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 
R600InstrInfo(const AMDGPUSubtarget & st)31 R600InstrInfo::R600InstrInfo(const AMDGPUSubtarget &st)
32     : AMDGPUInstrInfo(st), RI() {}
33 
getRegisterInfo() const34 const R600RegisterInfo &R600InstrInfo::getRegisterInfo() const {
35   return RI;
36 }
37 
isTrig(const MachineInstr & MI) const38 bool R600InstrInfo::isTrig(const MachineInstr &MI) const {
39   return get(MI.getOpcode()).TSFlags & R600_InstFlag::TRIG;
40 }
41 
isVector(const MachineInstr & MI) const42 bool R600InstrInfo::isVector(const MachineInstr &MI) const {
43   return get(MI.getOpcode()).TSFlags & R600_InstFlag::VECTOR;
44 }
45 
46 void
copyPhysReg(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,DebugLoc DL,unsigned DestReg,unsigned SrcReg,bool KillSrc) const47 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.
isLegalToSplitMBBAt(MachineBasicBlock & MBB,MachineBasicBlock::iterator MBBI) const82 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 
isMov(unsigned Opcode) const93 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.
isPlaceHolderOpcode(unsigned Opcode) const108 bool R600InstrInfo::isPlaceHolderOpcode(unsigned Opcode) const {
109   switch (Opcode) {
110   default: return false;
111   case AMDGPU::RETURN:
112     return true;
113   }
114 }
115 
isReductionOp(unsigned Opcode) const116 bool R600InstrInfo::isReductionOp(unsigned Opcode) const {
117   return false;
118 }
119 
isCubeOp(unsigned Opcode) const120 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 
isALUInstr(unsigned Opcode) const131 bool R600InstrInfo::isALUInstr(unsigned Opcode) const {
132   unsigned TargetFlags = get(Opcode).TSFlags;
133 
134   return (TargetFlags & R600_InstFlag::ALU_INST);
135 }
136 
hasInstrModifiers(unsigned Opcode) const137 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 
isLDSInstr(unsigned Opcode) const145 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 
isLDSNoRetInstr(unsigned Opcode) const153 bool R600InstrInfo::isLDSNoRetInstr(unsigned Opcode) const {
154   return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) == -1;
155 }
156 
isLDSRetInstr(unsigned Opcode) const157 bool R600InstrInfo::isLDSRetInstr(unsigned Opcode) const {
158   return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) != -1;
159 }
160 
canBeConsideredALU(const MachineInstr * MI) const161 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 
isTransOnly(unsigned Opcode) const179 bool R600InstrInfo::isTransOnly(unsigned Opcode) const {
180   if (ST.hasCaymanISA())
181     return false;
182   return (get(Opcode).getSchedClass() == AMDGPU::Sched::TransALU);
183 }
184 
isTransOnly(const MachineInstr * MI) const185 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
186   return isTransOnly(MI->getOpcode());
187 }
188 
isVectorOnly(unsigned Opcode) const189 bool R600InstrInfo::isVectorOnly(unsigned Opcode) const {
190   return (get(Opcode).getSchedClass() == AMDGPU::Sched::VecALU);
191 }
192 
isVectorOnly(const MachineInstr * MI) const193 bool R600InstrInfo::isVectorOnly(const MachineInstr *MI) const {
194   return isVectorOnly(MI->getOpcode());
195 }
196 
isExport(unsigned Opcode) const197 bool R600InstrInfo::isExport(unsigned Opcode) const {
198   return (get(Opcode).TSFlags & R600_InstFlag::IS_EXPORT);
199 }
200 
usesVertexCache(unsigned Opcode) const201 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
202   return ST.hasVertexCache() && IS_VTX(get(Opcode));
203 }
204 
usesVertexCache(const MachineInstr * MI) const205 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
206   const MachineFunction *MF = MI->getParent()->getParent();
207   const R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
208   return MFI->getShaderType() != ShaderType::COMPUTE &&
209     usesVertexCache(MI->getOpcode());
210 }
211 
usesTextureCache(unsigned Opcode) const212 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
213   return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode));
214 }
215 
usesTextureCache(const MachineInstr * MI) const216 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
217   const MachineFunction *MF = MI->getParent()->getParent();
218   const R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
219   return (MFI->getShaderType() == ShaderType::COMPUTE &&
220           usesVertexCache(MI->getOpcode())) ||
221     usesTextureCache(MI->getOpcode());
222 }
223 
mustBeLastInClause(unsigned Opcode) const224 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
225   switch (Opcode) {
226   case AMDGPU::KILLGT:
227   case AMDGPU::GROUP_BARRIER:
228     return true;
229   default:
230     return false;
231   }
232 }
233 
usesAddressRegister(MachineInstr * MI) const234 bool R600InstrInfo::usesAddressRegister(MachineInstr *MI) const {
235   return  MI->findRegisterUseOperandIdx(AMDGPU::AR_X) != -1;
236 }
237 
definesAddressRegister(MachineInstr * MI) const238 bool R600InstrInfo::definesAddressRegister(MachineInstr *MI) const {
239   return MI->findRegisterDefOperandIdx(AMDGPU::AR_X) != -1;
240 }
241 
readsLDSSrcReg(const MachineInstr * MI) const242 bool R600InstrInfo::readsLDSSrcReg(const MachineInstr *MI) const {
243   if (!isALUInstr(MI->getOpcode())) {
244     return false;
245   }
246   for (MachineInstr::const_mop_iterator I = MI->operands_begin(),
247                                         E = MI->operands_end(); I != E; ++I) {
248     if (!I->isReg() || !I->isUse() ||
249         TargetRegisterInfo::isVirtualRegister(I->getReg()))
250       continue;
251 
252     if (AMDGPU::R600_LDS_SRC_REGRegClass.contains(I->getReg()))
253       return true;
254   }
255   return false;
256 }
257 
getSrcIdx(unsigned Opcode,unsigned SrcNum) const258 int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const {
259   static const unsigned OpTable[] = {
260     AMDGPU::OpName::src0,
261     AMDGPU::OpName::src1,
262     AMDGPU::OpName::src2
263   };
264 
265   assert (SrcNum < 3);
266   return getOperandIdx(Opcode, OpTable[SrcNum]);
267 }
268 
getSelIdx(unsigned Opcode,unsigned SrcIdx) const269 int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const {
270   static const unsigned SrcSelTable[][2] = {
271     {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
272     {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
273     {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
274     {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
