lib/Target/CellSPU/SPUISelDAGToDAG.cpp

   1 //===-- SPUISelDAGToDAG.cpp - CellSPU 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 the Cell SPU,
  11 // converting from a legalized dag to a SPU-target dag.
  12 //
  13 //===----------------------------------------------------------------------===//
  14
  15 #include "SPU.h"
  16 #include "SPUTargetMachine.h"
  17 #include "SPUISelLowering.h"
  18 #include "SPUHazardRecognizers.h"
  19 #include "SPUFrameInfo.h"
  20 #include "SPURegisterNames.h"
  21 #include "SPUTargetMachine.h"
  22 #include "llvm/CodeGen/MachineConstantPool.h"
  23 #include "llvm/CodeGen/MachineInstrBuilder.h"
  24 #include "llvm/CodeGen/MachineFunction.h"
  25 #include "llvm/CodeGen/SelectionDAG.h"
  26 #include "llvm/CodeGen/SelectionDAGISel.h"
  27 #include "llvm/CodeGen/PseudoSourceValue.h"
  28 #include "llvm/Target/TargetOptions.h"
  29 #include "llvm/ADT/Statistic.h"
  30 #include "llvm/Constants.h"
  31 #include "llvm/GlobalValue.h"
  32 #include "llvm/Intrinsics.h"
  33 #include "llvm/LLVMContext.h"
  34 #include "llvm/Support/Debug.h"
  35 #include "llvm/Support/ErrorHandling.h"
  36 #include "llvm/Support/MathExtras.h"
  37 #include "llvm/Support/Compiler.h"
  38 #include "llvm/Support/raw_ostream.h"
  39
  40 using namespace llvm;
  41
  42 namespace {
  43   //! ConstantSDNode predicate for i32 sign-extended, 10-bit immediates
  44   bool
  45   isI64IntS10Immediate(ConstantSDNode *CN)
  46   {
  47     return isS10Constant(CN->getSExtValue());
  48   }
  49
  50   //! ConstantSDNode predicate for i32 sign-extended, 10-bit immediates
  51   bool
  52   isI32IntS10Immediate(ConstantSDNode *CN)
  53   {
  54     return isS10Constant(CN->getSExtValue());
  55   }
  56
  57   //! ConstantSDNode predicate for i32 unsigned 10-bit immediate values
  58   bool
  59   isI32IntU10Immediate(ConstantSDNode *CN)
  60   {
  61     return isU10Constant(CN->getSExtValue());
  62   }
  63
  64   //! ConstantSDNode predicate for i16 sign-extended, 10-bit immediate values
  65   bool
  66   isI16IntS10Immediate(ConstantSDNode *CN)
  67   {
  68     return isS10Constant(CN->getSExtValue());
  69   }
  70
  71   //! SDNode predicate for i16 sign-extended, 10-bit immediate values
  72   bool
  73   isI16IntS10Immediate(SDNode *N)
  74   {
  75     ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
  76     return (CN != 0 && isI16IntS10Immediate(CN));
  77   }
  78
  79   //! ConstantSDNode predicate for i16 unsigned 10-bit immediate values
  80   bool
  81   isI16IntU10Immediate(ConstantSDNode *CN)
  82   {
  83     return isU10Constant((short) CN->getZExtValue());
  84   }
  85
  86   //! SDNode predicate for i16 sign-extended, 10-bit immediate values
  87   bool
  88   isI16IntU10Immediate(SDNode *N)
  89   {
  90     return (N->getOpcode() == ISD::Constant
  91             && isI16IntU10Immediate(cast<ConstantSDNode>(N)));
  92   }
  93
  94   //! ConstantSDNode predicate for signed 16-bit values
  95   /*!
  96     \arg CN The constant SelectionDAG node holding the value
  97     \arg Imm The returned 16-bit value, if returning true
  98
  99     This predicate tests the value in \a CN to see whether it can be
 100     represented as a 16-bit, sign-extended quantity. Returns true if
 101     this is the case.
 102    */
 103   bool
 104   isIntS16Immediate(ConstantSDNode *CN, short &Imm)
 105   {
 106     EVT vt = CN->getValueType(0);
 107     Imm = (short) CN->getZExtValue();
 108     if (vt.getSimpleVT() >= MVT::i1 && vt.getSimpleVT() <= MVT::i16) {
 109       return true;
 110     } else if (vt == MVT::i32) {
 111       int32_t i_val = (int32_t) CN->getZExtValue();
 112       short s_val = (short) i_val;
 113       return i_val == s_val;
 114     } else {
 115       int64_t i_val = (int64_t) CN->getZExtValue();
 116       short s_val = (short) i_val;
 117       return i_val == s_val;
 118     }
 119
 120     return false;
 121   }
 122
 123   //! SDNode predicate for signed 16-bit values.
 124   bool
 125   isIntS16Immediate(SDNode *N, short &Imm)
 126   {
 127     return (N->getOpcode() == ISD::Constant
 128             && isIntS16Immediate(cast<ConstantSDNode>(N), Imm));
 129   }
 130
 131   //! ConstantFPSDNode predicate for representing floats as 16-bit sign ext.
 132   static bool
 133   isFPS16Immediate(ConstantFPSDNode *FPN, short &Imm)
 134   {
 135     EVT vt = FPN->getValueType(0);
 136     if (vt == MVT::f32) {
 137       int val = FloatToBits(FPN->getValueAPF().convertToFloat());
 138       int sval = (int) ((val << 16) >> 16);
 139       Imm = (short) val;
 140       return val == sval;
 141     }
 142
 143     return false;
 144   }
 145
 146   bool
 147   isHighLow(const SDValue &Op)
 148   {
 149     return (Op.getOpcode() == SPUISD::IndirectAddr
 150             && ((Op.getOperand(0).getOpcode() == SPUISD::Hi
 151                  && Op.getOperand(1).getOpcode() == SPUISD::Lo)
 152                 || (Op.getOperand(0).getOpcode() == SPUISD::Lo
 153                     && Op.getOperand(1).getOpcode() == SPUISD::Hi)));
 154   }
 155
 156   //===------------------------------------------------------------------===//
 157   //! EVT to "useful stuff" mapping structure:
 158
 159   struct valtype_map_s {
 160     EVT VT;
 161     unsigned ldresult_ins;      /// LDRESULT instruction (0 = undefined)
 162     bool ldresult_imm;          /// LDRESULT instruction requires immediate?
 163     unsigned lrinst;            /// LR instruction
 164   };
 165
 166   const valtype_map_s valtype_map[] = {
 167     { MVT::i8,    SPU::ORBIr8,  true,  SPU::LRr8 },
 168     { MVT::i16,   SPU::ORHIr16, true,  SPU::LRr16 },
 169     { MVT::i32,   SPU::ORIr32,  true,  SPU::LRr32 },
 170     { MVT::i64,   SPU::ORr64,   false, SPU::LRr64 },
 171     { MVT::f32,   SPU::ORf32,   false, SPU::LRf32 },
 172     { MVT::f64,   SPU::ORf64,   false, SPU::LRf64 },
 173     // vector types... (sigh!)
