SIISelLowering.cpp revision b89a467559d3eaade14993c7332afca3539f95d9 
1//===-- SIISelLowering.cpp - SI DAG Lowering Implementation ---------------===//
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 Custom DAG lowering for SI
12//
13//===----------------------------------------------------------------------===//
14
15#include "SIISelLowering.h"
16#include "AMDGPU.h"
17#include "AMDIL.h"
18#include "AMDILIntrinsicInfo.h"
19#include "SIInstrInfo.h"
20#include "SIMachineFunctionInfo.h"
21#include "SIRegisterInfo.h"
22#include "llvm/CodeGen/CallingConvLower.h"
23#include "llvm/CodeGen/MachineInstrBuilder.h"
24#include "llvm/CodeGen/MachineRegisterInfo.h"
25#include "llvm/CodeGen/SelectionDAG.h"
26#include "llvm/IR/Function.h"
27
28const uint64_t RSRC_DATA_FORMAT = 0xf00000000000LL;
29
30using namespace llvm;
31
32SITargetLowering::SITargetLowering(TargetMachine &TM) :
33    AMDGPUTargetLowering(TM),
34    TII(static_cast<const SIInstrInfo*>(TM.getInstrInfo())),
35    TRI(TM.getRegisterInfo()) {
36
37  addRegisterClass(MVT::i1, &AMDGPU::SReg_64RegClass);
38  addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
39
40  addRegisterClass(MVT::v16i8, &AMDGPU::SReg_128RegClass);
41  addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
42  addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
43
44  addRegisterClass(MVT::i32, &AMDGPU::VReg_32RegClass);
45  addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
46
47  addRegisterClass(MVT::v1i32, &AMDGPU::VReg_32RegClass);
48
49  addRegisterClass(MVT::v2i32, &AMDGPU::VReg_64RegClass);
50  addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
51
52  addRegisterClass(MVT::v4i32, &AMDGPU::VReg_128RegClass);
53  addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
54  addRegisterClass(MVT::i128, &AMDGPU::SReg_128RegClass);
55
56  addRegisterClass(MVT::v8i32, &AMDGPU::VReg_256RegClass);
57  addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
58
59  addRegisterClass(MVT::v16i32, &AMDGPU::VReg_512RegClass);
60  addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
61
62  computeRegisterProperties();
63
64  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
65  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
66  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
67  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
68
69  setOperationAction(ISD::ADD, MVT::i64, Legal);
70  setOperationAction(ISD::ADD, MVT::i32, Legal);
71
72  setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
73  setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
74
75  setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
76
77  setTargetDAGCombine(ISD::SELECT_CC);
78
79  setTargetDAGCombine(ISD::SETCC);
80
81  setSchedulingPreference(Sched::RegPressure);
82}
83
84SDValue SITargetLowering::LowerFormalArguments(
85                                      SDValue Chain,
86                                      CallingConv::ID CallConv,
87                                      bool isVarArg,
88                                      const SmallVectorImpl<ISD::InputArg> &Ins,
89                                      SDLoc DL, SelectionDAG &DAG,
90                                      SmallVectorImpl<SDValue> &InVals) const {
91
92  const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
93
94  MachineFunction &MF = DAG.getMachineFunction();
95  FunctionType *FType = MF.getFunction()->getFunctionType();
96  SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
97
98  assert(CallConv == CallingConv::C);
99
100  SmallVector<ISD::InputArg, 16> Splits;
101  uint32_t Skipped = 0;
102
103  for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
104    const ISD::InputArg &Arg = Ins[i];
105
106    // First check if it's a PS input addr
107    if (Info->ShaderType == ShaderType::PIXEL && !Arg.Flags.isInReg()) {
108
109      assert((PSInputNum <= 15) && "Too many PS inputs!");
110
111      if (!Arg.Used) {
112        // We can savely skip PS inputs
113        Skipped |= 1 << i;
114        ++PSInputNum;
115        continue;
116      }
117
118      Info->PSInputAddr |= 1 << PSInputNum++;
119    }
120
121    // Second split vertices into their elements
122    if (Arg.VT.isVector()) {
123      ISD::InputArg NewArg = Arg;
124      NewArg.Flags.setSplit();
125      NewArg.VT = Arg.VT.getVectorElementType();
126
127      // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
128      // three or five element vertex only needs three or five registers,
129      // NOT four or eigth.
