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/CodeGen/MachineLICM.h" |
19 | #include "llvm/ADT/BitVector.h" |
20 | #include "llvm/ADT/DenseMap.h" |
21 | #include "llvm/ADT/STLExtras.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/MachineDomTreeUpdater.h" |
28 | #include "llvm/CodeGen/MachineDominators.h" |
29 | #include "llvm/CodeGen/MachineFrameInfo.h" |
30 | #include "llvm/CodeGen/MachineFunction.h" |
31 | #include "llvm/CodeGen/MachineFunctionPass.h" |
32 | #include "llvm/CodeGen/MachineInstr.h" |
33 | #include "llvm/CodeGen/MachineLoopInfo.h" |
34 | #include "llvm/CodeGen/MachineMemOperand.h" |
35 | #include "llvm/CodeGen/MachineOperand.h" |
36 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
37 | #include "llvm/CodeGen/PseudoSourceValue.h" |
38 | #include "llvm/CodeGen/TargetInstrInfo.h" |
39 | #include "llvm/CodeGen/TargetLowering.h" |
40 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
41 | #include "llvm/CodeGen/TargetSchedule.h" |
42 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
43 | #include "llvm/IR/DebugLoc.h" |
44 | #include "llvm/InitializePasses.h" |
45 | #include "llvm/MC/MCInstrDesc.h" |
46 | #include "llvm/MC/MCRegister.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(Val: 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(Val: false), cl::Hidden); |
70 | |
71 | static cl::opt<bool> |
72 | HoistConstStores("hoist-const-stores" , |
73 | cl::desc("Hoist invariant stores" ), |
74 | cl::init(Val: true), cl::Hidden); |
75 | |
76 | static cl::opt<bool> HoistConstLoads("hoist-const-loads" , |
77 | cl::desc("Hoist invariant loads" ), |
78 | cl::init(Val: 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(Val: 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(Val: 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 MachineLICMImpl { |
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 | Pass *LegacyPass; |
130 | MachineFunctionAnalysisManager *MFAM; |
131 | |
132 | // Various analyses that we use... |
133 | AliasAnalysis *AA = nullptr; // Alias analysis info. |
134 | MachineBlockFrequencyInfo *MBFI = nullptr; // Machine block frequncy info |
135 | MachineLoopInfo *MLI = nullptr; // Current MachineLoopInfo |
136 | MachineDomTreeUpdater *MDTU = nullptr; // Wraps current dominator tree |
137 | |
138 | // State that is updated as we process loops |
139 | bool Changed = false; // True if a loop is changed. |
140 | bool FirstInLoop = false; // True if it's the first LICM in the loop. |
141 | |
142 | // Holds information about whether it is allowed to move load instructions |
143 | // out of the loop |
144 | SmallDenseMap<MachineLoop *, bool> AllowedToHoistLoads; |
145 | |
146 | // Exit blocks of each Loop. |
147 | DenseMap<MachineLoop *, SmallVector<MachineBasicBlock *, 8>> ExitBlockMap; |
148 | |
149 | bool isExitBlock(MachineLoop *CurLoop, const MachineBasicBlock *MBB) { |
150 | auto [It, Inserted] = ExitBlockMap.try_emplace(Key: CurLoop); |
151 | if (Inserted) { |
152 | SmallVector<MachineBasicBlock *, 8> ExitBlocks; |
153 | CurLoop->getExitBlocks(ExitBlocks); |
154 | It->second = std::move(ExitBlocks); |
155 | } |
156 | return is_contained(Range&: It->second, Element: MBB); |
157 | } |
158 | |
159 | // Track 'estimated' register pressure. |
160 | SmallDenseSet<Register> RegSeen; |
161 | SmallVector<unsigned, 8> RegPressure; |
162 | |
163 | // Register pressure "limit" per register pressure set. If the pressure |
164 | // is higher than the limit, then it's considered high. |
165 | SmallVector<unsigned, 8> RegLimit; |
166 | |
167 | // Register pressure on path leading from loop preheader to current BB. |
168 | SmallVector<SmallVector<unsigned, 8>, 16> BackTrace; |
169 | |
170 | // For each opcode per preheader, keep a list of potential CSE instructions. |
171 | DenseMap<MachineBasicBlock *, |
172 | DenseMap<unsigned, std::vector<MachineInstr *>>> |
173 | CSEMap; |
174 | |
175 | enum { |
176 | SpeculateFalse = 0, |
177 | SpeculateTrue = 1, |
178 | SpeculateUnknown = 2 |
179 | }; |
180 | |
181 | // If a MBB does not dominate loop exiting blocks then it may not safe |
182 | // to hoist loads from this block. |
183 | // Tri-state: 0 - false, 1 - true, 2 - unknown |
184 | unsigned SpeculationState = SpeculateUnknown; |
185 | |
186 | public: |
187 | MachineLICMImpl(bool PreRegAlloc, Pass *LegacyPass, |
188 | MachineFunctionAnalysisManager *MFAM) |
189 | : PreRegAlloc(PreRegAlloc), LegacyPass(LegacyPass), MFAM(MFAM) { |
190 | assert((LegacyPass || MFAM) && "LegacyPass or MFAM must be provided" ); |
191 | assert(!(LegacyPass && MFAM) && |
192 | "LegacyPass and MFAM cannot be provided at the same time" ); |
193 | } |
194 | |
195 | bool run(MachineFunction &MF); |
196 | |
197 | void releaseMemory() { |
198 | RegSeen.clear(); |
199 | RegPressure.clear(); |
200 | RegLimit.clear(); |
201 | BackTrace.clear(); |
202 | CSEMap.clear(); |
203 | ExitBlockMap.clear(); |
204 | } |
205 | |
206 | private: |
207 | /// Keep track of information about hoisting candidates. |
208 | struct CandidateInfo { |
209 | MachineInstr *MI; |
210 | Register Def; |
211 | int FI; |
212 | |
213 | CandidateInfo(MachineInstr *mi, Register def, int fi) |
214 | : MI(mi), Def(def), FI(fi) {} |
215 | }; |
216 | |
217 | void HoistRegionPostRA(MachineLoop *CurLoop); |
218 | |
219 | void HoistPostRA(MachineInstr *MI, Register Def, MachineLoop *CurLoop); |
220 | |
221 | void ProcessMI(MachineInstr *MI, BitVector &RUDefs, BitVector &RUClobbers, |
222 | SmallDenseSet<int> &StoredFIs, |
223 | SmallVectorImpl<CandidateInfo> &Candidates, |
224 | MachineLoop *CurLoop); |
225 | |
226 | void AddToLiveIns(MCRegister Reg, MachineLoop *CurLoop); |
227 | |
228 | bool IsLICMCandidate(MachineInstr &I, MachineLoop *CurLoop); |
229 | |
230 | bool IsLoopInvariantInst(MachineInstr &I, MachineLoop *CurLoop); |
231 | |
232 | bool HasLoopPHIUse(const MachineInstr *MI, MachineLoop *CurLoop); |
233 | |
234 | bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx, Register Reg, |
235 | MachineLoop *CurLoop) const; |
236 | |
237 | bool IsCheapInstruction(MachineInstr &MI) const; |
238 | |
239 | bool CanCauseHighRegPressure(const SmallDenseMap<unsigned, int> &Cost, |
240 | bool Cheap); |
241 | |
242 | void UpdateBackTraceRegPressure(const MachineInstr *MI); |
243 | |
244 | bool IsProfitableToHoist(MachineInstr &MI, MachineLoop *CurLoop); |
245 | |
246 | bool IsGuaranteedToExecute(MachineBasicBlock *BB, MachineLoop *CurLoop); |
247 | |
248 | bool isTriviallyReMaterializable(const MachineInstr &MI) const; |
249 | |
250 | void EnterScope(MachineBasicBlock *MBB); |
251 | |
252 | void ExitScope(MachineBasicBlock *MBB); |
253 | |
254 | void ExitScopeIfDone( |
255 | MachineDomTreeNode *Node, |
256 | DenseMap<MachineDomTreeNode *, unsigned> &OpenChildren, |
257 | const DenseMap<MachineDomTreeNode *, MachineDomTreeNode *> &ParentMap); |
258 | |
259 | void HoistOutOfLoop(MachineDomTreeNode *, MachineLoop *CurLoop); |
260 | |
261 | void InitRegPressure(MachineBasicBlock *BB); |
262 | |
263 | SmallDenseMap<unsigned, int> calcRegisterCost(const MachineInstr *MI, |
264 | bool ConsiderSeen, |
265 | bool ConsiderUnseenAsDef); |
266 | |
267 | void UpdateRegPressure(const MachineInstr *MI, |
268 | bool ConsiderUnseenAsDef = false); |
269 | |
270 | MachineInstr *ExtractHoistableLoad(MachineInstr *MI, MachineLoop *CurLoop); |
271 | |
272 | MachineInstr *LookForDuplicate(const MachineInstr *MI, |
273 | std::vector<MachineInstr *> &PrevMIs); |
274 | |
275 | bool |
276 | EliminateCSE(MachineInstr *MI, |
277 | DenseMap<unsigned, std::vector<MachineInstr *>>::iterator &CI); |
278 | |
279 | bool MayCSE(MachineInstr *MI); |
280 | |
281 | unsigned Hoist(MachineInstr *MI, MachineBasicBlock *, |
282 | MachineLoop *CurLoop); |
283 | |
284 | void InitCSEMap(MachineBasicBlock *BB); |
285 | |
286 | void InitializeLoadsHoistableLoops(); |
287 | |
288 | bool isTgtHotterThanSrc(MachineBasicBlock *SrcBlock, |
289 | MachineBasicBlock *TgtBlock); |
290 | MachineBasicBlock *getOrCreatePreheader(MachineLoop *CurLoop); |
291 | }; |
292 | |
293 | class MachineLICMBase : public MachineFunctionPass { |
294 | bool PreRegAlloc; |
295 | |
296 | public: |
297 | MachineLICMBase(char &ID, bool PreRegAlloc) |
298 | : MachineFunctionPass(ID), PreRegAlloc(PreRegAlloc) {} |
299 | |
300 | bool runOnMachineFunction(MachineFunction &MF) override; |
301 | |
302 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
303 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
304 | if (DisableHoistingToHotterBlocks != UseBFI::None) |
305 | AU.addRequired<MachineBlockFrequencyInfoWrapperPass>(); |
306 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
307 | AU.addRequired<AAResultsWrapperPass>(); |
308 | AU.addPreserved<MachineLoopInfoWrapperPass>(); |
