1 //===- SIInsertWaitcnts.cpp - Insert Wait Instructions --------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// Insert wait instructions for memory reads and writes.
13 /// Memory reads and writes are issued asynchronously, so we need to insert
14 /// S_WAITCNT instructions when we want to access any of their results or
15 /// overwrite any register that's used asynchronously.
17 /// TODO: This pass currently keeps one timeline per hardware counter. A more
18 /// finely-grained approach that keeps one timeline per event type could
19 /// sometimes get away with generating weaker s_waitcnt instructions. For
20 /// example, when both SMEM and LDS are in flight and we need to wait for
21 /// the i-th-last LDS instruction, then an lgkmcnt(i) is actually sufficient,
22 /// but the pass will currently generate a conservative lgkmcnt(0) because
23 /// multiple event types are in flight.
25 //===----------------------------------------------------------------------===//
28 #include "AMDGPUSubtarget.h"
29 #include "SIDefines.h"
30 #include "SIInstrInfo.h"
31 #include "SIMachineFunctionInfo.h"
32 #include "SIRegisterInfo.h"
33 #include "Utils/AMDGPUBaseInfo.h"
34 #include "llvm/ADT/DenseMap.h"
35 #include "llvm/ADT/DenseSet.h"
36 #include "llvm/ADT/PostOrderIterator.h"
37 #include "llvm/ADT/STLExtras.h"
38 #include "llvm/ADT/SmallVector.h"
39 #include "llvm/CodeGen/MachineBasicBlock.h"
40 #include "llvm/CodeGen/MachineFunction.h"
41 #include "llvm/CodeGen/MachineFunctionPass.h"
42 #include "llvm/CodeGen/MachineInstr.h"
43 #include "llvm/CodeGen/MachineInstrBuilder.h"
44 #include "llvm/CodeGen/MachineMemOperand.h"
45 #include "llvm/CodeGen/MachineOperand.h"
46 #include "llvm/CodeGen/MachineRegisterInfo.h"
47 #include "llvm/IR/DebugLoc.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/DebugCounter.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/raw_ostream.h"
63 #define DEBUG_TYPE "si-insert-waitcnts"
65 DEBUG_COUNTER(ForceExpCounter, DEBUG_TYPE"-forceexp",
66 "Force emit s_waitcnt expcnt(0) instrs");
67 DEBUG_COUNTER(ForceLgkmCounter, DEBUG_TYPE"-forcelgkm",
68 "Force emit s_waitcnt lgkmcnt(0) instrs");
69 DEBUG_COUNTER(ForceVMCounter, DEBUG_TYPE"-forcevm",
70 "Force emit s_waitcnt vmcnt(0) instrs");
72 static cl::opt<unsigned> ForceEmitZeroFlag(
73 "amdgpu-waitcnt-forcezero",
74 cl::desc("Force all waitcnt instrs to be emitted as s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)"),
75 cl::init(0), cl::Hidden);
79 template <typename EnumT>
81 : public iterator_facade_base<enum_iterator<EnumT>,
82 std::forward_iterator_tag, const EnumT> {
85 enum_iterator() = default;
86 enum_iterator(EnumT Value) : Value(Value) {}
88 enum_iterator &operator++() {
89 Value = static_cast<EnumT>(Value + 1);
93 bool operator==(const enum_iterator &RHS) const { return Value == RHS.Value; }
95 EnumT operator*() const { return Value; }
98 // Class of object that encapsulates latest instruction counter score
99 // associated with the operand. Used for determining whether
100 // s_waitcnt instruction needs to be emited.
102 #define CNT_MASK(t) (1u << (t))
104 enum InstCounterType { VM_CNT = 0, LGKM_CNT, EXP_CNT, NUM_INST_CNTS };
106 iterator_range<enum_iterator<InstCounterType>> inst_counter_types() {
107 return make_range(enum_iterator<InstCounterType>(VM_CNT),
108 enum_iterator<InstCounterType>(NUM_INST_CNTS));
111 using RegInterval = std::pair<signed, signed>;
129 VMEM_ACCESS, // vector-memory read & write
130 LDS_ACCESS, // lds read & write
131 GDS_ACCESS, // gds read & write
132 SQ_MESSAGE, // send message
133 SMEM_ACCESS, // scalar-memory read & write
134 EXP_GPR_LOCK, // export holding on its data src
135 GDS_GPR_LOCK, // GDS holding on its data and addr src
136 EXP_POS_ACCESS, // write to export position
137 EXP_PARAM_ACCESS, // write to export parameter
138 VMW_GPR_LOCK, // vector-memory write holding on its data src
142 static const uint32_t WaitEventMaskForInst[NUM_INST_CNTS] = {
144 (1 << SMEM_ACCESS) | (1 << LDS_ACCESS) | (1 << GDS_ACCESS) |
146 (1 << EXP_GPR_LOCK) | (1 << GDS_GPR_LOCK) | (1 << VMW_GPR_LOCK) |
147 (1 << EXP_PARAM_ACCESS) | (1 << EXP_POS_ACCESS),
151 // 0 .. SQ_MAX_PGM_VGPRS-1 real VGPRs
152 // SQ_MAX_PGM_VGPRS .. NUM_ALL_VGPRS-1 extra VGPR-like slots
153 // NUM_ALL_VGPRS .. NUM_ALL_VGPRS+SQ_MAX_PGM_SGPRS-1 real SGPRs
154 // We reserve a fixed number of VGPR slots in the scoring tables for
155 // special tokens like SCMEM_LDS (needed for buffer load to LDS).
156 enum RegisterMapping {
157 SQ_MAX_PGM_VGPRS = 256, // Maximum programmable VGPRs across all targets.
158 SQ_MAX_PGM_SGPRS = 256, // Maximum programmable SGPRs across all targets.
159 NUM_EXTRA_VGPRS = 1, // A reserved slot for DS.
160 EXTRA_VGPR_LDS = 0, // This is a placeholder the Shader algorithm uses.
161 NUM_ALL_VGPRS = SQ_MAX_PGM_VGPRS + NUM_EXTRA_VGPRS, // Where SGPR starts.
164 void addWait(AMDGPU::Waitcnt &Wait, InstCounterType T, unsigned Count) {
167 Wait.VmCnt = std::min(Wait.VmCnt, Count);
170 Wait.ExpCnt = std::min(Wait.ExpCnt, Count);
173 Wait.LgkmCnt = std::min(Wait.LgkmCnt, Count);
176 llvm_unreachable("bad InstCounterType");
180 // This objects maintains the current score brackets of each wait counter, and
181 // a per-register scoreboard for each wait counter.
183 // We also maintain the latest score for every event type that can change the
184 // waitcnt in order to know if there are multiple types of events within
185 // the brackets. When multiple types of event happen in the bracket,
186 // wait count may get decreased out of order, therefore we need to put in
187 // "s_waitcnt 0" before use.
188 class WaitcntBrackets {
190 WaitcntBrackets(const GCNSubtarget *SubTarget) : ST(SubTarget) {
191 for (auto T : inst_counter_types())
192 memset(VgprScores[T], 0, sizeof(VgprScores[T]));
195 static uint32_t getWaitCountMax(InstCounterType T) {
198 return HardwareLimits.VmcntMax;
200 return HardwareLimits.LgkmcntMax;
202 return HardwareLimits.ExpcntMax;
209 uint32_t getScoreLB(InstCounterType T) const {
210 assert(T < NUM_INST_CNTS);
211 if (T >= NUM_INST_CNTS)
216 uint32_t getScoreUB(InstCounterType T) const {
217 assert(T < NUM_INST_CNTS);
218 if (T >= NUM_INST_CNTS)
223 // Mapping from event to counter.
