1 //===- GVNHoist.cpp - Hoist scalar and load expressions -------------------===//
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 hoists expressions from branches to a common dominator. It uses
10 // GVN (global value numbering) to discover expressions computing the same
11 // values. The primary goals of code-hoisting are:
12 // 1. To reduce the code size.
13 // 2. In some cases reduce critical path (by exposing more ILP).
14 //
15 // The algorithm factors out the reachability of values such that multiple
16 // queries to find reachability of values are fast. This is based on finding the
17 // ANTIC points in the CFG which do not change during hoisting. The ANTIC points
18 // are basically the dominance-frontiers in the inverse graph. So we introduce a
19 // data structure (CHI nodes) to keep track of values flowing out of a basic
20 // block. We only do this for values with multiple occurrences in the function
21 // as they are the potential hoistable candidates. This approach allows us to
22 // hoist instructions to a basic block with more than two successors, as well as
23 // deal with infinite loops in a trivial way.
24 //
25 // Limitations: This pass does not hoist fully redundant expressions because
26 // they are already handled by GVN-PRE. It is advisable to run gvn-hoist before
27 // and after gvn-pre because gvn-pre creates opportunities for more instructions
28 // to be hoisted.
29 //
30 // Hoisting may affect the performance in some cases. To mitigate that, hoisting
31 // is disabled in the following cases.
32 // 1. Scalars across calls.
33 // 2. geps when corresponding load/store cannot be hoisted.
34 //===----------------------------------------------------------------------===//
35
36 #include "llvm/ADT/DenseMap.h"
37 #include "llvm/ADT/DenseSet.h"
38 #include "llvm/ADT/STLExtras.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/SmallVector.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/ADT/iterator_range.h"
43 #include "llvm/Analysis/AliasAnalysis.h"
44 #include "llvm/Analysis/GlobalsModRef.h"
45 #include "llvm/Analysis/IteratedDominanceFrontier.h"
46 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
47 #include "llvm/Analysis/MemorySSA.h"
48 #include "llvm/Analysis/MemorySSAUpdater.h"
49 #include "llvm/Analysis/PostDominators.h"
50 #include "llvm/Analysis/ValueTracking.h"
51 #include "llvm/IR/Argument.h"
52 #include "llvm/IR/BasicBlock.h"
53 #include "llvm/IR/CFG.h"
54 #include "llvm/IR/Constants.h"
55 #include "llvm/IR/Dominators.h"
56 #include "llvm/IR/Function.h"
57 #include "llvm/IR/Instruction.h"
58 #include "llvm/IR/Instructions.h"
59 #include "llvm/IR/IntrinsicInst.h"
60 #include "llvm/IR/LLVMContext.h"
61 #include "llvm/IR/PassManager.h"
62 #include "llvm/IR/Use.h"
63 #include "llvm/IR/User.h"
64 #include "llvm/IR/Value.h"
65 #include "llvm/InitializePasses.h"
66 #include "llvm/Pass.h"
67 #include "llvm/Support/Casting.h"
68 #include "llvm/Support/CommandLine.h"
69 #include "llvm/Support/Debug.h"
70 #include "llvm/Support/raw_ostream.h"
71 #include "llvm/Transforms/Scalar.h"
72 #include "llvm/Transforms/Scalar/GVN.h"
73 #include "llvm/Transforms/Utils/Local.h"
74 #include <algorithm>
75 #include <cassert>
76 #include <iterator>
77 #include <memory>
78 #include <utility>
79 #include <vector>
80
81 using namespace llvm;
82
83 #define DEBUG_TYPE "gvn-hoist"
84
85 STATISTIC(NumHoisted, "Number of instructions hoisted");
86 STATISTIC(NumRemoved, "Number of instructions removed");
87 STATISTIC(NumLoadsHoisted, "Number of loads hoisted");
88 STATISTIC(NumLoadsRemoved, "Number of loads removed");
89 STATISTIC(NumStoresHoisted, "Number of stores hoisted");
90 STATISTIC(NumStoresRemoved, "Number of stores removed");
91 STATISTIC(NumCallsHoisted, "Number of calls hoisted");
92 STATISTIC(NumCallsRemoved, "Number of calls removed");
93
94 static cl::opt<int>
95 MaxHoistedThreshold("gvn-max-hoisted", cl::Hidden, cl::init(-1),
96 cl::desc("Max number of instructions to hoist "
97 "(default unlimited = -1)"));
98
99 static cl::opt<int> MaxNumberOfBBSInPath(
100 "gvn-hoist-max-bbs", cl::Hidden, cl::init(4),
101 cl::desc("Max number of basic blocks on the path between "
102 "hoisting locations (default = 4, unlimited = -1)"));
103
104 static cl::opt<int> MaxDepthInBB(
105 "gvn-hoist-max-depth", cl::Hidden, cl::init(100),
106 cl::desc("Hoist instructions from the beginning of the BB up to the "
107 "maximum specified depth (default = 100, unlimited = -1)"));
108
109 static cl::opt<int>
110 MaxChainLength("gvn-hoist-max-chain-length", cl::Hidden, cl::init(10),
111 cl::desc("Maximum length of dependent chains to hoist "
112 "(default = 10, unlimited = -1)"));
113
114 namespace llvm {
115
116 using BBSideEffectsSet = DenseMap<const BasicBlock *, bool>;
117 using SmallVecInsn = SmallVector<Instruction *, 4>;
118 using SmallVecImplInsn = SmallVectorImpl<Instruction *>;
119
120 // Each element of a hoisting list contains the basic block where to hoist and
121 // a list of instructions to be hoisted.
122 using HoistingPointInfo = std::pair<BasicBlock *, SmallVecInsn>;
123
124 using HoistingPointList = SmallVector<HoistingPointInfo, 4>;
125
126 // A map from a pair of VNs to all the instructions with those VNs.
127 using VNType = std::pair<unsigned, uintptr_t>;
128
129 using VNtoInsns = DenseMap<VNType, SmallVector<Instruction *, 4>>;
130
131 // CHI keeps information about values flowing out of a basic block. It is
132 // similar to PHI but in the inverse graph, and used for outgoing values on each
133 // edge. For conciseness, it is computed only for instructions with multiple
134 // occurrences in the CFG because they are the only hoistable candidates.
135 // A (CHI[{V, B, I1}, {V, C, I2}]
136 // / \
137 // / \
138 // B(I1) C (I2)
139 // The Value number for both I1 and I2 is V, the CHI node will save the
140 // instruction as well as the edge where the value is flowing to.
141 struct CHIArg {
142 VNType VN;
143
144 // Edge destination (shows the direction of flow), may not be where the I is.
145 BasicBlock *Dest;
146
147 // The instruction (VN) which uses the values flowing out of CHI.
148 Instruction *I;
149
operator ==llvm::CHIArg150 bool operator==(const CHIArg &A) const { return VN == A.VN; }
operator !=llvm::CHIArg151 bool operator!=(const CHIArg &A) const { return !(*this == A); }
152 };
153
154 using CHIIt = SmallVectorImpl<CHIArg>::iterator;
155 using CHIArgs = iterator_range<CHIIt>;
156 using OutValuesType = DenseMap<BasicBlock *, SmallVector<CHIArg, 2>>;
157 using InValuesType =
158 DenseMap<BasicBlock *, SmallVector<std::pair<VNType, Instruction *>, 2>>;
159
160 // An invalid value number Used when inserting a single value number into
161 // VNtoInsns.