309 | MachineFunctionPass::getAnalysisUsage(AU); |
310 | } |
311 | }; |
312 | |
313 | class MachineLICM : public MachineLICMBase { |
314 | public: |
315 | static char ID; |
316 | MachineLICM() : MachineLICMBase(ID, false) { |
317 | initializeMachineLICMPass(*PassRegistry::getPassRegistry()); |
318 | } |
319 | }; |
320 | |
321 | class EarlyMachineLICM : public MachineLICMBase { |
322 | public: |
323 | static char ID; |
324 | EarlyMachineLICM() : MachineLICMBase(ID, true) { |
325 | initializeEarlyMachineLICMPass(*PassRegistry::getPassRegistry()); |
326 | } |
327 | }; |
328 | |
329 | } // end anonymous namespace |
330 | |
331 | char MachineLICM::ID; |
332 | char EarlyMachineLICM::ID; |
333 | |
334 | char &llvm::MachineLICMID = MachineLICM::ID; |
335 | char &llvm::EarlyMachineLICMID = EarlyMachineLICM::ID; |
336 | |
337 | INITIALIZE_PASS_BEGIN(MachineLICM, DEBUG_TYPE, |
338 | "Machine Loop Invariant Code Motion" , false, false) |
339 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
340 | INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfoWrapperPass) |
341 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
342 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
343 | INITIALIZE_PASS_END(MachineLICM, DEBUG_TYPE, |
344 | "Machine Loop Invariant Code Motion" , false, false) |
345 | |
346 | INITIALIZE_PASS_BEGIN(EarlyMachineLICM, "early-machinelicm" , |
347 | "Early Machine Loop Invariant Code Motion" , false, false) |
348 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
349 | INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfoWrapperPass) |
350 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
351 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
352 | INITIALIZE_PASS_END(EarlyMachineLICM, "early-machinelicm" , |
353 | "Early Machine Loop Invariant Code Motion" , false, false) |
354 | |
355 | bool MachineLICMBase::runOnMachineFunction(MachineFunction &MF) { |
356 | if (skipFunction(F: MF.getFunction())) |
357 | return false; |
358 | |
359 | MachineLICMImpl Impl(PreRegAlloc, this, nullptr); |
360 | return Impl.run(MF); |
361 | } |
362 | |
363 | #define GET_RESULT(RESULT, GETTER, INFIX) \ |
364 | ((LegacyPass) \ |
365 | ? &LegacyPass->getAnalysis<RESULT##INFIX##WrapperPass>().GETTER() \ |
366 | : &MFAM->getResult<RESULT##Analysis>(MF)) |
367 | |
368 | bool MachineLICMImpl::run(MachineFunction &MF) { |
369 | AA = MFAM != nullptr |
370 | ? &MFAM->getResult<FunctionAnalysisManagerMachineFunctionProxy>(IR&: MF) |
371 | .getManager() |
372 | .getResult<AAManager>(IR&: MF.getFunction()) |
373 | : &LegacyPass->getAnalysis<AAResultsWrapperPass>().getAAResults(); |
374 | MachineDomTreeUpdater DTU(GET_RESULT(MachineDominatorTree, getDomTree, ), |
375 | MachineDomTreeUpdater::UpdateStrategy::Lazy); |
376 | MDTU = &DTU; |
377 | MLI = GET_RESULT(MachineLoop, getLI, Info); |
378 | MBFI = DisableHoistingToHotterBlocks != UseBFI::None |
379 | ? GET_RESULT(MachineBlockFrequency, getMBFI, Info) |
380 | : nullptr; |
381 | |
382 | Changed = FirstInLoop = false; |
383 | const TargetSubtargetInfo &ST = MF.getSubtarget(); |
384 | TII = ST.getInstrInfo(); |
385 | TLI = ST.getTargetLowering(); |
386 | TRI = ST.getRegisterInfo(); |
387 | MFI = &MF.getFrameInfo(); |
388 | MRI = &MF.getRegInfo(); |
389 | SchedModel.init(TSInfo: &ST); |
390 | |
391 | HasProfileData = MF.getFunction().hasProfileData(); |
392 | |
393 | if (PreRegAlloc) |
394 | LLVM_DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: " ); |
395 | else |
396 | LLVM_DEBUG(dbgs() << "******** Post-regalloc Machine LICM: " ); |
397 | LLVM_DEBUG(dbgs() << MF.getName() << " ********\n" ); |
398 | |
399 | if (PreRegAlloc) { |
400 | // Estimate register pressure during pre-regalloc pass. |
401 | unsigned NumRPS = TRI->getNumRegPressureSets(); |
402 | RegPressure.resize(N: NumRPS); |
403 | llvm::fill(Range&: RegPressure, Value: 0); |
404 | RegLimit.resize(N: NumRPS); |
405 | for (unsigned i = 0, e = NumRPS; i != e; ++i) |
406 | RegLimit[i] = TRI->getRegPressureSetLimit(MF, Idx: i); |
407 | } |
408 | |
409 | if (HoistConstLoads) |
410 | InitializeLoadsHoistableLoops(); |
411 | |
412 | SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end()); |
413 | while (!Worklist.empty()) { |
414 | MachineLoop *CurLoop = Worklist.pop_back_val(); |
415 | |
416 | if (!PreRegAlloc) { |
417 | HoistRegionPostRA(CurLoop); |
418 | } else { |
419 | // CSEMap is initialized for loop header when the first instruction is |
420 | // being hoisted. |
421 | MachineDomTreeNode *N = MDTU->getDomTree().getNode(BB: CurLoop->getHeader()); |
422 | FirstInLoop = true; |
423 | HoistOutOfLoop(HeaderN: N, CurLoop); |
424 | CSEMap.clear(); |
425 | } |
426 | } |
427 | releaseMemory(); |
428 | return Changed; |
429 | } |
430 | |
431 | /// Return true if instruction stores to the specified frame. |
432 | static bool InstructionStoresToFI(const MachineInstr *MI, int FI) { |
433 | // Check mayStore before memory operands so that e.g. DBG_VALUEs will return |
434 | // true since they have no memory operands. |
435 | if (!MI->mayStore()) |
436 | return false; |
437 | // If we lost memory operands, conservatively assume that the instruction |
438 | // writes to all slots. |
439 | if (MI->memoperands_empty()) |
440 | return true; |
441 | for (const MachineMemOperand *MemOp : MI->memoperands()) { |
442 | if (!MemOp->isStore() || !MemOp->getPseudoValue()) |
443 | continue; |
444 | if (const FixedStackPseudoSourceValue *Value = |
445 | dyn_cast<FixedStackPseudoSourceValue>(Val: MemOp->getPseudoValue())) { |
446 | if (Value->getFrameIndex() == FI) |
447 | return true; |
448 | } |
449 | } |
450 | return false; |
451 | } |
452 | |
453 | static void applyBitsNotInRegMaskToRegUnitsMask(const TargetRegisterInfo &TRI, |
454 | BitVector &RUs, |
455 | const uint32_t *Mask) { |
456 | // FIXME: This intentionally works in reverse due to some issues with the |
457 | // Register Units infrastructure. |
458 | // |
459 | // This is used to apply callee-saved-register masks to the clobbered regunits |
460 | // mask. |
461 | // |
462 | // The right way to approach this is to start with a BitVector full of ones, |
463 | // then reset all the bits of the regunits of each register that is set in the |
464 | // mask (registers preserved), then OR the resulting bits with the Clobbers |
465 | // mask. This correctly prioritizes the saved registers, so if a RU is shared |
466 | // between a register that is preserved, and one that is NOT preserved, that |
467 | // RU will not be set in the output vector (the clobbers). |
468 | // |
469 | // What we have to do for now is the opposite: we have to assume that the |
470 | // regunits of all registers that are NOT preserved are clobbered, even if |
471 | // those regunits are preserved by another register. So if a RU is shared |
472 | // like described previously, that RU will be set. |
473 | // |
474 | // This is to work around an issue which appears in AArch64, but isn't |
475 | // exclusive to that target: AArch64's Qn registers (128 bits) have Dn |
476 | // register (lower 64 bits). A few Dn registers are preserved by some calling |
477 | // conventions, but Qn and Dn share exactly the same reg units. |
478 | // |
479 | // If we do this the right way, Qn will be marked as NOT clobbered even though |
480 | // its upper 64 bits are NOT preserved. The conservative approach handles this |
481 | // correctly at the cost of some missed optimizations on other targets. |
482 | // |
483 | // This is caused by how RegUnits are handled within TableGen. Ideally, Qn |
484 | // should have an extra RegUnit to model the "unknown" bits not covered by the |
485 | // subregs. |
486 | BitVector RUsFromRegsNotInMask(TRI.getNumRegUnits()); |
487 | const unsigned NumRegs = TRI.getNumRegs(); |
488 | const unsigned MaskWords = (NumRegs + 31) / 32; |
489 | for (unsigned K = 0; K < MaskWords; ++K) { |
490 | const uint32_t Word = Mask[K]; |
491 | for (unsigned Bit = 0; Bit < 32; ++Bit) { |
492 | const unsigned PhysReg = (K * 32) + Bit; |
493 | if (PhysReg == NumRegs) |
494 | break; |
495 | |
496 | if (PhysReg && !((Word >> Bit) & 1)) { |
497 | for (MCRegUnit Unit : TRI.regunits(Reg: PhysReg)) |
498 | RUsFromRegsNotInMask.set(Unit); |
499 | } |
500 | } |
501 | } |
502 | |
503 | RUs |= RUsFromRegsNotInMask; |
504 | } |
505 | |
506 | /// Examine the instruction for potential LICM candidate. Also |
507 | /// gather register def and frame object update information. |
508 | void MachineLICMImpl::ProcessMI(MachineInstr *MI, BitVector &RUDefs, |
509 | BitVector &RUClobbers, |
510 | SmallDenseSet<int> &StoredFIs, |
511 | SmallVectorImpl<CandidateInfo> &Candidates, |
512 | MachineLoop *CurLoop) { |
513 | bool RuledOut = false; |