224 InstCounterType eventCounter(WaitEventType E) {
225 if (E == VMEM_ACCESS)
227 if (WaitEventMaskForInst[LGKM_CNT] & (1 << E))
229 assert(WaitEventMaskForInst[EXP_CNT] & (1 << E));
233 uint32_t getRegScore(int GprNo, InstCounterType T) {
234 if (GprNo < NUM_ALL_VGPRS) {
235 return VgprScores[T][GprNo];
237 assert(T == LGKM_CNT);
238 return SgprScores[GprNo - NUM_ALL_VGPRS];
242 memset(ScoreLBs, 0, sizeof(ScoreLBs));
243 memset(ScoreUBs, 0, sizeof(ScoreUBs));
245 memset(MixedPendingEvents, 0, sizeof(MixedPendingEvents));
246 for (auto T : inst_counter_types())
247 memset(VgprScores[T], 0, sizeof(VgprScores[T]));
248 memset(SgprScores, 0, sizeof(SgprScores));
251 bool merge(const WaitcntBrackets &Other);
253 RegInterval getRegInterval(const MachineInstr *MI, const SIInstrInfo *TII,
254 const MachineRegisterInfo *MRI,
255 const SIRegisterInfo *TRI, unsigned OpNo,
258 int32_t getMaxVGPR() const { return VgprUB; }
259 int32_t getMaxSGPR() const { return SgprUB; }
261 bool counterOutOfOrder(InstCounterType T) const;
262 bool simplifyWaitcnt(AMDGPU::Waitcnt &Wait) const;
263 bool simplifyWaitcnt(InstCounterType T, unsigned &Count) const;
264 void determineWait(InstCounterType T, uint32_t ScoreToWait,
265 AMDGPU::Waitcnt &Wait) const;
266 void applyWaitcnt(const AMDGPU::Waitcnt &Wait);
267 void applyWaitcnt(InstCounterType T, unsigned Count);
268 void updateByEvent(const SIInstrInfo *TII, const SIRegisterInfo *TRI,
269 const MachineRegisterInfo *MRI, WaitEventType E,
272 bool hasPending() const { return PendingEvents != 0; }
273 bool hasPendingEvent(WaitEventType E) const {
274 return PendingEvents & (1 << E);
277 bool hasPendingFlat() const {
278 return ((LastFlat[LGKM_CNT] > ScoreLBs[LGKM_CNT] &&
279 LastFlat[LGKM_CNT] <= ScoreUBs[LGKM_CNT]) ||
280 (LastFlat[VM_CNT] > ScoreLBs[VM_CNT] &&
281 LastFlat[VM_CNT] <= ScoreUBs[VM_CNT]));
284 void setPendingFlat() {
285 LastFlat[VM_CNT] = ScoreUBs[VM_CNT];
286 LastFlat[LGKM_CNT] = ScoreUBs[LGKM_CNT];
289 void print(raw_ostream &);
290 void dump() { print(dbgs()); }
299 static bool mergeScore(const MergeInfo &M, uint32_t &Score,
300 uint32_t OtherScore);
302 void setScoreLB(InstCounterType T, uint32_t Val) {
303 assert(T < NUM_INST_CNTS);
304 if (T >= NUM_INST_CNTS)
309 void setScoreUB(InstCounterType T, uint32_t Val) {
310 assert(T < NUM_INST_CNTS);
311 if (T >= NUM_INST_CNTS)
315 uint32_t UB = ScoreUBs[T] - getWaitCountMax(EXP_CNT);
316 if (ScoreLBs[T] < UB && UB < ScoreUBs[T])
321 void setRegScore(int GprNo, InstCounterType T, uint32_t Val) {
322 if (GprNo < NUM_ALL_VGPRS) {
323 if (GprNo > VgprUB) {
326 VgprScores[T][GprNo] = Val;
328 assert(T == LGKM_CNT);
329 if (GprNo - NUM_ALL_VGPRS > SgprUB) {
330 SgprUB = GprNo - NUM_ALL_VGPRS;
332 SgprScores[GprNo - NUM_ALL_VGPRS] = Val;
336 void setExpScore(const MachineInstr *MI, const SIInstrInfo *TII,
337 const SIRegisterInfo *TRI, const MachineRegisterInfo *MRI,
338 unsigned OpNo, uint32_t Val);
340 const GCNSubtarget *ST = nullptr;
341 uint32_t ScoreLBs[NUM_INST_CNTS] = {0};
342 uint32_t ScoreUBs[NUM_INST_CNTS] = {0};
343 uint32_t PendingEvents = 0;
344 bool MixedPendingEvents[NUM_INST_CNTS] = {false};
345 // Remember the last flat memory operation.
346 uint32_t LastFlat[NUM_INST_CNTS] = {0};
347 // wait_cnt scores for every vgpr.
348 // Keep track of the VgprUB and SgprUB to make merge at join efficient.
351 uint32_t VgprScores[NUM_INST_CNTS][NUM_ALL_VGPRS];
352 // Wait cnt scores for every sgpr, only lgkmcnt is relevant.
353 uint32_t SgprScores[SQ_MAX_PGM_SGPRS] = {0};
356 class SIInsertWaitcnts : public MachineFunctionPass {
358 const GCNSubtarget *ST = nullptr;
359 const SIInstrInfo *TII = nullptr;
360 const SIRegisterInfo *TRI = nullptr;
361 const MachineRegisterInfo *MRI = nullptr;
362 AMDGPU::IsaVersion IV;
364 DenseSet<MachineInstr *> TrackedWaitcntSet;
365 DenseSet<MachineInstr *> VCCZBugHandledSet;
368 MachineBasicBlock *MBB;
369 std::unique_ptr<WaitcntBrackets> Incoming;
372 explicit BlockInfo(MachineBasicBlock *MBB) : MBB(MBB) {}
375 std::vector<BlockInfo> BlockInfos; // by reverse post-order traversal index
376 DenseMap<MachineBasicBlock *, unsigned> RpotIdxMap;
378 // ForceEmitZeroWaitcnts: force all waitcnts insts to be s_waitcnt 0
379 // because of amdgpu-waitcnt-forcezero flag
380 bool ForceEmitZeroWaitcnts;
381 bool ForceEmitWaitcnt[NUM_INST_CNTS];
386 SIInsertWaitcnts() : MachineFunctionPass(ID) {
387 (void)ForceExpCounter;
388 (void)ForceLgkmCounter;
389 (void)ForceVMCounter;
392 bool runOnMachineFunction(MachineFunction &MF) override;
394 StringRef getPassName() const override {
395 return "SI insert wait instructions";
398 void getAnalysisUsage(AnalysisUsage &AU) const override {
399 AU.setPreservesCFG();
400 MachineFunctionPass::getAnalysisUsage(AU);
403 bool isForceEmitWaitcnt() const {
404 for (auto T : inst_counter_types())
405 if (ForceEmitWaitcnt[T])
410 void setForceEmitWaitcnt() {
411 // For non-debug builds, ForceEmitWaitcnt has been initialized to false;
412 // For debug builds, get the debug counter info and adjust if need be
414 if (DebugCounter::isCounterSet(ForceExpCounter) &&
415 DebugCounter::shouldExecute(ForceExpCounter)) {
416 ForceEmitWaitcnt[EXP_CNT] = true;
418 ForceEmitWaitcnt[EXP_CNT] = false;
421 if (DebugCounter::isCounterSet(ForceLgkmCounter) &&
422 DebugCounter::shouldExecute(ForceLgkmCounter)) {
423 ForceEmitWaitcnt[LGKM_CNT] = true;
425 ForceEmitWaitcnt[LGKM_CNT] = false;
428 if (DebugCounter::isCounterSet(ForceVMCounter) &&
429 DebugCounter::shouldExecute(ForceVMCounter)) {
430 ForceEmitWaitcnt[VM_CNT] = true;
432 ForceEmitWaitcnt[VM_CNT] = false;
437 bool mayAccessLDSThroughFlat(const MachineInstr &MI) const;
438 bool generateWaitcntInstBefore(MachineInstr &MI,
439 WaitcntBrackets &ScoreBrackets,
440 MachineInstr *OldWaitcntInstr);
441 void updateEventWaitcntAfter(MachineInstr &Inst,
442 WaitcntBrackets *ScoreBrackets);
443 bool insertWaitcntInBlock(MachineFunction &MF, MachineBasicBlock &Block,
444 WaitcntBrackets &ScoreBrackets);
447 } // end anonymous namespace
449 RegInterval WaitcntBrackets::getRegInterval(const MachineInstr *MI,
450 const SIInstrInfo *TII,
451 const MachineRegisterInfo *MRI,
452 const SIRegisterInfo *TRI,
453 unsigned OpNo, bool Def) const {
454 const MachineOperand &Op = MI->getOperand(OpNo);
455 if (!Op.isReg() || !TRI->isInAllocatableClass(Op.getReg()) ||
456 (Def && !Op.isDef()))
459 // A use via a PW operand does not need a waitcnt.
