1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
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 global value numbering to eliminate fully redundant
10 // instructions. It also performs simple dead load elimination.
11 //
12 // Note that this pass does the value numbering itself; it does not use the
13 // ValueNumbering analysis passes.
14 //
15 //===----------------------------------------------------------------------===//
16
17 #include "llvm/Transforms/Scalar/GVN.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/DepthFirstIterator.h"
20 #include "llvm/ADT/Hashing.h"
21 #include "llvm/ADT/MapVector.h"
22 #include "llvm/ADT/PointerIntPair.h"
23 #include "llvm/ADT/PostOrderIterator.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/SetVector.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/Analysis/AliasAnalysis.h"
30 #include "llvm/Analysis/AssumptionCache.h"
31 #include "llvm/Analysis/CFG.h"
32 #include "llvm/Analysis/DomTreeUpdater.h"
33 #include "llvm/Analysis/GlobalsModRef.h"
34 #include "llvm/Analysis/InstructionSimplify.h"
35 #include "llvm/Analysis/LoopInfo.h"
36 #include "llvm/Analysis/MemoryBuiltins.h"
37 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
38 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
39 #include "llvm/Analysis/PHITransAddr.h"
40 #include "llvm/Analysis/TargetLibraryInfo.h"
41 #include "llvm/Analysis/ValueTracking.h"
42 #include "llvm/Config/llvm-config.h"
43 #include "llvm/IR/Attributes.h"
44 #include "llvm/IR/BasicBlock.h"
45 #include "llvm/IR/CallSite.h"
46 #include "llvm/IR/Constant.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DataLayout.h"
49 #include "llvm/IR/DebugInfoMetadata.h"
50 #include "llvm/IR/DebugLoc.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/Function.h"
53 #include "llvm/IR/InstrTypes.h"
54 #include "llvm/IR/Instruction.h"
55 #include "llvm/IR/Instructions.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/Intrinsics.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/Metadata.h"
60 #include "llvm/IR/Module.h"
61 #include "llvm/IR/Operator.h"
62 #include "llvm/IR/PassManager.h"
63 #include "llvm/IR/PatternMatch.h"
64 #include "llvm/IR/Type.h"
65 #include "llvm/IR/Use.h"
66 #include "llvm/IR/Value.h"
67 #include "llvm/Pass.h"
68 #include "llvm/Support/Casting.h"
69 #include "llvm/Support/CommandLine.h"
70 #include "llvm/Support/Compiler.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/raw_ostream.h"
73 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
74 #include "llvm/Transforms/Utils/Local.h"
75 #include "llvm/Transforms/Utils/SSAUpdater.h"
76 #include "llvm/Transforms/Utils/VNCoercion.h"
77 #include <algorithm>
78 #include <cassert>
79 #include <cstdint>
80 #include <utility>
81 #include <vector>
82
83 using namespace llvm;
84 using namespace llvm::gvn;
85 using namespace llvm::VNCoercion;
86 using namespace PatternMatch;
87
88 #define DEBUG_TYPE "gvn"
89
90 STATISTIC(NumGVNInstr, "Number of instructions deleted");
91 STATISTIC(NumGVNLoad, "Number of loads deleted");
92 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
93 STATISTIC(NumGVNBlocks, "Number of blocks merged");
94 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
95 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
96 STATISTIC(NumPRELoad, "Number of loads PRE'd");
97
98 static cl::opt<bool> EnablePRE("enable-pre",
99 cl::init(true), cl::Hidden);
100 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
101 static cl::opt<bool> EnableMemDep("enable-gvn-memdep", cl::init(true));
102
103 // Maximum allowed recursion depth.
104 static cl::opt<uint32_t>
105 MaxRecurseDepth("gvn-max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
106 cl::desc("Max recurse depth in GVN (default = 1000)"));
107
108 static cl::opt<uint32_t> MaxNumDeps(
109 "gvn-max-num-deps", cl::Hidden, cl::init(100), cl::ZeroOrMore,
110 cl::desc("Max number of dependences to attempt Load PRE (default = 100)"));
111
112 struct llvm::GVN::Expression {
113 uint32_t opcode;
114 Type *type;
115 bool commutative = false;
116 SmallVector<uint32_t, 4> varargs;
117
Expressionllvm::GVN::Expression118 Expression(uint32_t o = ~2U) : opcode(o) {}
119
operator ==llvm::GVN::Expression120 bool operator==(const Expression &other) const {
121 if (opcode != other.opcode)
122 return false;
123 if (opcode == ~0U || opcode == ~1U)
124 return true;
125 if (type != other.type)
126 return false;
127 if (varargs != other.varargs)
128 return false;
129 return true;
130 }
131
hash_value(const Expression & Value)132 friend hash_code hash_value(const Expression &Value) {
133 return hash_combine(
134 Value.opcode, Value.type,
135 hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
136 }
137 };
138
139 namespace llvm {
140
141 template <> struct DenseMapInfo<GVN::Expression> {
getEmptyKeyllvm::DenseMapInfo142 static inline GVN::Expression getEmptyKey() { return ~0U; }
getTombstoneKeyllvm::DenseMapInfo143 static inline GVN::Expression getTombstoneKey() { return ~1U; }
144
getHashValuellvm::DenseMapInfo145 static unsigned getHashValue(const GVN::Expression &e) {
146 using llvm::hash_value;
147
148 return static_cast<unsigned>(hash_value(e));
149 }
150
isEqualllvm::DenseMapInfo151 static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
152 return LHS == RHS;
153 }
154 };
155
156 } // end namespace llvm
157
158 /// Represents a particular available value that we know how to materialize.
159 /// Materialization of an AvailableValue never fails. An AvailableValue is
160 /// implicitly associated with a rematerialization point which is the
161 /// location of the instruction from which it was formed.
162 struct llvm::gvn::AvailableValue {
163 enum ValType {
164 SimpleVal, // A simple offsetted value that is accessed.
165 LoadVal, // A value produced by a load.
166 MemIntrin, // A memory intrinsic which is loaded from.
167 UndefVal // A UndefValue representing a value from dead block (which
168 // is not yet physically removed from the CFG).
169 };
170
171 /// V - The value that is live out of the block.
172 PointerIntPair<Value *, 2, ValType> Val;
173
174 /// Offset - The byte offset in Val that is interesting for the load query.
175 unsigned Offset;
176
getllvm::gvn::AvailableValue177 static AvailableValue get(Value *V, unsigned Offset = 0) {
178 AvailableValue Res;
179 Res.Val.setPointer(V);
180 Res.Val.setInt(SimpleVal);
181 Res.Offset = Offset;
182 return Res;
183 }
184
getMIllvm::gvn::AvailableValue185 static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
186 AvailableValue Res;
187 Res.Val.setPointer(MI);
188 Res.Val.setInt(MemIntrin);
189 Res.Offset = Offset;
190 return Res;
191 }
192
getLoadllvm::gvn::AvailableValue193 static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
194 AvailableValue Res;
195 Res.Val.setPointer(LI);
196 Res.Val.setInt(LoadVal);
197 Res.Offset = Offset;
198 return Res;
199 }
200
getUndefllvm::gvn::AvailableValue201 static AvailableValue getUndef() {
202 AvailableValue Res;
203 Res.Val.setPointer(nullptr);
204 Res.Val.setInt(UndefVal);
205 Res.Offset = 0;
206 return Res;
207 }
208
isSimpleValuellvm::gvn::AvailableValue209 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
isCoercedLoadValuellvm::gvn::AvailableValue210 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
isMemIntrinValuellvm::gvn::AvailableValue211 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
isUndefValuellvm::gvn::AvailableValue212 bool isUndefValue() const { return Val.getInt() == UndefVal; }
213
getSimpleValuellvm::gvn::AvailableValue214 Value *getSimpleValue() const {
215 assert(isSimpleValue() && "Wrong accessor");
216 return Val.getPointer();
217 }
218
getCoercedLoadValuellvm::gvn::AvailableValue219 LoadInst *getCoercedLoadValue() const {
220 assert(isCoercedLoadValue() && "Wrong accessor");
221 return cast<LoadInst>(Val.getPointer());
222 }
223
getMemIntrinValuellvm::gvn::AvailableValue224 MemIntrinsic *getMemIntrinValue() const {
225 assert(isMemIntrinValue() && "Wrong accessor");
226 return cast<MemIntrinsic>(Val.getPointer());
227 }
228
229 /// Emit code at the specified insertion point to adjust the value defined
230 /// here to the specified type. This handles various coercion cases.
231 Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
232 GVN &gvn) const;
233 };
234
235 /// Represents an AvailableValue which can be rematerialized at the end of
236 /// the associated BasicBlock.
237 struct llvm::gvn::AvailableValueInBlock {
238 /// BB - The basic block in question.
239 BasicBlock *BB;
240
241 /// AV - The actual available value
242 AvailableValue AV;
243
getllvm::gvn::AvailableValueInBlock244 static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) {
245 AvailableValueInBlock Res;
246 Res.BB = BB;
247 Res.AV = std::move(AV);
248 return Res;
249 }
250
getllvm::gvn::AvailableValueInBlock251 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
252 unsigned Offset = 0) {
253 return get(BB, AvailableValue::get(V, Offset));
254 }
255
getUndefllvm::gvn::AvailableValueInBlock256 static AvailableValueInBlock getUndef(BasicBlock *BB) {
257 return get(BB, AvailableValue::getUndef());
258 }
259
260 /// Emit code at the end of this block to adjust the value defined here to
261 /// the specified type. This handles various coercion cases.
MaterializeAdjustedValuellvm::gvn::AvailableValueInBlock262 Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const {
263 return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
264 }
265 };
266
267 //===----------------------------------------------------------------------===//
268 // ValueTable Internal Functions
269 //===----------------------------------------------------------------------===//
270
createExpr(Instruction * I)271 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
272 Expression e;
273 e.type = I->getType();
274 e.opcode = I->getOpcode();
275 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
276 OI != OE; ++OI)
277 e.varargs.push_back(lookupOrAdd(*OI));
278 if (I->isCommutative()) {
279 // Ensure that commutative instructions that only differ by a permutation
280 // of their operands get the same value number by sorting the operand value
281 // numbers. Since all commutative instructions have two operands it is more
282 // efficient to sort by hand rather than using, say, std::sort.
