1 //===-- AMDILISelDAGToDAG.cpp - A dag to dag inst selector for AMDIL ------===//
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 /// \file
11 /// \brief Defines an instruction selector for the AMDGPU target.
12 //
13 //===----------------------------------------------------------------------===//
14 #include "AMDGPUInstrInfo.h"
15 #include "AMDGPUISelLowering.h" // For AMDGPUISD
16 #include "AMDGPURegisterInfo.h"
17 #include "AMDGPUSubtarget.h"
18 #include "R600InstrInfo.h"
19 #include "SIDefines.h"
20 #include "SIISelLowering.h"
21 #include "SIMachineFunctionInfo.h"
22 #include "llvm/CodeGen/FunctionLoweringInfo.h"
23 #include "llvm/CodeGen/PseudoSourceValue.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
27 #include "llvm/CodeGen/SelectionDAGISel.h"
28 #include "llvm/IR/Function.h"
29 
30 using namespace llvm;
31 
32 //===----------------------------------------------------------------------===//
33 // Instruction Selector Implementation
34 //===----------------------------------------------------------------------===//
35 
36 namespace {
37 /// AMDGPU specific code to select AMDGPU machine instructions for
38 /// SelectionDAG operations.
39 class AMDGPUDAGToDAGISel : public SelectionDAGISel {
40   // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can
41   // make the right decision when generating code for different targets.
42   const AMDGPUSubtarget *Subtarget;
43 public:
44   AMDGPUDAGToDAGISel(TargetMachine &TM);
45   virtual ~AMDGPUDAGToDAGISel();
46   bool runOnMachineFunction(MachineFunction &MF) override;
47   SDNode *Select(SDNode *N) override;
48   const char *getPassName() const override;
49   void PostprocessISelDAG() override;
50 
51 private:
52   bool isInlineImmediate(SDNode *N) const;
53   bool FoldOperand(SDValue &Src, SDValue &Sel, SDValue &Neg, SDValue &Abs,
54                    const R600InstrInfo *TII);
55   bool FoldOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &);
56   bool FoldDotOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &);
57 
58   // Complex pattern selectors
59   bool SelectADDRParam(SDValue Addr, SDValue& R1, SDValue& R2);
60   bool SelectADDR(SDValue N, SDValue &R1, SDValue &R2);
61   bool SelectADDR64(SDValue N, SDValue &R1, SDValue &R2);
62 
63   static bool checkType(const Value *ptr, unsigned int addrspace);
64   static bool checkPrivateAddress(const MachineMemOperand *Op);
65 
66   static bool isGlobalStore(const StoreSDNode *N);
67   static bool isFlatStore(const StoreSDNode *N);
68   static bool isPrivateStore(const StoreSDNode *N);
69   static bool isLocalStore(const StoreSDNode *N);
70   static bool isRegionStore(const StoreSDNode *N);
71 
72   bool isCPLoad(const LoadSDNode *N) const;
73   bool isConstantLoad(const LoadSDNode *N, int cbID) const;
74   bool isGlobalLoad(const LoadSDNode *N) const;
75   bool isFlatLoad(const LoadSDNode *N) const;
76   bool isParamLoad(const LoadSDNode *N) const;
77   bool isPrivateLoad(const LoadSDNode *N) const;
78   bool isLocalLoad(const LoadSDNode *N) const;
79   bool isRegionLoad(const LoadSDNode *N) const;
80 
81   SDNode *glueCopyToM0(SDNode *N) const;
82 
83   const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const;
84   bool SelectGlobalValueConstantOffset(SDValue Addr, SDValue& IntPtr);
85   bool SelectGlobalValueVariableOffset(SDValue Addr, SDValue &BaseReg,
86                                        SDValue& Offset);
87   bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
88   bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset);
89   bool isDSOffsetLegal(const SDValue &Base, unsigned Offset,
90                        unsigned OffsetBits) const;
91   bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const;
92   bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
93                                  SDValue &Offset1) const;
94   void SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
95                    SDValue &SOffset, SDValue &Offset, SDValue &Offen,
96                    SDValue &Idxen, SDValue &Addr64, SDValue &GLC, SDValue &SLC,
97                    SDValue &TFE) const;
98   bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
99                          SDValue &SOffset, SDValue &Offset, SDValue &GLC,
100                          SDValue &SLC, SDValue &TFE) const;
101   bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
102                          SDValue &VAddr, SDValue &SOffset, SDValue &Offset,
103                          SDValue &SLC) const;
104   bool SelectMUBUFScratch(SDValue Addr, SDValue &RSrc, SDValue &VAddr,
105                           SDValue &SOffset, SDValue &ImmOffset) const;
106   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
107                          SDValue &Offset, SDValue &GLC, SDValue &SLC,
108                          SDValue &TFE) const;
109   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
110                          SDValue &Offset, SDValue &GLC) const;
111   SDNode *SelectAddrSpaceCast(SDNode *N);
112   bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
113   bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
114                        SDValue &Clamp, SDValue &Omod) const;
115 
116   bool SelectVOP3Mods0Clamp(SDValue In, SDValue &Src, SDValue &SrcMods,
117                             SDValue &Omod) const;
118   bool SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src, SDValue &SrcMods,
119                                  SDValue &Clamp,
120                                  SDValue &Omod) const;
121 
122   SDNode *SelectADD_SUB_I64(SDNode *N);
123   SDNode *SelectDIV_SCALE(SDNode *N);
124 
125   SDNode *getS_BFE(unsigned Opcode, SDLoc DL, SDValue Val,
126                    uint32_t Offset, uint32_t Width);
127   SDNode *SelectS_BFEFromShifts(SDNode *N);
128   SDNode *SelectS_BFE(SDNode *N);
129 
130   // Include the pieces autogenerated from the target description.
131 #include "AMDGPUGenDAGISel.inc"
132 };
133 }  // end anonymous namespace
134 
135 /// \brief This pass converts a legalized DAG into a AMDGPU-specific
136 // DAG, ready for instruction scheduling.
