1 //===- MachineLICM.cpp - Machine Loop Invariant Code Motion Pass ----------===//
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 // This pass performs loop invariant code motion on machine instructions. We
10 // attempt to remove as much code from the body of a loop as possible.
11 //
12 // This pass is not intended to be a replacement or a complete alternative
13 // for the LLVM-IR-level LICM pass. It is only designed to hoist simple
14 // constructs that are not exposed before lowering and instruction selection.
15 //
16 //===----------------------------------------------------------------------===//
17
18 #include "llvm/ADT/BitVector.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/CodeGen/MachineBasicBlock.h"
26 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
27 #include "llvm/CodeGen/MachineDominators.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineFunctionPass.h"
31 #include "llvm/CodeGen/MachineInstr.h"
32 #include "llvm/CodeGen/MachineLoopInfo.h"
33 #include "llvm/CodeGen/MachineMemOperand.h"
34 #include "llvm/CodeGen/MachineOperand.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/CodeGen/TargetInstrInfo.h"
38 #include "llvm/CodeGen/TargetLowering.h"
39 #include "llvm/CodeGen/TargetRegisterInfo.h"
40 #include "llvm/CodeGen/TargetSchedule.h"
41 #include "llvm/CodeGen/TargetSubtargetInfo.h"
42 #include "llvm/IR/DebugLoc.h"
43 #include "llvm/InitializePasses.h"
44 #include "llvm/MC/MCInstrDesc.h"
45 #include "llvm/MC/MCRegister.h"
46 #include "llvm/MC/MCRegisterInfo.h"
47 #include "llvm/Pass.h"
48 #include "llvm/Support/Casting.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include <algorithm>
53 #include <cassert>
54 #include <limits>
55 #include <vector>
56
57 using namespace llvm;
58
59 #define DEBUG_TYPE "machinelicm"
60
61 static cl::opt<bool>
62 AvoidSpeculation("avoid-speculation",
63 cl::desc("MachineLICM should avoid speculation"),
64 cl::init(true), cl::Hidden);
65
66 static cl::opt<bool>
67 HoistCheapInsts("hoist-cheap-insts",
68 cl::desc("MachineLICM should hoist even cheap instructions"),
69 cl::init(false), cl::Hidden);
70
71 static cl::opt<bool>
72 HoistConstStores("hoist-const-stores",
73 cl::desc("Hoist invariant stores"),
74 cl::init(true), cl::Hidden);
75
76 static cl::opt<bool> HoistConstLoads("hoist-const-loads",
77 cl::desc("Hoist invariant loads"),
78 cl::init(true), cl::Hidden);
79
80 // The default threshold of 100 (i.e. if target block is 100 times hotter)
81 // is based on empirical data on a single target and is subject to tuning.
82 static cl::opt<unsigned>
83 BlockFrequencyRatioThreshold("block-freq-ratio-threshold",
84 cl::desc("Do not hoist instructions if target"
85 "block is N times hotter than the source."),
86 cl::init(100), cl::Hidden);
87
88 enum class UseBFI { None, PGO, All };
89
90 static cl::opt<UseBFI>
91 DisableHoistingToHotterBlocks("disable-hoisting-to-hotter-blocks",
92 cl::desc("Disable hoisting instructions to"
93 " hotter blocks"),
94 cl::init(UseBFI::PGO), cl::Hidden,
95 cl::values(clEnumValN(UseBFI::None, "none",
96 "disable the feature"),
97 clEnumValN(UseBFI::PGO, "pgo",
98 "enable the feature when using profile data"),
99 clEnumValN(UseBFI::All, "all",
100 "enable the feature with/wo profile data")));
101
102 STATISTIC(NumHoisted,
103 "Number of machine instructions hoisted out of loops");
104 STATISTIC(NumLowRP,
105 "Number of instructions hoisted in low reg pressure situation");
106 STATISTIC(NumHighLatency,
107 "Number of high latency instructions hoisted");
108 STATISTIC(NumCSEed,
109 "Number of hoisted machine instructions CSEed");
110 STATISTIC(NumPostRAHoisted,
111 "Number of machine instructions hoisted out of loops post regalloc");
112 STATISTIC(NumStoreConst,
113 "Number of stores of const phys reg hoisted out of loops");
114 STATISTIC(NumNotHoistedDueToHotness,
115 "Number of instructions not hoisted due to block frequency");
116
117 namespace {
118 enum HoistResult { NotHoisted = 1, Hoisted = 2, ErasedMI = 4 };
119
120 class MachineLICMBase : public MachineFunctionPass {
121 const TargetInstrInfo *TII = nullptr;
122 const TargetLoweringBase *TLI = nullptr;
123 const TargetRegisterInfo *TRI = nullptr;
124 const MachineFrameInfo *MFI = nullptr;
125 MachineRegisterInfo *MRI = nullptr;
126 TargetSchedModel SchedModel;
127 bool PreRegAlloc = false;
128 bool HasProfileData = false;
129
130 // Various analyses that we use...
131 AliasAnalysis *AA = nullptr; // Alias analysis info.
132 MachineBlockFrequencyInfo *MBFI = nullptr; // Machine block frequncy info
133 MachineLoopInfo *MLI = nullptr; // Current MachineLoopInfo
134 MachineDominatorTree *DT = nullptr; // Machine dominator tree for the cur loop
135
136 // State that is updated as we process loops
137 bool Changed = false; // True if a loop is changed.
138 bool FirstInLoop = false; // True if it's the first LICM in the loop.
139
140 // Holds information about whether it is allowed to move load instructions
141 // out of the loop
142 SmallDenseMap<MachineLoop *, bool> AllowedToHoistLoads;
143
144 // Exit blocks of each Loop.
145 DenseMap<MachineLoop *, SmallVector<MachineBasicBlock *, 8>> ExitBlockMap;
146
isExitBlock(MachineLoop * CurLoop,const MachineBasicBlock * MBB)147 bool isExitBlock(MachineLoop *CurLoop, const MachineBasicBlock *MBB) {
148 if (ExitBlockMap.contains(CurLoop))
149 return is_contained(ExitBlockMap[CurLoop], MBB);
150
151 SmallVector<MachineBasicBlock *, 8> ExitBlocks;
152 CurLoop->getExitBlocks(ExitBlocks);
153 ExitBlockMap[CurLoop] = ExitBlocks;
154 return is_contained(ExitBlocks, MBB);
155 }
156
157 // Track 'estimated' register pressure.
158 SmallDenseSet<Register> RegSeen;
159 SmallVector<unsigned, 8> RegPressure;
160
161 // Register pressure "limit" per register pressure set. If the pressure
162 // is higher than the limit, then it's considered high.
163 SmallVector<unsigned, 8> RegLimit;
164
165 // Register pressure on path leading from loop preheader to current BB.
166 SmallVector<SmallVector<unsigned, 8>, 16> BackTrace;
167
168 // For each opcode per preheader, keep a list of potential CSE instructions.
169 DenseMap<MachineBasicBlock *,
170 DenseMap<unsigned, std::vector<MachineInstr *>>>
171 CSEMap;
172
173 enum {
174 SpeculateFalse = 0,
175 SpeculateTrue = 1,
176 SpeculateUnknown = 2
177 };
178
179 // If a MBB does not dominate loop exiting blocks then it may not safe
180 // to hoist loads from this block.
181 // Tri-state: 0 - false, 1 - true, 2 - unknown
182 unsigned SpeculationState = SpeculateUnknown;
183
184 public:
MachineLICMBase(char & PassID,bool PreRegAlloc)185 MachineLICMBase(char &PassID, bool PreRegAlloc)
186 : MachineFunctionPass(PassID), PreRegAlloc(PreRegAlloc) {}
187
188 bool runOnMachineFunction(MachineFunction &MF) override;
189
getAnalysisUsage(AnalysisUsage & AU) const190 void getAnalysisUsage(AnalysisUsage &AU) const override {
191 AU.addRequired<MachineLoopInfoWrapperPass>();
192 if (DisableHoistingToHotterBlocks != UseBFI::None)
193 AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
194 AU.addRequired<MachineDominatorTreeWrapperPass>();
195 AU.addRequired<AAResultsWrapperPass>();
196 AU.addPreserved<MachineLoopInfoWrapperPass>();
197 MachineFunctionPass::getAnalysisUsage(AU);
198 }
199
releaseMemory()200 void releaseMemory() override {
201 RegSeen.clear();
202 RegPressure.clear();
203 RegLimit.clear();
204 BackTrace.clear();
205 CSEMap.clear();
206 ExitBlockMap.clear();
207 }
208
209 private:
210 /// Keep track of information about hoisting candidates.
