xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp (revision 7ef62cebc2f965b0f640263e179276928885e33d)
1 //===- SIInsertWaitcnts.cpp - Insert Wait Instructions --------------------===//
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 /// Insert wait instructions for memory reads and writes.
11 ///
12 /// Memory reads and writes are issued asynchronously, so we need to insert
13 /// S_WAITCNT instructions when we want to access any of their results or
14 /// overwrite any register that's used asynchronously.
15 ///
16 /// TODO: This pass currently keeps one timeline per hardware counter. A more
17 /// finely-grained approach that keeps one timeline per event type could
18 /// sometimes get away with generating weaker s_waitcnt instructions. For
19 /// example, when both SMEM and LDS are in flight and we need to wait for
20 /// the i-th-last LDS instruction, then an lgkmcnt(i) is actually sufficient,
21 /// but the pass will currently generate a conservative lgkmcnt(0) because
22 /// multiple event types are in flight.
23 //
24 //===----------------------------------------------------------------------===//
25 
26 #include "AMDGPU.h"
27 #include "GCNSubtarget.h"
28 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
29 #include "SIMachineFunctionInfo.h"
30 #include "Utils/AMDGPUBaseInfo.h"
31 #include "llvm/ADT/MapVector.h"
32 #include "llvm/ADT/PostOrderIterator.h"
33 #include "llvm/ADT/Sequence.h"
34 #include "llvm/CodeGen/MachineLoopInfo.h"
35 #include "llvm/CodeGen/MachinePostDominators.h"
36 #include "llvm/InitializePasses.h"
37 #include "llvm/Support/DebugCounter.h"
38 #include "llvm/Support/TargetParser.h"
39 using namespace llvm;
40 
41 #define DEBUG_TYPE "si-insert-waitcnts"
42 
43 DEBUG_COUNTER(ForceExpCounter, DEBUG_TYPE"-forceexp",
44               "Force emit s_waitcnt expcnt(0) instrs");
45 DEBUG_COUNTER(ForceLgkmCounter, DEBUG_TYPE"-forcelgkm",
46               "Force emit s_waitcnt lgkmcnt(0) instrs");
47 DEBUG_COUNTER(ForceVMCounter, DEBUG_TYPE"-forcevm",
48               "Force emit s_waitcnt vmcnt(0) instrs");
49 
50 static cl::opt<bool> ForceEmitZeroFlag(
51   "amdgpu-waitcnt-forcezero",
52   cl::desc("Force all waitcnt instrs to be emitted as s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)"),
53   cl::init(false), cl::Hidden);
54 
55 namespace {
56 // Class of object that encapsulates latest instruction counter score
57 // associated with the operand.  Used for determining whether
58 // s_waitcnt instruction needs to be emitted.
59 
60 #define CNT_MASK(t) (1u << (t))
61 
62 enum InstCounterType { VM_CNT = 0, LGKM_CNT, EXP_CNT, VS_CNT, NUM_INST_CNTS };
63 } // namespace
64 
65 namespace llvm {
66 template <> struct enum_iteration_traits<InstCounterType> {
67   static constexpr bool is_iterable = true;
68 };
69 } // namespace llvm
70 
71 namespace {
72 auto inst_counter_types() { return enum_seq(VM_CNT, NUM_INST_CNTS); }
73 
74 using RegInterval = std::pair<int, int>;
75 
76 struct HardwareLimits {
77   unsigned VmcntMax;
78   unsigned ExpcntMax;
79   unsigned LgkmcntMax;
80   unsigned VscntMax;
81 };
82 
83 struct RegisterEncoding {
84   unsigned VGPR0;
85   unsigned VGPRL;
86   unsigned SGPR0;
87   unsigned SGPRL;
88 };
89 
90 enum WaitEventType {
91   VMEM_ACCESS,       // vector-memory read & write
92   VMEM_READ_ACCESS,  // vector-memory read
93   VMEM_WRITE_ACCESS, // vector-memory write
94   LDS_ACCESS,        // lds read & write
95   GDS_ACCESS,        // gds read & write
96   SQ_MESSAGE,        // send message
97   SMEM_ACCESS,       // scalar-memory read & write
98   EXP_GPR_LOCK,      // export holding on its data src
99   GDS_GPR_LOCK,      // GDS holding on its data and addr src
100   EXP_POS_ACCESS,    // write to export position
101   EXP_PARAM_ACCESS,  // write to export parameter
102   VMW_GPR_LOCK,      // vector-memory write holding on its data src
103   EXP_LDS_ACCESS,    // read by ldsdir counting as export
104   NUM_WAIT_EVENTS,
105 };
106 
107 static const unsigned WaitEventMaskForInst[NUM_INST_CNTS] = {
108     (1 << VMEM_ACCESS) | (1 << VMEM_READ_ACCESS),
109     (1 << SMEM_ACCESS) | (1 << LDS_ACCESS) | (1 << GDS_ACCESS) |
110         (1 << SQ_MESSAGE),
111     (1 << EXP_GPR_LOCK) | (1 << GDS_GPR_LOCK) | (1 << VMW_GPR_LOCK) |
112         (1 << EXP_PARAM_ACCESS) | (1 << EXP_POS_ACCESS) | (1 << EXP_LDS_ACCESS),
113     (1 << VMEM_WRITE_ACCESS)};
114 
115 // The mapping is:
116 //  0                .. SQ_MAX_PGM_VGPRS-1               real VGPRs
117 //  SQ_MAX_PGM_VGPRS .. NUM_ALL_VGPRS-1                  extra VGPR-like slots
118 //  NUM_ALL_VGPRS    .. NUM_ALL_VGPRS+SQ_MAX_PGM_SGPRS-1 real SGPRs
119 // We reserve a fixed number of VGPR slots in the scoring tables for
120 // special tokens like SCMEM_LDS (needed for buffer load to LDS).
121 enum RegisterMapping {
122   SQ_MAX_PGM_VGPRS = 512, // Maximum programmable VGPRs across all targets.
123   AGPR_OFFSET = 256,      // Maximum programmable ArchVGPRs across all targets.
124   SQ_MAX_PGM_SGPRS = 256, // Maximum programmable SGPRs across all targets.
125   NUM_EXTRA_VGPRS = 1,    // A reserved slot for DS.
126   EXTRA_VGPR_LDS = 0,     // An artificial register to track LDS writes.
127   NUM_ALL_VGPRS = SQ_MAX_PGM_VGPRS + NUM_EXTRA_VGPRS, // Where SGPR starts.
128 };
129 
130 // Enumerate different types of result-returning VMEM operations. Although
131 // s_waitcnt orders them all with a single vmcnt counter, in the absence of
132 // s_waitcnt only instructions of the same VmemType are guaranteed to write
133 // their results in order -- so there is no need to insert an s_waitcnt between
134 // two instructions of the same type that write the same vgpr.
135 enum VmemType {
136   // BUF instructions and MIMG instructions without a sampler.
137   VMEM_NOSAMPLER,
138   // MIMG instructions with a sampler.
139   VMEM_SAMPLER,
140   // BVH instructions
141   VMEM_BVH
142 };
143 
144 static bool updateVMCntOnly(const MachineInstr &Inst) {
145   return SIInstrInfo::isVMEM(Inst) || SIInstrInfo::isFLATGlobal(Inst) ||
146          SIInstrInfo::isFLATScratch(Inst);
147 }
148 
149 VmemType getVmemType(const MachineInstr &Inst) {
150   assert(updateVMCntOnly(Inst));
151   if (!SIInstrInfo::isMIMG(Inst))
152     return VMEM_NOSAMPLER;
153   const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(Inst.getOpcode());
154   const AMDGPU::MIMGBaseOpcodeInfo *BaseInfo =
155       AMDGPU::getMIMGBaseOpcodeInfo(Info->BaseOpcode);
156   return BaseInfo->BVH ? VMEM_BVH
157                        : BaseInfo->Sampler ? VMEM_SAMPLER : VMEM_NOSAMPLER;
158 }
159 
160 void addWait(AMDGPU::Waitcnt &Wait, InstCounterType T, unsigned Count) {
161   switch (T) {
162   case VM_CNT:
163     Wait.VmCnt = std::min(Wait.VmCnt, Count);
164     break;
165   case EXP_CNT:
166     Wait.ExpCnt = std::min(Wait.ExpCnt, Count);
167     break;
168   case LGKM_CNT:
169     Wait.LgkmCnt = std::min(Wait.LgkmCnt, Count);
170     break;
171   case VS_CNT:
172     Wait.VsCnt = std::min(Wait.VsCnt, Count);
173     break;
174   default:
175     llvm_unreachable("bad InstCounterType");
176   }
177 }
178 
179 // This objects maintains the current score brackets of each wait counter, and
180 // a per-register scoreboard for each wait counter.
181 //
182 // We also maintain the latest score for every event type that can change the
183 // waitcnt in order to know if there are multiple types of events within
184 // the brackets. When multiple types of event happen in the bracket,
185 // wait count may get decreased out of order, therefore we need to put in
186 // "s_waitcnt 0" before use.
187 class WaitcntBrackets {
188 public:
189   WaitcntBrackets(const GCNSubtarget *SubTarget, HardwareLimits Limits,
190                   RegisterEncoding Encoding)
191       : ST(SubTarget), Limits(Limits), Encoding(Encoding) {}
192 
193   unsigned getWaitCountMax(InstCounterType T) const {
194     switch (T) {
195     case VM_CNT:
196       return Limits.VmcntMax;
197     case LGKM_CNT:
198       return Limits.LgkmcntMax;
199     case EXP_CNT:
200       return Limits.ExpcntMax;
201     case VS_CNT:
202       return Limits.VscntMax;
203     default:
204       break;
205     }
206     return 0;
207   }
208 
209   unsigned getScoreLB(InstCounterType T) const {
210     assert(T < NUM_INST_CNTS);
211     return ScoreLBs[T];
212   }
213 
214   unsigned getScoreUB(InstCounterType T) const {
215     assert(T < NUM_INST_CNTS);
216     return ScoreUBs[T];
217   }
218 
219   unsigned getScoreRange(InstCounterType T) const {
220     return getScoreUB(T) - getScoreLB(T);
221   }
222 
223   // Mapping from event to counter.