137 FunctionPass *llvm::createAMDGPUISelDag(TargetMachine &TM) {
138   return new AMDGPUDAGToDAGISel(TM);
139 }
140 
141 AMDGPUDAGToDAGISel::AMDGPUDAGToDAGISel(TargetMachine &TM)
142     : SelectionDAGISel(TM) {}
143 
144 bool AMDGPUDAGToDAGISel::runOnMachineFunction(MachineFunction &MF) {
145   Subtarget = &static_cast<const AMDGPUSubtarget &>(MF.getSubtarget());
146   return SelectionDAGISel::runOnMachineFunction(MF);
147 }
148 
149 AMDGPUDAGToDAGISel::~AMDGPUDAGToDAGISel() {
150 }
151 
152 bool AMDGPUDAGToDAGISel::isInlineImmediate(SDNode *N) const {
153   const SITargetLowering *TL
154       = static_cast<const SITargetLowering *>(getTargetLowering());
155   return TL->analyzeImmediate(N) == 0;
156 }
157 
158 /// \brief Determine the register class for \p OpNo
159 /// \returns The register class of the virtual register that will be used for
160 /// the given operand number \OpNo or NULL if the register class cannot be
161 /// determined.
162 const TargetRegisterClass *AMDGPUDAGToDAGISel::getOperandRegClass(SDNode *N,
163                                                           unsigned OpNo) const {
164   if (!N->isMachineOpcode())
165     return nullptr;
166 
167   switch (N->getMachineOpcode()) {
168   default: {
169     const MCInstrDesc &Desc =
170         Subtarget->getInstrInfo()->get(N->getMachineOpcode());
171     unsigned OpIdx = Desc.getNumDefs() + OpNo;
172     if (OpIdx >= Desc.getNumOperands())
173       return nullptr;
174     int RegClass = Desc.OpInfo[OpIdx].RegClass;
175     if (RegClass == -1)
176       return nullptr;
177 
178     return Subtarget->getRegisterInfo()->getRegClass(RegClass);
179   }
180   case AMDGPU::REG_SEQUENCE: {
181     unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
182     const TargetRegisterClass *SuperRC =
183         Subtarget->getRegisterInfo()->getRegClass(RCID);
184 
185     SDValue SubRegOp = N->getOperand(OpNo + 1);
186     unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue();
187     return Subtarget->getRegisterInfo()->getSubClassWithSubReg(SuperRC,
188                                                               SubRegIdx);
189   }
190   }
191 }
192 
193 bool AMDGPUDAGToDAGISel::SelectADDRParam(
194   SDValue Addr, SDValue& R1, SDValue& R2) {
195 
196   if (Addr.getOpcode() == ISD::FrameIndex) {
197     if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
198       R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i32);
199       R2 = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i32);
200     } else {
201       R1 = Addr;
202       R2 = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i32);
203     }
204   } else if (Addr.getOpcode() == ISD::ADD) {
205     R1 = Addr.getOperand(0);
206     R2 = Addr.getOperand(1);
207   } else {
208     R1 = Addr;
209     R2 = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i32);
210   }
211   return true;
212 }
213 
214 bool AMDGPUDAGToDAGISel::SelectADDR(SDValue Addr, SDValue& R1, SDValue& R2) {
215   if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
216       Addr.getOpcode() == ISD::TargetGlobalAddress) {
217     return false;
218   }
219   return SelectADDRParam(Addr, R1, R2);
220 }
221 
222 
223 bool AMDGPUDAGToDAGISel::SelectADDR64(SDValue Addr, SDValue& R1, SDValue& R2) {
224   if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
225       Addr.getOpcode() == ISD::TargetGlobalAddress) {
226     return false;
227   }
228 
229   if (Addr.getOpcode() == ISD::FrameIndex) {
230     if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
231       R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i64);
232       R2 = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i64);
233     } else {
234       R1 = Addr;
235       R2 = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i64);
236     }
237   } else if (Addr.getOpcode() == ISD::ADD) {
238     R1 = Addr.getOperand(0);
239     R2 = Addr.getOperand(1);
240   } else {
241     R1 = Addr;
242     R2 = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i64);
243   }
244   return true;
245 }
246 
247 SDNode *AMDGPUDAGToDAGISel::glueCopyToM0(SDNode *N) const {
248   if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
249       !checkType(cast<MemSDNode>(N)->getMemOperand()->getValue(),
250                  AMDGPUAS::LOCAL_ADDRESS))
251     return N;
252 
253   const SITargetLowering& Lowering =
254       *static_cast<const SITargetLowering*>(getTargetLowering());
255 
256   // Write max value to m0 before each load operation
257 
258   SDValue M0 = Lowering.copyToM0(*CurDAG, CurDAG->getEntryNode(), SDLoc(N),
259                                  CurDAG->getTargetConstant(-1, SDLoc(N), MVT::i32));
260 
261   SDValue Glue = M0.getValue(1);
262 
263   SmallVector <SDValue, 8> Ops;
264   for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
265      Ops.push_back(N->getOperand(i));
266   }
267   Ops.push_back(Glue);
268   CurDAG->MorphNodeTo(N, N->getOpcode(), N->getVTList(), Ops);
269 
270   return N;
271 }
272 
273 SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
274   unsigned int Opc = N->getOpcode();
275   if (N->isMachineOpcode()) {
276     N->setNodeId(-1);
277     return nullptr;   // Already selected.
278   }
279 
280   if (isa<AtomicSDNode>(N))
281     N = glueCopyToM0(N);
282 
283   switch (Opc) {
284   default: break;
285   // We are selecting i64 ADD here instead of custom lower it during
286   // DAG legalization, so we can fold some i64 ADDs used for address
287   // calculation into the LOAD and STORE instructions.
288   case ISD::ADD:
289   case ISD::SUB: {
290     if (N->getValueType(0) != MVT::i64 ||
291         Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
292       break;
293 
294     return SelectADD_SUB_I64(N);
295   }
296   case ISD::SCALAR_TO_VECTOR:
297   case AMDGPUISD::BUILD_VERTICAL_VECTOR:
298   case ISD::BUILD_VECTOR: {
299     unsigned RegClassID;
300     const AMDGPURegisterInfo *TRI = Subtarget->getRegisterInfo();
301     EVT VT = N->getValueType(0);
302     unsigned NumVectorElts = VT.getVectorNumElements();
303     EVT EltVT = VT.getVectorElementType();
304     assert(EltVT.bitsEq(MVT::i32));
305     if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
306       bool UseVReg = true;
307       for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
308                                                     U != E; ++U) {
309         if (!U->isMachineOpcode()) {
310           continue;
311         }
312         const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
313         if (!RC) {
314           continue;
315         }
316         if (static_cast<const SIRegisterInfo *>(TRI)->isSGPRClass(RC)) {
317           UseVReg = false;
318         }
319       }
320       switch(NumVectorElts) {
321       case 1: RegClassID = UseVReg ? AMDGPU::VGPR_32RegClassID :
322                                      AMDGPU::SReg_32RegClassID;
323         break;
324       case 2: RegClassID = UseVReg ? AMDGPU::VReg_64RegClassID :
325                                      AMDGPU::SReg_64RegClassID;
326         break;
327       case 4: RegClassID = UseVReg ? AMDGPU::VReg_128RegClassID :
328                                      AMDGPU::SReg_128RegClassID;
329         break;
330       case 8: RegClassID = UseVReg ? AMDGPU::VReg_256RegClassID :
331                                      AMDGPU::SReg_256RegClassID;
332         break;
333       case 16: RegClassID = UseVReg ? AMDGPU::VReg_512RegClassID :
334                                       AMDGPU::SReg_512RegClassID;
335         break;
336       default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
337       }
338     } else {
339       // BUILD_VECTOR was lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
340       // that adds a 128 bits reg copy when going through TwoAddressInstructions
341       // pass. We want to avoid 128 bits copies as much as possible because they
342       // can't be bundled by our scheduler.