162 enum : uintptr_t { InvalidVN = ~(uintptr_t)2 };
163
164 // Records all scalar instructions candidate for code hoisting.
165 class InsnInfo {
166 VNtoInsns VNtoScalars;
167
168 public:
169 // Inserts I and its value number in VNtoScalars.
insert(Instruction * I,GVNPass::ValueTable & VN)170 void insert(Instruction *I, GVNPass::ValueTable &VN) {
171 // Scalar instruction.
172 unsigned V = VN.lookupOrAdd(I);
173 VNtoScalars[{V, InvalidVN}].push_back(I);
174 }
175
getVNTable() const176 const VNtoInsns &getVNTable() const { return VNtoScalars; }
177 };
178
179 // Records all load instructions candidate for code hoisting.
180 class LoadInfo {
181 VNtoInsns VNtoLoads;
182
183 public:
184 // Insert Load and the value number of its memory address in VNtoLoads.
insert(LoadInst * Load,GVNPass::ValueTable & VN)185 void insert(LoadInst *Load, GVNPass::ValueTable &VN) {
186 if (Load->isSimple()) {
187 unsigned V = VN.lookupOrAdd(Load->getPointerOperand());
188 // With opaque pointers we may have loads from the same pointer with
189 // different result types, which should be disambiguated.
190 VNtoLoads[{V, (uintptr_t)Load->getType()}].push_back(Load);
191 }
192 }
193
getVNTable() const194 const VNtoInsns &getVNTable() const { return VNtoLoads; }
195 };
196
197 // Records all store instructions candidate for code hoisting.
198 class StoreInfo {
199 VNtoInsns VNtoStores;
200
201 public:
202 // Insert the Store and a hash number of the store address and the stored
203 // value in VNtoStores.
insert(StoreInst * Store,GVNPass::ValueTable & VN)204 void insert(StoreInst *Store, GVNPass::ValueTable &VN) {
205 if (!Store->isSimple())
206 return;
207 // Hash the store address and the stored value.
208 Value *Ptr = Store->getPointerOperand();
209 Value *Val = Store->getValueOperand();
210 VNtoStores[{VN.lookupOrAdd(Ptr), VN.lookupOrAdd(Val)}].push_back(Store);
211 }
212
getVNTable() const213 const VNtoInsns &getVNTable() const { return VNtoStores; }
214 };
215
216 // Records all call instructions candidate for code hoisting.
217 class CallInfo {
218 VNtoInsns VNtoCallsScalars;
219 VNtoInsns VNtoCallsLoads;
220 VNtoInsns VNtoCallsStores;
221
222 public:
223 // Insert Call and its value numbering in one of the VNtoCalls* containers.
insert(CallInst * Call,GVNPass::ValueTable & VN)224 void insert(CallInst *Call, GVNPass::ValueTable &VN) {
225 // A call that doesNotAccessMemory is handled as a Scalar,
226 // onlyReadsMemory will be handled as a Load instruction,
227 // all other calls will be handled as stores.
228 unsigned V = VN.lookupOrAdd(Call);
229 auto Entry = std::make_pair(V, InvalidVN);
230
231 if (Call->doesNotAccessMemory())
232 VNtoCallsScalars[Entry].push_back(Call);
233 else if (Call->onlyReadsMemory())
234 VNtoCallsLoads[Entry].push_back(Call);
235 else
236 VNtoCallsStores[Entry].push_back(Call);
237 }
238
getScalarVNTable() const239 const VNtoInsns &getScalarVNTable() const { return VNtoCallsScalars; }
getLoadVNTable() const240 const VNtoInsns &getLoadVNTable() const { return VNtoCallsLoads; }
getStoreVNTable() const241 const VNtoInsns &getStoreVNTable() const { return VNtoCallsStores; }
242 };
243
combineKnownMetadata(Instruction * ReplInst,Instruction * I)244 static void combineKnownMetadata(Instruction *ReplInst, Instruction *I) {
245 static const unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
246 LLVMContext::MD_alias_scope,
247 LLVMContext::MD_noalias,
248 LLVMContext::MD_range,
249 LLVMContext::MD_fpmath,
250 LLVMContext::MD_invariant_load,
251 LLVMContext::MD_invariant_group,
252 LLVMContext::MD_access_group};
253 combineMetadata(ReplInst, I, KnownIDs, true);
254 }
255
256 // This pass hoists common computations across branches sharing common
257 // dominator. The primary goal is to reduce the code size, and in some
258 // cases reduce critical path (by exposing more ILP).
259 class GVNHoist {
260 public:
GVNHoist(DominatorTree * DT,PostDominatorTree * PDT,AliasAnalysis * AA,MemoryDependenceResults * MD,MemorySSA * MSSA)261 GVNHoist(DominatorTree *DT, PostDominatorTree *PDT, AliasAnalysis *AA,
262 MemoryDependenceResults *MD, MemorySSA *MSSA)
263 : DT(DT), PDT(PDT), AA(AA), MD(MD), MSSA(MSSA),
264 MSSAUpdater(std::make_unique<MemorySSAUpdater>(MSSA)) {
265 MSSA->ensureOptimizedUses();
266 }
267
268 bool run(Function &F);
269
270 // Copied from NewGVN.cpp
271 // This function provides global ranking of operations so that we can place
272 // them in a canonical order. Note that rank alone is not necessarily enough
273 // for a complete ordering, as constants all have the same rank. However,
274 // generally, we will simplify an operation with all constants so that it
275 // doesn't matter what order they appear in.
276 unsigned int rank(const Value *V) const;
277
278 private:
279 GVNPass::ValueTable VN;
280 DominatorTree *DT;
281 PostDominatorTree *PDT;
282 AliasAnalysis *AA;
283 MemoryDependenceResults *MD;
284 MemorySSA *MSSA;
285 std::unique_ptr<MemorySSAUpdater> MSSAUpdater;
286 DenseMap<const Value *, unsigned> DFSNumber;
287 BBSideEffectsSet BBSideEffects;
288 DenseSet<const BasicBlock *> HoistBarrier;
289 SmallVector<BasicBlock *, 32> IDFBlocks;
290 unsigned NumFuncArgs;
291 const bool HoistingGeps = false;
292
293 enum InsKind { Unknown, Scalar, Load, Store };
294
295 // Return true when there are exception handling in BB.
296 bool hasEH(const BasicBlock *BB);
297
298 // Return true when I1 appears before I2 in the instructions of BB.
firstInBB(const Instruction * I1,const Instruction * I2)299 bool firstInBB(const Instruction *I1, const Instruction *I2) {
300 assert(I1->getParent() == I2->getParent());
301 unsigned I1DFS = DFSNumber.lookup(I1);
302 unsigned I2DFS = DFSNumber.lookup(I2);
303 assert(I1DFS && I2DFS);
304 return I1DFS < I2DFS;
305 }
306
307 // Return true when there are memory uses of Def in BB.
308 bool hasMemoryUse(const Instruction *NewPt, MemoryDef *Def,
309 const BasicBlock *BB);
310
311 bool hasEHhelper(const BasicBlock *BB, const BasicBlock *SrcBB,
312 int &NBBsOnAllPaths);
313
314 // Return true when there are exception handling or loads of memory Def
315 // between Def and NewPt. This function is only called for stores: Def is
316 // the MemoryDef of the store to be hoisted.
317
318 // Decrement by 1 NBBsOnAllPaths for each block between HoistPt and BB, and
319 // return true when the counter NBBsOnAllPaths reaces 0, except when it is
320 // initialized to -1 which is unlimited.
