1 //===-- R600MachineScheduler.cpp - R600 Scheduler Interface -*- C++ -*-----===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
10 /// R600 Machine Scheduler interface
12 //===----------------------------------------------------------------------===//
14 #include "R600MachineScheduler.h"
15 #include "MCTargetDesc/R600MCTargetDesc.h"
16 #include "R600Subtarget.h"
20 #define DEBUG_TYPE "machine-scheduler"
22 void R600SchedStrategy::initialize(ScheduleDAGMI *dag) {
23 assert(dag->hasVRegLiveness() && "R600SchedStrategy needs vreg liveness");
24 DAG = static_cast<ScheduleDAGMILive*>(dag);
25 const R600Subtarget &ST = DAG->MF.getSubtarget<R600Subtarget>();
26 TII = static_cast<const R600InstrInfo*>(DAG->TII);
27 TRI = static_cast<const R600RegisterInfo*>(DAG->TRI);
28 VLIW5 = !ST.hasCaymanISA();
30 CurInstKind = IDOther;
32 OccupiedSlotsMask = 31;
33 InstKindLimit[IDAlu] = TII->getMaxAlusPerClause();
34 InstKindLimit[IDOther] = 32;
35 InstKindLimit[IDFetch] = ST.getTexVTXClauseSize();
40 void R600SchedStrategy::MoveUnits(std::vector<SUnit *> &QSrc,
41 std::vector<SUnit *> &QDst)
43 llvm::append_range(QDst, QSrc);
47 static unsigned getWFCountLimitedByGPR(unsigned GPRCount) {
48 assert (GPRCount && "GPRCount cannot be 0");
49 return 248 / GPRCount;
52 SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) {
54 NextInstKind = IDOther;
58 // check if we might want to switch current clause type
59 bool AllowSwitchToAlu = (CurEmitted >= InstKindLimit[CurInstKind]) ||
60 (Available[CurInstKind].empty());
61 bool AllowSwitchFromAlu = (CurEmitted >= InstKindLimit[CurInstKind]) &&
62 (!Available[IDFetch].empty() || !Available[IDOther].empty());
64 if (CurInstKind == IDAlu && !Available[IDFetch].empty()) {
65 // We use the heuristic provided by AMD Accelerated Parallel Processing
66 // OpenCL Programming Guide :
67 // The approx. number of WF that allows TEX inst to hide ALU inst is :
68 // 500 (cycles for TEX) / (AluFetchRatio * 8 (cycles for ALU))
69 float ALUFetchRationEstimate =
70 (AluInstCount + AvailablesAluCount() + Pending[IDAlu].size()) /
71 (FetchInstCount + Available[IDFetch].size());
72 if (ALUFetchRationEstimate == 0) {
73 AllowSwitchFromAlu = true;
75 unsigned NeededWF = 62.5f / ALUFetchRationEstimate;
76 LLVM_DEBUG(dbgs() << NeededWF << " approx. Wavefronts Required\n");
77 // We assume the local GPR requirements to be "dominated" by the requirement
78 // of the TEX clause (which consumes 128 bits regs) ; ALU inst before and
79 // after TEX are indeed likely to consume or generate values from/for the
81 // Available[IDFetch].size() * 2 : GPRs required in the Fetch clause
82 // We assume that fetch instructions are either TnXYZW = TEX TnXYZW (need
83 // one GPR) or TmXYZW = TnXYZW (need 2 GPR).
84 // (TODO : use RegisterPressure)
85 // If we are going too use too many GPR, we flush Fetch instruction to lower
86 // register pressure on 128 bits regs.