343       switch(NumVectorElts) {
344       case 2: RegClassID = AMDGPU::R600_Reg64RegClassID; break;
345       case 4:
346         if (Opc == AMDGPUISD::BUILD_VERTICAL_VECTOR)
347           RegClassID = AMDGPU::R600_Reg128VerticalRegClassID;
348         else
349           RegClassID = AMDGPU::R600_Reg128RegClassID;
350         break;
351       default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
352       }
353     }
354 
355     SDLoc DL(N);
356     SDValue RegClass = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32);
357 
358     if (NumVectorElts == 1) {
359       return CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS, EltVT,
360                                   N->getOperand(0), RegClass);
361     }
362 
363     assert(NumVectorElts <= 16 && "Vectors with more than 16 elements not "
364                                   "supported yet");
365     // 16 = Max Num Vector Elements
366     // 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
367     // 1 = Vector Register Class
368     SmallVector<SDValue, 16 * 2 + 1> RegSeqArgs(NumVectorElts * 2 + 1);
369 
370     RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32);
371     bool IsRegSeq = true;
372     unsigned NOps = N->getNumOperands();
373     for (unsigned i = 0; i < NOps; i++) {
374       // XXX: Why is this here?
375       if (isa<RegisterSDNode>(N->getOperand(i))) {
376         IsRegSeq = false;
377         break;
378       }
379       RegSeqArgs[1 + (2 * i)] = N->getOperand(i);
380       RegSeqArgs[1 + (2 * i) + 1] =
381               CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), DL,
382                                         MVT::i32);
383     }
384 
385     if (NOps != NumVectorElts) {
386       // Fill in the missing undef elements if this was a scalar_to_vector.
387       assert(Opc == ISD::SCALAR_TO_VECTOR && NOps < NumVectorElts);
388 
389       MachineSDNode *ImpDef = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
390                                                      DL, EltVT);
391       for (unsigned i = NOps; i < NumVectorElts; ++i) {
392         RegSeqArgs[1 + (2 * i)] = SDValue(ImpDef, 0);
393         RegSeqArgs[1 + (2 * i) + 1] =
394           CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), DL, MVT::i32);
395       }
396     }
397 
398     if (!IsRegSeq)
399       break;
400     return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(),
401                                 RegSeqArgs);
402   }
403   case ISD::BUILD_PAIR: {
404     SDValue RC, SubReg0, SubReg1;
405     if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
406       break;
407     }
408     SDLoc DL(N);
409     if (N->getValueType(0) == MVT::i128) {
410       RC = CurDAG->getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32);
411       SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32);
412       SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32);
413     } else if (N->getValueType(0) == MVT::i64) {
414       RC = CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32);
415       SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
416       SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
417     } else {
418       llvm_unreachable("Unhandled value type for BUILD_PAIR");
419     }
420     const SDValue Ops[] = { RC, N->getOperand(0), SubReg0,
421                             N->getOperand(1), SubReg1 };
422     return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
423                                   DL, N->getValueType(0), Ops);
424   }
425 
426   case ISD::Constant:
427   case ISD::ConstantFP: {
428     if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
429         N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N))
430       break;
431 
432     uint64_t Imm;
433     if (ConstantFPSDNode *FP = dyn_cast<ConstantFPSDNode>(N))
434       Imm = FP->getValueAPF().bitcastToAPInt().getZExtValue();
435     else {
436       ConstantSDNode *C = cast<ConstantSDNode>(N);
437       Imm = C->getZExtValue();
438     }
439 
440     SDLoc DL(N);
441     SDNode *Lo = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32,
442                                 CurDAG->getConstant(Imm & 0xFFFFFFFF, DL,
443                                                     MVT::i32));
444     SDNode *Hi = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32,
445                                 CurDAG->getConstant(Imm >> 32, DL, MVT::i32));
446     const SDValue Ops[] = {
447       CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
448       SDValue(Lo, 0), CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
449       SDValue(Hi, 0), CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32)
450     };
451 
452     return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL,
453                                   N->getValueType(0), Ops);
454   }
455 
456   case ISD::LOAD: {
457     LoadSDNode *LD = cast<LoadSDNode>(N);
458     SDLoc SL(N);
459     EVT VT = N->getValueType(0);
460 
461     if (VT != MVT::i64 || LD->getExtensionType() != ISD::NON_EXTLOAD) {
462       N = glueCopyToM0(N);
463       break;
464     }
465 
466     // To simplify the TableGen patters, we replace all i64 loads with
467     // v2i32 loads.  Alternatively, we could promote i64 loads to v2i32
468     // during DAG legalization, however, so places (ExpandUnalignedLoad)
469     // in the DAG legalizer assume that if i64 is legal, so doing this
470     // promotion early can cause problems.
471 
472     SDValue NewLoad = CurDAG->getLoad(MVT::v2i32, SDLoc(N), LD->getChain(),
473                                       LD->getBasePtr(), LD->getMemOperand());
474     SDValue BitCast = CurDAG->getNode(ISD::BITCAST, SL,
475                                       MVT::i64, NewLoad);
476     CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLoad.getValue(1));
477     CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), BitCast);
478     SDNode *Load = glueCopyToM0(NewLoad.getNode());
479     SelectCode(Load);
480     N = BitCast.getNode();
481     break;
482   }
483 
484   case ISD::STORE: {
485     // Handle i64 stores here for the same reason mentioned above for loads.