224   InstCounterType eventCounter(WaitEventType E) const {
225     for (auto T : inst_counter_types()) {
226       if (WaitEventMaskForInst[T] & (1 << E))
227         return T;
228     }
229     llvm_unreachable("event type has no associated counter");
230   }
231 
232   unsigned getRegScore(int GprNo, InstCounterType T) const {
233     if (GprNo < NUM_ALL_VGPRS) {
234       return VgprScores[T][GprNo];
235     }
236     assert(T == LGKM_CNT);
237     return SgprScores[GprNo - NUM_ALL_VGPRS];
238   }
239 
240   bool merge(const WaitcntBrackets &Other);
241 
242   RegInterval getRegInterval(const MachineInstr *MI, const SIInstrInfo *TII,
243                              const MachineRegisterInfo *MRI,
244                              const SIRegisterInfo *TRI, unsigned OpNo) const;
245 
246   bool counterOutOfOrder(InstCounterType T) const;
247   void simplifyWaitcnt(AMDGPU::Waitcnt &Wait) const;
248   void simplifyWaitcnt(InstCounterType T, unsigned &Count) const;
249   void determineWait(InstCounterType T, int RegNo, AMDGPU::Waitcnt &Wait) const;
250   void applyWaitcnt(const AMDGPU::Waitcnt &Wait);
251   void applyWaitcnt(InstCounterType T, unsigned Count);
252   void updateByEvent(const SIInstrInfo *TII, const SIRegisterInfo *TRI,
253                      const MachineRegisterInfo *MRI, WaitEventType E,
254                      MachineInstr &MI);
255 
256   unsigned hasPendingEvent() const { return PendingEvents; }
257   unsigned hasPendingEvent(WaitEventType E) const {
258     return PendingEvents & (1 << E);
259   }
260   unsigned hasPendingEvent(InstCounterType T) const {
261     unsigned HasPending = PendingEvents & WaitEventMaskForInst[T];
262     assert((HasPending != 0) == (getScoreRange(T) != 0));
263     return HasPending;
264   }
265 
266   bool hasMixedPendingEvents(InstCounterType T) const {
267     unsigned Events = hasPendingEvent(T);
268     // Return true if more than one bit is set in Events.
269     return Events & (Events - 1);
270   }
271 
272   bool hasPendingFlat() const {
273     return ((LastFlat[LGKM_CNT] > ScoreLBs[LGKM_CNT] &&
274              LastFlat[LGKM_CNT] <= ScoreUBs[LGKM_CNT]) ||
275             (LastFlat[VM_CNT] > ScoreLBs[VM_CNT] &&
276              LastFlat[VM_CNT] <= ScoreUBs[VM_CNT]));
277   }
278 
279   void setPendingFlat() {
280     LastFlat[VM_CNT] = ScoreUBs[VM_CNT];
281     LastFlat[LGKM_CNT] = ScoreUBs[LGKM_CNT];
282   }
283 
284   // Return true if there might be pending writes to the specified vgpr by VMEM
285   // instructions with types different from V.
286   bool hasOtherPendingVmemTypes(int GprNo, VmemType V) const {
287     assert(GprNo < NUM_ALL_VGPRS);
288     return VgprVmemTypes[GprNo] & ~(1 << V);
289   }
290 
291   void clearVgprVmemTypes(int GprNo) {
292     assert(GprNo < NUM_ALL_VGPRS);
293     VgprVmemTypes[GprNo] = 0;
294   }
295 
296   void print(raw_ostream &);
297   void dump() { print(dbgs()); }
298 
299 private:
300   struct MergeInfo {
301     unsigned OldLB;
302     unsigned OtherLB;
303     unsigned MyShift;
304     unsigned OtherShift;
305   };
306   static bool mergeScore(const MergeInfo &M, unsigned &Score,
307                          unsigned OtherScore);
308 
309   void setScoreLB(InstCounterType T, unsigned Val) {
310     assert(T < NUM_INST_CNTS);
311     ScoreLBs[T] = Val;
312   }
313 
314   void setScoreUB(InstCounterType T, unsigned Val) {
315     assert(T < NUM_INST_CNTS);
316     ScoreUBs[T] = Val;
317 
318     if (T != EXP_CNT)
319       return;
320 
321     if (getScoreRange(EXP_CNT) > getWaitCountMax(EXP_CNT))
322       ScoreLBs[EXP_CNT] = ScoreUBs[EXP_CNT] - getWaitCountMax(EXP_CNT);
323   }
324 
325   void setRegScore(int GprNo, InstCounterType T, unsigned Val) {
326     if (GprNo < NUM_ALL_VGPRS) {
327       VgprUB = std::max(VgprUB, GprNo);
328       VgprScores[T][GprNo] = Val;
329     } else {
330       assert(T == LGKM_CNT);
331       SgprUB = std::max(SgprUB, GprNo - NUM_ALL_VGPRS);
332       SgprScores[GprNo - NUM_ALL_VGPRS] = Val;
333     }
334   }
335 
336   void setExpScore(const MachineInstr *MI, const SIInstrInfo *TII,
337                    const SIRegisterInfo *TRI, const MachineRegisterInfo *MRI,
338                    unsigned OpNo, unsigned Val);
339 
340   const GCNSubtarget *ST = nullptr;
341   HardwareLimits Limits = {};
342   RegisterEncoding Encoding = {};
343   unsigned ScoreLBs[NUM_INST_CNTS] = {0};
344   unsigned ScoreUBs[NUM_INST_CNTS] = {0};
345   unsigned PendingEvents = 0;
346   // Remember the last flat memory operation.
347   unsigned LastFlat[NUM_INST_CNTS] = {0};
348   // wait_cnt scores for every vgpr.
349   // Keep track of the VgprUB and SgprUB to make merge at join efficient.
350   int VgprUB = -1;
351   int SgprUB = -1;
352   unsigned VgprScores[NUM_INST_CNTS][NUM_ALL_VGPRS] = {{0}};
353   // Wait cnt scores for every sgpr, only lgkmcnt is relevant.
354   unsigned SgprScores[SQ_MAX_PGM_SGPRS] = {0};
355   // Bitmask of the VmemTypes of VMEM instructions that might have a pending
356   // write to each vgpr.
357   unsigned char VgprVmemTypes[NUM_ALL_VGPRS] = {0};
358 };
359 
360 class SIInsertWaitcnts : public MachineFunctionPass {
361 private:
362   const GCNSubtarget *ST = nullptr;
363   const SIInstrInfo *TII = nullptr;
364   const SIRegisterInfo *TRI = nullptr;
365   const MachineRegisterInfo *MRI = nullptr;
366   AMDGPU::IsaVersion IV;
367 
368   DenseSet<MachineInstr *> TrackedWaitcntSet;
369   DenseMap<const Value *, MachineBasicBlock *> SLoadAddresses;
370   DenseMap<MachineBasicBlock *, bool> PreheadersToFlush;
371   MachineLoopInfo *MLI;
372   MachinePostDominatorTree *PDT;
373 
374   struct BlockInfo {
375     MachineBasicBlock *MBB;
376     std::unique_ptr<WaitcntBrackets> Incoming;
377     bool Dirty = true;
378 
379     explicit BlockInfo(MachineBasicBlock *MBB) : MBB(MBB) {}
380   };
381 
382   MapVector<MachineBasicBlock *, BlockInfo> BlockInfos;
383 
384   // ForceEmitZeroWaitcnts: force all waitcnts insts to be s_waitcnt 0
385   // because of amdgpu-waitcnt-forcezero flag
386   bool ForceEmitZeroWaitcnts;
387   bool ForceEmitWaitcnt[NUM_INST_CNTS];
388 
389 public:
390   static char ID;
391 
392   SIInsertWaitcnts() : MachineFunctionPass(ID) {
393     (void)ForceExpCounter;
394     (void)ForceLgkmCounter;
395     (void)ForceVMCounter;
396   }
397 
398   bool shouldFlushVmCnt(MachineLoop *ML, WaitcntBrackets &Brackets);
399   bool isPreheaderToFlush(MachineBasicBlock &MBB,
400                           WaitcntBrackets &ScoreBrackets);
401   bool runOnMachineFunction(MachineFunction &MF) override;
402 
403   StringRef getPassName() const override {
404     return "SI insert wait instructions";
405   }
406 
407   void getAnalysisUsage(AnalysisUsage &AU) const override {
408     AU.setPreservesCFG();
409     AU.addRequired<MachineLoopInfo>();
410     AU.addRequired<MachinePostDominatorTree>();
411     MachineFunctionPass::getAnalysisUsage(AU);
412   }
413 
414   bool isForceEmitWaitcnt() const {
415     for (auto T : inst_counter_types())
416       if (ForceEmitWaitcnt[T])
417         return true;
418     return false;
419   }
420 
421   AMDGPU::Waitcnt allZeroWaitcnt() const {
422     return AMDGPU::Waitcnt::allZero(ST->hasVscnt());
423   }
424 
425   void setForceEmitWaitcnt() {
426 // For non-debug builds, ForceEmitWaitcnt has been initialized to false;
427 // For debug builds, get the debug counter info and adjust if need be
428 #ifndef NDEBUG
429     if (DebugCounter::isCounterSet(ForceExpCounter) &&
430         DebugCounter::shouldExecute(ForceExpCounter)) {
431       ForceEmitWaitcnt[EXP_CNT] = true;
432     } else {
433       ForceEmitWaitcnt[EXP_CNT] = false;
434     }
435 
436     if (DebugCounter::isCounterSet(ForceLgkmCounter) &&
437         DebugCounter::shouldExecute(ForceLgkmCounter)) {
438       ForceEmitWaitcnt[LGKM_CNT] = true;
439     } else {
440       ForceEmitWaitcnt[LGKM_CNT] = false;
441     }
442 
443     if (DebugCounter::isCounterSet(ForceVMCounter) &&
444         DebugCounter::shouldExecute(ForceVMCounter)) {
445       ForceEmitWaitcnt[VM_CNT] = true;
446     } else {
447       ForceEmitWaitcnt[VM_CNT] = false;
448     }
449 #endif // NDEBUG
450   }
451 
452   // Return the appropriate VMEM_*_ACCESS type for Inst, which must be a VMEM or
453   // FLAT instruction.
454   WaitEventType getVmemWaitEventType(const MachineInstr &Inst) const {
455     assert(SIInstrInfo::isVMEM(Inst) || SIInstrInfo::isFLAT(Inst));
456     if (!ST->hasVscnt())
457       return VMEM_ACCESS;
458     if (Inst.mayStore() && !SIInstrInfo::isAtomicRet(Inst))
459       return VMEM_WRITE_ACCESS;
460     return VMEM_READ_ACCESS;
461   }
462 
463   bool mayAccessVMEMThroughFlat(const MachineInstr &MI) const;
464   bool mayAccessLDSThroughFlat(const MachineInstr &MI) const;
465   bool generateWaitcntInstBefore(MachineInstr &MI,
466                                  WaitcntBrackets &ScoreBrackets,
467                                  MachineInstr *OldWaitcntInstr,
468                                  bool FlushVmCnt);
469   bool generateWaitcntBlockEnd(MachineBasicBlock &Block,
470                                WaitcntBrackets &ScoreBrackets,
471                                MachineInstr *OldWaitcntInstr);
472   bool generateWaitcnt(AMDGPU::Waitcnt Wait,
473                        MachineBasicBlock::instr_iterator It,
474                        MachineBasicBlock &Block, WaitcntBrackets &ScoreBrackets,
475                        MachineInstr *OldWaitcntInstr);
476   void updateEventWaitcntAfter(MachineInstr &Inst,
477                                WaitcntBrackets *ScoreBrackets);
478   bool insertWaitcntInBlock(MachineFunction &MF, MachineBasicBlock &Block,
479                             WaitcntBrackets &ScoreBrackets);
480   bool applyPreexistingWaitcnt(WaitcntBrackets &ScoreBrackets,
481                                MachineInstr &OldWaitcntInstr,
482                                AMDGPU::Waitcnt &Wait,
483                                MachineBasicBlock::instr_iterator It) const;
484 };
485 
486 } // end anonymous namespace
487 
488 RegInterval WaitcntBrackets::getRegInterval(const MachineInstr *MI,
489                                             const SIInstrInfo *TII,
490                                             const MachineRegisterInfo *MRI,
491                                             const SIRegisterInfo *TRI,
492                                             unsigned OpNo) const {
493   const MachineOperand &Op = MI->getOperand(OpNo);
494   if (!TRI->isInAllocatableClass(Op.getReg()))
495     return {-1, -1};
496 
497   // A use via a PW operand does not need a waitcnt.
