1//===-- R600MachineScheduler.cpp - R600 Scheduler Interface -*- C++ -*-----===//
2//
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
6//
7//===----------------------------------------------------------------------===//
8//
9/// \file
10/// R600 Machine Scheduler interface
11//
12//===----------------------------------------------------------------------===//
13
14#include "R600MachineScheduler.h"
15#include "MCTargetDesc/R600MCTargetDesc.h"
16#include "R600Subtarget.h"
17
18using namespace llvm;
19
20#define DEBUG_TYPE "machine-scheduler"
21
22void 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();
29 MRI = &DAG->MRI;
30 CurInstKind = IDOther;
31 CurEmitted = 0;
32 OccupiedSlotsMask = 31;
33 InstKindLimit[IDAlu] = TII->getMaxAlusPerClause();
34 InstKindLimit[IDOther] = 32;
35 InstKindLimit[IDFetch] = ST.getTexVTXClauseSize();
36 AluInstCount = 0;
37 FetchInstCount = 0;
38}
39
40void R600SchedStrategy::MoveUnits(std::vector<SUnit *> &QSrc,
41 std::vector<SUnit *> &QDst)
42{
43 llvm::append_range(C&: QDst, R&: QSrc);
44 QSrc.clear();
45}
46
47static unsigned getWFCountLimitedByGPR(unsigned GPRCount) {
48 assert (GPRCount && "GPRCount cannot be 0");
49 return 248 / GPRCount;
50}
51
52SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) {
53 SUnit *SU = nullptr;
54 NextInstKind = IDOther;
55
56 IsTopNode = false;
57
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());
63
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;
74 } else {
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
80 // TEX clause.
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(GPRCount: NearRegisterRequirement))
89 AllowSwitchFromAlu = true;
90 }
91 }
92
93 if (!SU && ((AllowSwitchToAlu && CurInstKind != IDAlu) ||
94 (!AllowSwitchFromAlu && CurInstKind == IDAlu))) {
95 // try to pick ALU
96 SU = pickAlu();
97 if (!SU && !PhysicalRegCopy.empty()) {
98 SU = PhysicalRegCopy.front();
99 PhysicalRegCopy.erase(position: PhysicalRegCopy.begin());
100 }
101 if (SU) {
102 if (CurEmitted >= InstKindLimit[IDAlu])
103 CurEmitted = 0;
104 NextInstKind = IDAlu;
105 }
106 }
107
108 if (!SU) {
109 // try to pick FETCH
110 SU = pickOther(QID: IDFetch);
111 if (SU)
112 NextInstKind = IDFetch;
113 }
114
115 // try to pick other
116 if (!SU) {
117 SU = pickOther(QID: IDOther);
118 if (SU)
119 NextInstKind = IDOther;
120 }
121
122 LLVM_DEBUG(if (SU) {
123 dbgs() << " ** Pick node **\n";
124 DAG->dumpNode(*SU);
125 } else {
126 dbgs() << "NO NODE \n";
127 for (const SUnit &S : DAG->SUnits)
128 if (!S.isScheduled)
129 DAG->dumpNode(S);
130 });
131
132 return SU;
133}
134
135void 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;
140 CurEmitted = 0;
141 CurInstKind = NextInstKind;
142 }
143
144 if (CurInstKind == IDAlu) {
145 AluInstCount ++;
146 switch (getAluKind(SU)) {
147 case AluT_XYZW:
148 CurEmitted += 4;
149 break;
150 case AluDiscarded:
151 break;
152 default: {
153 ++CurEmitted;
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)
158 ++CurEmitted;
159 }
160 }
161 }
162 } else {
163 ++CurEmitted;
164 }
165
166 LLVM_DEBUG(dbgs() << CurEmitted << " Instructions Emitted in this clause\n");
167
168 if (CurInstKind != IDFetch) {