275     {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
276     {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
277     {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
278     {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
279     {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
280     {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
281     {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}
282   };
283 
284   for (const auto &Row : SrcSelTable) {
285     if (getOperandIdx(Opcode, Row[0]) == (int)SrcIdx) {
286       return getOperandIdx(Opcode, Row[1]);
287     }
288   }
289   return -1;
290 }
291 
292 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
getSrcs(MachineInstr * MI) const293 R600InstrInfo::getSrcs(MachineInstr *MI) const {
294   SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
295 
296   if (MI->getOpcode() == AMDGPU::DOT_4) {
297     static const unsigned OpTable[8][2] = {
298       {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
299       {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
300       {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
301       {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
302       {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
303       {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
304       {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
305       {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W},
306     };
307 
308     for (unsigned j = 0; j < 8; j++) {
309       MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
310                                                         OpTable[j][0]));
311       unsigned Reg = MO.getReg();
312       if (Reg == AMDGPU::ALU_CONST) {
313         unsigned Sel = MI->getOperand(getOperandIdx(MI->getOpcode(),
314                                                     OpTable[j][1])).getImm();
315         Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
316         continue;
317       }
318 
319     }
320     return Result;
321   }
322 
323   static const unsigned OpTable[3][2] = {
324     {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
325     {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
326     {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
327   };
328 
329   for (unsigned j = 0; j < 3; j++) {
330     int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]);
331     if (SrcIdx < 0)
332       break;
333     MachineOperand &MO = MI->getOperand(SrcIdx);
334     unsigned Reg = MI->getOperand(SrcIdx).getReg();
335     if (Reg == AMDGPU::ALU_CONST) {
336       unsigned Sel = MI->getOperand(
337           getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm();
338       Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
339       continue;
340     }
341     if (Reg == AMDGPU::ALU_LITERAL_X) {
342       unsigned Imm = MI->getOperand(
343           getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal)).getImm();
344       Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Imm));
345       continue;
346     }
347     Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, 0));
348   }
349   return Result;
350 }
351 
352 std::vector<std::pair<int, unsigned> >
ExtractSrcs(MachineInstr * MI,const DenseMap<unsigned,unsigned> & PV,unsigned & ConstCount) const353 R600InstrInfo::ExtractSrcs(MachineInstr *MI,
354                            const DenseMap<unsigned, unsigned> &PV,
355                            unsigned &ConstCount) const {
356   ConstCount = 0;
357   ArrayRef<std::pair<MachineOperand *, int64_t>> Srcs = getSrcs(MI);
358   const std::pair<int, unsigned> DummyPair(-1, 0);
359   std::vector<std::pair<int, unsigned> > Result;
360   unsigned i = 0;
361   for (unsigned n = Srcs.size(); i < n; ++i) {
362     unsigned Reg = Srcs[i].first->getReg();
363     unsigned Index = RI.getEncodingValue(Reg) & 0xff;
364     if (Reg == AMDGPU::OQAP) {
365       Result.push_back(std::pair<int, unsigned>(Index, 0));
366     }
367     if (PV.find(Reg) != PV.end()) {
368       // 255 is used to tells its a PS/PV reg
369       Result.push_back(std::pair<int, unsigned>(255, 0));
370       continue;
371     }
372     if (Index > 127) {
373       ConstCount++;
374       Result.push_back(DummyPair);
375       continue;
376     }
377     unsigned Chan = RI.getHWRegChan(Reg);
378     Result.push_back(std::pair<int, unsigned>(Index, Chan));
379   }
380   for (; i < 3; ++i)
381     Result.push_back(DummyPair);
382   return Result;
383 }
384 
385 static std::vector<std::pair<int, unsigned> >
Swizzle(std::vector<std::pair<int,unsigned>> Src,R600InstrInfo::BankSwizzle Swz)386 Swizzle(std::vector<std::pair<int, unsigned> > Src,
387         R600InstrInfo::BankSwizzle Swz) {
388   if (Src[0] == Src[1])
389     Src[1].first = -1;
390   switch (Swz) {
391   case R600InstrInfo::ALU_VEC_012_SCL_210:
392     break;
393   case R600InstrInfo::ALU_VEC_021_SCL_122:
394     std::swap(Src[1], Src[2]);
395     break;
396   case R600InstrInfo::ALU_VEC_102_SCL_221:
397     std::swap(Src[0], Src[1]);
398     break;
399   case R600InstrInfo::ALU_VEC_120_SCL_212:
400     std::swap(Src[0], Src[1]);
401     std::swap(Src[0], Src[2]);
402     break;
403   case R600InstrInfo::ALU_VEC_201:
404     std::swap(Src[0], Src[2]);
405     std::swap(Src[0], Src[1]);
406     break;
407   case R600InstrInfo::ALU_VEC_210:
408     std::swap(Src[0], Src[2]);
409     break;
410   }
411   return Src;
412 }
413 
414 static unsigned
getTransSwizzle(R600InstrInfo::BankSwizzle Swz,unsigned Op)415 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) {
416   switch (Swz) {
417   case R600InstrInfo::ALU_VEC_012_SCL_210: {
418     unsigned Cycles[3] = { 2, 1, 0};
419     return Cycles[Op];
420   }
421   case R600InstrInfo::ALU_VEC_021_SCL_122: {
422     unsigned Cycles[3] = { 1, 2, 2};
423     return Cycles[Op];
424   }
425   case R600InstrInfo::ALU_VEC_120_SCL_212: {
426     unsigned Cycles[3] = { 2, 1, 2};
427     return Cycles[Op];
428   }
429   case R600InstrInfo::ALU_VEC_102_SCL_221: {
430     unsigned Cycles[3] = { 2, 2, 1};
431     return Cycles[Op];
432   }
433   default:
434     llvm_unreachable("Wrong Swizzle for Trans Slot");
435     return 0;
436   }
437 }
438 
439 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed
440 /// in the same Instruction Group while meeting read port limitations given a
441 /// Swz swizzle sequence.