 174     { MVT::v16i8, 0,            false, SPU::LRv16i8 },
 175     { MVT::v8i16, 0,            false, SPU::LRv8i16 },
 176     { MVT::v4i32, 0,            false, SPU::LRv4i32 },
 177     { MVT::v2i64, 0,            false, SPU::LRv2i64 },
 178     { MVT::v4f32, 0,            false, SPU::LRv4f32 },
 179     { MVT::v2f64, 0,            false, SPU::LRv2f64 }
 180   };
 181
 182   const size_t n_valtype_map = sizeof(valtype_map) / sizeof(valtype_map[0]);
 183
 184   const valtype_map_s *getValueTypeMapEntry(EVT VT)
 185   {
 186     const valtype_map_s *retval = 0;
 187     for (size_t i = 0; i < n_valtype_map; ++i) {
 188       if (valtype_map[i].VT == VT) {
 189         retval = valtype_map + i;
 190         break;
 191       }
 192     }
 193
 194
 195 #ifndef NDEBUG
 196     if (retval == 0) {
 197       std::string msg;
 198       raw_string_ostream Msg(msg);
 199       Msg << "SPUISelDAGToDAG.cpp: getValueTypeMapEntry returns NULL for "
 200            << VT.getEVTString();
 201       llvm_report_error(Msg.str());
 202     }
 203 #endif
 204
 205     return retval;
 206   }
 207
 208   //! Generate the carry-generate shuffle mask.
 209   SDValue getCarryGenerateShufMask(SelectionDAG &DAG, DebugLoc dl) {
 210     SmallVector<SDValue, 16 > ShufBytes;
 211
 212     // Create the shuffle mask for "rotating" the borrow up one register slot
 213     // once the borrow is generated.
 214     ShufBytes.push_back(DAG.getConstant(0x04050607, MVT::i32));
 215     ShufBytes.push_back(DAG.getConstant(0x80808080, MVT::i32));
 216     ShufBytes.push_back(DAG.getConstant(0x0c0d0e0f, MVT::i32));
 217     ShufBytes.push_back(DAG.getConstant(0x80808080, MVT::i32));
 218
 219     return DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
 220                        &ShufBytes[0], ShufBytes.size());
 221   }
 222
 223   //! Generate the borrow-generate shuffle mask
 224   SDValue getBorrowGenerateShufMask(SelectionDAG &DAG, DebugLoc dl) {
 225     SmallVector<SDValue, 16 > ShufBytes;
 226
 227     // Create the shuffle mask for "rotating" the borrow up one register slot
 228     // once the borrow is generated.
 229     ShufBytes.push_back(DAG.getConstant(0x04050607, MVT::i32));
 230     ShufBytes.push_back(DAG.getConstant(0xc0c0c0c0, MVT::i32));
 231     ShufBytes.push_back(DAG.getConstant(0x0c0d0e0f, MVT::i32));
 232     ShufBytes.push_back(DAG.getConstant(0xc0c0c0c0, MVT::i32));
 233
 234     return DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
 235                        &ShufBytes[0], ShufBytes.size());
 236   }
 237
 238   //===------------------------------------------------------------------===//
 239   /// SPUDAGToDAGISel - Cell SPU-specific code to select SPU machine
 240   /// instructions for SelectionDAG operations.
 241   ///
 242   class SPUDAGToDAGISel :
 243     public SelectionDAGISel
 244   {
 245     SPUTargetMachine &TM;
 246     SPUTargetLowering &SPUtli;
 247     unsigned GlobalBaseReg;
 248
 249   public:
 250     explicit SPUDAGToDAGISel(SPUTargetMachine &tm) :
 251       SelectionDAGISel(tm),
 252       TM(tm),
 253       SPUtli(*tm.getTargetLowering())
 254     { }
 255
 256     virtual bool runOnMachineFunction(MachineFunction &MF) {
 257       // Make sure we re-emit a set of the global base reg if necessary
 258       GlobalBaseReg = 0;
 259       SelectionDAGISel::runOnMachineFunction(MF);
 260       return true;
 261     }
 262
 263     /// getI32Imm - Return a target constant with the specified value, of type
 264     /// i32.
 265     inline SDValue getI32Imm(uint32_t Imm) {
 266       return CurDAG->getTargetConstant(Imm, MVT::i32);
 267     }
 268
 269     /// getI64Imm - Return a target constant with the specified value, of type
 270     /// i64.
 271     inline SDValue getI64Imm(uint64_t Imm) {
 272       return CurDAG->getTargetConstant(Imm, MVT::i64);
 273     }
 274
 275     /// getSmallIPtrImm - Return a target constant of pointer type.
 276     inline SDValue getSmallIPtrImm(unsigned Imm) {
 277       return CurDAG->getTargetConstant(Imm, SPUtli.getPointerTy());
 278       }
 279
 280     SDNode *emitBuildVector(SDValue build_vec) {
 281       EVT vecVT = build_vec.getValueType();
 282       EVT eltVT = vecVT.getVectorElementType();
 283       SDNode *bvNode = build_vec.getNode();
 284       DebugLoc dl = bvNode->getDebugLoc();
 285
 286       // Check to see if this vector can be represented as a CellSPU immediate
 287       // constant by invoking all of the instruction selection predicates:
 288       if (((vecVT == MVT::v8i16) &&
 289            (SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i16).getNode() != 0)) ||
 290           ((vecVT == MVT::v4i32) &&
 291            ((SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
 292             (SPU::get_ILHUvec_imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
 293             (SPU::get_vec_u18imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
 294             (SPU::get_v4i32_imm(bvNode, *CurDAG).getNode() != 0))) ||
 295           ((vecVT == MVT::v2i64) &&
 296            ((SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i64).getNode() != 0) ||
 297             (SPU::get_ILHUvec_imm(bvNode, *CurDAG, MVT::i64).getNode() != 0) ||
 298             (SPU::get_vec_u18imm(bvNode, *CurDAG, MVT::i64).getNode() != 0))))
 299         return Select(build_vec);
 300
 301       // No, need to emit a constant pool spill:
 302       std::vector<Constant*> CV;
 303
 304       for (size_t i = 0; i < build_vec.getNumOperands(); ++i) {
 305         ConstantSDNode *V = dyn_cast<ConstantSDNode > (build_vec.getOperand(i));
 306         CV.push_back(const_cast<ConstantInt *> (V->getConstantIntValue()));
 307       }
 308
 309       Constant *CP = ConstantVector::get(CV);
 310       SDValue CPIdx = CurDAG->getConstantPool(CP, SPUtli.getPointerTy());
 311       unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
 312       SDValue CGPoolOffset =
 313               SPU::LowerConstantPool(CPIdx, *CurDAG,
 314                                      SPUtli.getSPUTargetMachine());
 315       return SelectCode(CurDAG->getLoad(build_vec.getValueType(), dl,
 316                                         CurDAG->getEntryNode(), CGPoolOffset,
 317                                         PseudoSourceValue::getConstantPool(), 0,
 318                                         false, Alignment));
 319     }
 320
 321     /// Select - Convert the specified operand from a target-independent to a
 322     /// target-specific node if it hasn't already been changed.