283 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
284 if (e.varargs[0] > e.varargs[1])
285 std::swap(e.varargs[0], e.varargs[1]);
286 e.commutative = true;
287 }
288
289 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
290 // Sort the operand value numbers so x<y and y>x get the same value number.
291 CmpInst::Predicate Predicate = C->getPredicate();
292 if (e.varargs[0] > e.varargs[1]) {
293 std::swap(e.varargs[0], e.varargs[1]);
294 Predicate = CmpInst::getSwappedPredicate(Predicate);
295 }
296 e.opcode = (C->getOpcode() << 8) | Predicate;
297 e.commutative = true;
298 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
299 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
300 II != IE; ++II)
301 e.varargs.push_back(*II);
302 }
303
304 return e;
305 }
306
createCmpExpr(unsigned Opcode,CmpInst::Predicate Predicate,Value * LHS,Value * RHS)307 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
308 CmpInst::Predicate Predicate,
309 Value *LHS, Value *RHS) {
310 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
311 "Not a comparison!");
312 Expression e;
313 e.type = CmpInst::makeCmpResultType(LHS->getType());
314 e.varargs.push_back(lookupOrAdd(LHS));
315 e.varargs.push_back(lookupOrAdd(RHS));
316
317 // Sort the operand value numbers so x<y and y>x get the same value number.
318 if (e.varargs[0] > e.varargs[1]) {
319 std::swap(e.varargs[0], e.varargs[1]);
320 Predicate = CmpInst::getSwappedPredicate(Predicate);
321 }
322 e.opcode = (Opcode << 8) | Predicate;
323 e.commutative = true;
324 return e;
325 }
326
createExtractvalueExpr(ExtractValueInst * EI)327 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
328 assert(EI && "Not an ExtractValueInst?");
329 Expression e;
330 e.type = EI->getType();
331 e.opcode = 0;
332
333 WithOverflowInst *WO = dyn_cast<WithOverflowInst>(EI->getAggregateOperand());
334 if (WO != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
335 // EI is an extract from one of our with.overflow intrinsics. Synthesize
336 // a semantically equivalent expression instead of an extract value
337 // expression.
338 e.opcode = WO->getBinaryOp();
339 e.varargs.push_back(lookupOrAdd(WO->getLHS()));
340 e.varargs.push_back(lookupOrAdd(WO->getRHS()));
341 return e;
342 }
343
344 // Not a recognised intrinsic. Fall back to producing an extract value
345 // expression.
346 e.opcode = EI->getOpcode();
347 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
348 OI != OE; ++OI)
349 e.varargs.push_back(lookupOrAdd(*OI));
350
351 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
352 II != IE; ++II)
353 e.varargs.push_back(*II);
354
355 return e;
356 }
357
358 //===----------------------------------------------------------------------===//
359 // ValueTable External Functions
360 //===----------------------------------------------------------------------===//
361
362 GVN::ValueTable::ValueTable() = default;
363 GVN::ValueTable::ValueTable(const ValueTable &) = default;
364 GVN::ValueTable::ValueTable(ValueTable &&) = default;
365 GVN::ValueTable::~ValueTable() = default;
366
367 /// add - Insert a value into the table with a specified value number.
add(Value * V,uint32_t num)368 void GVN::ValueTable::add(Value *V, uint32_t num) {
369 valueNumbering.insert(std::make_pair(V, num));
370 if (PHINode *PN = dyn_cast<PHINode>(V))
371 NumberingPhi[num] = PN;
372 }
373
lookupOrAddCall(CallInst * C)374 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
375 if (AA->doesNotAccessMemory(C)) {
376 Expression exp = createExpr(C);
377 uint32_t e = assignExpNewValueNum(exp).first;
378 valueNumbering[C] = e;
379 return e;
380 } else if (MD && AA->onlyReadsMemory(C)) {
381 Expression exp = createExpr(C);
382 auto ValNum = assignExpNewValueNum(exp);
383 if (ValNum.second) {
384 valueNumbering[C] = ValNum.first;
385 return ValNum.first;
386 }
387
388 MemDepResult local_dep = MD->getDependency(C);
389
390 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
391 valueNumbering[C] = nextValueNumber;
392 return nextValueNumber++;
393 }
394
395 if (local_dep.isDef()) {
396 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
397
398 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
399 valueNumbering[C] = nextValueNumber;
400 return nextValueNumber++;
401 }
402
403 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
404 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
405 uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
406 if (c_vn != cd_vn) {
407 valueNumbering[C] = nextValueNumber;
408 return nextValueNumber++;
409 }
410 }
411
412 uint32_t v = lookupOrAdd(local_cdep);
413 valueNumbering[C] = v;
414 return v;
415 }
416
417 // Non-local case.
418 const MemoryDependenceResults::NonLocalDepInfo &deps =
419 MD->getNonLocalCallDependency(C);
420 // FIXME: Move the checking logic to MemDep!
421 CallInst* cdep = nullptr;
422
423 // Check to see if we have a single dominating call instruction that is
424 // identical to C.
425 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
426 const NonLocalDepEntry *I = &deps[i];
427 if (I->getResult().isNonLocal())
428 continue;
429
430 // We don't handle non-definitions. If we already have a call, reject
431 // instruction dependencies.
432 if (!I->getResult().isDef() || cdep != nullptr) {
433 cdep = nullptr;
434 break;
435 }
436
437 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
438 // FIXME: All duplicated with non-local case.
439 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
440 cdep = NonLocalDepCall;
441 continue;
442 }
443
444 cdep = nullptr;
445 break;
446 }
447
448 if (!cdep) {
449 valueNumbering[C] = nextValueNumber;
450 return nextValueNumber++;
451 }
452
453 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
454 valueNumbering[C] = nextValueNumber;
455 return nextValueNumber++;
456 }
457 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
458 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
459 uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
460 if (c_vn != cd_vn) {
461 valueNumbering[C] = nextValueNumber;
462 return nextValueNumber++;
463 }
464 }
465
466 uint32_t v = lookupOrAdd(cdep);
467 valueNumbering[C] = v;
468 return v;
469 } else {
470 valueNumbering[C] = nextValueNumber;
471 return nextValueNumber++;
472 }
473 }
474
475 /// Returns true if a value number exists for the specified value.
exists(Value * V) const476 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
477
478 /// lookup_or_add - Returns the value number for the specified value, assigning
479 /// it a new number if it did not have one before.
lookupOrAdd(Value * V)480 uint32_t GVN::ValueTable::lookupOrAdd(Value *V) {
481 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
482 if (VI != valueNumbering.end())
483 return VI->second;
484
485 if (!isa<Instruction>(V)) {
486 valueNumbering[V] = nextValueNumber;
487 return nextValueNumber++;
488 }
489
490 Instruction* I = cast<Instruction>(V);
491 Expression exp;
492 switch (I->getOpcode()) {
493 case Instruction::Call:
494 return lookupOrAddCall(cast<CallInst>(I));
495 case Instruction::FNeg:
496 case Instruction::Add:
497 case Instruction::FAdd:
498 case Instruction::Sub:
499 case Instruction::FSub:
500 case Instruction::Mul:
501 case Instruction::FMul:
502 case Instruction::UDiv:
503 case Instruction::SDiv:
504 case Instruction::FDiv:
505 case Instruction::URem:
506 case Instruction::SRem:
507 case Instruction::FRem:
508 case Instruction::Shl:
509 case Instruction::LShr:
510 case Instruction::AShr:
511 case Instruction::And:
512 case Instruction::Or:
513 case Instruction::Xor:
514 case Instruction::ICmp:
515 case Instruction::FCmp:
516 case Instruction::Trunc:
517 case Instruction::ZExt:
518 case Instruction::SExt:
519 case Instruction::FPToUI:
520 case Instruction::FPToSI:
521 case Instruction::UIToFP:
522 case Instruction::SIToFP:
523 case Instruction::FPTrunc:
524 case Instruction::FPExt:
525 case Instruction::PtrToInt:
526 case Instruction::IntToPtr:
527 case Instruction::AddrSpaceCast:
528 case Instruction::BitCast:
529 case Instruction::Select:
530 case Instruction::ExtractElement:
531 case Instruction::InsertElement:
532 case Instruction::ShuffleVector:
533 case Instruction::InsertValue:
534 case Instruction::GetElementPtr:
535 exp = createExpr(I);
536 break;
537 case Instruction::ExtractValue:
538 exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
539 break;
540 case Instruction::PHI:
541 valueNumbering[V] = nextValueNumber;
542 NumberingPhi[nextValueNumber] = cast<PHINode>(V);
543 return nextValueNumber++;
544 default:
545 valueNumbering[V] = nextValueNumber;
546 return nextValueNumber++;
547 }
548
549 uint32_t e = assignExpNewValueNum(exp).first;
550 valueNumbering[V] = e;
551 return e;
552 }
553
554 /// Returns the value number of the specified value. Fails if
555 /// the value has not yet been numbered.
lookup(Value * V,bool Verify) const556 uint32_t GVN::ValueTable::lookup(Value *V, bool Verify) const {
557 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
558 if (Verify) {
559 assert(VI != valueNumbering.end() && "Value not numbered?");
560 return VI->second;
561 }
562 return (VI != valueNumbering.end()) ? VI->second : 0;
563 }
564
565 /// Returns the value number of the given comparison,
566 /// assigning it a new number if it did not have one before. Useful when
567 /// we deduced the result of a comparison, but don't immediately have an
568 /// instruction realizing that comparison to hand.
lookupOrAddCmp(unsigned Opcode,CmpInst::Predicate Predicate,Value * LHS,Value * RHS)569 uint32_t GVN::ValueTable::lookupOrAddCmp(unsigned Opcode,
570 CmpInst::Predicate Predicate,
571 Value *LHS, Value *RHS) {
572 Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
573 return assignExpNewValueNum(exp).first;
574 }
575
576 /// Remove all entries from the ValueTable.
clear()577 void GVN::ValueTable::clear() {
578 valueNumbering.clear();
579 expressionNumbering.clear();
580 NumberingPhi.clear();
581 PhiTranslateTable.clear();
582 nextValueNumber = 1;
583 Expressions.clear();
584 ExprIdx.clear();
585 nextExprNumber = 0;
586 }
587
588 /// Remove a value from the value numbering.
erase(Value * V)589 void GVN::ValueTable::erase(Value *V) {
590 uint32_t Num = valueNumbering.lookup(V);
591 valueNumbering.erase(V);
592 // If V is PHINode, V <--> value number is an one-to-one mapping.