169 MoveUnits(QSrc&: Pending[IDFetch], QDst&: Available[IDFetch]);
170 } else
171 FetchInstCount++;
172}
173
174static bool
175isPhysicalRegCopy(MachineInstr *MI) {
176 if (MI->getOpcode() != R600::COPY)
177 return false;
178
179 return !MI->getOperand(i: 1).getReg().isVirtual();
180}
181
182void R600SchedStrategy::releaseTopNode(SUnit *SU) {
183 LLVM_DEBUG(dbgs() << "Top Releasing "; DAG->dumpNode(*SU));
184}
185
186void R600SchedStrategy::releaseBottomNode(SUnit *SU) {
187 LLVM_DEBUG(dbgs() << "Bottom Releasing "; DAG->dumpNode(*SU));
188 if (isPhysicalRegCopy(MI: SU->getInstr())) {
189 PhysicalRegCopy.push_back(x: SU);
190 return;
191 }
192
193 int IK = getInstKind(SU);
194
195 // There is no export clause, we can schedule one as soon as its ready
196 if (IK == IDOther)
197 Available[IDOther].push_back(x: SU);
198 else
199 Pending[IK].push_back(x: SU);
200
201}
202
203bool R600SchedStrategy::regBelongsToClass(Register Reg,
204 const TargetRegisterClass *RC) const {
205 if (!Reg.isVirtual())
206 return RC->contains(Reg);
207 return MRI->getRegClass(Reg) == RC;
208}
209
210R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
211 MachineInstr *MI = SU->getInstr();
212
213 if (TII->isTransOnly(MI: *MI))
214 return AluTrans;
215
216 switch (MI->getOpcode()) {
217 case R600::PRED_X:
218 return AluPredX;
219 case R600::INTERP_PAIR_XY:
220 case R600::INTERP_PAIR_ZW:
221 case R600::INTERP_VEC_LOAD:
222 case R600::DOT_4:
223 return AluT_XYZW;
224 case R600::COPY:
225 if (MI->getOperand(i: 1).isUndef()) {
226 // MI will become a KILL, don't considers it in scheduling
227 return AluDiscarded;
228 }
229 break;
230 default:
231 break;
232 }
233
234 // Does the instruction take a whole IG ?
235 // XXX: Is it possible to add a helper function in R600InstrInfo that can
236 // be used here and in R600PacketizerList::isSoloInstruction() ?
237 if(TII->isVector(MI: *MI) ||
238 TII->isCubeOp(opcode: MI->getOpcode()) ||
239 TII->isReductionOp(opcode: MI->getOpcode()) ||
240 MI->getOpcode() == R600::GROUP_BARRIER) {
241 return AluT_XYZW;
242 }
243
244 if (TII->isLDSInstr(Opcode: MI->getOpcode())) {
245 return AluT_X;
246 }
247
248 // Is the result already assigned to a channel ?
249 unsigned DestSubReg = MI->getOperand(i: 0).getSubReg();
250 switch (DestSubReg) {
251 case R600::sub0:
252 return AluT_X;
253 case R600::sub1:
254 return AluT_Y;
255 case R600::sub2:
256 return AluT_Z;
257 case R600::sub3:
258 return AluT_W;
259 default:
260 break;
261 }
262
263 // Is the result already member of a X/Y/Z/W class ?
264 Register DestReg = MI->getOperand(i: 0).getReg();
265 if (regBelongsToClass(Reg: DestReg, RC: &R600::R600_TReg32_XRegClass) ||
266 regBelongsToClass(Reg: DestReg, RC: &R600::R600_AddrRegClass))
267 return AluT_X;
268 if (regBelongsToClass(Reg: DestReg, RC: &R600::R600_TReg32_YRegClass))
269 return AluT_Y;
270 if (regBelongsToClass(Reg: DestReg, RC: &R600::R600_TReg32_ZRegClass))
271 return AluT_Z;
272 if (regBelongsToClass(Reg: DestReg, RC: &R600::R600_TReg32_WRegClass))
273 return AluT_W;
274 if (regBelongsToClass(Reg: DestReg, RC: &R600::R600_Reg128RegClass))
275 return AluT_XYZW;
276
277 // LDS src registers cannot be used in the Trans slot.