486     StoreSDNode *ST = cast<StoreSDNode>(N);
487     SDValue Value = ST->getValue();
488     if (Value.getValueType() == MVT::i64 && !ST->isTruncatingStore()) {
489 
490       SDValue NewValue = CurDAG->getNode(ISD::BITCAST, SDLoc(N),
491                                         MVT::v2i32, Value);
492       SDValue NewStore = CurDAG->getStore(ST->getChain(), SDLoc(N), NewValue,
493                                           ST->getBasePtr(), ST->getMemOperand());
494 
495       CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), NewStore);
496 
497       if (NewValue.getOpcode() == ISD::BITCAST) {
498         Select(NewStore.getNode());
499         return SelectCode(NewValue.getNode());
500       }
501 
502       // getNode() may fold the bitcast if its input was another bitcast.  If that
503       // happens we should only select the new store.
504       N = NewStore.getNode();
505     }
506 
507     N = glueCopyToM0(N);
508     break;
509   }
510 
511   case AMDGPUISD::REGISTER_LOAD: {
512     if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
513       break;
514     SDValue Addr, Offset;
515 
516     SDLoc DL(N);
517     SelectADDRIndirect(N->getOperand(1), Addr, Offset);
518     const SDValue Ops[] = {
519       Addr,
520       Offset,
521       CurDAG->getTargetConstant(0, DL, MVT::i32),
522       N->getOperand(0),
523     };
524     return CurDAG->getMachineNode(AMDGPU::SI_RegisterLoad, DL,
525                                   CurDAG->getVTList(MVT::i32, MVT::i64,
526                                                     MVT::Other),
527                                   Ops);
528   }
529   case AMDGPUISD::REGISTER_STORE: {
530     if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
531       break;
532     SDValue Addr, Offset;
533     SelectADDRIndirect(N->getOperand(2), Addr, Offset);
534     SDLoc DL(N);
535     const SDValue Ops[] = {
536       N->getOperand(1),
537       Addr,
538       Offset,
539       CurDAG->getTargetConstant(0, DL, MVT::i32),
540       N->getOperand(0),
541     };
542     return CurDAG->getMachineNode(AMDGPU::SI_RegisterStorePseudo, DL,
543                                         CurDAG->getVTList(MVT::Other),
544                                         Ops);
545   }
546 
547   case AMDGPUISD::BFE_I32:
548   case AMDGPUISD::BFE_U32: {
549     if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
550       break;
551 
552     // There is a scalar version available, but unlike the vector version which
553     // has a separate operand for the offset and width, the scalar version packs
554     // the width and offset into a single operand. Try to move to the scalar
555     // version if the offsets are constant, so that we can try to keep extended
556     // loads of kernel arguments in SGPRs.
557 
558     // TODO: Technically we could try to pattern match scalar bitshifts of
559     // dynamic values, but it's probably not useful.
560     ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
561     if (!Offset)
562       break;
563 
564     ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
565     if (!Width)
566       break;
567 
568     bool Signed = Opc == AMDGPUISD::BFE_I32;
569 
570     uint32_t OffsetVal = Offset->getZExtValue();
571     uint32_t WidthVal = Width->getZExtValue();
572 
573     return getS_BFE(Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32, SDLoc(N),
574                     N->getOperand(0), OffsetVal, WidthVal);
575 
576   }
577   case AMDGPUISD::DIV_SCALE: {
578     return SelectDIV_SCALE(N);
579   }
580   case ISD::CopyToReg: {
581     const SITargetLowering& Lowering =
582       *static_cast<const SITargetLowering*>(getTargetLowering());
583     Lowering.legalizeTargetIndependentNode(N, *CurDAG);
584     break;
585   }
586   case ISD::ADDRSPACECAST:
587     return SelectAddrSpaceCast(N);
588   case ISD::AND:
589   case ISD::SRL:
590   case ISD::SRA:
591     if (N->getValueType(0) != MVT::i32 ||
592         Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
593       break;
594 
595     return SelectS_BFE(N);
596   }
597 
598   return SelectCode(N);
599 }
600 
601 
602 bool AMDGPUDAGToDAGISel::checkType(const Value *Ptr, unsigned AS) {
603   assert(AS != 0 && "Use checkPrivateAddress instead.");
604   if (!Ptr)
605     return false;
606 
607   return Ptr->getType()->getPointerAddressSpace() == AS;
608 }
609 
610 bool AMDGPUDAGToDAGISel::checkPrivateAddress(const MachineMemOperand *Op) {
611   if (Op->getPseudoValue())
612     return true;
613 
614   if (PointerType *PT = dyn_cast<PointerType>(Op->getValue()->getType()))
615     return PT->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS;
616 
617   return false;
618 }
619 
620 bool AMDGPUDAGToDAGISel::isGlobalStore(const StoreSDNode *N) {
621   return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
622 }
623 
624 bool AMDGPUDAGToDAGISel::isPrivateStore(const StoreSDNode *N) {
625   const Value *MemVal = N->getMemOperand()->getValue();
626   return (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
627           !checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
628           !checkType(MemVal, AMDGPUAS::REGION_ADDRESS));
629 }
630 
631 bool AMDGPUDAGToDAGISel::isLocalStore(const StoreSDNode *N) {
632   return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
633 }
634 
635 bool AMDGPUDAGToDAGISel::isFlatStore(const StoreSDNode *N) {
636   return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
637 }
638 
639 bool AMDGPUDAGToDAGISel::isRegionStore(const StoreSDNode *N) {
640   return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
641 }
642 
643 bool AMDGPUDAGToDAGISel::isConstantLoad(const LoadSDNode *N, int CbId) const {
644   const Value *MemVal = N->getMemOperand()->getValue();
645   if (CbId == -1)
646     return checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS);
647 
648   return checkType(MemVal, AMDGPUAS::CONSTANT_BUFFER_0 + CbId);
649 }
650 
651 bool AMDGPUDAGToDAGISel::isGlobalLoad(const LoadSDNode *N) const {
652   if (N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS)
653     if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
654         N->getMemoryVT().bitsLT(MVT::i32))
655       return true;
656 
657   return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
658 }
659 
660 bool AMDGPUDAGToDAGISel::isParamLoad(const LoadSDNode *N) const {
661   return checkType(N->getMemOperand()->getValue(), AMDGPUAS::PARAM_I_ADDRESS);
662 }
663 
664 bool AMDGPUDAGToDAGISel::isLocalLoad(const  LoadSDNode *N) const {
665   return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
666 }
667 
668 bool AMDGPUDAGToDAGISel::isFlatLoad(const  LoadSDNode *N) const {
669   return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
670 }
671 
672 bool AMDGPUDAGToDAGISel::isRegionLoad(const  LoadSDNode *N) const {
673   return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
674 }
675 
676 bool AMDGPUDAGToDAGISel::isCPLoad(const LoadSDNode *N) const {
677   MachineMemOperand *MMO = N->getMemOperand();
678   if (checkPrivateAddress(N->getMemOperand())) {
679     if (MMO) {
680       const PseudoSourceValue *PSV = MMO->getPseudoValue();