130      Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
131      unsigned NumElements = ParamType->getVectorNumElements();
132
133      for (unsigned j = 0; j != NumElements; ++j) {
134        Splits.push_back(NewArg);
135        NewArg.PartOffset += NewArg.VT.getStoreSize();
136      }
137
138    } else {
139      Splits.push_back(Arg);
140    }
141  }
142
143  SmallVector<CCValAssign, 16> ArgLocs;
144  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
145                 getTargetMachine(), ArgLocs, *DAG.getContext());
146
147  // At least one interpolation mode must be enabled or else the GPU will hang.
148  if (Info->ShaderType == ShaderType::PIXEL && (Info->PSInputAddr & 0x7F) == 0) {
149    Info->PSInputAddr |= 1;
150    CCInfo.AllocateReg(AMDGPU::VGPR0);
151    CCInfo.AllocateReg(AMDGPU::VGPR1);
152  }
153
154  AnalyzeFormalArguments(CCInfo, Splits);
155
156  for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
157
158    const ISD::InputArg &Arg = Ins[i];
159    if (Skipped & (1 << i)) {
160      InVals.push_back(DAG.getUNDEF(Arg.VT));
161      continue;
162    }
163
164    CCValAssign &VA = ArgLocs[ArgIdx++];
165    assert(VA.isRegLoc() && "Parameter must be in a register!");
166
167    unsigned Reg = VA.getLocReg();
168    MVT VT = VA.getLocVT();
169
170    if (VT == MVT::i64) {
171      // For now assume it is a pointer
172      Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
173                                     &AMDGPU::SReg_64RegClass);
174      Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
175      InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
176      continue;
177    }
178
179    const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
180
181    Reg = MF.addLiveIn(Reg, RC);
182    SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
183
184    if (Arg.VT.isVector()) {
185
186      // Build a vector from the registers
187      Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
188      unsigned NumElements = ParamType->getVectorNumElements();
189
190      SmallVector<SDValue, 4> Regs;
191      Regs.push_back(Val);
192      for (unsigned j = 1; j != NumElements; ++j) {
193        Reg = ArgLocs[ArgIdx++].getLocReg();
194        Reg = MF.addLiveIn(Reg, RC);
195        Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
196      }
197
198      // Fill up the missing vector elements
199      NumElements = Arg.VT.getVectorNumElements() - NumElements;
200      for (unsigned j = 0; j != NumElements; ++j)
201        Regs.push_back(DAG.getUNDEF(VT));
202
203      InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT,
204                                   Regs.data(), Regs.size()));
205      continue;
206    }
207
208    InVals.push_back(Val);
209  }
210  return Chain;
211}
212
213MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
214    MachineInstr * MI, MachineBasicBlock * BB) const {
215
216  MachineBasicBlock::iterator I = *MI;
217
218  switch (MI->getOpcode()) {
219  default:
220    return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
221  case AMDGPU::BRANCH: return BB;
222  case AMDGPU::SI_ADDR64_RSRC: {
223    MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
224    unsigned SuperReg = MI->getOperand(0).getReg();
225    unsigned SubRegLo = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
226    unsigned SubRegHi = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
227    unsigned SubRegHiHi = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
228    unsigned SubRegHiLo = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
229    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B64), SubRegLo)
230            .addOperand(MI->getOperand(1));
231    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiLo)
232            .addImm(0);
233    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiHi)
234            .addImm(RSRC_DATA_FORMAT >> 32);
235    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SubRegHi)
236            .addReg(SubRegHiLo)
237            .addImm(AMDGPU::sub0)
238            .addReg(SubRegHiHi)
239            .addImm(AMDGPU::sub1);
240    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SuperReg)
241            .addReg(SubRegLo)
242            .addImm(AMDGPU::sub0_sub1)
243            .addReg(SubRegHi)
244            .addImm(AMDGPU::sub2_sub3);