593 if (isa<PHINode>(V))
594 NumberingPhi.erase(Num);
595 }
596
597 /// verifyRemoved - Verify that the value is removed from all internal data
598 /// structures.
verifyRemoved(const Value * V) const599 void GVN::ValueTable::verifyRemoved(const Value *V) const {
600 for (DenseMap<Value*, uint32_t>::const_iterator
601 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
602 assert(I->first != V && "Inst still occurs in value numbering map!");
603 }
604 }
605
606 //===----------------------------------------------------------------------===//
607 // GVN Pass
608 //===----------------------------------------------------------------------===//
609
run(Function & F,FunctionAnalysisManager & AM)610 PreservedAnalyses GVN::run(Function &F, FunctionAnalysisManager &AM) {
611 // FIXME: The order of evaluation of these 'getResult' calls is very
612 // significant! Re-ordering these variables will cause GVN when run alone to
613 // be less effective! We should fix memdep and basic-aa to not exhibit this
614 // behavior, but until then don't change the order here.
615 auto &AC = AM.getResult<AssumptionAnalysis>(F);
616 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
617 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
618 auto &AA = AM.getResult<AAManager>(F);
619 auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
620 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
621 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
622 bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
623 if (!Changed)
624 return PreservedAnalyses::all();
625 PreservedAnalyses PA;
626 PA.preserve<DominatorTreeAnalysis>();
627 PA.preserve<GlobalsAA>();
628 PA.preserve<TargetLibraryAnalysis>();
629 return PA;
630 }
631
632 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump(DenseMap<uint32_t,Value * > & d) const633 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) const {
634 errs() << "{\n";
635 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
636 E = d.end(); I != E; ++I) {
637 errs() << I->first << "\n";
638 I->second->dump();
639 }
640 errs() << "}\n";
641 }
642 #endif
643
644 /// Return true if we can prove that the value
645 /// we're analyzing is fully available in the specified block. As we go, keep
646 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
647 /// map is actually a tri-state map with the following values:
648 /// 0) we know the block *is not* fully available.
649 /// 1) we know the block *is* fully available.
650 /// 2) we do not know whether the block is fully available or not, but we are
651 /// currently speculating that it will be.
652 /// 3) we are speculating for this block and have used that to speculate for
653 /// other blocks.
IsValueFullyAvailableInBlock(BasicBlock * BB,DenseMap<BasicBlock *,char> & FullyAvailableBlocks,uint32_t RecurseDepth)654 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
655 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
656 uint32_t RecurseDepth) {
657 if (RecurseDepth > MaxRecurseDepth)
658 return false;
659
660 // Optimistically assume that the block is fully available and check to see
661 // if we already know about this block in one lookup.
662 std::pair<DenseMap<BasicBlock*, char>::iterator, bool> IV =
663 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
664
665 // If the entry already existed for this block, return the precomputed value.
666 if (!IV.second) {
667 // If this is a speculative "available" value, mark it as being used for
668 // speculation of other blocks.
669 if (IV.first->second == 2)
670 IV.first->second = 3;
671 return IV.first->second != 0;
672 }
673
674 // Otherwise, see if it is fully available in all predecessors.
675 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
676
677 // If this block has no predecessors, it isn't live-in here.
678 if (PI == PE)
679 goto SpeculationFailure;
680
681 for (; PI != PE; ++PI)
682 // If the value isn't fully available in one of our predecessors, then it
683 // isn't fully available in this block either. Undo our previous
684 // optimistic assumption and bail out.
685 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
686 goto SpeculationFailure;
687
688 return true;
689
690 // If we get here, we found out that this is not, after
691 // all, a fully-available block. We have a problem if we speculated on this and
692 // used the speculation to mark other blocks as available.
693 SpeculationFailure:
694 char &BBVal = FullyAvailableBlocks[BB];
695
696 // If we didn't speculate on this, just return with it set to false.
697 if (BBVal == 2) {
698 BBVal = 0;
699 return false;
700 }
701
702 // If we did speculate on this value, we could have blocks set to 1 that are
703 // incorrect. Walk the (transitive) successors of this block and mark them as
704 // 0 if set to one.
705 SmallVector<BasicBlock*, 32> BBWorklist;
706 BBWorklist.push_back(BB);
707
708 do {
709 BasicBlock *Entry = BBWorklist.pop_back_val();
710 // Note that this sets blocks to 0 (unavailable) if they happen to not
711 // already be in FullyAvailableBlocks. This is safe.
712 char &EntryVal = FullyAvailableBlocks[Entry];
713 if (EntryVal == 0) continue; // Already unavailable.
714
715 // Mark as unavailable.
716 EntryVal = 0;
717
718 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
719 } while (!BBWorklist.empty());
720
721 return false;
722 }
723
724 /// Given a set of loads specified by ValuesPerBlock,
725 /// construct SSA form, allowing us to eliminate LI. This returns the value
726 /// that should be used at LI's definition site.
ConstructSSAForLoadSet(LoadInst * LI,SmallVectorImpl<AvailableValueInBlock> & ValuesPerBlock,GVN & gvn)727 static Value *ConstructSSAForLoadSet(LoadInst *LI,
728 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
729 GVN &gvn) {
730 // Check for the fully redundant, dominating load case. In this case, we can
731 // just use the dominating value directly.
732 if (ValuesPerBlock.size() == 1 &&
733 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
734 LI->getParent())) {
735 assert(!ValuesPerBlock[0].AV.isUndefValue() &&
736 "Dead BB dominate this block");
737 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
738 }
739
740 // Otherwise, we have to construct SSA form.
741 SmallVector<PHINode*, 8> NewPHIs;
742 SSAUpdater SSAUpdate(&NewPHIs);
743 SSAUpdate.Initialize(LI->getType(), LI->getName());
744
745 for (const AvailableValueInBlock &AV : ValuesPerBlock) {
746 BasicBlock *BB = AV.BB;
747
748 if (SSAUpdate.HasValueForBlock(BB))
749 continue;
750
751 // If the value is the load that we will be eliminating, and the block it's
752 // available in is the block that the load is in, then don't add it as
753 // SSAUpdater will resolve the value to the relevant phi which may let it
754 // avoid phi construction entirely if there's actually only one value.
755 if (BB == LI->getParent() &&
756 ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == LI) ||
757 (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == LI)))
758 continue;
759
760 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
761 }
762
763 // Perform PHI construction.
764 return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
765 }
766
MaterializeAdjustedValue(LoadInst * LI,Instruction * InsertPt,GVN & gvn) const767 Value *AvailableValue::MaterializeAdjustedValue(LoadInst *LI,
768 Instruction *InsertPt,
769 GVN &gvn) const {
770 Value *Res;
771 Type *LoadTy = LI->getType();
772 const DataLayout &DL = LI->getModule()->getDataLayout();
773 if (isSimpleValue()) {
774 Res = getSimpleValue();
775 if (Res->getType() != LoadTy) {
776 Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
777
778 LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset
779 << " " << *getSimpleValue() << '\n'
780 << *Res << '\n'
781 << "\n\n\n");
782 }
783 } else if (isCoercedLoadValue()) {
784 LoadInst *Load = getCoercedLoadValue();
785 if (Load->getType() == LoadTy && Offset == 0) {
786 Res = Load;
787 } else {
788 Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
789 // We would like to use gvn.markInstructionForDeletion here, but we can't
790 // because the load is already memoized into the leader map table that GVN
791 // tracks. It is potentially possible to remove the load from the table,
792 // but then there all of the operations based on it would need to be
793 // rehashed. Just leave the dead load around.
794 gvn.getMemDep().removeInstruction(Load);
795 LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset
796 << " " << *getCoercedLoadValue() << '\n'
797 << *Res << '\n'
798 << "\n\n\n");
799 }
800 } else if (isMemIntrinValue()) {
801 Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
802 InsertPt, DL);
803 LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
804 << " " << *getMemIntrinValue() << '\n'
805 << *Res << '\n'
806 << "\n\n\n");
807 } else {
808 assert(isUndefValue() && "Should be UndefVal");
809 LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
810 return UndefValue::get(LoadTy);
811 }
812 assert(Res && "failed to materialize?");
813 return Res;
814 }
815
isLifetimeStart(const Instruction * Inst)816 static bool isLifetimeStart(const Instruction *Inst) {
817 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
818 return II->getIntrinsicID() == Intrinsic::lifetime_start;
819 return false;
820 }
821
822 /// Try to locate the three instruction involved in a missed
823 /// load-elimination case that is due to an intervening store.
reportMayClobberedLoad(LoadInst * LI,MemDepResult DepInfo,DominatorTree * DT,OptimizationRemarkEmitter * ORE)824 static void reportMayClobberedLoad(LoadInst *LI, MemDepResult DepInfo,
825 DominatorTree *DT,
826 OptimizationRemarkEmitter *ORE) {
827 using namespace ore;
828
829 User *OtherAccess = nullptr;
830
831 OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
832 R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
833 << setExtraArgs();
834
835 for (auto *U : LI->getPointerOperand()->users())
836 if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
837 DT->dominates(cast<Instruction>(U), LI)) {
838 // FIXME: for now give up if there are multiple memory accesses that
839 // dominate the load. We need further analysis to decide which one is
840 // that we're forwarding from.
841 if (OtherAccess)
842 OtherAccess = nullptr;
843 else
844 OtherAccess = U;
845 }
846
847 if (OtherAccess)
848 R << " in favor of " << NV("OtherAccess", OtherAccess);
849
850 R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
851
852 ORE->emit(R);
853 }
854
AnalyzeLoadAvailability(LoadInst * LI,MemDepResult DepInfo,Value * Address,AvailableValue & Res)855 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
856 Value *Address, AvailableValue &Res) {
857 assert((DepInfo.isDef() || DepInfo.isClobber()) &&
858 "expected a local dependence");
859 assert(LI->isUnordered() && "rules below are incorrect for ordered access");
860
861 const DataLayout &DL = LI->getModule()->getDataLayout();
862
863 Instruction *DepInst = DepInfo.getInst();
864 if (DepInfo.isClobber()) {
865 // If the dependence is to a store that writes to a superset of the bits
866 // read by the load, we can extract the bits we need for the load from the
867 // stored value.
