1 //===-- R600InstrInfo.cpp - R600 Instruction Information ------------------===//
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 R600 Implementation of TargetInstrInfo.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "R600InstrInfo.h"
16 #include "AMDGPU.h"
17 #include "AMDGPUSubtarget.h"
18 #include "AMDGPUTargetMachine.h"
19 #include "R600Defines.h"
20 #include "R600MachineFunctionInfo.h"
21 #include "R600RegisterInfo.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 
26 using namespace llvm;
27 
28 #define GET_INSTRINFO_CTOR_DTOR
29 #include "AMDGPUGenDFAPacketizer.inc"
30 
31 R600InstrInfo::R600InstrInfo(const AMDGPUSubtarget &st)
32     : AMDGPUInstrInfo(st), RI() {}
33 
34 const R600RegisterInfo &R600InstrInfo::getRegisterInfo() const {
35   return RI;
36 }
37 
38 bool R600InstrInfo::isTrig(const MachineInstr &MI) const {
39   return get(MI.getOpcode()).TSFlags & R600_InstFlag::TRIG;
40 }
41 
42 bool R600InstrInfo::isVector(const MachineInstr &MI) const {
43   return get(MI.getOpcode()).TSFlags & R600_InstFlag::VECTOR;
44 }
45 
46 void
47 R600InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
48                            MachineBasicBlock::iterator MI, DebugLoc DL,
49                            unsigned DestReg, unsigned SrcReg,
50                            bool KillSrc) const {
51   unsigned VectorComponents = 0;
52   if ((AMDGPU::R600_Reg128RegClass.contains(DestReg) ||
53       AMDGPU::R600_Reg128VerticalRegClass.contains(DestReg)) &&
54       (AMDGPU::R600_Reg128RegClass.contains(SrcReg) ||
55        AMDGPU::R600_Reg128VerticalRegClass.contains(SrcReg))) {
56     VectorComponents = 4;
57   } else if((AMDGPU::R600_Reg64RegClass.contains(DestReg) ||
58             AMDGPU::R600_Reg64VerticalRegClass.contains(DestReg)) &&
59             (AMDGPU::R600_Reg64RegClass.contains(SrcReg) ||
60              AMDGPU::R600_Reg64VerticalRegClass.contains(SrcReg))) {
61     VectorComponents = 2;
62   }
63 
64   if (VectorComponents > 0) {
65     for (unsigned I = 0; I < VectorComponents; I++) {
66       unsigned SubRegIndex = RI.getSubRegFromChannel(I);
67       buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
68                               RI.getSubReg(DestReg, SubRegIndex),
69                               RI.getSubReg(SrcReg, SubRegIndex))
70                               .addReg(DestReg,
71                                       RegState::Define | RegState::Implicit);
72     }
73   } else {
74     MachineInstr *NewMI = buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
75                                                   DestReg, SrcReg);
76     NewMI->getOperand(getOperandIdx(*NewMI, AMDGPU::OpName::src0))
77                                     .setIsKill(KillSrc);
78   }
79 }
80 
81 /// \returns true if \p MBBI can be moved into a new basic.
82 bool R600InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB,
83                                        MachineBasicBlock::iterator MBBI) const {
84   for (MachineInstr::const_mop_iterator I = MBBI->operands_begin(),
85                                         E = MBBI->operands_end(); I != E; ++I) {
86     if (I->isReg() && !TargetRegisterInfo::isVirtualRegister(I->getReg()) &&
87         I->isUse() && RI.isPhysRegLiveAcrossClauses(I->getReg()))
88       return false;
89   }
90   return true;
91 }
92 
93 bool R600InstrInfo::isMov(unsigned Opcode) const {
94 
95 
96   switch(Opcode) {
97   default: return false;
98   case AMDGPU::MOV:
99   case AMDGPU::MOV_IMM_F32:
100   case AMDGPU::MOV_IMM_I32:
101     return true;
102   }
103 }
104 
105 // Some instructions act as place holders to emulate operations that the GPU
106 // hardware does automatically. This function can be used to check if
107 // an opcode falls into this category.
108 bool R600InstrInfo::isPlaceHolderOpcode(unsigned Opcode) const {
109   switch (Opcode) {
110   default: return false;
111   case AMDGPU::RETURN:
112     return true;
113   }
114 }
115 
116 bool R600InstrInfo::isReductionOp(unsigned Opcode) const {
117   return false;
118 }
119 
120 bool R600InstrInfo::isCubeOp(unsigned Opcode) const {
121   switch(Opcode) {
122     default: return false;
123     case AMDGPU::CUBE_r600_pseudo:
124     case AMDGPU::CUBE_r600_real:
125     case AMDGPU::CUBE_eg_pseudo:
126     case AMDGPU::CUBE_eg_real:
127       return true;
128   }
129 }
130 
131 bool R600InstrInfo::isALUInstr(unsigned Opcode) const {
132   unsigned TargetFlags = get(Opcode).TSFlags;
133 
134   return (TargetFlags & R600_InstFlag::ALU_INST);
135 }
136 
137 bool R600InstrInfo::hasInstrModifiers(unsigned Opcode) const {
138   unsigned TargetFlags = get(Opcode).TSFlags;
139 
140   return ((TargetFlags & R600_InstFlag::OP1) |
141           (TargetFlags & R600_InstFlag::OP2) |
142           (TargetFlags & R600_InstFlag::OP3));
143 }
144 
145 bool R600InstrInfo::isLDSInstr(unsigned Opcode) const {
146   unsigned TargetFlags = get(Opcode).TSFlags;
147 
148   return ((TargetFlags & R600_InstFlag::LDS_1A) |
149           (TargetFlags & R600_InstFlag::LDS_1A1D) |
150           (TargetFlags & R600_InstFlag::LDS_1A2D));
151 }
152 
153 bool R600InstrInfo::isLDSNoRetInstr(unsigned Opcode) const {
154   return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) == -1;
155 }
156 
157 bool R600InstrInfo::isLDSRetInstr(unsigned Opcode) const {
158   return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) != -1;
159 }
160 
161 bool R600InstrInfo::canBeConsideredALU(const MachineInstr *MI) const {
162   if (isALUInstr(MI->getOpcode()))
163     return true;
164   if (isVector(*MI) || isCubeOp(MI->getOpcode()))
165     return true;
166   switch (MI->getOpcode()) {
167   case AMDGPU::PRED_X:
168   case AMDGPU::INTERP_PAIR_XY:
169   case AMDGPU::INTERP_PAIR_ZW:
170   case AMDGPU::INTERP_VEC_LOAD:
171   case AMDGPU::COPY:
172   case AMDGPU::DOT_4:
173     return true;
174   default:
175     return false;
176   }
177 }
178 
179 bool R600InstrInfo::isTransOnly(unsigned Opcode) const {
180   if (ST.hasCaymanISA())
181     return false;
182   return (get(Opcode).getSchedClass() == AMDGPU::Sched::TransALU);
183 }
184 
185 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
186   return isTransOnly(MI->getOpcode());
187 }
188 
189 bool R600InstrInfo::isVectorOnly(unsigned Opcode) const {