514 | bool HasNonInvariantUse = false; |
515 | Register Def; |
516 | for (const MachineOperand &MO : MI->operands()) { |
517 | if (MO.isFI()) { |
518 | // Remember if the instruction stores to the frame index. |
519 | int FI = MO.getIndex(); |
520 | if (!StoredFIs.count(V: FI) && |
521 | MFI->isSpillSlotObjectIndex(ObjectIdx: FI) && |
522 | InstructionStoresToFI(MI, FI)) |
523 | StoredFIs.insert(V: FI); |
524 | HasNonInvariantUse = true; |
525 | continue; |
526 | } |
527 | |
528 | // We can't hoist an instruction defining a physreg that is clobbered in |
529 | // the loop. |
530 | if (MO.isRegMask()) { |
531 | applyBitsNotInRegMaskToRegUnitsMask(TRI: *TRI, RUs&: RUClobbers, Mask: MO.getRegMask()); |
532 | continue; |
533 | } |
534 | |
535 | if (!MO.isReg()) |
536 | continue; |
537 | Register Reg = MO.getReg(); |
538 | if (!Reg) |
539 | continue; |
540 | assert(Reg.isPhysical() && "Not expecting virtual register!" ); |
541 | |
542 | if (!MO.isDef()) { |
543 | if (!HasNonInvariantUse) { |
544 | for (MCRegUnit Unit : TRI->regunits(Reg)) { |
545 | // If it's using a non-loop-invariant register, then it's obviously |
546 | // not safe to hoist. |
547 | if (RUDefs.test(Idx: Unit) || RUClobbers.test(Idx: Unit)) { |
548 | HasNonInvariantUse = true; |
549 | break; |
550 | } |
551 | } |
552 | } |
553 | continue; |
554 | } |
555 | |
556 | if (MO.isImplicit()) { |
557 | for (MCRegUnit Unit : TRI->regunits(Reg)) |
558 | RUClobbers.set(Unit); |
559 | if (!MO.isDead()) |
560 | // Non-dead implicit def? This cannot be hoisted. |
561 | RuledOut = true; |
562 | // No need to check if a dead implicit def is also defined by |
563 | // another instruction. |
564 | continue; |
565 | } |
566 | |
567 | // FIXME: For now, avoid instructions with multiple defs, unless |
568 | // it's a dead implicit def. |
569 | if (Def) |
570 | RuledOut = true; |
571 | else |
572 | Def = Reg; |
573 | |
574 | // If we have already seen another instruction that defines the same |
575 | // register, then this is not safe. Two defs is indicated by setting a |
576 | // PhysRegClobbers bit. |
577 | for (MCRegUnit Unit : TRI->regunits(Reg)) { |
578 | if (RUDefs.test(Idx: Unit)) { |
579 | RUClobbers.set(Unit); |
580 | RuledOut = true; |
581 | } else if (RUClobbers.test(Idx: Unit)) { |
582 | // MI defined register is seen defined by another instruction in |
583 | // the loop, it cannot be a LICM candidate. |
584 | RuledOut = true; |
585 | } |
586 | |
587 | RUDefs.set(Unit); |
588 | } |
589 | } |
590 | |
591 | // Only consider reloads for now and remats which do not have register |
592 | // operands. FIXME: Consider unfold load folding instructions. |
593 | if (Def && !RuledOut) { |
594 | int FI = std::numeric_limits<int>::min(); |
595 | if ((!HasNonInvariantUse && IsLICMCandidate(I&: *MI, CurLoop)) || |
596 | (TII->isLoadFromStackSlot(MI: *MI, FrameIndex&: FI) && MFI->isSpillSlotObjectIndex(ObjectIdx: FI))) |
597 | Candidates.push_back(Elt: CandidateInfo(MI, Def, FI)); |
598 | } |
599 | } |
600 | |
601 | /// Walk the specified region of the CFG and hoist loop invariants out to the |
602 | /// preheader. |
603 | void MachineLICMImpl::HoistRegionPostRA(MachineLoop *CurLoop) { |
604 | MachineBasicBlock * = getOrCreatePreheader(CurLoop); |
605 | if (!Preheader) |
606 | return; |
607 | |
608 | unsigned NumRegUnits = TRI->getNumRegUnits(); |
609 | BitVector RUDefs(NumRegUnits); // RUs defined once in the loop. |
610 | BitVector RUClobbers(NumRegUnits); // RUs defined more than once. |
611 | |
612 | SmallVector<CandidateInfo, 32> Candidates; |
613 | SmallDenseSet<int> StoredFIs; |
614 | |
615 | // Walk the entire region, count number of defs for each register, and |
616 | // collect potential LICM candidates. |
617 | for (MachineBasicBlock *BB : CurLoop->getBlocks()) { |
618 | // If the header of the loop containing this basic block is a landing pad, |
619 | // then don't try to hoist instructions out of this loop. |
620 | const MachineLoop *ML = MLI->getLoopFor(BB); |
621 | if (ML && ML->getHeader()->isEHPad()) continue; |
622 | |
623 | // Conservatively treat live-in's as an external def. |
624 | // FIXME: That means a reload that're reused in successor block(s) will not |
625 | // be LICM'ed. |
626 | for (const auto &LI : BB->liveins()) { |
627 | for (MCRegUnit Unit : TRI->regunits(Reg: LI.PhysReg)) |
628 | RUDefs.set(Unit); |
629 | } |
630 | |
631 | // Funclet entry blocks will clobber all registers |
632 | if (const uint32_t *Mask = BB->getBeginClobberMask(TRI)) |
633 | applyBitsNotInRegMaskToRegUnitsMask(TRI: *TRI, RUs&: RUClobbers, Mask); |
634 | |
635 | // EH landing pads clobber exception pointer/selector registers. |
636 | if (BB->isEHPad()) { |
637 | const MachineFunction &MF = *BB->getParent(); |
638 | const Constant *PersonalityFn = MF.getFunction().getPersonalityFn(); |
639 | const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering(); |
640 | if (MCRegister Reg = TLI.getExceptionPointerRegister(PersonalityFn)) |
641 | for (MCRegUnit Unit : TRI->regunits(Reg)) |
642 | RUClobbers.set(Unit); |
643 | if (MCRegister Reg = TLI.getExceptionSelectorRegister(PersonalityFn)) |
644 | for (MCRegUnit Unit : TRI->regunits(Reg)) |
645 | RUClobbers.set(Unit); |
646 | } |
647 | |
648 | SpeculationState = SpeculateUnknown; |
649 | for (MachineInstr &MI : *BB) |
650 | ProcessMI(MI: &MI, RUDefs, RUClobbers, StoredFIs, Candidates, CurLoop); |
651 | } |
652 | |
653 | // Gather the registers read / clobbered by the terminator. |
654 | BitVector TermRUs(NumRegUnits); |
655 | MachineBasicBlock::iterator TI = Preheader->getFirstTerminator(); |
656 | if (TI != Preheader->end()) { |
657 | for (const MachineOperand &MO : TI->operands()) { |
658 | if (!MO.isReg()) |
659 | continue; |
660 | Register Reg = MO.getReg(); |
661 | if (!Reg) |
662 | continue; |
663 | for (MCRegUnit Unit : TRI->regunits(Reg)) |
664 | TermRUs.set(Unit); |
665 | } |
666 | } |
667 | |
668 | // Now evaluate whether the potential candidates qualify. |
669 | // 1. Check if the candidate defined register is defined by another |
670 | // instruction in the loop. |
671 | // 2. If the candidate is a load from stack slot (always true for now), |
672 | // check if the slot is stored anywhere in the loop. |
673 | // 3. Make sure candidate def should not clobber |
674 | // registers read by the terminator. Similarly its def should not be |
675 | // clobbered by the terminator. |
676 | for (CandidateInfo &Candidate : Candidates) { |
677 | if (Candidate.FI != std::numeric_limits<int>::min() && |
678 | StoredFIs.count(V: Candidate.FI)) |
679 | continue; |
680 | |
681 | Register Def = Candidate.Def; |
682 | bool Safe = true; |
683 | for (MCRegUnit Unit : TRI->regunits(Reg: Def)) { |
684 | if (RUClobbers.test(Idx: Unit) || TermRUs.test(Idx: Unit)) { |
685 | Safe = false; |
686 | break; |
687 | } |
688 | } |
689 | |
690 | if (!Safe) |
691 | continue; |
692 | |
693 | MachineInstr *MI = Candidate.MI; |
694 | for (const MachineOperand &MO : MI->all_uses()) { |
695 | if (!MO.getReg()) |
696 | continue; |
697 | for (MCRegUnit Unit : TRI->regunits(Reg: MO.getReg())) { |
698 | if (RUDefs.test(Idx: Unit) || RUClobbers.test(Idx: Unit)) { |
699 | // If it's using a non-loop-invariant register, then it's obviously |
700 | // not safe to hoist. |
701 | Safe = false; |
702 | break; |
703 | } |
704 | } |
705 | |
706 | if (!Safe) |
707 | break; |
708 | } |
709 | |
710 | if (Safe) |
711 | HoistPostRA(MI, Def: Candidate.Def, CurLoop); |
712 | } |
713 | } |
714 | |
715 | /// Add register 'Reg' to the livein sets of BBs in the current loop, and make |
716 | /// sure it is not killed by any instructions in the loop. |
717 | void MachineLICMImpl::AddToLiveIns(MCRegister Reg, MachineLoop *CurLoop) { |
718 | for (MachineBasicBlock *BB : CurLoop->getBlocks()) { |
719 | if (!BB->isLiveIn(Reg)) |
720 | BB->addLiveIn(PhysReg: Reg); |
721 | for (MachineInstr &MI : *BB) { |
722 | for (MachineOperand &MO : MI.all_uses()) { |
723 | if (!MO.getReg()) |
724 | continue; |
725 | if (TRI->regsOverlap(RegA: Reg, RegB: MO.getReg())) |
726 | MO.setIsKill(false); |
727 | } |
728 | } |
729 | } |
730 | } |
731 | |
732 | /// When an instruction is found to only use loop invariant operands that is |
733 | /// safe to hoist, this instruction is called to do the dirty work. |
734 | void MachineLICMImpl::HoistPostRA(MachineInstr *MI, Register Def, |
735 | MachineLoop *CurLoop) { |
736 | MachineBasicBlock * = CurLoop->getLoopPreheader(); |
737 | |
738 | // Now move the instructions to the predecessor, inserting it before any |
739 | // terminator instructions. |
740 | LLVM_DEBUG(dbgs() << "Hoisting to " << printMBBReference(*Preheader) |
741 | << " from " << printMBBReference(*MI->getParent()) << ": " |
742 | << *MI); |
743 | |