 323     SDNode *Select(SDValue Op);
 324
 325     //! Emit the instruction sequence for i64 shl
 326     SDNode *SelectSHLi64(SDValue &Op, EVT OpVT);
 327
 328     //! Emit the instruction sequence for i64 srl
 329     SDNode *SelectSRLi64(SDValue &Op, EVT OpVT);
 330
 331     //! Emit the instruction sequence for i64 sra
 332     SDNode *SelectSRAi64(SDValue &Op, EVT OpVT);
 333
 334     //! Emit the necessary sequence for loading i64 constants:
 335     SDNode *SelectI64Constant(SDValue &Op, EVT OpVT, DebugLoc dl);
 336
 337     //! Alternate instruction emit sequence for loading i64 constants
 338     SDNode *SelectI64Constant(uint64_t i64const, EVT OpVT, DebugLoc dl);
 339
 340     //! Returns true if the address N is an A-form (local store) address
 341     bool SelectAFormAddr(SDValue Op, SDValue N, SDValue &Base,
 342                          SDValue &Index);
 343
 344     //! D-form address predicate
 345     bool SelectDFormAddr(SDValue Op, SDValue N, SDValue &Base,
 346                          SDValue &Index);
 347
 348     /// Alternate D-form address using i7 offset predicate
 349     bool SelectDForm2Addr(SDValue Op, SDValue N, SDValue &Disp,
 350                           SDValue &Base);
 351
 352     /// D-form address selection workhorse
 353     bool DFormAddressPredicate(SDValue Op, SDValue N, SDValue &Disp,
 354                                SDValue &Base, int minOffset, int maxOffset);
 355
 356     //! Address predicate if N can be expressed as an indexed [r+r] operation.
 357     bool SelectXFormAddr(SDValue Op, SDValue N, SDValue &Base,
 358                          SDValue &Index);
 359
 360     /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
 361     /// inline asm expressions.
 362     virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
 363                                               char ConstraintCode,
 364                                               std::vector<SDValue> &OutOps) {
 365       SDValue Op0, Op1;
 366       switch (ConstraintCode) {
 367       default: return true;
 368       case 'm':   // memory
 369         if (!SelectDFormAddr(Op, Op, Op0, Op1)
 370             && !SelectAFormAddr(Op, Op, Op0, Op1))
 371           SelectXFormAddr(Op, Op, Op0, Op1);
 372         break;
 373       case 'o':   // offsetable
 374         if (!SelectDFormAddr(Op, Op, Op0, Op1)
 375             && !SelectAFormAddr(Op, Op, Op0, Op1)) {
 376           Op0 = Op;
 377           Op1 = getSmallIPtrImm(0);
 378         }
 379         break;
 380       case 'v':   // not offsetable
 381 #if 1
 382         llvm_unreachable("InlineAsmMemoryOperand 'v' constraint not handled.");
 383 #else
 384         SelectAddrIdxOnly(Op, Op, Op0, Op1);
 385 #endif
 386         break;
 387       }
 388
 389       OutOps.push_back(Op0);
 390       OutOps.push_back(Op1);
 391       return false;
 392     }
 393
 394     /// InstructionSelect - This callback is invoked by
 395     /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
 396     virtual void InstructionSelect();
 397
 398     virtual const char *getPassName() const {
 399       return "Cell SPU DAG->DAG Pattern Instruction Selection";
 400     }
 401
 402     /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
 403     /// this target when scheduling the DAG.
 404     virtual ScheduleHazardRecognizer *CreateTargetHazardRecognizer() {
 405       const TargetInstrInfo *II = TM.getInstrInfo();
 406       assert(II && "No InstrInfo?");
 407       return new SPUHazardRecognizer(*II);
 408     }
 409
 410     // Include the pieces autogenerated from the target description.
 411 #include "SPUGenDAGISel.inc"
 412   };
 413 }
 414
 415 /// InstructionSelect - This callback is invoked by
 416 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
 417 void
 418 SPUDAGToDAGISel::InstructionSelect()
 419 {
 420   DEBUG(BB->dump());
 421
 422   // Select target instructions for the DAG.
 423   SelectRoot(*CurDAG);
 424   CurDAG->RemoveDeadNodes();
 425 }
 426
 427 /*!
 428  \arg Op The ISD instruction operand
 429  \arg N The address to be tested
 430  \arg Base The base address
 431  \arg Index The base address index
 432  */
 433 bool
 434 SPUDAGToDAGISel::SelectAFormAddr(SDValue Op, SDValue N, SDValue &Base,
 435                     SDValue &Index) {
 436   // These match the addr256k operand type:
 437   EVT OffsVT = MVT::i16;
 438   SDValue Zero = CurDAG->getTargetConstant(0, OffsVT);
 439
 440   switch (N.getOpcode()) {
 441   case ISD::Constant:
 442   case ISD::ConstantPool:
 443   case ISD::GlobalAddress:
 444     llvm_report_error("SPU SelectAFormAddr: Constant/Pool/Global not lowered.");
 445     /*NOTREACHED*/
 446
 447   case ISD::TargetConstant:
 448   case ISD::TargetGlobalAddress:
 449   case ISD::TargetJumpTable:
 450     llvm_report_error("SPUSelectAFormAddr: Target Constant/Pool/Global "
 451                       "not wrapped as A-form address.");
 452     /*NOTREACHED*/
 453
 454   case SPUISD::AFormAddr:
 455     // Just load from memory if there's only a single use of the location,
 456     // otherwise, this will get handled below with D-form offset addresses
 457     if (N.hasOneUse()) {
 458       SDValue Op0 = N.getOperand(0);
 459       switch (Op0.getOpcode()) {
 460       case ISD::TargetConstantPool:
 461       case ISD::TargetJumpTable:
 462         Base = Op0;
 463         Index = Zero;
 464         return true;
 465
 466       case ISD::TargetGlobalAddress: {
 467         GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op0);
 468         GlobalValue *GV = GSDN->getGlobal();
 469         if (GV->getAlignment() == 16) {
 470           Base = Op0;
 471           Index = Zero;
 472           return true;
 473         }
 474         break;
 475       }
 476       }
 477     }
 478     break;
 479   }
 480   return false;
 481 }
 482
 483 bool
 484 SPUDAGToDAGISel::SelectDForm2Addr(SDValue Op, SDValue N, SDValue &Disp,
 485                                   SDValue &Base) {
 486   const int minDForm2Offset = -(1 << 7);
 487   const int maxDForm2Offset = (1 << 7) - 1;
 488   return DFormAddressPredicate(Op, N, Disp, Base, minDForm2Offset,
 489                                maxDForm2Offset);
 490 }
 491
 492 /*!
 493   \arg Op The ISD instruction (ignored)
 494   \arg N The address to be tested
 495   \arg Base Base address register/pointer
 496   \arg Index Base address index
 497
 498   Examine the input address by a base register plus a signed 10-bit
 499   displacement, [r+I10] (D-form address).
 500
 501   \return true if \a N is a D-form address with \a Base and \a Index set
 502   to non-empty SDValue instances.