498   // A partial write is not a WAW.
499   assert(!Op.getSubReg() || !Op.isUndef());
500 
501   RegInterval Result;
502 
503   unsigned Reg = TRI->getEncodingValue(AMDGPU::getMCReg(Op.getReg(), *ST));
504 
505   if (TRI->isVectorRegister(*MRI, Op.getReg())) {
506     assert(Reg >= Encoding.VGPR0 && Reg <= Encoding.VGPRL);
507     Result.first = Reg - Encoding.VGPR0;
508     if (TRI->isAGPR(*MRI, Op.getReg()))
509       Result.first += AGPR_OFFSET;
510     assert(Result.first >= 0 && Result.first < SQ_MAX_PGM_VGPRS);
511   } else if (TRI->isSGPRReg(*MRI, Op.getReg())) {
512     assert(Reg >= Encoding.SGPR0 && Reg < SQ_MAX_PGM_SGPRS);
513     Result.first = Reg - Encoding.SGPR0 + NUM_ALL_VGPRS;
514     assert(Result.first >= NUM_ALL_VGPRS &&
515            Result.first < SQ_MAX_PGM_SGPRS + NUM_ALL_VGPRS);
516   }
517   // TODO: Handle TTMP
518   // else if (TRI->isTTMP(*MRI, Reg.getReg())) ...
519   else
520     return {-1, -1};
521 
522   const TargetRegisterClass *RC = TII->getOpRegClass(*MI, OpNo);
523   unsigned Size = TRI->getRegSizeInBits(*RC);
524   Result.second = Result.first + ((Size + 16) / 32);
525 
526   return Result;
527 }
528 
529 void WaitcntBrackets::setExpScore(const MachineInstr *MI,
530                                   const SIInstrInfo *TII,
531                                   const SIRegisterInfo *TRI,
532                                   const MachineRegisterInfo *MRI, unsigned OpNo,
533                                   unsigned Val) {
534   RegInterval Interval = getRegInterval(MI, TII, MRI, TRI, OpNo);
535   assert(TRI->isVectorRegister(*MRI, MI->getOperand(OpNo).getReg()));
536   for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
537     setRegScore(RegNo, EXP_CNT, Val);
538   }
539 }
540 
541 // MUBUF and FLAT LDS DMA operations need a wait on vmcnt before LDS written
542 // can be accessed. A load from LDS to VMEM does not need a wait.
543 static bool mayWriteLDSThroughDMA(const MachineInstr &MI) {
544   return SIInstrInfo::isVALU(MI) &&
545          (SIInstrInfo::isMUBUF(MI) || SIInstrInfo::isFLAT(MI)) &&
546          MI.getOpcode() != AMDGPU::BUFFER_STORE_LDS_DWORD;
547 }
548 
549 void WaitcntBrackets::updateByEvent(const SIInstrInfo *TII,
550                                     const SIRegisterInfo *TRI,
551                                     const MachineRegisterInfo *MRI,
552                                     WaitEventType E, MachineInstr &Inst) {
553   InstCounterType T = eventCounter(E);
554   unsigned CurrScore = getScoreUB(T) + 1;
555   if (CurrScore == 0)
556     report_fatal_error("InsertWaitcnt score wraparound");
557   // PendingEvents and ScoreUB need to be update regardless if this event
558   // changes the score of a register or not.
559   // Examples including vm_cnt when buffer-store or lgkm_cnt when send-message.
560   PendingEvents |= 1 << E;
561   setScoreUB(T, CurrScore);
562 
563   if (T == EXP_CNT) {
564     // Put score on the source vgprs. If this is a store, just use those
565     // specific register(s).
566     if (TII->isDS(Inst) && (Inst.mayStore() || Inst.mayLoad())) {
567       int AddrOpIdx =
568           AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::addr);
569       // All GDS operations must protect their address register (same as
570       // export.)
571       if (AddrOpIdx != -1) {
572         setExpScore(&Inst, TII, TRI, MRI, AddrOpIdx, CurrScore);
573       }
574 
575       if (Inst.mayStore()) {
576         if (AMDGPU::hasNamedOperand(Inst.getOpcode(), AMDGPU::OpName::data0)) {
577           setExpScore(
578               &Inst, TII, TRI, MRI,
579               AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data0),
580               CurrScore);
581         }
582         if (AMDGPU::hasNamedOperand(Inst.getOpcode(), AMDGPU::OpName::data1)) {
583           setExpScore(&Inst, TII, TRI, MRI,
584                       AMDGPU::getNamedOperandIdx(Inst.getOpcode(),
585                                                  AMDGPU::OpName::data1),
586                       CurrScore);
587         }
588       } else if (SIInstrInfo::isAtomicRet(Inst) &&
589                  Inst.getOpcode() != AMDGPU::DS_GWS_INIT &&
590                  Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_V &&
591                  Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_BR &&
592                  Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_P &&
593                  Inst.getOpcode() != AMDGPU::DS_GWS_BARRIER &&
594                  Inst.getOpcode() != AMDGPU::DS_APPEND &&
595                  Inst.getOpcode() != AMDGPU::DS_CONSUME &&
596                  Inst.getOpcode() != AMDGPU::DS_ORDERED_COUNT) {
597         for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
598           const MachineOperand &Op = Inst.getOperand(I);
599           if (Op.isReg() && !Op.isDef() &&
600               TRI->isVectorRegister(*MRI, Op.getReg())) {
601             setExpScore(&Inst, TII, TRI, MRI, I, CurrScore);
602           }
603         }
604       }
605     } else if (TII->isFLAT(Inst)) {
606       if (Inst.mayStore()) {
607         setExpScore(
608             &Inst, TII, TRI, MRI,
609             AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
610             CurrScore);
611       } else if (SIInstrInfo::isAtomicRet(Inst)) {
612         setExpScore(
613             &Inst, TII, TRI, MRI,
614             AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
615             CurrScore);
616       }
617     } else if (TII->isMIMG(Inst)) {
618       if (Inst.mayStore()) {
619         setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore);
620       } else if (SIInstrInfo::isAtomicRet(Inst)) {
621         setExpScore(
622             &Inst, TII, TRI, MRI,
623             AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
624             CurrScore);
625       }
626     } else if (TII->isMTBUF(Inst)) {
627       if (Inst.mayStore()) {
628         setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore);
629       }
630     } else if (TII->isMUBUF(Inst)) {
631       if (Inst.mayStore()) {
632         setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore);
633       } else if (SIInstrInfo::isAtomicRet(Inst)) {
634         setExpScore(
635             &Inst, TII, TRI, MRI,
636             AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
637             CurrScore);
638       }
639     } else if (TII->isLDSDIR(Inst)) {
640       // LDSDIR instructions attach the score to the destination.
641       setExpScore(
642           &Inst, TII, TRI, MRI,
643           AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::vdst),
644           CurrScore);
645     } else {
646       if (TII->isEXP(Inst)) {
647         // For export the destination registers are really temps that
648         // can be used as the actual source after export patching, so
649         // we need to treat them like sources and set the EXP_CNT
650         // score.
651         for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
652           MachineOperand &DefMO = Inst.getOperand(I);
653           if (DefMO.isReg() && DefMO.isDef() &&
654               TRI->isVGPR(*MRI, DefMO.getReg())) {
655             setRegScore(
656                 TRI->getEncodingValue(AMDGPU::getMCReg(DefMO.getReg(), *ST)),
657                 EXP_CNT, CurrScore);
658           }
659         }
660       }
661       for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
662         MachineOperand &MO = Inst.getOperand(I);
663         if (MO.isReg() && !MO.isDef() &&
664             TRI->isVectorRegister(*MRI, MO.getReg())) {
665           setExpScore(&Inst, TII, TRI, MRI, I, CurrScore);
666         }
667       }
668     }
669 #if 0 // TODO: check if this is handled by MUBUF code above.
670   } else if (Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORD ||
671        Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORDX2 ||
672        Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORDX4) {
673     MachineOperand *MO = TII->getNamedOperand(Inst, AMDGPU::OpName::data);
674     unsigned OpNo;//TODO: find the OpNo for this operand;
675     RegInterval Interval = getRegInterval(&Inst, TII, MRI, TRI, OpNo);
676     for (int RegNo = Interval.first; RegNo < Interval.second;
677     ++RegNo) {
678       setRegScore(RegNo + NUM_ALL_VGPRS, t, CurrScore);
679     }
680 #endif
681   } else {
682     // Match the score to the destination registers.
683     for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
684       auto &Op = Inst.getOperand(I);
685       if (!Op.isReg() || !Op.isDef())
686         continue;
687       RegInterval Interval = getRegInterval(&Inst, TII, MRI, TRI, I);
688       if (T == VM_CNT) {
689         if (Interval.first >= NUM_ALL_VGPRS)
690           continue;
691         if (updateVMCntOnly(Inst)) {
692           VmemType V = getVmemType(Inst);
693           for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo)
694             VgprVmemTypes[RegNo] |= 1 << V;
695         }
696       }
697       for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
698         setRegScore(RegNo, T, CurrScore);
699       }
700     }
701     if (Inst.mayStore() && (TII->isDS(Inst) || mayWriteLDSThroughDMA(Inst))) {
702       setRegScore(SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS, T, CurrScore);
703     }
704   }
705 }
706 
707 void WaitcntBrackets::print(raw_ostream &OS) {
708   OS << '\n';
709   for (auto T : inst_counter_types()) {
710     unsigned SR = getScoreRange(T);
711 
712     switch (T) {
713     case VM_CNT:
714       OS << "    VM_CNT(" << SR << "): ";
715       break;
716     case LGKM_CNT:
717       OS << "    LGKM_CNT(" << SR << "): ";
718       break;
719     case EXP_CNT:
720       OS << "    EXP_CNT(" << SR << "): ";
721       break;
722     case VS_CNT:
723       OS << "    VS_CNT(" << SR << "): ";
724       break;
725     default:
726       OS << "    UNKNOWN(" << SR << "): ";
727       break;
728     }
729 
730     if (SR != 0) {
731       // Print vgpr scores.
732       unsigned LB = getScoreLB(T);
733 
734       for (int J = 0; J <= VgprUB; J++) {
735         unsigned RegScore = getRegScore(J, T);
736         if (RegScore <= LB)
737           continue;
738         unsigned RelScore = RegScore - LB - 1;
739         if (J < SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS) {
740           OS << RelScore << ":v" << J << " ";
741         } else {
742           OS << RelScore << ":ds ";
743         }
744       }
745       // Also need to print sgpr scores for lgkm_cnt.
746       if (T == LGKM_CNT) {
747         for (int J = 0; J <= SgprUB; J++) {
748           unsigned RegScore = getRegScore(J + NUM_ALL_VGPRS, LGKM_CNT);
749           if (RegScore <= LB)
750             continue;
751           unsigned RelScore = RegScore - LB - 1;
752           OS << RelScore << ":s" << J << " ";
753         }
754       }
755     }
756     OS << '\n';
757   }
758   OS << '\n';
759 }
760 
761 /// Simplify the waitcnt, in the sense of removing redundant counts, and return
762 /// whether a waitcnt instruction is needed at all.
