xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUInsertDelayAlu.cpp (revision 7ef62cebc2f965b0f640263e179276928885e33d)
1 //===- AMDGPUInsertDelayAlu.cpp - Insert s_delay_alu 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 s_delay_alu instructions to avoid stalls on GFX11+.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "AMDGPU.h"
15 #include "GCNSubtarget.h"
16 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
17 #include "SIInstrInfo.h"
18 #include "llvm/ADT/SetVector.h"
19 
20 using namespace llvm;
21 
22 #define DEBUG_TYPE "amdgpu-insert-delay-alu"
23 
24 namespace {
25 
26 class AMDGPUInsertDelayAlu : public MachineFunctionPass {
27 public:
28   static char ID;
29 
30   const SIInstrInfo *SII;
31   const TargetRegisterInfo *TRI;
32 
33   TargetSchedModel SchedModel;
34 
35   AMDGPUInsertDelayAlu() : MachineFunctionPass(ID) {}
36 
37   void getAnalysisUsage(AnalysisUsage &AU) const override {
38     AU.setPreservesCFG();
39     MachineFunctionPass::getAnalysisUsage(AU);
40   }
41 
42   // Return true if MI waits for all outstanding VALU instructions to complete.
43   static bool instructionWaitsForVALU(const MachineInstr &MI) {
44     // These instruction types wait for VA_VDST==0 before issuing.
45     const uint64_t VA_VDST_0 = SIInstrFlags::DS | SIInstrFlags::EXP |
46                                SIInstrFlags::FLAT | SIInstrFlags::MIMG |
47                                SIInstrFlags::MTBUF | SIInstrFlags::MUBUF;
48     if (MI.getDesc().TSFlags & VA_VDST_0)
49       return true;
50     if (MI.getOpcode() == AMDGPU::S_SENDMSG_RTN_B32 ||
51         MI.getOpcode() == AMDGPU::S_SENDMSG_RTN_B64)
52       return true;
53     if (MI.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR &&
54         (MI.getOperand(0).getImm() & 0xf000) == 0)
55       return true;
56     return false;
57   }
58 
59   // Types of delay that can be encoded in an s_delay_alu instruction.
60   enum DelayType { VALU, TRANS, SALU, OTHER };
61 
62   // Get the delay type for an instruction with the specified TSFlags.
63   static DelayType getDelayType(uint64_t TSFlags) {
64     if (TSFlags & SIInstrFlags::TRANS)
65       return TRANS;
66     if (TSFlags & SIInstrFlags::VALU)
67       return VALU;
68     if (TSFlags & SIInstrFlags::SALU)
69       return SALU;
70     return OTHER;
71   }
72 
73   // Information about the last instruction(s) that wrote to a particular
74   // regunit. In straight-line code there will only be one such instruction, but
75   // when control flow converges we merge the delay information from each path
76   // to represent the union of the worst-case delays of each type.
77   struct DelayInfo {
78     // One larger than the maximum number of (non-TRANS) VALU instructions we
79     // can encode in an s_delay_alu instruction.
80     static const unsigned VALU_MAX = 5;
81 
82     // One larger than the maximum number of TRANS instructions we can encode in
83     // an s_delay_alu instruction.
84     static const unsigned TRANS_MAX = 4;
85 
86     // If it was written by a (non-TRANS) VALU, remember how many clock cycles
87     // are left until it completes, and how many other (non-TRANS) VALU we have
88     // seen since it was issued.
89     uint8_t VALUCycles = 0;
90     uint8_t VALUNum = VALU_MAX;
91 
92     // If it was written by a TRANS, remember how many clock cycles are left
93     // until it completes, and how many other TRANS we have seen since it was
94     // issued.
95     uint8_t TRANSCycles = 0;
96     uint8_t TRANSNum = TRANS_MAX;
97     // Also remember how many other (non-TRANS) VALU we have seen since it was
98     // issued. When an instruction depends on both a prior TRANS and a prior
99     // non-TRANS VALU, this is used to decide whether to encode a wait for just
100     // one or both of them.
101     uint8_t TRANSNumVALU = VALU_MAX;
102 
103     // If it was written by an SALU, remember how many clock cycles are left
104     // until it completes.
