; 00944 } 00945 } 00946 00947 // Intermediate operand. 00948 SDNode *Result; 00949 00950 // Handle first 32 bits. 00951 unsigned Lo = Imm & 0xFFFF; 00952 unsigned Hi = (Imm >> 16) & 0xFFFF; 00953 00954 // Simple value. 00955 if (isInt<16>(Imm)) { 00956 // Just the Lo bits. 00957 Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo)); 00958 } else if (Lo) { 00959 // Handle the Hi bits. 00960 unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8; 00961 Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi)); 00962 // And Lo bits. 00963 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, 00964 SDValue(Result, 0), getI32Imm(Lo)); 00965 } else { 00966 // Just the Hi bits. 00967 Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi)); 00968 } 00969 00970 // If no shift, we're done. 00971 if (!Shift) return Result; 00972 00973 // Shift for next step if the upper 32-bits were not zero. 00974 if (Imm) { 00975 Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, 00976 SDValue(Result, 0), 00977 getI32Imm(Shift), 00978 getI32Imm(63 - Shift)); 00979 } 00980 00981 // Add in the last bits as required. 00982 if ((Hi = (Remainder >> 16) & 0xFFFF)) { 00983 Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64, 00984 SDValue(Result, 0), getI32Imm(Hi)); 00985 } 00986 if ((Lo = Remainder & 0xFFFF)) { 00987 Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, 00988 SDValue(Result, 0), getI32Imm(Lo)); 00989 } 00990 00991 return Result; 00992 } 00993 break; 00994 } 00995 00996 case ISD::SETCC: { 00997 SDNode *SN = SelectSETCC(N); 00998 if (SN) 00999 return SN; 01000 break; 01001 } 01002 case PPCISD::GlobalBaseReg: 01003 return getGlobalBaseReg(); 01004 01005 case ISD::FrameIndex: { 01006 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 01007 SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0)); 01008 unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8; 01009 if (N->hasOneUse()) 01010 return CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), TFI, 01011 getSmallIPtrImm(0)); 01012 return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI, 01013 getSmallIPtrImm(0)); 01014 } 01015 01016 case PPCISD::MFOCRF: { 01017 SDValue InFlag = N->getOperand(1); 01018 return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, 01019 N->getOperand(0), InFlag); 01020 } 01021 01022 case ISD::SDIV: { 01023 // FIXME: since this depends on the setting of the carry flag from the srawi 01024 // we should really be making notes about that for the scheduler. 01025 // FIXME: It sure would be nice if we could cheaply recognize the 01026 // srl/add/sra pattern the dag combiner will generate for this as 01027 // sra/addze rather than having to handle sdiv ourselves. oh well. 01028 unsigned Imm; 01029 if (isInt32Immediate(N->getOperand(1), Imm)) { 01030 SDValue N0 = N->getOperand(0); 01031 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) { 01032 SDNode *Op = 01033 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue, 01034 N0, getI32Imm(Log2_32(Imm))); 01035 return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, 01036 SDValue(Op, 0), SDValue(Op, 1)); 01037 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) { 01038 SDNode *Op = 01039 CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue, 01040 N0, getI32Imm(Log2_32(-Imm))); 01041 SDValue PT = 01042 SDValue(CurDAG->getMachineNode(PPC::ADDZE, dl, MVT::i32, 01043 SDValue(Op, 0), SDValue(Op, 1)), 01044 0); 01045 return CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT); 01046 } 01047 } 01048 01049 // Other cases are autogenerated. 01050 break; 01051 } 01052 01053 case ISD::LOAD: { 01054 // Handle preincrement loads. 01055 LoadSDNode *LD = cast<LoadSDNode>(N); 01056 EVT LoadedVT = LD->getMemoryVT(); 01057 01058 // Normal loads are handled by code generated from the .td file. 01059 if (LD->getAddressingMode() != ISD::PRE_INC) 01060 break; 01061 01062 SDValue Offset = LD->getOffset(); 01063 if (Offset.getOpcode() == ISD::TargetConstant || 01064 Offset.getOpcode() == ISD::TargetGlobalAddress) { 01065 01066 unsigned Opcode; 01067 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; 01068 if (LD->getValueType(0) != MVT::i64) { 01069 // Handle PPC32 integer and normal FP loads. 01070 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 01071 switch (LoadedVT.getSimpleVT().SimpleTy) { 01072 default: llvm_unreachable("Invalid PPC load type!"); 01073 case MVT::f64: Opcode = PPC::LFDU; break; 01074 case MVT::f32: Opcode = PPC::LFSU; break; 01075 case MVT::i32: Opcode = PPC::LWZU; break; 01076 case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break; 01077 case MVT::i1: 01078 case MVT::i8: Opcode = PPC::LBZU; break; 01079 } 01080 } else { 01081 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!"); 01082 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 01083 switch (LoadedVT.getSimpleVT().SimpleTy) { 01084 default: llvm_unreachable("Invalid PPC load type!"); 01085 case MVT::i64: Opcode = PPC::LDU; break; 01086 case MVT::i32: Opcode = PPC::LWZU8; break; 01087 case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break; 01088 case MVT::i1: 01089 case MVT::i8: Opcode = PPC::LBZU8; break; 01090 } 01091 } 01092 01093 SDValue Chain = LD->getChain(); 01094 SDValue Base = LD->getBasePtr(); 01095 SDValue Ops[] = { Offset, Base, Chain }; 01096 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0), 01097 PPCLowering->getPointerTy(), 01098 MVT::Other, Ops); 01099 } else { 01100 unsigned Opcode; 01101 bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; 01102 if (LD->getValueType(0) != MVT::i64) { 01103 // Handle PPC32 integer and normal FP loads. 