763 void WaitcntBrackets::simplifyWaitcnt(AMDGPU::Waitcnt &Wait) const {
764   simplifyWaitcnt(VM_CNT, Wait.VmCnt);
765   simplifyWaitcnt(EXP_CNT, Wait.ExpCnt);
766   simplifyWaitcnt(LGKM_CNT, Wait.LgkmCnt);
767   simplifyWaitcnt(VS_CNT, Wait.VsCnt);
768 }
769 
770 void WaitcntBrackets::simplifyWaitcnt(InstCounterType T,
771                                       unsigned &Count) const {
772   // The number of outstanding events for this type, T, can be calculated
773   // as (UB - LB). If the current Count is greater than or equal to the number
774   // of outstanding events, then the wait for this counter is redundant.
775   if (Count >= getScoreRange(T))
776     Count = ~0u;
777 }
778 
779 void WaitcntBrackets::determineWait(InstCounterType T, int RegNo,
780                                     AMDGPU::Waitcnt &Wait) const {
781   unsigned ScoreToWait = getRegScore(RegNo, T);
782 
783   // If the score of src_operand falls within the bracket, we need an
784   // s_waitcnt instruction.
785   const unsigned LB = getScoreLB(T);
786   const unsigned UB = getScoreUB(T);
787   if ((UB >= ScoreToWait) && (ScoreToWait > LB)) {
788     if ((T == VM_CNT || T == LGKM_CNT) &&
789         hasPendingFlat() &&
790         !ST->hasFlatLgkmVMemCountInOrder()) {
791       // If there is a pending FLAT operation, and this is a VMem or LGKM
792       // waitcnt and the target can report early completion, then we need
793       // to force a waitcnt 0.
794       addWait(Wait, T, 0);
795     } else if (counterOutOfOrder(T)) {
796       // Counter can get decremented out-of-order when there
797       // are multiple types event in the bracket. Also emit an s_wait counter
798       // with a conservative value of 0 for the counter.
799       addWait(Wait, T, 0);
800     } else {
801       // If a counter has been maxed out avoid overflow by waiting for
802       // MAX(CounterType) - 1 instead.
803       unsigned NeededWait = std::min(UB - ScoreToWait, getWaitCountMax(T) - 1);
804       addWait(Wait, T, NeededWait);
805     }
806   }
807 }
808 
809 void WaitcntBrackets::applyWaitcnt(const AMDGPU::Waitcnt &Wait) {
810   applyWaitcnt(VM_CNT, Wait.VmCnt);
811   applyWaitcnt(EXP_CNT, Wait.ExpCnt);
812   applyWaitcnt(LGKM_CNT, Wait.LgkmCnt);
813   applyWaitcnt(VS_CNT, Wait.VsCnt);
814 }
815 
816 void WaitcntBrackets::applyWaitcnt(InstCounterType T, unsigned Count) {
817   const unsigned UB = getScoreUB(T);
818   if (Count >= UB)
819     return;
820   if (Count != 0) {
821     if (counterOutOfOrder(T))
822       return;
823     setScoreLB(T, std::max(getScoreLB(T), UB - Count));
824   } else {
825     setScoreLB(T, UB);
826     PendingEvents &= ~WaitEventMaskForInst[T];
827   }
828 }
829 
830 // Where there are multiple types of event in the bracket of a counter,
831 // the decrement may go out of order.
832 bool WaitcntBrackets::counterOutOfOrder(InstCounterType T) const {
833   // Scalar memory read always can go out of order.
834   if (T == LGKM_CNT && hasPendingEvent(SMEM_ACCESS))
835     return true;
836   return hasMixedPendingEvents(T);
837 }
838 
839 INITIALIZE_PASS_BEGIN(SIInsertWaitcnts, DEBUG_TYPE, "SI Insert Waitcnts", false,
840                       false)
841 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
842 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
843 INITIALIZE_PASS_END(SIInsertWaitcnts, DEBUG_TYPE, "SI Insert Waitcnts", false,
844                     false)
845 
846 char SIInsertWaitcnts::ID = 0;
847 
848 char &llvm::SIInsertWaitcntsID = SIInsertWaitcnts::ID;
849 
850 FunctionPass *llvm::createSIInsertWaitcntsPass() {
851   return new SIInsertWaitcnts();
852 }
853 
854 static bool updateOperandIfDifferent(MachineInstr &MI, uint16_t OpName,
855                                      unsigned NewEnc) {
856   int OpIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OpName);
857   assert(OpIdx >= 0);
858 
859   MachineOperand &MO = MI.getOperand(OpIdx);
860 
861   if (NewEnc == MO.getImm())
862     return false;
863 
864   MO.setImm(NewEnc);
865   return true;
866 }
867 
868 /// Combine consecutive waitcnt instructions that precede \p It and follow
869 /// \p OldWaitcntInstr and apply any extra wait from waitcnt that were added
870 /// by previous passes. Currently this pass conservatively assumes that these
871 /// preexisting waitcnt are required for correctness.
872 bool SIInsertWaitcnts::applyPreexistingWaitcnt(
873     WaitcntBrackets &ScoreBrackets, MachineInstr &OldWaitcntInstr,
874     AMDGPU::Waitcnt &Wait, MachineBasicBlock::instr_iterator It) const {
875   bool Modified = false;
876   MachineInstr *WaitcntInstr = nullptr;
877   MachineInstr *WaitcntVsCntInstr = nullptr;
878 
879   for (auto &II :
880        make_early_inc_range(make_range(OldWaitcntInstr.getIterator(), It))) {
881     if (II.isMetaInstruction())
882       continue;
883 
884     if (II.getOpcode() == AMDGPU::S_WAITCNT) {
885       // Conservatively update required wait if this waitcnt was added in an
886       // earlier pass. In this case it will not exist in the tracked waitcnt
887       // set.
888       if (!TrackedWaitcntSet.count(&II)) {
889         unsigned IEnc = II.getOperand(0).getImm();
890         AMDGPU::Waitcnt OldWait = AMDGPU::decodeWaitcnt(IV, IEnc);
891         Wait = Wait.combined(OldWait);
892       }
893 
894       // Merge consecutive waitcnt of the same type by erasing multiples.
895       if (!WaitcntInstr) {
896         WaitcntInstr = &II;
897       } else {
898         II.eraseFromParent();
899         Modified = true;
900       }
901 
902     } else {
903       assert(II.getOpcode() == AMDGPU::S_WAITCNT_VSCNT);
904       assert(II.getOperand(0).getReg() == AMDGPU::SGPR_NULL);
905       if (!TrackedWaitcntSet.count(&II)) {
906         unsigned OldVSCnt =
907             TII->getNamedOperand(II, AMDGPU::OpName::simm16)->getImm();
908         Wait.VsCnt = std::min(Wait.VsCnt, OldVSCnt);
909       }
910 
911       if (!WaitcntVsCntInstr) {
912         WaitcntVsCntInstr = &II;
913       } else {
914         II.eraseFromParent();
915         Modified = true;
916       }
917     }
918   }
919 
920   // Updated encoding of merged waitcnt with the required wait.
921   if (WaitcntInstr) {
922     if (Wait.hasWaitExceptVsCnt()) {
923       Modified |=
924           updateOperandIfDifferent(*WaitcntInstr, AMDGPU::OpName::simm16,
925                                    AMDGPU::encodeWaitcnt(IV, Wait));
926       ScoreBrackets.applyWaitcnt(Wait);
927       Wait.VmCnt = ~0u;
928       Wait.LgkmCnt = ~0u;
929       Wait.ExpCnt = ~0u;
930 
931       LLVM_DEBUG(It == OldWaitcntInstr.getParent()->end()
932                      ? dbgs() << "applyPreexistingWaitcnt\n"
933                               << "New Instr at block end: " << *WaitcntInstr
934                               << '\n'
935                      : dbgs() << "applyPreexistingWaitcnt\n"
936                               << "Old Instr: " << *It
937                               << "New Instr: " << *WaitcntInstr << '\n');
938 
939     } else {
940       WaitcntInstr->eraseFromParent();
941       Modified = true;
942     }
943   }
944 
945   if (WaitcntVsCntInstr) {
946     if (Wait.hasWaitVsCnt()) {
947       assert(ST->hasVscnt());
948       Modified |= updateOperandIfDifferent(*WaitcntVsCntInstr,
949                                            AMDGPU::OpName::simm16, Wait.VsCnt);
950       ScoreBrackets.applyWaitcnt(Wait);
951       Wait.VsCnt = ~0u;
952 
953       LLVM_DEBUG(It == OldWaitcntInstr.getParent()->end()
954                      ? dbgs() << "applyPreexistingWaitcnt\n"
955                               << "New Instr at block end: "
956                               << *WaitcntVsCntInstr << '\n'
957                      : dbgs() << "applyPreexistingWaitcnt\n"
958                               << "Old Instr: " << *It
959                               << "New Instr: " << *WaitcntVsCntInstr << '\n');
960     } else {
961       WaitcntVsCntInstr->eraseFromParent();
962       Modified = true;
963     }
964   }
965 
966   return Modified;
967 }
968 
969 static bool readsVCCZ(const MachineInstr &MI) {
970   unsigned Opc = MI.getOpcode();
971   return (Opc == AMDGPU::S_CBRANCH_VCCNZ || Opc == AMDGPU::S_CBRANCH_VCCZ) &&
972          !MI.getOperand(1).isUndef();
973 }
974 
975 /// \returns true if the callee inserts an s_waitcnt 0 on function entry.
976 static bool callWaitsOnFunctionEntry(const MachineInstr &MI) {
977   // Currently all conventions wait, but this may not always be the case.
978   //
979   // TODO: If IPRA is enabled, and the callee is isSafeForNoCSROpt, it may make
980   // senses to omit the wait and do it in the caller.
981   return true;
982 }
983 
984 /// \returns true if the callee is expected to wait for any outstanding waits
985 /// before returning.
986 static bool callWaitsOnFunctionReturn(const MachineInstr &MI) {
987   return true;
988 }
989 
990 ///  Generate s_waitcnt instruction to be placed before cur_Inst.
991 ///  Instructions of a given type are returned in order,
992 ///  but instructions of different types can complete out of order.
993 ///  We rely on this in-order completion
994 ///  and simply assign a score to the memory access instructions.
995 ///  We keep track of the active "score bracket" to determine
996 ///  if an access of a memory read requires an s_waitcnt
997 ///  and if so what the value of each counter is.
998 ///  The "score bracket" is bound by the lower bound and upper bound
999 ///  scores (*_score_LB and *_score_ub respectively).
1000 ///  If FlushVmCnt is true, that means that we want to generate a s_waitcnt to
1001 ///  flush the vmcnt counter here.