211 struct CandidateInfo {
212 MachineInstr *MI;
213 unsigned Def;
214 int FI;
215
CandidateInfo__anona9a475ae0111::MachineLICMBase::CandidateInfo216 CandidateInfo(MachineInstr *mi, unsigned def, int fi)
217 : MI(mi), Def(def), FI(fi) {}
218 };
219
220 void HoistRegionPostRA(MachineLoop *CurLoop,
221 MachineBasicBlock *CurPreheader);
222
223 void HoistPostRA(MachineInstr *MI, unsigned Def, MachineLoop *CurLoop,
224 MachineBasicBlock *CurPreheader);
225
226 void ProcessMI(MachineInstr *MI, BitVector &RUDefs, BitVector &RUClobbers,
227 SmallDenseSet<int> &StoredFIs,
228 SmallVectorImpl<CandidateInfo> &Candidates,
229 MachineLoop *CurLoop);
230
231 void AddToLiveIns(MCRegister Reg, MachineLoop *CurLoop);
232
233 bool IsLICMCandidate(MachineInstr &I, MachineLoop *CurLoop);
234
235 bool IsLoopInvariantInst(MachineInstr &I, MachineLoop *CurLoop);
236
237 bool HasLoopPHIUse(const MachineInstr *MI, MachineLoop *CurLoop);
238
239 bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx, Register Reg,
240 MachineLoop *CurLoop) const;
241
242 bool IsCheapInstruction(MachineInstr &MI) const;
243
244 bool CanCauseHighRegPressure(const DenseMap<unsigned, int> &Cost,
245 bool Cheap);
246
247 void UpdateBackTraceRegPressure(const MachineInstr *MI);
248
249 bool IsProfitableToHoist(MachineInstr &MI, MachineLoop *CurLoop);
250
251 bool IsGuaranteedToExecute(MachineBasicBlock *BB, MachineLoop *CurLoop);
252
253 bool isTriviallyReMaterializable(const MachineInstr &MI) const;
254
255 void EnterScope(MachineBasicBlock *MBB);
256
257 void ExitScope(MachineBasicBlock *MBB);
258
259 void ExitScopeIfDone(
260 MachineDomTreeNode *Node,
261 DenseMap<MachineDomTreeNode *, unsigned> &OpenChildren,
262 const DenseMap<MachineDomTreeNode *, MachineDomTreeNode *> &ParentMap);
263
264 void HoistOutOfLoop(MachineDomTreeNode *HeaderN, MachineLoop *CurLoop,
265 MachineBasicBlock *CurPreheader);
266
267 void InitRegPressure(MachineBasicBlock *BB);
268
269 DenseMap<unsigned, int> calcRegisterCost(const MachineInstr *MI,
270 bool ConsiderSeen,
271 bool ConsiderUnseenAsDef);
272
273 void UpdateRegPressure(const MachineInstr *MI,
274 bool ConsiderUnseenAsDef = false);
275
276 MachineInstr *ExtractHoistableLoad(MachineInstr *MI, MachineLoop *CurLoop);
277
278 MachineInstr *LookForDuplicate(const MachineInstr *MI,
279 std::vector<MachineInstr *> &PrevMIs);
280
281 bool
282 EliminateCSE(MachineInstr *MI,
283 DenseMap<unsigned, std::vector<MachineInstr *>>::iterator &CI);
284
285 bool MayCSE(MachineInstr *MI);
286
287 unsigned Hoist(MachineInstr *MI, MachineBasicBlock *Preheader,
288 MachineLoop *CurLoop);
289
290 void InitCSEMap(MachineBasicBlock *BB);
291
292 void InitializeLoadsHoistableLoops();
293
294 bool isTgtHotterThanSrc(MachineBasicBlock *SrcBlock,
295 MachineBasicBlock *TgtBlock);
296 MachineBasicBlock *getCurPreheader(MachineLoop *CurLoop,
297 MachineBasicBlock *CurPreheader);
298 };
299
300 class MachineLICM : public MachineLICMBase {
301 public:
302 static char ID;
MachineLICM()303 MachineLICM() : MachineLICMBase(ID, false) {
304 initializeMachineLICMPass(*PassRegistry::getPassRegistry());
305 }
306 };
307
308 class EarlyMachineLICM : public MachineLICMBase {
309 public:
310 static char ID;
EarlyMachineLICM()311 EarlyMachineLICM() : MachineLICMBase(ID, true) {
312 initializeEarlyMachineLICMPass(*PassRegistry::getPassRegistry());
313 }
314 };
315
316 } // end anonymous namespace
317
318 char MachineLICM::ID;
319 char EarlyMachineLICM::ID;
320
321 char &llvm::MachineLICMID = MachineLICM::ID;
322 char &llvm::EarlyMachineLICMID = EarlyMachineLICM::ID;
323
324 INITIALIZE_PASS_BEGIN(MachineLICM, DEBUG_TYPE,
325 "Machine Loop Invariant Code Motion", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass)326 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass)
327 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfoWrapperPass)
328 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
329 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
330 INITIALIZE_PASS_END(MachineLICM, DEBUG_TYPE,
331 "Machine Loop Invariant Code Motion", false, false)
332
333 INITIALIZE_PASS_BEGIN(EarlyMachineLICM, "early-machinelicm",
334 "Early Machine Loop Invariant Code Motion", false, false)
335 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass)
336 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfoWrapperPass)
337 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
338 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
339 INITIALIZE_PASS_END(EarlyMachineLICM, "early-machinelicm",
340 "Early Machine Loop Invariant Code Motion", false, false)
341
342 bool MachineLICMBase::runOnMachineFunction(MachineFunction &MF) {
343 if (skipFunction(MF.getFunction()))
344 return false;
345
346 Changed = FirstInLoop = false;
347 const TargetSubtargetInfo &ST = MF.getSubtarget();
348 TII = ST.getInstrInfo();
349 TLI = ST.getTargetLowering();
350 TRI = ST.getRegisterInfo();
351 MFI = &MF.getFrameInfo();
352 MRI = &MF.getRegInfo();
353 SchedModel.init(&ST);
354
355 PreRegAlloc = MRI->isSSA();
356 HasProfileData = MF.getFunction().hasProfileData();
357
358 if (PreRegAlloc)
359 LLVM_DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: ");
360 else
361 LLVM_DEBUG(dbgs() << "******** Post-regalloc Machine LICM: ");
362 LLVM_DEBUG(dbgs() << MF.getName() << " ********\n");
363
364 if (PreRegAlloc) {
365 // Estimate register pressure during pre-regalloc pass.
366 unsigned NumRPS = TRI->getNumRegPressureSets();
367 RegPressure.resize(NumRPS);
368 std::fill(RegPressure.begin(), RegPressure.end(), 0);
369 RegLimit.resize(NumRPS);
370 for (unsigned i = 0, e = NumRPS; i != e; ++i)
371 RegLimit[i] = TRI->getRegPressureSetLimit(MF, i);
372 }
373
374 // Get our Loop information...
375 if (DisableHoistingToHotterBlocks != UseBFI::None)
376 MBFI = &getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI();
377 MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI();
378 DT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree();
379 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
380
381 if (HoistConstLoads)
382 InitializeLoadsHoistableLoops();
383
384 SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end());
385 while (!Worklist.empty()) {
386 MachineLoop *CurLoop = Worklist.pop_back_val();
387 MachineBasicBlock *CurPreheader = nullptr;
388
389 if (!PreRegAlloc)
390 HoistRegionPostRA(CurLoop, CurPreheader);
391 else {
392 // CSEMap is initialized for loop header when the first instruction is
393 // being hoisted.
394 MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader());
395 FirstInLoop = true;
396 HoistOutOfLoop(N, CurLoop, CurPreheader);
397 CSEMap.clear();
398 }
399 }
400
401 return Changed;
402 }
403
404 /// Return true if instruction stores to the specified frame.
InstructionStoresToFI(const MachineInstr * MI,int FI)405 static bool InstructionStoresToFI(const MachineInstr *MI, int FI) {
406 // Check mayStore before memory operands so that e.g. DBG_VALUEs will return
407 // true since they have no memory operands.
408 if (!MI->mayStore())
409 return false;
410 // If we lost memory operands, conservatively assume that the instruction
411 // writes to all slots.
412 if (MI->memoperands_empty())
413 return true;
414 for (const MachineMemOperand *MemOp : MI->memoperands()) {
415 if (!MemOp->isStore() || !MemOp->getPseudoValue())
416 continue;
417 if (const FixedStackPseudoSourceValue *Value =
418 dyn_cast<FixedStackPseudoSourceValue>(MemOp->getPseudoValue())) {
419 if (Value->getFrameIndex() == FI)
420 return true;
421 }
422 }
423 return false;
424 }
425
applyBitsNotInRegMaskToRegUnitsMask(const TargetRegisterInfo & TRI,BitVector & RUs,const uint32_t * Mask)426 static void applyBitsNotInRegMaskToRegUnitsMask(const TargetRegisterInfo &TRI,
427 BitVector &RUs,
428 const uint32_t *Mask) {
429 // FIXME: This intentionally works in reverse due to some issues with the
430 // Register Units infrastructure.
431 //
432 // This is used to apply callee-saved-register masks to the clobbered regunits
433 // mask.
434 //
435 // The right way to approach this is to start with a BitVector full of ones,
436 // then reset all the bits of the regunits of each register that is set in the
437 // mask (registers preserved), then OR the resulting bits with the Clobbers
438 // mask. This correctly prioritizes the saved registers, so if a RU is shared
439 // between a register that is preserved, and one that is NOT preserved, that
440 // RU will not be set in the output vector (the clobbers).
441 //
442 // What we have to do for now is the opposite: we have to assume that the
443 // regunits of all registers that are NOT preserved are clobbered, even if
444 // those regunits are preserved by another register. So if a RU is shared
445 // like described previously, that RU will be set.
446 //
447 // This is to work around an issue which appears in AArch64, but isn't
448 // exclusive to that target: AArch64's Qn registers (128 bits) have Dn
449 // register (lower 64 bits). A few Dn registers are preserved by some calling
450 // conventions, but Qn and Dn share exactly the same reg units.
451 //
452 // If we do this the right way, Qn will be marked as NOT clobbered even though
453 // its upper 64 bits are NOT preserved. The conservative approach handles this
454 // correctly at the cost of some missed optimizations on other targets.
455 //
456 // This is caused by how RegUnits are handled within TableGen. Ideally, Qn
457 // should have an extra RegUnit to model the "unknown" bits not covered by the
458 // subregs.