01104 assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); 01105 switch (LoadedVT.getSimpleVT().SimpleTy) { 01106 default: llvm_unreachable("Invalid PPC load type!"); 01107 case MVT::f64: Opcode = PPC::LFDUX; break; 01108 case MVT::f32: Opcode = PPC::LFSUX; break; 01109 case MVT::i32: Opcode = PPC::LWZUX; break; 01110 case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break; 01111 case MVT::i1: 01112 case MVT::i8: Opcode = PPC::LBZUX; break; 01113 } 01114 } else { 01115 assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!"); 01116 assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) && 01117 "Invalid sext update load"); 01118 switch (LoadedVT.getSimpleVT().SimpleTy) { 01119 default: llvm_unreachable("Invalid PPC load type!"); 01120 case MVT::i64: Opcode = PPC::LDUX; break; 01121 case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break; 01122 case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break; 01123 case MVT::i1: 01124 case MVT::i8: Opcode = PPC::LBZUX8; break; 01125 } 01126 } 01127 01128 SDValue Chain = LD->getChain(); 01129 SDValue Base = LD->getBasePtr(); 01130 SDValue Ops[] = { Base, Offset, Chain }; 01131 return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0), 01132 PPCLowering->getPointerTy(), 01133 MVT::Other, Ops); 01134 } 01135 } 01136 01137 case ISD::AND: { 01138 unsigned Imm, Imm2, SH, MB, ME; 01139 uint64_t Imm64; 01140 01141 // If this is an and of a value rotated between 0 and 31 bits and then and'd 01142 // with a mask, emit rlwinm 01143 if (isInt32Immediate(N->getOperand(1), Imm) && 01144 isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) { 01145 SDValue Val = N->getOperand(0).getOperand(0); 01146 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; 01147 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 01148 } 01149 // If this is just a masked value where the input is not handled above, and 01150 // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm 01151 if (isInt32Immediate(N->getOperand(1), Imm) && 01152 isRunOfOnes(Imm, MB, ME) && 01153 N->getOperand(0).getOpcode() != ISD::ROTL) { 01154 SDValue Val = N->getOperand(0); 01155 SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB), getI32Imm(ME) }; 01156 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 01157 } 01158 // If this is a 64-bit zero-extension mask, emit rldicl. 01159 if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) && 01160 isMask_64(Imm64)) { 01161 SDValue Val = N->getOperand(0); 01162 MB = 64 - CountTrailingOnes_64(Imm64); 01163 SH = 0; 01164 01165 // If the operand is a logical right shift, we can fold it into this 01166 // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb) 01167 // for n <= mb. The right shift is really a left rotate followed by a 01168 // mask, and this mask is a more-restrictive sub-mask of the mask implied 01169 // by the shift. 01170 if (Val.getOpcode() == ISD::SRL && 01171 isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) { 01172 assert(Imm < 64 && "Illegal shift amount"); 01173 Val = Val.getOperand(0); 01174 SH = 64 - Imm; 01175 } 01176 01177 SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB) }; 01178 return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops); 01179 } 01180 // AND X, 0 -> 0, not "rlwinm 32". 01181 if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) { 01182 ReplaceUses(SDValue(N, 0), N->getOperand(1)); 01183 return nullptr; 01184 } 01185 // ISD::OR doesn't get all the bitfield insertion fun. 01186 // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert 01187 if (isInt32Immediate(N->getOperand(1), Imm) && 01188 N->getOperand(0).getOpcode() == ISD::OR && 01189 isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) { 01190 unsigned MB, ME; 01191 Imm = ~(Imm^Imm2); 01192 if (isRunOfOnes(Imm, MB, ME)) { 01193 SDValue Ops[] = { N->getOperand(0).getOperand(0), 01194 N->getOperand(0).getOperand(1), 01195 getI32Imm(0), getI32Imm(MB),getI32Imm(ME) }; 01196 return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); 01197 } 01198 } 01199 01200 // Other cases are autogenerated. 01201 break; 01202 } 01203 case ISD::OR: 01204 if (N->getValueType(0) == MVT::i32) 01205 if (SDNode *I = SelectBitfieldInsert(N)) 01206 return I; 01207 01208 // Other cases are autogenerated. 01209 break; 01210 case ISD::SHL: { 01211 unsigned Imm, SH, MB, ME; 01212 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && 01213 isRotateAndMask(N, Imm, true, SH, MB, ME)) { 01214 SDValue Ops[] = { N->getOperand(0).getOperand(0), 01215 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; 01216 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 01217 } 01218 01219 // Other cases are autogenerated. 01220 break; 01221 } 01222 case ISD::SRL: { 01223 unsigned Imm, SH, MB, ME; 01224 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && 01225 isRotateAndMask(N, Imm, true, SH, MB, ME)) { 01226 SDValue Ops[] = { N->getOperand(0).getOperand(0), 01227 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; 01228 return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); 01229 } 01230 01231 // Other cases are autogenerated. 01232 break; 01233 } 01234 // FIXME: Remove this once the ANDI glue bug is fixed: 01235 case PPCISD::ANDIo_1_EQ_BIT: 01236 case PPCISD::ANDIo_1_GT_BIT: { 01237 if (!ANDIGlueBug) 01238 break; 01239 01240 EVT InVT = N->getOperand(0).getValueType(); 01241 assert((InVT == MVT::i64 || InVT == MVT::i32) && 01242 "Invalid input type for ANDIo_1_EQ_BIT"); 01243 01244 unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo; 01245 SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue, 01246 N->getOperand(0), 01247 CurDAG->getTargetConstant(1, InVT)), 0); 01248 SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32); 01249 SDValue SRIdxVal = 01250 CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ? 