460 // A partial write is not a WAW.
461 assert(!Op.getSubReg() || !Op.isUndef());
464 const MachineRegisterInfo &MRIA = *MRI;
466 unsigned Reg = TRI->getEncodingValue(Op.getReg());
468 if (TRI->isVGPR(MRIA, Op.getReg())) {
469 assert(Reg >= RegisterEncoding.VGPR0 && Reg <= RegisterEncoding.VGPRL);
470 Result.first = Reg - RegisterEncoding.VGPR0;
471 assert(Result.first >= 0 && Result.first < SQ_MAX_PGM_VGPRS);
472 } else if (TRI->isSGPRReg(MRIA, Op.getReg())) {
473 assert(Reg >= RegisterEncoding.SGPR0 && Reg < SQ_MAX_PGM_SGPRS);
474 Result.first = Reg - RegisterEncoding.SGPR0 + NUM_ALL_VGPRS;
475 assert(Result.first >= NUM_ALL_VGPRS &&
476 Result.first < SQ_MAX_PGM_SGPRS + NUM_ALL_VGPRS);
479 // else if (TRI->isTTMP(MRIA, Reg.getReg())) ...
483 const MachineInstr &MIA = *MI;
484 const TargetRegisterClass *RC = TII->getOpRegClass(MIA, OpNo);
485 unsigned Size = TRI->getRegSizeInBits(*RC);
486 Result.second = Result.first + (Size / 32);
491 void WaitcntBrackets::setExpScore(const MachineInstr *MI,
492 const SIInstrInfo *TII,
493 const SIRegisterInfo *TRI,
494 const MachineRegisterInfo *MRI, unsigned OpNo,
496 RegInterval Interval = getRegInterval(MI, TII, MRI, TRI, OpNo, false);
498 const MachineOperand &Opnd = MI->getOperand(OpNo);
499 assert(TRI->isVGPR(*MRI, Opnd.getReg()));
501 for (signed RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
502 setRegScore(RegNo, EXP_CNT, Val);
506 void WaitcntBrackets::updateByEvent(const SIInstrInfo *TII,
507 const SIRegisterInfo *TRI,
508 const MachineRegisterInfo *MRI,
509 WaitEventType E, MachineInstr &Inst) {
510 const MachineRegisterInfo &MRIA = *MRI;
511 InstCounterType T = eventCounter(E);
512 uint32_t CurrScore = getScoreUB(T) + 1;
514 report_fatal_error("InsertWaitcnt score wraparound");
515 // PendingEvents and ScoreUB need to be update regardless if this event
516 // changes the score of a register or not.
517 // Examples including vm_cnt when buffer-store or lgkm_cnt when send-message.
518 if (!hasPendingEvent(E)) {
519 if (PendingEvents & WaitEventMaskForInst[T])
520 MixedPendingEvents[T] = true;
521 PendingEvents |= 1 << E;
523 setScoreUB(T, CurrScore);
526 // Put score on the source vgprs. If this is a store, just use those
527 // specific register(s).
528 if (TII->isDS(Inst) && (Inst.mayStore() || Inst.mayLoad())) {
529 // All GDS operations must protect their address register (same as
531 if (Inst.getOpcode() != AMDGPU::DS_APPEND &&
532 Inst.getOpcode() != AMDGPU::DS_CONSUME) {
534 &Inst, TII, TRI, MRI,
535 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::addr),
538 if (Inst.mayStore()) {
539 if (AMDGPU::getNamedOperandIdx(Inst.getOpcode(),
540 AMDGPU::OpName::data0) != -1) {
542 &Inst, TII, TRI, MRI,
543 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data0),
546 if (AMDGPU::getNamedOperandIdx(Inst.getOpcode(),
547 AMDGPU::OpName::data1) != -1) {
548 setExpScore(&Inst, TII, TRI, MRI,
549 AMDGPU::getNamedOperandIdx(Inst.getOpcode(),
550 AMDGPU::OpName::data1),
553 } else if (AMDGPU::getAtomicNoRetOp(Inst.getOpcode()) != -1 &&
554 Inst.getOpcode() != AMDGPU::DS_GWS_INIT &&
555 Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_V &&
556 Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_BR &&
557 Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_P &&
558 Inst.getOpcode() != AMDGPU::DS_GWS_BARRIER &&
559 Inst.getOpcode() != AMDGPU::DS_APPEND &&
560 Inst.getOpcode() != AMDGPU::DS_CONSUME &&
561 Inst.getOpcode() != AMDGPU::DS_ORDERED_COUNT) {
562 for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
563 const MachineOperand &Op = Inst.getOperand(I);
564 if (Op.isReg() && !Op.isDef() && TRI->isVGPR(MRIA, Op.getReg())) {
565 setExpScore(&Inst, TII, TRI, MRI, I, CurrScore);
569 } else if (TII->isFLAT(Inst)) {
570 if (Inst.mayStore()) {
572 &Inst, TII, TRI, MRI,
573 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
575 } else if (AMDGPU::getAtomicNoRetOp(Inst.getOpcode()) != -1) {
577 &Inst, TII, TRI, MRI,
578 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
581 } else if (TII->isMIMG(Inst)) {
582 if (Inst.mayStore()) {
583 setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore);
584 } else if (AMDGPU::getAtomicNoRetOp(Inst.getOpcode()) != -1) {
586 &Inst, TII, TRI, MRI,
587 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
590 } else if (TII->isMTBUF(Inst)) {
591 if (Inst.mayStore()) {
592 setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore);
594 } else if (TII->isMUBUF(Inst)) {
595 if (Inst.mayStore()) {
596 setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore);
597 } else if (AMDGPU::getAtomicNoRetOp(Inst.getOpcode()) != -1) {
599 &Inst, TII, TRI, MRI,
600 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
604 if (TII->isEXP(Inst)) {
605 // For export the destination registers are really temps that
606 // can be used as the actual source after export patching, so
607 // we need to treat them like sources and set the EXP_CNT
609 for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
610 MachineOperand &DefMO = Inst.getOperand(I);
611 if (DefMO.isReg() && DefMO.isDef() &&
612 TRI->isVGPR(MRIA, DefMO.getReg())) {
613 setRegScore(TRI->getEncodingValue(DefMO.getReg()), EXP_CNT,
618 for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
619 MachineOperand &MO = Inst.getOperand(I);
620 if (MO.isReg() && !MO.isDef() && TRI->isVGPR(MRIA, MO.getReg())) {
621 setExpScore(&Inst, TII, TRI, MRI, I, CurrScore);
625 #if 0 // TODO: check if this is handled by MUBUF code above.
