xref: /llvm-project/llvm/lib/Target/AMDGPU/SIOptimizeExecMaskingPreRA.cpp (revision 476e26b5d34dfffa9e6995f1bb8a562d4a2bfa52)
1 //===-- SIOptimizeExecMaskingPreRA.cpp ------------------------------------===//
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 /// This pass removes redundant S_OR_B64 instructions enabling lanes in
11 /// the exec. If two SI_END_CF (lowered as S_OR_B64) come together without any
12 /// vector instructions between them we can only keep outer SI_END_CF, given
13 /// that CFG is structured and exec bits of the outer end statement are always
14 /// not less than exec bit of the inner one.
15 ///
16 /// This needs to be done before the RA to eliminate saved exec bits registers
17 /// but after register coalescer to have no vector registers copies in between
18 /// of different end cf statements.
19 ///
20 //===----------------------------------------------------------------------===//
21 
22 #include "AMDGPU.h"
23 #include "AMDGPUSubtarget.h"
24 #include "SIInstrInfo.h"
25 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
26 #include "llvm/CodeGen/LiveIntervals.h"
27 #include "llvm/CodeGen/MachineFunctionPass.h"
28 
29 using namespace llvm;
30 
31 #define DEBUG_TYPE "si-optimize-exec-masking-pre-ra"
32 
33 namespace {
34 
35 class SIOptimizeExecMaskingPreRA : public MachineFunctionPass {
36 public:
37   static char ID;
38 
39 public:
40   SIOptimizeExecMaskingPreRA() : MachineFunctionPass(ID) {
41     initializeSIOptimizeExecMaskingPreRAPass(*PassRegistry::getPassRegistry());
42   }
43 
44   bool runOnMachineFunction(MachineFunction &MF) override;
45 
46   StringRef getPassName() const override {
47     return "SI optimize exec mask operations pre-RA";
48   }
49 
50   void getAnalysisUsage(AnalysisUsage &AU) const override {
51     AU.addRequired<LiveIntervals>();
52     AU.setPreservesAll();
53     MachineFunctionPass::getAnalysisUsage(AU);
54   }
55 };
56 
57 } // End anonymous namespace.
58 
59 INITIALIZE_PASS_BEGIN(SIOptimizeExecMaskingPreRA, DEBUG_TYPE,
60                       "SI optimize exec mask operations pre-RA", false, false)
61 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
62 INITIALIZE_PASS_END(SIOptimizeExecMaskingPreRA, DEBUG_TYPE,
63                     "SI optimize exec mask operations pre-RA", false, false)
64 
65 char SIOptimizeExecMaskingPreRA::ID = 0;
66 
67 char &llvm::SIOptimizeExecMaskingPreRAID = SIOptimizeExecMaskingPreRA::ID;
68 
69 FunctionPass *llvm::createSIOptimizeExecMaskingPreRAPass() {
70   return new SIOptimizeExecMaskingPreRA();
71 }
72 
73 static bool isEndCF(const MachineInstr& MI, const SIRegisterInfo* TRI) {
74   return MI.getOpcode() == AMDGPU::S_OR_B64 &&
75          MI.modifiesRegister(AMDGPU::EXEC, TRI);
76 }
77 
78 static bool isFullExecCopy(const MachineInstr& MI) {
79   return MI.isFullCopy() && MI.getOperand(1).getReg() == AMDGPU::EXEC;
80 }
81 
82 static unsigned getOrNonExecReg(const MachineInstr &MI,
83                                 const SIInstrInfo &TII) {
84   auto Op = TII.getNamedOperand(MI, AMDGPU::OpName::src1);
85   if (Op->isReg() && Op->getReg() != AMDGPU::EXEC)
86      return Op->getReg();
87   Op = TII.getNamedOperand(MI, AMDGPU::OpName::src0);
88   if (Op->isReg() && Op->getReg() != AMDGPU::EXEC)
89      return Op->getReg();
90   return AMDGPU::NoRegister;
91 }
92 
93 static MachineInstr* getOrExecSource(const MachineInstr &MI,
94                                      const SIInstrInfo &TII,
95                                      const MachineRegisterInfo &MRI) {
96   auto SavedExec = getOrNonExecReg(MI, TII);
97   if (SavedExec == AMDGPU::NoRegister)
98     return nullptr;
99   auto SaveExecInst = MRI.getUniqueVRegDef(SavedExec);
100   if (!SaveExecInst || !isFullExecCopy(*SaveExecInst))
101     return nullptr;
102   return SaveExecInst;
103 }
104 
105 // Optimize sequence
106 //    %sel = V_CNDMASK_B32_e64 0, 1, %cc
107 //    %cmp = V_CMP_NE_U32 1, %1
108 //    $vcc = S_AND_B64 $exec, %cmp
109 //    S_CBRANCH_VCC[N]Z
110 // =>
111 //    $vcc = S_ANDN2_B64 $exec, %cc
112 //    S_CBRANCH_VCC[N]Z
113 //
114 // It is the negation pattern inserted by DAGCombiner::visitBRCOND() in the
115 // rebuildSetCC(). We start with S_CBRANCH to avoid exhaustive search, but
116 // only 3 first instructions are really needed. S_AND_B64 with exec is a
117 // required part of the pattern since V_CNDMASK_B32 writes zeroes for inactive
118 // lanes.