105     uint8_t SALUCycles = 0;
106 
107     DelayInfo() = default;
108 
109     DelayInfo(DelayType Type, unsigned Cycles) {
110       switch (Type) {
111       default:
112         llvm_unreachable("unexpected type");
113       case VALU:
114         VALUCycles = Cycles;
115         VALUNum = 0;
116         break;
117       case TRANS:
118         TRANSCycles = Cycles;
119         TRANSNum = 0;
120         TRANSNumVALU = 0;
121         break;
122       case SALU:
123         SALUCycles = Cycles;
124         break;
125       }
126     }
127 
128     bool operator==(const DelayInfo &RHS) const {
129       return VALUCycles == RHS.VALUCycles && VALUNum == RHS.VALUNum &&
130              TRANSCycles == RHS.TRANSCycles && TRANSNum == RHS.TRANSNum &&
131              TRANSNumVALU == RHS.TRANSNumVALU && SALUCycles == RHS.SALUCycles;
132     }
133 
134     bool operator!=(const DelayInfo &RHS) const { return !(*this == RHS); }
135 
136     // Merge another DelayInfo into this one, to represent the union of the
137     // worst-case delays of each type.
138     void merge(const DelayInfo &RHS) {
139       VALUCycles = std::max(VALUCycles, RHS.VALUCycles);
140       VALUNum = std::min(VALUNum, RHS.VALUNum);
141       TRANSCycles = std::max(TRANSCycles, RHS.TRANSCycles);
142       TRANSNum = std::min(TRANSNum, RHS.TRANSNum);
143       TRANSNumVALU = std::min(TRANSNumVALU, RHS.TRANSNumVALU);
144       SALUCycles = std::max(SALUCycles, RHS.SALUCycles);
145     }
146 
147     // Update this DelayInfo after issuing an instruction. IsVALU should be 1
148     // when issuing a (non-TRANS) VALU, else 0. IsTRANS should be 1 when issuing
149     // a TRANS, else 0. Cycles is the number of cycles it takes to issue the
150     // instruction.  Return true if there is no longer any useful delay info.
151     bool advance(DelayType Type, unsigned Cycles) {
152       bool Erase = true;
153 
154       VALUNum += (Type == VALU);
155       if (VALUNum >= VALU_MAX || VALUCycles <= Cycles) {
156         // Forget about the VALU instruction. It was too far back or has
157         // definitely completed by now.
158         VALUNum = VALU_MAX;
159         VALUCycles = 0;
160       } else {
161         VALUCycles -= Cycles;
162         Erase = false;
163       }
164 
165       TRANSNum += (Type == TRANS);
166       TRANSNumVALU += (Type == VALU);
167       if (TRANSNum >= TRANS_MAX || TRANSCycles <= Cycles) {
168         // Forget about any TRANS instruction. It was too far back or has
169         // definitely completed by now.
170         TRANSNum = TRANS_MAX;
171         TRANSNumVALU = VALU_MAX;
172         TRANSCycles = 0;
173       } else {
174         TRANSCycles -= Cycles;
175         Erase = false;
176       }
177 
178       if (SALUCycles <= Cycles) {
179         // Forget about any SALU instruction. It has definitely completed by
180         // now.
181         SALUCycles = 0;
182       } else {
183         SALUCycles -= Cycles;
184         Erase = false;
185       }
186 
187       return Erase;
188     }
189 
190 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
191     void dump() const {
192       if (VALUCycles)
193         dbgs() << " VALUCycles=" << (int)VALUCycles;
194       if (VALUNum < VALU_MAX)
195         dbgs() << " VALUNum=" << (int)VALUNum;
196       if (TRANSCycles)
197         dbgs() << " TRANSCycles=" << (int)TRANSCycles;
198       if (TRANSNum < TRANS_MAX)
199         dbgs() << " TRANSNum=" << (int)TRANSNum;
200       if (TRANSNumVALU < VALU_MAX)
201         dbgs() << " TRANSNumVALU=" << (int)TRANSNumVALU;
202       if (SALUCycles)
203         dbgs() << " SALUCycles=" << (int)SALUCycles;
204     }
205 #endif
206   };
207 
208   // A map from regunits to the delay info for that regunit.
209   struct DelayState : DenseMap<unsigned, DelayInfo> {
210     // Merge another DelayState into this one by merging the delay info for each
211     // regunit.
212     void merge(const DelayState &RHS) {
213       for (const auto &KV : RHS) {
214         iterator It;
215         bool Inserted;
216         std::tie(It, Inserted) = insert(KV);
217         if (!Inserted)
218           It->second.merge(KV.second);
219       }
220     }
221 
222     // Advance the delay info for each regunit, erasing any that are no longer
223     // useful.