 503 */
 504 bool
 505 SPUDAGToDAGISel::SelectDFormAddr(SDValue Op, SDValue N, SDValue &Base,
 506                                  SDValue &Index) {
 507   return DFormAddressPredicate(Op, N, Base, Index,
 508                                SPUFrameInfo::minFrameOffset(),
 509                                SPUFrameInfo::maxFrameOffset());
 510 }
 511
 512 bool
 513 SPUDAGToDAGISel::DFormAddressPredicate(SDValue Op, SDValue N, SDValue &Base,
 514                                       SDValue &Index, int minOffset,
 515                                       int maxOffset) {
 516   unsigned Opc = N.getOpcode();
 517   EVT PtrTy = SPUtli.getPointerTy();
 518
 519   if (Opc == ISD::FrameIndex) {
 520     // Stack frame index must be less than 512 (divided by 16):
 521     FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(N);
 522     int FI = int(FIN->getIndex());
 523     DEBUG(errs() << "SelectDFormAddr: ISD::FrameIndex = "
 524                << FI << "\n");
 525     if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
 526       Base = CurDAG->getTargetConstant(0, PtrTy);
 527       Index = CurDAG->getTargetFrameIndex(FI, PtrTy);
 528       return true;
 529     }
 530   } else if (Opc == ISD::ADD) {
 531     // Generated by getelementptr
 532     const SDValue Op0 = N.getOperand(0);
 533     const SDValue Op1 = N.getOperand(1);
 534
 535     if ((Op0.getOpcode() == SPUISD::Hi && Op1.getOpcode() == SPUISD::Lo)
 536         || (Op1.getOpcode() == SPUISD::Hi && Op0.getOpcode() == SPUISD::Lo)) {
 537       Base = CurDAG->getTargetConstant(0, PtrTy);
 538       Index = N;
 539       return true;
 540     } else if (Op1.getOpcode() == ISD::Constant
 541                || Op1.getOpcode() == ISD::TargetConstant) {
 542       ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1);
 543       int32_t offset = int32_t(CN->getSExtValue());
 544
 545       if (Op0.getOpcode() == ISD::FrameIndex) {
 546         FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op0);
 547         int FI = int(FIN->getIndex());
 548         DEBUG(errs() << "SelectDFormAddr: ISD::ADD offset = " << offset
 549                    << " frame index = " << FI << "\n");
 550
 551         if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
 552           Base = CurDAG->getTargetConstant(offset, PtrTy);
 553           Index = CurDAG->getTargetFrameIndex(FI, PtrTy);
 554           return true;
 555         }
 556       } else if (offset > minOffset && offset < maxOffset) {
 557         Base = CurDAG->getTargetConstant(offset, PtrTy);
 558         Index = Op0;
 559         return true;
 560       }
 561     } else if (Op0.getOpcode() == ISD::Constant
 562                || Op0.getOpcode() == ISD::TargetConstant) {
 563       ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op0);
 564       int32_t offset = int32_t(CN->getSExtValue());
 565
 566       if (Op1.getOpcode() == ISD::FrameIndex) {
 567         FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op1);
 568         int FI = int(FIN->getIndex());
 569         DEBUG(errs() << "SelectDFormAddr: ISD::ADD offset = " << offset
 570                    << " frame index = " << FI << "\n");
 571
 572         if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
 573           Base = CurDAG->getTargetConstant(offset, PtrTy);
 574           Index = CurDAG->getTargetFrameIndex(FI, PtrTy);
 575           return true;
 576         }
 577       } else if (offset > minOffset && offset < maxOffset) {
 578         Base = CurDAG->getTargetConstant(offset, PtrTy);
 579         Index = Op1;
 580         return true;
 581       }
 582     }
 583   } else if (Opc == SPUISD::IndirectAddr) {
 584     // Indirect with constant offset -> D-Form address
 585     const SDValue Op0 = N.getOperand(0);
 586     const SDValue Op1 = N.getOperand(1);
 587
 588     if (Op0.getOpcode() == SPUISD::Hi
 589         && Op1.getOpcode() == SPUISD::Lo) {
 590       // (SPUindirect (SPUhi <arg>, 0), (SPUlo <arg>, 0))
 591       Base = CurDAG->getTargetConstant(0, PtrTy);
 592       Index = N;
 593       return true;
 594     } else if (isa<ConstantSDNode>(Op0) || isa<ConstantSDNode>(Op1)) {
 595       int32_t offset = 0;
 596       SDValue idxOp;
 597
 598       if (isa<ConstantSDNode>(Op1)) {
 599         ConstantSDNode *CN = cast<ConstantSDNode>(Op1);
 600         offset = int32_t(CN->getSExtValue());
 601         idxOp = Op0;
 602       } else if (isa<ConstantSDNode>(Op0)) {
 603         ConstantSDNode *CN = cast<ConstantSDNode>(Op0);
 604         offset = int32_t(CN->getSExtValue());
 605         idxOp = Op1;
 606       }
 607
 608       if (offset >= minOffset && offset <= maxOffset) {
 609         Base = CurDAG->getTargetConstant(offset, PtrTy);
 610         Index = idxOp;
 611         return true;
 612       }
 613     }
 614   } else if (Opc == SPUISD::AFormAddr) {
 615     Base = CurDAG->getTargetConstant(0, N.getValueType());
 616     Index = N;
 617     return true;
 618   } else if (Opc == SPUISD::LDRESULT) {
 619     Base = CurDAG->getTargetConstant(0, N.getValueType());
 620     Index = N;
 621     return true;
 622   } else if (Opc == ISD::Register || Opc == ISD::CopyFromReg) {
 623     unsigned OpOpc = Op.getOpcode();
 624
 625     if (OpOpc == ISD::STORE || OpOpc == ISD::LOAD) {
 626       // Direct load/store without getelementptr
 627       SDValue Addr, Offs;
 628
 629       // Get the register from CopyFromReg
 630       if (Opc == ISD::CopyFromReg)
 631         Addr = N.getOperand(1);
 632       else
 633         Addr = N;                       // Register
 634
 635       Offs = ((OpOpc == ISD::STORE) ? Op.getOperand(3) : Op.getOperand(2));
 636
 637       if (Offs.getOpcode() == ISD::Constant || Offs.getOpcode() == ISD::UNDEF) {
 638         if (Offs.getOpcode() == ISD::UNDEF)
 639           Offs = CurDAG->getTargetConstant(0, Offs.getValueType());
 640
 641         Base = Offs;
 642         Index = Addr;
 643         return true;
 644       }
 645     } else {
 646       /* If otherwise unadorned, default to D-form address with 0 offset: */
 647       if (Opc == ISD::CopyFromReg) {
 648         Index = N.getOperand(1);
 649       } else {
 650         Index = N;
 651       }
 652
 653       Base = CurDAG->getTargetConstant(0, Index.getValueType());
 654       return true;
 655     }
 656   }
 657
 658   return false;
 659 }
 660
 661 /*!
 662   \arg Op The ISD instruction operand
 663   \arg N The address operand
 664   \arg Base The base pointer operand
 665   \arg Index The offset/index operand
 666
 667   If the address \a N can be expressed as an A-form or D-form address, returns
 668   false.  Otherwise, creates two operands, Base and Index that will become the
 669   (r)(r) X-form address.