1002 bool SIInsertWaitcnts::generateWaitcntInstBefore(MachineInstr &MI,
1003                                                  WaitcntBrackets &ScoreBrackets,
1004                                                  MachineInstr *OldWaitcntInstr,
1005                                                  bool FlushVmCnt) {
1006   setForceEmitWaitcnt();
1007 
1008   if (MI.isMetaInstruction())
1009     return false;
1010 
1011   AMDGPU::Waitcnt Wait;
1012 
1013   // FIXME: This should have already been handled by the memory legalizer.
1014   // Removing this currently doesn't affect any lit tests, but we need to
1015   // verify that nothing was relying on this. The number of buffer invalidates
1016   // being handled here should not be expanded.
1017   if (MI.getOpcode() == AMDGPU::BUFFER_WBINVL1 ||
1018       MI.getOpcode() == AMDGPU::BUFFER_WBINVL1_SC ||
1019       MI.getOpcode() == AMDGPU::BUFFER_WBINVL1_VOL ||
1020       MI.getOpcode() == AMDGPU::BUFFER_GL0_INV ||
1021       MI.getOpcode() == AMDGPU::BUFFER_GL1_INV) {
1022     Wait.VmCnt = 0;
1023   }
1024 
1025   // All waits must be resolved at call return.
1026   // NOTE: this could be improved with knowledge of all call sites or
1027   //   with knowledge of the called routines.
1028   if (MI.getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG ||
1029       MI.getOpcode() == AMDGPU::SI_RETURN ||
1030       MI.getOpcode() == AMDGPU::S_SETPC_B64_return ||
1031       (MI.isReturn() && MI.isCall() && !callWaitsOnFunctionEntry(MI))) {
1032     Wait = Wait.combined(allZeroWaitcnt());
1033   }
1034   // Resolve vm waits before gs-done.
1035   else if ((MI.getOpcode() == AMDGPU::S_SENDMSG ||
1036             MI.getOpcode() == AMDGPU::S_SENDMSGHALT) &&
1037            ST->hasLegacyGeometry() &&
1038            ((MI.getOperand(0).getImm() & AMDGPU::SendMsg::ID_MASK_PreGFX11_) ==
1039             AMDGPU::SendMsg::ID_GS_DONE_PreGFX11)) {
1040     Wait.VmCnt = 0;
1041   }
1042 #if 0 // TODO: the following blocks of logic when we have fence.
1043   else if (MI.getOpcode() == SC_FENCE) {
1044     const unsigned int group_size =
1045       context->shader_info->GetMaxThreadGroupSize();
1046     // group_size == 0 means thread group size is unknown at compile time
1047     const bool group_is_multi_wave =
1048       (group_size == 0 || group_size > target_info->GetWaveFrontSize());
1049     const bool fence_is_global = !((SCInstInternalMisc*)Inst)->IsGroupFence();
1050 
1051     for (unsigned int i = 0; i < Inst->NumSrcOperands(); i++) {
1052       SCRegType src_type = Inst->GetSrcType(i);
1053       switch (src_type) {
1054         case SCMEM_LDS:
1055           if (group_is_multi_wave ||
1056             context->OptFlagIsOn(OPT_R1100_LDSMEM_FENCE_CHICKEN_BIT)) {
1057             EmitWaitcnt |= ScoreBrackets->updateByWait(LGKM_CNT,
1058                                ScoreBrackets->getScoreUB(LGKM_CNT));
1059             // LDS may have to wait for VM_CNT after buffer load to LDS
1060             if (target_info->HasBufferLoadToLDS()) {
1061               EmitWaitcnt |= ScoreBrackets->updateByWait(VM_CNT,
1062                                  ScoreBrackets->getScoreUB(VM_CNT));
1063             }
1064           }
1065           break;
1066 
1067         case SCMEM_GDS:
1068           if (group_is_multi_wave || fence_is_global) {
1069             EmitWaitcnt |= ScoreBrackets->updateByWait(EXP_CNT,
1070               ScoreBrackets->getScoreUB(EXP_CNT));
1071             EmitWaitcnt |= ScoreBrackets->updateByWait(LGKM_CNT,
1072               ScoreBrackets->getScoreUB(LGKM_CNT));
1073           }
1074           break;
1075 
1076         case SCMEM_UAV:
1077         case SCMEM_TFBUF:
1078         case SCMEM_RING:
1079         case SCMEM_SCATTER:
1080           if (group_is_multi_wave || fence_is_global) {
1081             EmitWaitcnt |= ScoreBrackets->updateByWait(EXP_CNT,
1082               ScoreBrackets->getScoreUB(EXP_CNT));
1083             EmitWaitcnt |= ScoreBrackets->updateByWait(VM_CNT,
1084               ScoreBrackets->getScoreUB(VM_CNT));
1085           }
1086           break;
1087 
1088         case SCMEM_SCRATCH:
1089         default:
1090           break;
1091       }
1092     }
1093   }
1094 #endif
1095 
1096   // Export & GDS instructions do not read the EXEC mask until after the export
1097   // is granted (which can occur well after the instruction is issued).
1098   // The shader program must flush all EXP operations on the export-count
1099   // before overwriting the EXEC mask.
1100   else {
1101     if (MI.modifiesRegister(AMDGPU::EXEC, TRI)) {
1102       // Export and GDS are tracked individually, either may trigger a waitcnt
1103       // for EXEC.
1104       if (ScoreBrackets.hasPendingEvent(EXP_GPR_LOCK) ||
1105           ScoreBrackets.hasPendingEvent(EXP_PARAM_ACCESS) ||
1106           ScoreBrackets.hasPendingEvent(EXP_POS_ACCESS) ||
1107           ScoreBrackets.hasPendingEvent(GDS_GPR_LOCK)) {
1108         Wait.ExpCnt = 0;
1109       }
1110     }
1111 
1112     if (MI.isCall() && callWaitsOnFunctionEntry(MI)) {
1113       // The function is going to insert a wait on everything in its prolog.
1114       // This still needs to be careful if the call target is a load (e.g. a GOT
1115       // load). We also need to check WAW dependency with saved PC.
1116       Wait = AMDGPU::Waitcnt();
1117 
1118       int CallAddrOpIdx =
1119           AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::src0);
1120 
1121       if (MI.getOperand(CallAddrOpIdx).isReg()) {
1122         RegInterval CallAddrOpInterval =
1123           ScoreBrackets.getRegInterval(&MI, TII, MRI, TRI, CallAddrOpIdx);
1124 
1125         for (int RegNo = CallAddrOpInterval.first;
1126              RegNo < CallAddrOpInterval.second; ++RegNo)
1127           ScoreBrackets.determineWait(LGKM_CNT, RegNo, Wait);
1128 
1129         int RtnAddrOpIdx =
1130           AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::dst);
1131         if (RtnAddrOpIdx != -1) {
1132           RegInterval RtnAddrOpInterval =
1133             ScoreBrackets.getRegInterval(&MI, TII, MRI, TRI, RtnAddrOpIdx);
1134 
1135           for (int RegNo = RtnAddrOpInterval.first;
1136                RegNo < RtnAddrOpInterval.second; ++RegNo)
1137             ScoreBrackets.determineWait(LGKM_CNT, RegNo, Wait);
1138         }
1139       }
1140     } else {
1141       // FIXME: Should not be relying on memoperands.
1142       // Look at the source operands of every instruction to see if
1143       // any of them results from a previous memory operation that affects
1144       // its current usage. If so, an s_waitcnt instruction needs to be
1145       // emitted.
1146       // If the source operand was defined by a load, add the s_waitcnt
1147       // instruction.
1148       //
1149       // Two cases are handled for destination operands:
1150       // 1) If the destination operand was defined by a load, add the s_waitcnt
1151       // instruction to guarantee the right WAW order.
1152       // 2) If a destination operand that was used by a recent export/store ins,
1153       // add s_waitcnt on exp_cnt to guarantee the WAR order.
1154       for (const MachineMemOperand *Memop : MI.memoperands()) {
1155         const Value *Ptr = Memop->getValue();
1156         if (Memop->isStore() && SLoadAddresses.count(Ptr)) {
1157           addWait(Wait, LGKM_CNT, 0);
1158           if (PDT->dominates(MI.getParent(), SLoadAddresses.find(Ptr)->second))
1159             SLoadAddresses.erase(Ptr);
1160         }
1161         unsigned AS = Memop->getAddrSpace();
1162         if (AS != AMDGPUAS::LOCAL_ADDRESS && AS != AMDGPUAS::FLAT_ADDRESS)
1163           continue;
1164         // No need to wait before load from VMEM to LDS.
1165         if (mayWriteLDSThroughDMA(MI))
1166           continue;
1167         unsigned RegNo = SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS;
1168         // VM_CNT is only relevant to vgpr or LDS.
1169         ScoreBrackets.determineWait(VM_CNT, RegNo, Wait);
1170         if (Memop->isStore()) {
1171           ScoreBrackets.determineWait(EXP_CNT, RegNo, Wait);
1172         }
1173       }
1174 
1175       // Loop over use and def operands.
1176       for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) {
1177         MachineOperand &Op = MI.getOperand(I);
1178         if (!Op.isReg())
1179           continue;
1180 
1181         // If the instruction does not read tied source, skip the operand.
1182         if (Op.isTied() && Op.isUse() && TII->doesNotReadTiedSource(MI))
1183           continue;
1184 
1185         RegInterval Interval =
1186             ScoreBrackets.getRegInterval(&MI, TII, MRI, TRI, I);
1187 
1188         const bool IsVGPR = TRI->isVectorRegister(*MRI, Op.getReg());
1189         for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
1190           if (IsVGPR) {
1191             // RAW always needs an s_waitcnt. WAW needs an s_waitcnt unless the
1192             // previous write and this write are the same type of VMEM
1193             // instruction, in which case they're guaranteed to write their
1194             // results in order anyway.
1195             if (Op.isUse() || !updateVMCntOnly(MI) ||
1196                 ScoreBrackets.hasOtherPendingVmemTypes(RegNo,
1197                                                        getVmemType(MI))) {
1198               ScoreBrackets.determineWait(VM_CNT, RegNo, Wait);
1199               ScoreBrackets.clearVgprVmemTypes(RegNo);
1200             }
1201             if (Op.isDef() || ScoreBrackets.hasPendingEvent(EXP_LDS_ACCESS)) {
1202               ScoreBrackets.determineWait(EXP_CNT, RegNo, Wait);
1203             }
1204           }
1205           ScoreBrackets.determineWait(LGKM_CNT, RegNo, Wait);
1206         }
1207       }
1208     }
1209   }
1210 
1211   // The subtarget may have an implicit S_WAITCNT 0 before barriers. If it does
1212   // not, we need to ensure the subtarget is capable of backing off barrier
1213   // instructions in case there are any outstanding memory operations that may
1214   // cause an exception. Otherwise, insert an explicit S_WAITCNT 0 here.
1215   if (MI.getOpcode() == AMDGPU::S_BARRIER &&
1216       !ST->hasAutoWaitcntBeforeBarrier() && !ST->supportsBackOffBarrier()) {
1217     Wait = Wait.combined(allZeroWaitcnt());
1218   }
1219 
1220   // TODO: Remove this work-around, enable the assert for Bug 457939
1221   //       after fixing the scheduler. Also, the Shader Compiler code is
1222   //       independent of target.