744 | // Splice the instruction to the preheader. |
745 | MachineBasicBlock *MBB = MI->getParent(); |
746 | Preheader->splice(Where: Preheader->getFirstTerminator(), Other: MBB, From: MI); |
747 | |
748 | // Since we are moving the instruction out of its basic block, we do not |
749 | // retain its debug location. Doing so would degrade the debugging |
750 | // experience and adversely affect the accuracy of profiling information. |
751 | assert(!MI->isDebugInstr() && "Should not hoist debug inst" ); |
752 | MI->setDebugLoc(DebugLoc()); |
753 | |
754 | // Add register to livein list to all the BBs in the current loop since a |
755 | // loop invariant must be kept live throughout the whole loop. This is |
756 | // important to ensure later passes do not scavenge the def register. |
757 | AddToLiveIns(Reg: Def, CurLoop); |
758 | |
759 | ++NumPostRAHoisted; |
760 | Changed = true; |
761 | } |
762 | |
763 | /// Check if this mbb is guaranteed to execute. If not then a load from this mbb |
764 | /// may not be safe to hoist. |
765 | bool MachineLICMImpl::IsGuaranteedToExecute(MachineBasicBlock *BB, |
766 | MachineLoop *CurLoop) { |
767 | if (SpeculationState != SpeculateUnknown) |
768 | return SpeculationState == SpeculateFalse; |
769 | |
770 | if (BB != CurLoop->getHeader()) { |
771 | // Check loop exiting blocks. |
772 | SmallVector<MachineBasicBlock*, 8> CurrentLoopExitingBlocks; |
773 | CurLoop->getExitingBlocks(ExitingBlocks&: CurrentLoopExitingBlocks); |
774 | for (MachineBasicBlock *CurrentLoopExitingBlock : CurrentLoopExitingBlocks) |
775 | if (!MDTU->getDomTree().dominates(A: BB, B: CurrentLoopExitingBlock)) { |
776 | SpeculationState = SpeculateTrue; |
777 | return false; |
778 | } |
779 | } |
780 | |
781 | SpeculationState = SpeculateFalse; |
782 | return true; |
783 | } |
784 | |
785 | /// Check if \p MI is trivially remateralizable and if it does not have any |
786 | /// virtual register uses. Even though rematerializable RA might not actually |
787 | /// rematerialize it in this scenario. In that case we do not want to hoist such |
788 | /// instruction out of the loop in a belief RA will sink it back if needed. |
789 | bool MachineLICMImpl::isTriviallyReMaterializable( |
790 | const MachineInstr &MI) const { |
791 | if (!TII->isTriviallyReMaterializable(MI)) |
792 | return false; |
793 | |
794 | for (const MachineOperand &MO : MI.all_uses()) { |
795 | if (MO.getReg().isVirtual()) |
796 | return false; |
797 | } |
798 | |
799 | return true; |
800 | } |
801 | |
802 | void MachineLICMImpl::EnterScope(MachineBasicBlock *MBB) { |
803 | LLVM_DEBUG(dbgs() << "Entering " << printMBBReference(*MBB) << '\n'); |
804 | |
805 | // Remember livein register pressure. |
806 | BackTrace.push_back(Elt: RegPressure); |
807 | } |
808 | |
809 | void MachineLICMImpl::ExitScope(MachineBasicBlock *MBB) { |
810 | LLVM_DEBUG(dbgs() << "Exiting " << printMBBReference(*MBB) << '\n'); |
811 | BackTrace.pop_back(); |
812 | } |
813 | |
814 | /// Destroy scope for the MBB that corresponds to the given dominator tree node |
815 | /// if its a leaf or all of its children are done. Walk up the dominator tree to |
816 | /// destroy ancestors which are now done. |
817 | void MachineLICMImpl::ExitScopeIfDone( |
818 | MachineDomTreeNode *Node, |
819 | DenseMap<MachineDomTreeNode *, unsigned> &OpenChildren, |
820 | const DenseMap<MachineDomTreeNode *, MachineDomTreeNode *> &ParentMap) { |
821 | if (OpenChildren[Node]) |
822 | return; |
823 | |
824 | for(;;) { |
825 | ExitScope(MBB: Node->getBlock()); |
826 | // Now traverse upwards to pop ancestors whose offsprings are all done. |
827 | MachineDomTreeNode *Parent = ParentMap.lookup(Val: Node); |
828 | if (!Parent || --OpenChildren[Parent] != 0) |
829 | break; |
830 | Node = Parent; |
831 | } |
832 | } |
833 | |
834 | /// Walk the specified loop in the CFG (defined by all blocks dominated by the |
835 | /// specified header block, and that are in the current loop) in depth first |
836 | /// order w.r.t the DominatorTree. This allows us to visit definitions before |
837 | /// uses, allowing us to hoist a loop body in one pass without iteration. |
838 | void MachineLICMImpl::HoistOutOfLoop(MachineDomTreeNode *, |
839 | MachineLoop *CurLoop) { |
840 | MachineBasicBlock * = getOrCreatePreheader(CurLoop); |
841 | if (!Preheader) |
842 | return; |
843 | |
844 | SmallVector<MachineDomTreeNode*, 32> Scopes; |
845 | SmallVector<MachineDomTreeNode*, 8> WorkList; |
846 | DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> ParentMap; |
847 | DenseMap<MachineDomTreeNode*, unsigned> OpenChildren; |
848 | |
849 | // Perform a DFS walk to determine the order of visit. |
850 | WorkList.push_back(Elt: HeaderN); |
851 | while (!WorkList.empty()) { |
852 | MachineDomTreeNode *Node = WorkList.pop_back_val(); |
853 | assert(Node && "Null dominator tree node?" ); |
854 | MachineBasicBlock *BB = Node->getBlock(); |
855 | |
856 | // If the header of the loop containing this basic block is a landing pad, |
857 | // then don't try to hoist instructions out of this loop. |
858 | const MachineLoop *ML = MLI->getLoopFor(BB); |
859 | if (ML && ML->getHeader()->isEHPad()) |
860 | continue; |
861 | |
862 | // If this subregion is not in the top level loop at all, exit. |
863 | if (!CurLoop->contains(BB)) |
864 | continue; |
865 | |
866 | Scopes.push_back(Elt: Node); |
867 | unsigned NumChildren = Node->getNumChildren(); |
868 | |
869 | // Don't hoist things out of a large switch statement. This often causes |
870 | // code to be hoisted that wasn't going to be executed, and increases |
871 | // register pressure in a situation where it's likely to matter. |
872 | if (BB->succ_size() >= 25) |
873 | NumChildren = 0; |
874 | |
875 | OpenChildren[Node] = NumChildren; |
876 | if (NumChildren) { |
877 | // Add children in reverse order as then the next popped worklist node is |
878 | // the first child of this node. This means we ultimately traverse the |
879 | // DOM tree in exactly the same order as if we'd recursed. |
880 | for (MachineDomTreeNode *Child : reverse(C: Node->children())) { |
881 | ParentMap[Child] = Node; |
882 | WorkList.push_back(Elt: Child); |
883 | } |
884 | } |
885 | } |
886 | |
887 | if (Scopes.size() == 0) |
888 | return; |
889 | |
890 | // Compute registers which are livein into the loop headers. |
891 | RegSeen.clear(); |
892 | BackTrace.clear(); |
893 | InitRegPressure(BB: Preheader); |
894 | |
895 | // Now perform LICM. |
896 | for (MachineDomTreeNode *Node : Scopes) { |
897 | MachineBasicBlock *MBB = Node->getBlock(); |
898 | |
899 | EnterScope(MBB); |
900 | |
901 | // Process the block |
902 | SpeculationState = SpeculateUnknown; |
903 | for (MachineInstr &MI : llvm::make_early_inc_range(Range&: *MBB)) { |
904 | unsigned HoistRes = HoistResult::NotHoisted; |
905 | HoistRes = Hoist(MI: &MI, Preheader, CurLoop); |
906 | if (HoistRes & HoistResult::NotHoisted) { |
907 | // We have failed to hoist MI to outermost loop's preheader. If MI is in |
908 | // a subloop, try to hoist it to subloop's preheader. |
909 | SmallVector<MachineLoop *> InnerLoopWorkList; |
910 | for (MachineLoop *L = MLI->getLoopFor(BB: MI.getParent()); L != CurLoop; |
911 | L = L->getParentLoop()) |
912 | InnerLoopWorkList.push_back(Elt: L); |
913 | |
914 | while (!InnerLoopWorkList.empty()) { |
915 | MachineLoop *InnerLoop = InnerLoopWorkList.pop_back_val(); |
916 | MachineBasicBlock * = InnerLoop->getLoopPreheader(); |
917 | if (InnerLoopPreheader) { |
918 | HoistRes = Hoist(MI: &MI, Preheader: InnerLoopPreheader, CurLoop: InnerLoop); |
919 | if (HoistRes & HoistResult::Hoisted) |
920 | break; |
921 | } |
922 | } |
923 | } |
924 | |
925 | if (HoistRes & HoistResult::ErasedMI) |
926 | continue; |
927 | |
928 | UpdateRegPressure(MI: &MI); |
929 | } |
930 | |
931 | // If it's a leaf node, it's done. Traverse upwards to pop ancestors. |
932 | ExitScopeIfDone(Node, OpenChildren, ParentMap); |
933 | } |
934 | } |
935 | |
936 | static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) { |
937 | return MO.isKill() || MRI->hasOneNonDBGUse(RegNo: MO.getReg()); |
938 | } |
939 | |
940 | /// Find all virtual register references that are liveout of the preheader to |
941 | /// initialize the starting "register pressure". Note this does not count live |
942 | /// through (livein but not used) registers. |
943 | void MachineLICMImpl::InitRegPressure(MachineBasicBlock *BB) { |
944 | llvm::fill(Range&: RegPressure, Value: 0); |
945 | |
946 | // If the preheader has only a single predecessor and it ends with a |
947 | // fallthrough or an unconditional branch, then scan its predecessor for live |
948 | // defs as well. This happens whenever the preheader is created by splitting |