 670 */
 671 bool
 672 SPUDAGToDAGISel::SelectXFormAddr(SDValue Op, SDValue N, SDValue &Base,
 673                                  SDValue &Index) {
 674   if (!SelectAFormAddr(Op, N, Base, Index)
 675       && !SelectDFormAddr(Op, N, Base, Index)) {
 676     // If the address is neither A-form or D-form, punt and use an X-form
 677     // address:
 678     Base = N.getOperand(1);
 679     Index = N.getOperand(0);
 680     return true;
 681   }
 682
 683   return false;
 684 }
 685
 686 //! Convert the operand from a target-independent to a target-specific node
 687 /*!
 688  */
 689 SDNode *
 690 SPUDAGToDAGISel::Select(SDValue Op) {
 691   SDNode *N = Op.getNode();
 692   unsigned Opc = N->getOpcode();
 693   int n_ops = -1;
 694   unsigned NewOpc;
 695   EVT OpVT = Op.getValueType();
 696   SDValue Ops[8];
 697   DebugLoc dl = N->getDebugLoc();
 698
 699   if (N->isMachineOpcode()) {
 700     return NULL;   // Already selected.
 701   }
 702
 703   if (Opc == ISD::FrameIndex) {
 704     int FI = cast<FrameIndexSDNode>(N)->getIndex();
 705     SDValue TFI = CurDAG->getTargetFrameIndex(FI, Op.getValueType());
 706     SDValue Imm0 = CurDAG->getTargetConstant(0, Op.getValueType());
 707
 708     if (FI < 128) {
 709       NewOpc = SPU::AIr32;
 710       Ops[0] = TFI;
 711       Ops[1] = Imm0;
 712       n_ops = 2;
 713     } else {
 714       NewOpc = SPU::Ar32;
 715       Ops[0] = CurDAG->getRegister(SPU::R1, Op.getValueType());
 716       Ops[1] = SDValue(CurDAG->getTargetNode(SPU::ILAr32, dl, Op.getValueType(),
 717                                              TFI, Imm0), 0);
 718       n_ops = 2;
 719     }
 720   } else if (Opc == ISD::Constant && OpVT == MVT::i64) {
 721     // Catch the i64 constants that end up here. Note: The backend doesn't
 722     // attempt to legalize the constant (it's useless because DAGCombiner
 723     // will insert 64-bit constants and we can't stop it).
 724     return SelectI64Constant(Op, OpVT, Op.getDebugLoc());
 725   } else if ((Opc == ISD::ZERO_EXTEND || Opc == ISD::ANY_EXTEND)
 726              && OpVT == MVT::i64) {
 727     SDValue Op0 = Op.getOperand(0);
 728     EVT Op0VT = Op0.getValueType();
 729     EVT Op0VecVT = EVT::getVectorVT(*CurDAG->getContext(),
 730                                     Op0VT, (128 / Op0VT.getSizeInBits()));
 731     EVT OpVecVT = EVT::getVectorVT(*CurDAG->getContext(),
 732                                    OpVT, (128 / OpVT.getSizeInBits()));
 733     SDValue shufMask;
 734
 735     switch (Op0VT.getSimpleVT().SimpleTy) {
 736     default:
 737       llvm_report_error("CellSPU Select: Unhandled zero/any extend EVT");
 738       /*NOTREACHED*/
 739     case MVT::i32:
 740       shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
 741                                  CurDAG->getConstant(0x80808080, MVT::i32),
 742                                  CurDAG->getConstant(0x00010203, MVT::i32),
 743                                  CurDAG->getConstant(0x80808080, MVT::i32),
 744                                  CurDAG->getConstant(0x08090a0b, MVT::i32));
 745       break;
 746
 747     case MVT::i16:
 748       shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
 749                                  CurDAG->getConstant(0x80808080, MVT::i32),
 750                                  CurDAG->getConstant(0x80800203, MVT::i32),
 751                                  CurDAG->getConstant(0x80808080, MVT::i32),
 752                                  CurDAG->getConstant(0x80800a0b, MVT::i32));
 753       break;
 754
 755     case MVT::i8:
 756       shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
 757                                  CurDAG->getConstant(0x80808080, MVT::i32),
 758                                  CurDAG->getConstant(0x80808003, MVT::i32),
 759                                  CurDAG->getConstant(0x80808080, MVT::i32),
 760                                  CurDAG->getConstant(0x8080800b, MVT::i32));
 761       break;
 762     }
 763
 764     SDNode *shufMaskLoad = emitBuildVector(shufMask);
 765     SDNode *PromoteScalar =
 766             SelectCode(CurDAG->getNode(SPUISD::PREFSLOT2VEC, dl, Op0VecVT, Op0));
 767
 768     SDValue zextShuffle =
 769             CurDAG->getNode(SPUISD::SHUFB, dl, OpVecVT,
 770                             SDValue(PromoteScalar, 0),
 771                             SDValue(PromoteScalar, 0),
 772                             SDValue(shufMaskLoad, 0));
 773
 774     // N.B.: BIT_CONVERT replaces and updates the zextShuffle node, so we
 775     // re-use it in the VEC2PREFSLOT selection without needing to explicitly
 776     // call SelectCode (it's already done for us.)