87 unsigned NearRegisterRequirement = 2 * Available[IDFetch].size();
88 if (NeededWF > getWFCountLimitedByGPR(NearRegisterRequirement))
89 AllowSwitchFromAlu = true;
93 if (!SU && ((AllowSwitchToAlu && CurInstKind != IDAlu) ||
94 (!AllowSwitchFromAlu && CurInstKind == IDAlu))) {
97 if (!SU && !PhysicalRegCopy.empty()) {
98 SU = PhysicalRegCopy.front();
99 PhysicalRegCopy.erase(PhysicalRegCopy.begin());
102 if (CurEmitted >= InstKindLimit[IDAlu])
104 NextInstKind = IDAlu;
110 SU = pickOther(IDFetch);
112 NextInstKind = IDFetch;
117 SU = pickOther(IDOther);
119 NextInstKind = IDOther;
123 dbgs() << " ** Pick node **\n";
126 dbgs() << "NO NODE \n";
127 for (const SUnit &S : DAG->SUnits)
135 void R600SchedStrategy::schedNode(SUnit *SU, bool IsTopNode) {
136 if (NextInstKind != CurInstKind) {
137 LLVM_DEBUG(dbgs() << "Instruction Type Switch\n");
138 if (NextInstKind != IDAlu)
139 OccupiedSlotsMask |= 31;
141 CurInstKind = NextInstKind;
144 if (CurInstKind == IDAlu) {
146 switch (getAluKind(SU)) {
154 for (MachineInstr::mop_iterator It = SU->getInstr()->operands_begin(),
155 E = SU->getInstr()->operands_end(); It != E; ++It) {
156 MachineOperand &MO = *It;
157 if (MO.isReg() && MO.getReg() == R600::ALU_LITERAL_X)
166 LLVM_DEBUG(dbgs() << CurEmitted << " Instructions Emitted in this clause\n");
168 if (CurInstKind != IDFetch) {
169 MoveUnits(Pending[IDFetch], Available[IDFetch]);
175 isPhysicalRegCopy(MachineInstr *MI) {
176 if (MI->getOpcode() != R600::COPY)
179 return !MI->getOperand(1).getReg().isVirtual();
182 void R600SchedStrategy::releaseTopNode(SUnit *SU) {
183 LLVM_DEBUG(dbgs() << "Top Releasing "; DAG->dumpNode(*SU));
186 void R600SchedStrategy::releaseBottomNode(SUnit *SU) {
187 LLVM_DEBUG(dbgs() << "Bottom Releasing "; DAG->dumpNode(*SU));
188 if (isPhysicalRegCopy(SU->getInstr())) {
189 PhysicalRegCopy.push_back(SU);
193 int IK = getInstKind(SU);
195 // There is no export clause, we can schedule one as soon as its ready
197 Available[IDOther].push_back(SU);
199 Pending[IK].push_back(SU);
203 bool R600SchedStrategy::regBelongsToClass(Register Reg,
204 const TargetRegisterClass *RC) const {
205 if (!Reg.isVirtual()) {
206 return RC->contains(Reg);
208 return MRI->getRegClass(Reg) == RC;
212 R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
213 MachineInstr *MI = SU->getInstr();
215 if (TII->isTransOnly(*MI))
218 switch (MI->getOpcode()) {
221 case R600::INTERP_PAIR_XY:
222 case R600::INTERP_PAIR_ZW:
223 case R600::INTERP_VEC_LOAD:
227 if (MI->getOperand(1).isUndef()) {
228 // MI will become a KILL, don't considers it in scheduling
236 // Does the instruction take a whole IG ?
237 // XXX: Is it possible to add a helper function in R600InstrInfo that can
238 // be used here and in R600PacketizerList::isSoloInstruction() ?
239 if(TII->isVector(*MI) ||
240 TII->isCubeOp(MI->getOpcode()) ||
241 TII->isReductionOp(MI->getOpcode()) ||
242 MI->getOpcode() == R600::GROUP_BARRIER) {
246 if (TII->isLDSInstr(MI->getOpcode())) {
250 // Is the result already assigned to a channel ?
251 unsigned DestSubReg = MI->getOperand(0).getSubReg();
252 switch (DestSubReg) {
265 // Is the result already member of a X/Y/Z/W class ?
266 Register DestReg = MI->getOperand(0).getReg();
267 if (regBelongsToClass(DestReg, &R600::R600_TReg32_XRegClass) ||
268 regBelongsToClass(DestReg, &R600::R600_AddrRegClass))
270 if (regBelongsToClass(DestReg, &R600::R600_TReg32_YRegClass))
272 if (regBelongsToClass(DestReg, &R600::R600_TReg32_ZRegClass))
274 if (regBelongsToClass(DestReg, &R600::R600_TReg32_WRegClass))
276 if (regBelongsToClass(DestReg, &R600::R600_Reg128RegClass))
279 // LDS src registers cannot be used in the Trans slot.