626 } else if (Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORD ||
627 Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORDX2 ||
628 Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORDX4) {
629 MachineOperand *MO = TII->getNamedOperand(Inst, AMDGPU::OpName::data);
630 unsigned OpNo;//TODO: find the OpNo for this operand;
631 RegInterval Interval = getRegInterval(&Inst, TII, MRI, TRI, OpNo, false);
632 for (signed RegNo = Interval.first; RegNo < Interval.second;
634 setRegScore(RegNo + NUM_ALL_VGPRS, t, CurrScore);
638 // Match the score to the destination registers.
639 for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
640 RegInterval Interval = getRegInterval(&Inst, TII, MRI, TRI, I, true);
641 if (T == VM_CNT && Interval.first >= NUM_ALL_VGPRS)
643 for (signed RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
644 setRegScore(RegNo, T, CurrScore);
647 if (TII->isDS(Inst) && Inst.mayStore()) {
648 setRegScore(SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS, T, CurrScore);
653 void WaitcntBrackets::print(raw_ostream &OS) {
655 for (auto T : inst_counter_types()) {
656 uint32_t LB = getScoreLB(T);
657 uint32_t UB = getScoreUB(T);
661 OS << " VM_CNT(" << UB - LB << "): ";
664 OS << " LGKM_CNT(" << UB - LB << "): ";
667 OS << " EXP_CNT(" << UB - LB << "): ";
670 OS << " UNKNOWN(" << UB - LB << "): ";
675 // Print vgpr scores.
676 for (int J = 0; J <= getMaxVGPR(); J++) {
677 uint32_t RegScore = getRegScore(J, T);
680 uint32_t RelScore = RegScore - LB - 1;
681 if (J < SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS) {
682 OS << RelScore << ":v" << J << " ";
684 OS << RelScore << ":ds ";
687 // Also need to print sgpr scores for lgkm_cnt.
689 for (int J = 0; J <= getMaxSGPR(); J++) {
690 uint32_t RegScore = getRegScore(J + NUM_ALL_VGPRS, LGKM_CNT);
693 uint32_t RelScore = RegScore - LB - 1;
694 OS << RelScore << ":s" << J << " ";
703 /// Simplify the waitcnt, in the sense of removing redundant counts, and return
704 /// whether a waitcnt instruction is needed at all.
705 bool WaitcntBrackets::simplifyWaitcnt(AMDGPU::Waitcnt &Wait) const {
706 return simplifyWaitcnt(VM_CNT, Wait.VmCnt) |
707 simplifyWaitcnt(EXP_CNT, Wait.ExpCnt) |
708 simplifyWaitcnt(LGKM_CNT, Wait.LgkmCnt);
711 bool WaitcntBrackets::simplifyWaitcnt(InstCounterType T,
712 unsigned &Count) const {
713 const uint32_t LB = getScoreLB(T);
714 const uint32_t UB = getScoreUB(T);
715 if (Count < UB && UB - Count > LB)
722 void WaitcntBrackets::determineWait(InstCounterType T, uint32_t ScoreToWait,
723 AMDGPU::Waitcnt &Wait) const {
724 // If the score of src_operand falls within the bracket, we need an
725 // s_waitcnt instruction.
726 const uint32_t LB = getScoreLB(T);
727 const uint32_t UB = getScoreUB(T);
728 if ((UB >= ScoreToWait) && (ScoreToWait > LB)) {
729 if ((T == VM_CNT || T == LGKM_CNT) &&
731 !ST->hasFlatLgkmVMemCountInOrder()) {
732 // If there is a pending FLAT operation, and this is a VMem or LGKM
733 // waitcnt and the target can report early completion, then we need
734 // to force a waitcnt 0.
736 } else if (counterOutOfOrder(T)) {
737 // Counter can get decremented out-of-order when there
738 // are multiple types event in the bracket. Also emit an s_wait counter
739 // with a conservative value of 0 for the counter.
742 addWait(Wait, T, UB - ScoreToWait);
747 void WaitcntBrackets::applyWaitcnt(const AMDGPU::Waitcnt &Wait) {
748 applyWaitcnt(VM_CNT, Wait.VmCnt);
749 applyWaitcnt(EXP_CNT, Wait.ExpCnt);
750 applyWaitcnt(LGKM_CNT, Wait.LgkmCnt);
753 void WaitcntBrackets::applyWaitcnt(InstCounterType T, unsigned Count) {
754 const uint32_t UB = getScoreUB(T);
758 if (counterOutOfOrder(T))
760 setScoreLB(T, std::max(getScoreLB(T), UB - Count));
763 MixedPendingEvents[T] = false;
764 PendingEvents &= ~WaitEventMaskForInst[T];
768 // Where there are multiple types of event in the bracket of a counter,
769 // the decrement may go out of order.
770 bool WaitcntBrackets::counterOutOfOrder(InstCounterType T) const {
771 // Scalar memory read always can go out of order.
772 if (T == LGKM_CNT && hasPendingEvent(SMEM_ACCESS))
774 return MixedPendingEvents[T];
777 INITIALIZE_PASS_BEGIN(SIInsertWaitcnts, DEBUG_TYPE, "SI Insert Waitcnts", false,
779 INITIALIZE_PASS_END(SIInsertWaitcnts, DEBUG_TYPE, "SI Insert Waitcnts", false,
782 char SIInsertWaitcnts::ID = 0;
784 char &llvm::SIInsertWaitcntsID = SIInsertWaitcnts::ID;
786 FunctionPass *llvm::createSIInsertWaitcntsPass() {
787 return new SIInsertWaitcnts();
790 static bool readsVCCZ(const MachineInstr &MI) {
791 unsigned Opc = MI.getOpcode();
792 return (Opc == AMDGPU::S_CBRANCH_VCCNZ || Opc == AMDGPU::S_CBRANCH_VCCZ) &&
793 !MI.getOperand(1).isUndef();
796 /// Generate s_waitcnt instruction to be placed before cur_Inst.
797 /// Instructions of a given type are returned in order,
798 /// but instructions of different types can complete out of order.
799 /// We rely on this in-order completion
800 /// and simply assign a score to the memory access instructions.
801 /// We keep track of the active "score bracket" to determine
802 /// if an access of a memory read requires an s_waitcnt
803 /// and if so what the value of each counter is.