321 bool hasEHOrLoadsOnPath(const Instruction *NewPt, MemoryDef *Def,
322 int &NBBsOnAllPaths);
323
324 // Return true when there are exception handling between HoistPt and BB.
325 // Decrement by 1 NBBsOnAllPaths for each block between HoistPt and BB, and
326 // return true when the counter NBBsOnAllPaths reaches 0, except when it is
327 // initialized to -1 which is unlimited.
328 bool hasEHOnPath(const BasicBlock *HoistPt, const BasicBlock *SrcBB,
329 int &NBBsOnAllPaths);
330
331 // Return true when it is safe to hoist a memory load or store U from OldPt
332 // to NewPt.
333 bool safeToHoistLdSt(const Instruction *NewPt, const Instruction *OldPt,
334 MemoryUseOrDef *U, InsKind K, int &NBBsOnAllPaths);
335
336 // Return true when it is safe to hoist scalar instructions from all blocks in
337 // WL to HoistBB.
safeToHoistScalar(const BasicBlock * HoistBB,const BasicBlock * BB,int & NBBsOnAllPaths)338 bool safeToHoistScalar(const BasicBlock *HoistBB, const BasicBlock *BB,
339 int &NBBsOnAllPaths) {
340 return !hasEHOnPath(HoistBB, BB, NBBsOnAllPaths);
341 }
342
343 // In the inverse CFG, the dominance frontier of basic block (BB) is the
344 // point where ANTIC needs to be computed for instructions which are going
345 // to be hoisted. Since this point does not change during gvn-hoist,
346 // we compute it only once (on demand).
347 // The ides is inspired from:
348 // "Partial Redundancy Elimination in SSA Form"
349 // ROBERT KENNEDY, SUN CHAN, SHIN-MING LIU, RAYMOND LO, PENG TU and FRED CHOW
350 // They use similar idea in the forward graph to find fully redundant and
351 // partially redundant expressions, here it is used in the inverse graph to
352 // find fully anticipable instructions at merge point (post-dominator in
353 // the inverse CFG).
354 // Returns the edge via which an instruction in BB will get the values from.
355
356 // Returns true when the values are flowing out to each edge.
357 bool valueAnticipable(CHIArgs C, Instruction *TI) const;
358
359 // Check if it is safe to hoist values tracked by CHI in the range
360 // [Begin, End) and accumulate them in Safe.
361 void checkSafety(CHIArgs C, BasicBlock *BB, InsKind K,
362 SmallVectorImpl<CHIArg> &Safe);
363
364 using RenameStackType = DenseMap<VNType, SmallVector<Instruction *, 2>>;
365
366 // Push all the VNs corresponding to BB into RenameStack.
367 void fillRenameStack(BasicBlock *BB, InValuesType &ValueBBs,
368 RenameStackType &RenameStack);
369
370 void fillChiArgs(BasicBlock *BB, OutValuesType &CHIBBs,
371 RenameStackType &RenameStack);
372
373 // Walk the post-dominator tree top-down and use a stack for each value to
374 // store the last value you see. When you hit a CHI from a given edge, the
375 // value to use as the argument is at the top of the stack, add the value to
376 // CHI and pop.
insertCHI(InValuesType & ValueBBs,OutValuesType & CHIBBs)377 void insertCHI(InValuesType &ValueBBs, OutValuesType &CHIBBs) {
378 auto Root = PDT->getNode(nullptr);
379 if (!Root)
380 return;
381 // Depth first walk on PDom tree to fill the CHIargs at each PDF.
382 for (auto *Node : depth_first(Root)) {
383 BasicBlock *BB = Node->getBlock();
384 if (!BB)
385 continue;
386
387 RenameStackType RenameStack;
388 // Collect all values in BB and push to stack.
389 fillRenameStack(BB, ValueBBs, RenameStack);
390
391 // Fill outgoing values in each CHI corresponding to BB.
392 fillChiArgs(BB, CHIBBs, RenameStack);
393 }
394 }
395
396 // Walk all the CHI-nodes to find ones which have a empty-entry and remove
397 // them Then collect all the instructions which are safe to hoist and see if
398 // they form a list of anticipable values. OutValues contains CHIs
399 // corresponding to each basic block.
400 void findHoistableCandidates(OutValuesType &CHIBBs, InsKind K,
401 HoistingPointList &HPL);
402
403 // Compute insertion points for each values which can be fully anticipated at
404 // a dominator. HPL contains all such values.
computeInsertionPoints(const VNtoInsns & Map,HoistingPointList & HPL,InsKind K)405 void computeInsertionPoints(const VNtoInsns &Map, HoistingPointList &HPL,
406 InsKind K) {
407 // Sort VNs based on their rankings
408 std::vector<VNType> Ranks;
409 for (const auto &Entry : Map) {
410 Ranks.push_back(Entry.first);
411 }
412
413 // TODO: Remove fully-redundant expressions.
414 // Get instruction from the Map, assume that all the Instructions
415 // with same VNs have same rank (this is an approximation).
416 llvm::sort(Ranks, [this, &Map](const VNType &r1, const VNType &r2) {
417 return (rank(*Map.lookup(r1).begin()) < rank(*Map.lookup(r2).begin()));
418 });
419
420 // - Sort VNs according to their rank, and start with lowest ranked VN
421 // - Take a VN and for each instruction with same VN
422 // - Find the dominance frontier in the inverse graph (PDF)
423 // - Insert the chi-node at PDF
424 // - Remove the chi-nodes with missing entries
425 // - Remove values from CHI-nodes which do not truly flow out, e.g.,
426 // modified along the path.
427 // - Collect the remaining values that are still anticipable
428 SmallVector<BasicBlock *, 2> IDFBlocks;
429 ReverseIDFCalculator IDFs(*PDT);
430 OutValuesType OutValue;
431 InValuesType InValue;
432 for (const auto &R : Ranks) {
433 const SmallVecInsn &V = Map.lookup(R);
434 if (V.size() < 2)
435 continue;
436 const VNType &VN = R;
437 SmallPtrSet<BasicBlock *, 2> VNBlocks;
438 for (const auto &I : V) {
439 BasicBlock *BBI = I->getParent();
440 if (!hasEH(BBI))
441 VNBlocks.insert(BBI);
442 }
443 // Compute the Post Dominance Frontiers of each basic block
444 // The dominance frontier of a live block X in the reverse
445 // control graph is the set of blocks upon which X is control
446 // dependent. The following sequence computes the set of blocks
447 // which currently have dead terminators that are control
448 // dependence sources of a block which is in NewLiveBlocks.
449 IDFs.setDefiningBlocks(VNBlocks);
450 IDFBlocks.clear();
451 IDFs.calculate(IDFBlocks);
452
453 // Make a map of BB vs instructions to be hoisted.
454 for (unsigned i = 0; i < V.size(); ++i) {
455 InValue[V[i]->getParent()].push_back(std::make_pair(VN, V[i]));
456 }
457 // Insert empty CHI node for this VN. This is used to factor out
458 // basic blocks where the ANTIC can potentially change.
459 CHIArg EmptyChi = {VN, nullptr, nullptr};
460 for (auto *IDFBB : IDFBlocks) {
461 for (unsigned i = 0; i < V.size(); ++i) {
462 // Ignore spurious PDFs.