isLegalUpTo(const std::vector<std::vector<std::pair<int,unsigned>>> & IGSrcs,const std::vector<R600InstrInfo::BankSwizzle> & Swz,const std::vector<std::pair<int,unsigned>> & TransSrcs,R600InstrInfo::BankSwizzle TransSwz) const442 unsigned  R600InstrInfo::isLegalUpTo(
443     const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
444     const std::vector<R600InstrInfo::BankSwizzle> &Swz,
445     const std::vector<std::pair<int, unsigned> > &TransSrcs,
446     R600InstrInfo::BankSwizzle TransSwz) const {
447   int Vector[4][3];
448   memset(Vector, -1, sizeof(Vector));
449   for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) {
450     const std::vector<std::pair<int, unsigned> > &Srcs =
451         Swizzle(IGSrcs[i], Swz[i]);
452     for (unsigned j = 0; j < 3; j++) {
453       const std::pair<int, unsigned> &Src = Srcs[j];
454       if (Src.first < 0 || Src.first == 255)
455         continue;
456       if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) {
457         if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 &&
458             Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) {
459             // The value from output queue A (denoted by register OQAP) can
460             // only be fetched during the first cycle.
461             return false;
462         }
463         // OQAP does not count towards the normal read port restrictions
464         continue;
465       }
466       if (Vector[Src.second][j] < 0)
467         Vector[Src.second][j] = Src.first;
468       if (Vector[Src.second][j] != Src.first)
469         return i;
470     }
471   }
472   // Now check Trans Alu
473   for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) {
474     const std::pair<int, unsigned> &Src = TransSrcs[i];
475     unsigned Cycle = getTransSwizzle(TransSwz, i);
476     if (Src.first < 0)
477       continue;
478     if (Src.first == 255)
479       continue;
480     if (Vector[Src.second][Cycle] < 0)
481       Vector[Src.second][Cycle] = Src.first;
482     if (Vector[Src.second][Cycle] != Src.first)
483       return IGSrcs.size() - 1;
484   }
485   return IGSrcs.size();
486 }
487 
488 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next
489 /// (in lexicographic term) swizzle sequence assuming that all swizzles after
490 /// Idx can be skipped
491 static bool
NextPossibleSolution(std::vector<R600InstrInfo::BankSwizzle> & SwzCandidate,unsigned Idx)492 NextPossibleSolution(
493     std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
494     unsigned Idx) {
495   assert(Idx < SwzCandidate.size());
496   int ResetIdx = Idx;
497   while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210)
498     ResetIdx --;
499   for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) {
500     SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210;
501   }
502   if (ResetIdx == -1)
503     return false;
504   int NextSwizzle = SwzCandidate[ResetIdx] + 1;
505   SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle;
506   return true;
507 }
508 
509 /// Enumerate all possible Swizzle sequence to find one that can meet all
510 /// read port requirements.
FindSwizzleForVectorSlot(const std::vector<std::vector<std::pair<int,unsigned>>> & IGSrcs,std::vector<R600InstrInfo::BankSwizzle> & SwzCandidate,const std::vector<std::pair<int,unsigned>> & TransSrcs,R600InstrInfo::BankSwizzle TransSwz) const511 bool R600InstrInfo::FindSwizzleForVectorSlot(
512     const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
513     std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
514     const std::vector<std::pair<int, unsigned> > &TransSrcs,
515     R600InstrInfo::BankSwizzle TransSwz) const {
516   unsigned ValidUpTo = 0;
517   do {
518     ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz);
519     if (ValidUpTo == IGSrcs.size())
520       return true;
521   } while (NextPossibleSolution(SwzCandidate, ValidUpTo));
522   return false;
523 }
524 
525 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read
526 /// a const, and can't read a gpr at cycle 1 if they read 2 const.