681       if (PSV && PSV == PseudoSourceValue::getConstantPool()) {
682         return true;
683       }
684     }
685   }
686   return false;
687 }
688 
689 bool AMDGPUDAGToDAGISel::isPrivateLoad(const LoadSDNode *N) const {
690   if (checkPrivateAddress(N->getMemOperand())) {
691     // Check to make sure we are not a constant pool load or a constant load
692     // that is marked as a private load
693     if (isCPLoad(N) || isConstantLoad(N, -1)) {
694       return false;
695     }
696   }
697 
698   const Value *MemVal = N->getMemOperand()->getValue();
699   if (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
700       !checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
701       !checkType(MemVal, AMDGPUAS::FLAT_ADDRESS) &&
702       !checkType(MemVal, AMDGPUAS::REGION_ADDRESS) &&
703       !checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS) &&
704       !checkType(MemVal, AMDGPUAS::PARAM_D_ADDRESS) &&
705       !checkType(MemVal, AMDGPUAS::PARAM_I_ADDRESS)) {
706     return true;
707   }
708   return false;
709 }
710 
711 const char *AMDGPUDAGToDAGISel::getPassName() const {
712   return "AMDGPU DAG->DAG Pattern Instruction Selection";
713 }
714 
715 #ifdef DEBUGTMP
716 #undef INT64_C
717 #endif
718 #undef DEBUGTMP
719 
720 //===----------------------------------------------------------------------===//
721 // Complex Patterns
722 //===----------------------------------------------------------------------===//
723 
724 bool AMDGPUDAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr,
725                                                          SDValue& IntPtr) {
726   if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) {
727     IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, SDLoc(Addr),
728                                        true);
729     return true;
730   }
731   return false;
732 }
733 
734 bool AMDGPUDAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr,
735     SDValue& BaseReg, SDValue &Offset) {
736   if (!isa<ConstantSDNode>(Addr)) {
737     BaseReg = Addr;
738     Offset = CurDAG->getIntPtrConstant(0, SDLoc(Addr), true);
739     return true;
740   }
741   return false;
742 }
743 
744 bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
745                                            SDValue &Offset) {
746   ConstantSDNode *IMMOffset;
747 
748   if (Addr.getOpcode() == ISD::ADD
749       && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))
750       && isInt<16>(IMMOffset->getZExtValue())) {
751 
752       Base = Addr.getOperand(0);
753       Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr),
754                                          MVT::i32);
755       return true;
756   // If the pointer address is constant, we can move it to the offset field.
757   } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr))
758              && isInt<16>(IMMOffset->getZExtValue())) {
759     Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
760                                   SDLoc(CurDAG->getEntryNode()),
761                                   AMDGPU::ZERO, MVT::i32);
762     Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr),
763                                        MVT::i32);
764     return true;
765   }
766 
767   // Default case, no offset
768   Base = Addr;
769   Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i32);
770   return true;
771 }
772 
773 bool AMDGPUDAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
774                                             SDValue &Offset) {
775   ConstantSDNode *C;
776   SDLoc DL(Addr);
777 
778   if ((C = dyn_cast<ConstantSDNode>(Addr))) {
779     Base = CurDAG->getRegister(AMDGPU::INDIRECT_BASE_ADDR, MVT::i32);
780     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
781   } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
782             (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) {
783     Base = Addr.getOperand(0);
784     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
785   } else {
786     Base = Addr;
787     Offset = CurDAG->getTargetConstant(0, DL, MVT::i32);
788   }
789 
790   return true;
791 }
792 
793 SDNode *AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) {
794   SDLoc DL(N);
795   SDValue LHS = N->getOperand(0);
796   SDValue RHS = N->getOperand(1);
797 
798   bool IsAdd = (N->getOpcode() == ISD::ADD);
799 
800   SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
801   SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
802 
803   SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
804                                        DL, MVT::i32, LHS, Sub0);
805   SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
806                                        DL, MVT::i32, LHS, Sub1);
807 
808   SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
809                                        DL, MVT::i32, RHS, Sub0);
810   SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
811                                        DL, MVT::i32, RHS, Sub1);
812 
813   SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
814   SDValue AddLoArgs[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) };
815 
816 
817   unsigned Opc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32;
818   unsigned CarryOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32;
819 
820   SDNode *AddLo = CurDAG->getMachineNode( Opc, DL, VTList, AddLoArgs);
821   SDValue Carry(AddLo, 1);
822   SDNode *AddHi
823     = CurDAG->getMachineNode(CarryOpc, DL, MVT::i32,
824                              SDValue(Hi0, 0), SDValue(Hi1, 0), Carry);
825 
826   SDValue Args[5] = {
827     CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
828     SDValue(AddLo,0),
829     Sub0,
830     SDValue(AddHi,0),
831     Sub1,
832   };
833   return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, MVT::i64, Args);
834 }
835 
836 // We need to handle this here because tablegen doesn't support matching
837 // instructions with multiple outputs.
838 SDNode *AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) {
839   SDLoc SL(N);
840   EVT VT = N->getValueType(0);
841 
842   assert(VT == MVT::f32 || VT == MVT::f64);
843 
844   unsigned Opc
845     = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32;
846 
847   // src0_modifiers, src0, src1_modifiers, src1, src2_modifiers, src2, clamp, omod
848   SDValue Ops[8];
849 
850   SelectVOP3Mods0(N->getOperand(0), Ops[1], Ops[0], Ops[6], Ops[7]);
851   SelectVOP3Mods(N->getOperand(1), Ops[3], Ops[2]);
852   SelectVOP3Mods(N->getOperand(2), Ops[5], Ops[4]);
853   return CurDAG->SelectNodeTo(N, Opc, VT, MVT::i1, Ops);
854 }
855 
856 bool AMDGPUDAGToDAGISel::isDSOffsetLegal(const SDValue &Base, unsigned Offset,
857                                          unsigned OffsetBits) const {
858   if ((OffsetBits == 16 && !isUInt<16>(Offset)) ||
859       (OffsetBits == 8 && !isUInt<8>(Offset)))
860     return false;
861 
862   if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS)
863     return true;
864 
865   // On Southern Islands instruction with a negative base value and an offset
866   // don't seem to work.