949 | // the critical edge from the loop predecessor to the loop header. |
950 | if (BB->pred_size() == 1) { |
951 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; |
952 | SmallVector<MachineOperand, 4> Cond; |
953 | if (!TII->analyzeBranch(MBB&: *BB, TBB, FBB, Cond, AllowModify: false) && Cond.empty()) |
954 | InitRegPressure(BB: *BB->pred_begin()); |
955 | } |
956 | |
957 | for (const MachineInstr &MI : *BB) |
958 | UpdateRegPressure(MI: &MI, /*ConsiderUnseenAsDef=*/true); |
959 | } |
960 | |
961 | /// Update estimate of register pressure after the specified instruction. |
962 | void MachineLICMImpl::UpdateRegPressure(const MachineInstr *MI, |
963 | bool ConsiderUnseenAsDef) { |
964 | auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/true, ConsiderUnseenAsDef); |
965 | for (const auto &RPIdAndCost : Cost) { |
966 | unsigned Class = RPIdAndCost.first; |
967 | if (static_cast<int>(RegPressure[Class]) < -RPIdAndCost.second) |
968 | RegPressure[Class] = 0; |
969 | else |
970 | RegPressure[Class] += RPIdAndCost.second; |
971 | } |
972 | } |
973 | |
974 | /// Calculate the additional register pressure that the registers used in MI |
975 | /// cause. |
976 | /// |
977 | /// If 'ConsiderSeen' is true, updates 'RegSeen' and uses the information to |
978 | /// figure out which usages are live-ins. |
979 | /// FIXME: Figure out a way to consider 'RegSeen' from all code paths. |
980 | SmallDenseMap<unsigned, int> |
981 | MachineLICMImpl::calcRegisterCost(const MachineInstr *MI, bool ConsiderSeen, |
982 | bool ConsiderUnseenAsDef) { |
983 | SmallDenseMap<unsigned, int> Cost; |
984 | if (MI->isImplicitDef()) |
985 | return Cost; |
986 | for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) { |
987 | const MachineOperand &MO = MI->getOperand(i); |
988 | if (!MO.isReg() || MO.isImplicit()) |
989 | continue; |
990 | Register Reg = MO.getReg(); |
991 | if (!Reg.isVirtual()) |
992 | continue; |
993 | |
994 | // FIXME: It seems bad to use RegSeen only for some of these calculations. |
995 | bool isNew = ConsiderSeen ? RegSeen.insert(V: Reg).second : false; |
996 | const TargetRegisterClass *RC = MRI->getRegClass(Reg); |
997 | |
998 | RegClassWeight W = TRI->getRegClassWeight(RC); |
999 | int RCCost = 0; |
1000 | if (MO.isDef()) |
1001 | RCCost = W.RegWeight; |
1002 | else { |
1003 | bool isKill = isOperandKill(MO, MRI); |
1004 | if (isNew && !isKill && ConsiderUnseenAsDef) |
1005 | // Haven't seen this, it must be a livein. |
1006 | RCCost = W.RegWeight; |
1007 | else if (!isNew && isKill) |
1008 | RCCost = -W.RegWeight; |
1009 | } |
1010 | if (RCCost == 0) |
1011 | continue; |
1012 | const int *PS = TRI->getRegClassPressureSets(RC); |
1013 | for (; *PS != -1; ++PS) |
1014 | Cost[*PS] += RCCost; |
1015 | } |
1016 | return Cost; |
1017 | } |
1018 | |
1019 | /// Return true if this machine instruction loads from global offset table or |
1020 | /// constant pool. |
1021 | static bool mayLoadFromGOTOrConstantPool(MachineInstr &MI) { |
1022 | assert(MI.mayLoad() && "Expected MI that loads!" ); |
1023 | |
1024 | // If we lost memory operands, conservatively assume that the instruction |
1025 | // reads from everything.. |
1026 | if (MI.memoperands_empty()) |
1027 | return true; |
1028 | |
1029 | for (MachineMemOperand *MemOp : MI.memoperands()) |
1030 | if (const PseudoSourceValue *PSV = MemOp->getPseudoValue()) |
1031 | if (PSV->isGOT() || PSV->isConstantPool()) |
1032 | return true; |
1033 | |
1034 | return false; |
1035 | } |
1036 | |
1037 | // This function iterates through all the operands of the input store MI and |
1038 | // checks that each register operand statisfies isCallerPreservedPhysReg. |
1039 | // This means, the value being stored and the address where it is being stored |
1040 | // is constant throughout the body of the function (not including prologue and |
1041 | // epilogue). When called with an MI that isn't a store, it returns false. |
1042 | // A future improvement can be to check if the store registers are constant |
1043 | // throughout the loop rather than throughout the funtion. |
1044 | static bool isInvariantStore(const MachineInstr &MI, |
1045 | const TargetRegisterInfo *TRI, |
1046 | const MachineRegisterInfo *MRI) { |
1047 | |
1048 | bool FoundCallerPresReg = false; |
1049 | if (!MI.mayStore() || MI.hasUnmodeledSideEffects() || |
1050 | (MI.getNumOperands() == 0)) |
1051 | return false; |
1052 | |
1053 | // Check that all register operands are caller-preserved physical registers. |
1054 | for (const MachineOperand &MO : MI.operands()) { |
1055 | if (MO.isReg()) { |
1056 | Register Reg = MO.getReg(); |
1057 | // If operand is a virtual register, check if it comes from a copy of a |
1058 | // physical register. |
1059 | if (Reg.isVirtual()) |
1060 | Reg = TRI->lookThruCopyLike(SrcReg: MO.getReg(), MRI); |
1061 | if (Reg.isVirtual()) |
1062 | return false; |
1063 | if (!TRI->isCallerPreservedPhysReg(PhysReg: Reg.asMCReg(), MF: *MI.getMF())) |
1064 | return false; |
1065 | else |
1066 | FoundCallerPresReg = true; |
1067 | } else if (!MO.isImm()) { |
1068 | return false; |
1069 | } |
1070 | } |
1071 | return FoundCallerPresReg; |
1072 | } |
1073 | |
1074 | // Return true if the input MI is a copy instruction that feeds an invariant |
1075 | // store instruction. This means that the src of the copy has to satisfy |
1076 | // isCallerPreservedPhysReg and atleast one of it's users should satisfy |
1077 | // isInvariantStore. |
1078 | static bool isCopyFeedingInvariantStore(const MachineInstr &MI, |
1079 | const MachineRegisterInfo *MRI, |
1080 | const TargetRegisterInfo *TRI) { |
1081 | |
1082 | // FIXME: If targets would like to look through instructions that aren't |
1083 | // pure copies, this can be updated to a query. |
1084 | if (!MI.isCopy()) |
1085 | return false; |
1086 | |
1087 | const MachineFunction *MF = MI.getMF(); |
1088 | // Check that we are copying a constant physical register. |
1089 | Register CopySrcReg = MI.getOperand(i: 1).getReg(); |
1090 | if (CopySrcReg.isVirtual()) |
1091 | return false; |
1092 | |
1093 | if (!TRI->isCallerPreservedPhysReg(PhysReg: CopySrcReg.asMCReg(), MF: *MF)) |
1094 | return false; |
1095 | |
1096 | Register CopyDstReg = MI.getOperand(i: 0).getReg(); |
1097 | // Check if any of the uses of the copy are invariant stores. |
1098 | assert(CopyDstReg.isVirtual() && "copy dst is not a virtual reg" ); |
1099 | |
1100 | for (MachineInstr &UseMI : MRI->use_instructions(Reg: CopyDstReg)) { |
1101 | if (UseMI.mayStore() && isInvariantStore(MI: UseMI, TRI, MRI)) |
1102 | return true; |
1103 | } |
1104 | return false; |
1105 | } |
1106 | |
1107 | /// Returns true if the instruction may be a suitable candidate for LICM. |
1108 | /// e.g. If the instruction is a call, then it's obviously not safe to hoist it. |
1109 | bool MachineLICMImpl::IsLICMCandidate(MachineInstr &I, MachineLoop *CurLoop) { |
1110 | // Check if it's safe to move the instruction. |
1111 | bool DontMoveAcrossStore = !HoistConstLoads || !AllowedToHoistLoads[CurLoop]; |
1112 | if ((!I.isSafeToMove(SawStore&: DontMoveAcrossStore)) && |
1113 | !(HoistConstStores && isInvariantStore(MI: I, TRI, MRI))) { |
1114 | LLVM_DEBUG(dbgs() << "LICM: Instruction not safe to move.\n" ); |
1115 | return false; |
1116 | } |
1117 | |
1118 | // If it is a load then check if it is guaranteed to execute by making sure |
1119 | // that it dominates all exiting blocks. If it doesn't, then there is a path |
1120 | // out of the loop which does not execute this load, so we can't hoist it. |
1121 | // Loads from constant memory are safe to speculate, for example indexed load |
1122 | // from a jump table. |
1123 | // Stores and side effects are already checked by isSafeToMove. |
1124 | if (I.mayLoad() && !mayLoadFromGOTOrConstantPool(MI&: I) && |
1125 | !IsGuaranteedToExecute(BB: I.getParent(), CurLoop)) { |
1126 | LLVM_DEBUG(dbgs() << "LICM: Load not guaranteed to execute.\n" ); |
1127 | return false; |
1128 | } |
1129 | |
1130 | // Convergent attribute has been used on operations that involve inter-thread |
1131 | // communication which results are implicitly affected by the enclosing |
1132 | // control flows. It is not safe to hoist or sink such operations across |
1133 | // control flow. |
1134 | if (I.isConvergent()) |
1135 | return false; |
1136 | |
1137 | if (!TII->shouldHoist(MI: I, FromLoop: CurLoop)) |
1138 | return false; |
1139 | |
1140 | return true; |
1141 | } |
1142 | |
1143 | /// Returns true if the instruction is loop invariant. |
1144 | bool MachineLICMImpl::IsLoopInvariantInst(MachineInstr &I, |
1145 | MachineLoop *CurLoop) { |
1146 | if (!IsLICMCandidate(I, CurLoop)) { |
1147 | LLVM_DEBUG(dbgs() << "LICM: Instruction not a LICM candidate\n" ); |