190   return (get(Opcode).getSchedClass() == AMDGPU::Sched::VecALU);
191 }
192 
193 bool R600InstrInfo::isVectorOnly(const MachineInstr *MI) const {
194   return isVectorOnly(MI->getOpcode());
195 }
196 
197 bool R600InstrInfo::isExport(unsigned Opcode) const {
198   return (get(Opcode).TSFlags & R600_InstFlag::IS_EXPORT);
199 }
200 
201 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
202   return ST.hasVertexCache() && IS_VTX(get(Opcode));
203 }
204 
205 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
206   const MachineFunction *MF = MI->getParent()->getParent();
207   return !AMDGPU::isCompute(MF->getFunction()->getCallingConv()) &&
208     usesVertexCache(MI->getOpcode());
209 }
210 
211 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
212   return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode));
213 }
214 
215 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
216   const MachineFunction *MF = MI->getParent()->getParent();
217   return (AMDGPU::isCompute(MF->getFunction()->getCallingConv()) &&
218           usesVertexCache(MI->getOpcode())) ||
219     usesTextureCache(MI->getOpcode());
220 }
221 
222 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
223   switch (Opcode) {
224   case AMDGPU::KILLGT:
225   case AMDGPU::GROUP_BARRIER:
226     return true;
227   default:
228     return false;
229   }
230 }
231 
232 bool R600InstrInfo::usesAddressRegister(MachineInstr *MI) const {
233   return  MI->findRegisterUseOperandIdx(AMDGPU::AR_X) != -1;
234 }
235 
236 bool R600InstrInfo::definesAddressRegister(MachineInstr *MI) const {
237   return MI->findRegisterDefOperandIdx(AMDGPU::AR_X) != -1;
238 }
239 
240 bool R600InstrInfo::readsLDSSrcReg(const MachineInstr *MI) const {
241   if (!isALUInstr(MI->getOpcode())) {
242     return false;
243   }
244   for (MachineInstr::const_mop_iterator I = MI->operands_begin(),
245                                         E = MI->operands_end(); I != E; ++I) {
246     if (!I->isReg() || !I->isUse() ||
247         TargetRegisterInfo::isVirtualRegister(I->getReg()))
248       continue;
249 
250     if (AMDGPU::R600_LDS_SRC_REGRegClass.contains(I->getReg()))
251       return true;
252   }
253   return false;
254 }
255 
256 int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const {
257   static const unsigned OpTable[] = {
258     AMDGPU::OpName::src0,
259     AMDGPU::OpName::src1,
260     AMDGPU::OpName::src2
261   };
262 
263   assert (SrcNum < 3);
264   return getOperandIdx(Opcode, OpTable[SrcNum]);
265 }
266 
267 int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const {
268   static const unsigned SrcSelTable[][2] = {
269     {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
270     {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
271     {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
272     {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
273     {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
274     {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
275     {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
276     {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
277     {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
278     {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
279     {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}
280   };
281 
282   for (const auto &Row : SrcSelTable) {
283     if (getOperandIdx(Opcode, Row[0]) == (int)SrcIdx) {
284       return getOperandIdx(Opcode, Row[1]);
285     }
286   }
287   return -1;
288 }
289 
290 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
291 R600InstrInfo::getSrcs(MachineInstr *MI) const {
292   SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
293 
294   if (MI->getOpcode() == AMDGPU::DOT_4) {
295     static const unsigned OpTable[8][2] = {
296       {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
297       {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
298       {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
299       {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
300       {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
301       {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
302       {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
303       {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W},
304     };
305 
306     for (unsigned j = 0; j < 8; j++) {
307       MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
308                                                         OpTable[j][0]));
309       unsigned Reg = MO.getReg();
310       if (Reg == AMDGPU::ALU_CONST) {
311         MachineOperand &Sel = MI->getOperand(getOperandIdx(MI->getOpcode(),
312                                                     OpTable[j][1]));
313         Result.push_back(std::make_pair(&MO, Sel.getImm()));
314         continue;
315       }
316 
317     }
318     return Result;
319   }
320 
321   static const unsigned OpTable[3][2] = {
322     {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
323     {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
324     {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
325   };
326 
327   for (unsigned j = 0; j < 3; j++) {
328     int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]);
329     if (SrcIdx < 0)
330       break;
331     MachineOperand &MO = MI->getOperand(SrcIdx);
332     unsigned Reg = MO.getReg();
333     if (Reg == AMDGPU::ALU_CONST) {
334       MachineOperand &Sel = MI->getOperand(
335           getOperandIdx(MI->getOpcode(), OpTable[j][1]));
336       Result.push_back(std::make_pair(&MO, Sel.getImm()));
337       continue;
338     }
339     if (Reg == AMDGPU::ALU_LITERAL_X) {
340       MachineOperand &Operand = MI->getOperand(
341           getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal));
342       if (Operand.isImm()) {
343         Result.push_back(std::make_pair(&MO, Operand.getImm()));
344         continue;
345       }
346       assert(Operand.isGlobal());
347     }
348     Result.push_back(std::make_pair(&MO, 0));
349   }
350   return Result;
351 }
352 
353 std::vector<std::pair<int, unsigned> >
354 R600InstrInfo::ExtractSrcs(MachineInstr *MI,
355                            const DenseMap<unsigned, unsigned> &PV,