804 /// The "score bracket" is bound by the lower bound and upper bound
805 /// scores (*_score_LB and *_score_ub respectively).
806 bool SIInsertWaitcnts::generateWaitcntInstBefore(
807 MachineInstr &MI, WaitcntBrackets &ScoreBrackets,
808 MachineInstr *OldWaitcntInstr) {
809 setForceEmitWaitcnt();
810 bool IsForceEmitWaitcnt = isForceEmitWaitcnt();
812 if (MI.isDebugInstr())
815 AMDGPU::Waitcnt Wait;
817 // See if this instruction has a forced S_WAITCNT VM.
818 // TODO: Handle other cases of NeedsWaitcntVmBefore()
819 if (MI.getOpcode() == AMDGPU::BUFFER_WBINVL1 ||
820 MI.getOpcode() == AMDGPU::BUFFER_WBINVL1_SC ||
821 MI.getOpcode() == AMDGPU::BUFFER_WBINVL1_VOL) {
825 // All waits must be resolved at call return.
826 // NOTE: this could be improved with knowledge of all call sites or
827 // with knowledge of the called routines.
828 if (MI.getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG ||
829 MI.getOpcode() == AMDGPU::S_SETPC_B64_return) {
830 Wait = AMDGPU::Waitcnt::allZero();
832 // Resolve vm waits before gs-done.
833 else if ((MI.getOpcode() == AMDGPU::S_SENDMSG ||
834 MI.getOpcode() == AMDGPU::S_SENDMSGHALT) &&
835 ((MI.getOperand(0).getImm() & AMDGPU::SendMsg::ID_MASK_) ==
836 AMDGPU::SendMsg::ID_GS_DONE)) {
839 #if 0 // TODO: the following blocks of logic when we have fence.
840 else if (MI.getOpcode() == SC_FENCE) {
841 const unsigned int group_size =
842 context->shader_info->GetMaxThreadGroupSize();
843 // group_size == 0 means thread group size is unknown at compile time
844 const bool group_is_multi_wave =
845 (group_size == 0 || group_size > target_info->GetWaveFrontSize());
846 const bool fence_is_global = !((SCInstInternalMisc*)Inst)->IsGroupFence();
848 for (unsigned int i = 0; i < Inst->NumSrcOperands(); i++) {
849 SCRegType src_type = Inst->GetSrcType(i);
852 if (group_is_multi_wave ||
853 context->OptFlagIsOn(OPT_R1100_LDSMEM_FENCE_CHICKEN_BIT)) {
854 EmitWaitcnt |= ScoreBrackets->updateByWait(LGKM_CNT,
855 ScoreBrackets->getScoreUB(LGKM_CNT));
856 // LDS may have to wait for VM_CNT after buffer load to LDS
857 if (target_info->HasBufferLoadToLDS()) {
858 EmitWaitcnt |= ScoreBrackets->updateByWait(VM_CNT,
859 ScoreBrackets->getScoreUB(VM_CNT));
865 if (group_is_multi_wave || fence_is_global) {
866 EmitWaitcnt |= ScoreBrackets->updateByWait(EXP_CNT,
867 ScoreBrackets->getScoreUB(EXP_CNT));
868 EmitWaitcnt |= ScoreBrackets->updateByWait(LGKM_CNT,
869 ScoreBrackets->getScoreUB(LGKM_CNT));
877 if (group_is_multi_wave || fence_is_global) {
878 EmitWaitcnt |= ScoreBrackets->updateByWait(EXP_CNT,
879 ScoreBrackets->getScoreUB(EXP_CNT));
880 EmitWaitcnt |= ScoreBrackets->updateByWait(VM_CNT,
881 ScoreBrackets->getScoreUB(VM_CNT));
893 // Export & GDS instructions do not read the EXEC mask until after the export
894 // is granted (which can occur well after the instruction is issued).
895 // The shader program must flush all EXP operations on the export-count
896 // before overwriting the EXEC mask.
898 if (MI.modifiesRegister(AMDGPU::EXEC, TRI)) {
899 // Export and GDS are tracked individually, either may trigger a waitcnt
901 if (ScoreBrackets.hasPendingEvent(EXP_GPR_LOCK) ||
902 ScoreBrackets.hasPendingEvent(EXP_PARAM_ACCESS) ||
903 ScoreBrackets.hasPendingEvent(EXP_POS_ACCESS) ||
904 ScoreBrackets.hasPendingEvent(GDS_GPR_LOCK)) {
909 #if 0 // TODO: the following code to handle CALL.
910 // The argument passing for CALLs should suffice for VM_CNT and LGKM_CNT.
911 // However, there is a problem with EXP_CNT, because the call cannot
912 // easily tell if a register is used in the function, and if it did, then
913 // the referring instruction would have to have an S_WAITCNT, which is
914 // dependent on all call sites. So Instead, force S_WAITCNT for EXP_CNTs
916 if (MI.getOpcode() == SC_CALL) {
917 if (ScoreBrackets->getScoreUB(EXP_CNT) >
918 ScoreBrackets->getScoreLB(EXP_CNT)) {
919 ScoreBrackets->setScoreLB(EXP_CNT, ScoreBrackets->getScoreUB(EXP_CNT));
920 EmitWaitcnt |= CNT_MASK(EXP_CNT);
925 // FIXME: Should not be relying on memoperands.
926 // Look at the source operands of every instruction to see if
927 // any of them results from a previous memory operation that affects
928 // its current usage. If so, an s_waitcnt instruction needs to be
930 // If the source operand was defined by a load, add the s_waitcnt
932 for (const MachineMemOperand *Memop : MI.memoperands()) {
933 unsigned AS = Memop->getAddrSpace();
934 if (AS != AMDGPUAS::LOCAL_ADDRESS)
936 unsigned RegNo = SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS;
937 // VM_CNT is only relevant to vgpr or LDS.
938 ScoreBrackets.determineWait(
939 VM_CNT, ScoreBrackets.getRegScore(RegNo, VM_CNT), Wait);
942 for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) {
943 const MachineOperand &Op = MI.getOperand(I);
944 const MachineRegisterInfo &MRIA = *MRI;
945 RegInterval Interval =
946 ScoreBrackets.getRegInterval(&MI, TII, MRI, TRI, I, false);
947 for (signed RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
948 if (TRI->isVGPR(MRIA, Op.getReg())) {
949 // VM_CNT is only relevant to vgpr or LDS.
950 ScoreBrackets.determineWait(
951 VM_CNT, ScoreBrackets.getRegScore(RegNo, VM_CNT), Wait);
953 ScoreBrackets.determineWait(
954 LGKM_CNT, ScoreBrackets.getRegScore(RegNo, LGKM_CNT), Wait);
957 // End of for loop that looks at all source operands to decide vm_wait_cnt
960 // Two cases are handled for destination operands:
961 // 1) If the destination operand was defined by a load, add the s_waitcnt
962 // instruction to guarantee the right WAW order.
963 // 2) If a destination operand that was used by a recent export/store ins,
964 // add s_waitcnt on exp_cnt to guarantee the WAR order.
966 // FIXME: Should not be relying on memoperands.