459 BitVector RUsFromRegsNotInMask(TRI.getNumRegUnits());
460 const unsigned NumRegs = TRI.getNumRegs();
461 const unsigned MaskWords = (NumRegs + 31) / 32;
462 for (unsigned K = 0; K < MaskWords; ++K) {
463 const uint32_t Word = Mask[K];
464 for (unsigned Bit = 0; Bit < 32; ++Bit) {
465 const unsigned PhysReg = (K * 32) + Bit;
466 if (PhysReg == NumRegs)
467 break;
468
469 if (PhysReg && !((Word >> Bit) & 1)) {
470 for (MCRegUnitIterator RUI(PhysReg, &TRI); RUI.isValid(); ++RUI)
471 RUsFromRegsNotInMask.set(*RUI);
472 }
473 }
474 }
475
476 RUs |= RUsFromRegsNotInMask;
477 }
478
479 /// Examine the instruction for potentai LICM candidate. Also
480 /// gather register def and frame object update information.
ProcessMI(MachineInstr * MI,BitVector & RUDefs,BitVector & RUClobbers,SmallDenseSet<int> & StoredFIs,SmallVectorImpl<CandidateInfo> & Candidates,MachineLoop * CurLoop)481 void MachineLICMBase::ProcessMI(MachineInstr *MI, BitVector &RUDefs,
482 BitVector &RUClobbers,
483 SmallDenseSet<int> &StoredFIs,
484 SmallVectorImpl<CandidateInfo> &Candidates,
485 MachineLoop *CurLoop) {
486 bool RuledOut = false;
487 bool HasNonInvariantUse = false;
488 unsigned Def = 0;
489 for (const MachineOperand &MO : MI->operands()) {
490 if (MO.isFI()) {
491 // Remember if the instruction stores to the frame index.
492 int FI = MO.getIndex();
493 if (!StoredFIs.count(FI) &&
494 MFI->isSpillSlotObjectIndex(FI) &&
495 InstructionStoresToFI(MI, FI))
496 StoredFIs.insert(FI);
497 HasNonInvariantUse = true;
498 continue;
499 }
500
501 // We can't hoist an instruction defining a physreg that is clobbered in
502 // the loop.
503 if (MO.isRegMask()) {
504 applyBitsNotInRegMaskToRegUnitsMask(*TRI, RUClobbers, MO.getRegMask());
505 continue;
506 }
507
508 if (!MO.isReg())
509 continue;
510 Register Reg = MO.getReg();
511 if (!Reg)
512 continue;
513 assert(Reg.isPhysical() && "Not expecting virtual register!");
514
515 if (!MO.isDef()) {
516 if (!HasNonInvariantUse) {
517 for (MCRegUnitIterator RUI(Reg, TRI); RUI.isValid(); ++RUI) {
518 // If it's using a non-loop-invariant register, then it's obviously
519 // not safe to hoist.
520 if (RUDefs.test(*RUI) || RUClobbers.test(*RUI)) {
521 HasNonInvariantUse = true;
522 break;
523 }
524 }
525 }
526 continue;
527 }
528
529 if (MO.isImplicit()) {
530 for (MCRegUnitIterator RUI(Reg, TRI); RUI.isValid(); ++RUI)
531 RUClobbers.set(*RUI);
532 if (!MO.isDead())
533 // Non-dead implicit def? This cannot be hoisted.
534 RuledOut = true;
535 // No need to check if a dead implicit def is also defined by
536 // another instruction.
537 continue;
538 }
539
540 // FIXME: For now, avoid instructions with multiple defs, unless
541 // it's a dead implicit def.
542 if (Def)
543 RuledOut = true;
544 else
545 Def = Reg;
546
547 // If we have already seen another instruction that defines the same
548 // register, then this is not safe. Two defs is indicated by setting a
549 // PhysRegClobbers bit.
550 for (MCRegUnitIterator RUI(Reg, TRI); RUI.isValid(); ++RUI) {
551 if (RUDefs.test(*RUI)) {
552 RUClobbers.set(*RUI);
553 RuledOut = true;
554 } else if (RUClobbers.test(*RUI)) {
555 // MI defined register is seen defined by another instruction in
556 // the loop, it cannot be a LICM candidate.
557 RuledOut = true;
558 }
559
560 RUDefs.set(*RUI);
561 }
562 }
563
564 // Only consider reloads for now and remats which do not have register
565 // operands. FIXME: Consider unfold load folding instructions.
566 if (Def && !RuledOut) {
567 int FI = std::numeric_limits<int>::min();
568 if ((!HasNonInvariantUse && IsLICMCandidate(*MI, CurLoop)) ||
569 (TII->isLoadFromStackSlot(*MI, FI) && MFI->isSpillSlotObjectIndex(FI)))
570 Candidates.push_back(CandidateInfo(MI, Def, FI));
571 }
572 }
573
574 /// Walk the specified region of the CFG and hoist loop invariants out to the
575 /// preheader.
HoistRegionPostRA(MachineLoop * CurLoop,MachineBasicBlock * CurPreheader)576 void MachineLICMBase::HoistRegionPostRA(MachineLoop *CurLoop,
577 MachineBasicBlock *CurPreheader) {
578 MachineBasicBlock *Preheader = getCurPreheader(CurLoop, CurPreheader);
579 if (!Preheader)
580 return;
581
582 unsigned NumRegUnits = TRI->getNumRegUnits();
583 BitVector RUDefs(NumRegUnits); // RUs defined once in the loop.
584 BitVector RUClobbers(NumRegUnits); // RUs defined more than once.
585
586 SmallVector<CandidateInfo, 32> Candidates;
587 SmallDenseSet<int> StoredFIs;
588
589 // Walk the entire region, count number of defs for each register, and
590 // collect potential LICM candidates.
591 for (MachineBasicBlock *BB : CurLoop->getBlocks()) {
592 // If the header of the loop containing this basic block is a landing pad,
593 // then don't try to hoist instructions out of this loop.
594 const MachineLoop *ML = MLI->getLoopFor(BB);
595 if (ML && ML->getHeader()->isEHPad()) continue;
596
597 // Conservatively treat live-in's as an external def.
598 // FIXME: That means a reload that're reused in successor block(s) will not
599 // be LICM'ed.
600 for (const auto &LI : BB->liveins()) {
601 for (MCRegUnitIterator RUI(LI.PhysReg, TRI); RUI.isValid(); ++RUI)
602 RUDefs.set(*RUI);
603 }
604
605 // Funclet entry blocks will clobber all registers
606 if (const uint32_t *Mask = BB->getBeginClobberMask(TRI))
607 applyBitsNotInRegMaskToRegUnitsMask(*TRI, RUClobbers, Mask);
608
609 SpeculationState = SpeculateUnknown;
610 for (MachineInstr &MI : *BB)
611 ProcessMI(&MI, RUDefs, RUClobbers, StoredFIs, Candidates, CurLoop);
612 }
613
614 // Gather the registers read / clobbered by the terminator.
615 BitVector TermRUs(NumRegUnits);
616 MachineBasicBlock::iterator TI = Preheader->getFirstTerminator();
617 if (TI != Preheader->end()) {
618 for (const MachineOperand &MO : TI->operands()) {
619 if (!MO.isReg())
620 continue;
621 Register Reg = MO.getReg();
622 if (!Reg)
623 continue;
624 for (MCRegUnitIterator RUI(Reg, TRI); RUI.isValid(); ++RUI)
625 TermRUs.set(*RUI);
626 }
627 }
628
629 // Now evaluate whether the potential candidates qualify.
630 // 1. Check if the candidate defined register is defined by another
631 // instruction in the loop.
632 // 2. If the candidate is a load from stack slot (always true for now),
633 // check if the slot is stored anywhere in the loop.
634 // 3. Make sure candidate def should not clobber
635 // registers read by the terminator. Similarly its def should not be
636 // clobbered by the terminator.
637 for (CandidateInfo &Candidate : Candidates) {
638 if (Candidate.FI != std::numeric_limits<int>::min() &&
639 StoredFIs.count(Candidate.FI))
640 continue;
641
642 unsigned Def = Candidate.Def;
643 bool Safe = true;
644 for (MCRegUnitIterator RUI(Def, TRI); RUI.isValid(); ++RUI) {
645 if (RUClobbers.test(*RUI) || TermRUs.test(*RUI)) {
646 Safe = false;
647 break;
648 }
649 }
650
651 if (!Safe)
652 continue;
653
654 MachineInstr *MI = Candidate.MI;
655 for (const MachineOperand &MO : MI->all_uses()) {
656 if (!MO.getReg())
657 continue;
658 for (MCRegUnitIterator RUI(MO.getReg(), TRI); RUI.isValid(); ++RUI) {
659 if (RUDefs.test(*RUI) || RUClobbers.test(*RUI)) {
660 // If it's using a non-loop-invariant register, then it's obviously
661 // not safe to hoist.
662 Safe = false;
663 break;
664 }
665 }
666
667 if (!Safe)
668 break;
669 }
670
671 if (Safe)
672 HoistPostRA(MI, Candidate.Def, CurLoop, CurPreheader);
673 }
674 }
675
676 /// Add register 'Reg' to the livein sets of BBs in the current loop, and make
677 /// sure it is not killed by any instructions in the loop.