868 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
869 // Can't forward from non-atomic to atomic without violating memory model.
870 if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
871 int Offset =
872 analyzeLoadFromClobberingStore(LI->getType(), Address, DepSI, DL);
873 if (Offset != -1) {
874 Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
875 return true;
876 }
877 }
878 }
879
880 // Check to see if we have something like this:
881 // load i32* P
882 // load i8* (P+1)
883 // if we have this, replace the later with an extraction from the former.
884 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
885 // If this is a clobber and L is the first instruction in its block, then
886 // we have the first instruction in the entry block.
887 // Can't forward from non-atomic to atomic without violating memory model.
888 if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
889 int Offset =
890 analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
891
892 if (Offset != -1) {
893 Res = AvailableValue::getLoad(DepLI, Offset);
894 return true;
895 }
896 }
897 }
898
899 // If the clobbering value is a memset/memcpy/memmove, see if we can
900 // forward a value on from it.
901 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
902 if (Address && !LI->isAtomic()) {
903 int Offset = analyzeLoadFromClobberingMemInst(LI->getType(), Address,
904 DepMI, DL);
905 if (Offset != -1) {
906 Res = AvailableValue::getMI(DepMI, Offset);
907 return true;
908 }
909 }
910 }
911 // Nothing known about this clobber, have to be conservative
912 LLVM_DEBUG(
913 // fast print dep, using operator<< on instruction is too slow.
914 dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
915 dbgs() << " is clobbered by " << *DepInst << '\n';);
916 if (ORE->allowExtraAnalysis(DEBUG_TYPE))
917 reportMayClobberedLoad(LI, DepInfo, DT, ORE);
918
919 return false;
920 }
921 assert(DepInfo.isDef() && "follows from above");
922
923 // Loading the allocation -> undef.
924 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
925 // Loading immediately after lifetime begin -> undef.
926 isLifetimeStart(DepInst)) {
927 Res = AvailableValue::get(UndefValue::get(LI->getType()));
928 return true;
929 }
930
931 // Loading from calloc (which zero initializes memory) -> zero
932 if (isCallocLikeFn(DepInst, TLI)) {
933 Res = AvailableValue::get(Constant::getNullValue(LI->getType()));
934 return true;
935 }
936
937 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
938 // Reject loads and stores that are to the same address but are of
939 // different types if we have to. If the stored value is larger or equal to
940 // the loaded value, we can reuse it.
941 if (!canCoerceMustAliasedValueToLoad(S->getValueOperand(), LI->getType(),
942 DL))
943 return false;
944
945 // Can't forward from non-atomic to atomic without violating memory model.
946 if (S->isAtomic() < LI->isAtomic())
947 return false;
948
949 Res = AvailableValue::get(S->getValueOperand());
950 return true;
951 }
952
953 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
954 // If the types mismatch and we can't handle it, reject reuse of the load.
955 // If the stored value is larger or equal to the loaded value, we can reuse
956 // it.
957 if (!canCoerceMustAliasedValueToLoad(LD, LI->getType(), DL))
958 return false;
959
960 // Can't forward from non-atomic to atomic without violating memory model.
961 if (LD->isAtomic() < LI->isAtomic())
962 return false;
963
964 Res = AvailableValue::getLoad(LD);
965 return true;
966 }
967
968 // Unknown def - must be conservative
969 LLVM_DEBUG(
970 // fast print dep, using operator<< on instruction is too slow.
971 dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
972 dbgs() << " has unknown def " << *DepInst << '\n';);
973 return false;
974 }
975
AnalyzeLoadAvailability(LoadInst * LI,LoadDepVect & Deps,AvailValInBlkVect & ValuesPerBlock,UnavailBlkVect & UnavailableBlocks)976 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
977 AvailValInBlkVect &ValuesPerBlock,
978 UnavailBlkVect &UnavailableBlocks) {
979 // Filter out useless results (non-locals, etc). Keep track of the blocks
980 // where we have a value available in repl, also keep track of whether we see
981 // dependencies that produce an unknown value for the load (such as a call
982 // that could potentially clobber the load).
983 unsigned NumDeps = Deps.size();
984 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
985 BasicBlock *DepBB = Deps[i].getBB();
986 MemDepResult DepInfo = Deps[i].getResult();
987
988 if (DeadBlocks.count(DepBB)) {
989 // Dead dependent mem-op disguise as a load evaluating the same value
990 // as the load in question.
991 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
992 continue;
993 }
994
995 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
996 UnavailableBlocks.push_back(DepBB);
997 continue;
998 }
999
1000 // The address being loaded in this non-local block may not be the same as
1001 // the pointer operand of the load if PHI translation occurs. Make sure
1002 // to consider the right address.
1003 Value *Address = Deps[i].getAddress();
1004
1005 AvailableValue AV;
1006 if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1007 // subtlety: because we know this was a non-local dependency, we know
1008 // it's safe to materialize anywhere between the instruction within
1009 // DepInfo and the end of it's block.
1010 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1011 std::move(AV)));
1012 } else {
1013 UnavailableBlocks.push_back(DepBB);
1014 }
1015 }
1016
1017 assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1018 "post condition violation");
1019 }
1020
PerformLoadPRE(LoadInst * LI,AvailValInBlkVect & ValuesPerBlock,UnavailBlkVect & UnavailableBlocks)1021 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1022 UnavailBlkVect &UnavailableBlocks) {
1023 // Okay, we have *some* definitions of the value. This means that the value
1024 // is available in some of our (transitive) predecessors. Lets think about
1025 // doing PRE of this load. This will involve inserting a new load into the
1026 // predecessor when it's not available. We could do this in general, but
1027 // prefer to not increase code size. As such, we only do this when we know
1028 // that we only have to insert *one* load (which means we're basically moving
1029 // the load, not inserting a new one).
1030
1031 SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1032 UnavailableBlocks.end());
1033
1034 // Let's find the first basic block with more than one predecessor. Walk
1035 // backwards through predecessors if needed.
1036 BasicBlock *LoadBB = LI->getParent();
1037 BasicBlock *TmpBB = LoadBB;
1038 bool IsSafeToSpeculativelyExecute = isSafeToSpeculativelyExecute(LI);
1039
1040 // Check that there is no implicit control flow instructions above our load in
1041 // its block. If there is an instruction that doesn't always pass the
1042 // execution to the following instruction, then moving through it may become
1043 // invalid. For example:
1044 //
1045 // int arr[LEN];
1046 // int index = ???;
1047 // ...
1048 // guard(0 <= index && index < LEN);
1049 // use(arr[index]);
1050 //
1051 // It is illegal to move the array access to any point above the guard,
1052 // because if the index is out of bounds we should deoptimize rather than
1053 // access the array.
1054 // Check that there is no guard in this block above our instruction.
1055 if (!IsSafeToSpeculativelyExecute && ICF->isDominatedByICFIFromSameBlock(LI))
1056 return false;
1057 while (TmpBB->getSinglePredecessor()) {
1058 TmpBB = TmpBB->getSinglePredecessor();
1059 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1060 return false;
1061 if (Blockers.count(TmpBB))
1062 return false;
1063
1064 // If any of these blocks has more than one successor (i.e. if the edge we
1065 // just traversed was critical), then there are other paths through this
1066 // block along which the load may not be anticipated. Hoisting the load
1067 // above this block would be adding the load to execution paths along
1068 // which it was not previously executed.
1069 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1070 return false;
1071
1072 // Check that there is no implicit control flow in a block above.
1073 if (!IsSafeToSpeculativelyExecute && ICF->hasICF(TmpBB))
1074 return false;
1075 }
1076
1077 assert(TmpBB);
1078 LoadBB = TmpBB;
1079
1080 // Check to see how many predecessors have the loaded value fully
1081 // available.
1082 MapVector<BasicBlock *, Value *> PredLoads;
1083 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1084 for (const AvailableValueInBlock &AV : ValuesPerBlock)
1085 FullyAvailableBlocks[AV.BB] = true;
1086 for (BasicBlock *UnavailableBB : UnavailableBlocks)
1087 FullyAvailableBlocks[UnavailableBB] = false;
1088
1089 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1090 for (BasicBlock *Pred : predecessors(LoadBB)) {
1091 // If any predecessor block is an EH pad that does not allow non-PHI
1092 // instructions before the terminator, we can't PRE the load.
1093 if (Pred->getTerminator()->isEHPad()) {
1094 LLVM_DEBUG(
1095 dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1096 << Pred->getName() << "': " << *LI << '\n');
1097 return false;
1098 }
1099
1100 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1101 continue;
1102 }
1103
1104 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1105 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1106 LLVM_DEBUG(
1107 dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1108 << Pred->getName() << "': " << *LI << '\n');
1109 return false;
1110 }
1111
1112 // FIXME: Can we support the fallthrough edge?
1113 if (isa<CallBrInst>(Pred->getTerminator())) {
1114 LLVM_DEBUG(
1115 dbgs() << "COULD NOT PRE LOAD BECAUSE OF CALLBR CRITICAL EDGE '"
1116 << Pred->getName() << "': " << *LI << '\n');
1117 return false;
1118 }
1119
1120 if (LoadBB->isEHPad()) {
1121 LLVM_DEBUG(
1122 dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1123 << Pred->getName() << "': " << *LI << '\n');
1124 return false;
1125 }
1126
1127 CriticalEdgePred.push_back(Pred);
1128 } else {
1129 // Only add the predecessors that will not be split for now.
1130 PredLoads[Pred] = nullptr;
1131 }
1132 }
1133
1134 // Decide whether PRE is profitable for this load.
1135 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1136 assert(NumUnavailablePreds != 0 &&
1137 "Fully available value should already be eliminated!");
1138
1139 // If this load is unavailable in multiple predecessors, reject it.
1140 // FIXME: If we could restructure the CFG, we could make a common pred with
1141 // all the preds that don't have an available LI and insert a new load into
1142 // that one block.