356                            unsigned &ConstCount) const {
357   ConstCount = 0;
358   ArrayRef<std::pair<MachineOperand *, int64_t>> Srcs = getSrcs(MI);
359   const std::pair<int, unsigned> DummyPair(-1, 0);
360   std::vector<std::pair<int, unsigned> > Result;
361   unsigned i = 0;
362   for (unsigned n = Srcs.size(); i < n; ++i) {
363     unsigned Reg = Srcs[i].first->getReg();
364     int Index = RI.getEncodingValue(Reg) & 0xff;
365     if (Reg == AMDGPU::OQAP) {
366       Result.push_back(std::make_pair(Index, 0U));
367     }
368     if (PV.find(Reg) != PV.end()) {
369       // 255 is used to tells its a PS/PV reg
370       Result.push_back(std::make_pair(255, 0U));
371       continue;
372     }
373     if (Index > 127) {
374       ConstCount++;
375       Result.push_back(DummyPair);
376       continue;
377     }
378     unsigned Chan = RI.getHWRegChan(Reg);
379     Result.push_back(std::make_pair(Index, Chan));
380   }
381   for (; i < 3; ++i)
382     Result.push_back(DummyPair);
383   return Result;
384 }
385 
386 static std::vector<std::pair<int, unsigned> >
387 Swizzle(std::vector<std::pair<int, unsigned> > Src,
388         R600InstrInfo::BankSwizzle Swz) {
389   if (Src[0] == Src[1])
390     Src[1].first = -1;
391   switch (Swz) {
392   case R600InstrInfo::ALU_VEC_012_SCL_210:
393     break;
394   case R600InstrInfo::ALU_VEC_021_SCL_122:
395     std::swap(Src[1], Src[2]);
396     break;
397   case R600InstrInfo::ALU_VEC_102_SCL_221:
398     std::swap(Src[0], Src[1]);
399     break;
400   case R600InstrInfo::ALU_VEC_120_SCL_212:
401     std::swap(Src[0], Src[1]);
402     std::swap(Src[0], Src[2]);
403     break;
404   case R600InstrInfo::ALU_VEC_201:
405     std::swap(Src[0], Src[2]);
406     std::swap(Src[0], Src[1]);
407     break;
408   case R600InstrInfo::ALU_VEC_210:
409     std::swap(Src[0], Src[2]);
410     break;
411   }
412   return Src;
413 }
414 
415 static unsigned
416 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) {
417   switch (Swz) {
418   case R600InstrInfo::ALU_VEC_012_SCL_210: {
419     unsigned Cycles[3] = { 2, 1, 0};
420     return Cycles[Op];
421   }
422   case R600InstrInfo::ALU_VEC_021_SCL_122: {
423     unsigned Cycles[3] = { 1, 2, 2};
424     return Cycles[Op];
425   }
426   case R600InstrInfo::ALU_VEC_120_SCL_212: {
427     unsigned Cycles[3] = { 2, 1, 2};
428     return Cycles[Op];
429   }
430   case R600InstrInfo::ALU_VEC_102_SCL_221: {
431     unsigned Cycles[3] = { 2, 2, 1};
432     return Cycles[Op];
433   }
434   default:
435     llvm_unreachable("Wrong Swizzle for Trans Slot");
436     return 0;
437   }
438 }
439 
440 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed
441 /// in the same Instruction Group while meeting read port limitations given a
442 /// Swz swizzle sequence.
443 unsigned  R600InstrInfo::isLegalUpTo(
444     const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
445     const std::vector<R600InstrInfo::BankSwizzle> &Swz,
446     const std::vector<std::pair<int, unsigned> > &TransSrcs,
447     R600InstrInfo::BankSwizzle TransSwz) const {
448   int Vector[4][3];
449   memset(Vector, -1, sizeof(Vector));
450   for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) {
451     const std::vector<std::pair<int, unsigned> > &Srcs =
452         Swizzle(IGSrcs[i], Swz[i]);
453     for (unsigned j = 0; j < 3; j++) {
454       const std::pair<int, unsigned> &Src = Srcs[j];
455       if (Src.first < 0 || Src.first == 255)
456         continue;
457       if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) {
458         if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 &&
459             Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) {
460             // The value from output queue A (denoted by register OQAP) can
461             // only be fetched during the first cycle.
462             return false;
463         }
464         // OQAP does not count towards the normal read port restrictions
465         continue;
466       }
467       if (Vector[Src.second][j] < 0)
468         Vector[Src.second][j] = Src.first;
469       if (Vector[Src.second][j] != Src.first)
470         return i;
471     }
472   }
473   // Now check Trans Alu
474   for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) {
475     const std::pair<int, unsigned> &Src = TransSrcs[i];
476     unsigned Cycle = getTransSwizzle(TransSwz, i);
477     if (Src.first < 0)
478       continue;
479     if (Src.first == 255)
480       continue;
481     if (Vector[Src.second][Cycle] < 0)
482       Vector[Src.second][Cycle] = Src.first;
483     if (Vector[Src.second][Cycle] != Src.first)
484       return IGSrcs.size() - 1;
485   }
486   return IGSrcs.size();
487 }
488 
489 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next
490 /// (in lexicographic term) swizzle sequence assuming that all swizzles after
491 /// Idx can be skipped
492 static bool
493 NextPossibleSolution(
494     std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
495     unsigned Idx) {
496   assert(Idx < SwzCandidate.size());
497   int ResetIdx = Idx;
498   while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210)
499     ResetIdx --;
500   for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) {
501     SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210;
502   }
503   if (ResetIdx == -1)
504     return false;
505   int NextSwizzle = SwzCandidate[ResetIdx] + 1;
506   SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle;
507   return true;
508 }
509 
510 /// Enumerate all possible Swizzle sequence to find one that can meet all
511 /// read port requirements.
512 bool R600InstrInfo::FindSwizzleForVectorSlot(
513     const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
514     std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
515     const std::vector<std::pair<int, unsigned> > &TransSrcs,
516     R600InstrInfo::BankSwizzle TransSwz) const {
517   unsigned ValidUpTo = 0;
518   do {
519     ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz);
520     if (ValidUpTo == IGSrcs.size())
521       return true;
522   } while (NextPossibleSolution(SwzCandidate, ValidUpTo));
523   return false;
524 }
525 
526 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read
527 /// a const, and can't read a gpr at cycle 1 if they read 2 const.