119 //
120 // Returns %cc register on success.
121 static unsigned optimizeVcndVcmpPair(MachineBasicBlock &MBB,
122                                      const GCNSubtarget &ST,
123                                      MachineRegisterInfo &MRI,
124                                      LiveIntervals *LIS) {
125   const SIRegisterInfo *TRI = ST.getRegisterInfo();
126   const SIInstrInfo *TII = ST.getInstrInfo();
127   const unsigned AndOpc = AMDGPU::S_AND_B64;
128   const unsigned Andn2Opc = AMDGPU::S_ANDN2_B64;
129   const unsigned CondReg = AMDGPU::VCC;
130   const unsigned ExecReg = AMDGPU::EXEC;
131 
132   auto I = llvm::find_if(MBB.terminators(), [](const MachineInstr &MI) {
133                            unsigned Opc = MI.getOpcode();
134                            return Opc == AMDGPU::S_CBRANCH_VCCZ ||
135                                   Opc == AMDGPU::S_CBRANCH_VCCNZ; });
136   if (I == MBB.terminators().end())
137     return AMDGPU::NoRegister;
138 
139   auto *And = TRI->findReachingDef(CondReg, AMDGPU::NoSubRegister,
140                                    *I, MRI, LIS);
141   if (!And || And->getOpcode() != AndOpc ||
142       !And->getOperand(1).isReg() || !And->getOperand(2).isReg())
143     return AMDGPU::NoRegister;
144 
145   MachineOperand *AndCC = &And->getOperand(1);
146   unsigned CmpReg = AndCC->getReg();
147   unsigned CmpSubReg = AndCC->getSubReg();
148   if (CmpReg == ExecReg) {
149     AndCC = &And->getOperand(2);
150     CmpReg = AndCC->getReg();
151     CmpSubReg = AndCC->getSubReg();
152   } else if (And->getOperand(2).getReg() != ExecReg) {
153     return AMDGPU::NoRegister;
154   }
155 
156   auto *Cmp = TRI->findReachingDef(CmpReg, CmpSubReg, *And, MRI, LIS);
157   if (!Cmp || !(Cmp->getOpcode() == AMDGPU::V_CMP_NE_U32_e32 ||
158                 Cmp->getOpcode() == AMDGPU::V_CMP_NE_U32_e64) ||
159       Cmp->getParent() != And->getParent())
160     return AMDGPU::NoRegister;
161 
162   MachineOperand *Op1 = TII->getNamedOperand(*Cmp, AMDGPU::OpName::src0);
163   MachineOperand *Op2 = TII->getNamedOperand(*Cmp, AMDGPU::OpName::src1);
164   if (Op1->isImm() && Op2->isReg())
165     std::swap(Op1, Op2);
166   if (!Op1->isReg() || !Op2->isImm() || Op2->getImm() != 1)
167     return AMDGPU::NoRegister;
168 
169   unsigned SelReg = Op1->getReg();
170   auto *Sel = TRI->findReachingDef(SelReg, Op1->getSubReg(), *Cmp, MRI, LIS);
171   if (!Sel || Sel->getOpcode() != AMDGPU::V_CNDMASK_B32_e64)
172     return AMDGPU::NoRegister;
173 
174   Op1 = TII->getNamedOperand(*Sel, AMDGPU::OpName::src0);
175   Op2 = TII->getNamedOperand(*Sel, AMDGPU::OpName::src1);
176   MachineOperand *CC = TII->getNamedOperand(*Sel, AMDGPU::OpName::src2);
177   if (!Op1->isImm() || !Op2->isImm() || !CC->isReg() ||
178       Op1->getImm() != 0 || Op2->getImm() != 1)