 777     SelectCode(CurDAG->getNode(ISD::BIT_CONVERT, dl, OpVecVT, zextShuffle));
 778     return SelectCode(CurDAG->getNode(SPUISD::VEC2PREFSLOT, dl, OpVT,
 779                                       zextShuffle));
 780   } else if (Opc == ISD::ADD && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
 781     SDNode *CGLoad =
 782             emitBuildVector(getCarryGenerateShufMask(*CurDAG, dl));
 783
 784     return SelectCode(CurDAG->getNode(SPUISD::ADD64_MARKER, dl, OpVT,
 785                                       Op.getOperand(0), Op.getOperand(1),
 786                                       SDValue(CGLoad, 0)));
 787   } else if (Opc == ISD::SUB && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
 788     SDNode *CGLoad =
 789             emitBuildVector(getBorrowGenerateShufMask(*CurDAG, dl));
 790
 791     return SelectCode(CurDAG->getNode(SPUISD::SUB64_MARKER, dl, OpVT,
 792                                       Op.getOperand(0), Op.getOperand(1),
 793                                       SDValue(CGLoad, 0)));
 794   } else if (Opc == ISD::MUL && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
 795     SDNode *CGLoad =
 796             emitBuildVector(getCarryGenerateShufMask(*CurDAG, dl));
 797
 798     return SelectCode(CurDAG->getNode(SPUISD::MUL64_MARKER, dl, OpVT,
 799                                       Op.getOperand(0), Op.getOperand(1),
 800                                       SDValue(CGLoad, 0)));
 801   } else if (Opc == ISD::TRUNCATE) {
 802     SDValue Op0 = Op.getOperand(0);
 803     if ((Op0.getOpcode() == ISD::SRA || Op0.getOpcode() == ISD::SRL)
 804         && OpVT == MVT::i32
 805         && Op0.getValueType() == MVT::i64) {
 806       // Catch (truncate:i32 ([sra|srl]:i64 arg, c), where c >= 32
 807       //
 808       // Take advantage of the fact that the upper 32 bits are in the
 809       // i32 preferred slot and avoid shuffle gymnastics:
 810       ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op0.getOperand(1));
 811       if (CN != 0) {
 812         unsigned shift_amt = unsigned(CN->getZExtValue());
 813
 814         if (shift_amt >= 32) {
 815           SDNode *hi32 =
 816                   CurDAG->getTargetNode(SPU::ORr32_r64, dl, OpVT,
 817                                         Op0.getOperand(0));
 818
 819           shift_amt -= 32;
 820           if (shift_amt > 0) {
 821             // Take care of the additional shift, if present:
 822             SDValue shift = CurDAG->getTargetConstant(shift_amt, MVT::i32);
 823             unsigned Opc = SPU::ROTMAIr32_i32;
 824
 825             if (Op0.getOpcode() == ISD::SRL)
 826               Opc = SPU::ROTMr32;
 827
 828             hi32 = CurDAG->getTargetNode(Opc, dl, OpVT, SDValue(hi32, 0),
 829                                          shift);
 830           }
 831
 832           return hi32;
 833         }
 834       }
 835     }
 836   } else if (Opc == ISD::SHL) {
 837     if (OpVT == MVT::i64) {
 838       return SelectSHLi64(Op, OpVT);
 839     }
 840   } else if (Opc == ISD::SRL) {
 841     if (OpVT == MVT::i64) {
 842       return SelectSRLi64(Op, OpVT);
 843     }
 844   } else if (Opc == ISD::SRA) {
 845     if (OpVT == MVT::i64) {
 846       return SelectSRAi64(Op, OpVT);
 847     }
 848   } else if (Opc == ISD::FNEG
 849              && (OpVT == MVT::f64 || OpVT == MVT::v2f64)) {
 850     DebugLoc dl = Op.getDebugLoc();
 851     // Check if the pattern is a special form of DFNMS:
 852     // (fneg (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))
 853     SDValue Op0 = Op.getOperand(0);
 854     if (Op0.getOpcode() == ISD::FSUB) {
 855       SDValue Op00 = Op0.getOperand(0);
 856       if (Op00.getOpcode() == ISD::FMUL) {
 857         unsigned Opc = SPU::DFNMSf64;
 858         if (OpVT == MVT::v2f64)
 859           Opc = SPU::DFNMSv2f64;
 860
 861         return CurDAG->getTargetNode(Opc, dl, OpVT,
 862                                      Op00.getOperand(0),
 863                                      Op00.getOperand(1),
 864                                      Op0.getOperand(1));
 865       }
 866     }
 867
 868     SDValue negConst = CurDAG->getConstant(0x8000000000000000ULL, MVT::i64);
 869     SDNode *signMask = 0;
 870     unsigned Opc = SPU::XORfneg64;
 871
 872     if (OpVT == MVT::f64) {
 873       signMask = SelectI64Constant(negConst, MVT::i64, dl);
 874     } else if (OpVT == MVT::v2f64) {
 875       Opc = SPU::XORfnegvec;
 876       signMask = emitBuildVector(CurDAG->getNode(ISD::BUILD_VECTOR, dl,
 877                                                  MVT::v2i64,
 878                                                  negConst, negConst));
 879     }
 880
 881     return CurDAG->getTargetNode(Opc, dl, OpVT,
 882                                  Op.getOperand(0), SDValue(signMask, 0));
 883   } else if (Opc == ISD::FABS) {
 884     if (OpVT == MVT::f64) {
 885       SDNode *signMask = SelectI64Constant(0x7fffffffffffffffULL, MVT::i64, dl);
 886       return CurDAG->getTargetNode(SPU::ANDfabs64, dl, OpVT,
 887                                    Op.getOperand(0), SDValue(signMask, 0));
 888     } else if (OpVT == MVT::v2f64) {
 889       SDValue absConst = CurDAG->getConstant(0x7fffffffffffffffULL, MVT::i64);
 890       SDValue absVec = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64,
 891                                        absConst, absConst);
 892       SDNode *signMask = emitBuildVector(absVec);
 893       return CurDAG->getTargetNode(SPU::ANDfabsvec, dl, OpVT,
 894                                    Op.getOperand(0), SDValue(signMask, 0));
 895     }
 896   } else if (Opc == SPUISD::LDRESULT) {
 897     // Custom select instructions for LDRESULT
 898     EVT VT = N->getValueType(0);
 899     SDValue Arg = N->getOperand(0);
 900     SDValue Chain = N->getOperand(1);
 901     SDNode *Result;
 902     const valtype_map_s *vtm = getValueTypeMapEntry(VT);
 903
 904     if (vtm->ldresult_ins == 0) {
 905       std::string msg;
 906       raw_string_ostream Msg(msg);
 907       Msg << "LDRESULT for unsupported type: "
 908            << VT.getEVTString();
 909       llvm_report_error(Msg.str());
 910     }
 911
 912     Opc = vtm->ldresult_ins;
 913     if (vtm->ldresult_imm) {
 914       SDValue Zero = CurDAG->getTargetConstant(0, VT);
 915
 916       Result = CurDAG->getTargetNode(Opc, dl, VT, MVT::Other, Arg, Zero, Chain);
 917     } else {
 918       Result = CurDAG->getTargetNode(Opc, dl, VT, MVT::Other, Arg, Arg, Chain);
 919     }
 920
 921     return Result;
 922   } else if (Opc == SPUISD::IndirectAddr) {
 923     // Look at the operands: SelectCode() will catch the cases that aren't
 924     // specifically handled here.
 925     //
 926     // SPUInstrInfo catches the following patterns:
 927     // (SPUindirect (SPUhi ...), (SPUlo ...))
 928     // (SPUindirect $sp, imm)
 929     EVT VT = Op.getValueType();
 930     SDValue Op0 = N->getOperand(0);
 931     SDValue Op1 = N->getOperand(1);
 932     RegisterSDNode *RN;
 933
 934     if ((Op0.getOpcode() != SPUISD::Hi && Op1.getOpcode() != SPUISD::Lo)
 935         || (Op0.getOpcode() == ISD::Register
 936             && ((RN = dyn_cast<RegisterSDNode>(Op0.getNode())) != 0
 937                 && RN->getReg() != SPU::R1))) {
 938       NewOpc = SPU::Ar32;
 939       if (Op1.getOpcode() == ISD::Constant) {
 940         ConstantSDNode *CN = cast<ConstantSDNode>(Op1);
 941         Op1 = CurDAG->getTargetConstant(CN->getSExtValue(), VT);
 942         NewOpc = (isI32IntS10Immediate(CN) ? SPU::AIr32 : SPU::Ar32);
 943       }
 944       Ops[0] = Op0;
 945       Ops[1] = Op1;
 946       n_ops = 2;
 947     }
 948   }
 949
 950   if (n_ops > 0) {
 951     if (N->hasOneUse())
 952       return CurDAG->SelectNodeTo(N, NewOpc, OpVT, Ops, n_ops);
 953     else
 954       return CurDAG->getTargetNode(NewOpc, dl, OpVT, Ops, n_ops);
 955   } else
 956     return SelectCode(Op);
 957 }
 958
 959 /*!
 960  * Emit the instruction sequence for i64 left shifts. The basic algorithm
 961  * is to fill the bottom two word slots with zeros so that zeros are shifted
 962  * in as the entire quadword is shifted left.