527 static bool
isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,const std::vector<std::pair<int,unsigned>> & TransOps,unsigned ConstCount)528 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
529                   const std::vector<std::pair<int, unsigned> > &TransOps,
530                   unsigned ConstCount) {
531   // TransALU can't read 3 constants
532   if (ConstCount > 2)
533     return false;
534   for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
535     const std::pair<int, unsigned> &Src = TransOps[i];
536     unsigned Cycle = getTransSwizzle(TransSwz, i);
537     if (Src.first < 0)
538       continue;
539     if (ConstCount > 0 && Cycle == 0)
540       return false;
541     if (ConstCount > 1 && Cycle == 1)
542       return false;
543   }
544   return true;
545 }
546 
547 bool
fitsReadPortLimitations(const std::vector<MachineInstr * > & IG,const DenseMap<unsigned,unsigned> & PV,std::vector<BankSwizzle> & ValidSwizzle,bool isLastAluTrans) const548 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
549                                        const DenseMap<unsigned, unsigned> &PV,
550                                        std::vector<BankSwizzle> &ValidSwizzle,
551                                        bool isLastAluTrans)
552     const {
553   //Todo : support shared src0 - src1 operand
554 
555   std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
556   ValidSwizzle.clear();
557   unsigned ConstCount;
558   BankSwizzle TransBS = ALU_VEC_012_SCL_210;
559   for (unsigned i = 0, e = IG.size(); i < e; ++i) {
560     IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount));
561     unsigned Op = getOperandIdx(IG[i]->getOpcode(),
562         AMDGPU::OpName::bank_swizzle);
563     ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
564         IG[i]->getOperand(Op).getImm());
565   }
566   std::vector<std::pair<int, unsigned> > TransOps;
567   if (!isLastAluTrans)
568     return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
569 
570   TransOps = std::move(IGSrcs.back());
571   IGSrcs.pop_back();
572   ValidSwizzle.pop_back();
573 
574   static const R600InstrInfo::BankSwizzle TransSwz[] = {
575     ALU_VEC_012_SCL_210,
576     ALU_VEC_021_SCL_122,
577     ALU_VEC_120_SCL_212,
578     ALU_VEC_102_SCL_221
579   };
580   for (unsigned i = 0; i < 4; i++) {
581     TransBS = TransSwz[i];
582     if (!isConstCompatible(TransBS, TransOps, ConstCount))
583       continue;
584     bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps,
585         TransBS);
586     if (Result) {
587       ValidSwizzle.push_back(TransBS);
588       return true;
589     }
590   }
591 
592   return false;
593 }
594 
595 
596 bool
fitsConstReadLimitations(const std::vector<unsigned> & Consts) const597 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
598     const {
599   assert (Consts.size() <= 12 && "Too many operands in instructions group");
600   unsigned Pair1 = 0, Pair2 = 0;
601   for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
602     unsigned ReadConstHalf = Consts[i] & 2;
603     unsigned ReadConstIndex = Consts[i] & (~3);
604     unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
605     if (!Pair1) {
606       Pair1 = ReadHalfConst;
607       continue;
608     }
609     if (Pair1 == ReadHalfConst)
610       continue;
611     if (!Pair2) {
612       Pair2 = ReadHalfConst;
613       continue;
614     }
615     if (Pair2 != ReadHalfConst)
616       return false;
617   }
618   return true;
619 }
620 
621 bool
fitsConstReadLimitations(const std::vector<MachineInstr * > & MIs) const622 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
623     const {
624   std::vector<unsigned> Consts;
625   SmallSet<int64_t, 4> Literals;
626   for (unsigned i = 0, n = MIs.size(); i < n; i++) {
627     MachineInstr *MI = MIs[i];
628     if (!isALUInstr(MI->getOpcode()))
629       continue;
630 
631     ArrayRef<std::pair<MachineOperand *, int64_t>> Srcs = getSrcs(MI);
632 
633     for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
634       std::pair<MachineOperand *, unsigned> Src = Srcs[j];
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 *
CreateTargetScheduleState(const TargetSubtargetInfo & STI) const653 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
isPredicateSetter(unsigned Opcode)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 *
findFirstPredicateSetterFrom(MachineBasicBlock & MBB,MachineBasicBlock::iterator I)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
isJump(unsigned Opcode)682 bool isJump(unsigned Opcode) {
683   return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
684 }
685 
isBranch(unsigned Opcode)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
AnalyzeBranch(MachineBasicBlock & MBB,MachineBasicBlock * & TBB,MachineBasicBlock * & FBB,SmallVectorImpl<MachineOperand> & Cond,bool AllowModify) const692 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
FindLastAluClause(MachineBasicBlock & MBB)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
InsertBranch(MachineBasicBlock & MBB,MachineBasicBlock * TBB,MachineBasicBlock * FBB,ArrayRef<MachineOperand> Cond,DebugLoc DL) const776 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
RemoveBranch(MachineBasicBlock & MBB) const822 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
isPredicated(const MachineInstr * MI) const880 R600InstrInfo::isPredicated(const MachineInstr *MI) const {
881   int idx = MI->findFirstPredOperandIdx();
882   if (idx < 0)
883     return false;
884 
885   unsigned Reg = MI->getOperand(idx).getReg();
886   switch (Reg) {
887   default: return false;
888   case AMDGPU::PRED_SEL_ONE:
889   case AMDGPU::PRED_SEL_ZERO:
890   case AMDGPU::PREDICATE_BIT:
891     return true;
892   }
893 }
894 
895 bool
isPredicable(MachineInstr * MI) const896 R600InstrInfo::isPredicable(MachineInstr *MI) const {
897   // XXX: KILL* instructions can be predicated, but they must be the last
898   // instruction in a clause, so this means any instructions after them cannot
899   // be predicated.  Until we have proper support for instruction clauses in the
900   // backend, we will mark KILL* instructions as unpredicable.