245    MI->eraseFromParent();
246    break;
247  }
248  }
249  return BB;
250}
251
252EVT SITargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
253  return MVT::i1;
254}
255
256MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
257  return MVT::i32;
258}
259
260//===----------------------------------------------------------------------===//
261// Custom DAG Lowering Operations
262//===----------------------------------------------------------------------===//
263
264SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
265  switch (Op.getOpcode()) {
266  default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
267  case ISD::BRCOND: return LowerBRCOND(Op, DAG);
268  case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
269  }
270  return SDValue();
271}
272
273/// \brief Helper function for LowerBRCOND
274static SDNode *findUser(SDValue Value, unsigned Opcode) {
275
276  SDNode *Parent = Value.getNode();
277  for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
278       I != E; ++I) {
279
280    if (I.getUse().get() != Value)
281      continue;
282
283    if (I->getOpcode() == Opcode)
284      return *I;
285  }
286  return 0;
287}
288
289/// This transforms the control flow intrinsics to get the branch destination as
290/// last parameter, also switches branch target with BR if the need arise
291SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
292                                      SelectionDAG &DAG) const {
293
294  SDLoc DL(BRCOND);
295
296  SDNode *Intr = BRCOND.getOperand(1).getNode();
297  SDValue Target = BRCOND.getOperand(2);
298  SDNode *BR = 0;
299
300  if (Intr->getOpcode() == ISD::SETCC) {
301    // As long as we negate the condition everything is fine
302    SDNode *SetCC = Intr;
303    assert(SetCC->getConstantOperandVal(1) == 1);
304    assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
305           ISD::SETNE);
306    Intr = SetCC->getOperand(0).getNode();
307
308  } else {
309    // Get the target from BR if we don't negate the condition
310    BR = findUser(BRCOND, ISD::BR);
311    Target = BR->getOperand(1);
312  }
313
314  assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
315
316  // Build the result and
317  SmallVector<EVT, 4> Res;
318  for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
319    Res.push_back(Intr->getValueType(i));
320
321  // operands of the new intrinsic call
322  SmallVector<SDValue, 4> Ops;
323  Ops.push_back(BRCOND.getOperand(0));
324  for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
325    Ops.push_back(Intr->getOperand(i));
326  Ops.push_back(Target);
327
328  // build the new intrinsic call
329  SDNode *Result = DAG.getNode(
330    Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
331    DAG.getVTList(Res.data(), Res.size()), Ops.data(), Ops.size()).getNode();
332
333  if (BR) {
334    // Give the branch instruction our target
335    SDValue Ops[] = {
336      BR->getOperand(0),
337      BRCOND.getOperand(2)
338    };
339    DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops, 2);
340  }
341
342  SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
343
344  // Copy the intrinsic results to registers
345  for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
346    SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
347    if (!CopyToReg)
348      continue;
349
350    Chain = DAG.getCopyToReg(
351      Chain, DL,
352      CopyToReg->getOperand(1),
353      SDValue(Result, i - 1),
354      SDValue());
355
356    DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
357  }
358
359  // Remove the old intrinsic from the chain
360  DAG.ReplaceAllUsesOfValueWith(
361    SDValue(Intr, Intr->getNumValues() - 1),
362    Intr->getOperand(0));
363
364  return Chain;
365}
366
367SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
368  SDValue LHS = Op.getOperand(0);
369  SDValue RHS = Op.getOperand(1);
370  SDValue True = Op.getOperand(2);
371  SDValue False = Op.getOperand(3);
372  SDValue CC = Op.getOperand(4);
373  EVT VT = Op.getValueType();
374  SDLoc DL(Op);
375
376  // Possible Min/Max pattern
377  SDValue MinMax = LowerMinMax(Op, DAG);
378  if (MinMax.getNode()) {
379    return MinMax;
380  }
381
382  SDValue Cond = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, CC);
383  return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False);