 963  *
 964  * \note This code could also be used to implement v2i64 shl.
 965  *
 966  * @param Op The shl operand
 967  * @param OpVT Op's machine value value type (doesn't need to be passed, but
 968  * makes life easier.)
 969  * @return The SDNode with the entire instruction sequence
 970  */
 971 SDNode *
 972 SPUDAGToDAGISel::SelectSHLi64(SDValue &Op, EVT OpVT) {
 973   SDValue Op0 = Op.getOperand(0);
 974   EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
 975                                OpVT, (128 / OpVT.getSizeInBits()));
 976   SDValue ShiftAmt = Op.getOperand(1);
 977   EVT ShiftAmtVT = ShiftAmt.getValueType();
 978   SDNode *VecOp0, *SelMask, *ZeroFill, *Shift = 0;
 979   SDValue SelMaskVal;
 980   DebugLoc dl = Op.getDebugLoc();
 981
 982   VecOp0 = CurDAG->getTargetNode(SPU::ORv2i64_i64, dl, VecVT, Op0);
 983   SelMaskVal = CurDAG->getTargetConstant(0xff00ULL, MVT::i16);
 984   SelMask = CurDAG->getTargetNode(SPU::FSMBIv2i64, dl, VecVT, SelMaskVal);
 985   ZeroFill = CurDAG->getTargetNode(SPU::ILv2i64, dl, VecVT,
 986                                    CurDAG->getTargetConstant(0, OpVT));
 987   VecOp0 = CurDAG->getTargetNode(SPU::SELBv2i64, dl, VecVT,
 988                                  SDValue(ZeroFill, 0),
 989                                  SDValue(VecOp0, 0),
 990                                  SDValue(SelMask, 0));
 991
 992   if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
 993     unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
 994     unsigned bits = unsigned(CN->getZExtValue()) & 7;
 995
 996     if (bytes > 0) {
 997       Shift =
 998         CurDAG->getTargetNode(SPU::SHLQBYIv2i64, dl, VecVT,
 999                               SDValue(VecOp0, 0),
1000                               CurDAG->getTargetConstant(bytes, ShiftAmtVT));
1001     }
1002
1003     if (bits > 0) {
1004       Shift =
1005         CurDAG->getTargetNode(SPU::SHLQBIIv2i64, dl, VecVT,
1006                               SDValue((Shift != 0 ? Shift : VecOp0), 0),
1007                               CurDAG->getTargetConstant(bits, ShiftAmtVT));
1008     }
1009   } else {
1010     SDNode *Bytes =
1011       CurDAG->getTargetNode(SPU::ROTMIr32, dl, ShiftAmtVT,
1012                             ShiftAmt,
1013                             CurDAG->getTargetConstant(3, ShiftAmtVT));
1014     SDNode *Bits =
1015       CurDAG->getTargetNode(SPU::ANDIr32, dl, ShiftAmtVT,
1016                             ShiftAmt,
1017                             CurDAG->getTargetConstant(7, ShiftAmtVT));
1018     Shift =
1019       CurDAG->getTargetNode(SPU::SHLQBYv2i64, dl, VecVT,
1020                             SDValue(VecOp0, 0), SDValue(Bytes, 0));
1021     Shift =
1022       CurDAG->getTargetNode(SPU::SHLQBIv2i64, dl, VecVT,
1023                             SDValue(Shift, 0), SDValue(Bits, 0));
1024   }
1025
1026   return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT, SDValue(Shift, 0));
1027 }
1028
1029 /*!
1030  * Emit the instruction sequence for i64 logical right shifts.
1031  *
1032  * @param Op The shl operand
1033  * @param OpVT Op's machine value value type (doesn't need to be passed, but
1034  * makes life easier.)
1035  * @return The SDNode with the entire instruction sequence
1036  */
1037 SDNode *
1038 SPUDAGToDAGISel::SelectSRLi64(SDValue &Op, EVT OpVT) {
1039   SDValue Op0 = Op.getOperand(0);
1040   EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
1041                                OpVT, (128 / OpVT.getSizeInBits()));
1042   SDValue ShiftAmt = Op.getOperand(1);
1043   EVT ShiftAmtVT = ShiftAmt.getValueType();
1044   SDNode *VecOp0, *Shift = 0;
1045   DebugLoc dl = Op.getDebugLoc();
1046
1047   VecOp0 = CurDAG->getTargetNode(SPU::ORv2i64_i64, dl, VecVT, Op0);
1048
1049   if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
1050     unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
1051     unsigned bits = unsigned(CN->getZExtValue()) & 7;
1052
1053     if (bytes > 0) {
1054       Shift =
1055         CurDAG->getTargetNode(SPU::ROTQMBYIv2i64, dl, VecVT,
1056                               SDValue(VecOp0, 0),
1057                               CurDAG->getTargetConstant(bytes, ShiftAmtVT));
1058     }
1059
1060     if (bits > 0) {
1061       Shift =
1062         CurDAG->getTargetNode(SPU::ROTQMBIIv2i64, dl, VecVT,
1063                               SDValue((Shift != 0 ? Shift : VecOp0), 0),
1064                               CurDAG->getTargetConstant(bits, ShiftAmtVT));
1065     }
1066   } else {
1067     SDNode *Bytes =
1068       CurDAG->getTargetNode(SPU::ROTMIr32, dl, ShiftAmtVT,
1069                             ShiftAmt,
1070                             CurDAG->getTargetConstant(3, ShiftAmtVT));
1071     SDNode *Bits =
1072       CurDAG->getTargetNode(SPU::ANDIr32, dl, ShiftAmtVT,
1073                             ShiftAmt,
1074                             CurDAG->getTargetConstant(7, ShiftAmtVT));
1075
1076     // Ensure that the shift amounts are negated!
1077     Bytes = CurDAG->getTargetNode(SPU::SFIr32, dl, ShiftAmtVT,
1078                                   SDValue(Bytes, 0),
1079                                   CurDAG->getTargetConstant(0, ShiftAmtVT));
1080
1081     Bits = CurDAG->getTargetNode(SPU::SFIr32, dl, ShiftAmtVT,
1082                                  SDValue(Bits, 0),
1083                                  CurDAG->getTargetConstant(0, ShiftAmtVT));
1084
1085     Shift =
1086       CurDAG->getTargetNode(SPU::ROTQMBYv2i64, dl, VecVT,
1087                             SDValue(VecOp0, 0), SDValue(Bytes, 0));
1088     Shift =
1089       CurDAG->getTargetNode(SPU::ROTQMBIv2i64, dl, VecVT,
1090                             SDValue(Shift, 0), SDValue(Bits, 0));
1091   }
1092
1093   return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT, SDValue(Shift, 0));
1094 }
1095
1096 /*!
1097  * Emit the instruction sequence for i64 arithmetic right shifts.
1098  *
1099  * @param Op The shl operand
1100  * @param OpVT Op's machine value value type (doesn't need to be passed, but
1101  * makes life easier.)