278 if (TII->readsLDSSrcReg(MI: *MI))
279 return AluT_XYZW;
280
281 return AluAny;
282}
283
284int R600SchedStrategy::getInstKind(SUnit* SU) {
285 int Opcode = SU->getInstr()->getOpcode();
286
287 if (TII->usesTextureCache(Opcode) || TII->usesVertexCache(Opcode))
288 return IDFetch;
289
290 if (TII->isALUInstr(Opcode)) {
291 return IDAlu;
292 }
293
294 switch (Opcode) {
295 case R600::PRED_X:
296 case R600::COPY:
297 case R600::CONST_COPY:
298 case R600::INTERP_PAIR_XY:
299 case R600::INTERP_PAIR_ZW:
300 case R600::INTERP_VEC_LOAD:
301 case R600::DOT_4:
302 return IDAlu;
303 default:
304 return IDOther;
305 }
306}
307
308SUnit *R600SchedStrategy::PopInst(std::vector<SUnit *> &Q, bool AnyALU) {
309 if (Q.empty())
310 return nullptr;
311 for (std::vector<SUnit *>::reverse_iterator It = Q.rbegin(), E = Q.rend();
312 It != E; ++It) {
313 SUnit *SU = *It;
314 InstructionsGroupCandidate.push_back(x: SU->getInstr());
315 if (TII->fitsConstReadLimitations(InstructionsGroupCandidate) &&
316 (!AnyALU || !TII->isVectorOnly(MI: *SU->getInstr()))) {
317 InstructionsGroupCandidate.pop_back();
318 Q.erase(position: (It + 1).base());
319 return SU;
320 }
321 InstructionsGroupCandidate.pop_back();
322 }
323 return nullptr;
324}
325
326void R600SchedStrategy::LoadAlu() {
327 std::vector<SUnit *> &QSrc = Pending[IDAlu];
328 for (SUnit *SU : QSrc) {
329 AluKind AK = getAluKind(SU);
330 AvailableAlus[AK].push_back(x: SU);
331 }
332 QSrc.clear();
333}
334
335void R600SchedStrategy::PrepareNextSlot() {
336 LLVM_DEBUG(dbgs() << "New Slot\n");
337 assert(OccupiedSlotsMask && "Slot wasn't filled");
338 OccupiedSlotsMask = 0;
339 // if (HwGen == AMDGPUSubtarget::NORTHERN_ISLANDS)
340 // OccupiedSlotsMask |= 16;
341 InstructionsGroupCandidate.clear();
342 LoadAlu();
343}
344
345void R600SchedStrategy::AssignSlot(MachineInstr* MI, unsigned Slot) {
346 int DstIndex = TII->getOperandIdx(Opcode: MI->getOpcode(), Op: R600::OpName::dst);
347 if (DstIndex == -1) {
348 return;
349 }
350 Register DestReg = MI->getOperand(i: DstIndex).getReg();
351 // PressureRegister crashes if an operand is def and used in the same inst
352 // and we try to constraint its regclass
353 for (const MachineOperand &MO : MI->all_uses())
354 if (MO.getReg() == DestReg)
355 return;
356 // Constrains the regclass of DestReg to assign it to Slot
357 switch (Slot) {
358 case 0:
359 MRI->constrainRegClass(Reg: DestReg, RC: &R600::R600_TReg32_XRegClass);
360 break;
361 case 1:
362 MRI->constrainRegClass(Reg: DestReg, RC: &R600::R600_TReg32_YRegClass);
363 break;
364 case 2:
365 MRI->constrainRegClass(Reg: DestReg, RC: &R600::R600_TReg32_ZRegClass);
366 break;
367 case 3:
368 MRI->constrainRegClass(Reg: DestReg, RC: &R600::R600_TReg32_WRegClass);
369 break;
370 }
371}
372
373SUnit *R600SchedStrategy::AttemptFillSlot(unsigned Slot, bool AnyAlu) {
374 static const AluKind IndexToID[] = {AluT_X, AluT_Y, AluT_Z, AluT_W};
375 SUnit *SlotedSU = PopInst(Q&: AvailableAlus[IndexToID[Slot]], AnyALU: AnyAlu);
376 if (SlotedSU)
377 return SlotedSU;
378 SUnit *UnslotedSU = PopInst(Q&: AvailableAlus[AluAny], AnyALU: AnyAlu);
379 if (UnslotedSU)
380 AssignSlot(MI: UnslotedSU->getInstr(), Slot);
381 return UnslotedSU;
382}
383
384unsigned R600SchedStrategy::AvailablesAluCount() const {
385 return AvailableAlus[AluAny].size() + AvailableAlus[AluT_XYZW].size() +
386 AvailableAlus[AluT_X].size() + AvailableAlus[AluT_Y].size() +
387 AvailableAlus[AluT_Z].size() + AvailableAlus[AluT_W].size() +
388 AvailableAlus[AluTrans].size() + AvailableAlus[AluDiscarded].size() +
389 AvailableAlus[AluPredX].size();
390}
391
392SUnit* R600SchedStrategy::pickAlu() {
393 while (AvailablesAluCount() || !Pending[IDAlu].empty()) {
394 if (!OccupiedSlotsMask) {
395 // Bottom up scheduling : predX must comes first
396 if (!AvailableAlus[AluPredX].empty()) {
397 OccupiedSlotsMask |= 31;
398 return PopInst(Q&: AvailableAlus[AluPredX], AnyALU: false);
399 }
400 // Flush physical reg copies (RA will discard them)
401 if (!AvailableAlus[AluDiscarded].empty()) {
402 OccupiedSlotsMask |= 31;
403 return PopInst(Q&: AvailableAlus[AluDiscarded], AnyALU: false);
404 }
405 // If there is a T_XYZW alu available, use it
406 if (!AvailableAlus[AluT_XYZW].empty()) {
407 OccupiedSlotsMask |= 15;
408 return PopInst(Q&: AvailableAlus[AluT_XYZW], AnyALU: false);
409 }
410 }
411 bool TransSlotOccupied = OccupiedSlotsMask & 16;
412 if (!TransSlotOccupied && VLIW5) {
413 if (!AvailableAlus[AluTrans].empty()) {
414 OccupiedSlotsMask |= 16;
415 return PopInst(Q&: AvailableAlus[AluTrans], AnyALU: false);
416 }
417 SUnit *SU = AttemptFillSlot(Slot: 3, AnyAlu: true);
418 if (SU) {
419 OccupiedSlotsMask |= 16;
420 return SU;
421 }
422 }
423 for (int Chan = 3; Chan > -1; --Chan) {
424 bool isOccupied = OccupiedSlotsMask & (1 << Chan);
425 if (!isOccupied) {
426 SUnit *SU = AttemptFillSlot(Slot: Chan, AnyAlu: false);
427 if (SU) {
428 OccupiedSlotsMask |= (1 << Chan);
429 InstructionsGroupCandidate.push_back(x: SU->getInstr());
430 return SU;
431 }
432 }
433 }
434 PrepareNextSlot();
435 }
436 return nullptr;
437}
438
439SUnit* R600SchedStrategy::pickOther(int QID) {
440 SUnit *SU = nullptr;
441 std::vector<SUnit *> &AQ = Available[QID];
442
443 if (AQ.empty()) {
444 MoveUnits(QSrc&: Pending[QID], QDst&: AQ);
445 }
446 if (!AQ.empty()) {
447 SU = AQ.back();
448 AQ.pop_back();
449 }
450 return SU;
451}
452