528 static bool
529 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
530                   const std::vector<std::pair<int, unsigned> > &TransOps,
531                   unsigned ConstCount) {
532   // TransALU can't read 3 constants
533   if (ConstCount > 2)
534     return false;
535   for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
536     const std::pair<int, unsigned> &Src = TransOps[i];
537     unsigned Cycle = getTransSwizzle(TransSwz, i);
538     if (Src.first < 0)
539       continue;
540     if (ConstCount > 0 && Cycle == 0)
541       return false;
542     if (ConstCount > 1 && Cycle == 1)
543       return false;
544   }
545   return true;
546 }
547 
548 bool
549 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
550                                        const DenseMap<unsigned, unsigned> &PV,
551                                        std::vector<BankSwizzle> &ValidSwizzle,
552                                        bool isLastAluTrans)
553     const {
554   //Todo : support shared src0 - src1 operand
555 
556   std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
557   ValidSwizzle.clear();
558   unsigned ConstCount;
559   BankSwizzle TransBS = ALU_VEC_012_SCL_210;
560   for (unsigned i = 0, e = IG.size(); i < e; ++i) {
561     IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount));
562     unsigned Op = getOperandIdx(IG[i]->getOpcode(),
563         AMDGPU::OpName::bank_swizzle);
564     ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
565         IG[i]->getOperand(Op).getImm());
566   }
567   std::vector<std::pair<int, unsigned> > TransOps;
568   if (!isLastAluTrans)
569     return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
570 
571   TransOps = std::move(IGSrcs.back());
572   IGSrcs.pop_back();
573   ValidSwizzle.pop_back();
574 
575   static const R600InstrInfo::BankSwizzle TransSwz[] = {
576     ALU_VEC_012_SCL_210,
577     ALU_VEC_021_SCL_122,
578     ALU_VEC_120_SCL_212,
579     ALU_VEC_102_SCL_221
580   };
581   for (unsigned i = 0; i < 4; i++) {
582     TransBS = TransSwz[i];
583     if (!isConstCompatible(TransBS, TransOps, ConstCount))
584       continue;
585     bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps,
586         TransBS);
587     if (Result) {
588       ValidSwizzle.push_back(TransBS);
589       return true;
590     }
591   }
592 
593   return false;
594 }
595 
596 
597 bool
598 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
599     const {
600   assert (Consts.size() <= 12 && "Too many operands in instructions group");
601   unsigned Pair1 = 0, Pair2 = 0;
602   for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
603     unsigned ReadConstHalf = Consts[i] & 2;
604     unsigned ReadConstIndex = Consts[i] & (~3);
605     unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
606     if (!Pair1) {
607       Pair1 = ReadHalfConst;
608       continue;
609     }
610     if (Pair1 == ReadHalfConst)
611       continue;
612     if (!Pair2) {
613       Pair2 = ReadHalfConst;
614       continue;
615     }
616     if (Pair2 != ReadHalfConst)
617       return false;
618   }
619   return true;
620 }
621 
622 bool
623 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
624     const {
625   std::vector<unsigned> Consts;
626   SmallSet<int64_t, 4> Literals;
627   for (unsigned i = 0, n = MIs.size(); i < n; i++) {
628     MachineInstr *MI = MIs[i];
629     if (!isALUInstr(MI->getOpcode()))
630       continue;
631 
632     ArrayRef<std::pair<MachineOperand *, int64_t>> Srcs = getSrcs(MI);
633 
634     for (const auto &Src:Srcs) {
635       if (Src.first->getReg() == AMDGPU::ALU_LITERAL_X)
636         Literals.insert(Src.second);
637       if (Literals.size() > 4)
638         return false;
639       if (Src.first->getReg() == AMDGPU::ALU_CONST)
640         Consts.push_back(Src.second);
641       if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
642           AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) {
643         unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff;
644         unsigned Chan = RI.getHWRegChan(Src.first->getReg());
645         Consts.push_back((Index << 2) | Chan);
646       }
647     }
648   }
649   return fitsConstReadLimitations(Consts);
650 }
651 
652 DFAPacketizer *
653 R600InstrInfo::CreateTargetScheduleState(const TargetSubtargetInfo &STI) const {
654   const InstrItineraryData *II = STI.getInstrItineraryData();
655   return static_cast<const AMDGPUSubtarget &>(STI).createDFAPacketizer(II);
656 }
657 
658 static bool
659 isPredicateSetter(unsigned Opcode) {
660   switch (Opcode) {
661   case AMDGPU::PRED_X:
662     return true;
663   default:
664     return false;
665   }
666 }
667 
668 static MachineInstr *
669 findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
670                              MachineBasicBlock::iterator I) {
671   while (I != MBB.begin()) {
672     --I;
673     MachineInstr *MI = I;
674     if (isPredicateSetter(MI->getOpcode()))
675       return MI;
676   }
677 
678   return nullptr;
679 }
680 
681 static
682 bool isJump(unsigned Opcode) {
683   return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
684 }
685 
686 static bool isBranch(unsigned Opcode) {
687   return Opcode == AMDGPU::BRANCH || Opcode == AMDGPU::BRANCH_COND_i32 ||
688       Opcode == AMDGPU::BRANCH_COND_f32;
689 }
690 
691 bool
692 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
693                              MachineBasicBlock *&TBB,
694                              MachineBasicBlock *&FBB,
695                              SmallVectorImpl<MachineOperand> &Cond,
696                              bool AllowModify) const {
697   // Most of the following comes from the ARM implementation of AnalyzeBranch
698 
699   // If the block has no terminators, it just falls into the block after it.
700   MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
701   if (I == MBB.end())
702     return false;
703 
704   // AMDGPU::BRANCH* instructions are only available after isel and are not
705   // handled
706   if (isBranch(I->getOpcode()))
707     return true;
708   if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
709     return false;
710   }
711 
712   // Remove successive JUMP
713   while (I != MBB.begin() && std::prev(I)->getOpcode() == AMDGPU::JUMP) {
714       MachineBasicBlock::iterator PriorI = std::prev(I);
715       if (AllowModify)
716         I->removeFromParent();
717       I = PriorI;
718   }
719   MachineInstr *LastInst = I;
720 
721   // If there is only one terminator instruction, process it.
722   unsigned LastOpc = LastInst->getOpcode();
723   if (I == MBB.begin() ||
724           !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) {
725     if (LastOpc == AMDGPU::JUMP) {
726       TBB = LastInst->getOperand(0).getMBB();
727       return false;
728     } else if (LastOpc == AMDGPU::JUMP_COND) {
729       MachineInstr *predSet = I;
730       while (!isPredicateSetter(predSet->getOpcode())) {
731         predSet = --I;
732       }
733       TBB = LastInst->getOperand(0).getMBB();
734       Cond.push_back(predSet->getOperand(1));
735       Cond.push_back(predSet->getOperand(2));
736       Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
737       return false;
738     }
739     return true;  // Can't handle indirect branch.
740   }
741 
742   // Get the instruction before it if it is a terminator.