AddToLiveIns(MCRegister Reg,MachineLoop * CurLoop)678 void MachineLICMBase::AddToLiveIns(MCRegister Reg, MachineLoop *CurLoop) {
679 for (MachineBasicBlock *BB : CurLoop->getBlocks()) {
680 if (!BB->isLiveIn(Reg))
681 BB->addLiveIn(Reg);
682 for (MachineInstr &MI : *BB) {
683 for (MachineOperand &MO : MI.all_uses()) {
684 if (!MO.getReg())
685 continue;
686 if (TRI->regsOverlap(Reg, MO.getReg()))
687 MO.setIsKill(false);
688 }
689 }
690 }
691 }
692
693 /// When an instruction is found to only use loop invariant operands that is
694 /// safe to hoist, this instruction is called to do the dirty work.
HoistPostRA(MachineInstr * MI,unsigned Def,MachineLoop * CurLoop,MachineBasicBlock * CurPreheader)695 void MachineLICMBase::HoistPostRA(MachineInstr *MI, unsigned Def,
696 MachineLoop *CurLoop,
697 MachineBasicBlock *CurPreheader) {
698 MachineBasicBlock *Preheader = getCurPreheader(CurLoop, CurPreheader);
699
700 // Now move the instructions to the predecessor, inserting it before any
701 // terminator instructions.
702 LLVM_DEBUG(dbgs() << "Hoisting to " << printMBBReference(*Preheader)
703 << " from " << printMBBReference(*MI->getParent()) << ": "
704 << *MI);
705
706 // Splice the instruction to the preheader.
707 MachineBasicBlock *MBB = MI->getParent();
708 Preheader->splice(Preheader->getFirstTerminator(), MBB, MI);
709
710 // Since we are moving the instruction out of its basic block, we do not
711 // retain its debug location. Doing so would degrade the debugging
712 // experience and adversely affect the accuracy of profiling information.
713 assert(!MI->isDebugInstr() && "Should not hoist debug inst");
714 MI->setDebugLoc(DebugLoc());
715
716 // Add register to livein list to all the BBs in the current loop since a
717 // loop invariant must be kept live throughout the whole loop. This is
718 // important to ensure later passes do not scavenge the def register.
719 AddToLiveIns(Def, CurLoop);
720
721 ++NumPostRAHoisted;
722 Changed = true;
723 }
724
725 /// Check if this mbb is guaranteed to execute. If not then a load from this mbb
726 /// may not be safe to hoist.
IsGuaranteedToExecute(MachineBasicBlock * BB,MachineLoop * CurLoop)727 bool MachineLICMBase::IsGuaranteedToExecute(MachineBasicBlock *BB,
728 MachineLoop *CurLoop) {
729 if (SpeculationState != SpeculateUnknown)
730 return SpeculationState == SpeculateFalse;
731
732 if (BB != CurLoop->getHeader()) {
733 // Check loop exiting blocks.
734 SmallVector<MachineBasicBlock*, 8> CurrentLoopExitingBlocks;
735 CurLoop->getExitingBlocks(CurrentLoopExitingBlocks);
736 for (MachineBasicBlock *CurrentLoopExitingBlock : CurrentLoopExitingBlocks)
737 if (!DT->dominates(BB, CurrentLoopExitingBlock)) {
738 SpeculationState = SpeculateTrue;
739 return false;
740 }
741 }
742
743 SpeculationState = SpeculateFalse;
744 return true;
745 }
746
747 /// Check if \p MI is trivially remateralizable and if it does not have any
748 /// virtual register uses. Even though rematerializable RA might not actually
749 /// rematerialize it in this scenario. In that case we do not want to hoist such
750 /// instruction out of the loop in a belief RA will sink it back if needed.
isTriviallyReMaterializable(const MachineInstr & MI) const751 bool MachineLICMBase::isTriviallyReMaterializable(
752 const MachineInstr &MI) const {
753 if (!TII->isTriviallyReMaterializable(MI))
754 return false;
755
756 for (const MachineOperand &MO : MI.all_uses()) {
757 if (MO.getReg().isVirtual())
758 return false;
759 }
760
761 return true;
762 }
763
EnterScope(MachineBasicBlock * MBB)764 void MachineLICMBase::EnterScope(MachineBasicBlock *MBB) {
765 LLVM_DEBUG(dbgs() << "Entering " << printMBBReference(*MBB) << '\n');
766
767 // Remember livein register pressure.
768 BackTrace.push_back(RegPressure);
769 }
770
ExitScope(MachineBasicBlock * MBB)771 void MachineLICMBase::ExitScope(MachineBasicBlock *MBB) {
772 LLVM_DEBUG(dbgs() << "Exiting " << printMBBReference(*MBB) << '\n');
773 BackTrace.pop_back();
774 }
775
776 /// Destroy scope for the MBB that corresponds to the given dominator tree node
777 /// if its a leaf or all of its children are done. Walk up the dominator tree to
778 /// destroy ancestors which are now done.
ExitScopeIfDone(MachineDomTreeNode * Node,DenseMap<MachineDomTreeNode *,unsigned> & OpenChildren,const DenseMap<MachineDomTreeNode *,MachineDomTreeNode * > & ParentMap)779 void MachineLICMBase::ExitScopeIfDone(MachineDomTreeNode *Node,
780 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
781 const DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap) {
782 if (OpenChildren[Node])
783 return;
784
785 for(;;) {
786 ExitScope(Node->getBlock());
787 // Now traverse upwards to pop ancestors whose offsprings are all done.
788 MachineDomTreeNode *Parent = ParentMap.lookup(Node);
789 if (!Parent || --OpenChildren[Parent] != 0)
790 break;
791 Node = Parent;
792 }
793 }
794
795 /// Walk the specified loop in the CFG (defined by all blocks dominated by the
796 /// specified header block, and that are in the current loop) in depth first
797 /// order w.r.t the DominatorTree. This allows us to visit definitions before
798 /// uses, allowing us to hoist a loop body in one pass without iteration.
HoistOutOfLoop(MachineDomTreeNode * HeaderN,MachineLoop * CurLoop,MachineBasicBlock * CurPreheader)799 void MachineLICMBase::HoistOutOfLoop(MachineDomTreeNode *HeaderN,
800 MachineLoop *CurLoop,
801 MachineBasicBlock *CurPreheader) {
802 MachineBasicBlock *Preheader = getCurPreheader(CurLoop, CurPreheader);
803 if (!Preheader)
804 return;
805
806 SmallVector<MachineDomTreeNode*, 32> Scopes;
807 SmallVector<MachineDomTreeNode*, 8> WorkList;
808 DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> ParentMap;
809 DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
810
811 // Perform a DFS walk to determine the order of visit.
812 WorkList.push_back(HeaderN);
813 while (!WorkList.empty()) {
814 MachineDomTreeNode *Node = WorkList.pop_back_val();
815 assert(Node && "Null dominator tree node?");
816 MachineBasicBlock *BB = Node->getBlock();
817
818 // If the header of the loop containing this basic block is a landing pad,
819 // then don't try to hoist instructions out of this loop.
820 const MachineLoop *ML = MLI->getLoopFor(BB);
821 if (ML && ML->getHeader()->isEHPad())
822 continue;
823
824 // If this subregion is not in the top level loop at all, exit.
825 if (!CurLoop->contains(BB))
826 continue;
827
828 Scopes.push_back(Node);
829 unsigned NumChildren = Node->getNumChildren();
830
831 // Don't hoist things out of a large switch statement. This often causes
832 // code to be hoisted that wasn't going to be executed, and increases
833 // register pressure in a situation where it's likely to matter.
834 if (BB->succ_size() >= 25)
835 NumChildren = 0;
836
837 OpenChildren[Node] = NumChildren;
838 if (NumChildren) {
839 // Add children in reverse order as then the next popped worklist node is
840 // the first child of this node. This means we ultimately traverse the
841 // DOM tree in exactly the same order as if we'd recursed.
842 for (MachineDomTreeNode *Child : reverse(Node->children())) {
843 ParentMap[Child] = Node;
844 WorkList.push_back(Child);
845 }
846 }
847 }
848
849 if (Scopes.size() == 0)
850 return;
851
852 // Compute registers which are livein into the loop headers.
853 RegSeen.clear();
854 BackTrace.clear();
855 InitRegPressure(Preheader);
856
857 // Now perform LICM.
858 for (MachineDomTreeNode *Node : Scopes) {
859 MachineBasicBlock *MBB = Node->getBlock();
860
861 EnterScope(MBB);
862
863 // Process the block
864 SpeculationState = SpeculateUnknown;
865 for (MachineInstr &MI : llvm::make_early_inc_range(*MBB)) {
866 unsigned HoistRes = HoistResult::NotHoisted;
867 HoistRes = Hoist(&MI, Preheader, CurLoop);
868 if (HoistRes & HoistResult::NotHoisted) {
869 // We have failed to hoist MI to outermost loop's preheader. If MI is in
870 // a subloop, try to hoist it to subloop's preheader.
871 SmallVector<MachineLoop *> InnerLoopWorkList;
872 for (MachineLoop *L = MLI->getLoopFor(MI.getParent()); L != CurLoop;
873 L = L->getParentLoop())
874 InnerLoopWorkList.push_back(L);
875
876 while (!InnerLoopWorkList.empty()) {
877 MachineLoop *InnerLoop = InnerLoopWorkList.pop_back_val();
878 MachineBasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader();
879 if (InnerLoopPreheader) {
880 HoistRes = Hoist(&MI, InnerLoopPreheader, InnerLoop);
881 if (HoistRes & HoistResult::Hoisted)
882 break;
883 }
884 }
885 }
886
887 if (HoistRes & HoistResult::ErasedMI)
888 continue;
889
890 UpdateRegPressure(&MI);
891 }
892
893 // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
894 ExitScopeIfDone(Node, OpenChildren, ParentMap);
895 }
896 }
897
isOperandKill(const MachineOperand & MO,MachineRegisterInfo * MRI)898 static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) {
899 return MO.isKill() || MRI->hasOneNonDBGUse(MO.getReg());
900 }
901
902 /// Find all virtual register references that are liveout of the preheader to
903 /// initialize the starting "register pressure". Note this does not count live
904 /// through (livein but not used) registers.