179     return AMDGPU::NoRegister;
180 
181   LLVM_DEBUG(dbgs() << "Folding sequence:\n\t" << *Sel << '\t'
182                     << *Cmp << '\t' << *And);
183 
184   unsigned CCReg = CC->getReg();
185   LIS->RemoveMachineInstrFromMaps(*And);
186   MachineInstr *Andn2 = BuildMI(MBB, *And, And->getDebugLoc(),
187                                 TII->get(Andn2Opc), And->getOperand(0).getReg())
188                             .addReg(ExecReg)
189                             .addReg(CCReg, CC->getSubReg());
190   And->eraseFromParent();
191   LIS->InsertMachineInstrInMaps(*Andn2);
192 
193   LLVM_DEBUG(dbgs() << "=>\n\t" << *Andn2 << '\n');
194 
195   // Try to remove compare. Cmp value should not used in between of cmp
196   // and s_and_b64 if VCC or just unused if any other register.
197   if ((TargetRegisterInfo::isVirtualRegister(CmpReg) &&
198        MRI.use_nodbg_empty(CmpReg)) ||
199       (CmpReg == CondReg &&
200        std::none_of(std::next(Cmp->getIterator()), Andn2->getIterator(),
201                     [&](const MachineInstr &MI) {
202                       return MI.readsRegister(CondReg, TRI); }))) {
203     LLVM_DEBUG(dbgs() << "Erasing: " << *Cmp << '\n');
204 
205     LIS->RemoveMachineInstrFromMaps(*Cmp);
206     Cmp->eraseFromParent();
207 
208     // Try to remove v_cndmask_b32.
209     if (TargetRegisterInfo::isVirtualRegister(SelReg) &&
210         MRI.use_nodbg_empty(SelReg)) {
211       LLVM_DEBUG(dbgs() << "Erasing: " << *Sel << '\n');
212 
213       LIS->RemoveMachineInstrFromMaps(*Sel);
214       Sel->eraseFromParent();
215     }
216   }
217 
218   return CCReg;
219 }
220 
221 bool SIOptimizeExecMaskingPreRA::runOnMachineFunction(MachineFunction &MF) {
222   if (skipFunction(MF.getFunction()))
223     return false;
224 
225   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
226   const SIRegisterInfo *TRI = ST.getRegisterInfo();
227   const SIInstrInfo *TII = ST.getInstrInfo();
228   MachineRegisterInfo &MRI = MF.getRegInfo();
229   LiveIntervals *LIS = &getAnalysis<LiveIntervals>();
230   DenseSet<unsigned> RecalcRegs({AMDGPU::EXEC_LO, AMDGPU::EXEC_HI});
231   bool Changed = false;
232 
233   for (MachineBasicBlock &MBB : MF) {
234 
235     if (unsigned Reg = optimizeVcndVcmpPair(MBB, ST, MRI, LIS)) {
236       RecalcRegs.insert(Reg);
237       RecalcRegs.insert(AMDGPU::VCC_LO);
238       RecalcRegs.insert(AMDGPU::VCC_HI);
239       RecalcRegs.insert(AMDGPU::SCC);
240       Changed = true;
241     }
242 
243     // Try to remove unneeded instructions before s_endpgm.
244     if (MBB.succ_empty()) {
245       if (MBB.empty())
246         continue;
247 
248       // Skip this if the endpgm has any implicit uses, otherwise we would need
249       // to be careful to update / remove them.