1223   if (readsVCCZ(MI) && ST->hasReadVCCZBug()) {
1224     if (ScoreBrackets.hasPendingEvent(SMEM_ACCESS)) {
1225       Wait.LgkmCnt = 0;
1226     }
1227   }
1228 
1229   // Verify that the wait is actually needed.
1230   ScoreBrackets.simplifyWaitcnt(Wait);
1231 
1232   if (ForceEmitZeroWaitcnts)
1233     Wait = allZeroWaitcnt();
1234 
1235   if (ForceEmitWaitcnt[VM_CNT])
1236     Wait.VmCnt = 0;
1237   if (ForceEmitWaitcnt[EXP_CNT])
1238     Wait.ExpCnt = 0;
1239   if (ForceEmitWaitcnt[LGKM_CNT])
1240     Wait.LgkmCnt = 0;
1241   if (ForceEmitWaitcnt[VS_CNT])
1242     Wait.VsCnt = 0;
1243 
1244   if (FlushVmCnt) {
1245     if (ScoreBrackets.hasPendingEvent(VM_CNT))
1246       Wait.VmCnt = 0;
1247   }
1248 
1249   return generateWaitcnt(Wait, MI.getIterator(), *MI.getParent(), ScoreBrackets,
1250                          OldWaitcntInstr);
1251 }
1252 
1253 // Add a waitcnt to flush the vmcnt counter at the end of the given block if
1254 // needed.
1255 bool SIInsertWaitcnts::generateWaitcntBlockEnd(MachineBasicBlock &Block,
1256                                                WaitcntBrackets &ScoreBrackets,
1257                                                MachineInstr *OldWaitcntInstr) {
1258   AMDGPU::Waitcnt Wait;
1259 
1260   if (!ScoreBrackets.hasPendingEvent(VM_CNT))
1261     return false;
1262 
1263   Wait.VmCnt = 0;
1264 
1265   return generateWaitcnt(Wait, Block.instr_end(), Block, ScoreBrackets,
1266                          OldWaitcntInstr);
1267 }
1268 
1269 bool SIInsertWaitcnts::generateWaitcnt(AMDGPU::Waitcnt Wait,
1270                                        MachineBasicBlock::instr_iterator It,
1271                                        MachineBasicBlock &Block,
1272                                        WaitcntBrackets &ScoreBrackets,
1273                                        MachineInstr *OldWaitcntInstr) {
1274   bool Modified = false;
1275   const DebugLoc &DL = Block.findDebugLoc(It);
1276 
1277   if (OldWaitcntInstr)
1278     // Try to merge the required wait with preexisting waitcnt instructions.
1279     // Also erase redundant waitcnt.
1280     Modified =
1281         applyPreexistingWaitcnt(ScoreBrackets, *OldWaitcntInstr, Wait, It);
1282   else
1283     ScoreBrackets.applyWaitcnt(Wait);
1284 
1285   // ExpCnt can be merged into VINTERP.
1286   if (Wait.ExpCnt != ~0u && It != Block.instr_end() &&
1287       SIInstrInfo::isVINTERP(*It)) {
1288     MachineOperand *WaitExp =
1289         TII->getNamedOperand(*It, AMDGPU::OpName::waitexp);
1290     if (Wait.ExpCnt < WaitExp->getImm()) {
1291       WaitExp->setImm(Wait.ExpCnt);
1292       Modified = true;
1293     }
1294     Wait.ExpCnt = ~0u;
1295 
1296     LLVM_DEBUG(dbgs() << "generateWaitcntInstBefore\n"
1297                       << "Update Instr: " << *It);
1298   }
1299 
1300   // Build new waitcnt instructions unless no wait is needed or the old waitcnt
1301   // instruction was modified to handle the required wait.
1302   if (Wait.hasWaitExceptVsCnt()) {
1303     unsigned Enc = AMDGPU::encodeWaitcnt(IV, Wait);
1304     auto SWaitInst =
1305         BuildMI(Block, It, DL, TII->get(AMDGPU::S_WAITCNT)).addImm(Enc);
1306     TrackedWaitcntSet.insert(SWaitInst);
1307     Modified = true;
1308 
1309     LLVM_DEBUG(dbgs() << "generateWaitcnt\n";
1310                if (It != Block.instr_end()) dbgs() << "Old Instr: " << *It;
1311                dbgs() << "New Instr: " << *SWaitInst << '\n');
1312   }
1313 
1314   if (Wait.hasWaitVsCnt()) {
1315     assert(ST->hasVscnt());
1316 
1317     auto SWaitInst = BuildMI(Block, It, DL, TII->get(AMDGPU::S_WAITCNT_VSCNT))
1318                          .addReg(AMDGPU::SGPR_NULL, RegState::Undef)
1319                          .addImm(Wait.VsCnt);
1320     TrackedWaitcntSet.insert(SWaitInst);
1321     Modified = true;
1322 
1323     LLVM_DEBUG(dbgs() << "generateWaitcnt\n";
1324                if (It != Block.instr_end()) dbgs() << "Old Instr: " << *It;
1325                dbgs() << "New Instr: " << *SWaitInst << '\n');
1326   }
1327   return Modified;
1328 }
1329 
1330 // This is a flat memory operation. Check to see if it has memory tokens other
1331 // than LDS. Other address spaces supported by flat memory operations involve
1332 // global memory.
1333 bool SIInsertWaitcnts::mayAccessVMEMThroughFlat(const MachineInstr &MI) const {
1334   assert(TII->isFLAT(MI));
1335 
1336   // All flat instructions use the VMEM counter.
1337   assert(TII->usesVM_CNT(MI));
1338 
1339   // If there are no memory operands then conservatively assume the flat
1340   // operation may access VMEM.
1341   if (MI.memoperands_empty())
1342     return true;
1343 
1344   // See if any memory operand specifies an address space that involves VMEM.
1345   // Flat operations only supported FLAT, LOCAL (LDS), or address spaces
1346   // involving VMEM such as GLOBAL, CONSTANT, PRIVATE (SCRATCH), etc. The REGION
1347   // (GDS) address space is not supported by flat operations. Therefore, simply
1348   // return true unless only the LDS address space is found.
1349   for (const MachineMemOperand *Memop : MI.memoperands()) {
1350     unsigned AS = Memop->getAddrSpace();
1351     assert(AS != AMDGPUAS::REGION_ADDRESS);
1352     if (AS != AMDGPUAS::LOCAL_ADDRESS)
1353       return true;
1354   }
1355 
1356   return false;
1357 }
1358 
1359 // This is a flat memory operation. Check to see if it has memory tokens for
1360 // either LDS or FLAT.
1361 bool SIInsertWaitcnts::mayAccessLDSThroughFlat(const MachineInstr &MI) const {
1362   assert(TII->isFLAT(MI));
1363 
1364   // Flat instruction such as SCRATCH and GLOBAL do not use the lgkm counter.
1365   if (!TII->usesLGKM_CNT(MI))
1366     return false;
1367 
1368   // If in tgsplit mode then there can be no use of LDS.
1369   if (ST->isTgSplitEnabled())
1370     return false;
1371 
1372   // If there are no memory operands then conservatively assume the flat
1373   // operation may access LDS.
1374   if (MI.memoperands_empty())
1375     return true;
1376 
1377   // See if any memory operand specifies an address space that involves LDS.
1378   for (const MachineMemOperand *Memop : MI.memoperands()) {
1379     unsigned AS = Memop->getAddrSpace();
1380     if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS)
1381       return true;
1382   }
1383 
1384   return false;
1385 }
1386 
1387 void SIInsertWaitcnts::updateEventWaitcntAfter(MachineInstr &Inst,
1388                                                WaitcntBrackets *ScoreBrackets) {
1389   // Now look at the instruction opcode. If it is a memory access
1390   // instruction, update the upper-bound of the appropriate counter's
1391   // bracket and the destination operand scores.
1392   // TODO: Use the (TSFlags & SIInstrFlags::LGKM_CNT) property everywhere.
1393   if (TII->isDS(Inst) && TII->usesLGKM_CNT(Inst)) {
1394     if (TII->isAlwaysGDS(Inst.getOpcode()) ||
1395         TII->hasModifiersSet(Inst, AMDGPU::OpName::gds)) {
1396       ScoreBrackets->updateByEvent(TII, TRI, MRI, GDS_ACCESS, Inst);
1397       ScoreBrackets->updateByEvent(TII, TRI, MRI, GDS_GPR_LOCK, Inst);
1398     } else {
1399       ScoreBrackets->updateByEvent(TII, TRI, MRI, LDS_ACCESS, Inst);
1400     }
1401   } else if (TII->isFLAT(Inst)) {
1402     assert(Inst.mayLoadOrStore());
1403 
1404     int FlatASCount = 0;
1405 
1406     if (mayAccessVMEMThroughFlat(Inst)) {
1407       ++FlatASCount;
1408       ScoreBrackets->updateByEvent(TII, TRI, MRI, getVmemWaitEventType(Inst),
1409                                    Inst);
1410     }
1411 
1412     if (mayAccessLDSThroughFlat(Inst)) {
1413       ++FlatASCount;
1414       ScoreBrackets->updateByEvent(TII, TRI, MRI, LDS_ACCESS, Inst);
1415     }
1416 
1417     // A Flat memory operation must access at least one address space.
1418     assert(FlatASCount);
1419 
1420     // This is a flat memory operation that access both VMEM and LDS, so note it
1421     // - it will require that both the VM and LGKM be flushed to zero if it is
1422     // pending when a VM or LGKM dependency occurs.
1423     if (FlatASCount > 1)
1424       ScoreBrackets->setPendingFlat();
1425   } else if (SIInstrInfo::isVMEM(Inst) &&
1426              !llvm::AMDGPU::getMUBUFIsBufferInv(Inst.getOpcode())) {
1427     ScoreBrackets->updateByEvent(TII, TRI, MRI, getVmemWaitEventType(Inst),
1428                                  Inst);
1429 
1430     if (ST->vmemWriteNeedsExpWaitcnt() &&
1431         (Inst.mayStore() || SIInstrInfo::isAtomicRet(Inst))) {
1432       ScoreBrackets->updateByEvent(TII, TRI, MRI, VMW_GPR_LOCK, Inst);
1433     }
1434   } else if (TII->isSMRD(Inst)) {
1435     ScoreBrackets->updateByEvent(TII, TRI, MRI, SMEM_ACCESS, Inst);
1436   } else if (Inst.isCall()) {
1437     if (callWaitsOnFunctionReturn(Inst)) {
1438       // Act as a wait on everything
1439       ScoreBrackets->applyWaitcnt(allZeroWaitcnt());
1440     } else {
1441       // May need to way wait for anything.