867   return CurDAG->SignBitIsZero(Base);
868 }
869 
870 bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base,
871                                               SDValue &Offset) const {
872   if (CurDAG->isBaseWithConstantOffset(Addr)) {
873     SDValue N0 = Addr.getOperand(0);
874     SDValue N1 = Addr.getOperand(1);
875     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
876     if (isDSOffsetLegal(N0, C1->getSExtValue(), 16)) {
877       // (add n0, c0)
878       Base = N0;
879       Offset = N1;
880       return true;
881     }
882   }
883 
884   SDLoc DL(Addr);
885 
886   // If we have a constant address, prefer to put the constant into the
887   // offset. This can save moves to load the constant address since multiple
888   // operations can share the zero base address register, and enables merging
889   // into read2 / write2 instructions.
890   if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
891     if (isUInt<16>(CAddr->getZExtValue())) {
892       SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
893       MachineSDNode *MovZero = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
894                                  DL, MVT::i32, Zero);
895       Base = SDValue(MovZero, 0);
896       Offset = Addr;
897       return true;
898     }
899   }
900 
901   // default case
902   Base = Addr;
903   Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
904   return true;
905 }
906 
907 bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base,
908                                                    SDValue &Offset0,
909                                                    SDValue &Offset1) const {
910   SDLoc DL(Addr);
911 
912   if (CurDAG->isBaseWithConstantOffset(Addr)) {
913     SDValue N0 = Addr.getOperand(0);
914     SDValue N1 = Addr.getOperand(1);
915     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
916     unsigned DWordOffset0 = C1->getZExtValue() / 4;
917     unsigned DWordOffset1 = DWordOffset0 + 1;
918     // (add n0, c0)
919     if (isDSOffsetLegal(N0, DWordOffset1, 8)) {
920       Base = N0;
921       Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
922       Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
923       return true;
924     }
925   }
926 
927   if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
928     unsigned DWordOffset0 = CAddr->getZExtValue() / 4;
929     unsigned DWordOffset1 = DWordOffset0 + 1;
930     assert(4 * DWordOffset0 == CAddr->getZExtValue());
931 
932     if (isUInt<8>(DWordOffset0) && isUInt<8>(DWordOffset1)) {
933       SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
934       MachineSDNode *MovZero
935         = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
936                                  DL, MVT::i32, Zero);
937       Base = SDValue(MovZero, 0);
938       Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
939       Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
940       return true;
941     }
942   }
943 
944   // default case
945   Base = Addr;
946   Offset0 = CurDAG->getTargetConstant(0, DL, MVT::i8);
947   Offset1 = CurDAG->getTargetConstant(1, DL, MVT::i8);
948   return true;
949 }
950 
951 static bool isLegalMUBUFImmOffset(const ConstantSDNode *Imm) {
952   return isUInt<12>(Imm->getZExtValue());
953 }
954 
955 void AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr,
956                                      SDValue &VAddr, SDValue &SOffset,
957                                      SDValue &Offset, SDValue &Offen,
958                                      SDValue &Idxen, SDValue &Addr64,
959                                      SDValue &GLC, SDValue &SLC,
960                                      SDValue &TFE) const {
961   SDLoc DL(Addr);
962 
963   GLC = CurDAG->getTargetConstant(0, DL, MVT::i1);
964   SLC = CurDAG->getTargetConstant(0, DL, MVT::i1);
965   TFE = CurDAG->getTargetConstant(0, DL, MVT::i1);
966 
967   Idxen = CurDAG->getTargetConstant(0, DL, MVT::i1);
968   Offen = CurDAG->getTargetConstant(0, DL, MVT::i1);
969   Addr64 = CurDAG->getTargetConstant(0, DL, MVT::i1);
970   SOffset = CurDAG->getTargetConstant(0, DL, MVT::i32);
971 
972   if (CurDAG->isBaseWithConstantOffset(Addr)) {
973     SDValue N0 = Addr.getOperand(0);
974     SDValue N1 = Addr.getOperand(1);
975     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
976 
977     if (N0.getOpcode() == ISD::ADD) {
978       // (add (add N2, N3), C1) -> addr64
979       SDValue N2 = N0.getOperand(0);
980       SDValue N3 = N0.getOperand(1);
981       Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1);
982       Ptr = N2;
983       VAddr = N3;
984     } else {
985 
986       // (add N0, C1) -> offset
987       VAddr = CurDAG->getTargetConstant(0, DL, MVT::i32);
988       Ptr = N0;
989     }
990 
991     if (isLegalMUBUFImmOffset(C1)) {
992         Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
993         return;
994     } else if (isUInt<32>(C1->getZExtValue())) {
995       // Illegal offset, store it in soffset.