1148 | return false; |
1149 | } |
1150 | return CurLoop->isLoopInvariant(I); |
1151 | } |
1152 | |
1153 | /// Return true if the specified instruction is used by a phi node and hoisting |
1154 | /// it could cause a copy to be inserted. |
1155 | bool MachineLICMImpl::HasLoopPHIUse(const MachineInstr *MI, |
1156 | MachineLoop *CurLoop) { |
1157 | SmallVector<const MachineInstr *, 8> Work(1, MI); |
1158 | do { |
1159 | MI = Work.pop_back_val(); |
1160 | for (const MachineOperand &MO : MI->all_defs()) { |
1161 | Register Reg = MO.getReg(); |
1162 | if (!Reg.isVirtual()) |
1163 | continue; |
1164 | for (MachineInstr &UseMI : MRI->use_instructions(Reg)) { |
1165 | // A PHI may cause a copy to be inserted. |
1166 | if (UseMI.isPHI()) { |
1167 | // A PHI inside the loop causes a copy because the live range of Reg is |
1168 | // extended across the PHI. |
1169 | if (CurLoop->contains(Inst: &UseMI)) |
1170 | return true; |
1171 | // A PHI in an exit block can cause a copy to be inserted if the PHI |
1172 | // has multiple predecessors in the loop with different values. |
1173 | // For now, approximate by rejecting all exit blocks. |
1174 | if (isExitBlock(CurLoop, MBB: UseMI.getParent())) |
1175 | return true; |
1176 | continue; |
1177 | } |
1178 | // Look past copies as well. |
1179 | if (UseMI.isCopy() && CurLoop->contains(Inst: &UseMI)) |
1180 | Work.push_back(Elt: &UseMI); |
1181 | } |
1182 | } |
1183 | } while (!Work.empty()); |
1184 | return false; |
1185 | } |
1186 | |
1187 | /// Compute operand latency between a def of 'Reg' and an use in the current |
1188 | /// loop, return true if the target considered it high. |
1189 | bool MachineLICMImpl::HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx, |
1190 | Register Reg, |
1191 | MachineLoop *CurLoop) const { |
1192 | if (MRI->use_nodbg_empty(RegNo: Reg)) |
1193 | return false; |
1194 | |
1195 | for (MachineInstr &UseMI : MRI->use_nodbg_instructions(Reg)) { |
1196 | if (UseMI.isCopyLike()) |
1197 | continue; |
1198 | if (!CurLoop->contains(BB: UseMI.getParent())) |
1199 | continue; |
1200 | for (unsigned i = 0, e = UseMI.getNumOperands(); i != e; ++i) { |
1201 | const MachineOperand &MO = UseMI.getOperand(i); |
1202 | if (!MO.isReg() || !MO.isUse()) |
1203 | continue; |
1204 | Register MOReg = MO.getReg(); |
1205 | if (MOReg != Reg) |
1206 | continue; |
1207 | |
1208 | if (TII->hasHighOperandLatency(SchedModel, MRI, DefMI: MI, DefIdx, UseMI, UseIdx: i)) |
1209 | return true; |
1210 | } |
1211 | |
1212 | // Only look at the first in loop use. |
1213 | break; |
1214 | } |
1215 | |
1216 | return false; |
1217 | } |
1218 | |
1219 | /// Return true if the instruction is marked "cheap" or the operand latency |
1220 | /// between its def and a use is one or less. |
1221 | bool MachineLICMImpl::IsCheapInstruction(MachineInstr &MI) const { |
1222 | if (TII->isAsCheapAsAMove(MI) || MI.isCopyLike()) |
1223 | return true; |
1224 | |
1225 | bool isCheap = false; |
1226 | unsigned NumDefs = MI.getDesc().getNumDefs(); |
1227 | for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) { |
1228 | MachineOperand &DefMO = MI.getOperand(i); |
1229 | if (!DefMO.isReg() || !DefMO.isDef()) |
1230 | continue; |
1231 | --NumDefs; |
1232 | Register Reg = DefMO.getReg(); |
1233 | if (Reg.isPhysical()) |
1234 | continue; |
1235 | |
1236 | if (!TII->hasLowDefLatency(SchedModel, DefMI: MI, DefIdx: i)) |
1237 | return false; |
1238 | isCheap = true; |
1239 | } |
1240 | |
1241 | return isCheap; |
1242 | } |
1243 | |
1244 | /// Visit BBs from header to current BB, check if hoisting an instruction of the |
1245 | /// given cost matrix can cause high register pressure. |
1246 | bool MachineLICMImpl::CanCauseHighRegPressure( |
1247 | const SmallDenseMap<unsigned, int> &Cost, bool CheapInstr) { |
1248 | for (const auto &RPIdAndCost : Cost) { |
1249 | if (RPIdAndCost.second <= 0) |
1250 | continue; |
1251 | |
1252 | unsigned Class = RPIdAndCost.first; |
1253 | int Limit = RegLimit[Class]; |
1254 | |
1255 | // Don't hoist cheap instructions if they would increase register pressure, |
1256 | // even if we're under the limit. |
1257 | if (CheapInstr && !HoistCheapInsts) |
1258 | return true; |
1259 | |
1260 | for (const auto &RP : BackTrace) |
1261 | if (static_cast<int>(RP[Class]) + RPIdAndCost.second >= Limit) |
1262 | return true; |
1263 | } |
1264 | |
1265 | return false; |
1266 | } |
1267 | |
1268 | /// Traverse the back trace from header to the current block and update their |
1269 | /// register pressures to reflect the effect of hoisting MI from the current |
1270 | /// block to the preheader. |
1271 | void MachineLICMImpl::UpdateBackTraceRegPressure(const MachineInstr *MI) { |
1272 | // First compute the 'cost' of the instruction, i.e. its contribution |
1273 | // to register pressure. |
1274 | auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/false, |
1275 | /*ConsiderUnseenAsDef=*/false); |
1276 | |
1277 | // Update register pressure of blocks from loop header to current block. |
1278 | for (auto &RP : BackTrace) |
1279 | for (const auto &RPIdAndCost : Cost) |
1280 | RP[RPIdAndCost.first] += RPIdAndCost.second; |
1281 | } |
1282 | |
1283 | /// Return true if it is potentially profitable to hoist the given loop |
1284 | /// invariant. |
1285 | bool MachineLICMImpl::IsProfitableToHoist(MachineInstr &MI, |
1286 | MachineLoop *CurLoop) { |
1287 | if (MI.isImplicitDef()) |
1288 | return true; |
1289 | |
1290 | // Besides removing computation from the loop, hoisting an instruction has |
1291 | // these effects: |
1292 | // |
1293 | // - The value defined by the instruction becomes live across the entire |
1294 | // loop. This increases register pressure in the loop. |
1295 | // |
1296 | // - If the value is used by a PHI in the loop, a copy will be required for |
1297 | // lowering the PHI after extending the live range. |
1298 | // |
1299 | // - When hoisting the last use of a value in the loop, that value no longer |
1300 | // needs to be live in the loop. This lowers register pressure in the loop. |
1301 | |
1302 | if (HoistConstStores && isCopyFeedingInvariantStore(MI, MRI, TRI)) |
1303 | return true; |
1304 | |
1305 | bool CheapInstr = IsCheapInstruction(MI); |
1306 | bool CreatesCopy = HasLoopPHIUse(MI: &MI, CurLoop); |
1307 | |
1308 | // Don't hoist a cheap instruction if it would create a copy in the loop. |
1309 | if (CheapInstr && CreatesCopy) { |
1310 | LLVM_DEBUG(dbgs() << "Won't hoist cheap instr with loop PHI use: " << MI); |
1311 | return false; |
1312 | } |
1313 | |
1314 | // Rematerializable instructions should always be hoisted providing the |
1315 | // register allocator can just pull them down again when needed. |
1316 | if (isTriviallyReMaterializable(MI)) |
1317 | return true; |
1318 | |
1319 | // FIXME: If there are long latency loop-invariant instructions inside the |
1320 | // loop at this point, why didn't the optimizer's LICM hoist them? |
1321 | for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) { |
1322 | const MachineOperand &MO = MI.getOperand(i); |
1323 | if (!MO.isReg() || MO.isImplicit()) |
1324 | continue; |
1325 | Register Reg = MO.getReg(); |
1326 | if (!Reg.isVirtual()) |
1327 | continue; |
1328 | if (MO.isDef() && HasHighOperandLatency(MI, DefIdx: i, Reg, CurLoop)) { |
1329 | LLVM_DEBUG(dbgs() << "Hoist High Latency: " << MI); |
1330 | ++NumHighLatency; |
1331 | return true; |
1332 | } |
1333 | } |
1334 | |
1335 | // Estimate register pressure to determine whether to LICM the instruction. |
1336 | // In low register pressure situation, we can be more aggressive about |
1337 | // hoisting. Also, favors hoisting long latency instructions even in |
1338 | // moderately high pressure situation. |
1339 | // Cheap instructions will only be hoisted if they don't increase register |
1340 | // pressure at all. |
1341 | auto Cost = calcRegisterCost(MI: &MI, /*ConsiderSeen=*/false, |
1342 | /*ConsiderUnseenAsDef=*/false); |
1343 | |
1344 | // Visit BBs from header to current BB, if hoisting this doesn't cause |
1345 | // high register pressure, then it's safe to proceed. |
1346 | if (!CanCauseHighRegPressure(Cost, CheapInstr)) { |
1347 | LLVM_DEBUG(dbgs() << "Hoist non-reg-pressure: " << MI); |
1348 | ++NumLowRP; |
1349 | return true; |
1350 | } |
1351 | |
1352 | // Don't risk increasing register pressure if it would create copies. |
1353 | if (CreatesCopy) { |
1354 | LLVM_DEBUG(dbgs() << "Won't hoist instr with loop PHI use: " << MI); |
1355 | return false; |
1356 | } |
1357 | |
1358 | // Do not "speculate" in high register pressure situation. If an |
1359 | // instruction is not guaranteed to be executed in the loop, it's best to be |
1360 | // conservative. |
1361 | if (AvoidSpeculation && |