1442       ScoreBrackets->applyWaitcnt(AMDGPU::Waitcnt());
1443     }
1444   } else if (SIInstrInfo::isLDSDIR(Inst)) {
1445     ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_LDS_ACCESS, Inst);
1446   } else if (TII->isVINTERP(Inst)) {
1447     int64_t Imm = TII->getNamedOperand(Inst, AMDGPU::OpName::waitexp)->getImm();
1448     ScoreBrackets->applyWaitcnt(EXP_CNT, Imm);
1449   } else if (SIInstrInfo::isEXP(Inst)) {
1450     unsigned Imm = TII->getNamedOperand(Inst, AMDGPU::OpName::tgt)->getImm();
1451     if (Imm >= AMDGPU::Exp::ET_PARAM0 && Imm <= AMDGPU::Exp::ET_PARAM31)
1452       ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_PARAM_ACCESS, Inst);
1453     else if (Imm >= AMDGPU::Exp::ET_POS0 && Imm <= AMDGPU::Exp::ET_POS_LAST)
1454       ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_POS_ACCESS, Inst);
1455     else
1456       ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_GPR_LOCK, Inst);
1457   } else {
1458     switch (Inst.getOpcode()) {
1459     case AMDGPU::S_SENDMSG:
1460     case AMDGPU::S_SENDMSG_RTN_B32:
1461     case AMDGPU::S_SENDMSG_RTN_B64:
1462     case AMDGPU::S_SENDMSGHALT:
1463       ScoreBrackets->updateByEvent(TII, TRI, MRI, SQ_MESSAGE, Inst);
1464       break;
1465     case AMDGPU::S_MEMTIME:
1466     case AMDGPU::S_MEMREALTIME:
1467       ScoreBrackets->updateByEvent(TII, TRI, MRI, SMEM_ACCESS, Inst);
1468       break;
1469     }
1470   }
1471 }
1472 
1473 bool WaitcntBrackets::mergeScore(const MergeInfo &M, unsigned &Score,
1474                                  unsigned OtherScore) {
1475   unsigned MyShifted = Score <= M.OldLB ? 0 : Score + M.MyShift;
1476   unsigned OtherShifted =
1477       OtherScore <= M.OtherLB ? 0 : OtherScore + M.OtherShift;
1478   Score = std::max(MyShifted, OtherShifted);
1479   return OtherShifted > MyShifted;
1480 }
1481 
1482 /// Merge the pending events and associater score brackets of \p Other into
1483 /// this brackets status.
1484 ///
1485 /// Returns whether the merge resulted in a change that requires tighter waits
1486 /// (i.e. the merged brackets strictly dominate the original brackets).
1487 bool WaitcntBrackets::merge(const WaitcntBrackets &Other) {
1488   bool StrictDom = false;
1489 
1490   VgprUB = std::max(VgprUB, Other.VgprUB);
1491   SgprUB = std::max(SgprUB, Other.SgprUB);
1492 
1493   for (auto T : inst_counter_types()) {
1494     // Merge event flags for this counter
1495     const unsigned OldEvents = PendingEvents & WaitEventMaskForInst[T];
1496     const unsigned OtherEvents = Other.PendingEvents & WaitEventMaskForInst[T];
1497     if (OtherEvents & ~OldEvents)
1498       StrictDom = true;
1499     PendingEvents |= OtherEvents;
1500 
1501     // Merge scores for this counter
1502     const unsigned MyPending = ScoreUBs[T] - ScoreLBs[T];
1503     const unsigned OtherPending = Other.ScoreUBs[T] - Other.ScoreLBs[T];
1504     const unsigned NewUB = ScoreLBs[T] + std::max(MyPending, OtherPending);
1505     if (NewUB < ScoreLBs[T])
1506       report_fatal_error("waitcnt score overflow");
1507 
1508     MergeInfo M;
1509     M.OldLB = ScoreLBs[T];
1510     M.OtherLB = Other.ScoreLBs[T];
1511     M.MyShift = NewUB - ScoreUBs[T];
1512     M.OtherShift = NewUB - Other.ScoreUBs[T];
1513 
1514     ScoreUBs[T] = NewUB;
1515 
1516     StrictDom |= mergeScore(M, LastFlat[T], Other.LastFlat[T]);
1517 
1518     for (int J = 0; J <= VgprUB; J++)
1519       StrictDom |= mergeScore(M, VgprScores[T][J], Other.VgprScores[T][J]);
1520 
1521     if (T == LGKM_CNT) {
1522       for (int J = 0; J <= SgprUB; J++)
1523         StrictDom |= mergeScore(M, SgprScores[J], Other.SgprScores[J]);
1524     }
1525   }
1526 
1527   for (int J = 0; J <= VgprUB; J++) {
1528     unsigned char NewVmemTypes = VgprVmemTypes[J] | Other.VgprVmemTypes[J];
1529     StrictDom |= NewVmemTypes != VgprVmemTypes[J];
1530     VgprVmemTypes[J] = NewVmemTypes;
1531   }
1532 
1533   return StrictDom;
1534 }
1535 
1536 static bool isWaitInstr(MachineInstr &Inst) {
1537   return Inst.getOpcode() == AMDGPU::S_WAITCNT ||
1538          (Inst.getOpcode() == AMDGPU::S_WAITCNT_VSCNT &&
1539           Inst.getOperand(0).isReg() &&
1540           Inst.getOperand(0).getReg() == AMDGPU::SGPR_NULL);
1541 }
1542 
1543 // Generate s_waitcnt instructions where needed.
1544 bool SIInsertWaitcnts::insertWaitcntInBlock(MachineFunction &MF,
1545                                             MachineBasicBlock &Block,
1546                                             WaitcntBrackets &ScoreBrackets) {
1547   bool Modified = false;
1548 
1549   LLVM_DEBUG({
1550     dbgs() << "*** Block" << Block.getNumber() << " ***";
1551     ScoreBrackets.dump();
1552   });
1553 
1554   // Track the correctness of vccz through this basic block. There are two
1555   // reasons why it might be incorrect; see ST->hasReadVCCZBug() and
1556   // ST->partialVCCWritesUpdateVCCZ().
1557   bool VCCZCorrect = true;
1558   if (ST->hasReadVCCZBug()) {
1559     // vccz could be incorrect at a basic block boundary if a predecessor wrote
1560     // to vcc and then issued an smem load.
1561     VCCZCorrect = false;
1562   } else if (!ST->partialVCCWritesUpdateVCCZ()) {
1563     // vccz could be incorrect at a basic block boundary if a predecessor wrote
1564     // to vcc_lo or vcc_hi.
1565     VCCZCorrect = false;
1566   }
1567 
1568   // Walk over the instructions.
1569   MachineInstr *OldWaitcntInstr = nullptr;
1570 
1571   for (MachineBasicBlock::instr_iterator Iter = Block.instr_begin(),
1572                                          E = Block.instr_end();
1573        Iter != E;) {
1574     MachineInstr &Inst = *Iter;
1575 
1576     // Track pre-existing waitcnts that were added in earlier iterations or by
1577     // the memory legalizer.
1578     if (isWaitInstr(Inst)) {
1579       if (!OldWaitcntInstr)
1580         OldWaitcntInstr = &Inst;
1581       ++Iter;
1582       continue;
1583     }
1584 
1585     bool FlushVmCnt = Block.getFirstTerminator() == Inst &&
1586                       isPreheaderToFlush(Block, ScoreBrackets);
1587 
1588     // Generate an s_waitcnt instruction to be placed before Inst, if needed.
1589     Modified |= generateWaitcntInstBefore(Inst, ScoreBrackets, OldWaitcntInstr,
1590                                           FlushVmCnt);
1591     OldWaitcntInstr = nullptr;
1592 
1593     // Restore vccz if it's not known to be correct already.
1594     bool RestoreVCCZ = !VCCZCorrect && readsVCCZ(Inst);
1595 
1596     // Don't examine operands unless we need to track vccz correctness.
1597     if (ST->hasReadVCCZBug() || !ST->partialVCCWritesUpdateVCCZ()) {
1598       if (Inst.definesRegister(AMDGPU::VCC_LO) ||
1599           Inst.definesRegister(AMDGPU::VCC_HI)) {
1600         // Up to gfx9, writes to vcc_lo and vcc_hi don't update vccz.
1601         if (!ST->partialVCCWritesUpdateVCCZ())
1602           VCCZCorrect = false;
1603       } else if (Inst.definesRegister(AMDGPU::VCC)) {
1604         // There is a hardware bug on CI/SI where SMRD instruction may corrupt
1605         // vccz bit, so when we detect that an instruction may read from a
1606         // corrupt vccz bit, we need to:
1607         // 1. Insert s_waitcnt lgkm(0) to wait for all outstanding SMRD
1608         //    operations to complete.
1609         // 2. Restore the correct value of vccz by writing the current value
1610         //    of vcc back to vcc.
1611         if (ST->hasReadVCCZBug() &&
1612             ScoreBrackets.hasPendingEvent(SMEM_ACCESS)) {
1613           // Writes to vcc while there's an outstanding smem read may get
1614           // clobbered as soon as any read completes.
1615           VCCZCorrect = false;
1616         } else {
1617           // Writes to vcc will fix any incorrect value in vccz.
1618           VCCZCorrect = true;
1619         }
1620       }
1621     }
1622 
1623     if (TII->isSMRD(Inst)) {
1624       for (const MachineMemOperand *Memop : Inst.memoperands()) {
1625         // No need to handle invariant loads when avoiding WAR conflicts, as
1626         // there cannot be a vector store to the same memory location.
1627         if (!Memop->isInvariant()) {
1628           const Value *Ptr = Memop->getValue();
1629           SLoadAddresses.insert(std::pair(Ptr, Inst.getParent()));
1630         }
1631       }
1632       if (ST->hasReadVCCZBug()) {
1633         // This smem read could complete and clobber vccz at any time.
1634         VCCZCorrect = false;
1635       }
1636     }
1637 
1638     updateEventWaitcntAfter(Inst, &ScoreBrackets);
1639 
1640 #if 0 // TODO: implement resource type check controlled by options with ub = LB.
1641     // If this instruction generates a S_SETVSKIP because it is an
1642     // indexed resource, and we are on Tahiti, then it will also force
1643     // an S_WAITCNT vmcnt(0)
1644     if (RequireCheckResourceType(Inst, context)) {
1645       // Force the score to as if an S_WAITCNT vmcnt(0) is emitted.
1646       ScoreBrackets->setScoreLB(VM_CNT,
1647       ScoreBrackets->getScoreUB(VM_CNT));
1648     }
1649 #endif
1650 
1651     LLVM_DEBUG({
1652       Inst.print(dbgs());
1653       ScoreBrackets.dump();
1654     });
1655 
1656     // TODO: Remove this work-around after fixing the scheduler and enable the
1657     // assert above.
1658     if (RestoreVCCZ) {
1659       // Restore the vccz bit.  Any time a value is written to vcc, the vcc
1660       // bit is updated, so we can restore the bit by reading the value of
1661       // vcc and then writing it back to the register.
1662       BuildMI(Block, Inst, Inst.getDebugLoc(),
1663               TII->get(ST->isWave32() ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64),
1664               TRI->getVCC())
1665           .addReg(TRI->getVCC());
1666       VCCZCorrect = true;
1667       Modified = true;
1668     }
1669 
1670     ++Iter;
1671   }
1672 
1673   if (Block.getFirstTerminator() == Block.end() &&
1674       isPreheaderToFlush(Block, ScoreBrackets))
1675     Modified |= generateWaitcntBlockEnd(Block, ScoreBrackets, OldWaitcntInstr);
1676 
1677   return Modified;
1678 }
1679 
1680 // Return true if the given machine basic block is a preheader of a loop in
1681 // which we want to flush the vmcnt counter, and false otherwise.