996       Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
997       SOffset = SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32,
998                    CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32)),
999                         0);
1000       return;
1001     }
1002   }
1003 
1004   if (Addr.getOpcode() == ISD::ADD) {
1005     // (add N0, N1) -> addr64
1006     SDValue N0 = Addr.getOperand(0);
1007     SDValue N1 = Addr.getOperand(1);
1008     Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1);
1009     Ptr = N0;
1010     VAddr = N1;
1011     Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1012     return;
1013   }
1014 
1015   // default case -> offset
1016   VAddr = CurDAG->getTargetConstant(0, DL, MVT::i32);
1017   Ptr = Addr;
1018   Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1019 
1020 }
1021 
1022 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
1023                                            SDValue &VAddr, SDValue &SOffset,
1024                                            SDValue &Offset, SDValue &GLC,
1025                                            SDValue &SLC, SDValue &TFE) const {
1026   SDValue Ptr, Offen, Idxen, Addr64;
1027 
1028   SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
1029               GLC, SLC, TFE);
1030 
1031   ConstantSDNode *C = cast<ConstantSDNode>(Addr64);
1032   if (C->getSExtValue()) {
1033     SDLoc DL(Addr);
1034 
1035     const SITargetLowering& Lowering =
1036       *static_cast<const SITargetLowering*>(getTargetLowering());
1037 
1038     SRsrc = SDValue(Lowering.wrapAddr64Rsrc(*CurDAG, DL, Ptr), 0);
1039     return true;
1040   }
1041 
1042   return false;
1043 }
1044 
1045 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
1046                                            SDValue &VAddr, SDValue &SOffset,
1047 					   SDValue &Offset,
1048 					   SDValue &SLC) const {
1049   SLC = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i1);
1050   SDValue GLC, TFE;
1051 
1052   return SelectMUBUFAddr64(Addr, SRsrc, VAddr, SOffset, Offset, GLC, SLC, TFE);
1053 }
1054 
1055 bool AMDGPUDAGToDAGISel::SelectMUBUFScratch(SDValue Addr, SDValue &Rsrc,
1056                                             SDValue &VAddr, SDValue &SOffset,
1057                                             SDValue &ImmOffset) const {
1058 
1059   SDLoc DL(Addr);
1060   MachineFunction &MF = CurDAG->getMachineFunction();
1061   const SIRegisterInfo *TRI =
1062       static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
1063   MachineRegisterInfo &MRI = MF.getRegInfo();
1064   const SITargetLowering& Lowering =
1065     *static_cast<const SITargetLowering*>(getTargetLowering());
1066 
1067   unsigned ScratchOffsetReg =
1068       TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
1069   Lowering.CreateLiveInRegister(*CurDAG, &AMDGPU::SReg_32RegClass,
1070                                 ScratchOffsetReg, MVT::i32);
1071   SDValue Sym0 = CurDAG->getExternalSymbol("SCRATCH_RSRC_DWORD0", MVT::i32);
1072   SDValue ScratchRsrcDword0 =
1073       SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, Sym0), 0);
1074 
1075   SDValue Sym1 = CurDAG->getExternalSymbol("SCRATCH_RSRC_DWORD1", MVT::i32);
1076   SDValue ScratchRsrcDword1 =
1077       SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, Sym1), 0);
1078 
1079   const SDValue RsrcOps[] = {
1080       CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
1081       ScratchRsrcDword0,
1082       CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
1083       ScratchRsrcDword1,
1084       CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32),
1085   };
1086   SDValue ScratchPtr = SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
1087                                               MVT::v2i32, RsrcOps), 0);
1088   Rsrc = SDValue(Lowering.buildScratchRSRC(*CurDAG, DL, ScratchPtr), 0);
1089   SOffset = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
1090       MRI.getLiveInVirtReg(ScratchOffsetReg), MVT::i32);
1091 
1092   // (add n0, c1)
1093   if (CurDAG->isBaseWithConstantOffset(Addr)) {
1094     SDValue N1 = Addr.getOperand(1);
1095     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1096 
1097     if (isLegalMUBUFImmOffset(C1)) {
1098       VAddr = Addr.getOperand(0);
1099       ImmOffset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
1100       return true;
1101     }
1102   }
1103 
1104   // (node)
1105   VAddr = Addr;
1106   ImmOffset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1107   return true;
1108 }
1109 
1110 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1111                                            SDValue &SOffset, SDValue &Offset,
1112                                            SDValue &GLC, SDValue &SLC,
1113                                            SDValue &TFE) const {
1114   SDValue Ptr, VAddr, Offen, Idxen, Addr64;
1115   const SIInstrInfo *TII =
1116     static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
1117 
1118   SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
1119               GLC, SLC, TFE);
1120 
1121   if (!cast<ConstantSDNode>(Offen)->getSExtValue() &&
1122       !cast<ConstantSDNode>(Idxen)->getSExtValue() &&
1123       !cast<ConstantSDNode>(Addr64)->getSExtValue()) {
1124     uint64_t Rsrc = TII->getDefaultRsrcDataFormat() |
1125                     APInt::getAllOnesValue(32).getZExtValue(); // Size
1126     SDLoc DL(Addr);
1127 
1128     const SITargetLowering& Lowering =
1129       *static_cast<const SITargetLowering*>(getTargetLowering());
1130 
1131     SRsrc = SDValue(Lowering.buildRSRC(*CurDAG, DL, Ptr, 0, Rsrc), 0);
1132     return true;
1133   }
1134   return false;
1135 }
1136 
1137 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1138                                            SDValue &Soffset, SDValue &Offset,
1139                                            SDValue &GLC) const {
1140   SDValue SLC, TFE;
1141 
1142   return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE);
1143 }
1144 
1145 // FIXME: This is incorrect and only enough to be able to compile.
1146 SDNode *AMDGPUDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) {
1147   AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(N);
1148   SDLoc DL(N);
1149 
1150   assert(Subtarget->hasFlatAddressSpace() &&
1151          "addrspacecast only supported with flat address space!");
1152 
1153   assert((ASC->getSrcAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS &&
1154           ASC->getDestAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS) &&
1155          "Cannot cast address space to / from constant address!");
1156 
1157   assert((ASC->getSrcAddressSpace() == AMDGPUAS::FLAT_ADDRESS ||
1158           ASC->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS) &&
1159          "Can only cast to / from flat address space!");
1160 
1161   // The flat instructions read the address as the index of the VGPR holding the
1162   // address, so casting should just be reinterpreting the base VGPR, so just
1163   // insert trunc / bitcast / zext.
1164 
1165   SDValue Src = ASC->getOperand(0);
1166   EVT DestVT = ASC->getValueType(0);
1167   EVT SrcVT = Src.getValueType();
1168 
1169   unsigned SrcSize = SrcVT.getSizeInBits();
1170   unsigned DestSize = DestVT.getSizeInBits();
1171 
1172   if (SrcSize > DestSize) {
1173     assert(SrcSize == 64 && DestSize == 32);
1174     return CurDAG->getMachineNode(
1175       TargetOpcode::EXTRACT_SUBREG,
1176       DL,
1177       DestVT,
1178       Src,
1179       CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32));
1180   }
1181 
1182 
1183   if (DestSize > SrcSize) {
1184     assert(SrcSize == 32 && DestSize == 64);
1185 
1186     // FIXME: This is probably wrong, we should never be defining
1187     // a register class with both VGPRs and SGPRs
1188     SDValue RC = CurDAG->getTargetConstant(AMDGPU::VS_64RegClassID, DL,
1189                                            MVT::i32);
1190 
1191     const SDValue Ops[] = {
1192       RC,
1193       Src,
1194       CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
1195       SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32,
1196                                      CurDAG->getConstant(0, DL, MVT::i32)), 0),
1197       CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32)
1198     };
1199 
1200     return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
1201                                   DL, N->getValueType(0), Ops);
1202   }
1203 
1204   assert(SrcSize == 64 && DestSize == 64);
1205   return CurDAG->getNode(ISD::BITCAST, DL, DestVT, Src).getNode();
1206 }
1207 
1208 SDNode *AMDGPUDAGToDAGISel::getS_BFE(unsigned Opcode, SDLoc DL, SDValue Val,
1209                                      uint32_t Offset, uint32_t Width) {
1210   // Transformation function, pack the offset and width of a BFE into
1211   // the format expected by the S_BFE_I32 / S_BFE_U32. In the second
1212   // source, bits [5:0] contain the offset and bits [22:16] the width.