901 
902   if (MI->getOpcode() == AMDGPU::KILLGT) {
903     return false;
904   } else if (MI->getOpcode() == AMDGPU::CF_ALU) {
905     // If the clause start in the middle of MBB then the MBB has more
906     // than a single clause, unable to predicate several clauses.
907     if (MI->getParent()->begin() != MachineBasicBlock::iterator(MI))
908       return false;
909     // TODO: We don't support KC merging atm
910     if (MI->getOperand(3).getImm() != 0 || MI->getOperand(4).getImm() != 0)
911       return false;
912     return true;
913   } else if (isVector(*MI)) {
914     return false;
915   } else {
916     return AMDGPUInstrInfo::isPredicable(MI);
917   }
918 }
919 
920 
921 bool
isProfitableToIfCvt(MachineBasicBlock & MBB,unsigned NumCyles,unsigned ExtraPredCycles,BranchProbability Probability) const922 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
923                                    unsigned NumCyles,
924                                    unsigned ExtraPredCycles,
925                                    BranchProbability Probability) const{
926   return true;
927 }
928 
929 bool
isProfitableToIfCvt(MachineBasicBlock & TMBB,unsigned NumTCycles,unsigned ExtraTCycles,MachineBasicBlock & FMBB,unsigned NumFCycles,unsigned ExtraFCycles,BranchProbability Probability) const930 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
931                                    unsigned NumTCycles,
932                                    unsigned ExtraTCycles,
933                                    MachineBasicBlock &FMBB,
934                                    unsigned NumFCycles,
935                                    unsigned ExtraFCycles,
936                                    BranchProbability Probability) const {
937   return true;
938 }
939 
940 bool
isProfitableToDupForIfCvt(MachineBasicBlock & MBB,unsigned NumCyles,BranchProbability Probability) const941 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
942                                          unsigned NumCyles,
943                                          BranchProbability Probability)
944                                          const {
945   return true;
946 }
947 
948 bool
isProfitableToUnpredicate(MachineBasicBlock & TMBB,MachineBasicBlock & FMBB) const949 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
950                                          MachineBasicBlock &FMBB) const {
951   return false;
952 }
953 
954 
955 bool
ReverseBranchCondition(SmallVectorImpl<MachineOperand> & Cond) const956 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
957   MachineOperand &MO = Cond[1];
958   switch (MO.getImm()) {
959   case OPCODE_IS_ZERO_INT:
960     MO.setImm(OPCODE_IS_NOT_ZERO_INT);
961     break;
962   case OPCODE_IS_NOT_ZERO_INT:
963     MO.setImm(OPCODE_IS_ZERO_INT);
964     break;
965   case OPCODE_IS_ZERO:
966     MO.setImm(OPCODE_IS_NOT_ZERO);
967     break;
968   case OPCODE_IS_NOT_ZERO:
969     MO.setImm(OPCODE_IS_ZERO);
970     break;
971   default:
972     return true;
973   }
974 
975   MachineOperand &MO2 = Cond[2];
976   switch (MO2.getReg()) {
977   case AMDGPU::PRED_SEL_ZERO:
978     MO2.setReg(AMDGPU::PRED_SEL_ONE);
979     break;
980   case AMDGPU::PRED_SEL_ONE:
981     MO2.setReg(AMDGPU::PRED_SEL_ZERO);
982     break;
983   default:
984     return true;
985   }
986   return false;
987 }
988 
989 bool
DefinesPredicate(MachineInstr * MI,std::vector<MachineOperand> & Pred) const990 R600InstrInfo::DefinesPredicate(MachineInstr *MI,
991                                 std::vector<MachineOperand> &Pred) const {
992   return isPredicateSetter(MI->getOpcode());
993 }
994 
995 
996 bool
SubsumesPredicate(ArrayRef<MachineOperand> Pred1,ArrayRef<MachineOperand> Pred2) const997 R600InstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
998                                  ArrayRef<MachineOperand> Pred2) const {
999   return false;
1000 }
1001 
1002 
1003 bool
PredicateInstruction(MachineInstr * MI,ArrayRef<MachineOperand> Pred) const1004 R600InstrInfo::PredicateInstruction(MachineInstr *MI,
1005                                     ArrayRef<MachineOperand> Pred) const {
1006   int PIdx = MI->findFirstPredOperandIdx();
1007 
1008   if (MI->getOpcode() == AMDGPU::CF_ALU) {
1009     MI->getOperand(8).setImm(0);
1010     return true;
1011   }
1012 
1013   if (MI->getOpcode() == AMDGPU::DOT_4) {
1014     MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_X))
1015         .setReg(Pred[2].getReg());
1016     MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Y))
1017         .setReg(Pred[2].getReg());
1018     MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Z))
1019         .setReg(Pred[2].getReg());
1020     MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_W))
1021         .setReg(Pred[2].getReg());
1022     MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1023     MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
1024     return true;
1025   }
1026 
1027   if (PIdx != -1) {
1028     MachineOperand &PMO = MI->getOperand(PIdx);
1029     PMO.setReg(Pred[2].getReg());
1030     MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1031     MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
1032     return true;
1033   }
1034 
1035   return false;
1036 }
1037 
getPredicationCost(const MachineInstr *) const1038 unsigned int R600InstrInfo::getPredicationCost(const MachineInstr *) const {
1039   return 2;
1040 }
1041 
getInstrLatency(const InstrItineraryData * ItinData,const MachineInstr * MI,unsigned * PredCost) const1042 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1043                                             const MachineInstr *MI,
1044                                             unsigned *PredCost) const {
1045   if (PredCost)
1046     *PredCost = 2;
1047   return 2;
1048 }
1049 