743   MachineInstr *SecondLastInst = I;
744   unsigned SecondLastOpc = SecondLastInst->getOpcode();
745 
746   // If the block ends with a B and a Bcc, handle it.
747   if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) {
748     MachineInstr *predSet = --I;
749     while (!isPredicateSetter(predSet->getOpcode())) {
750       predSet = --I;
751     }
752     TBB = SecondLastInst->getOperand(0).getMBB();
753     FBB = LastInst->getOperand(0).getMBB();
754     Cond.push_back(predSet->getOperand(1));
755     Cond.push_back(predSet->getOperand(2));
756     Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
757     return false;
758   }
759 
760   // Otherwise, can't handle this.
761   return true;
762 }
763 
764 static
765 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) {
766   for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend();
767       It != E; ++It) {
768     if (It->getOpcode() == AMDGPU::CF_ALU ||
769         It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
770       return std::prev(It.base());
771   }
772   return MBB.end();
773 }
774 
775 unsigned
776 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
777                             MachineBasicBlock *TBB,
778                             MachineBasicBlock *FBB,
779                             ArrayRef<MachineOperand> Cond,
780                             DebugLoc DL) const {
781   assert(TBB && "InsertBranch must not be told to insert a fallthrough");
782 
783   if (!FBB) {
784     if (Cond.empty()) {
785       BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
786       return 1;
787     } else {
788       MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
789       assert(PredSet && "No previous predicate !");
790       addFlag(PredSet, 0, MO_FLAG_PUSH);
791       PredSet->getOperand(2).setImm(Cond[1].getImm());
792 
793       BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
794              .addMBB(TBB)
795              .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
796       MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
797       if (CfAlu == MBB.end())
798         return 1;
799       assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
800       CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
801       return 1;
802     }
803   } else {
804     MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
805     assert(PredSet && "No previous predicate !");
806     addFlag(PredSet, 0, MO_FLAG_PUSH);
807     PredSet->getOperand(2).setImm(Cond[1].getImm());
808     BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
809             .addMBB(TBB)
810             .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
811     BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB);
812     MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
813     if (CfAlu == MBB.end())
814       return 2;
815     assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
816     CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
817     return 2;
818   }
819 }
820 
821 unsigned
822 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
823 
824   // Note : we leave PRED* instructions there.
825   // They may be needed when predicating instructions.
826 
827   MachineBasicBlock::iterator I = MBB.end();
828 
829   if (I == MBB.begin()) {
830     return 0;
831   }
832   --I;
833   switch (I->getOpcode()) {
834   default:
835     return 0;
836   case AMDGPU::JUMP_COND: {
837     MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
838     clearFlag(predSet, 0, MO_FLAG_PUSH);
839     I->eraseFromParent();
840     MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
841     if (CfAlu == MBB.end())
842       break;
843     assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
844     CfAlu->setDesc(get(AMDGPU::CF_ALU));
845     break;
846   }
847   case AMDGPU::JUMP:
848     I->eraseFromParent();
849     break;
850   }
851   I = MBB.end();
852 
853   if (I == MBB.begin()) {
854     return 1;
855   }
856   --I;
857   switch (I->getOpcode()) {
858     // FIXME: only one case??
859   default:
860     return 1;
861   case AMDGPU::JUMP_COND: {
862     MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
863     clearFlag(predSet, 0, MO_FLAG_PUSH);
864     I->eraseFromParent();
865     MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
866     if (CfAlu == MBB.end())
867       break;
868     assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
869     CfAlu->setDesc(get(AMDGPU::CF_ALU));
870     break;
871   }
872   case AMDGPU::JUMP:
873     I->eraseFromParent();
874     break;
875   }
876   return 2;
877 }
878 
879 bool R600InstrInfo::isPredicated(const MachineInstr &MI) const {
880   int idx = MI.findFirstPredOperandIdx();
881   if (idx < 0)
882     return false;
883 
884   unsigned Reg = MI.getOperand(idx).getReg();
885   switch (Reg) {
886   default: return false;
887   case AMDGPU::PRED_SEL_ONE:
888   case AMDGPU::PRED_SEL_ZERO:
889   case AMDGPU::PREDICATE_BIT:
890     return true;
891   }
892 }
893 
894 bool R600InstrInfo::isPredicable(MachineInstr &MI) const {
895   // XXX: KILL* instructions can be predicated, but they must be the last
896   // instruction in a clause, so this means any instructions after them cannot
897   // be predicated.  Until we have proper support for instruction clauses in the
898   // backend, we will mark KILL* instructions as unpredicable.
899 
900   if (MI.getOpcode() == AMDGPU::KILLGT) {
901     return false;
902   } else if (MI.getOpcode() == AMDGPU::CF_ALU) {
903     // If the clause start in the middle of MBB then the MBB has more
904     // than a single clause, unable to predicate several clauses.