384}
385
386//===----------------------------------------------------------------------===//
387// Custom DAG optimizations
388//===----------------------------------------------------------------------===//
389
390SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
391                                            DAGCombinerInfo &DCI) const {
392  SelectionDAG &DAG = DCI.DAG;
393  SDLoc DL(N);
394  EVT VT = N->getValueType(0);
395
396  switch (N->getOpcode()) {
397    default: break;
398    case ISD::SELECT_CC: {
399      N->dump();
400      ConstantSDNode *True, *False;
401      // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
402      if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
403          && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
404          && True->isAllOnesValue()
405          && False->isNullValue()
406          && VT == MVT::i1) {
407        return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
408                           N->getOperand(1), N->getOperand(4));
409
410      }
411      break;
412    }
413    case ISD::SETCC: {
414      SDValue Arg0 = N->getOperand(0);
415      SDValue Arg1 = N->getOperand(1);
416      SDValue CC = N->getOperand(2);
417      ConstantSDNode * C = NULL;
418      ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
419
420      // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
421      if (VT == MVT::i1
422          && Arg0.getOpcode() == ISD::SIGN_EXTEND
423          && Arg0.getOperand(0).getValueType() == MVT::i1
424          && (C = dyn_cast<ConstantSDNode>(Arg1))
425          && C->isNullValue()
426          && CCOp == ISD::SETNE) {
427        return SimplifySetCC(VT, Arg0.getOperand(0),
428                             DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
429      }
430      break;
431    }
432  }
433  return SDValue();
434}
435
436/// \brief Test if RegClass is one of the VSrc classes
437static bool isVSrc(unsigned RegClass) {
438  return AMDGPU::VSrc_32RegClassID == RegClass ||
439         AMDGPU::VSrc_64RegClassID == RegClass;
440}
441
442/// \brief Test if RegClass is one of the SSrc classes
443static bool isSSrc(unsigned RegClass) {
444  return AMDGPU::SSrc_32RegClassID == RegClass ||
445         AMDGPU::SSrc_64RegClassID == RegClass;
446}
447
448/// \brief Analyze the possible immediate value Op
449///
450/// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
451/// and the immediate value if it's a literal immediate
452int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
453
454  union {
455    int32_t I;
456    float F;
457  } Imm;
458
459  if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
460    if (Node->getZExtValue() >> 32) {
461        return -1;
462    }
463    Imm.I = Node->getSExtValue();
464  } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N))
465    Imm.F = Node->getValueAPF().convertToFloat();
466  else
467    return -1; // It isn't an immediate
468
469  if ((Imm.I >= -16 && Imm.I <= 64) ||
470      Imm.F == 0.5f || Imm.F == -0.5f ||
471      Imm.F == 1.0f || Imm.F == -1.0f ||
472      Imm.F == 2.0f || Imm.F == -2.0f ||
473      Imm.F == 4.0f || Imm.F == -4.0f)
474    return 0; // It's an inline immediate
475
476  return Imm.I; // It's a literal immediate
477}
478
479/// \brief Try to fold an immediate directly into an instruction
480bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
481                               bool &ScalarSlotUsed) const {
482
483  MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
484  if (Mov == 0 || !TII->isMov(Mov->getMachineOpcode()))
485    return false;
486
487  const SDValue &Op = Mov->getOperand(0);
488  int32_t Value = analyzeImmediate(Op.getNode());
489  if (Value == -1) {
490    // Not an immediate at all
491    return false;
492
493  } else if (Value == 0) {
494    // Inline immediates can always be fold
495    Operand = Op;
496    return true;
497
498  } else if (Value == Immediate) {
499    // Already fold literal immediate
500    Operand = Op;
501    return true;
502
503  } else if (!ScalarSlotUsed && !Immediate) {
504    // Fold this literal immediate
505    ScalarSlotUsed = true;
506    Immediate = Value;
507    Operand = Op;
508    return true;
509
510  }
511
512  return false;
513}
514
515/// \brief Does "Op" fit into register class "RegClass" ?
516bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
517                                    unsigned RegClass) const {
518
519  MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
520  SDNode *Node = Op.getNode();
521
522  const TargetRegisterClass *OpClass;
523  if (MachineSDNode *MN = dyn_cast<MachineSDNode>(Node)) {
524    const MCInstrDesc &Desc = TII->get(MN->getMachineOpcode());
525    int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
526    if (OpClassID == -1) {
527      switch (MN->getMachineOpcode()) {
528      case AMDGPU::REG_SEQUENCE:
529        // Operand 0 is the register class id for REG_SEQUENCE instructions.
530        OpClass = TRI->getRegClass(
531                       cast<ConstantSDNode>(MN->getOperand(0))->getZExtValue());
532        break;
533      default:
534        OpClass = getRegClassFor(Op.getSimpleValueType());
535        break;
536      }
537    } else {
538      OpClass = TRI->getRegClass(OpClassID);
539    }
540
541  } else if (Node->getOpcode() == ISD::CopyFromReg) {
542    RegisterSDNode *Reg = cast<RegisterSDNode>(Node->getOperand(1).getNode());
543    OpClass = MRI.getRegClass(Reg->getReg());
544
545  } else
546    return false;
547
548  return TRI->getRegClass(RegClass)->hasSubClassEq(OpClass);
549}
550
551/// \brief Make sure that we don't exeed the number of allowed scalars
552void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
553                                       unsigned RegClass,
554                                       bool &ScalarSlotUsed) const {
555
556  // First map the operands register class to a destination class
557  if (RegClass == AMDGPU::VSrc_32RegClassID)
558    RegClass = AMDGPU::VReg_32RegClassID;
559  else if (RegClass == AMDGPU::VSrc_64RegClassID)
560    RegClass = AMDGPU::VReg_64RegClassID;
561  else
562    return;
563
564  // Nothing todo if they fit naturaly
565  if (fitsRegClass(DAG, Operand, RegClass))
566    return;
567
568  // If the scalar slot isn't used yet use it now
569  if (!ScalarSlotUsed) {
570    ScalarSlotUsed = true;
571    return;
572  }
573
574  // This is a conservative aproach, it is possible that we can't determine
575  // the correct register class and copy too often, but better save than sorry.