463 if (DT->properlyDominates(IDFBB, V[i]->getParent())) {
464 OutValue[IDFBB].push_back(EmptyChi);
465 LLVM_DEBUG(dbgs() << "\nInserting a CHI for BB: "
466 << IDFBB->getName() << ", for Insn: " << *V[i]);
467 }
468 }
469 }
470 }
471
472 // Insert CHI args at each PDF to iterate on factored graph of
473 // control dependence.
474 insertCHI(InValue, OutValue);
475 // Using the CHI args inserted at each PDF, find fully anticipable values.
476 findHoistableCandidates(OutValue, K, HPL);
477 }
478
479 // Return true when all operands of Instr are available at insertion point
480 // HoistPt. When limiting the number of hoisted expressions, one could hoist
481 // a load without hoisting its access function. So before hoisting any
482 // expression, make sure that all its operands are available at insert point.
483 bool allOperandsAvailable(const Instruction *I,
484 const BasicBlock *HoistPt) const;
485
486 // Same as allOperandsAvailable with recursive check for GEP operands.
487 bool allGepOperandsAvailable(const Instruction *I,
488 const BasicBlock *HoistPt) const;
489
490 // Make all operands of the GEP available.
491 void makeGepsAvailable(Instruction *Repl, BasicBlock *HoistPt,
492 const SmallVecInsn &InstructionsToHoist,
493 Instruction *Gep) const;
494
495 void updateAlignment(Instruction *I, Instruction *Repl);
496
497 // Remove all the instructions in Candidates and replace their usage with
498 // Repl. Returns the number of instructions removed.
499 unsigned rauw(const SmallVecInsn &Candidates, Instruction *Repl,
500 MemoryUseOrDef *NewMemAcc);
501
502 // Replace all Memory PHI usage with NewMemAcc.
503 void raMPHIuw(MemoryUseOrDef *NewMemAcc);
504
505 // Remove all other instructions and replace them with Repl.
506 unsigned removeAndReplace(const SmallVecInsn &Candidates, Instruction *Repl,
507 BasicBlock *DestBB, bool MoveAccess);
508
509 // In the case Repl is a load or a store, we make all their GEPs
510 // available: GEPs are not hoisted by default to avoid the address
511 // computations to be hoisted without the associated load or store.
512 bool makeGepOperandsAvailable(Instruction *Repl, BasicBlock *HoistPt,
513 const SmallVecInsn &InstructionsToHoist) const;
514
515 std::pair<unsigned, unsigned> hoist(HoistingPointList &HPL);
516
517 // Hoist all expressions. Returns Number of scalars hoisted
518 // and number of non-scalars hoisted.
519 std::pair<unsigned, unsigned> hoistExpressions(Function &F);
520 };
521
522 class GVNHoistLegacyPass : public FunctionPass {
523 public:
524 static char ID;
525
GVNHoistLegacyPass()526 GVNHoistLegacyPass() : FunctionPass(ID) {
527 initializeGVNHoistLegacyPassPass(*PassRegistry::getPassRegistry());
528 }
529
runOnFunction(Function & F)530 bool runOnFunction(Function &F) override {
531 if (skipFunction(F))
532 return false;
533 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
534 auto &PDT = getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree();
535 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
536 auto &MD = getAnalysis<MemoryDependenceWrapperPass>().getMemDep();
537 auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
538
539 GVNHoist G(&DT, &PDT, &AA, &MD, &MSSA);
540 return G.run(F);
541 }
542
getAnalysisUsage(AnalysisUsage & AU) const543 void getAnalysisUsage(AnalysisUsage &AU) const override {
544 AU.addRequired<DominatorTreeWrapperPass>();
545 AU.addRequired<PostDominatorTreeWrapperPass>();
546 AU.addRequired<AAResultsWrapperPass>();
547 AU.addRequired<MemoryDependenceWrapperPass>();
548 AU.addRequired<MemorySSAWrapperPass>();
549 AU.addPreserved<DominatorTreeWrapperPass>();
550 AU.addPreserved<MemorySSAWrapperPass>();
551 AU.addPreserved<GlobalsAAWrapperPass>();
552 }
553 };
554
run(Function & F)555 bool GVNHoist::run(Function &F) {
556 NumFuncArgs = F.arg_size();
557 VN.setDomTree(DT);
558 VN.setAliasAnalysis(AA);
559 VN.setMemDep(MD);
560 bool Res = false;
561 // Perform DFS Numbering of instructions.
562 unsigned BBI = 0;
563 for (const BasicBlock *BB : depth_first(&F.getEntryBlock())) {
564 DFSNumber[BB] = ++BBI;
565 unsigned I = 0;
566 for (const auto &Inst : *BB)
567 DFSNumber[&Inst] = ++I;
568 }
569
570 int ChainLength = 0;
571
572 // FIXME: use lazy evaluation of VN to avoid the fix-point computation.
573 while (true) {
574 if (MaxChainLength != -1 && ++ChainLength >= MaxChainLength)
575 return Res;
576
577 auto HoistStat = hoistExpressions(F);
578 if (HoistStat.first + HoistStat.second == 0)
579 return Res;
580
581 if (HoistStat.second > 0)
582 // To address a limitation of the current GVN, we need to rerun the
583 // hoisting after we hoisted loads or stores in order to be able to
584 // hoist all scalars dependent on the hoisted ld/st.
585 VN.clear();
586
587 Res = true;
588 }
589
590 return Res;
591 }
592
rank(const Value * V) const593 unsigned int GVNHoist::rank(const Value *V) const {
594 // Prefer constants to undef to anything else
595 // Undef is a constant, have to check it first.
596 // Prefer smaller constants to constantexprs
597 if (isa<ConstantExpr>(V))
598 return 2;
599 if (isa<UndefValue>(V))
600 return 1;
601 if (isa<Constant>(V))
602 return 0;
603 else if (auto *A = dyn_cast<Argument>(V))
604 return 3 + A->getArgNo();
605
606 // Need to shift the instruction DFS by number of arguments + 3 to account
607 // for the constant and argument ranking above.
608 auto Result = DFSNumber.lookup(V);
609 if (Result > 0)
610 return 4 + NumFuncArgs + Result;
611 // Unreachable or something else, just return a really large number.
612 return ~0;
613 }
614
hasEH(const BasicBlock * BB)615 bool GVNHoist::hasEH(const BasicBlock *BB) {
616 auto It = BBSideEffects.find(BB);
617 if (It != BBSideEffects.end())
618 return It->second;
619
620 if (BB->isEHPad() || BB->hasAddressTaken()) {
621 BBSideEffects[BB] = true;
622 return true;
623 }
624
625 if (BB->getTerminator()->mayThrow()) {
626 BBSideEffects[BB] = true;
627 return true;
628 }
629
630 BBSideEffects[BB] = false;
631 return false;
632 }
633
hasMemoryUse(const Instruction * NewPt,MemoryDef * Def,const BasicBlock * BB)634 bool GVNHoist::hasMemoryUse(const Instruction *NewPt, MemoryDef *Def,
635 const BasicBlock *BB) {
636 const MemorySSA::AccessList *Acc = MSSA->getBlockAccesses(BB);
637 if (!Acc)
638 return false;
639
640 Instruction *OldPt = Def->getMemoryInst();
641 const BasicBlock *OldBB = OldPt->getParent();
642 const BasicBlock *NewBB = NewPt->getParent();
643 bool ReachedNewPt = false;
644
645 for (const MemoryAccess &MA : *Acc)
646 if (const MemoryUse *MU = dyn_cast<MemoryUse>(&MA)) {
647 Instruction *Insn = MU->getMemoryInst();
648
649 // Do not check whether MU aliases Def when MU occurs after OldPt.