01251 PPC::sub_eq : PPC::sub_gt, MVT::i32); 01252 01253 return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1, 01254 CR0Reg, SRIdxVal, 01255 SDValue(AndI.getNode(), 1) /* glue */); 01256 } 01257 case ISD::SELECT_CC: { 01258 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get(); 01259 EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy(); 01260 bool isPPC64 = (PtrVT == MVT::i64); 01261 01262 // If this is a select of i1 operands, we'll pattern match it. 01263 if (PPCSubTarget->useCRBits() && 01264 N->getOperand(0).getValueType() == MVT::i1) 01265 break; 01266 01267 // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc 01268 if (!isPPC64) 01269 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1))) 01270 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2))) 01271 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3))) 01272 if (N1C->isNullValue() && N3C->isNullValue() && 01273 N2C->getZExtValue() == 1ULL && CC == ISD::SETNE && 01274 // FIXME: Implement this optzn for PPC64. 01275 N->getValueType(0) == MVT::i32) { 01276 SDNode *Tmp = 01277 CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, 01278 N->getOperand(0), getI32Imm(~0U)); 01279 return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, 01280 SDValue(Tmp, 0), N->getOperand(0), 01281 SDValue(Tmp, 1)); 01282 } 01283 01284 SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl); 01285 01286 if (N->getValueType(0) == MVT::i1) { 01287 // An i1 select is: (c & t) | (!c & f). 01288 bool Inv; 01289 unsigned Idx = getCRIdxForSetCC(CC, Inv); 01290 01291 unsigned SRI; 01292 switch (Idx) { 01293 default: llvm_unreachable("Invalid CC index"); 01294 case 0: SRI = PPC::sub_lt; break; 01295 case 1: SRI = PPC::sub_gt; break; 01296 case 2: SRI = PPC::sub_eq; break; 01297 case 3: SRI = PPC::sub_un; break; 01298 } 01299 01300 SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg); 01301 01302 SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1, 01303 CCBit, CCBit), 0); 01304 SDValue C = Inv ? NotCCBit : CCBit, 01305 NotC = Inv ? CCBit : NotCCBit; 01306 01307 SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1, 01308 C, N->getOperand(2)), 0); 01309 SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1, 01310 NotC, N->getOperand(3)), 0); 01311 01312 return CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF); 01313 } 01314 01315 unsigned BROpc = getPredicateForSetCC(CC); 01316 01317 unsigned SelectCCOp; 01318 if (N->getValueType(0) == MVT::i32) 01319 SelectCCOp = PPC::SELECT_CC_I4; 01320 else if (N->getValueType(0) == MVT::i64) 01321 SelectCCOp = PPC::SELECT_CC_I8; 01322 else if (N->getValueType(0) == MVT::f32) 01323 SelectCCOp = PPC::SELECT_CC_F4; 01324 else if (N->getValueType(0) == MVT::f64) 01325 SelectCCOp = PPC::SELECT_CC_F8; 01326 else 01327 SelectCCOp = PPC::SELECT_CC_VRRC; 01328 01329 SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3), 01330 getI32Imm(BROpc) }; 01331 return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops); 01332 } 01333 case ISD::VSELECT: 01334 if (PPCSubTarget->hasVSX()) { 01335 SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) }; 01336 return CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops); 01337 } 01338 01339 break; 01340 case ISD::VECTOR_SHUFFLE: 01341 if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 || 01342 N->getValueType(0) == MVT::v2i64)) { 01343 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 01344 01345 SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1), 01346 Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1); 01347 unsigned DM[2]; 01348 01349 for (int i = 0; i < 2; ++i) 01350 if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2) 01351 DM[i] = 0; 01352 else 01353 DM[i] = 1; 01354 01355 SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), MVT::i32); 01356 01357 if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 && 01358 Op1.getOpcode() == ISD::SCALAR_TO_VECTOR && 01359 isa<LoadSDNode>(Op1.getOperand(0))) { 01360 LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0)); 01361 SDValue Base, Offset; 01362 01363 if (LD->isUnindexed() && 01364 SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) { 01365 SDValue Chain = LD->getChain(); 01366 SDValue Ops[] = { Base, Offset, Chain }; 01367 return CurDAG->SelectNodeTo(N, PPC::LXVDSX, 01368 N->getValueType(0), Ops); 01369 } 01370 } 01371 01372 SDValue Ops[] = { Op1, Op2, DMV }; 01373 return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops); 01374 } 01375 01376 break; 01377 case PPCISD::BDNZ: 01378 case PPCISD::BDZ: { 01379 bool IsPPC64 = PPCSubTarget->isPPC64(); 01380 SDValue Ops[] = { N->getOperand(1), N->getOperand(0) }; 01381 return CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ ? 01382 (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 01383 (IsPPC64 ? PPC::BDZ8 : PPC::BDZ), 01384 MVT::Other, Ops); 01385 } 01386 case PPCISD::COND_BRANCH: { 01387 // Op #0 is the Chain. 01388 // Op #1 is the PPC::PRED_* number. 01389 // Op #2 is the CR# 01390 // Op #3 is the Dest MBB 01391 // Op #4 is the Flag. 01392 // Prevent PPC::PRED_* from being selected into LI. 01393 SDValue Pred = 01394 getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()); 01395 SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3), 01396 N->getOperand(0), N->getOperand(4) }; 01397 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); 01398 } 01399 case ISD::BR_CC: { 01400 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get(); 01401 unsigned PCC = getPredicateForSetCC(CC); 01402 01403 if (N->getOperand(2).getValueType() == MVT::i1) { 01404 unsigned Opc; 01405 bool Swap; 01406 switch (PCC) { 01407 default: llvm_unreachable("Unexpected Boolean-operand predicate"); 01408 case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break; 01409 case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break; 01410 case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break; 01411 case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break; 01412 case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break; 01413 case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break; 01414 } 01415 01416 SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1, 01417 N->getOperand(Swap ? 