expandPostRAPseudo(MachineBasicBlock::iterator MI) const1050 bool R600InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
1051 
1052   switch(MI->getOpcode()) {
1053   default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
1054   case AMDGPU::R600_EXTRACT_ELT_V2:
1055   case AMDGPU::R600_EXTRACT_ELT_V4:
1056     buildIndirectRead(MI->getParent(), MI, MI->getOperand(0).getReg(),
1057                       RI.getHWRegIndex(MI->getOperand(1).getReg()), //  Address
1058                       MI->getOperand(2).getReg(),
1059                       RI.getHWRegChan(MI->getOperand(1).getReg()));
1060     break;
1061   case AMDGPU::R600_INSERT_ELT_V2:
1062   case AMDGPU::R600_INSERT_ELT_V4:
1063     buildIndirectWrite(MI->getParent(), MI, MI->getOperand(2).getReg(), // Value
1064                        RI.getHWRegIndex(MI->getOperand(1).getReg()),  // Address
1065                        MI->getOperand(3).getReg(),                    // Offset
1066                        RI.getHWRegChan(MI->getOperand(1).getReg()));  // Channel
1067     break;
1068   }
1069   MI->eraseFromParent();
1070   return true;
1071 }
1072 
reserveIndirectRegisters(BitVector & Reserved,const MachineFunction & MF) const1073 void  R600InstrInfo::reserveIndirectRegisters(BitVector &Reserved,
1074                                              const MachineFunction &MF) const {
1075   const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
1076       MF.getSubtarget().getFrameLowering());
1077 
1078   unsigned StackWidth = TFL->getStackWidth(MF);
1079   int End = getIndirectIndexEnd(MF);
1080 
1081   if (End == -1)
1082     return;
1083 
1084   for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
1085     unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
1086     Reserved.set(SuperReg);
1087     for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
1088       unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
1089       Reserved.set(Reg);
1090     }
1091   }
1092 }
1093 
calculateIndirectAddress(unsigned RegIndex,unsigned Channel) const1094 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
1095                                                  unsigned Channel) const {
1096   // XXX: Remove when we support a stack width > 2
1097   assert(Channel == 0);
1098   return RegIndex;
1099 }
1100 
getIndirectAddrRegClass() const1101 const TargetRegisterClass *R600InstrInfo::getIndirectAddrRegClass() const {
1102   return &AMDGPU::R600_TReg32_XRegClass;
1103 }
1104 
buildIndirectWrite(MachineBasicBlock * MBB,MachineBasicBlock::iterator I,unsigned ValueReg,unsigned Address,unsigned OffsetReg) const1105 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1106                                        MachineBasicBlock::iterator I,
1107                                        unsigned ValueReg, unsigned Address,
1108                                        unsigned OffsetReg) const {
1109   return buildIndirectWrite(MBB, I, ValueReg, Address, OffsetReg, 0);
1110 }
1111 
buildIndirectWrite(MachineBasicBlock * MBB,MachineBasicBlock::iterator I,unsigned ValueReg,unsigned Address,unsigned OffsetReg,unsigned AddrChan) const1112 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1113                                        MachineBasicBlock::iterator I,
1114                                        unsigned ValueReg, unsigned Address,
1115                                        unsigned OffsetReg,
1116                                        unsigned AddrChan) const {
1117   unsigned AddrReg;
1118   switch (AddrChan) {
1119     default: llvm_unreachable("Invalid Channel");
1120     case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break;
1121     case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break;
1122     case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break;
1123     case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break;
1124   }
1125   MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1126                                                AMDGPU::AR_X, OffsetReg);
1127   setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1128 
1129   MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1130                                       AddrReg, ValueReg)
1131                                       .addReg(AMDGPU::AR_X,
1132                                            RegState::Implicit | RegState::Kill);
1133   setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1);
1134   return Mov;
1135 }
1136 
buildIndirectRead(MachineBasicBlock * MBB,MachineBasicBlock::iterator I,unsigned ValueReg,unsigned Address,unsigned OffsetReg) const1137 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1138                                        MachineBasicBlock::iterator I,
1139                                        unsigned ValueReg, unsigned Address,
1140                                        unsigned OffsetReg) const {
1141   return buildIndirectRead(MBB, I, ValueReg, Address, OffsetReg, 0);
1142 }
1143 
buildIndirectRead(MachineBasicBlock * MBB,MachineBasicBlock::iterator I,unsigned ValueReg,unsigned Address,unsigned OffsetReg,unsigned AddrChan) const1144 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1145                                        MachineBasicBlock::iterator I,
1146                                        unsigned ValueReg, unsigned Address,
1147                                        unsigned OffsetReg,
1148                                        unsigned AddrChan) const {
1149   unsigned AddrReg;
1150   switch (AddrChan) {
1151     default: llvm_unreachable("Invalid Channel");
1152     case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break;
1153     case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break;
1154     case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break;
1155     case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break;
1156   }