576  SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
577  SDNode *Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, SDLoc(),
578                                    Operand.getValueType(), Operand, RC);
579  Operand = SDValue(Node, 0);
580}
581
582/// \brief Try to fold the Nodes operands into the Node
583SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
584                                       SelectionDAG &DAG) const {
585
586  // Original encoding (either e32 or e64)
587  int Opcode = Node->getMachineOpcode();
588  const MCInstrDesc *Desc = &TII->get(Opcode);
589
590  unsigned NumDefs = Desc->getNumDefs();
591  unsigned NumOps = Desc->getNumOperands();
592
593  // Commuted opcode if available
594  int OpcodeRev = Desc->isCommutable() ? TII->commuteOpcode(Opcode) : -1;
595  const MCInstrDesc *DescRev = OpcodeRev == -1 ? 0 : &TII->get(OpcodeRev);
596
597  assert(!DescRev || DescRev->getNumDefs() == NumDefs);
598  assert(!DescRev || DescRev->getNumOperands() == NumOps);
599
600  // e64 version if available, -1 otherwise
601  int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
602  const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? 0 : &TII->get(OpcodeE64);
603
604  assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
605  assert(!DescE64 || DescE64->getNumOperands() == (NumOps + 4));
606
607  int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
608  bool HaveVSrc = false, HaveSSrc = false;
609
610  // First figure out what we alread have in this instruction
611  for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
612       i != e && Op < NumOps; ++i, ++Op) {
613
614    unsigned RegClass = Desc->OpInfo[Op].RegClass;
615    if (isVSrc(RegClass))
616      HaveVSrc = true;
617    else if (isSSrc(RegClass))
618      HaveSSrc = true;
619    else
620      continue;
621
622    int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
623    if (Imm != -1 && Imm != 0) {
624      // Literal immediate
625      Immediate = Imm;
626    }
627  }
628
629  // If we neither have VSrc nor SSrc it makes no sense to continue
630  if (!HaveVSrc && !HaveSSrc)
631    return Node;
632
633  // No scalar allowed when we have both VSrc and SSrc
634  bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
635
636  // Second go over the operands and try to fold them
637  std::vector<SDValue> Ops;
638  bool Promote2e64 = false;
639  for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
640       i != e && Op < NumOps; ++i, ++Op) {
641
642    const SDValue &Operand = Node->getOperand(i);
643    Ops.push_back(Operand);
644
645    // Already folded immediate ?
646    if (isa<ConstantSDNode>(Operand.getNode()) ||
647        isa<ConstantFPSDNode>(Operand.getNode()))
648      continue;
649
650    // Is this a VSrc or SSrc operand ?
651    unsigned RegClass = Desc->OpInfo[Op].RegClass;
652    if (isVSrc(RegClass) || isSSrc(RegClass)) {
653      // Try to fold the immediates
654      if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
655        // Folding didn't worked, make sure we don't hit the SReg limit
656        ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
657      }
658      continue;
659    }
660
661    if (i == 1 && DescRev && fitsRegClass(DAG, Ops[0], RegClass)) {
662
663      unsigned OtherRegClass = Desc->OpInfo[NumDefs].RegClass;
664      assert(isVSrc(OtherRegClass) || isSSrc(OtherRegClass));
665
666      // Test if it makes sense to swap operands
667      if (foldImm(Ops[1], Immediate, ScalarSlotUsed) ||
668          (!fitsRegClass(DAG, Ops[1], RegClass) &&
669           fitsRegClass(DAG, Ops[1], OtherRegClass))) {
670
671        // Swap commutable operands
672        SDValue Tmp = Ops[1];
673        Ops[1] = Ops[0];
674        Ops[0] = Tmp;
675
676        Desc = DescRev;
677        DescRev = 0;
678        continue;
679      }
680    }
681
682    if (DescE64 && !Immediate) {
683
684      // Test if it makes sense to switch to e64 encoding
685      unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
686      if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
687        continue;
688
689      int32_t TmpImm = -1;
690      if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
691          (!fitsRegClass(DAG, Ops[i], RegClass) &&
692           fitsRegClass(DAG, Ops[1], OtherRegClass))) {
693
694        // Switch to e64 encoding
695        Immediate = -1;
696        Promote2e64 = true;
697        Desc = DescE64;
698        DescE64 = 0;
699      }
700    }
701  }
702
703  if (Promote2e64) {
704    // Add the modifier flags while promoting
705    for (unsigned i = 0; i < 4; ++i)
706      Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
707  }
708
709  // Add optional chain and glue
710  for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
711    Ops.push_back(Node->getOperand(i));
712
713  // Nodes that have a glue result are not CSE'd by getMachineNode(), so in
714  // this case a brand new node is always be created, even if the operands
715  // are the same as before.  So, manually check if anything has been changed.