280 if (TII->readsLDSSrcReg(*MI))
286 int R600SchedStrategy::getInstKind(SUnit* SU) {
287 int Opcode = SU->getInstr()->getOpcode();
289 if (TII->usesTextureCache(Opcode) || TII->usesVertexCache(Opcode))
292 if (TII->isALUInstr(Opcode)) {
299 case R600::CONST_COPY:
300 case R600::INTERP_PAIR_XY:
301 case R600::INTERP_PAIR_ZW:
302 case R600::INTERP_VEC_LOAD:
310 SUnit *R600SchedStrategy::PopInst(std::vector<SUnit *> &Q, bool AnyALU) {
313 for (std::vector<SUnit *>::reverse_iterator It = Q.rbegin(), E = Q.rend();
316 InstructionsGroupCandidate.push_back(SU->getInstr());
317 if (TII->fitsConstReadLimitations(InstructionsGroupCandidate) &&
318 (!AnyALU || !TII->isVectorOnly(*SU->getInstr()))) {
319 InstructionsGroupCandidate.pop_back();
320 Q.erase((It + 1).base());
323 InstructionsGroupCandidate.pop_back();
329 void R600SchedStrategy::LoadAlu() {
330 std::vector<SUnit *> &QSrc = Pending[IDAlu];
331 for (SUnit *SU : QSrc) {
332 AluKind AK = getAluKind(SU);
333 AvailableAlus[AK].push_back(SU);
338 void R600SchedStrategy::PrepareNextSlot() {
339 LLVM_DEBUG(dbgs() << "New Slot\n");
340 assert(OccupiedSlotsMask && "Slot wasn't filled");
341 OccupiedSlotsMask = 0;
342 // if (HwGen == AMDGPUSubtarget::NORTHERN_ISLANDS)
343 // OccupiedSlotsMask |= 16;
344 InstructionsGroupCandidate.clear();
348 void R600SchedStrategy::AssignSlot(MachineInstr* MI, unsigned Slot) {
349 int DstIndex = TII->getOperandIdx(MI->getOpcode(), R600::OpName::dst);
350 if (DstIndex == -1) {
353 Register DestReg = MI->getOperand(DstIndex).getReg();
354 // PressureRegister crashes if an operand is def and used in the same inst
355 // and we try to constraint its regclass
356 for (MachineInstr::mop_iterator It = MI->operands_begin(),
357 E = MI->operands_end(); It != E; ++It) {
358 MachineOperand &MO = *It;
359 if (MO.isReg() && !MO.isDef() &&
360 MO.getReg() == DestReg)
363 // Constrains the regclass of DestReg to assign it to Slot
366 MRI->constrainRegClass(DestReg, &R600::R600_TReg32_XRegClass);
369 MRI->constrainRegClass(DestReg, &R600::R600_TReg32_YRegClass);
372 MRI->constrainRegClass(DestReg, &R600::R600_TReg32_ZRegClass);
375 MRI->constrainRegClass(DestReg, &R600::R600_TReg32_WRegClass);
380 SUnit *R600SchedStrategy::AttemptFillSlot(unsigned Slot, bool AnyAlu) {
381 static const AluKind IndexToID[] = {AluT_X, AluT_Y, AluT_Z, AluT_W};
382 SUnit *SlotedSU = PopInst(AvailableAlus[IndexToID[Slot]], AnyAlu);
385 SUnit *UnslotedSU = PopInst(AvailableAlus[AluAny], AnyAlu);
387 AssignSlot(UnslotedSU->getInstr(), Slot);
391 unsigned R600SchedStrategy::AvailablesAluCount() const {
392 return AvailableAlus[AluAny].size() + AvailableAlus[AluT_XYZW].size() +
393 AvailableAlus[AluT_X].size() + AvailableAlus[AluT_Y].size() +
394 AvailableAlus[AluT_Z].size() + AvailableAlus[AluT_W].size() +
395 AvailableAlus[AluTrans].size() + AvailableAlus[AluDiscarded].size() +
396 AvailableAlus[AluPredX].size();
399 SUnit* R600SchedStrategy::pickAlu() {
400 while (AvailablesAluCount() || !Pending[IDAlu].empty()) {
401 if (!OccupiedSlotsMask) {
402 // Bottom up scheduling : predX must comes first
403 if (!AvailableAlus[AluPredX].empty()) {
404 OccupiedSlotsMask |= 31;
405 return PopInst(AvailableAlus[AluPredX], false);
407 // Flush physical reg copies (RA will discard them)
408 if (!AvailableAlus[AluDiscarded].empty()) {
409 OccupiedSlotsMask |= 31;
410 return PopInst(AvailableAlus[AluDiscarded], false);
412 // If there is a T_XYZW alu available, use it
413 if (!AvailableAlus[AluT_XYZW].empty()) {
414 OccupiedSlotsMask |= 15;
415 return PopInst(AvailableAlus[AluT_XYZW], false);
418 bool TransSlotOccupied = OccupiedSlotsMask & 16;
419 if (!TransSlotOccupied && VLIW5) {
420 if (!AvailableAlus[AluTrans].empty()) {
421 OccupiedSlotsMask |= 16;
422 return PopInst(AvailableAlus[AluTrans], false);
424 SUnit *SU = AttemptFillSlot(3, true);
426 OccupiedSlotsMask |= 16;
430 for (int Chan = 3; Chan > -1; --Chan) {
431 bool isOccupied = OccupiedSlotsMask & (1 << Chan);
433 SUnit *SU = AttemptFillSlot(Chan, false);
435 OccupiedSlotsMask |= (1 << Chan);
436 InstructionsGroupCandidate.push_back(SU->getInstr());
446 SUnit* R600SchedStrategy::pickOther(int QID) {
448 std::vector<SUnit *> &AQ = Available[QID];
451 MoveUnits(Pending[QID], AQ);