3 : 2), 01418 N->getOperand(Swap ? 2 : 3)), 0); 01419 return CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other, 01420 BitComp, N->getOperand(4), N->getOperand(0)); 01421 } 01422 01423 SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl); 01424 SDValue Ops[] = { getI32Imm(PCC), CondCode, 01425 N->getOperand(4), N->getOperand(0) }; 01426 return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); 01427 } 01428 case ISD::BRIND: { 01429 // FIXME: Should custom lower this. 01430 SDValue Chain = N->getOperand(0); 01431 SDValue Target = N->getOperand(1); 01432 unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8; 01433 unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8; 01434 Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target, 01435 Chain), 0); 01436 return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain); 01437 } 01438 case PPCISD::TOC_ENTRY: { 01439 if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) { 01440 SDValue GA = N->getOperand(0); 01441 return CurDAG->getMachineNode(PPC::LWZtoc, dl, MVT::i32, GA, 01442 N->getOperand(1)); 01443 } 01444 assert (PPCSubTarget->isPPC64() && 01445 "Only supported for 64-bit ABI and 32-bit SVR4"); 01446 01447 // For medium and large code model, we generate two instructions as 01448 // described below. Otherwise we allow SelectCodeCommon to handle this, 01449 // selecting one of LDtoc, LDtocJTI, and LDtocCPT. 01450 CodeModel::Model CModel = TM.getCodeModel(); 01451 if (CModel != CodeModel::Medium && CModel != CodeModel::Large) 01452 break; 01453 01454 // The first source operand is a TargetGlobalAddress or a TargetJumpTable. 01455 // If it is an externally defined symbol, a symbol with common linkage, 01456 // a non-local function address, or a jump table address, or if we are 01457 // generating code for large code model, we generate: 01458 // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>)) 01459 // Otherwise we generate: 01460 // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>) 01461 SDValue GA = N->getOperand(0); 01462 SDValue TOCbase = N->getOperand(1); 01463 SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64, 01464 TOCbase, GA); 01465 01466 if (isa<JumpTableSDNode>(GA) || CModel == CodeModel::Large) 01467 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA, 01468 SDValue(Tmp, 0)); 01469 01470 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) { 01471 const GlobalValue *GValue = G->getGlobal(); 01472 if ((GValue->getType()->getElementType()->isFunctionTy() && 01473 (GValue->isDeclaration() || GValue->isWeakForLinker())) || 01474 GValue->isDeclaration() || GValue->hasCommonLinkage() || 01475 GValue->hasAvailableExternallyLinkage()) 01476 return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA, 01477 SDValue(Tmp, 0)); 01478 } 01479 01480 return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64, 01481 SDValue(Tmp, 0), GA); 01482 } 01483 case PPCISD::PPC32_PICGOT: { 01484 // Generate a PIC-safe GOT reference. 01485 assert(!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() && 01486 "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4"); 01487 return CurDAG->SelectNodeTo(N, PPC::PPC32PICGOT, PPCLowering->getPointerTy(), MVT::i32); 01488 } 01489 case PPCISD::VADD_SPLAT: { 01490 // This expands into one of three sequences, depending on whether 01491 // the first operand is odd or even, positive or negative. 01492 assert(isa<ConstantSDNode>(N->getOperand(0)) && 01493 isa<ConstantSDNode>(N->getOperand(1)) && 01494 "Invalid operand on VADD_SPLAT!"); 01495 01496 int Elt = N->getConstantOperandVal(0); 01497 int EltSize = N->getConstantOperandVal(1); 01498 unsigned Opc1, Opc2, Opc3; 01499 EVT VT; 01500 01501 if (EltSize == 1) { 01502 Opc1 = PPC::VSPLTISB; 01503 Opc2 = PPC::VADDUBM; 01504 Opc3 = PPC::VSUBUBM; 01505 VT = MVT::v16i8; 01506 } else if (EltSize == 2) { 01507 Opc1 = PPC::VSPLTISH; 01508 Opc2 = PPC::VADDUHM; 01509 Opc3 = PPC::VSUBUHM; 01510 VT = MVT::v8i16; 01511 } else { 01512 assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!"); 01513 Opc1 = PPC::VSPLTISW; 01514 Opc2 = PPC::VADDUWM; 01515 Opc3 = PPC::VSUBUWM; 01516 VT = MVT::v4i32; 01517 } 01518 01519 if ((Elt & 1) == 0) { 01520 // Elt is even, in the range [-32,-18] + [16,30]. 01521 // 01522 // Convert: VADD_SPLAT elt, size 01523 // Into: tmp = VSPLTIS[BHW] elt 01524 // VADDU[BHW]M tmp, tmp 01525 // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4 01526 SDValue EltVal = getI32Imm(Elt >> 1); 01527 SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 01528 SDValue TmpVal = SDValue(Tmp, 0); 01529 return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal); 01530 01531 } else if (Elt > 0) { 01532 // Elt is odd and positive, in the range [17,31]. 01533 // 01534 // Convert: VADD_SPLAT elt, size 01535 // Into: tmp1 = VSPLTIS[BHW] elt-16 01536 // tmp2 = VSPLTIS[BHW] -16 01537 // VSUBU[BHW]M tmp1, tmp2 01538 SDValue EltVal = getI32Imm(Elt - 16); 01539 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 01540 EltVal = getI32Imm(-16); 01541 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 01542 return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0), 01543 SDValue(Tmp2, 0)); 01544 01545 } else { 01546 // Elt is odd and negative, in the range [-31,-17]. 