224     void advance(DelayType Type, unsigned Cycles) {
225       iterator Next;
226       for (auto I = begin(), E = end(); I != E; I = Next) {
227         Next = std::next(I);
228         if (I->second.advance(Type, Cycles))
229           erase(I);
230       }
231     }
232 
233 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
234     void dump(const TargetRegisterInfo *TRI) const {
235       if (empty()) {
236         dbgs() << "    empty\n";
237         return;
238       }
239 
240       // Dump DelayInfo for each RegUnit in numerical order.
241       SmallVector<const_iterator, 8> Order;
242       Order.reserve(size());
243       for (const_iterator I = begin(), E = end(); I != E; ++I)
244         Order.push_back(I);
245       llvm::sort(Order, [](const const_iterator &A, const const_iterator &B) {
246         return A->first < B->first;
247       });
248       for (const_iterator I : Order) {
249         dbgs() << "    " << printRegUnit(I->first, TRI);
250         I->second.dump();
251         dbgs() << "\n";
252       }
253     }
254 #endif
255   };
256 
257   // The saved delay state at the end of each basic block.
258   DenseMap<MachineBasicBlock *, DelayState> BlockState;
259 
260   // Emit an s_delay_alu instruction if necessary before MI.
261   MachineInstr *emitDelayAlu(MachineInstr &MI, DelayInfo Delay,
262                              MachineInstr *LastDelayAlu) {
263     unsigned Imm = 0;
264 
265     // Wait for a TRANS instruction.
266     if (Delay.TRANSNum < DelayInfo::TRANS_MAX)
267       Imm |= 4 + Delay.TRANSNum;
268 
269     // Wait for a VALU instruction (if it's more recent than any TRANS
270     // instruction that we're also waiting for).
271     if (Delay.VALUNum < DelayInfo::VALU_MAX &&
272         Delay.VALUNum <= Delay.TRANSNumVALU) {
273       if (Imm & 0xf)
274         Imm |= Delay.VALUNum << 7;
275       else
276         Imm |= Delay.VALUNum;
277     }
278 
279     // Wait for an SALU instruction.
280     if (Delay.SALUCycles) {
281       if (Imm & 0x780) {
282         // We have already encoded a VALU and a TRANS delay. There's no room in
283         // the encoding for an SALU delay as well, so just drop it.
284       } else if (Imm & 0xf) {
285         Imm |= (Delay.SALUCycles + 8) << 7;
286       } else {
287         Imm |= Delay.SALUCycles + 8;
288       }
289     }
290 
291     // Don't emit the s_delay_alu instruction if there's nothing to wait for.
292     if (!Imm)
293       return LastDelayAlu;
294 
295     // If we only need to wait for one instruction, try encoding it in the last
296     // s_delay_alu that we emitted.
297     if (!(Imm & 0x780) && LastDelayAlu) {
298       unsigned Skip = 0;
299       for (auto I = MachineBasicBlock::instr_iterator(LastDelayAlu),
300                 E = MachineBasicBlock::instr_iterator(MI);
301            ++I != E;) {
302         if (!I->isBundle() && !I->isMetaInstruction())
303           ++Skip;
304       }
305       if (Skip < 6) {
306         MachineOperand &Op = LastDelayAlu->getOperand(0);
307         unsigned LastImm = Op.getImm();
308         assert((LastImm & ~0xf) == 0 &&
309                "Remembered an s_delay_alu with no room for another delay!");
310         LastImm |= Imm << 7 | Skip << 4;
311         Op.setImm(LastImm);
312         return nullptr;
313       }
314     }
315 
316     auto &MBB = *MI.getParent();
317     MachineInstr *DelayAlu =
318         BuildMI(MBB, MI, DebugLoc(), SII->get(AMDGPU::S_DELAY_ALU)).addImm(Imm);
319     // Remember the s_delay_alu for next time if there is still room in it to
320     // encode another delay.
321     return (Imm & 0x780) ? nullptr : DelayAlu;
322   }
323 
324   bool runOnMachineBasicBlock(MachineBasicBlock &MBB, bool Emit) {
325     DelayState State;
326     for (auto *Pred : MBB.predecessors())
327       State.merge(BlockState[Pred]);
328 
329     LLVM_DEBUG(dbgs() << "  State at start of " << printMBBReference(MBB)
330                       << "\n";
331                State.dump(TRI););
332 
333     bool Changed = false;
334     MachineInstr *LastDelayAlu = nullptr;
335 
336     // Iterate over the contents of bundles, but don't emit any instructions
337     // inside a bundle.
338     for (auto &MI : MBB.instrs()) {
339       if (MI.isBundle() || MI.isMetaInstruction())
340         continue;
341 
342       // Ignore some more instructions that do not generate any code.