1157   MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1158                                                        AMDGPU::AR_X,
1159                                                        OffsetReg);
1160   setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1161   MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1162                                       ValueReg,
1163                                       AddrReg)
1164                                       .addReg(AMDGPU::AR_X,
1165                                            RegState::Implicit | RegState::Kill);
1166   setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1);
1167 
1168   return Mov;
1169 }
1170 
getMaxAlusPerClause() const1171 unsigned R600InstrInfo::getMaxAlusPerClause() const {
1172   return 115;
1173 }
1174 
buildDefaultInstruction(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,unsigned Opcode,unsigned DstReg,unsigned Src0Reg,unsigned Src1Reg) const1175 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
1176                                                   MachineBasicBlock::iterator I,
1177                                                   unsigned Opcode,
1178                                                   unsigned DstReg,
1179                                                   unsigned Src0Reg,
1180                                                   unsigned Src1Reg) const {
1181   MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
1182     DstReg);           // $dst
1183 
1184   if (Src1Reg) {
1185     MIB.addImm(0)     // $update_exec_mask
1186        .addImm(0);    // $update_predicate
1187   }
1188   MIB.addImm(1)        // $write
1189      .addImm(0)        // $omod
1190      .addImm(0)        // $dst_rel
1191      .addImm(0)        // $dst_clamp
1192      .addReg(Src0Reg)  // $src0
1193      .addImm(0)        // $src0_neg
1194      .addImm(0)        // $src0_rel
1195      .addImm(0)        // $src0_abs
1196      .addImm(-1);       // $src0_sel
1197 
1198   if (Src1Reg) {
1199     MIB.addReg(Src1Reg) // $src1
1200        .addImm(0)       // $src1_neg
1201        .addImm(0)       // $src1_rel
1202        .addImm(0)       // $src1_abs
1203        .addImm(-1);      // $src1_sel
1204   }
1205 
1206   //XXX: The r600g finalizer expects this to be 1, once we've moved the
1207   //scheduling to the backend, we can change the default to 0.
1208   MIB.addImm(1)        // $last
1209       .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
1210       .addImm(0)         // $literal
1211       .addImm(0);        // $bank_swizzle
1212 
1213   return MIB;
1214 }
1215 
1216 #define OPERAND_CASE(Label) \
1217   case Label: { \
1218     static const unsigned Ops[] = \
1219     { \
1220       Label##_X, \
1221       Label##_Y, \
1222       Label##_Z, \
1223       Label##_W \
1224     }; \
1225     return Ops[Slot]; \
1226   }
1227 
getSlotedOps(unsigned Op,unsigned Slot)1228 static unsigned getSlotedOps(unsigned  Op, unsigned Slot) {
1229   switch (Op) {
1230   OPERAND_CASE(AMDGPU::OpName::update_exec_mask)
1231   OPERAND_CASE(AMDGPU::OpName::update_pred)
1232   OPERAND_CASE(AMDGPU::OpName::write)
1233   OPERAND_CASE(AMDGPU::OpName::omod)
1234   OPERAND_CASE(AMDGPU::OpName::dst_rel)
1235   OPERAND_CASE(AMDGPU::OpName::clamp)
1236   OPERAND_CASE(AMDGPU::OpName::src0)
1237   OPERAND_CASE(AMDGPU::OpName::src0_neg)
1238   OPERAND_CASE(AMDGPU::OpName::src0_rel)
1239   OPERAND_CASE(AMDGPU::OpName::src0_abs)
1240   OPERAND_CASE(AMDGPU::OpName::src0_sel)
1241   OPERAND_CASE(AMDGPU::OpName::src1)
1242   OPERAND_CASE(AMDGPU::OpName::src1_neg)
1243   OPERAND_CASE(AMDGPU::OpName::src1_rel)
1244   OPERAND_CASE(AMDGPU::OpName::src1_abs)
1245   OPERAND_CASE(AMDGPU::OpName::src1_sel)
1246   OPERAND_CASE(AMDGPU::OpName::pred_sel)
1247   default:
1248     llvm_unreachable("Wrong Operand");
1249   }
1250 }
1251 
1252 #undef OPERAND_CASE
1253 
buildSlotOfVectorInstruction(MachineBasicBlock & MBB,MachineInstr * MI,unsigned Slot,unsigned DstReg) const1254 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
1255     MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
1256     const {
1257   assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
1258   unsigned Opcode;
1259   if (ST.getGeneration() <= AMDGPUSubtarget::R700)
1260     Opcode = AMDGPU::DOT4_r600;
1261   else
1262     Opcode = AMDGPU::DOT4_eg;
1263   MachineBasicBlock::iterator I = MI;
1264   MachineOperand &Src0 = MI->getOperand(
1265       getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot)));
1266   MachineOperand &Src1 = MI->getOperand(
1267       getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot)));
1268   MachineInstr *MIB = buildDefaultInstruction(
1269       MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
1270   static const unsigned  Operands[14] = {
1271     AMDGPU::OpName::update_exec_mask,
1272     AMDGPU::OpName::update_pred,
1273     AMDGPU::OpName::write,
1274     AMDGPU::OpName::omod,
1275     AMDGPU::OpName::dst_rel,
1276     AMDGPU::OpName::clamp,
1277     AMDGPU::OpName::src0_neg,
1278     AMDGPU::OpName::src0_rel,
1279     AMDGPU::OpName::src0_abs,
1280     AMDGPU::OpName::src0_sel,
1281     AMDGPU::OpName::src1_neg,
1282     AMDGPU::OpName::src1_rel,
1283     AMDGPU::OpName::src1_abs,
1284     AMDGPU::OpName::src1_sel,
1285   };
1286 
1287   MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
1288       getSlotedOps(AMDGPU::OpName::pred_sel, Slot)));
1289   MIB->getOperand(getOperandIdx(Opcode, AMDGPU::OpName::pred_sel))
1290       .setReg(MO.getReg());
1291 
1292   for (unsigned i = 0; i < 14; i++) {
1293     MachineOperand &MO = MI->getOperand(
1294         getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
1295     assert (MO.isImm());
1296     setImmOperand(MIB, Operands[i], MO.getImm());
1297   }
1298   MIB->getOperand(20).setImm(0);