InitRegPressure(MachineBasicBlock * BB)905 void MachineLICMBase::InitRegPressure(MachineBasicBlock *BB) {
906 std::fill(RegPressure.begin(), RegPressure.end(), 0);
907
908 // If the preheader has only a single predecessor and it ends with a
909 // fallthrough or an unconditional branch, then scan its predecessor for live
910 // defs as well. This happens whenever the preheader is created by splitting
911 // the critical edge from the loop predecessor to the loop header.
912 if (BB->pred_size() == 1) {
913 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
914 SmallVector<MachineOperand, 4> Cond;
915 if (!TII->analyzeBranch(*BB, TBB, FBB, Cond, false) && Cond.empty())
916 InitRegPressure(*BB->pred_begin());
917 }
918
919 for (const MachineInstr &MI : *BB)
920 UpdateRegPressure(&MI, /*ConsiderUnseenAsDef=*/true);
921 }
922
923 /// Update estimate of register pressure after the specified instruction.
UpdateRegPressure(const MachineInstr * MI,bool ConsiderUnseenAsDef)924 void MachineLICMBase::UpdateRegPressure(const MachineInstr *MI,
925 bool ConsiderUnseenAsDef) {
926 auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/true, ConsiderUnseenAsDef);
927 for (const auto &RPIdAndCost : Cost) {
928 unsigned Class = RPIdAndCost.first;
929 if (static_cast<int>(RegPressure[Class]) < -RPIdAndCost.second)
930 RegPressure[Class] = 0;
931 else
932 RegPressure[Class] += RPIdAndCost.second;
933 }
934 }
935
936 /// Calculate the additional register pressure that the registers used in MI
937 /// cause.
938 ///
939 /// If 'ConsiderSeen' is true, updates 'RegSeen' and uses the information to
940 /// figure out which usages are live-ins.
941 /// FIXME: Figure out a way to consider 'RegSeen' from all code paths.
942 DenseMap<unsigned, int>
calcRegisterCost(const MachineInstr * MI,bool ConsiderSeen,bool ConsiderUnseenAsDef)943 MachineLICMBase::calcRegisterCost(const MachineInstr *MI, bool ConsiderSeen,
944 bool ConsiderUnseenAsDef) {
945 DenseMap<unsigned, int> Cost;
946 if (MI->isImplicitDef())
947 return Cost;
948 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
949 const MachineOperand &MO = MI->getOperand(i);
950 if (!MO.isReg() || MO.isImplicit())
951 continue;
952 Register Reg = MO.getReg();
953 if (!Reg.isVirtual())
954 continue;
955
956 // FIXME: It seems bad to use RegSeen only for some of these calculations.
957 bool isNew = ConsiderSeen ? RegSeen.insert(Reg).second : false;
958 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
959
960 RegClassWeight W = TRI->getRegClassWeight(RC);
961 int RCCost = 0;
962 if (MO.isDef())
963 RCCost = W.RegWeight;
964 else {
965 bool isKill = isOperandKill(MO, MRI);
966 if (isNew && !isKill && ConsiderUnseenAsDef)
967 // Haven't seen this, it must be a livein.
968 RCCost = W.RegWeight;
969 else if (!isNew && isKill)
970 RCCost = -W.RegWeight;
971 }
972 if (RCCost == 0)
973 continue;
974 const int *PS = TRI->getRegClassPressureSets(RC);
975 for (; *PS != -1; ++PS) {
976 if (!Cost.contains(*PS))
977 Cost[*PS] = RCCost;
978 else
979 Cost[*PS] += RCCost;
980 }
981 }
982 return Cost;
983 }
984
985 /// Return true if this machine instruction loads from global offset table or
986 /// constant pool.
mayLoadFromGOTOrConstantPool(MachineInstr & MI)987 static bool mayLoadFromGOTOrConstantPool(MachineInstr &MI) {
988 assert(MI.mayLoad() && "Expected MI that loads!");
989
990 // If we lost memory operands, conservatively assume that the instruction
991 // reads from everything..
992 if (MI.memoperands_empty())
993 return true;
994
995 for (MachineMemOperand *MemOp : MI.memoperands())
996 if (const PseudoSourceValue *PSV = MemOp->getPseudoValue())
997 if (PSV->isGOT() || PSV->isConstantPool())
998 return true;
999
1000 return false;
1001 }
1002
1003 // This function iterates through all the operands of the input store MI and
1004 // checks that each register operand statisfies isCallerPreservedPhysReg.
1005 // This means, the value being stored and the address where it is being stored
1006 // is constant throughout the body of the function (not including prologue and
1007 // epilogue). When called with an MI that isn't a store, it returns false.
1008 // A future improvement can be to check if the store registers are constant
1009 // throughout the loop rather than throughout the funtion.
isInvariantStore(const MachineInstr & MI,const TargetRegisterInfo * TRI,const MachineRegisterInfo * MRI)1010 static bool isInvariantStore(const MachineInstr &MI,
1011 const TargetRegisterInfo *TRI,
1012 const MachineRegisterInfo *MRI) {
1013
1014 bool FoundCallerPresReg = false;
1015 if (!MI.mayStore() || MI.hasUnmodeledSideEffects() ||
1016 (MI.getNumOperands() == 0))
1017 return false;
1018
1019 // Check that all register operands are caller-preserved physical registers.
1020 for (const MachineOperand &MO : MI.operands()) {
1021 if (MO.isReg()) {
1022 Register Reg = MO.getReg();
1023 // If operand is a virtual register, check if it comes from a copy of a
1024 // physical register.
1025 if (Reg.isVirtual())
1026 Reg = TRI->lookThruCopyLike(MO.getReg(), MRI);
1027 if (Reg.isVirtual())
1028 return false;
1029 if (!TRI->isCallerPreservedPhysReg(Reg.asMCReg(), *MI.getMF()))
1030 return false;
1031 else
1032 FoundCallerPresReg = true;
1033 } else if (!MO.isImm()) {
1034 return false;
1035 }
1036 }
1037 return FoundCallerPresReg;
1038 }
1039
1040 // Return true if the input MI is a copy instruction that feeds an invariant
1041 // store instruction. This means that the src of the copy has to satisfy
1042 // isCallerPreservedPhysReg and atleast one of it's users should satisfy
1043 // isInvariantStore.
isCopyFeedingInvariantStore(const MachineInstr & MI,const MachineRegisterInfo * MRI,const TargetRegisterInfo * TRI)1044 static bool isCopyFeedingInvariantStore(const MachineInstr &MI,
1045 const MachineRegisterInfo *MRI,
1046 const TargetRegisterInfo *TRI) {
1047
1048 // FIXME: If targets would like to look through instructions that aren't
1049 // pure copies, this can be updated to a query.
1050 if (!MI.isCopy())
1051 return false;
1052
1053 const MachineFunction *MF = MI.getMF();
1054 // Check that we are copying a constant physical register.
1055 Register CopySrcReg = MI.getOperand(1).getReg();
1056 if (CopySrcReg.isVirtual())
1057 return false;
1058
1059 if (!TRI->isCallerPreservedPhysReg(CopySrcReg.asMCReg(), *MF))
1060 return false;
1061
1062 Register CopyDstReg = MI.getOperand(0).getReg();
1063 // Check if any of the uses of the copy are invariant stores.
1064 assert(CopyDstReg.isVirtual() && "copy dst is not a virtual reg");
1065
1066 for (MachineInstr &UseMI : MRI->use_instructions(CopyDstReg)) {
1067 if (UseMI.mayStore() && isInvariantStore(UseMI, TRI, MRI))
1068 return true;
1069 }
1070 return false;
1071 }
1072
1073 /// Returns true if the instruction may be a suitable candidate for LICM.
1074 /// e.g. If the instruction is a call, then it's obviously not safe to hoist it.
IsLICMCandidate(MachineInstr & I,MachineLoop * CurLoop)1075 bool MachineLICMBase::IsLICMCandidate(MachineInstr &I, MachineLoop *CurLoop) {
1076 // Check if it's safe to move the instruction.
1077 bool DontMoveAcrossStore = !HoistConstLoads || !AllowedToHoistLoads[CurLoop];
1078 if ((!I.isSafeToMove(AA, DontMoveAcrossStore)) &&
1079 !(HoistConstStores && isInvariantStore(I, TRI, MRI))) {
1080 LLVM_DEBUG(dbgs() << "LICM: Instruction not safe to move.\n");
1081 return false;
1082 }
1083
1084 // If it is a load then check if it is guaranteed to execute by making sure
1085 // that it dominates all exiting blocks. If it doesn't, then there is a path
1086 // out of the loop which does not execute this load, so we can't hoist it.
1087 // Loads from constant memory are safe to speculate, for example indexed load
1088 // from a jump table.
1089 // Stores and side effects are already checked by isSafeToMove.
1090 if (I.mayLoad() && !mayLoadFromGOTOrConstantPool(I) &&
1091 !IsGuaranteedToExecute(I.getParent(), CurLoop)) {
1092 LLVM_DEBUG(dbgs() << "LICM: Load not guaranteed to execute.\n");
1093 return false;
1094 }
1095
1096 // Convergent attribute has been used on operations that involve inter-thread
1097 // communication which results are implicitly affected by the enclosing
1098 // control flows. It is not safe to hoist or sink such operations across
1099 // control flow.