967 for (const MachineMemOperand *Memop : MI.memoperands()) {
968 unsigned AS = Memop->getAddrSpace();
969 if (AS != AMDGPUAS::LOCAL_ADDRESS)
971 unsigned RegNo = SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS;
972 ScoreBrackets.determineWait(
973 VM_CNT, ScoreBrackets.getRegScore(RegNo, VM_CNT), Wait);
974 ScoreBrackets.determineWait(
975 EXP_CNT, ScoreBrackets.getRegScore(RegNo, EXP_CNT), Wait);
978 for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) {
979 MachineOperand &Def = MI.getOperand(I);
980 const MachineRegisterInfo &MRIA = *MRI;
981 RegInterval Interval =
982 ScoreBrackets.getRegInterval(&MI, TII, MRI, TRI, I, true);
983 for (signed RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
984 if (TRI->isVGPR(MRIA, Def.getReg())) {
985 ScoreBrackets.determineWait(
986 VM_CNT, ScoreBrackets.getRegScore(RegNo, VM_CNT), Wait);
987 ScoreBrackets.determineWait(
988 EXP_CNT, ScoreBrackets.getRegScore(RegNo, EXP_CNT), Wait);
990 ScoreBrackets.determineWait(
991 LGKM_CNT, ScoreBrackets.getRegScore(RegNo, LGKM_CNT), Wait);
993 } // End of for loop that looks at all dest operands.
996 // Check to see if this is an S_BARRIER, and if an implicit S_WAITCNT 0
997 // occurs before the instruction. Doing it here prevents any additional
998 // S_WAITCNTs from being emitted if the instruction was marked as
999 // requiring a WAITCNT beforehand.
1000 if (MI.getOpcode() == AMDGPU::S_BARRIER &&
1001 !ST->hasAutoWaitcntBeforeBarrier()) {
1002 Wait = AMDGPU::Waitcnt::allZero();
1005 // TODO: Remove this work-around, enable the assert for Bug 457939
1006 // after fixing the scheduler. Also, the Shader Compiler code is
1007 // independent of target.
1008 if (readsVCCZ(MI) && ST->getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS) {
1009 if (ScoreBrackets.getScoreLB(LGKM_CNT) <
1010 ScoreBrackets.getScoreUB(LGKM_CNT) &&
1011 ScoreBrackets.hasPendingEvent(SMEM_ACCESS)) {
1016 // Early-out if no wait is indicated.
1017 if (!ScoreBrackets.simplifyWaitcnt(Wait) && !IsForceEmitWaitcnt) {
1018 bool Modified = false;
1019 if (OldWaitcntInstr) {
1020 if (TrackedWaitcntSet.count(OldWaitcntInstr)) {
1021 TrackedWaitcntSet.erase(OldWaitcntInstr);
1022 OldWaitcntInstr->eraseFromParent();
1025 int64_t Imm = OldWaitcntInstr->getOperand(0).getImm();
1026 ScoreBrackets.applyWaitcnt(AMDGPU::decodeWaitcnt(IV, Imm));
1033 if (ForceEmitZeroWaitcnts)
1034 Wait = AMDGPU::Waitcnt::allZero();
1036 if (ForceEmitWaitcnt[VM_CNT])
1038 if (ForceEmitWaitcnt[EXP_CNT])
1040 if (ForceEmitWaitcnt[LGKM_CNT])
1043 ScoreBrackets.applyWaitcnt(Wait);
1045 AMDGPU::Waitcnt OldWait;
1046 if (OldWaitcntInstr) {
1048 AMDGPU::decodeWaitcnt(IV, OldWaitcntInstr->getOperand(0).getImm());
1050 if (OldWait.dominates(Wait))
1053 if (OldWaitcntInstr && !TrackedWaitcntSet.count(OldWaitcntInstr))
1054 Wait = Wait.combined(OldWait);
1056 unsigned Enc = AMDGPU::encodeWaitcnt(IV, Wait);
1057 if (OldWaitcntInstr) {
1058 OldWaitcntInstr->getOperand(0).setImm(Enc);
1060 LLVM_DEBUG(dbgs() << "updateWaitcntInBlock\n"
1061 << "Old Instr: " << MI << '\n'
1062 << "New Instr: " << *OldWaitcntInstr << '\n');
1064 auto SWaitInst = BuildMI(*MI.getParent(), MI.getIterator(),
1065 MI.getDebugLoc(), TII->get(AMDGPU::S_WAITCNT))
1067 TrackedWaitcntSet.insert(SWaitInst);
1069 LLVM_DEBUG(dbgs() << "insertWaitcntInBlock\n"
1070 << "Old Instr: " << MI << '\n'
1071 << "New Instr: " << *SWaitInst << '\n');
1077 // This is a flat memory operation. Check to see if it has memory
1078 // tokens for both LDS and Memory, and if so mark it as a flat.
1079 bool SIInsertWaitcnts::mayAccessLDSThroughFlat(const MachineInstr &MI) const {
1080 if (MI.memoperands_empty())
1083 for (const MachineMemOperand *Memop : MI.memoperands()) {
1084 unsigned AS = Memop->getAddrSpace();
1085 if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS)
1092 void SIInsertWaitcnts::updateEventWaitcntAfter(MachineInstr &Inst,
1093 WaitcntBrackets *ScoreBrackets) {
1094 // Now look at the instruction opcode. If it is a memory access
1095 // instruction, update the upper-bound of the appropriate counter's
1096 // bracket and the destination operand scores.
1097 // TODO: Use the (TSFlags & SIInstrFlags::LGKM_CNT) property everywhere.
1098 if (TII->isDS(Inst) && TII->usesLGKM_CNT(Inst)) {
1099 if (TII->isAlwaysGDS(Inst.getOpcode()) ||
1100 TII->hasModifiersSet(Inst, AMDGPU::OpName::gds)) {
1101 ScoreBrackets->updateByEvent(TII, TRI, MRI, GDS_ACCESS, Inst);
1102 ScoreBrackets->updateByEvent(TII, TRI, MRI, GDS_GPR_LOCK, Inst);
1104 ScoreBrackets->updateByEvent(TII, TRI, MRI, LDS_ACCESS, Inst);
1106 } else if (TII->isFLAT(Inst)) {
1107 assert(Inst.mayLoad() || Inst.mayStore());
1109 if (TII->usesVM_CNT(Inst))
1110 ScoreBrackets->updateByEvent(TII, TRI, MRI, VMEM_ACCESS, Inst);
1112 if (TII->usesLGKM_CNT(Inst)) {
1113 ScoreBrackets->updateByEvent(TII, TRI, MRI, LDS_ACCESS, Inst);
1115 // This is a flat memory operation, so note it - it will require
1116 // that both the VM and LGKM be flushed to zero if it is pending when
1117 // a VM or LGKM dependency occurs.
1118 if (mayAccessLDSThroughFlat(Inst))
1119 ScoreBrackets->setPendingFlat();
1121 } else if (SIInstrInfo::isVMEM(Inst) &&
1122 // TODO: get a better carve out.