1299   return MIB;
1300 }
1301 
buildMovImm(MachineBasicBlock & BB,MachineBasicBlock::iterator I,unsigned DstReg,uint64_t Imm) const1302 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
1303                                          MachineBasicBlock::iterator I,
1304                                          unsigned DstReg,
1305                                          uint64_t Imm) const {
1306   MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
1307                                                   AMDGPU::ALU_LITERAL_X);
1308   setImmOperand(MovImm, AMDGPU::OpName::literal, Imm);
1309   return MovImm;
1310 }
1311 
buildMovInstr(MachineBasicBlock * MBB,MachineBasicBlock::iterator I,unsigned DstReg,unsigned SrcReg) const1312 MachineInstr *R600InstrInfo::buildMovInstr(MachineBasicBlock *MBB,
1313                                        MachineBasicBlock::iterator I,
1314                                        unsigned DstReg, unsigned SrcReg) const {
1315   return buildDefaultInstruction(*MBB, I, AMDGPU::MOV, DstReg, SrcReg);
1316 }
1317 
getOperandIdx(const MachineInstr & MI,unsigned Op) const1318 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
1319   return getOperandIdx(MI.getOpcode(), Op);
1320 }
1321 
getOperandIdx(unsigned Opcode,unsigned Op) const1322 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const {
1323   return AMDGPU::getNamedOperandIdx(Opcode, Op);
1324 }
1325 
setImmOperand(MachineInstr * MI,unsigned Op,int64_t Imm) const1326 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op,
1327                                   int64_t Imm) const {
1328   int Idx = getOperandIdx(*MI, Op);
1329   assert(Idx != -1 && "Operand not supported for this instruction.");
1330   assert(MI->getOperand(Idx).isImm());
1331   MI->getOperand(Idx).setImm(Imm);
1332 }
1333 
1334 //===----------------------------------------------------------------------===//
1335 // Instruction flag getters/setters
1336 //===----------------------------------------------------------------------===//
1337 
hasFlagOperand(const MachineInstr & MI) const1338 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1339   return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1340 }
1341 
getFlagOp(MachineInstr * MI,unsigned SrcIdx,unsigned Flag) const1342 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1343                                          unsigned Flag) const {
1344   unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1345   int FlagIndex = 0;
1346   if (Flag != 0) {
1347     // If we pass something other than the default value of Flag to this
1348     // function, it means we are want to set a flag on an instruction
1349     // that uses native encoding.
1350     assert(HAS_NATIVE_OPERANDS(TargetFlags));
1351     bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1352     switch (Flag) {
1353     case MO_FLAG_CLAMP:
1354       FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp);
1355       break;
1356     case MO_FLAG_MASK:
1357       FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write);
1358       break;
1359     case MO_FLAG_NOT_LAST:
1360     case MO_FLAG_LAST:
1361       FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last);
1362       break;
1363     case MO_FLAG_NEG:
1364       switch (SrcIdx) {
1365       case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break;
1366       case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break;
1367       case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break;
1368       }
1369       break;
1370 
1371     case MO_FLAG_ABS:
1372       assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1373                        "instructions.");
1374       (void)IsOP3;
1375       switch (SrcIdx) {
1376       case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break;
1377       case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break;
1378       }
1379       break;
1380 
1381     default:
1382       FlagIndex = -1;
1383       break;
1384     }
1385     assert(FlagIndex != -1 && "Flag not supported for this instruction");
1386   } else {
1387       FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1388       assert(FlagIndex != 0 &&
1389          "Instruction flags not supported for this instruction");
1390   }
1391 
1392   MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1393   assert(FlagOp.isImm());
1394   return FlagOp;
1395 }
1396 
addFlag(MachineInstr * MI,unsigned Operand,unsigned Flag) const1397 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1398                             unsigned Flag) const {
1399   unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1400   if (Flag == 0) {
1401     return;
1402   }
1403   if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1404     MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1405     if (Flag == MO_FLAG_NOT_LAST) {
1406       clearFlag(MI, Operand, MO_FLAG_LAST);
1407     } else if (Flag == MO_FLAG_MASK) {
1408       clearFlag(MI, Operand, Flag);
1409     } else {
1410       FlagOp.setImm(1);
1411     }
1412   } else {
1413       MachineOperand &FlagOp = getFlagOp(MI, Operand);
1414       FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1415   }
1416 }
1417 
clearFlag(MachineInstr * MI,unsigned Operand,unsigned Flag) const1418 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1419                               unsigned Flag) const {
1420   unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1421   if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1422     MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1423     FlagOp.setImm(0);
1424   } else {
1425     MachineOperand &FlagOp = getFlagOp(MI);
1426     unsigned InstFlags = FlagOp.getImm();
1427     InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1428     FlagOp.setImm(InstFlags);
1429   }
1430 }
1431