1213   uint32_t PackedVal = Offset | (Width << 16);
1214   SDValue PackedConst = CurDAG->getTargetConstant(PackedVal, DL, MVT::i32);
1215 
1216   return CurDAG->getMachineNode(Opcode, DL, MVT::i32, Val, PackedConst);
1217 }
1218 
1219 SDNode *AMDGPUDAGToDAGISel::SelectS_BFEFromShifts(SDNode *N) {
1220   // "(a << b) srl c)" ---> "BFE_U32 a, (c-b), (32-c)
1221   // "(a << b) sra c)" ---> "BFE_I32 a, (c-b), (32-c)
1222   // Predicate: 0 < b <= c < 32
1223 
1224   const SDValue &Shl = N->getOperand(0);
1225   ConstantSDNode *B = dyn_cast<ConstantSDNode>(Shl->getOperand(1));
1226   ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1227 
1228   if (B && C) {
1229     uint32_t BVal = B->getZExtValue();
1230     uint32_t CVal = C->getZExtValue();
1231 
1232     if (0 < BVal && BVal <= CVal && CVal < 32) {
1233       bool Signed = N->getOpcode() == ISD::SRA;
1234       unsigned Opcode = Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32;
1235 
1236       return getS_BFE(Opcode, SDLoc(N), Shl.getOperand(0),
1237                       CVal - BVal, 32 - CVal);
1238     }
1239   }
1240   return SelectCode(N);
1241 }
1242 
1243 SDNode *AMDGPUDAGToDAGISel::SelectS_BFE(SDNode *N) {
1244   switch (N->getOpcode()) {
1245   case ISD::AND:
1246     if (N->getOperand(0).getOpcode() == ISD::SRL) {
1247       // "(a srl b) & mask" ---> "BFE_U32 a, b, popcount(mask)"
1248       // Predicate: isMask(mask)
1249       const SDValue &Srl = N->getOperand(0);
1250       ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(Srl.getOperand(1));
1251       ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(N->getOperand(1));
1252 
1253       if (Shift && Mask) {
1254         uint32_t ShiftVal = Shift->getZExtValue();
1255         uint32_t MaskVal = Mask->getZExtValue();
1256 
1257         if (isMask_32(MaskVal)) {
1258           uint32_t WidthVal = countPopulation(MaskVal);
1259 
1260           return getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N), Srl.getOperand(0),
1261                           ShiftVal, WidthVal);
1262         }
1263       }
1264     }
1265     break;
1266   case ISD::SRL:
1267     if (N->getOperand(0).getOpcode() == ISD::AND) {
1268       // "(a & mask) srl b)" ---> "BFE_U32 a, b, popcount(mask >> b)"
1269       // Predicate: isMask(mask >> b)
1270       const SDValue &And = N->getOperand(0);
1271       ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(N->getOperand(1));
1272       ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(And->getOperand(1));
1273 
1274       if (Shift && Mask) {
1275         uint32_t ShiftVal = Shift->getZExtValue();
1276         uint32_t MaskVal = Mask->getZExtValue() >> ShiftVal;
1277 
1278         if (isMask_32(MaskVal)) {
1279           uint32_t WidthVal = countPopulation(MaskVal);
1280 
1281           return getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N), And.getOperand(0),
1282                           ShiftVal, WidthVal);
1283         }
1284       }
1285     } else if (N->getOperand(0).getOpcode() == ISD::SHL)
1286       return SelectS_BFEFromShifts(N);
1287     break;
1288   case ISD::SRA:
1289     if (N->getOperand(0).getOpcode() == ISD::SHL)
1290       return SelectS_BFEFromShifts(N);
1291     break;
1292   }
1293 
1294   return SelectCode(N);
1295 }
1296 
1297 bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src,
1298                                         SDValue &SrcMods) const {
1299 
1300   unsigned Mods = 0;
1301 
1302   Src = In;
1303 
1304   if (Src.getOpcode() == ISD::FNEG) {
1305     Mods |= SISrcMods::NEG;
1306     Src = Src.getOperand(0);
1307   }
1308 
1309   if (Src.getOpcode() == ISD::FABS) {
1310     Mods |= SISrcMods::ABS;
1311     Src = Src.getOperand(0);
1312   }
1313 
1314   SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
1315 
1316   return true;
1317 }
1318 
1319 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0(SDValue In, SDValue &Src,
1320                                          SDValue &SrcMods, SDValue &Clamp,
1321                                          SDValue &Omod) const {
1322   SDLoc DL(In);
1323   // FIXME: Handle Clamp and Omod
1324   Clamp = CurDAG->getTargetConstant(0, DL, MVT::i32);
1325   Omod = CurDAG->getTargetConstant(0, DL, MVT::i32);
1326 
1327   return SelectVOP3Mods(In, Src, SrcMods);
1328 }
1329 
1330 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp(SDValue In, SDValue &Src,
1331                                               SDValue &SrcMods,
1332                                               SDValue &Omod) const {
1333   // FIXME: Handle Omod
1334   Omod = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32);
1335 
1336   return SelectVOP3Mods(In, Src, SrcMods);
1337 }
1338 
1339 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src,
1340                                                    SDValue &SrcMods,
1341                                                    SDValue &Clamp,
1342                                                    SDValue &Omod) const {
1343   Clamp = Omod = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32);
1344   return SelectVOP3Mods(In, Src, SrcMods);
1345 }
1346 
1347 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
1348   const AMDGPUTargetLowering& Lowering =
1349     *static_cast<const AMDGPUTargetLowering*>(getTargetLowering());
1350   bool IsModified = false;
1351   do {
1352     IsModified = false;
1353     // Go over all selected nodes and try to fold them a bit more
1354     for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
1355          E = CurDAG->allnodes_end(); I != E; ++I) {
1356 
1357       SDNode *Node = I;
1358 
1359       MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I);
1360       if (!MachineNode)
1361         continue;
1362 
1363       SDNode *ResNode = Lowering.PostISelFolding(MachineNode, *CurDAG);
1364       if (ResNode != Node) {
1365         ReplaceUses(Node, ResNode);
1366         IsModified = true;
1367       }
1368     }
1369     CurDAG->RemoveDeadNodes();
1370   } while (IsModified);
1371 }
1372