1100 if (I.isConvergent())
1101 return false;
1102
1103 if (!TII->shouldHoist(I, CurLoop))
1104 return false;
1105
1106 return true;
1107 }
1108
1109 /// Returns true if the instruction is loop invariant.
IsLoopInvariantInst(MachineInstr & I,MachineLoop * CurLoop)1110 bool MachineLICMBase::IsLoopInvariantInst(MachineInstr &I,
1111 MachineLoop *CurLoop) {
1112 if (!IsLICMCandidate(I, CurLoop)) {
1113 LLVM_DEBUG(dbgs() << "LICM: Instruction not a LICM candidate\n");
1114 return false;
1115 }
1116 return CurLoop->isLoopInvariant(I);
1117 }
1118
1119 /// Return true if the specified instruction is used by a phi node and hoisting
1120 /// it could cause a copy to be inserted.
HasLoopPHIUse(const MachineInstr * MI,MachineLoop * CurLoop)1121 bool MachineLICMBase::HasLoopPHIUse(const MachineInstr *MI,
1122 MachineLoop *CurLoop) {
1123 SmallVector<const MachineInstr *, 8> Work(1, MI);
1124 do {
1125 MI = Work.pop_back_val();
1126 for (const MachineOperand &MO : MI->all_defs()) {
1127 Register Reg = MO.getReg();
1128 if (!Reg.isVirtual())
1129 continue;
1130 for (MachineInstr &UseMI : MRI->use_instructions(Reg)) {
1131 // A PHI may cause a copy to be inserted.
1132 if (UseMI.isPHI()) {
1133 // A PHI inside the loop causes a copy because the live range of Reg is
1134 // extended across the PHI.
1135 if (CurLoop->contains(&UseMI))
1136 return true;
1137 // A PHI in an exit block can cause a copy to be inserted if the PHI
1138 // has multiple predecessors in the loop with different values.
1139 // For now, approximate by rejecting all exit blocks.
1140 if (isExitBlock(CurLoop, UseMI.getParent()))
1141 return true;
1142 continue;
1143 }
1144 // Look past copies as well.
1145 if (UseMI.isCopy() && CurLoop->contains(&UseMI))
1146 Work.push_back(&UseMI);
1147 }
1148 }
1149 } while (!Work.empty());
1150 return false;
1151 }
1152
1153 /// Compute operand latency between a def of 'Reg' and an use in the current
1154 /// loop, return true if the target considered it high.
HasHighOperandLatency(MachineInstr & MI,unsigned DefIdx,Register Reg,MachineLoop * CurLoop) const1155 bool MachineLICMBase::HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx,
1156 Register Reg,
1157 MachineLoop *CurLoop) const {
1158 if (MRI->use_nodbg_empty(Reg))
1159 return false;
1160
1161 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(Reg)) {
1162 if (UseMI.isCopyLike())
1163 continue;
1164 if (!CurLoop->contains(UseMI.getParent()))
1165 continue;
1166 for (unsigned i = 0, e = UseMI.getNumOperands(); i != e; ++i) {
1167 const MachineOperand &MO = UseMI.getOperand(i);
1168 if (!MO.isReg() || !MO.isUse())
1169 continue;
1170 Register MOReg = MO.getReg();
1171 if (MOReg != Reg)
1172 continue;
1173
1174 if (TII->hasHighOperandLatency(SchedModel, MRI, MI, DefIdx, UseMI, i))
1175 return true;
1176 }
1177
1178 // Only look at the first in loop use.
1179 break;
1180 }
1181
1182 return false;
1183 }
1184
1185 /// Return true if the instruction is marked "cheap" or the operand latency
1186 /// between its def and a use is one or less.
IsCheapInstruction(MachineInstr & MI) const1187 bool MachineLICMBase::IsCheapInstruction(MachineInstr &MI) const {
1188 if (TII->isAsCheapAsAMove(MI) || MI.isCopyLike())
1189 return true;
1190
1191 bool isCheap = false;
1192 unsigned NumDefs = MI.getDesc().getNumDefs();
1193 for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) {
1194 MachineOperand &DefMO = MI.getOperand(i);
1195 if (!DefMO.isReg() || !DefMO.isDef())
1196 continue;
1197 --NumDefs;
1198 Register Reg = DefMO.getReg();
1199 if (Reg.isPhysical())
1200 continue;
1201
1202 if (!TII->hasLowDefLatency(SchedModel, MI, i))
1203 return false;
1204 isCheap = true;
1205 }
1206
1207 return isCheap;
1208 }
1209
1210 /// Visit BBs from header to current BB, check if hoisting an instruction of the
1211 /// given cost matrix can cause high register pressure.
1212 bool
CanCauseHighRegPressure(const DenseMap<unsigned,int> & Cost,bool CheapInstr)1213 MachineLICMBase::CanCauseHighRegPressure(const DenseMap<unsigned, int>& Cost,
1214 bool CheapInstr) {
1215 for (const auto &RPIdAndCost : Cost) {
1216 if (RPIdAndCost.second <= 0)
1217 continue;
1218
1219 unsigned Class = RPIdAndCost.first;
1220 int Limit = RegLimit[Class];
1221
1222 // Don't hoist cheap instructions if they would increase register pressure,
1223 // even if we're under the limit.
1224 if (CheapInstr && !HoistCheapInsts)
1225 return true;
1226
1227 for (const auto &RP : BackTrace)
1228 if (static_cast<int>(RP[Class]) + RPIdAndCost.second >= Limit)
1229 return true;
1230 }
1231
1232 return false;
1233 }
1234
1235 /// Traverse the back trace from header to the current block and update their
1236 /// register pressures to reflect the effect of hoisting MI from the current
1237 /// block to the preheader.
UpdateBackTraceRegPressure(const MachineInstr * MI)1238 void MachineLICMBase::UpdateBackTraceRegPressure(const MachineInstr *MI) {
1239 // First compute the 'cost' of the instruction, i.e. its contribution
1240 // to register pressure.
1241 auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/false,
1242 /*ConsiderUnseenAsDef=*/false);
1243
1244 // Update register pressure of blocks from loop header to current block.
1245 for (auto &RP : BackTrace)
1246 for (const auto &RPIdAndCost : Cost)
1247 RP[RPIdAndCost.first] += RPIdAndCost.second;
1248 }
1249
1250 /// Return true if it is potentially profitable to hoist the given loop
1251 /// invariant.
IsProfitableToHoist(MachineInstr & MI,MachineLoop * CurLoop)1252 bool MachineLICMBase::IsProfitableToHoist(MachineInstr &MI,
1253 MachineLoop *CurLoop) {
1254 if (MI.isImplicitDef())
1255 return true;
1256
1257 // Besides removing computation from the loop, hoisting an instruction has
1258 // these effects:
1259 //
1260 // - The value defined by the instruction becomes live across the entire
1261 // loop. This increases register pressure in the loop.
1262 //
1263 // - If the value is used by a PHI in the loop, a copy will be required for
1264 // lowering the PHI after extending the live range.
1265 //
1266 // - When hoisting the last use of a value in the loop, that value no longer
1267 // needs to be live in the loop. This lowers register pressure in the loop.
1268
1269 if (HoistConstStores && isCopyFeedingInvariantStore(MI, MRI, TRI))
1270 return true;
1271
1272 bool CheapInstr = IsCheapInstruction(MI);
1273 bool CreatesCopy = HasLoopPHIUse(&MI, CurLoop);
1274
1275 // Don't hoist a cheap instruction if it would create a copy in the loop.
1276 if (CheapInstr && CreatesCopy) {
1277 LLVM_DEBUG(dbgs() << "Won't hoist cheap instr with loop PHI use: " << MI);
1278 return false;
1279 }
1280
1281 // Rematerializable instructions should always be hoisted providing the
1282 // register allocator can just pull them down again when needed.
1283 if (isTriviallyReMaterializable(MI))
1284 return true;
1285
1286 // FIXME: If there are long latency loop-invariant instructions inside the
1287 // loop at this point, why didn't the optimizer's LICM hoist them?
1288 for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) {
1289 const MachineOperand &MO = MI.getOperand(i);
1290 if (!MO.isReg() || MO.isImplicit())
1291 continue;
1292 Register Reg = MO.getReg();
1293 if (!Reg.isVirtual())
1294 continue;
1295 if (MO.isDef() && HasHighOperandLatency(MI, i, Reg, CurLoop)) {
1296 LLVM_DEBUG(dbgs() << "Hoist High Latency: " << MI);
1297 ++NumHighLatency;
1298 return true;
1299 }
1300 }
1301
1302 // Estimate register pressure to determine whether to LICM the instruction.
1303 // In low register pressure situation, we can be more aggressive about
1304 // hoisting. Also, favors hoisting long latency instructions even in
1305 // moderately high pressure situation.
1306 // Cheap instructions will only be hoisted if they don't increase register
1307 // pressure at all.
1308 auto Cost = calcRegisterCost(&MI, /*ConsiderSeen=*/false,
1309 /*ConsiderUnseenAsDef=*/false);
1310
1311 // Visit BBs from header to current BB, if hoisting this doesn't cause
1312 // high register pressure, then it's safe to proceed.
1313 if (!CanCauseHighRegPressure(Cost, CheapInstr)) {
1314 LLVM_DEBUG(dbgs() << "Hoist non-reg-pressure: " << MI);
1315 ++NumLowRP;
1316 return true;
1317 }
1318
1319 // Don't risk increasing register pressure if it would create copies.