1123 Inst.getOpcode() != AMDGPU::BUFFER_WBINVL1 &&
1124 Inst.getOpcode() != AMDGPU::BUFFER_WBINVL1_SC &&
1125 Inst.getOpcode() != AMDGPU::BUFFER_WBINVL1_VOL) {
1126 ScoreBrackets->updateByEvent(TII, TRI, MRI, VMEM_ACCESS, Inst);
1127 if (ST->vmemWriteNeedsExpWaitcnt() &&
1128 (Inst.mayStore() || AMDGPU::getAtomicNoRetOp(Inst.getOpcode()) != -1)) {
1129 ScoreBrackets->updateByEvent(TII, TRI, MRI, VMW_GPR_LOCK, Inst);
1131 } else if (TII->isSMRD(Inst)) {
1132 ScoreBrackets->updateByEvent(TII, TRI, MRI, SMEM_ACCESS, Inst);
1134 switch (Inst.getOpcode()) {
1135 case AMDGPU::S_SENDMSG:
1136 case AMDGPU::S_SENDMSGHALT:
1137 ScoreBrackets->updateByEvent(TII, TRI, MRI, SQ_MESSAGE, Inst);
1140 case AMDGPU::EXP_DONE: {
1141 int Imm = TII->getNamedOperand(Inst, AMDGPU::OpName::tgt)->getImm();
1142 if (Imm >= 32 && Imm <= 63)
1143 ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_PARAM_ACCESS, Inst);
1144 else if (Imm >= 12 && Imm <= 15)
1145 ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_POS_ACCESS, Inst);
1147 ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_GPR_LOCK, Inst);
1150 case AMDGPU::S_MEMTIME:
1151 case AMDGPU::S_MEMREALTIME:
1152 ScoreBrackets->updateByEvent(TII, TRI, MRI, SMEM_ACCESS, Inst);
1160 bool WaitcntBrackets::mergeScore(const MergeInfo &M, uint32_t &Score,
1161 uint32_t OtherScore) {
1162 uint32_t MyShifted = Score <= M.OldLB ? 0 : Score + M.MyShift;
1163 uint32_t OtherShifted =
1164 OtherScore <= M.OtherLB ? 0 : OtherScore + M.OtherShift;
1165 Score = std::max(MyShifted, OtherShifted);
1166 return OtherShifted > MyShifted;
1169 /// Merge the pending events and associater score brackets of \p Other into
1170 /// this brackets status.
1172 /// Returns whether the merge resulted in a change that requires tighter waits
1173 /// (i.e. the merged brackets strictly dominate the original brackets).
1174 bool WaitcntBrackets::merge(const WaitcntBrackets &Other) {
1175 bool StrictDom = false;
1177 for (auto T : inst_counter_types()) {
1178 // Merge event flags for this counter
1179 const bool OldOutOfOrder = counterOutOfOrder(T);
1180 const uint32_t OldEvents = PendingEvents & WaitEventMaskForInst[T];
1181 const uint32_t OtherEvents = Other.PendingEvents & WaitEventMaskForInst[T];
1182 if (OtherEvents & ~OldEvents)
1184 if (Other.MixedPendingEvents[T] ||
1185 (OldEvents && OtherEvents && OldEvents != OtherEvents))
1186 MixedPendingEvents[T] = true;
1187 PendingEvents |= OtherEvents;
1189 // Merge scores for this counter
1190 const uint32_t MyPending = ScoreUBs[T] - ScoreLBs[T];
1191 const uint32_t OtherPending = Other.ScoreUBs[T] - Other.ScoreLBs[T];
1193 M.OldLB = ScoreLBs[T];
1194 M.OtherLB = Other.ScoreLBs[T];
1195 M.MyShift = OtherPending > MyPending ? OtherPending - MyPending : 0;
1196 M.OtherShift = ScoreUBs[T] - Other.ScoreUBs[T] + M.MyShift;
1198 const uint32_t NewUB = ScoreUBs[T] + M.MyShift;
1199 if (NewUB < ScoreUBs[T])
1200 report_fatal_error("waitcnt score overflow");
1201 ScoreUBs[T] = NewUB;
1202 ScoreLBs[T] = std::min(M.OldLB + M.MyShift, M.OtherLB + M.OtherShift);
1204 StrictDom |= mergeScore(M, LastFlat[T], Other.LastFlat[T]);
1206 bool RegStrictDom = false;
1207 for (int J = 0, E = std::max(getMaxVGPR(), Other.getMaxVGPR()) + 1; J != E;
1209 RegStrictDom |= mergeScore(M, VgprScores[T][J], Other.VgprScores[T][J]);
1212 if (T == LGKM_CNT) {
1213 for (int J = 0, E = std::max(getMaxSGPR(), Other.getMaxSGPR()) + 1;
1215 RegStrictDom |= mergeScore(M, SgprScores[J], Other.SgprScores[J]);
1219 if (RegStrictDom && !OldOutOfOrder)
1223 VgprUB = std::max(getMaxVGPR(), Other.getMaxVGPR());
1224 SgprUB = std::max(getMaxSGPR(), Other.getMaxSGPR());
1229 // Generate s_waitcnt instructions where needed.
1230 bool SIInsertWaitcnts::insertWaitcntInBlock(MachineFunction &MF,
1231 MachineBasicBlock &Block,
1232 WaitcntBrackets &ScoreBrackets) {
1233 bool Modified = false;
1236 dbgs() << "*** Block" << Block.getNumber() << " ***";
1237 ScoreBrackets.dump();
1240 // Walk over the instructions.
1241 MachineInstr *OldWaitcntInstr = nullptr;
1243 for (MachineBasicBlock::iterator Iter = Block.begin(), E = Block.end();
1245 MachineInstr &Inst = *Iter;
1247 // Remove any previously existing waitcnts.
1248 if (Inst.getOpcode() == AMDGPU::S_WAITCNT) {
1249 if (OldWaitcntInstr) {
1250 if (TrackedWaitcntSet.count(OldWaitcntInstr)) {
1251 TrackedWaitcntSet.erase(OldWaitcntInstr);
1252 OldWaitcntInstr->eraseFromParent();
1253 OldWaitcntInstr = nullptr;
1254 } else if (!TrackedWaitcntSet.count(&Inst)) {
1255 // Two successive s_waitcnt's, both of which are pre-existing and
1256 // are therefore preserved.
1257 int64_t Imm = OldWaitcntInstr->getOperand(0).getImm();
1258 ScoreBrackets.applyWaitcnt(AMDGPU::decodeWaitcnt(IV, Imm));
1261 Inst.eraseFromParent();
1267 OldWaitcntInstr = &Inst;
1272 bool VCCZBugWorkAround = false;
1273 if (readsVCCZ(Inst) &&
1274 (!VCCZBugHandledSet.count(&Inst))) {
1275 if (ScoreBrackets.getScoreLB(LGKM_CNT) <
1276 ScoreBrackets.getScoreUB(LGKM_CNT) &&
1277 ScoreBrackets.hasPendingEvent(SMEM_ACCESS)) {
1278 if (ST->getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS)
1279 VCCZBugWorkAround = true;
1283 // Generate an s_waitcnt instruction to be placed before
1284 // cur_Inst, if needed.
1285 Modified |= generateWaitcntInstBefore(Inst, ScoreBrackets, OldWaitcntInstr);
1286 OldWaitcntInstr = nullptr;
1288 updateEventWaitcntAfter(Inst, &ScoreBrackets);
1290 #if 0 // TODO: implement resource type check controlled by options with ub = LB.
1291 // If this instruction generates a S_SETVSKIP because it is an
1292 // indexed resource, and we are on Tahiti, then it will also force
1293 // an S_WAITCNT vmcnt(0)
1294 if (RequireCheckResourceType(Inst, context)) {
1295 // Force the score to as if an S_WAITCNT vmcnt(0) is emitted.