343       switch (MI.getOpcode()) {
344       case AMDGPU::SI_RETURN_TO_EPILOG:
345         continue;
346       }
347 
348       DelayType Type = getDelayType(MI.getDesc().TSFlags);
349 
350       if (instructionWaitsForVALU(MI)) {
351         // Forget about all outstanding VALU delays.
352         State = DelayState();
353       } else if (Type != OTHER) {
354         DelayInfo Delay;
355         // TODO: Scan implicit uses too?
356         for (const auto &Op : MI.explicit_uses()) {
357           if (Op.isReg()) {
358             // One of the operands of the writelane is also the output operand.
359             // This creates the insertion of redundant delays. Hence, we have to
360             // ignore this operand.
361             if (MI.getOpcode() == AMDGPU::V_WRITELANE_B32 && Op.isTied())
362               continue;
363             for (MCRegUnitIterator UI(Op.getReg(), TRI); UI.isValid(); ++UI) {
364               auto It = State.find(*UI);
365               if (It != State.end()) {
366                 Delay.merge(It->second);
367                 State.erase(*UI);
368               }
369             }
370           }
371         }
372         if (Emit && !MI.isBundledWithPred()) {
373           // TODO: For VALU->SALU delays should we use s_delay_alu or s_nop or
374           // just ignore them?
375           LastDelayAlu = emitDelayAlu(MI, Delay, LastDelayAlu);
376         }
377       }
378 
379       if (Type != OTHER) {
380         // TODO: Scan implicit defs too?
381         for (const auto &Op : MI.defs()) {
382           unsigned Latency = SchedModel.computeOperandLatency(
383               &MI, MI.getOperandNo(&Op), nullptr, 0);
384           for (MCRegUnitIterator UI(Op.getReg(), TRI); UI.isValid(); ++UI)
385             State[*UI] = DelayInfo(Type, Latency);
386         }
387       }
388 
389       // Advance by the number of cycles it takes to issue this instruction.
390       // TODO: Use a more advanced model that accounts for instructions that
391       // take multiple cycles to issue on a particular pipeline.
392       unsigned Cycles = SIInstrInfo::getNumWaitStates(MI);
393       // TODO: In wave64 mode, double the number of cycles for VALU and VMEM
394       // instructions on the assumption that they will usually have to be issued
395       // twice?
396       State.advance(Type, Cycles);
397 
398       LLVM_DEBUG(dbgs() << "  State after " << MI; State.dump(TRI););
399     }
400 
401     if (Emit) {
402       assert(State == BlockState[&MBB] &&
403              "Basic block state should not have changed on final pass!");
404     } else if (State != BlockState[&MBB]) {
405       BlockState[&MBB] = std::move(State);
406       Changed = true;
407     }
408     return Changed;
409   }
410 
411   bool runOnMachineFunction(MachineFunction &MF) override {
412     if (skipFunction(MF.getFunction()))
413       return false;
414 
415     LLVM_DEBUG(dbgs() << "AMDGPUInsertDelayAlu running on " << MF.getName()
416                       << "\n");
417 
418     const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
419     if (!ST.hasDelayAlu())
420       return false;
421 
422     SII = ST.getInstrInfo();
423     TRI = ST.getRegisterInfo();
424 
425     SchedModel.init(&ST);
426 
427     // Calculate the delay state for each basic block, iterating until we reach
428     // a fixed point.
429     SetVector<MachineBasicBlock *> WorkList;
430     for (auto &MBB : reverse(MF))
431       WorkList.insert(&MBB);
432     while (!WorkList.empty()) {
433       auto &MBB = *WorkList.pop_back_val();
434       bool Changed = runOnMachineBasicBlock(MBB, false);
435       if (Changed)
436         WorkList.insert(MBB.succ_begin(), MBB.succ_end());
437     }
438 
439     LLVM_DEBUG(dbgs() << "Final pass over all BBs\n");
440 
441     // Make one last pass over all basic blocks to emit s_delay_alu
442     // instructions.
443     bool Changed = false;
444     for (auto &MBB : MF)
445       Changed |= runOnMachineBasicBlock(MBB, true);
446     return Changed;
447   }
448 };
449 
450 } // namespace
451 
452 char AMDGPUInsertDelayAlu::ID = 0;
453 
454 char &llvm::AMDGPUInsertDelayAluID = AMDGPUInsertDelayAlu::ID;
455 
456 INITIALIZE_PASS(AMDGPUInsertDelayAlu, DEBUG_TYPE, "AMDGPU Insert Delay ALU",
457                 false, false)
458