01547 // 01548 // Convert: VADD_SPLAT elt, size 01549 // Into: tmp1 = VSPLTIS[BHW] elt+16 01550 // tmp2 = VSPLTIS[BHW] -16 01551 // VADDU[BHW]M tmp1, tmp2 01552 SDValue EltVal = getI32Imm(Elt + 16); 01553 SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 01554 EltVal = getI32Imm(-16); 01555 SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); 01556 return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0), 01557 SDValue(Tmp2, 0)); 01558 } 01559 } 01560 } 01561 01562 return SelectCode(N); 01563 } 01564 01565 /// PostprocessISelDAG - Perform some late peephole optimizations 01566 /// on the DAG representation. 01567 void PPCDAGToDAGISel::PostprocessISelDAG() { 01568 01569 // Skip peepholes at -O0. 01570 if (TM.getOptLevel() == CodeGenOpt::None) 01571 return; 01572 01573 PeepholePPC64(); 01574 PeepholeCROps(); 01575 } 01576 01577 // Check if all users of this node will become isel where the second operand 01578 // is the constant zero. If this is so, and if we can negate the condition, 01579 // then we can flip the true and false operands. This will allow the zero to 01580 // be folded with the isel so that we don't need to materialize a register 01581 // containing zero. 01582 bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) { 01583 // If we're not using isel, then this does not matter. 01584 if (!PPCSubTarget->hasISEL()) 01585 return false; 01586 01587 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 01588 UI != UE; ++UI) { 01589 SDNode *User = *UI; 01590 if (!User->isMachineOpcode()) 01591 return false; 01592 if (User->getMachineOpcode() != PPC::SELECT_I4 && 01593 User->getMachineOpcode() != PPC::SELECT_I8) 01594 return false; 01595 01596 SDNode *Op2 = User->getOperand(2).getNode(); 01597 if (!Op2->isMachineOpcode()) 01598 return false; 01599 01600 if (Op2->getMachineOpcode() != PPC::LI && 01601 Op2->getMachineOpcode() != PPC::LI8) 01602 return false; 01603 01604 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0)); 01605 if (!C) 01606 return false; 01607 01608 if (!C->isNullValue()) 01609 return false; 01610 } 01611 01612 return true; 01613 } 01614 01615 void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) { 01616 SmallVector<SDNode *, 4> ToReplace; 01617 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 01618 UI != UE; ++UI) { 01619 SDNode *User = *UI; 01620 assert((User->getMachineOpcode() == PPC::SELECT_I4 || 01621 User->getMachineOpcode() == PPC::SELECT_I8) && 01622 "Must have all select users"); 01623 ToReplace.push_back(User); 01624 } 01625 01626 for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(), 01627 UE = ToReplace.end(); UI != UE; ++UI) { 01628 SDNode *User = *UI; 01629 SDNode *ResNode = 01630 CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User), 01631 User->getValueType(0), User->getOperand(0), 01632 User->getOperand(2), 01633 User->getOperand(1)); 01634 01635 DEBUG(dbgs() << "CR Peephole replacing:\nOld: "); 01636 DEBUG(User->dump(CurDAG)); 01637 DEBUG(dbgs() << "\nNew: "); 01638 DEBUG(ResNode->dump(CurDAG)); 01639 DEBUG(dbgs() << "\n"); 01640 01641 ReplaceUses(User, ResNode); 01642 } 01643 } 01644 01645 void PPCDAGToDAGISel::PeepholeCROps() { 01646 bool IsModified; 01647 do { 01648 IsModified = false; 01649 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(), 01650 E = CurDAG->allnodes_end(); I != E; ++I) { 01651 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I); 01652 if (!MachineNode || MachineNode->use_empty()) 01653 continue; 01654 SDNode *ResNode = MachineNode; 01655 01656 bool Op1Set = false, Op1Unset = false, 01657 Op1Not = false, 01658 Op2Set = false, Op2Unset = false, 01659 Op2Not = false; 01660 01661 unsigned Opcode = MachineNode->getMachineOpcode(); 01662 switch (Opcode) { 01663 default: break; 01664 case PPC::CRAND: 01665 case PPC::CRNAND: 01666 case PPC::CROR: 01667 case PPC::CRXOR: 01668 case PPC::CRNOR: 01669 case PPC::CREQV: 01670 case PPC::CRANDC: 01671 case PPC::CRORC: { 01672 SDValue Op = MachineNode->getOperand(1); 01673 if (Op.isMachineOpcode()) { 01674 if (Op.getMachineOpcode() == PPC::CRSET) 01675 Op2Set = true; 01676 else if (Op.getMachineOpcode() == PPC::CRUNSET) 01677 Op2Unset = true; 01678 else if (Op.getMachineOpcode() == PPC::CRNOR && 01679 Op.getOperand(0) == Op.getOperand(1)) 01680 Op2Not = true; 01681 } 01682 } // fallthrough 01683 case PPC::BC: 01684 case PPC::BCn: 01685 case PPC::SELECT_I4: 01686 case PPC::SELECT_I8: 01687 case PPC::SELECT_F4: 01688 case PPC::SELECT_F8: 01689 case PPC::SELECT_VRRC: { 01690 SDValue Op = MachineNode->getOperand(0); 01691 if (Op.isMachineOpcode()) { 01692 if (Op.getMachineOpcode() == PPC::CRSET) 01693 Op1Set = true; 01694 else if (Op.getMachineOpcode() == PPC::CRUNSET) 01695 Op1Unset = true; 01696 else if (Op.getMachineOpcode() == PPC::CRNOR && 01697 Op.getOperand(0) == Op.getOperand(1)) 01698 Op1Not = true; 01699 } 01700 } 01701 break; 01702 } 01703 01704 bool SelectSwap = false; 01705 switch (Opcode) { 01706 default: break; 01707 case PPC::CRAND: 01708 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 01709 // x & x = x 01710 ResNode = MachineNode->getOperand(0).getNode(); 01711 else if (Op1Set) 01712 // 1 & y = y 01713 ResNode = MachineNode->getOperand(1).getNode(); 01714 else if (Op2Set) 01715 // x & 1 = x 01716 ResNode = MachineNode->getOperand(0).getNode(); 01717 else if (Op1Unset || Op2Unset) 01718 // x & 0 = 0 & y = 0 01719 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 01720 MVT::i1); 01721 else if (Op1Not) 01722 // ~x & y = andc(y, x) 01723 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 01724 MVT::i1, MachineNode->getOperand(1), 01725 MachineNode->getOperand(0). 