905     if (MI.getParent()->begin() != MachineBasicBlock::iterator(MI))
906       return false;
907     // TODO: We don't support KC merging atm
908     return MI.getOperand(3).getImm() == 0 && MI.getOperand(4).getImm() == 0;
909   } else if (isVector(MI)) {
910     return false;
911   } else {
912     return AMDGPUInstrInfo::isPredicable(MI);
913   }
914 }
915 
916 
917 bool
918 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
919                                    unsigned NumCyles,
920                                    unsigned ExtraPredCycles,
921                                    BranchProbability Probability) const{
922   return true;
923 }
924 
925 bool
926 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
927                                    unsigned NumTCycles,
928                                    unsigned ExtraTCycles,
929                                    MachineBasicBlock &FMBB,
930                                    unsigned NumFCycles,
931                                    unsigned ExtraFCycles,
932                                    BranchProbability Probability) const {
933   return true;
934 }
935 
936 bool
937 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
938                                          unsigned NumCyles,
939                                          BranchProbability Probability)
940                                          const {
941   return true;
942 }
943 
944 bool
945 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
946                                          MachineBasicBlock &FMBB) const {
947   return false;
948 }
949 
950 
951 bool
952 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
953   MachineOperand &MO = Cond[1];
954   switch (MO.getImm()) {
955   case OPCODE_IS_ZERO_INT:
956     MO.setImm(OPCODE_IS_NOT_ZERO_INT);
957     break;
958   case OPCODE_IS_NOT_ZERO_INT:
959     MO.setImm(OPCODE_IS_ZERO_INT);
960     break;
961   case OPCODE_IS_ZERO:
962     MO.setImm(OPCODE_IS_NOT_ZERO);
963     break;
964   case OPCODE_IS_NOT_ZERO:
965     MO.setImm(OPCODE_IS_ZERO);
966     break;
967   default:
968     return true;
969   }
970 
971   MachineOperand &MO2 = Cond[2];
972   switch (MO2.getReg()) {
973   case AMDGPU::PRED_SEL_ZERO:
974     MO2.setReg(AMDGPU::PRED_SEL_ONE);
975     break;
976   case AMDGPU::PRED_SEL_ONE:
977     MO2.setReg(AMDGPU::PRED_SEL_ZERO);
978     break;
979   default:
980     return true;
981   }
982   return false;
983 }
984 
985 bool R600InstrInfo::DefinesPredicate(MachineInstr &MI,
986                                      std::vector<MachineOperand> &Pred) const {
987   return isPredicateSetter(MI.getOpcode());
988 }
989 
990 
991 bool
992 R600InstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
993                                  ArrayRef<MachineOperand> Pred2) const {
994   return false;
995 }
996 
997 bool R600InstrInfo::PredicateInstruction(MachineInstr &MI,
998                                          ArrayRef<MachineOperand> Pred) const {
999   int PIdx = MI.findFirstPredOperandIdx();
1000 
1001   if (MI.getOpcode() == AMDGPU::CF_ALU) {
1002     MI.getOperand(8).setImm(0);
1003     return true;
1004   }
1005 
1006   if (MI.getOpcode() == AMDGPU::DOT_4) {
1007     MI.getOperand(getOperandIdx(MI, AMDGPU::OpName::pred_sel_X))
1008         .setReg(Pred[2].getReg());
1009     MI.getOperand(getOperandIdx(MI, AMDGPU::OpName::pred_sel_Y))
1010         .setReg(Pred[2].getReg());
1011     MI.getOperand(getOperandIdx(MI, AMDGPU::OpName::pred_sel_Z))
1012         .setReg(Pred[2].getReg());
1013     MI.getOperand(getOperandIdx(MI, AMDGPU::OpName::pred_sel_W))
1014         .setReg(Pred[2].getReg());
1015     MachineInstrBuilder MIB(*MI.getParent()->getParent(), MI);
1016     MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
1017     return true;
1018   }
1019 
1020   if (PIdx != -1) {
1021     MachineOperand &PMO = MI.getOperand(PIdx);
1022     PMO.setReg(Pred[2].getReg());
1023     MachineInstrBuilder MIB(*MI.getParent()->getParent(), MI);
1024     MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
1025     return true;
1026   }
1027 
1028   return false;
1029 }
1030 
1031 unsigned int R600InstrInfo::getPredicationCost(const MachineInstr &) const {
1032   return 2;
1033 }
1034 
1035 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1036                                             const MachineInstr *MI,
1037                                             unsigned *PredCost) const {
1038   if (PredCost)
1039     *PredCost = 2;
1040   return 2;
1041 }
1042 
1043 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
1044                                                    unsigned Channel) const {
1045   assert(Channel == 0);
1046   return RegIndex;
1047 }
1048 
1049 bool R600InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
1050 
1051   switch(MI->getOpcode()) {
1052   default: {
1053     MachineBasicBlock *MBB = MI->getParent();
1054     int OffsetOpIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1055                                                  AMDGPU::OpName::addr);
1056      // addr is a custom operand with multiple MI operands, and only the
1057      // first MI operand is given a name.
1058     int RegOpIdx = OffsetOpIdx + 1;
1059     int ChanOpIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1060                                                AMDGPU::OpName::chan);
1061     if (isRegisterLoad(*MI)) {
1062       int DstOpIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1063                                                 AMDGPU::OpName::dst);
1064       unsigned RegIndex = MI->getOperand(RegOpIdx).getImm();
1065       unsigned Channel = MI->getOperand(ChanOpIdx).getImm();
1066       unsigned Address = calculateIndirectAddress(RegIndex, Channel);
1067       unsigned OffsetReg = MI->getOperand(OffsetOpIdx).getReg();
1068       if (OffsetReg == AMDGPU::INDIRECT_BASE_ADDR) {
1069         buildMovInstr(MBB, MI, MI->getOperand(DstOpIdx).getReg(),
1070                       getIndirectAddrRegClass()->getRegister(Address));
1071       } else {
1072         buildIndirectRead(MBB, MI, MI->getOperand(DstOpIdx).getReg(),
1073                           Address, OffsetReg);
1074       }
1075     } else if (isRegisterStore(*MI)) {
1076       int ValOpIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1077                                                 AMDGPU::OpName::val);
1078       unsigned RegIndex = MI->getOperand(RegOpIdx).getImm();
1079       unsigned Channel = MI->getOperand(ChanOpIdx).getImm();
1080       unsigned Address = calculateIndirectAddress(RegIndex, Channel);
1081       unsigned OffsetReg = MI->getOperand(OffsetOpIdx).getReg();
1082       if (OffsetReg == AMDGPU::INDIRECT_BASE_ADDR) {
1083         buildMovInstr(MBB, MI, getIndirectAddrRegClass()->getRegister(Address),
1084                       MI->getOperand(ValOpIdx).getReg());
1085       } else {
1086         buildIndirectWrite(MBB, MI, MI->getOperand(ValOpIdx).getReg(),
1087                            calculateIndirectAddress(RegIndex, Channel),
1088                            OffsetReg);
1089       }
1090     } else {
1091       return false;
1092     }
1093 
1094     MBB->erase(MI);
1095     return true;
1096   }
1097   case AMDGPU::R600_EXTRACT_ELT_V2:
1098   case AMDGPU::R600_EXTRACT_ELT_V4:
1099     buildIndirectRead(MI->getParent(), MI, MI->getOperand(0).getReg(),
1100                       RI.getHWRegIndex(MI->getOperand(1).getReg()), //  Address
1101                       MI->getOperand(2).getReg(),
1102                       RI.getHWRegChan(MI->getOperand(1).getReg()));
1103     break;
1104   case AMDGPU::R600_INSERT_ELT_V2:
1105   case AMDGPU::R600_INSERT_ELT_V4:
1106     buildIndirectWrite(MI->getParent(), MI, MI->getOperand(2).getReg(), // Value
1107                        RI.getHWRegIndex(MI->getOperand(1).getReg()),  // Address
1108                        MI->getOperand(3).getReg(),                    // Offset
1109                        RI.getHWRegChan(MI->getOperand(1).getReg()));  // Channel
1110     break;
1111   }
1112   MI->eraseFromParent();
1113   return true;
1114 }
1115 
1116 void  R600InstrInfo::reserveIndirectRegisters(BitVector &Reserved,
1117                                              const MachineFunction &MF) const {
1118   const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
1119       MF.getSubtarget().getFrameLowering());
1120 
1121   unsigned StackWidth = TFL->getStackWidth(MF);
1122   int End = getIndirectIndexEnd(MF);
1123 
1124   if (End == -1)
1125     return;
1126 
1127   for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
1128     unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
1129     Reserved.set(SuperReg);
1130     for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
1131       unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
1132       Reserved.set(Reg);
1133     }
1134   }
1135 }
1136 
1137 const TargetRegisterClass *R600InstrInfo::getIndirectAddrRegClass() const {
1138   return &AMDGPU::R600_TReg32_XRegClass;
1139 }
1140 
1141 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1142                                        MachineBasicBlock::iterator I,
1143                                        unsigned ValueReg, unsigned Address,
1144                                        unsigned OffsetReg) const {
1145   return buildIndirectWrite(MBB, I, ValueReg, Address, OffsetReg, 0);
1146 }
1147 
1148 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1149                                        MachineBasicBlock::iterator I,
1150                                        unsigned ValueReg, unsigned Address,
1151                                        unsigned OffsetReg,
1152                                        unsigned AddrChan) const {
1153   unsigned AddrReg;
1154   switch (AddrChan) {
1155     default: llvm_unreachable("Invalid Channel");
1156     case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break;
1157     case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break;
1158     case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break;
1159     case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break;
1160   }
1161   MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1162                                                AMDGPU::AR_X, OffsetReg);
1163   setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1164 
1165   MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1166                                       AddrReg, ValueReg)
1167                                       .addReg(AMDGPU::AR_X,
1168                                            RegState::Implicit | RegState::Kill);
1169   setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1);
1170   return Mov;
1171 }
1172 
1173 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1174                                        MachineBasicBlock::iterator I,
1175                                        unsigned ValueReg, unsigned Address,
1176                                        unsigned OffsetReg) const {
1177   return buildIndirectRead(MBB, I, ValueReg, Address, OffsetReg, 0);
1178 }
1179 
1180 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1181                                        MachineBasicBlock::iterator I,
1182                                        unsigned ValueReg, unsigned Address,
1183                                        unsigned OffsetReg,
1184                                        unsigned AddrChan) const {
1185   unsigned AddrReg;
1186   switch (AddrChan) {
1187     default: llvm_unreachable("Invalid Channel");
1188     case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break;
1189     case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break;
1190     case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break;
1191     case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break;
1192   }
1193   MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1194                                                        AMDGPU::AR_X,
1195                                                        OffsetReg);
1196   setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1197   MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1198                                       ValueReg,
1199                                       AddrReg)
1200                                       .addReg(AMDGPU::AR_X,
1201                                            RegState::Implicit | RegState::Kill);
1202   setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1);
1203 
1204   return Mov;
1205 }
1206 
1207 unsigned R600InstrInfo::getMaxAlusPerClause() const {
1208   return 115;
1209 }
1210 
1211 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
1212                                                   MachineBasicBlock::iterator I,
1213                                                   unsigned Opcode,
1214                                                   unsigned DstReg,
1215                                                   unsigned Src0Reg,
1216                                                   unsigned Src1Reg) const {
1217   MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
1218     DstReg);           // $dst
1219 
1220   if (Src1Reg) {
1221     MIB.addImm(0)     // $update_exec_mask
1222        .addImm(0);    // $update_predicate
1223   }
1224   MIB.addImm(1)        // $write
1225      .addImm(0)        // $omod
1226      .addImm(0)        // $dst_rel
1227      .addImm(0)        // $dst_clamp
1228      .addReg(Src0Reg)  // $src0
1229      .addImm(0)        // $src0_neg
1230      .addImm(0)        // $src0_rel
1231      .addImm(0)        // $src0_abs
1232      .addImm(-1);       // $src0_sel
1233 
1234   if (Src1Reg) {
1235     MIB.addReg(Src1Reg) // $src1
1236        .addImm(0)       // $src1_neg
1237        .addImm(0)       // $src1_rel
1238        .addImm(0)       // $src1_abs
1239        .addImm(-1);      // $src1_sel
1240   }
1241 
1242   //XXX: The r600g finalizer expects this to be 1, once we've moved the
1243   //scheduling to the backend, we can change the default to 0.
1244   MIB.addImm(1)        // $last
1245       .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
1246       .addImm(0)         // $literal
1247       .addImm(0);        // $bank_swizzle
1248 
1249   return MIB;
1250 }
1251 
1252 #define OPERAND_CASE(Label) \
1253   case Label: { \
1254     static const unsigned Ops[] = \
1255     { \
1256       Label##_X, \
1257       Label##_Y, \
1258       Label##_Z, \
1259       Label##_W \
1260     }; \
1261     return Ops[Slot]; \
1262   }
1263 
1264 static unsigned getSlotedOps(unsigned  Op, unsigned Slot) {
1265   switch (Op) {
1266   OPERAND_CASE(AMDGPU::OpName::update_exec_mask)
1267   OPERAND_CASE(AMDGPU::OpName::update_pred)
1268   OPERAND_CASE(AMDGPU::OpName::write)
1269   OPERAND_CASE(AMDGPU::OpName::omod)
1270   OPERAND_CASE(AMDGPU::OpName::dst_rel)
1271   OPERAND_CASE(AMDGPU::OpName::clamp)
1272   OPERAND_CASE(AMDGPU::OpName::src0)
1273   OPERAND_CASE(AMDGPU::OpName::src0_neg)
1274   OPERAND_CASE(AMDGPU::OpName::src0_rel)
1275   OPERAND_CASE(AMDGPU::OpName::src0_abs)
1276   OPERAND_CASE(AMDGPU::OpName::src0_sel)
1277   OPERAND_CASE(AMDGPU::OpName::src1)
1278   OPERAND_CASE(AMDGPU::OpName::src1_neg)
1279   OPERAND_CASE(AMDGPU::OpName::src1_rel)