1296 ScoreBrackets->setScoreLB(VM_CNT,
1297 ScoreBrackets->getScoreUB(VM_CNT));
1303 ScoreBrackets.dump();
1306 // Check to see if this is a GWS instruction. If so, and if this is CI or
1307 // VI, then the generated code sequence will include an S_WAITCNT 0.
1308 // TODO: Are these the only GWS instructions?
1309 if (Inst.getOpcode() == AMDGPU::DS_GWS_INIT ||
1310 Inst.getOpcode() == AMDGPU::DS_GWS_SEMA_V ||
1311 Inst.getOpcode() == AMDGPU::DS_GWS_SEMA_BR ||
1312 Inst.getOpcode() == AMDGPU::DS_GWS_SEMA_P ||
1313 Inst.getOpcode() == AMDGPU::DS_GWS_BARRIER) {
1314 // TODO: && context->target_info->GwsRequiresMemViolTest() ) {
1315 ScoreBrackets.applyWaitcnt(AMDGPU::Waitcnt::allZero());
1318 // TODO: Remove this work-around after fixing the scheduler and enable the
1320 if (VCCZBugWorkAround) {
1321 // Restore the vccz bit. Any time a value is written to vcc, the vcc
1322 // bit is updated, so we can restore the bit by reading the value of
1323 // vcc and then writing it back to the register.
1324 BuildMI(Block, Inst, Inst.getDebugLoc(), TII->get(AMDGPU::S_MOV_B64),
1326 .addReg(AMDGPU::VCC);
1327 VCCZBugHandledSet.insert(&Inst);
1337 bool SIInsertWaitcnts::runOnMachineFunction(MachineFunction &MF) {
1338 ST = &MF.getSubtarget<GCNSubtarget>();
1339 TII = ST->getInstrInfo();
1340 TRI = &TII->getRegisterInfo();
1341 MRI = &MF.getRegInfo();
1342 IV = AMDGPU::getIsaVersion(ST->getCPU());
1343 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1345 ForceEmitZeroWaitcnts = ForceEmitZeroFlag;
1346 for (auto T : inst_counter_types())
1347 ForceEmitWaitcnt[T] = false;
1349 HardwareLimits.VmcntMax = AMDGPU::getVmcntBitMask(IV);
1350 HardwareLimits.ExpcntMax = AMDGPU::getExpcntBitMask(IV);
1351 HardwareLimits.LgkmcntMax = AMDGPU::getLgkmcntBitMask(IV);
1353 HardwareLimits.NumVGPRsMax = ST->getAddressableNumVGPRs();
1354 HardwareLimits.NumSGPRsMax = ST->getAddressableNumSGPRs();
1355 assert(HardwareLimits.NumVGPRsMax <= SQ_MAX_PGM_VGPRS);
1356 assert(HardwareLimits.NumSGPRsMax <= SQ_MAX_PGM_SGPRS);
1358 RegisterEncoding.VGPR0 = TRI->getEncodingValue(AMDGPU::VGPR0);
1359 RegisterEncoding.VGPRL =
1360 RegisterEncoding.VGPR0 + HardwareLimits.NumVGPRsMax - 1;
1361 RegisterEncoding.SGPR0 = TRI->getEncodingValue(AMDGPU::SGPR0);
1362 RegisterEncoding.SGPRL =
1363 RegisterEncoding.SGPR0 + HardwareLimits.NumSGPRsMax - 1;
1365 TrackedWaitcntSet.clear();
1366 VCCZBugHandledSet.clear();
1370 // Keep iterating over the blocks in reverse post order, inserting and
1371 // updating s_waitcnt where needed, until a fix point is reached.
1372 for (MachineBasicBlock *MBB :
1373 ReversePostOrderTraversal<MachineFunction *>(&MF)) {
1374 RpotIdxMap[MBB] = BlockInfos.size();
1375 BlockInfos.emplace_back(MBB);
1378 std::unique_ptr<WaitcntBrackets> Brackets;
1379 bool Modified = false;
1384 for (BlockInfo &BI : BlockInfos) {
1388 unsigned Idx = std::distance(&*BlockInfos.begin(), &BI);
1392 Brackets = llvm::make_unique<WaitcntBrackets>(*BI.Incoming);
1394 *Brackets = *BI.Incoming;
1397 Brackets = llvm::make_unique<WaitcntBrackets>(ST);
1402 Modified |= insertWaitcntInBlock(MF, *BI.MBB, *Brackets);
1405 if (Brackets->hasPending()) {
1406 BlockInfo *MoveBracketsToSucc = nullptr;
1407 for (MachineBasicBlock *Succ : BI.MBB->successors()) {
1408 unsigned SuccIdx = RpotIdxMap[Succ];
1409 BlockInfo &SuccBI = BlockInfos[SuccIdx];
1410 if (!SuccBI.Incoming) {
1411 SuccBI.Dirty = true;
1414 if (!MoveBracketsToSucc) {
1415 MoveBracketsToSucc = &SuccBI;
1417 SuccBI.Incoming = llvm::make_unique<WaitcntBrackets>(*Brackets);
1419 } else if (SuccBI.Incoming->merge(*Brackets)) {
1420 SuccBI.Dirty = true;
1425 if (MoveBracketsToSucc)
1426 MoveBracketsToSucc->Incoming = std::move(Brackets);
1431 SmallVector<MachineBasicBlock *, 4> EndPgmBlocks;
1433 bool HaveScalarStores = false;
1435 for (MachineFunction::iterator BI = MF.begin(), BE = MF.end(); BI != BE;
1437 MachineBasicBlock &MBB = *BI;
1439 for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end(); I != E;
1441 if (!HaveScalarStores && TII->isScalarStore(*I))
1442 HaveScalarStores = true;
1444 if (I->getOpcode() == AMDGPU::S_ENDPGM ||
1445 I->getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG)
1446 EndPgmBlocks.push_back(&MBB);
1450 if (HaveScalarStores) {
1451 // If scalar writes are used, the cache must be flushed or else the next
1452 // wave to reuse the same scratch memory can be clobbered.
1454 // Insert s_dcache_wb at wave termination points if there were any scalar
1455 // stores, and only if the cache hasn't already been flushed. This could be
1456 // improved by looking across blocks for flushes in postdominating blocks
1457 // from the stores but an explicitly requested flush is probably very rare.
1458 for (MachineBasicBlock *MBB : EndPgmBlocks) {
1459 bool SeenDCacheWB = false;
1461 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E;
1463 if (I->getOpcode() == AMDGPU::S_DCACHE_WB)
1464 SeenDCacheWB = true;
1465 else if (TII->isScalarStore(*I))
1466 SeenDCacheWB = false;
1468 // FIXME: It would be better to insert this before a waitcnt if any.
1469 if ((I->getOpcode() == AMDGPU::S_ENDPGM ||
1470 I->getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG) &&
1473 BuildMI(*MBB, I, I->getDebugLoc(), TII->get(AMDGPU::S_DCACHE_WB));
1479 if (!MFI->isEntryFunction()) {
1480 // Wait for any outstanding memory operations that the input registers may
1481 // depend on. We can't track them and it's better to the wait after the
1482 // costly call sequence.
1484 // TODO: Could insert earlier and schedule more liberally with operations
1485 // that only use caller preserved registers.
1486 MachineBasicBlock &EntryBB = MF.front();
1487 BuildMI(EntryBB, EntryBB.getFirstNonPHI(), DebugLoc(), TII->get(AMDGPU::S_WAITCNT))