1143 if (NumUnavailablePreds != 1)
1144 return false;
1145
1146 // Split critical edges, and update the unavailable predecessors accordingly.
1147 for (BasicBlock *OrigPred : CriticalEdgePred) {
1148 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1149 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1150 PredLoads[NewPred] = nullptr;
1151 LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1152 << LoadBB->getName() << '\n');
1153 }
1154
1155 // Check if the load can safely be moved to all the unavailable predecessors.
1156 bool CanDoPRE = true;
1157 const DataLayout &DL = LI->getModule()->getDataLayout();
1158 SmallVector<Instruction*, 8> NewInsts;
1159 for (auto &PredLoad : PredLoads) {
1160 BasicBlock *UnavailablePred = PredLoad.first;
1161
1162 // Do PHI translation to get its value in the predecessor if necessary. The
1163 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1164
1165 // If all preds have a single successor, then we know it is safe to insert
1166 // the load on the pred (?!?), so we can insert code to materialize the
1167 // pointer if it is not available.
1168 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1169 Value *LoadPtr = nullptr;
1170 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1171 *DT, NewInsts);
1172
1173 // If we couldn't find or insert a computation of this phi translated value,
1174 // we fail PRE.
1175 if (!LoadPtr) {
1176 LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1177 << *LI->getPointerOperand() << "\n");
1178 CanDoPRE = false;
1179 break;
1180 }
1181
1182 PredLoad.second = LoadPtr;
1183 }
1184
1185 if (!CanDoPRE) {
1186 while (!NewInsts.empty()) {
1187 Instruction *I = NewInsts.pop_back_val();
1188 markInstructionForDeletion(I);
1189 }
1190 // HINT: Don't revert the edge-splitting as following transformation may
1191 // also need to split these critical edges.
1192 return !CriticalEdgePred.empty();
1193 }
1194
1195 // Okay, we can eliminate this load by inserting a reload in the predecessor
1196 // and using PHI construction to get the value in the other predecessors, do
1197 // it.
1198 LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1199 LLVM_DEBUG(if (!NewInsts.empty()) dbgs()
1200 << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts.back()
1201 << '\n');
1202
1203 // Assign value numbers to the new instructions.
1204 for (Instruction *I : NewInsts) {
1205 // Instructions that have been inserted in predecessor(s) to materialize
1206 // the load address do not retain their original debug locations. Doing
1207 // so could lead to confusing (but correct) source attributions.
1208 if (const DebugLoc &DL = I->getDebugLoc())
1209 I->setDebugLoc(DebugLoc::get(0, 0, DL.getScope(), DL.getInlinedAt()));
1210
1211 // FIXME: We really _ought_ to insert these value numbers into their
1212 // parent's availability map. However, in doing so, we risk getting into
1213 // ordering issues. If a block hasn't been processed yet, we would be
1214 // marking a value as AVAIL-IN, which isn't what we intend.
1215 VN.lookupOrAdd(I);
1216 }
1217
1218 for (const auto &PredLoad : PredLoads) {
1219 BasicBlock *UnavailablePred = PredLoad.first;
1220 Value *LoadPtr = PredLoad.second;
1221
1222 auto *NewLoad =
1223 new LoadInst(LI->getType(), LoadPtr, LI->getName() + ".pre",
1224 LI->isVolatile(), LI->getAlignment(), LI->getOrdering(),
1225 LI->getSyncScopeID(), UnavailablePred->getTerminator());
1226 NewLoad->setDebugLoc(LI->getDebugLoc());
1227
1228 // Transfer the old load's AA tags to the new load.
1229 AAMDNodes Tags;
1230 LI->getAAMetadata(Tags);
1231 if (Tags)
1232 NewLoad->setAAMetadata(Tags);
1233
1234 if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1235 NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1236 if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1237 NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1238 if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1239 NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1240
1241 // We do not propagate the old load's debug location, because the new
1242 // load now lives in a different BB, and we want to avoid a jumpy line
1243 // table.
1244 // FIXME: How do we retain source locations without causing poor debugging
1245 // behavior?
1246
1247 // Add the newly created load.
1248 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1249 NewLoad));
1250 MD->invalidateCachedPointerInfo(LoadPtr);
1251 LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1252 }
1253
1254 // Perform PHI construction.
1255 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1256 LI->replaceAllUsesWith(V);
1257 if (isa<PHINode>(V))
1258 V->takeName(LI);
1259 if (Instruction *I = dyn_cast<Instruction>(V))
1260 I->setDebugLoc(LI->getDebugLoc());
1261 if (V->getType()->isPtrOrPtrVectorTy())
1262 MD->invalidateCachedPointerInfo(V);
1263 markInstructionForDeletion(LI);
1264 ORE->emit([&]() {
1265 return OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1266 << "load eliminated by PRE";
1267 });
1268 ++NumPRELoad;
1269 return true;
1270 }
1271
reportLoadElim(LoadInst * LI,Value * AvailableValue,OptimizationRemarkEmitter * ORE)1272 static void reportLoadElim(LoadInst *LI, Value *AvailableValue,
1273 OptimizationRemarkEmitter *ORE) {
1274 using namespace ore;
1275
1276 ORE->emit([&]() {
1277 return OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1278 << "load of type " << NV("Type", LI->getType()) << " eliminated"
1279 << setExtraArgs() << " in favor of "
1280 << NV("InfavorOfValue", AvailableValue);
1281 });
1282 }
1283
1284 /// Attempt to eliminate a load whose dependencies are
1285 /// non-local by performing PHI construction.
processNonLocalLoad(LoadInst * LI)1286 bool GVN::processNonLocalLoad(LoadInst *LI) {
1287 // non-local speculations are not allowed under asan.
1288 if (LI->getParent()->getParent()->hasFnAttribute(
1289 Attribute::SanitizeAddress) ||
1290 LI->getParent()->getParent()->hasFnAttribute(
1291 Attribute::SanitizeHWAddress))
1292 return false;
1293
1294 // Step 1: Find the non-local dependencies of the load.
1295 LoadDepVect Deps;
1296 MD->getNonLocalPointerDependency(LI, Deps);
1297
1298 // If we had to process more than one hundred blocks to find the
1299 // dependencies, this load isn't worth worrying about. Optimizing
1300 // it will be too expensive.
1301 unsigned NumDeps = Deps.size();
1302 if (NumDeps > MaxNumDeps)
1303 return false;
1304
1305 // If we had a phi translation failure, we'll have a single entry which is a
1306 // clobber in the current block. Reject this early.
1307 if (NumDeps == 1 &&
1308 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1309 LLVM_DEBUG(dbgs() << "GVN: non-local load "; LI->printAsOperand(dbgs());
1310 dbgs() << " has unknown dependencies\n";);
1311 return false;
1312 }
1313
1314 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1315 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1316 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1317 OE = GEP->idx_end();
1318 OI != OE; ++OI)
1319 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1320 performScalarPRE(I);
1321 }
1322
1323 // Step 2: Analyze the availability of the load
1324 AvailValInBlkVect ValuesPerBlock;
1325 UnavailBlkVect UnavailableBlocks;
1326 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1327
1328 // If we have no predecessors that produce a known value for this load, exit
1329 // early.
1330 if (ValuesPerBlock.empty())
1331 return false;
1332
1333 // Step 3: Eliminate fully redundancy.
1334 //
1335 // If all of the instructions we depend on produce a known value for this
1336 // load, then it is fully redundant and we can use PHI insertion to compute
1337 // its value. Insert PHIs and remove the fully redundant value now.
1338 if (UnavailableBlocks.empty()) {
1339 LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1340
1341 // Perform PHI construction.
1342 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1343 LI->replaceAllUsesWith(V);
1344
1345 if (isa<PHINode>(V))
1346 V->takeName(LI);
1347 if (Instruction *I = dyn_cast<Instruction>(V))
1348 // If instruction I has debug info, then we should not update it.
1349 // Also, if I has a null DebugLoc, then it is still potentially incorrect
1350 // to propagate LI's DebugLoc because LI may not post-dominate I.
1351 if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1352 I->setDebugLoc(LI->getDebugLoc());
1353 if (V->getType()->isPtrOrPtrVectorTy())
1354 MD->invalidateCachedPointerInfo(V);
1355 markInstructionForDeletion(LI);
1356 ++NumGVNLoad;
1357 reportLoadElim(LI, V, ORE);
1358 return true;
1359 }
1360
1361 // Step 4: Eliminate partial redundancy.
1362 if (!EnablePRE || !EnableLoadPRE)
1363 return false;
1364
1365 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1366 }
1367
processAssumeIntrinsic(IntrinsicInst * IntrinsicI)1368 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1369 assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1370 "This function can only be called with llvm.assume intrinsic");
1371 Value *V = IntrinsicI->getArgOperand(0);
1372
1373 if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1374 if (Cond->isZero()) {
1375 Type *Int8Ty = Type::getInt8Ty(V->getContext());
1376 // Insert a new store to null instruction before the load to indicate that
1377 // this code is not reachable. FIXME: We could insert unreachable
1378 // instruction directly because we can modify the CFG.
1379 new StoreInst(UndefValue::get(Int8Ty),
1380 Constant::getNullValue(Int8Ty->getPointerTo()),
1381 IntrinsicI);
1382 }
1383 markInstructionForDeletion(IntrinsicI);
1384 return false;
1385 } else if (isa<Constant>(V)) {
1386 // If it's not false, and constant, it must evaluate to true. This means our
1387 // assume is assume(true), and thus, pointless, and we don't want to do
1388 // anything more here.
1389 return false;
1390 }
1391
1392 Constant *True = ConstantInt::getTrue(V->getContext());
1393 bool Changed = false;
1394
1395 for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1396 BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1397
1398 // This property is only true in dominated successors, propagateEquality
1399 // will check dominance for us.
1400 Changed |= propagateEquality(V, True, Edge, false);
1401 }
1402
1403 // We can replace assume value with true, which covers cases like this:
1404 // call void @llvm.assume(i1 %cmp)
1405 // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1406 ReplaceWithConstMap[V] = True;
1407
1408 // If one of *cmp *eq operand is const, adding it to map will cover this:
1409 // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1410 // call void @llvm.assume(i1 %cmp)
1411 // ret float %0 ; will change it to ret float 3.000000e+00
1412 if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1413 if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1414 CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1415 (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1416 CmpI->getFastMathFlags().noNaNs())) {
1417 Value *CmpLHS = CmpI->getOperand(0);
1418 Value *CmpRHS = CmpI->getOperand(1);
1419 if (isa<Constant>(CmpLHS))
1420 std::swap(CmpLHS, CmpRHS);
1421 auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1422
1423 // If only one operand is constant.