1280   OPERAND_CASE(AMDGPU::OpName::src1_abs)
1281   OPERAND_CASE(AMDGPU::OpName::src1_sel)
1282   OPERAND_CASE(AMDGPU::OpName::pred_sel)
1283   default:
1284     llvm_unreachable("Wrong Operand");
1285   }
1286 }
1287 
1288 #undef OPERAND_CASE
1289 
1290 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
1291     MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
1292     const {
1293   assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
1294   unsigned Opcode;
1295   if (ST.getGeneration() <= AMDGPUSubtarget::R700)
1296     Opcode = AMDGPU::DOT4_r600;
1297   else
1298     Opcode = AMDGPU::DOT4_eg;
1299   MachineBasicBlock::iterator I = MI;
1300   MachineOperand &Src0 = MI->getOperand(
1301       getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot)));
1302   MachineOperand &Src1 = MI->getOperand(
1303       getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot)));
1304   MachineInstr *MIB = buildDefaultInstruction(
1305       MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
1306   static const unsigned  Operands[14] = {
1307     AMDGPU::OpName::update_exec_mask,
1308     AMDGPU::OpName::update_pred,
1309     AMDGPU::OpName::write,
1310     AMDGPU::OpName::omod,
1311     AMDGPU::OpName::dst_rel,
1312     AMDGPU::OpName::clamp,
1313     AMDGPU::OpName::src0_neg,
1314     AMDGPU::OpName::src0_rel,
1315     AMDGPU::OpName::src0_abs,
1316     AMDGPU::OpName::src0_sel,
1317     AMDGPU::OpName::src1_neg,
1318     AMDGPU::OpName::src1_rel,
1319     AMDGPU::OpName::src1_abs,
1320     AMDGPU::OpName::src1_sel,
1321   };
1322 
1323   MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
1324       getSlotedOps(AMDGPU::OpName::pred_sel, Slot)));
1325   MIB->getOperand(getOperandIdx(Opcode, AMDGPU::OpName::pred_sel))
1326       .setReg(MO.getReg());
1327 
1328   for (unsigned i = 0; i < 14; i++) {
1329     MachineOperand &MO = MI->getOperand(
1330         getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
1331     assert (MO.isImm());
1332     setImmOperand(MIB, Operands[i], MO.getImm());
1333   }
1334   MIB->getOperand(20).setImm(0);
1335   return MIB;
1336 }
1337 
1338 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
1339                                          MachineBasicBlock::iterator I,
1340                                          unsigned DstReg,
1341                                          uint64_t Imm) const {
1342   MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
1343                                                   AMDGPU::ALU_LITERAL_X);
1344   setImmOperand(MovImm, AMDGPU::OpName::literal, Imm);
1345   return MovImm;
1346 }
1347 
1348 MachineInstr *R600InstrInfo::buildMovInstr(MachineBasicBlock *MBB,
1349                                        MachineBasicBlock::iterator I,
1350                                        unsigned DstReg, unsigned SrcReg) const {
1351   return buildDefaultInstruction(*MBB, I, AMDGPU::MOV, DstReg, SrcReg);
1352 }
1353 
1354 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
1355   return getOperandIdx(MI.getOpcode(), Op);
1356 }
1357 
1358 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const {
1359   return AMDGPU::getNamedOperandIdx(Opcode, Op);
1360 }
1361 
1362 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op,
1363                                   int64_t Imm) const {
1364   int Idx = getOperandIdx(*MI, Op);
1365   assert(Idx != -1 && "Operand not supported for this instruction.");
1366   assert(MI->getOperand(Idx).isImm());
1367   MI->getOperand(Idx).setImm(Imm);
1368 }
1369 
1370 //===----------------------------------------------------------------------===//
1371 // Instruction flag getters/setters
1372 //===----------------------------------------------------------------------===//
1373 
1374 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1375   return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1376 }
1377 
1378 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1379                                          unsigned Flag) const {
1380   unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1381   int FlagIndex = 0;
1382   if (Flag != 0) {
1383     // If we pass something other than the default value of Flag to this
1384     // function, it means we are want to set a flag on an instruction
1385     // that uses native encoding.
1386     assert(HAS_NATIVE_OPERANDS(TargetFlags));
1387     bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1388     switch (Flag) {
1389     case MO_FLAG_CLAMP:
1390       FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp);
1391       break;
1392     case MO_FLAG_MASK:
1393       FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write);
1394       break;
1395     case MO_FLAG_NOT_LAST:
1396     case MO_FLAG_LAST:
1397       FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last);
1398       break;
1399     case MO_FLAG_NEG:
1400       switch (SrcIdx) {
1401       case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break;
1402       case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break;
1403       case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break;
1404       }
1405       break;
1406 
1407     case MO_FLAG_ABS:
1408       assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1409                        "instructions.");
1410       (void)IsOP3;
1411       switch (SrcIdx) {
1412       case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break;
1413       case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break;
1414       }
1415       break;
1416 
1417     default:
1418       FlagIndex = -1;
1419       break;
1420     }
1421     assert(FlagIndex != -1 && "Flag not supported for this instruction");
1422   } else {
1423       FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1424       assert(FlagIndex != 0 &&
1425          "Instruction flags not supported for this instruction");
1426   }
1427 
1428   MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1429   assert(FlagOp.isImm());
1430   return FlagOp;
1431 }
1432 
1433 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1434                             unsigned Flag) const {
1435   unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1436   if (Flag == 0) {
1437     return;
1438   }
1439   if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1440     MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1441     if (Flag == MO_FLAG_NOT_LAST) {
1442       clearFlag(MI, Operand, MO_FLAG_LAST);
1443     } else if (Flag == MO_FLAG_MASK) {
1444       clearFlag(MI, Operand, Flag);
1445     } else {
1446       FlagOp.setImm(1);
1447     }
1448   } else {
1449       MachineOperand &FlagOp = getFlagOp(MI, Operand);
1450       FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1451   }
1452 }
1453 
1454 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1455                               unsigned Flag) const {
1456   unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1457   if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1458     MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1459     FlagOp.setImm(0);
1460   } else {
1461     MachineOperand &FlagOp = getFlagOp(MI);
1462     unsigned InstFlags = FlagOp.getImm();
1463     InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1464     FlagOp.setImm(InstFlags);
1465   }
1466 }
1467 
1468 bool R600InstrInfo::isRegisterStore(const MachineInstr &MI) const {
1469   return get(MI.getOpcode()).TSFlags & AMDGPU_FLAG_REGISTER_STORE;
1470 }
1471 
1472 bool R600InstrInfo::isRegisterLoad(const MachineInstr &MI) const {
1473   return get(MI.getOpcode()).TSFlags & AMDGPU_FLAG_REGISTER_LOAD;
1474 }
1475