716  if (Desc->Opcode == Opcode) {
717    bool Changed = false;
718    for (unsigned i = 0, e = Node->getNumOperands(); i < e; ++i) {
719      if (Ops[i].getNode() != Node->getOperand(i).getNode()) {
720        Changed = true;
721        break;
722      }
723    }
724    if (!Changed) {
725      return Node;
726    }
727  }
728
729  // Create a complete new instruction
730  return DAG.getMachineNode(Desc->Opcode, SDLoc(Node), Node->getVTList(), Ops);
731}
732
733/// \brief Helper function for adjustWritemask
734static unsigned SubIdx2Lane(unsigned Idx) {
735  switch (Idx) {
736  default: return 0;
737  case AMDGPU::sub0: return 0;
738  case AMDGPU::sub1: return 1;
739  case AMDGPU::sub2: return 2;
740  case AMDGPU::sub3: return 3;
741  }
742}
743
744/// \brief Adjust the writemask of MIMG instructions
745void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
746                                       SelectionDAG &DAG) const {
747  SDNode *Users[4] = { };
748  unsigned Writemask = 0, Lane = 0;
749
750  // Try to figure out the used register components
751  for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
752       I != E; ++I) {
753
754    // Abort if we can't understand the usage
755    if (!I->isMachineOpcode() ||
756        I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
757      return;
758
759    Lane = SubIdx2Lane(I->getConstantOperandVal(1));
760
761    // Abort if we have more than one user per component
762    if (Users[Lane])
763      return;
764
765    Users[Lane] = *I;
766    Writemask |= 1 << Lane;
767  }
768
769  // Abort if all components are used
770  if (Writemask == 0xf)
771    return;
772
773  // Adjust the writemask in the node
774  std::vector<SDValue> Ops;
775  Ops.push_back(DAG.getTargetConstant(Writemask, MVT::i32));
776  for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
777    Ops.push_back(Node->getOperand(i));
778  Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops.data(), Ops.size());
779
780  // If we only got one lane, replace it with a copy
781  if (Writemask == (1U << Lane)) {
782    SDValue RC = DAG.getTargetConstant(AMDGPU::VReg_32RegClassID, MVT::i32);
783    SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
784                                      SDLoc(), Users[Lane]->getValueType(0),
785                                      SDValue(Node, 0), RC);
786    DAG.ReplaceAllUsesWith(Users[Lane], Copy);
787    return;
788  }
789
790  // Update the users of the node with the new indices
791  for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
792
793    SDNode *User = Users[i];
794    if (!User)
795      continue;
796
797    SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
798    DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
799
800    switch (Idx) {
801    default: break;
802    case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
803    case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
804    case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
805    }
806  }
807}
808
809/// \brief Fold the instructions after slecting them
810SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
811                                          SelectionDAG &DAG) const {
812
813  if (AMDGPU::isMIMG(Node->getMachineOpcode()) != -1)
814    adjustWritemask(Node, DAG);
815
816  return foldOperands(Node, DAG);
817}
818
819/// \brief Assign the register class depending on the number of
820/// bits set in the writemask
821void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
822                                                     SDNode *Node) const {
823  if (AMDGPU::isMIMG(MI->getOpcode()) == -1)
824    return;
825
826  unsigned VReg = MI->getOperand(0).getReg();
827  unsigned Writemask = MI->getOperand(1).getImm();
828  unsigned BitsSet = 0;
829  for (unsigned i = 0; i < 4; ++i)
830    BitsSet += Writemask & (1 << i) ? 1 : 0;
831
832  const TargetRegisterClass *RC;
833  switch (BitsSet) {
834  default: return;
835  case 1:  RC = &AMDGPU::VReg_32RegClass; break;
836  case 2:  RC = &AMDGPU::VReg_64RegClass; break;
837  case 3:  RC = &AMDGPU::VReg_96RegClass; break;
838  }
839
840  MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
841  MRI.setRegClass(VReg, RC);
842}
843