1424 if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1425 ReplaceWithConstMap[CmpLHS] = RHSConst;
1426 }
1427 }
1428 return Changed;
1429 }
1430
patchAndReplaceAllUsesWith(Instruction * I,Value * Repl)1431 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1432 patchReplacementInstruction(I, Repl);
1433 I->replaceAllUsesWith(Repl);
1434 }
1435
1436 /// Attempt to eliminate a load, first by eliminating it
1437 /// locally, and then attempting non-local elimination if that fails.
processLoad(LoadInst * L)1438 bool GVN::processLoad(LoadInst *L) {
1439 if (!MD)
1440 return false;
1441
1442 // This code hasn't been audited for ordered or volatile memory access
1443 if (!L->isUnordered())
1444 return false;
1445
1446 if (L->use_empty()) {
1447 markInstructionForDeletion(L);
1448 return true;
1449 }
1450
1451 // ... to a pointer that has been loaded from before...
1452 MemDepResult Dep = MD->getDependency(L);
1453
1454 // If it is defined in another block, try harder.
1455 if (Dep.isNonLocal())
1456 return processNonLocalLoad(L);
1457
1458 // Only handle the local case below
1459 if (!Dep.isDef() && !Dep.isClobber()) {
1460 // This might be a NonFuncLocal or an Unknown
1461 LLVM_DEBUG(
1462 // fast print dep, using operator<< on instruction is too slow.
1463 dbgs() << "GVN: load "; L->printAsOperand(dbgs());
1464 dbgs() << " has unknown dependence\n";);
1465 return false;
1466 }
1467
1468 AvailableValue AV;
1469 if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1470 Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1471
1472 // Replace the load!
1473 patchAndReplaceAllUsesWith(L, AvailableValue);
1474 markInstructionForDeletion(L);
1475 ++NumGVNLoad;
1476 reportLoadElim(L, AvailableValue, ORE);
1477 // Tell MDA to rexamine the reused pointer since we might have more
1478 // information after forwarding it.
1479 if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1480 MD->invalidateCachedPointerInfo(AvailableValue);
1481 return true;
1482 }
1483
1484 return false;
1485 }
1486
1487 /// Return a pair the first field showing the value number of \p Exp and the
1488 /// second field showing whether it is a value number newly created.
1489 std::pair<uint32_t, bool>
assignExpNewValueNum(Expression & Exp)1490 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1491 uint32_t &e = expressionNumbering[Exp];
1492 bool CreateNewValNum = !e;
1493 if (CreateNewValNum) {
1494 Expressions.push_back(Exp);
1495 if (ExprIdx.size() < nextValueNumber + 1)
1496 ExprIdx.resize(nextValueNumber * 2);
1497 e = nextValueNumber;
1498 ExprIdx[nextValueNumber++] = nextExprNumber++;
1499 }
1500 return {e, CreateNewValNum};
1501 }
1502
1503 /// Return whether all the values related with the same \p num are
1504 /// defined in \p BB.
areAllValsInBB(uint32_t Num,const BasicBlock * BB,GVN & Gvn)1505 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1506 GVN &Gvn) {
1507 LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1508 while (Vals && Vals->BB == BB)
1509 Vals = Vals->Next;
1510 return !Vals;
1511 }
1512
1513 /// Wrap phiTranslateImpl to provide caching functionality.
phiTranslate(const BasicBlock * Pred,const BasicBlock * PhiBlock,uint32_t Num,GVN & Gvn)1514 uint32_t GVN::ValueTable::phiTranslate(const BasicBlock *Pred,
1515 const BasicBlock *PhiBlock, uint32_t Num,
1516 GVN &Gvn) {
1517 auto FindRes = PhiTranslateTable.find({Num, Pred});
1518 if (FindRes != PhiTranslateTable.end())
1519 return FindRes->second;
1520 uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1521 PhiTranslateTable.insert({{Num, Pred}, NewNum});
1522 return NewNum;
1523 }
1524
1525 // Return true if the value number \p Num and NewNum have equal value.
1526 // Return false if the result is unknown.
areCallValsEqual(uint32_t Num,uint32_t NewNum,const BasicBlock * Pred,const BasicBlock * PhiBlock,GVN & Gvn)1527 bool GVN::ValueTable::areCallValsEqual(uint32_t Num, uint32_t NewNum,
1528 const BasicBlock *Pred,
1529 const BasicBlock *PhiBlock, GVN &Gvn) {
1530 CallInst *Call = nullptr;
1531 LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1532 while (Vals) {
1533 Call = dyn_cast<CallInst>(Vals->Val);
1534 if (Call && Call->getParent() == PhiBlock)
1535 break;
1536 Vals = Vals->Next;
1537 }
1538
1539 if (AA->doesNotAccessMemory(Call))
1540 return true;
1541
1542 if (!MD || !AA->onlyReadsMemory(Call))
1543 return false;
1544
1545 MemDepResult local_dep = MD->getDependency(Call);
1546 if (!local_dep.isNonLocal())
1547 return false;
1548
1549 const MemoryDependenceResults::NonLocalDepInfo &deps =
1550 MD->getNonLocalCallDependency(Call);
1551
1552 // Check to see if the Call has no function local clobber.
1553 for (unsigned i = 0; i < deps.size(); i++) {
1554 if (deps[i].getResult().isNonFuncLocal())
1555 return true;
1556 }
1557 return false;
1558 }
1559
1560 /// Translate value number \p Num using phis, so that it has the values of
1561 /// the phis in BB.
phiTranslateImpl(const BasicBlock * Pred,const BasicBlock * PhiBlock,uint32_t Num,GVN & Gvn)1562 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1563 const BasicBlock *PhiBlock,
1564 uint32_t Num, GVN &Gvn) {
1565 if (PHINode *PN = NumberingPhi[Num]) {
1566 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1567 if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1568 if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1569 return TransVal;
1570 }
1571 return Num;
1572 }
1573
1574 // If there is any value related with Num is defined in a BB other than
1575 // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1576 // a backedge. We can do an early exit in that case to save compile time.
1577 if (!areAllValsInBB(Num, PhiBlock, Gvn))
1578 return Num;
1579
1580 if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1581 return Num;
1582 Expression Exp = Expressions[ExprIdx[Num]];
1583
1584 for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1585 // For InsertValue and ExtractValue, some varargs are index numbers
1586 // instead of value numbers. Those index numbers should not be
1587 // translated.
1588 if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1589 (i > 0 && Exp.opcode == Instruction::ExtractValue))
1590 continue;
1591 Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1592 }
1593
1594 if (Exp.commutative) {
1595 assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1596 if (Exp.varargs[0] > Exp.varargs[1]) {
1597 std::swap(Exp.varargs[0], Exp.varargs[1]);
1598 uint32_t Opcode = Exp.opcode >> 8;
1599 if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1600 Exp.opcode = (Opcode << 8) |
1601 CmpInst::getSwappedPredicate(
1602 static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1603 }
1604 }
1605
1606 if (uint32_t NewNum = expressionNumbering[Exp]) {
1607 if (Exp.opcode == Instruction::Call && NewNum != Num)
1608 return areCallValsEqual(Num, NewNum, Pred, PhiBlock, Gvn) ? NewNum : Num;
1609 return NewNum;
1610 }
1611 return Num;
1612 }
1613
1614 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1615 /// again.
eraseTranslateCacheEntry(uint32_t Num,const BasicBlock & CurrBlock)1616 void GVN::ValueTable::eraseTranslateCacheEntry(uint32_t Num,
1617 const BasicBlock &CurrBlock) {
1618 for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1619 auto FindRes = PhiTranslateTable.find({Num, Pred});
1620 if (FindRes != PhiTranslateTable.end())
1621 PhiTranslateTable.erase(FindRes);
1622 }
1623 }
1624
1625 // In order to find a leader for a given value number at a
1626 // specific basic block, we first obtain the list of all Values for that number,
1627 // and then scan the list to find one whose block dominates the block in
1628 // question. This is fast because dominator tree queries consist of only
1629 // a few comparisons of DFS numbers.
findLeader(const BasicBlock * BB,uint32_t num)1630 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1631 LeaderTableEntry Vals = LeaderTable[num];
1632 if (!Vals.Val) return nullptr;
1633
1634 Value *Val = nullptr;
1635 if (DT->dominates(Vals.BB, BB)) {
1636 Val = Vals.Val;
1637 if (isa<Constant>(Val)) return Val;
1638 }
1639
1640 LeaderTableEntry* Next = Vals.Next;
1641 while (Next) {
1642 if (DT->dominates(Next->BB, BB)) {
1643 if (isa<Constant>(Next->Val)) return Next->Val;
1644 if (!Val) Val = Next->Val;
1645 }
1646
1647 Next = Next->Next;
1648 }
1649
1650 return Val;
1651 }
1652
1653 /// There is an edge from 'Src' to 'Dst'. Return
1654 /// true if every path from the entry block to 'Dst' passes via this edge. In
1655 /// particular 'Dst' must not be reachable via another edge from 'Src'.
isOnlyReachableViaThisEdge(const BasicBlockEdge & E,DominatorTree * DT)1656 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
1657 DominatorTree *DT) {
1658 // While in theory it is interesting to consider the case in which Dst has
1659 // more than one predecessor, because Dst might be part of a loop which is
1660 // only reachable from Src, in practice it is pointless since at the time
1661 // GVN runs all such loops have preheaders, which means that Dst will have
1662 // been changed to have only one predecessor, namely Src.