01726 getOperand(0)); 01727 else if (Op2Not) 01728 // x & ~y = andc(x, y) 01729 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 01730 MVT::i1, MachineNode->getOperand(0), 01731 MachineNode->getOperand(1). 01732 getOperand(0)); 01733 else if (AllUsersSelectZero(MachineNode)) 01734 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode), 01735 MVT::i1, MachineNode->getOperand(0), 01736 MachineNode->getOperand(1)), 01737 SelectSwap = true; 01738 break; 01739 case PPC::CRNAND: 01740 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 01741 // nand(x, x) -> nor(x, x) 01742 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01743 MVT::i1, MachineNode->getOperand(0), 01744 MachineNode->getOperand(0)); 01745 else if (Op1Set) 01746 // nand(1, y) -> nor(y, y) 01747 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01748 MVT::i1, MachineNode->getOperand(1), 01749 MachineNode->getOperand(1)); 01750 else if (Op2Set) 01751 // nand(x, 1) -> nor(x, x) 01752 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01753 MVT::i1, MachineNode->getOperand(0), 01754 MachineNode->getOperand(0)); 01755 else if (Op1Unset || Op2Unset) 01756 // nand(x, 0) = nand(0, y) = 1 01757 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 01758 MVT::i1); 01759 else if (Op1Not) 01760 // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y) 01761 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 01762 MVT::i1, MachineNode->getOperand(0). 01763 getOperand(0), 01764 MachineNode->getOperand(1)); 01765 else if (Op2Not) 01766 // nand(x, ~y) = ~x | y = orc(y, x) 01767 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 01768 MVT::i1, MachineNode->getOperand(1). 01769 getOperand(0), 01770 MachineNode->getOperand(0)); 01771 else if (AllUsersSelectZero(MachineNode)) 01772 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode), 01773 MVT::i1, MachineNode->getOperand(0), 01774 MachineNode->getOperand(1)), 01775 SelectSwap = true; 01776 break; 01777 case PPC::CROR: 01778 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 01779 // x | x = x 01780 ResNode = MachineNode->getOperand(0).getNode(); 01781 else if (Op1Set || Op2Set) 01782 // x | 1 = 1 | y = 1 01783 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 01784 MVT::i1); 01785 else if (Op1Unset) 01786 // 0 | y = y 01787 ResNode = MachineNode->getOperand(1).getNode(); 01788 else if (Op2Unset) 01789 // x | 0 = x 01790 ResNode = MachineNode->getOperand(0).getNode(); 01791 else if (Op1Not) 01792 // ~x | y = orc(y, x) 01793 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 01794 MVT::i1, MachineNode->getOperand(1), 01795 MachineNode->getOperand(0). 01796 getOperand(0)); 01797 else if (Op2Not) 01798 // x | ~y = orc(x, y) 01799 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 01800 MVT::i1, MachineNode->getOperand(0), 01801 MachineNode->getOperand(1). 01802 getOperand(0)); 01803 else if (AllUsersSelectZero(MachineNode)) 01804 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01805 MVT::i1, MachineNode->getOperand(0), 01806 MachineNode->getOperand(1)), 01807 SelectSwap = true; 01808 break; 01809 case PPC::CRXOR: 01810 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 01811 // xor(x, x) = 0 01812 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 01813 MVT::i1); 01814 else if (Op1Set) 01815 // xor(1, y) -> nor(y, y) 01816 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01817 MVT::i1, MachineNode->getOperand(1), 01818 MachineNode->getOperand(1)); 01819 else if (Op2Set) 01820 // xor(x, 1) -> nor(x, x) 01821 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01822 MVT::i1, MachineNode->getOperand(0), 01823 MachineNode->getOperand(0)); 01824 else if (Op1Unset) 01825 // xor(0, y) = y 01826 ResNode = MachineNode->getOperand(1).getNode(); 01827 else if (Op2Unset) 01828 // xor(x, 0) = x 01829 ResNode = MachineNode->getOperand(0).getNode(); 01830 else if (Op1Not) 01831 // xor(~x, y) = eqv(x, y) 01832 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 01833 MVT::i1, MachineNode->getOperand(0). 01834 getOperand(0), 01835 MachineNode->getOperand(1)); 01836 else if (Op2Not) 01837 // xor(x, ~y) = eqv(x, y) 01838 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 01839 MVT::i1, MachineNode->getOperand(0), 01840 MachineNode->getOperand(1). 01841 getOperand(0)); 01842 else if (AllUsersSelectZero(MachineNode)) 01843 ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), 01844 MVT::i1, MachineNode->getOperand(0), 01845 MachineNode->getOperand(1)), 01846 SelectSwap = true; 01847 break; 01848 case PPC::CRNOR: 01849 if (Op1Set || Op2Set) 01850 // nor(1, y) -> 0 01851 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 01852 MVT::i1); 01853 else if (Op1Unset) 01854 // nor(0, y) = ~y -> nor(y, y) 01855 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01856 MVT::i1, MachineNode->getOperand(1), 01857 MachineNode->getOperand(1)); 01858 else if (Op2Unset) 01859 // nor(x, 0) = ~x 01860 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01861 MVT::i1, MachineNode->getOperand(0), 01862 MachineNode->getOperand(0)); 01863 else if (Op1Not) 01864 // nor(~x, y) = andc(x, y) 01865 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 01866 MVT::i1, MachineNode->getOperand(0). 01867 getOperand(0), 01868 MachineNode->getOperand(1)); 01869 else if (Op2Not) 01870 // nor(x, ~y) = andc(y, x) 01871 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 01872 MVT::i1, MachineNode->getOperand(1). 