1362 | (!IsGuaranteedToExecute(BB: MI.getParent(), CurLoop) && !MayCSE(MI: &MI))) { |
1363 | LLVM_DEBUG(dbgs() << "Won't speculate: " << MI); |
1364 | return false; |
1365 | } |
1366 | |
1367 | // If we have a COPY with other uses in the loop, hoist to allow the users to |
1368 | // also be hoisted. |
1369 | // TODO: Handle all isCopyLike? |
1370 | if (MI.isCopy() || MI.isRegSequence()) { |
1371 | Register DefReg = MI.getOperand(i: 0).getReg(); |
1372 | if (DefReg.isVirtual() && |
1373 | all_of(Range: MI.uses(), |
1374 | P: [this](const MachineOperand &UseOp) { |
1375 | return !UseOp.isReg() || UseOp.getReg().isVirtual() || |
1376 | MRI->isConstantPhysReg(PhysReg: UseOp.getReg()); |
1377 | }) && |
1378 | IsLoopInvariantInst(I&: MI, CurLoop) && |
1379 | any_of(Range: MRI->use_nodbg_instructions(Reg: DefReg), |
1380 | P: [&CurLoop, this, DefReg, |
1381 | Cost = std::move(Cost)](MachineInstr &UseMI) { |
1382 | if (!CurLoop->contains(Inst: &UseMI)) |
1383 | return false; |
1384 | |
1385 | // COPY is a cheap instruction, but if moving it won't cause |
1386 | // high RP we're fine to hoist it even if the user can't be |
1387 | // hoisted later Otherwise we want to check the user if it's |
1388 | // hoistable |
1389 | if (CanCauseHighRegPressure(Cost, CheapInstr: false) && |
1390 | !CurLoop->isLoopInvariant(I&: UseMI, ExcludeReg: DefReg)) |
1391 | return false; |
1392 | |
1393 | return true; |
1394 | })) |
1395 | return true; |
1396 | } |
1397 | |
1398 | // High register pressure situation, only hoist if the instruction is going |
1399 | // to be remat'ed. |
1400 | if (!isTriviallyReMaterializable(MI) && |
1401 | !MI.isDereferenceableInvariantLoad()) { |
1402 | LLVM_DEBUG(dbgs() << "Can't remat / high reg-pressure: " << MI); |
1403 | return false; |
1404 | } |
1405 | |
1406 | return true; |
1407 | } |
1408 | |
1409 | /// Unfold a load from the given machineinstr if the load itself could be |
1410 | /// hoisted. Return the unfolded and hoistable load, or null if the load |
1411 | /// couldn't be unfolded or if it wouldn't be hoistable. |
1412 | MachineInstr *MachineLICMImpl::(MachineInstr *MI, |
1413 | MachineLoop *CurLoop) { |
1414 | // Don't unfold simple loads. |
1415 | if (MI->canFoldAsLoad()) |
1416 | return nullptr; |
1417 | |
1418 | // If not, we may be able to unfold a load and hoist that. |
1419 | // First test whether the instruction is loading from an amenable |
1420 | // memory location. |
1421 | if (!MI->isDereferenceableInvariantLoad()) |
1422 | return nullptr; |
1423 | |
1424 | // Next determine the register class for a temporary register. |
1425 | unsigned LoadRegIndex; |
1426 | unsigned NewOpc = |
1427 | TII->getOpcodeAfterMemoryUnfold(Opc: MI->getOpcode(), |
1428 | /*UnfoldLoad=*/true, |
1429 | /*UnfoldStore=*/false, |
1430 | LoadRegIndex: &LoadRegIndex); |
1431 | if (NewOpc == 0) return nullptr; |
1432 | const MCInstrDesc &MID = TII->get(Opcode: NewOpc); |
1433 | MachineFunction &MF = *MI->getMF(); |
1434 | const TargetRegisterClass *RC = TII->getRegClass(MCID: MID, OpNum: LoadRegIndex, TRI, MF); |
1435 | // Ok, we're unfolding. Create a temporary register and do the unfold. |
1436 | Register Reg = MRI->createVirtualRegister(RegClass: RC); |
1437 | |
1438 | SmallVector<MachineInstr *, 2> NewMIs; |
1439 | bool Success = TII->unfoldMemoryOperand(MF, MI&: *MI, Reg, |
1440 | /*UnfoldLoad=*/true, |
1441 | /*UnfoldStore=*/false, NewMIs); |
1442 | (void)Success; |
1443 | assert(Success && |
1444 | "unfoldMemoryOperand failed when getOpcodeAfterMemoryUnfold " |
1445 | "succeeded!" ); |
1446 | assert(NewMIs.size() == 2 && |
1447 | "Unfolded a load into multiple instructions!" ); |
1448 | MachineBasicBlock *MBB = MI->getParent(); |
1449 | MachineBasicBlock::iterator Pos = MI; |
1450 | MBB->insert(I: Pos, MI: NewMIs[0]); |
1451 | MBB->insert(I: Pos, MI: NewMIs[1]); |
1452 | // If unfolding produced a load that wasn't loop-invariant or profitable to |
1453 | // hoist, discard the new instructions and bail. |
1454 | if (!IsLoopInvariantInst(I&: *NewMIs[0], CurLoop) || |
1455 | !IsProfitableToHoist(MI&: *NewMIs[0], CurLoop)) { |
1456 | NewMIs[0]->eraseFromParent(); |
1457 | NewMIs[1]->eraseFromParent(); |
1458 | return nullptr; |
1459 | } |
1460 | |
1461 | // Update register pressure for the unfolded instruction. |
1462 | UpdateRegPressure(MI: NewMIs[1]); |
1463 | |
1464 | // Otherwise we successfully unfolded a load that we can hoist. |
1465 | |
1466 | // Update the call info. |
1467 | if (MI->shouldUpdateAdditionalCallInfo()) |
1468 | MF.eraseAdditionalCallInfo(MI); |
1469 | |
1470 | MI->eraseFromParent(); |
1471 | return NewMIs[0]; |
1472 | } |
1473 | |
1474 | /// Initialize the CSE map with instructions that are in the current loop |
1475 | /// preheader that may become duplicates of instructions that are hoisted |
1476 | /// out of the loop. |
1477 | void MachineLICMImpl::InitCSEMap(MachineBasicBlock *BB) { |
1478 | for (MachineInstr &MI : *BB) |
1479 | CSEMap[BB][MI.getOpcode()].push_back(x: &MI); |
1480 | } |
1481 | |
1482 | /// Initialize AllowedToHoistLoads with information about whether invariant |
1483 | /// loads can be moved outside a given loop |
1484 | void MachineLICMImpl::InitializeLoadsHoistableLoops() { |
1485 | SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end()); |
1486 | SmallVector<MachineLoop *, 8> LoopsInPreOrder; |
1487 | |
1488 | // Mark all loops as hoistable initially and prepare a list of loops in |
1489 | // pre-order DFS. |
1490 | while (!Worklist.empty()) { |
1491 | auto *L = Worklist.pop_back_val(); |
1492 | AllowedToHoistLoads[L] = true; |
1493 | LoopsInPreOrder.push_back(Elt: L); |
1494 | llvm::append_range(C&: Worklist, R: L->getSubLoops()); |
1495 | } |
1496 | |
1497 | // Going from the innermost to outermost loops, check if a loop has |
1498 | // instructions preventing invariant load hoisting. If such instruction is |
1499 | // found, mark this loop and its parent as non-hoistable and continue |
1500 | // investigating the next loop. |
1501 | // Visiting in a reversed pre-ordered DFS manner |
1502 | // allows us to not process all the instructions of the outer loop if the |
1503 | // inner loop is proved to be non-load-hoistable. |
1504 | for (auto *Loop : reverse(C&: LoopsInPreOrder)) { |
1505 | for (auto *MBB : Loop->blocks()) { |
1506 | // If this loop has already been marked as non-hoistable, skip it. |
1507 | if (!AllowedToHoistLoads[Loop]) |
1508 | continue; |
1509 | for (auto &MI : *MBB) { |
1510 | if (!MI.isLoadFoldBarrier() && !MI.mayStore() && !MI.isCall() && |
1511 | !(MI.mayLoad() && MI.hasOrderedMemoryRef())) |
1512 | continue; |
1513 | for (MachineLoop *L = Loop; L != nullptr; L = L->getParentLoop()) |
1514 | AllowedToHoistLoads[L] = false; |
1515 | break; |
1516 | } |
1517 | } |
1518 | } |
1519 | } |
1520 | |
1521 | /// Find an instruction amount PrevMIs that is a duplicate of MI. |
1522 | /// Return this instruction if it's found. |
1523 | MachineInstr * |
1524 | MachineLICMImpl::LookForDuplicate(const MachineInstr *MI, |
1525 | std::vector<MachineInstr *> &PrevMIs) { |
1526 | for (MachineInstr *PrevMI : PrevMIs) |
1527 | if (TII->produceSameValue(MI0: *MI, MI1: *PrevMI, MRI: (PreRegAlloc ? MRI : nullptr))) |
1528 | return PrevMI; |
1529 | |
1530 | return nullptr; |
1531 | } |
1532 | |
1533 | /// Given a LICM'ed instruction, look for an instruction on the preheader that |
1534 | /// computes the same value. If it's found, do a RAU on with the definition of |
1535 | /// the existing instruction rather than hoisting the instruction to the |
1536 | /// preheader. |
1537 | bool MachineLICMImpl::EliminateCSE( |
1538 | MachineInstr *MI, |
1539 | DenseMap<unsigned, std::vector<MachineInstr *>>::iterator &CI) { |
1540 | // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate |
1541 | // the undef property onto uses. |
1542 | if (MI->isImplicitDef()) |
1543 | return false; |
1544 | |
1545 | // Do not CSE normal loads because between them could be store instructions |
1546 | // that change the loaded value |
1547 | if (MI->mayLoad() && !MI->isDereferenceableInvariantLoad()) |
1548 | return false; |
1549 | |
1550 | if (MachineInstr *Dup = LookForDuplicate(MI, PrevMIs&: CI->second)) { |
1551 | LLVM_DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup); |
1552 | |
1553 | // Replace virtual registers defined by MI by their counterparts defined |
1554 | // by Dup. |
1555 | SmallVector<unsigned, 2> Defs; |
1556 | for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { |
1557 | const MachineOperand &MO = MI->getOperand(i); |
1558 | |
1559 | // Physical registers may not differ here. |
1560 | assert((!MO.isReg() || MO.getReg() == 0 || !MO.getReg().isPhysical() || |
1561 | MO.getReg() == Dup->getOperand(i).getReg()) && |