650 if (BB == OldBB && firstInBB(OldPt, Insn))
651 break;
652
653 // Do not check whether MU aliases Def when MU occurs before NewPt.
654 if (BB == NewBB) {
655 if (!ReachedNewPt) {
656 if (firstInBB(Insn, NewPt))
657 continue;
658 ReachedNewPt = true;
659 }
660 }
661 if (MemorySSAUtil::defClobbersUseOrDef(Def, MU, *AA))
662 return true;
663 }
664
665 return false;
666 }
667
hasEHhelper(const BasicBlock * BB,const BasicBlock * SrcBB,int & NBBsOnAllPaths)668 bool GVNHoist::hasEHhelper(const BasicBlock *BB, const BasicBlock *SrcBB,
669 int &NBBsOnAllPaths) {
670 // Stop walk once the limit is reached.
671 if (NBBsOnAllPaths == 0)
672 return true;
673
674 // Impossible to hoist with exceptions on the path.
675 if (hasEH(BB))
676 return true;
677
678 // No such instruction after HoistBarrier in a basic block was
679 // selected for hoisting so instructions selected within basic block with
680 // a hoist barrier can be hoisted.
681 if ((BB != SrcBB) && HoistBarrier.count(BB))
682 return true;
683
684 return false;
685 }
686
hasEHOrLoadsOnPath(const Instruction * NewPt,MemoryDef * Def,int & NBBsOnAllPaths)687 bool GVNHoist::hasEHOrLoadsOnPath(const Instruction *NewPt, MemoryDef *Def,
688 int &NBBsOnAllPaths) {
689 const BasicBlock *NewBB = NewPt->getParent();
690 const BasicBlock *OldBB = Def->getBlock();
691 assert(DT->dominates(NewBB, OldBB) && "invalid path");
692 assert(DT->dominates(Def->getDefiningAccess()->getBlock(), NewBB) &&
693 "def does not dominate new hoisting point");
694
695 // Walk all basic blocks reachable in depth-first iteration on the inverse
696 // CFG from OldBB to NewBB. These blocks are all the blocks that may be
697 // executed between the execution of NewBB and OldBB. Hoisting an expression
698 // from OldBB into NewBB has to be safe on all execution paths.
699 for (auto I = idf_begin(OldBB), E = idf_end(OldBB); I != E;) {
700 const BasicBlock *BB = *I;
701 if (BB == NewBB) {
702 // Stop traversal when reaching HoistPt.
703 I.skipChildren();
704 continue;
705 }
706
707 if (hasEHhelper(BB, OldBB, NBBsOnAllPaths))
708 return true;
709
710 // Check that we do not move a store past loads.
711 if (hasMemoryUse(NewPt, Def, BB))
712 return true;
713
714 // -1 is unlimited number of blocks on all paths.
715 if (NBBsOnAllPaths != -1)
716 --NBBsOnAllPaths;
717
718 ++I;
719 }
720
721 return false;
722 }
723
hasEHOnPath(const BasicBlock * HoistPt,const BasicBlock * SrcBB,int & NBBsOnAllPaths)724 bool GVNHoist::hasEHOnPath(const BasicBlock *HoistPt, const BasicBlock *SrcBB,
725 int &NBBsOnAllPaths) {
726 assert(DT->dominates(HoistPt, SrcBB) && "Invalid path");
727
728 // Walk all basic blocks reachable in depth-first iteration on
729 // the inverse CFG from BBInsn to NewHoistPt. These blocks are all the
730 // blocks that may be executed between the execution of NewHoistPt and
731 // BBInsn. Hoisting an expression from BBInsn into NewHoistPt has to be safe
732 // on all execution paths.
733 for (auto I = idf_begin(SrcBB), E = idf_end(SrcBB); I != E;) {
734 const BasicBlock *BB = *I;
735 if (BB == HoistPt) {
736 // Stop traversal when reaching NewHoistPt.
737 I.skipChildren();
738 continue;
739 }
740
741 if (hasEHhelper(BB, SrcBB, NBBsOnAllPaths))
742 return true;
743
744 // -1 is unlimited number of blocks on all paths.
745 if (NBBsOnAllPaths != -1)
746 --NBBsOnAllPaths;
747
748 ++I;
749 }
750
751 return false;
752 }
753
safeToHoistLdSt(const Instruction * NewPt,const Instruction * OldPt,MemoryUseOrDef * U,GVNHoist::InsKind K,int & NBBsOnAllPaths)754 bool GVNHoist::safeToHoistLdSt(const Instruction *NewPt,
755 const Instruction *OldPt, MemoryUseOrDef *U,
756 GVNHoist::InsKind K, int &NBBsOnAllPaths) {
757 // In place hoisting is safe.
758 if (NewPt == OldPt)
759 return true;
760
761 const BasicBlock *NewBB = NewPt->getParent();
762 const BasicBlock *OldBB = OldPt->getParent();
763 const BasicBlock *UBB = U->getBlock();
764
765 // Check for dependences on the Memory SSA.
766 MemoryAccess *D = U->getDefiningAccess();
767 BasicBlock *DBB = D->getBlock();
768 if (DT->properlyDominates(NewBB, DBB))
769 // Cannot move the load or store to NewBB above its definition in DBB.
770 return false;
771
772 if (NewBB == DBB && !MSSA->isLiveOnEntryDef(D))
773 if (auto *UD = dyn_cast<MemoryUseOrDef>(D))
774 if (!firstInBB(UD->getMemoryInst(), NewPt))
775 // Cannot move the load or store to NewPt above its definition in D.
776 return false;
777
778 // Check for unsafe hoistings due to side effects.
779 if (K == InsKind::Store) {
780 if (hasEHOrLoadsOnPath(NewPt, cast<MemoryDef>(U), NBBsOnAllPaths))
781 return false;
782 } else if (hasEHOnPath(NewBB, OldBB, NBBsOnAllPaths))
783 return false;
784
785 if (UBB == NewBB) {
786 if (DT->properlyDominates(DBB, NewBB))
787 return true;
788 assert(UBB == DBB);
789 assert(MSSA->locallyDominates(D, U));
790 }
791
792 // No side effects: it is safe to hoist.
793 return true;
794 }
795
valueAnticipable(CHIArgs C,Instruction * TI) const796 bool GVNHoist::valueAnticipable(CHIArgs C, Instruction *TI) const {
797 if (TI->getNumSuccessors() > (unsigned)size(C))
798 return false; // Not enough args in this CHI.
799
800 for (auto CHI : C) {
801 // Find if all the edges have values flowing out of BB.
802 if (!llvm::is_contained(successors(TI), CHI.Dest))
803 return false;
804 }
805 return true;
806 }
807
checkSafety(CHIArgs C,BasicBlock * BB,GVNHoist::InsKind K,SmallVectorImpl<CHIArg> & Safe)808 void GVNHoist::checkSafety(CHIArgs C, BasicBlock *BB, GVNHoist::InsKind K,
809 SmallVectorImpl<CHIArg> &Safe) {
810 int NumBBsOnAllPaths = MaxNumberOfBBSInPath;
811 for (auto CHI : C) {
812 Instruction *Insn = CHI.I;
813 if (!Insn) // No instruction was inserted in this CHI.