1320 if (CreatesCopy) {
1321 LLVM_DEBUG(dbgs() << "Won't hoist instr with loop PHI use: " << MI);
1322 return false;
1323 }
1324
1325 // Do not "speculate" in high register pressure situation. If an
1326 // instruction is not guaranteed to be executed in the loop, it's best to be
1327 // conservative.
1328 if (AvoidSpeculation &&
1329 (!IsGuaranteedToExecute(MI.getParent(), CurLoop) && !MayCSE(&MI))) {
1330 LLVM_DEBUG(dbgs() << "Won't speculate: " << MI);
1331 return false;
1332 }
1333
1334 // If we have a COPY with other uses in the loop, hoist to allow the users to
1335 // also be hoisted.
1336 // TODO: Handle all isCopyLike?
1337 if (MI.isCopy() || MI.isRegSequence()) {
1338 Register DefReg = MI.getOperand(0).getReg();
1339 if (DefReg.isVirtual() &&
1340 all_of(MI.uses(),
1341 [this](const MachineOperand &UseOp) {
1342 return !UseOp.isReg() || UseOp.getReg().isVirtual() ||
1343 MRI->isConstantPhysReg(UseOp.getReg());
1344 }) &&
1345 IsLoopInvariantInst(MI, CurLoop) &&
1346 any_of(MRI->use_nodbg_instructions(DefReg),
1347 [&CurLoop, this, DefReg, Cost](MachineInstr &UseMI) {
1348 if (!CurLoop->contains(&UseMI))
1349 return false;
1350
1351 // COPY is a cheap instruction, but if moving it won't cause
1352 // high RP we're fine to hoist it even if the user can't be
1353 // hoisted later Otherwise we want to check the user if it's
1354 // hoistable
1355 if (CanCauseHighRegPressure(Cost, false) &&
1356 !CurLoop->isLoopInvariant(UseMI, DefReg))
1357 return false;
1358
1359 return true;
1360 }))
1361 return true;
1362 }
1363
1364 // High register pressure situation, only hoist if the instruction is going
1365 // to be remat'ed.
1366 if (!isTriviallyReMaterializable(MI) &&
1367 !MI.isDereferenceableInvariantLoad()) {
1368 LLVM_DEBUG(dbgs() << "Can't remat / high reg-pressure: " << MI);
1369 return false;
1370 }
1371
1372 return true;
1373 }
1374
1375 /// Unfold a load from the given machineinstr if the load itself could be
1376 /// hoisted. Return the unfolded and hoistable load, or null if the load
1377 /// couldn't be unfolded or if it wouldn't be hoistable.
ExtractHoistableLoad(MachineInstr * MI,MachineLoop * CurLoop)1378 MachineInstr *MachineLICMBase::ExtractHoistableLoad(MachineInstr *MI,
1379 MachineLoop *CurLoop) {
1380 // Don't unfold simple loads.
1381 if (MI->canFoldAsLoad())
1382 return nullptr;
1383
1384 // If not, we may be able to unfold a load and hoist that.
1385 // First test whether the instruction is loading from an amenable
1386 // memory location.
1387 if (!MI->isDereferenceableInvariantLoad())
1388 return nullptr;
1389
1390 // Next determine the register class for a temporary register.
1391 unsigned LoadRegIndex;
1392 unsigned NewOpc =
1393 TII->getOpcodeAfterMemoryUnfold(MI->getOpcode(),
1394 /*UnfoldLoad=*/true,
1395 /*UnfoldStore=*/false,
1396 &LoadRegIndex);
1397 if (NewOpc == 0) return nullptr;
1398 const MCInstrDesc &MID = TII->get(NewOpc);
1399 MachineFunction &MF = *MI->getMF();
1400 const TargetRegisterClass *RC = TII->getRegClass(MID, LoadRegIndex, TRI, MF);
1401 // Ok, we're unfolding. Create a temporary register and do the unfold.
1402 Register Reg = MRI->createVirtualRegister(RC);
1403
1404 SmallVector<MachineInstr *, 2> NewMIs;
1405 bool Success = TII->unfoldMemoryOperand(MF, *MI, Reg,
1406 /*UnfoldLoad=*/true,
1407 /*UnfoldStore=*/false, NewMIs);
1408 (void)Success;
1409 assert(Success &&
1410 "unfoldMemoryOperand failed when getOpcodeAfterMemoryUnfold "
1411 "succeeded!");
1412 assert(NewMIs.size() == 2 &&
1413 "Unfolded a load into multiple instructions!");
1414 MachineBasicBlock *MBB = MI->getParent();
1415 MachineBasicBlock::iterator Pos = MI;
1416 MBB->insert(Pos, NewMIs[0]);
1417 MBB->insert(Pos, NewMIs[1]);
1418 // If unfolding produced a load that wasn't loop-invariant or profitable to
1419 // hoist, discard the new instructions and bail.
1420 if (!IsLoopInvariantInst(*NewMIs[0], CurLoop) ||
1421 !IsProfitableToHoist(*NewMIs[0], CurLoop)) {
1422 NewMIs[0]->eraseFromParent();
1423 NewMIs[1]->eraseFromParent();
1424 return nullptr;
1425 }
1426
1427 // Update register pressure for the unfolded instruction.
1428 UpdateRegPressure(NewMIs[1]);
1429
1430 // Otherwise we successfully unfolded a load that we can hoist.
1431
1432 // Update the call site info.
1433 if (MI->shouldUpdateCallSiteInfo())
1434 MF.eraseCallSiteInfo(MI);
1435
1436 MI->eraseFromParent();
1437 return NewMIs[0];
1438 }
1439
1440 /// Initialize the CSE map with instructions that are in the current loop
1441 /// preheader that may become duplicates of instructions that are hoisted
1442 /// out of the loop.
InitCSEMap(MachineBasicBlock * BB)1443 void MachineLICMBase::InitCSEMap(MachineBasicBlock *BB) {
1444 for (MachineInstr &MI : *BB)
1445 CSEMap[BB][MI.getOpcode()].push_back(&MI);
1446 }
1447
1448 /// Initialize AllowedToHoistLoads with information about whether invariant
1449 /// loads can be moved outside a given loop
InitializeLoadsHoistableLoops()1450 void MachineLICMBase::InitializeLoadsHoistableLoops() {
1451 SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end());
1452 SmallVector<MachineLoop *, 8> LoopsInPreOrder;
1453
1454 // Mark all loops as hoistable initially and prepare a list of loops in
1455 // pre-order DFS.
1456 while (!Worklist.empty()) {
1457 auto *L = Worklist.pop_back_val();
1458 AllowedToHoistLoads[L] = true;
1459 LoopsInPreOrder.push_back(L);
1460 Worklist.insert(Worklist.end(), L->getSubLoops().begin(),
1461 L->getSubLoops().end());
1462 }
1463
1464 // Going from the innermost to outermost loops, check if a loop has
1465 // instructions preventing invariant load hoisting. If such instruction is
1466 // found, mark this loop and its parent as non-hoistable and continue
1467 // investigating the next loop.
1468 // Visiting in a reversed pre-ordered DFS manner
1469 // allows us to not process all the instructions of the outer loop if the
1470 // inner loop is proved to be non-load-hoistable.
1471 for (auto *Loop : reverse(LoopsInPreOrder)) {
1472 for (auto *MBB : Loop->blocks()) {
1473 // If this loop has already been marked as non-hoistable, skip it.
1474 if (!AllowedToHoistLoads[Loop])
1475 continue;
1476 for (auto &MI : *MBB) {
1477 if (!MI.isLoadFoldBarrier() && !MI.mayStore() && !MI.isCall() &&
1478 !(MI.mayLoad() && MI.hasOrderedMemoryRef()))
1479 continue;
1480 for (MachineLoop *L = Loop; L != nullptr; L = L->getParentLoop())
1481 AllowedToHoistLoads[L] = false;
1482 break;
1483 }
1484 }
1485 }
1486 }
1487
1488 /// Find an instruction amount PrevMIs that is a duplicate of MI.
1489 /// Return this instruction if it's found.
1490 MachineInstr *
LookForDuplicate(const MachineInstr * MI,std::vector<MachineInstr * > & PrevMIs)1491 MachineLICMBase::LookForDuplicate(const MachineInstr *MI,
1492 std::vector<MachineInstr *> &PrevMIs) {
1493 for (MachineInstr *PrevMI : PrevMIs)
1494 if (TII->produceSameValue(*MI, *PrevMI, (PreRegAlloc ? MRI : nullptr)))
1495 return PrevMI;
1496
1497 return nullptr;
1498 }
1499
1500 /// Given a LICM'ed instruction, look for an instruction on the preheader that
1501 /// computes the same value. If it's found, do a RAU on with the definition of
1502 /// the existing instruction rather than hoisting the instruction to the
1503 /// preheader.
EliminateCSE(MachineInstr * MI,DenseMap<unsigned,std::vector<MachineInstr * >>::iterator & CI)1504 bool MachineLICMBase::EliminateCSE(
1505 MachineInstr *MI,
1506 DenseMap<unsigned, std::vector<MachineInstr *>>::iterator &CI) {
1507 // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1508 // the undef property onto uses.
1509 if (MI->isImplicitDef())
1510 return false;
1511
1512 // Do not CSE normal loads because between them could be store instructions
1513 // that change the loaded value
1514 if (MI->mayLoad() && !MI->isDereferenceableInvariantLoad())
1515 return false;
1516
1517 if (MachineInstr *Dup = LookForDuplicate(MI, CI->second)) {
1518 LLVM_DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup);
1519
1520 // Replace virtual registers defined by MI by their counterparts defined
1521 // by Dup.