250       MachineInstr &Term = MBB.back();
251       if (Term.getOpcode() != AMDGPU::S_ENDPGM ||
252           Term.getNumOperands() != 0)
253         continue;
254 
255       SmallVector<MachineBasicBlock*, 4> Blocks({&MBB});
256 
257       while (!Blocks.empty()) {
258         auto CurBB = Blocks.pop_back_val();
259         auto I = CurBB->rbegin(), E = CurBB->rend();
260         if (I != E) {
261           if (I->isUnconditionalBranch() || I->getOpcode() == AMDGPU::S_ENDPGM)
262             ++I;
263           else if (I->isBranch())
264             continue;
265         }
266 
267         while (I != E) {
268           if (I->isDebugInstr()) {
269             I = std::next(I);
270             continue;
271           }
272 
273           if (I->mayStore() || I->isBarrier() || I->isCall() ||
274               I->hasUnmodeledSideEffects() || I->hasOrderedMemoryRef())
275             break;
276 
277           LLVM_DEBUG(dbgs()
278                      << "Removing no effect instruction: " << *I << '\n');
279 
280           for (auto &Op : I->operands()) {
281             if (Op.isReg())
282               RecalcRegs.insert(Op.getReg());
283           }
284 
285           auto Next = std::next(I);
286           LIS->RemoveMachineInstrFromMaps(*I);
287           I->eraseFromParent();
288           I = Next;
289 
290           Changed = true;
291         }
292 
293         if (I != E)
294           continue;
295 
296         // Try to ascend predecessors.
297         for (auto *Pred : CurBB->predecessors()) {
298           if (Pred->succ_size() == 1)
299             Blocks.push_back(Pred);
300         }
301       }
302       continue;
303     }
304 
305     // Try to collapse adjacent endifs.
306     auto Lead = MBB.begin(), E = MBB.end();
307     if (MBB.succ_size() != 1 || Lead == E || !isEndCF(*Lead, TRI))
308       continue;
309 
310     const MachineBasicBlock* Succ = *MBB.succ_begin();
311     if (!MBB.isLayoutSuccessor(Succ))
312       continue;
313 
314     auto I = std::next(Lead);
315 
316     for ( ; I != E; ++I)
317       if (!TII->isSALU(*I) || I->readsRegister(AMDGPU::EXEC, TRI))
318         break;
319 
320     if (I != E)
321       continue;
322 
323     const auto NextLead = Succ->begin();
324     if (NextLead == Succ->end() || !isEndCF(*NextLead, TRI) ||
325         !getOrExecSource(*NextLead, *TII, MRI))
326       continue;
327 
328     LLVM_DEBUG(dbgs() << "Redundant EXEC = S_OR_B64 found: " << *Lead << '\n');
329 
330     auto SaveExec = getOrExecSource(*Lead, *TII, MRI);
331     unsigned SaveExecReg = getOrNonExecReg(*Lead, *TII);
332     for (auto &Op : Lead->operands()) {
333       if (Op.isReg())
334         RecalcRegs.insert(Op.getReg());
335     }
336 
337     LIS->RemoveMachineInstrFromMaps(*Lead);
338     Lead->eraseFromParent();
339     if (SaveExecReg) {
340       LIS->removeInterval(SaveExecReg);
341       LIS->createAndComputeVirtRegInterval(SaveExecReg);
342     }
343 
344     Changed = true;
345 
346     // If the only use of saved exec in the removed instruction is S_AND_B64
347     // fold the copy now.
348     if (!SaveExec || !SaveExec->isFullCopy())
349       continue;
350 
351     unsigned SavedExec = SaveExec->getOperand(0).getReg();
352     bool SafeToReplace = true;
353     for (auto& U : MRI.use_nodbg_instructions(SavedExec)) {
354       if (U.getParent() != SaveExec->getParent()) {
355         SafeToReplace = false;
356         break;
357       }
358 
359       LLVM_DEBUG(dbgs() << "Redundant EXEC COPY: " << *SaveExec << '\n');
360     }
361 
362     if (SafeToReplace) {
363       LIS->RemoveMachineInstrFromMaps(*SaveExec);
364       SaveExec->eraseFromParent();
365       MRI.replaceRegWith(SavedExec, AMDGPU::EXEC);
366       LIS->removeInterval(SavedExec);
367     }
368   }
369 
370   if (Changed) {
371     for (auto Reg : RecalcRegs) {
372       if (TargetRegisterInfo::isVirtualRegister(Reg)) {
373         LIS->removeInterval(Reg);
374         if (!MRI.reg_empty(Reg))
375           LIS->createAndComputeVirtRegInterval(Reg);
376       } else {
377         LIS->removeAllRegUnitsForPhysReg(Reg);
378       }
379     }
380   }
381 
382   return Changed;
383 }
384