1682 bool SIInsertWaitcnts::isPreheaderToFlush(MachineBasicBlock &MBB,
1683                                           WaitcntBrackets &ScoreBrackets) {
1684   if (PreheadersToFlush.count(&MBB))
1685     return PreheadersToFlush[&MBB];
1686 
1687   auto UpdateCache = [&](bool val) {
1688     PreheadersToFlush[&MBB] = val;
1689     return val;
1690   };
1691 
1692   MachineBasicBlock *Succ = MBB.getSingleSuccessor();
1693   if (!Succ)
1694     return UpdateCache(false);
1695 
1696   MachineLoop *Loop = MLI->getLoopFor(Succ);
1697   if (!Loop)
1698     return UpdateCache(false);
1699 
1700   if (Loop->getLoopPreheader() == &MBB && shouldFlushVmCnt(Loop, ScoreBrackets))
1701     return UpdateCache(true);
1702 
1703   return UpdateCache(false);
1704 }
1705 
1706 // Return true if it is better to flush the vmcnt counter in the preheader of
1707 // the given loop. We currently decide to flush in two situations:
1708 // 1. The loop contains vmem store(s), no vmem load and at least one use of a
1709 //    vgpr containing a value that is loaded outside of the loop. (Only on
1710 //    targets with no vscnt counter).
1711 // 2. The loop contains vmem load(s), but the loaded values are not used in the
1712 //    loop, and at least one use of a vgpr containing a value that is loaded
1713 //    outside of the loop.
1714 bool SIInsertWaitcnts::shouldFlushVmCnt(MachineLoop *ML,
1715                                         WaitcntBrackets &Brackets) {
1716   bool HasVMemLoad = false;
1717   bool HasVMemStore = false;
1718   bool UsesVgprLoadedOutside = false;
1719   DenseSet<Register> VgprUse;
1720   DenseSet<Register> VgprDef;
1721 
1722   for (MachineBasicBlock *MBB : ML->blocks()) {
1723     for (MachineInstr &MI : *MBB) {
1724       if (SIInstrInfo::isVMEM(MI)) {
1725         if (MI.mayLoad())
1726           HasVMemLoad = true;
1727         if (MI.mayStore())
1728           HasVMemStore = true;
1729       }
1730       for (unsigned I = 0; I < MI.getNumOperands(); I++) {
1731         MachineOperand &Op = MI.getOperand(I);
1732         if (!Op.isReg() || !TRI->isVectorRegister(*MRI, Op.getReg()))
1733           continue;
1734         RegInterval Interval = Brackets.getRegInterval(&MI, TII, MRI, TRI, I);
1735         // Vgpr use
1736         if (Op.isUse()) {
1737           for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
1738             // If we find a register that is loaded inside the loop, 1. and 2.
1739             // are invalidated and we can exit.
1740             if (VgprDef.contains(RegNo))
1741               return false;
1742             VgprUse.insert(RegNo);
1743             // If at least one of Op's registers is in the score brackets, the
1744             // value is likely loaded outside of the loop.
1745             if (Brackets.getRegScore(RegNo, VM_CNT) > Brackets.getScoreLB(VM_CNT)) {
1746               UsesVgprLoadedOutside = true;
1747               break;
1748             }
1749           }
1750         }
1751         // VMem load vgpr def
1752         else if (SIInstrInfo::isVMEM(MI) && MI.mayLoad() && Op.isDef())
1753           for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
1754             // If we find a register that is loaded inside the loop, 1. and 2.
1755             // are invalidated and we can exit.
1756             if (VgprUse.contains(RegNo))
1757               return false;
1758             VgprDef.insert(RegNo);
1759           }
1760       }
1761     }
1762   }
1763   if (!ST->hasVscnt() && HasVMemStore && !HasVMemLoad && UsesVgprLoadedOutside)
1764     return true;
1765   return HasVMemLoad && UsesVgprLoadedOutside;
1766 }
1767 
1768 bool SIInsertWaitcnts::runOnMachineFunction(MachineFunction &MF) {
1769   ST = &MF.getSubtarget<GCNSubtarget>();
1770   TII = ST->getInstrInfo();
1771   TRI = &TII->getRegisterInfo();
1772   MRI = &MF.getRegInfo();
1773   IV = AMDGPU::getIsaVersion(ST->getCPU());
1774   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1775   MLI = &getAnalysis<MachineLoopInfo>();
1776   PDT = &getAnalysis<MachinePostDominatorTree>();
1777 
1778   ForceEmitZeroWaitcnts = ForceEmitZeroFlag;
1779   for (auto T : inst_counter_types())
1780     ForceEmitWaitcnt[T] = false;
1781 
1782   HardwareLimits Limits = {};
1783   Limits.VmcntMax = AMDGPU::getVmcntBitMask(IV);
1784   Limits.ExpcntMax = AMDGPU::getExpcntBitMask(IV);
1785   Limits.LgkmcntMax = AMDGPU::getLgkmcntBitMask(IV);
1786   Limits.VscntMax = ST->hasVscnt() ? 63 : 0;
1787 
1788   unsigned NumVGPRsMax = ST->getAddressableNumVGPRs();
1789   unsigned NumSGPRsMax = ST->getAddressableNumSGPRs();
1790   assert(NumVGPRsMax <= SQ_MAX_PGM_VGPRS);
1791   assert(NumSGPRsMax <= SQ_MAX_PGM_SGPRS);
1792 
1793   RegisterEncoding Encoding = {};
1794   Encoding.VGPR0 = TRI->getEncodingValue(AMDGPU::VGPR0);
1795   Encoding.VGPRL = Encoding.VGPR0 + NumVGPRsMax - 1;
1796   Encoding.SGPR0 = TRI->getEncodingValue(AMDGPU::SGPR0);
1797   Encoding.SGPRL = Encoding.SGPR0 + NumSGPRsMax - 1;
1798 
1799   TrackedWaitcntSet.clear();
1800   BlockInfos.clear();
1801   bool Modified = false;
1802 
1803   if (!MFI->isEntryFunction()) {
1804     // Wait for any outstanding memory operations that the input registers may
1805     // depend on. We can't track them and it's better to do the wait after the
1806     // costly call sequence.
1807 
1808     // TODO: Could insert earlier and schedule more liberally with operations
1809     // that only use caller preserved registers.
1810     MachineBasicBlock &EntryBB = MF.front();
1811     MachineBasicBlock::iterator I = EntryBB.begin();
1812     for (MachineBasicBlock::iterator E = EntryBB.end();
1813          I != E && (I->isPHI() || I->isMetaInstruction()); ++I)
1814       ;
1815     BuildMI(EntryBB, I, DebugLoc(), TII->get(AMDGPU::S_WAITCNT)).addImm(0);
1816     if (ST->hasVscnt())
1817       BuildMI(EntryBB, I, DebugLoc(), TII->get(AMDGPU::S_WAITCNT_VSCNT))
1818           .addReg(AMDGPU::SGPR_NULL, RegState::Undef)
1819           .addImm(0);
1820 
1821     Modified = true;
1822   }
1823 
1824   // Keep iterating over the blocks in reverse post order, inserting and
1825   // updating s_waitcnt where needed, until a fix point is reached.
1826   for (auto *MBB : ReversePostOrderTraversal<MachineFunction *>(&MF))
1827     BlockInfos.insert({MBB, BlockInfo(MBB)});
1828 
1829   std::unique_ptr<WaitcntBrackets> Brackets;
1830   bool Repeat;
1831   do {
1832     Repeat = false;
1833 
1834     for (auto BII = BlockInfos.begin(), BIE = BlockInfos.end(); BII != BIE;
1835          ++BII) {
1836       BlockInfo &BI = BII->second;
1837       if (!BI.Dirty)
1838         continue;
1839 
1840       if (BI.Incoming) {
1841         if (!Brackets)
1842           Brackets = std::make_unique<WaitcntBrackets>(*BI.Incoming);
1843         else
1844           *Brackets = *BI.Incoming;
1845       } else {
1846         if (!Brackets)
1847           Brackets = std::make_unique<WaitcntBrackets>(ST, Limits, Encoding);
1848         else
1849           *Brackets = WaitcntBrackets(ST, Limits, Encoding);
1850       }
1851 
1852       Modified |= insertWaitcntInBlock(MF, *BI.MBB, *Brackets);
1853       BI.Dirty = false;
1854 
1855       if (Brackets->hasPendingEvent()) {
1856         BlockInfo *MoveBracketsToSucc = nullptr;
1857         for (MachineBasicBlock *Succ : BI.MBB->successors()) {
1858           auto SuccBII = BlockInfos.find(Succ);
1859           BlockInfo &SuccBI = SuccBII->second;
1860           if (!SuccBI.Incoming) {
1861             SuccBI.Dirty = true;
1862             if (SuccBII <= BII)
1863               Repeat = true;
1864             if (!MoveBracketsToSucc) {
1865               MoveBracketsToSucc = &SuccBI;
1866             } else {
1867               SuccBI.Incoming = std::make_unique<WaitcntBrackets>(*Brackets);
1868             }
1869           } else if (SuccBI.Incoming->merge(*Brackets)) {
1870             SuccBI.Dirty = true;
1871             if (SuccBII <= BII)
1872               Repeat = true;
1873           }
1874         }
1875         if (MoveBracketsToSucc)
1876           MoveBracketsToSucc->Incoming = std::move(Brackets);
1877       }
1878     }
1879   } while (Repeat);
1880 
1881   if (ST->hasScalarStores()) {
1882     SmallVector<MachineBasicBlock *, 4> EndPgmBlocks;
1883     bool HaveScalarStores = false;
1884 
1885     for (MachineBasicBlock &MBB : MF) {
1886       for (MachineInstr &MI : MBB) {
1887         if (!HaveScalarStores && TII->isScalarStore(MI))
1888           HaveScalarStores = true;
1889 
1890         if (MI.getOpcode() == AMDGPU::S_ENDPGM ||
1891             MI.getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG)
1892           EndPgmBlocks.push_back(&MBB);
1893       }
1894     }
1895 
1896     if (HaveScalarStores) {
1897       // If scalar writes are used, the cache must be flushed or else the next
1898       // wave to reuse the same scratch memory can be clobbered.
1899       //
1900       // Insert s_dcache_wb at wave termination points if there were any scalar
1901       // stores, and only if the cache hasn't already been flushed. This could
1902       // be improved by looking across blocks for flushes in postdominating
1903       // blocks from the stores but an explicitly requested flush is probably
1904       // very rare.
1905       for (MachineBasicBlock *MBB : EndPgmBlocks) {
1906         bool SeenDCacheWB = false;
1907 
1908         for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
1909              I != E; ++I) {
1910           if (I->getOpcode() == AMDGPU::S_DCACHE_WB)
1911             SeenDCacheWB = true;
1912           else if (TII->isScalarStore(*I))
1913             SeenDCacheWB = false;
1914 
1915           // FIXME: It would be better to insert this before a waitcnt if any.
1916           if ((I->getOpcode() == AMDGPU::S_ENDPGM ||
1917                I->getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG) &&
1918               !SeenDCacheWB) {
1919             Modified = true;
1920             BuildMI(*MBB, I, I->getDebugLoc(), TII->get(AMDGPU::S_DCACHE_WB));
1921           }
1922         }
1923       }
1924     }
1925   }
1926 
1927   return Modified;
1928 }
1929