814 continue;
815 if (K == InsKind::Scalar) {
816 if (safeToHoistScalar(BB, Insn->getParent(), NumBBsOnAllPaths))
817 Safe.push_back(CHI);
818 } else {
819 auto *T = BB->getTerminator();
820 if (MemoryUseOrDef *UD = MSSA->getMemoryAccess(Insn))
821 if (safeToHoistLdSt(T, Insn, UD, K, NumBBsOnAllPaths))
822 Safe.push_back(CHI);
823 }
824 }
825 }
826
fillRenameStack(BasicBlock * BB,InValuesType & ValueBBs,GVNHoist::RenameStackType & RenameStack)827 void GVNHoist::fillRenameStack(BasicBlock *BB, InValuesType &ValueBBs,
828 GVNHoist::RenameStackType &RenameStack) {
829 auto it1 = ValueBBs.find(BB);
830 if (it1 != ValueBBs.end()) {
831 // Iterate in reverse order to keep lower ranked values on the top.
832 LLVM_DEBUG(dbgs() << "\nVisiting: " << BB->getName()
833 << " for pushing instructions on stack";);
834 for (std::pair<VNType, Instruction *> &VI : reverse(it1->second)) {
835 // Get the value of instruction I
836 LLVM_DEBUG(dbgs() << "\nPushing on stack: " << *VI.second);
837 RenameStack[VI.first].push_back(VI.second);
838 }
839 }
840 }
841
fillChiArgs(BasicBlock * BB,OutValuesType & CHIBBs,GVNHoist::RenameStackType & RenameStack)842 void GVNHoist::fillChiArgs(BasicBlock *BB, OutValuesType &CHIBBs,
843 GVNHoist::RenameStackType &RenameStack) {
844 // For each *predecessor* (because Post-DOM) of BB check if it has a CHI
845 for (auto *Pred : predecessors(BB)) {
846 auto P = CHIBBs.find(Pred);
847 if (P == CHIBBs.end()) {
848 continue;
849 }
850 LLVM_DEBUG(dbgs() << "\nLooking at CHIs in: " << Pred->getName(););
851 // A CHI is found (BB -> Pred is an edge in the CFG)
852 // Pop the stack until Top(V) = Ve.
853 auto &VCHI = P->second;
854 for (auto It = VCHI.begin(), E = VCHI.end(); It != E;) {
855 CHIArg &C = *It;
856 if (!C.Dest) {
857 auto si = RenameStack.find(C.VN);
858 // The Basic Block where CHI is must dominate the value we want to
859 // track in a CHI. In the PDom walk, there can be values in the
860 // stack which are not control dependent e.g., nested loop.
861 if (si != RenameStack.end() && si->second.size() &&
862 DT->properlyDominates(Pred, si->second.back()->getParent())) {
863 C.Dest = BB; // Assign the edge
864 C.I = si->second.pop_back_val(); // Assign the argument
865 LLVM_DEBUG(dbgs()
866 << "\nCHI Inserted in BB: " << C.Dest->getName() << *C.I
867 << ", VN: " << C.VN.first << ", " << C.VN.second);
868 }
869 // Move to next CHI of a different value
870 It = std::find_if(It, VCHI.end(), [It](CHIArg &A) { return A != *It; });
871 } else
872 ++It;
873 }
874 }
875 }
876
findHoistableCandidates(OutValuesType & CHIBBs,GVNHoist::InsKind K,HoistingPointList & HPL)877 void GVNHoist::findHoistableCandidates(OutValuesType &CHIBBs,
878 GVNHoist::InsKind K,
879 HoistingPointList &HPL) {
880 auto cmpVN = [](const CHIArg &A, const CHIArg &B) { return A.VN < B.VN; };
881
882 // CHIArgs now have the outgoing values, so check for anticipability and
883 // accumulate hoistable candidates in HPL.
884 for (std::pair<BasicBlock *, SmallVector<CHIArg, 2>> &A : CHIBBs) {
885 BasicBlock *BB = A.first;
886 SmallVectorImpl<CHIArg> &CHIs = A.second;
887 // Vector of PHIs contains PHIs for different instructions.
888 // Sort the args according to their VNs, such that identical
889 // instructions are together.
890 llvm::stable_sort(CHIs, cmpVN);
891 auto TI = BB->getTerminator();
892 auto B = CHIs.begin();
893 // [PreIt, PHIIt) form a range of CHIs which have identical VNs.
894 auto PHIIt = llvm::find_if(CHIs, [B](CHIArg &A) { return A != *B; });
895 auto PrevIt = CHIs.begin();
896 while (PrevIt != PHIIt) {
897 // Collect values which satisfy safety checks.
898 SmallVector<CHIArg, 2> Safe;
899 // We check for safety first because there might be multiple values in
900 // the same path, some of which are not safe to be hoisted, but overall
901 // each edge has at least one value which can be hoisted, making the
902 // value anticipable along that path.
903 checkSafety(make_range(PrevIt, PHIIt), BB, K, Safe);
904
905 // List of safe values should be anticipable at TI.
906 if (valueAnticipable(make_range(Safe.begin(), Safe.end()), TI)) {
907 HPL.push_back({BB, SmallVecInsn()});
908 SmallVecInsn &V = HPL.back().second;
909 for (auto B : Safe)
910 V.push_back(B.I);
911 }
912
913 // Check other VNs
914 PrevIt = PHIIt;
915 PHIIt = std::find_if(PrevIt, CHIs.end(),
916 [PrevIt](CHIArg &A) { return A != *PrevIt; });
917 }
918 }
919 }
920
allOperandsAvailable(const Instruction * I,const BasicBlock * HoistPt) const921 bool GVNHoist::allOperandsAvailable(const Instruction *I,
922 const BasicBlock *HoistPt) const {
923 for (const Use &Op : I->operands())
924 if (const auto *Inst = dyn_cast<Instruction>(&Op))
925 if (!DT->dominates(Inst->getParent(), HoistPt))
926 return false;
927
928 return true;
929 }
930
allGepOperandsAvailable(const Instruction * I,const BasicBlock * HoistPt) const931 bool GVNHoist::allGepOperandsAvailable(const Instruction *I,
932 const BasicBlock *HoistPt) const {
933 for (const Use &Op : I->operands())
934 if (const auto *Inst = dyn_cast<Instruction>(&Op))
935 if (!DT->dominates(Inst->getParent(), HoistPt)) {
936 if (const GetElementPtrInst *GepOp =
937 dyn_cast<GetElementPtrInst>(Inst)) {
938 if (!allGepOperandsAvailable(GepOp, HoistPt))
939 return false;
940 // Gep is available if all operands of GepOp are available.
941 } else {
942 // Gep is not available if it has operands other than GEPs that are
943 // defined in blocks not dominating HoistPt.
944 return false;
945 }
946 }
947 return true;
948 }
949
makeGepsAvailable(Instruction * Repl,BasicBlock * HoistPt,const SmallVecInsn & InstructionsToHoist,Instruction * Gep) const950 void GVNHoist::makeGepsAvailable(Instruction *Repl, BasicBlock *HoistPt,
951 const SmallVecInsn &InstructionsToHoist,
952 Instruction *Gep) const {
953 assert(allGepOperandsAvailable(Gep, HoistPt) && "GEP operands not available");
954
955 Instruction *ClonedGep = Gep->clone();
956 for (unsigned i = 0, e = Gep->getNumOperands(); i != e; ++i)
957 if (Instruction *Op = dyn_cast<Instruction>(Gep->getOperand(i))) {
958 // Check whether the operand is already available.