1663 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1664 assert((!Pred || Pred == E.getStart()) &&
1665 "No edge between these basic blocks!");
1666 return Pred != nullptr;
1667 }
1668
assignBlockRPONumber(Function & F)1669 void GVN::assignBlockRPONumber(Function &F) {
1670 BlockRPONumber.clear();
1671 uint32_t NextBlockNumber = 1;
1672 ReversePostOrderTraversal<Function *> RPOT(&F);
1673 for (BasicBlock *BB : RPOT)
1674 BlockRPONumber[BB] = NextBlockNumber++;
1675 InvalidBlockRPONumbers = false;
1676 }
1677
1678 // Tries to replace instruction with const, using information from
1679 // ReplaceWithConstMap.
replaceOperandsWithConsts(Instruction * Instr) const1680 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1681 bool Changed = false;
1682 for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1683 Value *Operand = Instr->getOperand(OpNum);
1684 auto it = ReplaceWithConstMap.find(Operand);
1685 if (it != ReplaceWithConstMap.end()) {
1686 assert(!isa<Constant>(Operand) &&
1687 "Replacing constants with constants is invalid");
1688 LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "
1689 << *it->second << " in instruction " << *Instr << '\n');
1690 Instr->setOperand(OpNum, it->second);
1691 Changed = true;
1692 }
1693 }
1694 return Changed;
1695 }
1696
1697 /// The given values are known to be equal in every block
1698 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1699 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1700 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1701 /// value starting from the end of Root.Start.
propagateEquality(Value * LHS,Value * RHS,const BasicBlockEdge & Root,bool DominatesByEdge)1702 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1703 bool DominatesByEdge) {
1704 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
1705 Worklist.push_back(std::make_pair(LHS, RHS));
1706 bool Changed = false;
1707 // For speed, compute a conservative fast approximation to
1708 // DT->dominates(Root, Root.getEnd());
1709 const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1710
1711 while (!Worklist.empty()) {
1712 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1713 LHS = Item.first; RHS = Item.second;
1714
1715 if (LHS == RHS)
1716 continue;
1717 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1718
1719 // Don't try to propagate equalities between constants.
1720 if (isa<Constant>(LHS) && isa<Constant>(RHS))
1721 continue;
1722
1723 // Prefer a constant on the right-hand side, or an Argument if no constants.
1724 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1725 std::swap(LHS, RHS);
1726 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1727
1728 // If there is no obvious reason to prefer the left-hand side over the
1729 // right-hand side, ensure the longest lived term is on the right-hand side,
1730 // so the shortest lived term will be replaced by the longest lived.
1731 // This tends to expose more simplifications.
1732 uint32_t LVN = VN.lookupOrAdd(LHS);
1733 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1734 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1735 // Move the 'oldest' value to the right-hand side, using the value number
1736 // as a proxy for age.
1737 uint32_t RVN = VN.lookupOrAdd(RHS);
1738 if (LVN < RVN) {
1739 std::swap(LHS, RHS);
1740 LVN = RVN;
1741 }
1742 }
1743
1744 // If value numbering later sees that an instruction in the scope is equal
1745 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1746 // the invariant that instructions only occur in the leader table for their
1747 // own value number (this is used by removeFromLeaderTable), do not do this
1748 // if RHS is an instruction (if an instruction in the scope is morphed into
1749 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1750 // using the leader table is about compiling faster, not optimizing better).
1751 // The leader table only tracks basic blocks, not edges. Only add to if we
1752 // have the simple case where the edge dominates the end.
1753 if (RootDominatesEnd && !isa<Instruction>(RHS))
1754 addToLeaderTable(LVN, RHS, Root.getEnd());
1755
1756 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1757 // LHS always has at least one use that is not dominated by Root, this will
1758 // never do anything if LHS has only one use.
1759 if (!LHS->hasOneUse()) {
1760 unsigned NumReplacements =
1761 DominatesByEdge
1762 ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1763 : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1764
1765 Changed |= NumReplacements > 0;
1766 NumGVNEqProp += NumReplacements;
1767 // Cached information for anything that uses LHS will be invalid.
1768 if (MD)
1769 MD->invalidateCachedPointerInfo(LHS);
1770 }
1771
1772 // Now try to deduce additional equalities from this one. For example, if
1773 // the known equality was "(A != B)" == "false" then it follows that A and B
1774 // are equal in the scope. Only boolean equalities with an explicit true or
1775 // false RHS are currently supported.
1776 if (!RHS->getType()->isIntegerTy(1))
1777 // Not a boolean equality - bail out.
1778 continue;
1779 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1780 if (!CI)
1781 // RHS neither 'true' nor 'false' - bail out.
1782 continue;
1783 // Whether RHS equals 'true'. Otherwise it equals 'false'.
1784 bool isKnownTrue = CI->isMinusOne();
1785 bool isKnownFalse = !isKnownTrue;
1786
1787 // If "A && B" is known true then both A and B are known true. If "A || B"
1788 // is known false then both A and B are known false.
1789 Value *A, *B;
1790 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1791 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1792 Worklist.push_back(std::make_pair(A, RHS));
1793 Worklist.push_back(std::make_pair(B, RHS));
1794 continue;
1795 }
1796
1797 // If we are propagating an equality like "(A == B)" == "true" then also
1798 // propagate the equality A == B. When propagating a comparison such as
1799 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1800 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1801 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1802
1803 // If "A == B" is known true, or "A != B" is known false, then replace
1804 // A with B everywhere in the scope.
1805 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1806 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1807 Worklist.push_back(std::make_pair(Op0, Op1));
1808
1809 // Handle the floating point versions of equality comparisons too.
1810 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1811 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1812
1813 // Floating point -0.0 and 0.0 compare equal, so we can only
1814 // propagate values if we know that we have a constant and that
1815 // its value is non-zero.
1816
1817 // FIXME: We should do this optimization if 'no signed zeros' is
1818 // applicable via an instruction-level fast-math-flag or some other
1819 // indicator that relaxed FP semantics are being used.
1820
1821 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1822 Worklist.push_back(std::make_pair(Op0, Op1));
1823 }
1824
1825 // If "A >= B" is known true, replace "A < B" with false everywhere.
1826 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1827 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1828 // Since we don't have the instruction "A < B" immediately to hand, work
1829 // out the value number that it would have and use that to find an
1830 // appropriate instruction (if any).
1831 uint32_t NextNum = VN.getNextUnusedValueNumber();
1832 uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1833 // If the number we were assigned was brand new then there is no point in
1834 // looking for an instruction realizing it: there cannot be one!
1835 if (Num < NextNum) {
1836 Value *NotCmp = findLeader(Root.getEnd(), Num);
1837 if (NotCmp && isa<Instruction>(NotCmp)) {
1838 unsigned NumReplacements =
1839 DominatesByEdge
1840 ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1841 : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1842 Root.getStart());
1843 Changed |= NumReplacements > 0;
1844 NumGVNEqProp += NumReplacements;
1845 // Cached information for anything that uses NotCmp will be invalid.
1846 if (MD)
1847 MD->invalidateCachedPointerInfo(NotCmp);
1848 }
1849 }
1850 // Ensure that any instruction in scope that gets the "A < B" value number
1851 // is replaced with false.
1852 // The leader table only tracks basic blocks, not edges. Only add to if we
1853 // have the simple case where the edge dominates the end.
1854 if (RootDominatesEnd)
1855 addToLeaderTable(Num, NotVal, Root.getEnd());
1856
1857 continue;
1858 }
1859 }
1860
1861 return Changed;
1862 }
1863
1864 /// When calculating availability, handle an instruction
1865 /// by inserting it into the appropriate sets
processInstruction(Instruction * I)1866 bool GVN::processInstruction(Instruction *I) {
1867 // Ignore dbg info intrinsics.
1868 if (isa<DbgInfoIntrinsic>(I))
1869 return false;
1870
1871 // If the instruction can be easily simplified then do so now in preference
1872 // to value numbering it. Value numbering often exposes redundancies, for
1873 // example if it determines that %y is equal to %x then the instruction
1874 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1875 const DataLayout &DL = I->getModule()->getDataLayout();
1876 if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1877 bool Changed = false;
1878 if (!I->use_empty()) {
1879 I->replaceAllUsesWith(V);
1880 Changed = true;
1881 }
1882 if (isInstructionTriviallyDead(I, TLI)) {
1883 markInstructionForDeletion(I);
1884 Changed = true;
1885 }
1886 if (Changed) {
1887 if (MD && V->getType()->isPtrOrPtrVectorTy())
1888 MD->invalidateCachedPointerInfo(V);
1889 ++NumGVNSimpl;
1890 return true;
1891 }
1892 }
1893
1894 if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1895 if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1896 return processAssumeIntrinsic(IntrinsicI);
1897
1898 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1899 if (processLoad(LI))
1900 return true;
1901
1902 unsigned Num = VN.lookupOrAdd(LI);
1903 addToLeaderTable(Num, LI, LI->getParent());
1904 return false;
1905 }
1906
1907 // For conditional branches, we can perform simple conditional propagation on
1908 // the condition value itself.
1909 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1910 if (!BI->isConditional())
1911 return false;
1912
1913 if (isa<Constant>(BI->getCondition()))
1914 return processFoldableCondBr(BI);
1915
1916 Value *BranchCond = BI->getCondition();
1917 BasicBlock *TrueSucc = BI->getSuccessor(0);
1918 BasicBlock *FalseSucc = BI->getSuccessor(1);
1919 // Avoid multiple edges early.
1920 if (TrueSucc == FalseSucc)
1921 return false;
1922
1923 BasicBlock *Parent = BI->getParent();
1924 bool Changed = false;
1925
1926 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1927 BasicBlockEdge TrueE(Parent, TrueSucc);
1928 Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1929
1930 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1931 BasicBlockEdge FalseE(Parent, FalseSucc);
1932 Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1933
1934 return Changed;
1935 }
1936
1937 // For switches, propagate the case values into the case destinations.
1938 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1939 Value *SwitchCond = SI->getCondition();
1940 BasicBlock *Parent = SI->getParent();
1941 bool Changed = false;
1942
1943 // Remember how many outgoing edges there are to every successor.
1944 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
1945 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1946 ++SwitchEdges[SI->getSuccessor(i)];
1947
1948 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1949 i != e; ++i) {
1950 BasicBlock *Dst = i->getCaseSuccessor();
1951 // If there is only a single edge, propagate the case value into it.
1952 if (SwitchEdges.lookup(Dst) == 1) {
1953 BasicBlockEdge E(Parent, Dst);
1954 Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1955 }
1956 }
1957 return Changed;
1958 }
1959
1960 // Instructions with void type don't return a value, so there's
1961 // no point in trying to find redundancies in them.