01873 getOperand(0), 01874 MachineNode->getOperand(0)); 01875 else if (AllUsersSelectZero(MachineNode)) 01876 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode), 01877 MVT::i1, MachineNode->getOperand(0), 01878 MachineNode->getOperand(1)), 01879 SelectSwap = true; 01880 break; 01881 case PPC::CREQV: 01882 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 01883 // eqv(x, x) = 1 01884 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 01885 MVT::i1); 01886 else if (Op1Set) 01887 // eqv(1, y) = y 01888 ResNode = MachineNode->getOperand(1).getNode(); 01889 else if (Op2Set) 01890 // eqv(x, 1) = x 01891 ResNode = MachineNode->getOperand(0).getNode(); 01892 else if (Op1Unset) 01893 // eqv(0, y) = ~y -> nor(y, y) 01894 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01895 MVT::i1, MachineNode->getOperand(1), 01896 MachineNode->getOperand(1)); 01897 else if (Op2Unset) 01898 // eqv(x, 0) = ~x 01899 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01900 MVT::i1, MachineNode->getOperand(0), 01901 MachineNode->getOperand(0)); 01902 else if (Op1Not) 01903 // eqv(~x, y) = xor(x, y) 01904 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 01905 MVT::i1, MachineNode->getOperand(0). 01906 getOperand(0), 01907 MachineNode->getOperand(1)); 01908 else if (Op2Not) 01909 // eqv(x, ~y) = xor(x, y) 01910 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 01911 MVT::i1, MachineNode->getOperand(0), 01912 MachineNode->getOperand(1). 01913 getOperand(0)); 01914 else if (AllUsersSelectZero(MachineNode)) 01915 ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), 01916 MVT::i1, MachineNode->getOperand(0), 01917 MachineNode->getOperand(1)), 01918 SelectSwap = true; 01919 break; 01920 case PPC::CRANDC: 01921 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 01922 // andc(x, x) = 0 01923 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 01924 MVT::i1); 01925 else if (Op1Set) 01926 // andc(1, y) = ~y 01927 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01928 MVT::i1, MachineNode->getOperand(1), 01929 MachineNode->getOperand(1)); 01930 else if (Op1Unset || Op2Set) 01931 // andc(0, y) = andc(x, 1) = 0 01932 ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), 01933 MVT::i1); 01934 else if (Op2Unset) 01935 // andc(x, 0) = x 01936 ResNode = MachineNode->getOperand(0).getNode(); 01937 else if (Op1Not) 01938 // andc(~x, y) = ~(x | y) = nor(x, y) 01939 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01940 MVT::i1, MachineNode->getOperand(0). 01941 getOperand(0), 01942 MachineNode->getOperand(1)); 01943 else if (Op2Not) 01944 // andc(x, ~y) = x & y 01945 ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode), 01946 MVT::i1, MachineNode->getOperand(0), 01947 MachineNode->getOperand(1). 01948 getOperand(0)); 01949 else if (AllUsersSelectZero(MachineNode)) 01950 ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), 01951 MVT::i1, MachineNode->getOperand(1), 01952 MachineNode->getOperand(0)), 01953 SelectSwap = true; 01954 break; 01955 case PPC::CRORC: 01956 if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) 01957 // orc(x, x) = 1 01958 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 01959 MVT::i1); 01960 else if (Op1Set || Op2Unset) 01961 // orc(1, y) = orc(x, 0) = 1 01962 ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), 01963 MVT::i1); 01964 else if (Op2Set) 01965 // orc(x, 1) = x 01966 ResNode = MachineNode->getOperand(0).getNode(); 01967 else if (Op1Unset) 01968 // orc(0, y) = ~y 01969 ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), 01970 MVT::i1, MachineNode->getOperand(1), 01971 MachineNode->getOperand(1)); 01972 else if (Op1Not) 01973 // orc(~x, y) = ~(x & y) = nand(x, y) 01974 ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode), 01975 MVT::i1, MachineNode->getOperand(0). 01976 getOperand(0), 01977 MachineNode->getOperand(1)); 01978 else if (Op2Not) 01979 // orc(x, ~y) = x | y 01980 ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode), 01981 MVT::i1, MachineNode->getOperand(0), 01982 MachineNode->getOperand(1). 01983 getOperand(0)); 01984 else if (AllUsersSelectZero(MachineNode)) 01985 ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), 01986 MVT::i1, MachineNode->getOperand(1), 01987 MachineNode->getOperand(0)), 01988 SelectSwap = true; 01989 break; 01990 case PPC::SELECT_I4: 01991 case PPC::SELECT_I8: 01992 case PPC::SELECT_F4: 01993 case PPC::SELECT_F8: 01994 case PPC::SELECT_VRRC: 01995 if (Op1Set) 01996 ResNode = MachineNode->getOperand(1).getNode(); 01997 else if (Op1Unset) 01998 ResNode = MachineNode->getOperand(2).getNode(); 01999 else if (Op1Not) 02000 ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(), 02001 SDLoc(MachineNode), 02002 MachineNode->getValueType(0), 02003 MachineNode->getOperand(0). 02004 getOperand(0), 02005 MachineNode->getOperand(2), 02006 MachineNode->getOperand(1)); 02007 break; 02008 case PPC::BC: 02009 case PPC::BCn: 02010 if (Op1Not) 02011 ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn : 02012 PPC::BC, 02013 SDLoc(MachineNode), 02014 MVT::Other, 02015 MachineNode->getOperand(0). 02016 getOperand(0), 02017 MachineNode->getOperand(1), 02018 MachineNode->getOperand(2)); 02019 // FIXME: Handle Op1Set, Op1Unset here too. 02020 break; 02021 } 02022 02023 // If we're inverting this node because it is used only by selects that 02024 // we'd like to swap, then swap the selects before the node replacement. 02025 if (SelectSwap) 02026 SwapAllSelectUsers(MachineNode); 02027 02028 if (ResNode != MachineNode) { 02029 DEBUG(dbgs() << "CR Peephole replacing:\nOld: "); 02030 DEBUG(MachineNode->dump(CurDAG)); 02031 DEBUG(dbgs() << "\nNew: "); 02032 DEBUG(ResNode->dump(CurDAG)); 02033 DEBUG(dbgs() << "\n"); 02034 02035 ReplaceUses(MachineNode, ResNode); 02036 IsModified = true; 02037 } 02038 } 02039 if (IsModified) 02040 CurDAG->RemoveDeadNodes(); 02041 } while (IsModified); 02042 } 02043 02044 void PPCDAGToDAGISel::PeepholePPC64() { 02045 // These optimizations are currently supported only for 64-bit SVR4. 