959 if (DT->dominates(Op->getParent(), HoistPt))
960 continue;
961
962 // As a GEP can refer to other GEPs, recursively make all the operands
963 // of this GEP available at HoistPt.
964 if (GetElementPtrInst *GepOp = dyn_cast<GetElementPtrInst>(Op))
965 makeGepsAvailable(ClonedGep, HoistPt, InstructionsToHoist, GepOp);
966 }
967
968 // Copy Gep and replace its uses in Repl with ClonedGep.
969 ClonedGep->insertBefore(HoistPt->getTerminator());
970
971 // Conservatively discard any optimization hints, they may differ on the
972 // other paths.
973 ClonedGep->dropUnknownNonDebugMetadata();
974
975 // If we have optimization hints which agree with each other along different
976 // paths, preserve them.
977 for (const Instruction *OtherInst : InstructionsToHoist) {
978 const GetElementPtrInst *OtherGep;
979 if (auto *OtherLd = dyn_cast<LoadInst>(OtherInst))
980 OtherGep = cast<GetElementPtrInst>(OtherLd->getPointerOperand());
981 else
982 OtherGep = cast<GetElementPtrInst>(
983 cast<StoreInst>(OtherInst)->getPointerOperand());
984 ClonedGep->andIRFlags(OtherGep);
985 }
986
987 // Replace uses of Gep with ClonedGep in Repl.
988 Repl->replaceUsesOfWith(Gep, ClonedGep);
989 }
990
updateAlignment(Instruction * I,Instruction * Repl)991 void GVNHoist::updateAlignment(Instruction *I, Instruction *Repl) {
992 if (auto *ReplacementLoad = dyn_cast<LoadInst>(Repl)) {
993 ReplacementLoad->setAlignment(
994 std::min(ReplacementLoad->getAlign(), cast<LoadInst>(I)->getAlign()));
995 ++NumLoadsRemoved;
996 } else if (auto *ReplacementStore = dyn_cast<StoreInst>(Repl)) {
997 ReplacementStore->setAlignment(
998 std::min(ReplacementStore->getAlign(), cast<StoreInst>(I)->getAlign()));
999 ++NumStoresRemoved;
1000 } else if (auto *ReplacementAlloca = dyn_cast<AllocaInst>(Repl)) {
1001 ReplacementAlloca->setAlignment(std::max(ReplacementAlloca->getAlign(),
1002 cast<AllocaInst>(I)->getAlign()));
1003 } else if (isa<CallInst>(Repl)) {
1004 ++NumCallsRemoved;
1005 }
1006 }
1007
rauw(const SmallVecInsn & Candidates,Instruction * Repl,MemoryUseOrDef * NewMemAcc)1008 unsigned GVNHoist::rauw(const SmallVecInsn &Candidates, Instruction *Repl,
1009 MemoryUseOrDef *NewMemAcc) {
1010 unsigned NR = 0;
1011 for (Instruction *I : Candidates) {
1012 if (I != Repl) {
1013 ++NR;
1014 updateAlignment(I, Repl);
1015 if (NewMemAcc) {
1016 // Update the uses of the old MSSA access with NewMemAcc.
1017 MemoryAccess *OldMA = MSSA->getMemoryAccess(I);
1018 OldMA->replaceAllUsesWith(NewMemAcc);
1019 MSSAUpdater->removeMemoryAccess(OldMA);
1020 }
1021
1022 Repl->andIRFlags(I);
1023 combineKnownMetadata(Repl, I);
1024 I->replaceAllUsesWith(Repl);
1025 // Also invalidate the Alias Analysis cache.
1026 MD->removeInstruction(I);
1027 I->eraseFromParent();
1028 }
1029 }
1030 return NR;
1031 }
1032
raMPHIuw(MemoryUseOrDef * NewMemAcc)1033 void GVNHoist::raMPHIuw(MemoryUseOrDef *NewMemAcc) {
1034 SmallPtrSet<MemoryPhi *, 4> UsePhis;
1035 for (User *U : NewMemAcc->users())
1036 if (MemoryPhi *Phi = dyn_cast<MemoryPhi>(U))
1037 UsePhis.insert(Phi);
1038
1039 for (MemoryPhi *Phi : UsePhis) {
1040 auto In = Phi->incoming_values();
1041 if (llvm::all_of(In, [&](Use &U) { return U == NewMemAcc; })) {
1042 Phi->replaceAllUsesWith(NewMemAcc);
1043 MSSAUpdater->removeMemoryAccess(Phi);
1044 }
1045 }
1046 }
1047
removeAndReplace(const SmallVecInsn & Candidates,Instruction * Repl,BasicBlock * DestBB,bool MoveAccess)1048 unsigned GVNHoist::removeAndReplace(const SmallVecInsn &Candidates,
1049 Instruction *Repl, BasicBlock *DestBB,
1050 bool MoveAccess) {
1051 MemoryUseOrDef *NewMemAcc = MSSA->getMemoryAccess(Repl);
1052 if (MoveAccess && NewMemAcc) {
1053 // The definition of this ld/st will not change: ld/st hoisting is
1054 // legal when the ld/st is not moved past its current definition.
1055 MSSAUpdater->moveToPlace(NewMemAcc, DestBB, MemorySSA::BeforeTerminator);
1056 }
1057
1058 // Replace all other instructions with Repl with memory access NewMemAcc.
1059 unsigned NR = rauw(Candidates, Repl, NewMemAcc);
1060
1061 // Remove MemorySSA phi nodes with the same arguments.
1062 if (NewMemAcc)
1063 raMPHIuw(NewMemAcc);
1064 return NR;
1065 }
1066
makeGepOperandsAvailable(Instruction * Repl,BasicBlock * HoistPt,const SmallVecInsn & InstructionsToHoist) const1067 bool GVNHoist::makeGepOperandsAvailable(
1068 Instruction *Repl, BasicBlock *HoistPt,
1069 const SmallVecInsn &InstructionsToHoist) const {
1070 // Check whether the GEP of a ld/st can be synthesized at HoistPt.
1071 GetElementPtrInst *Gep = nullptr;
1072 Instruction *Val = nullptr;
1073 if (auto *Ld = dyn_cast<LoadInst>(Repl)) {
1074 Gep = dyn_cast<GetElementPtrInst>(Ld->getPointerOperand());
1075 } else if (auto *St = dyn_cast<StoreInst>(Repl)) {
1076 Gep = dyn_cast<GetElementPtrInst>(St->getPointerOperand());
1077 Val = dyn_cast<Instruction>(St->getValueOperand());
1078 // Check that the stored value is available.
1079 if (Val) {
1080 if (isa<GetElementPtrInst>(Val)) {
1081 // Check whether we can compute the GEP at HoistPt.
1082 if (!allGepOperandsAvailable(Val, HoistPt))
1083 return false;
1084 } else if (!DT->dominates(Val->getParent(), HoistPt))
1085 return false;
1086 }
1087 }
1088
1089 // Check whether we can compute the Gep at HoistPt.
1090 if (!Gep || !allGepOperandsAvailable(Gep, HoistPt))
1091 return false;
1092
1093 makeGepsAvailable(Repl, HoistPt, InstructionsToHoist, Gep);
1094
1095 if (Val && isa<GetElementPtrInst>(Val))
1096 makeGepsAvailable(Repl, HoistPt, InstructionsToHoist, Val);
1097
1098 return true;
1099 }
1100
hoist(HoistingPointList & HPL)1101 std::pair<unsigned, unsigned> GVNHoist::hoist(HoistingPointList &HPL) {
1102 unsigned NI = 0, NL = 0, NS = 0, NC = 0, NR = 0;
1103 for (const HoistingPointInfo &HP : HPL) {
1104 // Find out whether we already have one of the instructions in HoistPt,
1105 // in which case we do not have to move it.