1522 SmallVector<unsigned, 2> Defs;
1523 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1524 const MachineOperand &MO = MI->getOperand(i);
1525
1526 // Physical registers may not differ here.
1527 assert((!MO.isReg() || MO.getReg() == 0 || !MO.getReg().isPhysical() ||
1528 MO.getReg() == Dup->getOperand(i).getReg()) &&
1529 "Instructions with different phys regs are not identical!");
1530
1531 if (MO.isReg() && MO.isDef() && !MO.getReg().isPhysical())
1532 Defs.push_back(i);
1533 }
1534
1535 SmallVector<const TargetRegisterClass*, 2> OrigRCs;
1536 for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
1537 unsigned Idx = Defs[i];
1538 Register Reg = MI->getOperand(Idx).getReg();
1539 Register DupReg = Dup->getOperand(Idx).getReg();
1540 OrigRCs.push_back(MRI->getRegClass(DupReg));
1541
1542 if (!MRI->constrainRegClass(DupReg, MRI->getRegClass(Reg))) {
1543 // Restore old RCs if more than one defs.
1544 for (unsigned j = 0; j != i; ++j)
1545 MRI->setRegClass(Dup->getOperand(Defs[j]).getReg(), OrigRCs[j]);
1546 return false;
1547 }
1548 }
1549
1550 for (unsigned Idx : Defs) {
1551 Register Reg = MI->getOperand(Idx).getReg();
1552 Register DupReg = Dup->getOperand(Idx).getReg();
1553 MRI->replaceRegWith(Reg, DupReg);
1554 MRI->clearKillFlags(DupReg);
1555 // Clear Dup dead flag if any, we reuse it for Reg.
1556 if (!MRI->use_nodbg_empty(DupReg))
1557 Dup->getOperand(Idx).setIsDead(false);
1558 }
1559
1560 MI->eraseFromParent();
1561 ++NumCSEed;
1562 return true;
1563 }
1564 return false;
1565 }
1566
1567 /// Return true if the given instruction will be CSE'd if it's hoisted out of
1568 /// the loop.
MayCSE(MachineInstr * MI)1569 bool MachineLICMBase::MayCSE(MachineInstr *MI) {
1570 if (MI->mayLoad() && !MI->isDereferenceableInvariantLoad())
1571 return false;
1572
1573 unsigned Opcode = MI->getOpcode();
1574 for (auto &Map : CSEMap) {
1575 // Check this CSEMap's preheader dominates MI's basic block.
1576 if (DT->dominates(Map.first, MI->getParent())) {
1577 DenseMap<unsigned, std::vector<MachineInstr *>>::iterator CI =
1578 Map.second.find(Opcode);
1579 // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1580 // the undef property onto uses.
1581 if (CI == Map.second.end() || MI->isImplicitDef())
1582 continue;
1583 if (LookForDuplicate(MI, CI->second) != nullptr)
1584 return true;
1585 }
1586 }
1587
1588 return false;
1589 }
1590
1591 /// When an instruction is found to use only loop invariant operands
1592 /// that are safe to hoist, this instruction is called to do the dirty work.
1593 /// It returns true if the instruction is hoisted.
Hoist(MachineInstr * MI,MachineBasicBlock * Preheader,MachineLoop * CurLoop)1594 unsigned MachineLICMBase::Hoist(MachineInstr *MI, MachineBasicBlock *Preheader,
1595 MachineLoop *CurLoop) {
1596 MachineBasicBlock *SrcBlock = MI->getParent();
1597
1598 // Disable the instruction hoisting due to block hotness
1599 if ((DisableHoistingToHotterBlocks == UseBFI::All ||
1600 (DisableHoistingToHotterBlocks == UseBFI::PGO && HasProfileData)) &&
1601 isTgtHotterThanSrc(SrcBlock, Preheader)) {
1602 ++NumNotHoistedDueToHotness;
1603 return HoistResult::NotHoisted;
1604 }
1605 // First check whether we should hoist this instruction.
1606 bool HasExtractHoistableLoad = false;
1607 if (!IsLoopInvariantInst(*MI, CurLoop) ||
1608 !IsProfitableToHoist(*MI, CurLoop)) {
1609 // If not, try unfolding a hoistable load.
1610 MI = ExtractHoistableLoad(MI, CurLoop);
1611 if (!MI)
1612 return HoistResult::NotHoisted;
1613 HasExtractHoistableLoad = true;
1614 }
1615
1616 // If we have hoisted an instruction that may store, it can only be a constant
1617 // store.
1618 if (MI->mayStore())
1619 NumStoreConst++;
1620
1621 // Now move the instructions to the predecessor, inserting it before any
1622 // terminator instructions.
1623 LLVM_DEBUG({
1624 dbgs() << "Hoisting " << *MI;
1625 if (MI->getParent()->getBasicBlock())
1626 dbgs() << " from " << printMBBReference(*MI->getParent());
1627 if (Preheader->getBasicBlock())
1628 dbgs() << " to " << printMBBReference(*Preheader);
1629 dbgs() << "\n";
1630 });
1631
1632 // If this is the first instruction being hoisted to the preheader,
1633 // initialize the CSE map with potential common expressions.
1634 if (FirstInLoop) {
1635 InitCSEMap(Preheader);
1636 FirstInLoop = false;
1637 }
1638
1639 // Look for opportunity to CSE the hoisted instruction.
1640 unsigned Opcode = MI->getOpcode();
1641 bool HasCSEDone = false;
1642 for (auto &Map : CSEMap) {
1643 // Check this CSEMap's preheader dominates MI's basic block.
1644 if (DT->dominates(Map.first, MI->getParent())) {
1645 DenseMap<unsigned, std::vector<MachineInstr *>>::iterator CI =
1646 Map.second.find(Opcode);
1647 if (CI != Map.second.end()) {
1648 if (EliminateCSE(MI, CI)) {
1649 HasCSEDone = true;
1650 break;
1651 }
1652 }
1653 }
1654 }
1655
1656 if (!HasCSEDone) {
1657 // Otherwise, splice the instruction to the preheader.
1658 Preheader->splice(Preheader->getFirstTerminator(),MI->getParent(),MI);
1659
1660 // Since we are moving the instruction out of its basic block, we do not
1661 // retain its debug location. Doing so would degrade the debugging
1662 // experience and adversely affect the accuracy of profiling information.
1663 assert(!MI->isDebugInstr() && "Should not hoist debug inst");
1664 MI->setDebugLoc(DebugLoc());
1665
1666 // Update register pressure for BBs from header to this block.
1667 UpdateBackTraceRegPressure(MI);
1668
1669 // Clear the kill flags of any register this instruction defines,
1670 // since they may need to be live throughout the entire loop
1671 // rather than just live for part of it.
1672 for (MachineOperand &MO : MI->all_defs())
1673 if (!MO.isDead())
1674 MRI->clearKillFlags(MO.getReg());
1675
1676 CSEMap[Preheader][Opcode].push_back(MI);
1677 }
1678
1679 ++NumHoisted;
1680 Changed = true;
1681
1682 if (HasCSEDone || HasExtractHoistableLoad)
1683 return HoistResult::Hoisted | HoistResult::ErasedMI;
1684 return HoistResult::Hoisted;
1685 }
1686
1687 /// Get the preheader for the current loop, splitting a critical edge if needed.
1688 MachineBasicBlock *
getCurPreheader(MachineLoop * CurLoop,MachineBasicBlock * CurPreheader)1689 MachineLICMBase::getCurPreheader(MachineLoop *CurLoop,
1690 MachineBasicBlock *CurPreheader) {
1691 // Determine the block to which to hoist instructions. If we can't find a
1692 // suitable loop predecessor, we can't do any hoisting.
1693
1694 // If we've tried to get a preheader and failed, don't try again.
1695 if (CurPreheader == reinterpret_cast<MachineBasicBlock *>(-1))
1696 return nullptr;
1697
1698 if (!CurPreheader) {
1699 CurPreheader = CurLoop->getLoopPreheader();
1700 if (!CurPreheader) {
1701 MachineBasicBlock *Pred = CurLoop->getLoopPredecessor();
1702 if (!Pred) {
1703 CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1704 return nullptr;
1705 }
1706
1707 CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), *this);
1708 if (!CurPreheader) {
1709 CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1710 return nullptr;
1711 }
1712 }
1713 }
1714 return CurPreheader;
1715 }
1716
1717 /// Is the target basic block at least "BlockFrequencyRatioThreshold"
1718 /// times hotter than the source basic block.
isTgtHotterThanSrc(MachineBasicBlock * SrcBlock,MachineBasicBlock * TgtBlock)1719 bool MachineLICMBase::isTgtHotterThanSrc(MachineBasicBlock *SrcBlock,
1720 MachineBasicBlock *TgtBlock) {
1721 // Parse source and target basic block frequency from MBFI
1722 uint64_t SrcBF = MBFI->getBlockFreq(SrcBlock).getFrequency();
1723 uint64_t DstBF = MBFI->getBlockFreq(TgtBlock).getFrequency();
1724
1725 // Disable the hoisting if source block frequency is zero
1726 if (!SrcBF)
1727 return true;
1728
1729 double Ratio = (double)DstBF / SrcBF;
1730
1731 // Compare the block frequency ratio with the threshold
1732 return Ratio > BlockFrequencyRatioThreshold;
1733 }
1734