02046 if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64()) 02047 return; 02048 02049 SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode()); 02050 ++Position; 02051 02052 while (Position != CurDAG->allnodes_begin()) { 02053 SDNode *N = --Position; 02054 // Skip dead nodes and any non-machine opcodes. 02055 if (N->use_empty() || !N->isMachineOpcode()) 02056 continue; 02057 02058 unsigned FirstOp; 02059 unsigned StorageOpcode = N->getMachineOpcode(); 02060 02061 switch (StorageOpcode) { 02062 default: continue; 02063 02064 case PPC::LBZ: 02065 case PPC::LBZ8: 02066 case PPC::LD: 02067 case PPC::LFD: 02068 case PPC::LFS: 02069 case PPC::LHA: 02070 case PPC::LHA8: 02071 case PPC::LHZ: 02072 case PPC::LHZ8: 02073 case PPC::LWA: 02074 case PPC::LWZ: 02075 case PPC::LWZ8: 02076 FirstOp = 0; 02077 break; 02078 02079 case PPC::STB: 02080 case PPC::STB8: 02081 case PPC::STD: 02082 case PPC::STFD: 02083 case PPC::STFS: 02084 case PPC::STH: 02085 case PPC::STH8: 02086 case PPC::STW: 02087 case PPC::STW8: 02088 FirstOp = 1; 02089 break; 02090 } 02091 02092 // If this is a load or store with a zero offset, we may be able to 02093 // fold an add-immediate into the memory operation. 02094 if (!isa<ConstantSDNode>(N->getOperand(FirstOp)) || 02095 N->getConstantOperandVal(FirstOp) != 0) 02096 continue; 02097 02098 SDValue Base = N->getOperand(FirstOp + 1); 02099 if (!Base.isMachineOpcode()) 02100 continue; 02101 02102 unsigned Flags = 0; 02103 bool ReplaceFlags = true; 02104 02105 // When the feeding operation is an add-immediate of some sort, 02106 // determine whether we need to add relocation information to the 02107 // target flags on the immediate operand when we fold it into the 02108 // load instruction. 02109 // 02110 // For something like ADDItocL, the relocation information is 02111 // inferred from the opcode; when we process it in the AsmPrinter, 02112 // we add the necessary relocation there. A load, though, can receive 02113 // relocation from various flavors of ADDIxxx, so we need to carry 02114 // the relocation information in the target flags. 02115 switch (Base.getMachineOpcode()) { 02116 default: continue; 02117 02118 case PPC::ADDI8: 02119 case PPC::ADDI: 02120 // In some cases (such as TLS) the relocation information 02121 // is already in place on the operand, so copying the operand 02122 // is sufficient. 02123 ReplaceFlags = false; 02124 // For these cases, the immediate may not be divisible by 4, in 02125 // which case the fold is illegal for DS-form instructions. (The 02126 // other cases provide aligned addresses and are always safe.) 02127 if ((StorageOpcode == PPC::LWA || 02128 StorageOpcode == PPC::LD || 02129 StorageOpcode == PPC::STD) && 02130 (!isa<ConstantSDNode>(Base.getOperand(1)) || 02131 Base.getConstantOperandVal(1) % 4 != 0)) 02132 continue; 02133 break; 02134 case PPC::ADDIdtprelL: 02135 Flags = PPCII::MO_DTPREL_LO; 02136 break; 02137 case PPC::ADDItlsldL: 02138 Flags = PPCII::MO_TLSLD_LO; 02139 break; 02140 case PPC::ADDItocL: 02141 Flags = PPCII::MO_TOC_LO; 02142 break; 02143 } 02144 02145 // We found an opportunity. Reverse the operands from the add 02146 // immediate and substitute them into the load or store. If 02147 // needed, update the target flags for the immediate operand to 02148 // reflect the necessary relocation information. 02149 DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: "); 02150 DEBUG(Base->dump(CurDAG)); 02151 DEBUG(dbgs() << "\nN: "); 02152 DEBUG(N->dump(CurDAG)); 02153 DEBUG(dbgs() << "\n"); 02154 02155 SDValue ImmOpnd = Base.getOperand(1); 02156 02157 // If the relocation information isn't already present on the 02158 // immediate operand, add it now. 02159 if (ReplaceFlags) { 02160 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) { 02161 SDLoc dl(GA); 02162 const GlobalValue *GV = GA->getGlobal(); 02163 // We can't perform this optimization for data whose alignment 02164 // is insufficient for the instruction encoding. 02165 if (GV->getAlignment() < 4 && 02166 (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD || 02167 StorageOpcode == PPC::LWA)) { 02168 DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n"); 02169 continue; 02170 } 02171 ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, 0, Flags); 02172 } else if (ConstantPoolSDNode *CP = 02173 dyn_cast<ConstantPoolSDNode>(ImmOpnd)) { 02174 const Constant *C = CP->getConstVal(); 02175 ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64, 02176 CP->getAlignment(), 02177 0, Flags); 02178 } 02179 } 02180 02181 if (FirstOp == 1) // Store 02182 (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd, 02183 Base.getOperand(0), N->getOperand(3)); 02184 else // Load 02185 (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0), 02186 N->getOperand(2)); 02187 02188 // The add-immediate may now be dead, in which case remove it. 02189 if (Base.getNode()->use_empty()) 02190 CurDAG->RemoveDeadNode(Base.getNode()); 02191 } 02192 } 02193 02194 02195 /// createPPCISelDag - This pass converts a legalized DAG into a 02196 /// PowerPC-specific DAG, ready for instruction scheduling. 02197 /// 02198 FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) { 02199 return new PPCDAGToDAGISel(TM); 02200 } 02201 02202 static void initializePassOnce(PassRegistry &Registry) { 02203 const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection"; 02204 PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID, 02205 nullptr, false, false); 02206 Registry.registerPass(*PI, true); 02207 } 02208 02209 void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) { 02210 CALL_ONCE_INITIALIZATION(initializePassOnce); 02211 } 02212