1106 BasicBlock *DestBB = HP.first;
1107 const SmallVecInsn &InstructionsToHoist = HP.second;
1108 Instruction *Repl = nullptr;
1109 for (Instruction *I : InstructionsToHoist)
1110 if (I->getParent() == DestBB)
1111 // If there are two instructions in HoistPt to be hoisted in place:
1112 // update Repl to be the first one, such that we can rename the uses
1113 // of the second based on the first.
1114 if (!Repl || firstInBB(I, Repl))
1115 Repl = I;
1116
1117 // Keep track of whether we moved the instruction so we know whether we
1118 // should move the MemoryAccess.
1119 bool MoveAccess = true;
1120 if (Repl) {
1121 // Repl is already in HoistPt: it remains in place.
1122 assert(allOperandsAvailable(Repl, DestBB) &&
1123 "instruction depends on operands that are not available");
1124 MoveAccess = false;
1125 } else {
1126 // When we do not find Repl in HoistPt, select the first in the list
1127 // and move it to HoistPt.
1128 Repl = InstructionsToHoist.front();
1129
1130 // We can move Repl in HoistPt only when all operands are available.
1131 // The order in which hoistings are done may influence the availability
1132 // of operands.
1133 if (!allOperandsAvailable(Repl, DestBB)) {
1134 // When HoistingGeps there is nothing more we can do to make the
1135 // operands available: just continue.
1136 if (HoistingGeps)
1137 continue;
1138
1139 // When not HoistingGeps we need to copy the GEPs.
1140 if (!makeGepOperandsAvailable(Repl, DestBB, InstructionsToHoist))
1141 continue;
1142 }
1143
1144 // Move the instruction at the end of HoistPt.
1145 Instruction *Last = DestBB->getTerminator();
1146 MD->removeInstruction(Repl);
1147 Repl->moveBefore(Last);
1148
1149 DFSNumber[Repl] = DFSNumber[Last]++;
1150 }
1151
1152 // Drop debug location as per debug info update guide.
1153 Repl->dropLocation();
1154 NR += removeAndReplace(InstructionsToHoist, Repl, DestBB, MoveAccess);
1155
1156 if (isa<LoadInst>(Repl))
1157 ++NL;
1158 else if (isa<StoreInst>(Repl))
1159 ++NS;
1160 else if (isa<CallInst>(Repl))
1161 ++NC;
1162 else // Scalar
1163 ++NI;
1164 }
1165
1166 if (MSSA && VerifyMemorySSA)
1167 MSSA->verifyMemorySSA();
1168
1169 NumHoisted += NL + NS + NC + NI;
1170 NumRemoved += NR;
1171 NumLoadsHoisted += NL;
1172 NumStoresHoisted += NS;
1173 NumCallsHoisted += NC;
1174 return {NI, NL + NC + NS};
1175 }
1176
hoistExpressions(Function & F)1177 std::pair<unsigned, unsigned> GVNHoist::hoistExpressions(Function &F) {
1178 InsnInfo II;
1179 LoadInfo LI;
1180 StoreInfo SI;
1181 CallInfo CI;
1182 for (BasicBlock *BB : depth_first(&F.getEntryBlock())) {
1183 int InstructionNb = 0;
1184 for (Instruction &I1 : *BB) {
1185 // If I1 cannot guarantee progress, subsequent instructions
1186 // in BB cannot be hoisted anyways.
1187 if (!isGuaranteedToTransferExecutionToSuccessor(&I1)) {
1188 HoistBarrier.insert(BB);
1189 break;
1190 }
1191 // Only hoist the first instructions in BB up to MaxDepthInBB. Hoisting
1192 // deeper may increase the register pressure and compilation time.
1193 if (MaxDepthInBB != -1 && InstructionNb++ >= MaxDepthInBB)
1194 break;
1195
1196 // Do not value number terminator instructions.
1197 if (I1.isTerminator())
1198 break;
1199
1200 if (auto *Load = dyn_cast<LoadInst>(&I1))
1201 LI.insert(Load, VN);
1202 else if (auto *Store = dyn_cast<StoreInst>(&I1))
1203 SI.insert(Store, VN);
1204 else if (auto *Call = dyn_cast<CallInst>(&I1)) {
1205 if (auto *Intr = dyn_cast<IntrinsicInst>(Call)) {
1206 if (isa<DbgInfoIntrinsic>(Intr) ||
1207 Intr->getIntrinsicID() == Intrinsic::assume ||
1208 Intr->getIntrinsicID() == Intrinsic::sideeffect)
1209 continue;
1210 }
1211 if (Call->mayHaveSideEffects())
1212 break;
1213
1214 if (Call->isConvergent())
1215 break;
1216
1217 CI.insert(Call, VN);
1218 } else if (HoistingGeps || !isa<GetElementPtrInst>(&I1))
1219 // Do not hoist scalars past calls that may write to memory because
1220 // that could result in spills later. geps are handled separately.
1221 // TODO: We can relax this for targets like AArch64 as they have more
1222 // registers than X86.
1223 II.insert(&I1, VN);
1224 }
1225 }
1226
1227 HoistingPointList HPL;
1228 computeInsertionPoints(II.getVNTable(), HPL, InsKind::Scalar);
1229 computeInsertionPoints(LI.getVNTable(), HPL, InsKind::Load);
1230 computeInsertionPoints(SI.getVNTable(), HPL, InsKind::Store);
1231 computeInsertionPoints(CI.getScalarVNTable(), HPL, InsKind::Scalar);
1232 computeInsertionPoints(CI.getLoadVNTable(), HPL, InsKind::Load);
1233 computeInsertionPoints(CI.getStoreVNTable(), HPL, InsKind::Store);
1234 return hoist(HPL);
1235 }
1236
1237 } // end namespace llvm
1238
run(Function & F,FunctionAnalysisManager & AM)1239 PreservedAnalyses GVNHoistPass::run(Function &F, FunctionAnalysisManager &AM) {
1240 DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F);
1241 PostDominatorTree &PDT = AM.getResult<PostDominatorTreeAnalysis>(F);
1242 AliasAnalysis &AA = AM.getResult<AAManager>(F);
1243 MemoryDependenceResults &MD = AM.getResult<MemoryDependenceAnalysis>(F);
1244 MemorySSA &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
1245 GVNHoist G(&DT, &PDT, &AA, &MD, &MSSA);
1246 if (!G.run(F))
1247 return PreservedAnalyses::all();
1248
1249 PreservedAnalyses PA;
1250 PA.preserve<DominatorTreeAnalysis>();
1251 PA.preserve<MemorySSAAnalysis>();
1252 return PA;
1253 }
1254
1255 char GVNHoistLegacyPass::ID = 0;
1256
1257 INITIALIZE_PASS_BEGIN(GVNHoistLegacyPass, "gvn-hoist",
1258 "Early GVN Hoisting of Expressions", false, false)
INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)1259 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
1260 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
1261 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1262 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
1263 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1264 INITIALIZE_PASS_END(GVNHoistLegacyPass, "gvn-hoist",
1265 "Early GVN Hoisting of Expressions", false, false)
1266
1267 FunctionPass *llvm::createGVNHoistPass() { return new GVNHoistLegacyPass(); }
1268