1562 | "Instructions with different phys regs are not identical!" ); |
1563 | |
1564 | if (MO.isReg() && MO.isDef() && !MO.getReg().isPhysical()) |
1565 | Defs.push_back(Elt: i); |
1566 | } |
1567 | |
1568 | SmallVector<const TargetRegisterClass*, 2> OrigRCs; |
1569 | for (unsigned i = 0, e = Defs.size(); i != e; ++i) { |
1570 | unsigned Idx = Defs[i]; |
1571 | Register Reg = MI->getOperand(i: Idx).getReg(); |
1572 | Register DupReg = Dup->getOperand(i: Idx).getReg(); |
1573 | OrigRCs.push_back(Elt: MRI->getRegClass(Reg: DupReg)); |
1574 | |
1575 | if (!MRI->constrainRegClass(Reg: DupReg, RC: MRI->getRegClass(Reg))) { |
1576 | // Restore old RCs if more than one defs. |
1577 | for (unsigned j = 0; j != i; ++j) |
1578 | MRI->setRegClass(Reg: Dup->getOperand(i: Defs[j]).getReg(), RC: OrigRCs[j]); |
1579 | return false; |
1580 | } |
1581 | } |
1582 | |
1583 | for (unsigned Idx : Defs) { |
1584 | Register Reg = MI->getOperand(i: Idx).getReg(); |
1585 | Register DupReg = Dup->getOperand(i: Idx).getReg(); |
1586 | MRI->replaceRegWith(FromReg: Reg, ToReg: DupReg); |
1587 | MRI->clearKillFlags(Reg: DupReg); |
1588 | // Clear Dup dead flag if any, we reuse it for Reg. |
1589 | if (!MRI->use_nodbg_empty(RegNo: DupReg)) |
1590 | Dup->getOperand(i: Idx).setIsDead(false); |
1591 | } |
1592 | |
1593 | MI->eraseFromParent(); |
1594 | ++NumCSEed; |
1595 | return true; |
1596 | } |
1597 | return false; |
1598 | } |
1599 | |
1600 | /// Return true if the given instruction will be CSE'd if it's hoisted out of |
1601 | /// the loop. |
1602 | bool MachineLICMImpl::MayCSE(MachineInstr *MI) { |
1603 | if (MI->mayLoad() && !MI->isDereferenceableInvariantLoad()) |
1604 | return false; |
1605 | |
1606 | unsigned Opcode = MI->getOpcode(); |
1607 | for (auto &Map : CSEMap) { |
1608 | // Check this CSEMap's preheader dominates MI's basic block. |
1609 | if (MDTU->getDomTree().dominates(A: Map.first, B: MI->getParent())) { |
1610 | DenseMap<unsigned, std::vector<MachineInstr *>>::iterator CI = |
1611 | Map.second.find(Val: Opcode); |
1612 | // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate |
1613 | // the undef property onto uses. |
1614 | if (CI == Map.second.end() || MI->isImplicitDef()) |
1615 | continue; |
1616 | if (LookForDuplicate(MI, PrevMIs&: CI->second) != nullptr) |
1617 | return true; |
1618 | } |
1619 | } |
1620 | |
1621 | return false; |
1622 | } |
1623 | |
1624 | /// When an instruction is found to use only loop invariant operands |
1625 | /// that are safe to hoist, this instruction is called to do the dirty work. |
1626 | /// It returns true if the instruction is hoisted. |
1627 | unsigned MachineLICMImpl::Hoist(MachineInstr *MI, MachineBasicBlock *, |
1628 | MachineLoop *CurLoop) { |
1629 | MachineBasicBlock *SrcBlock = MI->getParent(); |
1630 | |
1631 | // Disable the instruction hoisting due to block hotness |
1632 | if ((DisableHoistingToHotterBlocks == UseBFI::All || |
1633 | (DisableHoistingToHotterBlocks == UseBFI::PGO && HasProfileData)) && |
1634 | isTgtHotterThanSrc(SrcBlock, TgtBlock: Preheader)) { |
1635 | ++NumNotHoistedDueToHotness; |
1636 | return HoistResult::NotHoisted; |
1637 | } |
1638 | // First check whether we should hoist this instruction. |
1639 | bool = false; |
1640 | if (!IsLoopInvariantInst(I&: *MI, CurLoop) || |
1641 | !IsProfitableToHoist(MI&: *MI, CurLoop)) { |
1642 | // If not, try unfolding a hoistable load. |
1643 | MI = ExtractHoistableLoad(MI, CurLoop); |
1644 | if (!MI) |
1645 | return HoistResult::NotHoisted; |
1646 | HasExtractHoistableLoad = true; |
1647 | } |
1648 | |
1649 | // If we have hoisted an instruction that may store, it can only be a constant |
1650 | // store. |
1651 | if (MI->mayStore()) |
1652 | NumStoreConst++; |
1653 | |
1654 | // Now move the instructions to the predecessor, inserting it before any |
1655 | // terminator instructions. |
1656 | LLVM_DEBUG({ |
1657 | dbgs() << "Hoisting " << *MI; |
1658 | if (MI->getParent()->getBasicBlock()) |
1659 | dbgs() << " from " << printMBBReference(*MI->getParent()); |
1660 | if (Preheader->getBasicBlock()) |
1661 | dbgs() << " to " << printMBBReference(*Preheader); |
1662 | dbgs() << "\n" ; |
1663 | }); |
1664 | |
1665 | // If this is the first instruction being hoisted to the preheader, |
1666 | // initialize the CSE map with potential common expressions. |
1667 | if (FirstInLoop) { |
1668 | InitCSEMap(BB: Preheader); |
1669 | FirstInLoop = false; |
1670 | } |
1671 | |
1672 | // Look for opportunity to CSE the hoisted instruction. |
1673 | unsigned Opcode = MI->getOpcode(); |
1674 | bool HasCSEDone = false; |
1675 | for (auto &Map : CSEMap) { |
1676 | // Check this CSEMap's preheader dominates MI's basic block. |
1677 | if (MDTU->getDomTree().dominates(A: Map.first, B: MI->getParent())) { |
1678 | DenseMap<unsigned, std::vector<MachineInstr *>>::iterator CI = |
1679 | Map.second.find(Val: Opcode); |
1680 | if (CI != Map.second.end()) { |
1681 | if (EliminateCSE(MI, CI)) { |
1682 | HasCSEDone = true; |
1683 | break; |
1684 | } |
1685 | } |
1686 | } |
1687 | } |
1688 | |
1689 | if (!HasCSEDone) { |
1690 | // Otherwise, splice the instruction to the preheader. |
1691 | Preheader->splice(Where: Preheader->getFirstTerminator(),Other: MI->getParent(),From: MI); |
1692 | |
1693 | // Since we are moving the instruction out of its basic block, we do not |
1694 | // retain its debug location. Doing so would degrade the debugging |
1695 | // experience and adversely affect the accuracy of profiling information. |
1696 | assert(!MI->isDebugInstr() && "Should not hoist debug inst" ); |
1697 | MI->setDebugLoc(DebugLoc()); |
1698 | |
1699 | // Update register pressure for BBs from header to this block. |
1700 | UpdateBackTraceRegPressure(MI); |
1701 | |
1702 | // Clear the kill flags of any register this instruction defines, |
1703 | // since they may need to be live throughout the entire loop |
1704 | // rather than just live for part of it. |
1705 | for (MachineOperand &MO : MI->all_defs()) |
1706 | if (!MO.isDead()) |
1707 | MRI->clearKillFlags(Reg: MO.getReg()); |
1708 | |
1709 | CSEMap[Preheader][Opcode].push_back(x: MI); |
1710 | } |
1711 | |
1712 | ++NumHoisted; |
1713 | Changed = true; |
1714 | |
1715 | if (HasCSEDone || HasExtractHoistableLoad) |
1716 | return HoistResult::Hoisted | HoistResult::ErasedMI; |
1717 | return HoistResult::Hoisted; |
1718 | } |
1719 | |
1720 | /// Get the preheader for the current loop, splitting a critical edge if needed. |
1721 | MachineBasicBlock *MachineLICMImpl::(MachineLoop *CurLoop) { |
1722 | // Determine the block to which to hoist instructions. If we can't find a |
1723 | // suitable loop predecessor, we can't do any hoisting. |
1724 | if (MachineBasicBlock * = CurLoop->getLoopPreheader()) |
1725 | return Preheader; |
1726 | |
1727 | // Try forming a preheader by splitting the critical edge between the single |
1728 | // predecessor and the loop header. |
1729 | if (MachineBasicBlock *Pred = CurLoop->getLoopPredecessor()) { |
1730 | MachineBasicBlock * = Pred->SplitCriticalEdge( |
1731 | Succ: CurLoop->getHeader(), P: LegacyPass, MFAM, LiveInSets: nullptr, MDTU); |
1732 | if (NewPreheader) |
1733 | Changed = true; |
1734 | return NewPreheader; |
1735 | } |
1736 | |
1737 | return nullptr; |
1738 | } |
1739 | |
1740 | /// Is the target basic block at least "BlockFrequencyRatioThreshold" |
1741 | /// times hotter than the source basic block. |
1742 | bool MachineLICMImpl::isTgtHotterThanSrc(MachineBasicBlock *SrcBlock, |
1743 | MachineBasicBlock *TgtBlock) { |
1744 | // Parse source and target basic block frequency from MBFI |
1745 | uint64_t SrcBF = MBFI->getBlockFreq(MBB: SrcBlock).getFrequency(); |
1746 | uint64_t DstBF = MBFI->getBlockFreq(MBB: TgtBlock).getFrequency(); |
1747 | |
1748 | // Disable the hoisting if source block frequency is zero |
1749 | if (!SrcBF) |
1750 | return true; |
1751 | |
1752 | double Ratio = (double)DstBF / SrcBF; |
1753 | |
1754 | // Compare the block frequency ratio with the threshold |
1755 | return Ratio > BlockFrequencyRatioThreshold; |
1756 | } |
1757 | |
1758 | template <typename DerivedT, bool PreRegAlloc> |
1759 | PreservedAnalyses MachineLICMBasePass<DerivedT, PreRegAlloc>::run( |
1760 | MachineFunction &MF, MachineFunctionAnalysisManager &MFAM) { |
1761 | bool Changed = MachineLICMImpl(PreRegAlloc, nullptr, &MFAM).run(MF); |
1762 | if (!Changed) |
1763 | return PreservedAnalyses::all(); |
1764 | auto PA = getMachineFunctionPassPreservedAnalyses(); |
1765 | PA.preserve<MachineLoopAnalysis>(); |
1766 | return PA; |
1767 | } |
1768 | |
1769 | template class llvm::MachineLICMBasePass<EarlyMachineLICMPass, true>; |
1770 | template class llvm::MachineLICMBasePass<MachineLICMPass, false>; |
1771 | |