1102  * @return The SDNode with the entire instruction sequence
1103  */
1104 SDNode *
1105 SPUDAGToDAGISel::SelectSRAi64(SDValue &Op, EVT OpVT) {
1106   // Promote Op0 to vector
1107   EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
1108                                OpVT, (128 / OpVT.getSizeInBits()));
1109   SDValue ShiftAmt = Op.getOperand(1);
1110   EVT ShiftAmtVT = ShiftAmt.getValueType();
1111   DebugLoc dl = Op.getDebugLoc();
1112
1113   SDNode *VecOp0 =
1114     CurDAG->getTargetNode(SPU::ORv2i64_i64, dl, VecVT, Op.getOperand(0));
1115
1116   SDValue SignRotAmt = CurDAG->getTargetConstant(31, ShiftAmtVT);
1117   SDNode *SignRot =
1118     CurDAG->getTargetNode(SPU::ROTMAIv2i64_i32, dl, MVT::v2i64,
1119                           SDValue(VecOp0, 0), SignRotAmt);
1120   SDNode *UpperHalfSign =
1121     CurDAG->getTargetNode(SPU::ORi32_v4i32, dl, MVT::i32, SDValue(SignRot, 0));
1122
1123   SDNode *UpperHalfSignMask =
1124     CurDAG->getTargetNode(SPU::FSM64r32, dl, VecVT, SDValue(UpperHalfSign, 0));
1125   SDNode *UpperLowerMask =
1126     CurDAG->getTargetNode(SPU::FSMBIv2i64, dl, VecVT,
1127                           CurDAG->getTargetConstant(0xff00ULL, MVT::i16));
1128   SDNode *UpperLowerSelect =
1129     CurDAG->getTargetNode(SPU::SELBv2i64, dl, VecVT,
1130                           SDValue(UpperHalfSignMask, 0),
1131                           SDValue(VecOp0, 0),
1132                           SDValue(UpperLowerMask, 0));
1133
1134   SDNode *Shift = 0;
1135
1136   if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
1137     unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
1138     unsigned bits = unsigned(CN->getZExtValue()) & 7;
1139
1140     if (bytes > 0) {
1141       bytes = 31 - bytes;
1142       Shift =
1143         CurDAG->getTargetNode(SPU::ROTQBYIv2i64, dl, VecVT,
1144                               SDValue(UpperLowerSelect, 0),
1145                               CurDAG->getTargetConstant(bytes, ShiftAmtVT));
1146     }
1147
1148     if (bits > 0) {
1149       bits = 8 - bits;
1150       Shift =
1151         CurDAG->getTargetNode(SPU::ROTQBIIv2i64, dl, VecVT,
1152                               SDValue((Shift != 0 ? Shift : UpperLowerSelect), 0),
1153                               CurDAG->getTargetConstant(bits, ShiftAmtVT));
1154     }
1155   } else {
1156     SDNode *NegShift =
1157       CurDAG->getTargetNode(SPU::SFIr32, dl, ShiftAmtVT,
1158                             ShiftAmt, CurDAG->getTargetConstant(0, ShiftAmtVT));
1159
1160     Shift =
1161       CurDAG->getTargetNode(SPU::ROTQBYBIv2i64_r32, dl, VecVT,
1162                             SDValue(UpperLowerSelect, 0), SDValue(NegShift, 0));
1163     Shift =
1164       CurDAG->getTargetNode(SPU::ROTQBIv2i64, dl, VecVT,
1165                             SDValue(Shift, 0), SDValue(NegShift, 0));
1166   }
1167
1168   return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT, SDValue(Shift, 0));
1169 }
1170
1171 /*!
1172  Do the necessary magic necessary to load a i64 constant
1173  */
1174 SDNode *SPUDAGToDAGISel::SelectI64Constant(SDValue& Op, EVT OpVT,
1175                                            DebugLoc dl) {
1176   ConstantSDNode *CN = cast<ConstantSDNode>(Op.getNode());
1177   return SelectI64Constant(CN->getZExtValue(), OpVT, dl);
1178 }
1179
1180 SDNode *SPUDAGToDAGISel::SelectI64Constant(uint64_t Value64, EVT OpVT,
1181                                            DebugLoc dl) {
1182   EVT OpVecVT = EVT::getVectorVT(*CurDAG->getContext(), OpVT, 2);
1183   SDValue i64vec =
1184           SPU::LowerV2I64Splat(OpVecVT, *CurDAG, Value64, dl);
1185
1186   // Here's where it gets interesting, because we have to parse out the
1187   // subtree handed back in i64vec:
1188
1189   if (i64vec.getOpcode() == ISD::BIT_CONVERT) {
1190     // The degenerate case where the upper and lower bits in the splat are
1191     // identical:
1192     SDValue Op0 = i64vec.getOperand(0);
1193
1194     ReplaceUses(i64vec, Op0);
1195     return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT,
1196                                  SDValue(emitBuildVector(Op0), 0));
1197   } else if (i64vec.getOpcode() == SPUISD::SHUFB) {
1198     SDValue lhs = i64vec.getOperand(0);
1199     SDValue rhs = i64vec.getOperand(1);
1200     SDValue shufmask = i64vec.getOperand(2);
1201
1202     if (lhs.getOpcode() == ISD::BIT_CONVERT) {
1203       ReplaceUses(lhs, lhs.getOperand(0));
1204       lhs = lhs.getOperand(0);
1205     }
1206
1207     SDNode *lhsNode = (lhs.getNode()->isMachineOpcode()
1208                        ? lhs.getNode()
1209                        : emitBuildVector(lhs));
1210
1211     if (rhs.getOpcode() == ISD::BIT_CONVERT) {
1212       ReplaceUses(rhs, rhs.getOperand(0));
1213       rhs = rhs.getOperand(0);
1214     }
1215
1216     SDNode *rhsNode = (rhs.getNode()->isMachineOpcode()
1217                        ? rhs.getNode()
1218                        : emitBuildVector(rhs));
1219
1220     if (shufmask.getOpcode() == ISD::BIT_CONVERT) {
1221       ReplaceUses(shufmask, shufmask.getOperand(0));
1222       shufmask = shufmask.getOperand(0);
1223     }
1224
1225     SDNode *shufMaskNode = (shufmask.getNode()->isMachineOpcode()
1226                             ? shufmask.getNode()
1227                             : emitBuildVector(shufmask));
1228
1229     SDNode *shufNode =
1230             Select(CurDAG->getNode(SPUISD::SHUFB, dl, OpVecVT,
1231                                    SDValue(lhsNode, 0), SDValue(rhsNode, 0),
1232                                    SDValue(shufMaskNode, 0)));
1233
1234     return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT,
1235                                  SDValue(shufNode, 0));
1236   } else if (i64vec.getOpcode() == ISD::BUILD_VECTOR) {
1237     return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT,
1238                                  SDValue(emitBuildVector(i64vec), 0));
1239   } else {
1240     llvm_report_error("SPUDAGToDAGISel::SelectI64Constant: Unhandled i64vec"
1241                       "condition");
1242   }
1243 }
1244
1245 /// createSPUISelDag - This pass converts a legalized DAG into a
1246 /// SPU-specific DAG, ready for instruction scheduling.
1247 ///
1248 FunctionPass *llvm::createSPUISelDag(SPUTargetMachine &TM) {
1249   return new SPUDAGToDAGISel(TM);
1250 }