1962 if (I->getType()->isVoidTy())
1963 return false;
1964
1965 uint32_t NextNum = VN.getNextUnusedValueNumber();
1966 unsigned Num = VN.lookupOrAdd(I);
1967
1968 // Allocations are always uniquely numbered, so we can save time and memory
1969 // by fast failing them.
1970 if (isa<AllocaInst>(I) || I->isTerminator() || isa<PHINode>(I)) {
1971 addToLeaderTable(Num, I, I->getParent());
1972 return false;
1973 }
1974
1975 // If the number we were assigned was a brand new VN, then we don't
1976 // need to do a lookup to see if the number already exists
1977 // somewhere in the domtree: it can't!
1978 if (Num >= NextNum) {
1979 addToLeaderTable(Num, I, I->getParent());
1980 return false;
1981 }
1982
1983 // Perform fast-path value-number based elimination of values inherited from
1984 // dominators.
1985 Value *Repl = findLeader(I->getParent(), Num);
1986 if (!Repl) {
1987 // Failure, just remember this instance for future use.
1988 addToLeaderTable(Num, I, I->getParent());
1989 return false;
1990 } else if (Repl == I) {
1991 // If I was the result of a shortcut PRE, it might already be in the table
1992 // and the best replacement for itself. Nothing to do.
1993 return false;
1994 }
1995
1996 // Remove it!
1997 patchAndReplaceAllUsesWith(I, Repl);
1998 if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1999 MD->invalidateCachedPointerInfo(Repl);
2000 markInstructionForDeletion(I);
2001 return true;
2002 }
2003
2004 /// runOnFunction - This is the main transformation entry point for a function.
runImpl(Function & F,AssumptionCache & RunAC,DominatorTree & RunDT,const TargetLibraryInfo & RunTLI,AAResults & RunAA,MemoryDependenceResults * RunMD,LoopInfo * LI,OptimizationRemarkEmitter * RunORE)2005 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
2006 const TargetLibraryInfo &RunTLI, AAResults &RunAA,
2007 MemoryDependenceResults *RunMD, LoopInfo *LI,
2008 OptimizationRemarkEmitter *RunORE) {
2009 AC = &RunAC;
2010 DT = &RunDT;
2011 VN.setDomTree(DT);
2012 TLI = &RunTLI;
2013 VN.setAliasAnalysis(&RunAA);
2014 MD = RunMD;
2015 ImplicitControlFlowTracking ImplicitCFT(DT);
2016 ICF = &ImplicitCFT;
2017 VN.setMemDep(MD);
2018 ORE = RunORE;
2019 InvalidBlockRPONumbers = true;
2020
2021 bool Changed = false;
2022 bool ShouldContinue = true;
2023
2024 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
2025 // Merge unconditional branches, allowing PRE to catch more
2026 // optimization opportunities.
2027 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2028 BasicBlock *BB = &*FI++;
2029
2030 bool removedBlock = MergeBlockIntoPredecessor(BB, &DTU, LI, nullptr, MD);
2031 if (removedBlock)
2032 ++NumGVNBlocks;
2033
2034 Changed |= removedBlock;
2035 }
2036
2037 unsigned Iteration = 0;
2038 while (ShouldContinue) {
2039 LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2040 ShouldContinue = iterateOnFunction(F);
2041 Changed |= ShouldContinue;
2042 ++Iteration;
2043 }
2044
2045 if (EnablePRE) {
2046 // Fabricate val-num for dead-code in order to suppress assertion in
2047 // performPRE().
2048 assignValNumForDeadCode();
2049 bool PREChanged = true;
2050 while (PREChanged) {
2051 PREChanged = performPRE(F);
2052 Changed |= PREChanged;
2053 }
2054 }
2055
2056 // FIXME: Should perform GVN again after PRE does something. PRE can move
2057 // computations into blocks where they become fully redundant. Note that
2058 // we can't do this until PRE's critical edge splitting updates memdep.
2059 // Actually, when this happens, we should just fully integrate PRE into GVN.
2060
2061 cleanupGlobalSets();
2062 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2063 // iteration.
2064 DeadBlocks.clear();
2065
2066 return Changed;
2067 }
2068
processBlock(BasicBlock * BB)2069 bool GVN::processBlock(BasicBlock *BB) {
2070 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2071 // (and incrementing BI before processing an instruction).
2072 assert(InstrsToErase.empty() &&
2073 "We expect InstrsToErase to be empty across iterations");
2074 if (DeadBlocks.count(BB))
2075 return false;
2076
2077 // Clearing map before every BB because it can be used only for single BB.
2078 ReplaceWithConstMap.clear();
2079 bool ChangedFunction = false;
2080
2081 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2082 BI != BE;) {
2083 if (!ReplaceWithConstMap.empty())
2084 ChangedFunction |= replaceOperandsWithConsts(&*BI);
2085 ChangedFunction |= processInstruction(&*BI);
2086
2087 if (InstrsToErase.empty()) {
2088 ++BI;
2089 continue;
2090 }
2091
2092 // If we need some instructions deleted, do it now.
2093 NumGVNInstr += InstrsToErase.size();
2094
2095 // Avoid iterator invalidation.
2096 bool AtStart = BI == BB->begin();
2097 if (!AtStart)
2098 --BI;
2099
2100 for (auto *I : InstrsToErase) {
2101 assert(I->getParent() == BB && "Removing instruction from wrong block?");
2102 LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n');
2103 salvageDebugInfo(*I);
2104 if (MD) MD->removeInstruction(I);
2105 LLVM_DEBUG(verifyRemoved(I));
2106 ICF->removeInstruction(I);
2107 I->eraseFromParent();
2108 }
2109 InstrsToErase.clear();
2110
2111 if (AtStart)
2112 BI = BB->begin();
2113 else
2114 ++BI;
2115 }
2116
2117 return ChangedFunction;
2118 }
2119
2120 // Instantiate an expression in a predecessor that lacked it.
performScalarPREInsertion(Instruction * Instr,BasicBlock * Pred,BasicBlock * Curr,unsigned int ValNo)2121 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2122 BasicBlock *Curr, unsigned int ValNo) {
2123 // Because we are going top-down through the block, all value numbers
2124 // will be available in the predecessor by the time we need them. Any
2125 // that weren't originally present will have been instantiated earlier
2126 // in this loop.
2127 bool success = true;
2128 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2129 Value *Op = Instr->getOperand(i);
2130 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2131 continue;
2132 // This could be a newly inserted instruction, in which case, we won't
2133 // find a value number, and should give up before we hurt ourselves.
2134 // FIXME: Rewrite the infrastructure to let it easier to value number
2135 // and process newly inserted instructions.
2136 if (!VN.exists(Op)) {
2137 success = false;
2138 break;
2139 }
2140 uint32_t TValNo =
2141 VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2142 if (Value *V = findLeader(Pred, TValNo)) {
2143 Instr->setOperand(i, V);
2144 } else {
2145 success = false;
2146 break;
2147 }
2148 }
2149
2150 // Fail out if we encounter an operand that is not available in
2151 // the PRE predecessor. This is typically because of loads which
2152 // are not value numbered precisely.
2153 if (!success)
2154 return false;
2155
2156 Instr->insertBefore(Pred->getTerminator());
2157 Instr->setName(Instr->getName() + ".pre");
2158 Instr->setDebugLoc(Instr->getDebugLoc());
2159
2160 unsigned Num = VN.lookupOrAdd(Instr);
2161 VN.add(Instr, Num);
2162
2163 // Update the availability map to include the new instruction.
2164 addToLeaderTable(Num, Instr, Pred);
2165 return true;
2166 }
2167
performScalarPRE(Instruction * CurInst)2168 bool GVN::performScalarPRE(Instruction *CurInst) {
2169 if (isa<AllocaInst>(CurInst) || CurInst->isTerminator() ||
2170 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2171 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2172 isa<DbgInfoIntrinsic>(CurInst))
2173 return false;
2174
2175 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2176 // sinking the compare again, and it would force the code generator to
2177 // move the i1 from processor flags or predicate registers into a general
2178 // purpose register.
2179 if (isa<CmpInst>(CurInst))
2180 return false;
2181
2182 // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from
2183 // sinking the addressing mode computation back to its uses. Extending the
2184 // GEP's live range increases the register pressure, and therefore it can
2185 // introduce unnecessary spills.
2186 //
2187 // This doesn't prevent Load PRE. PHI translation will make the GEP available
2188 // to the load by moving it to the predecessor block if necessary.
2189 if (isa<GetElementPtrInst>(CurInst))
2190 return false;
2191
2192 // We don't currently value number ANY inline asm calls.
2193 if (auto *CallB = dyn_cast<CallBase>(CurInst))
2194 if (CallB->isInlineAsm())
2195 return false;
2196
2197 uint32_t ValNo = VN.lookup(CurInst);
2198
2199 // Look for the predecessors for PRE opportunities. We're
2200 // only trying to solve the basic diamond case, where
2201 // a value is computed in the successor and one predecessor,
2202 // but not the other. We also explicitly disallow cases
2203 // where the successor is its own predecessor, because they're
2204 // more complicated to get right.
2205 unsigned NumWith = 0;
2206 unsigned NumWithout = 0;
2207 BasicBlock *PREPred = nullptr;
2208 BasicBlock *CurrentBlock = CurInst->getParent();
2209
2210 // Update the RPO numbers for this function.
2211 if (InvalidBlockRPONumbers)
2212 assignBlockRPONumber(*CurrentBlock->getParent());
2213
2214 SmallVector<std::pair<Value *, BasicBlock *>, 8> predMap;
2215 for (BasicBlock *P : predecessors(CurrentBlock)) {
2216 // We're not interested in PRE where blocks with predecessors that are
2217 // not reachable.
2218 if (!DT->isReachableFromEntry(P)) {
2219 NumWithout = 2;
2220 break;
2221 }
2222 // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2223 // when CurInst has operand defined in CurrentBlock (so it may be defined
2224 // by phi in the loop header).
2225 assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) &&
2226 "Invalid BlockRPONumber map.");
2227 if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2228 llvm::any_of(CurInst->operands(), [&](const Use &U) {
2229 if (auto *Inst = dyn_cast<Instruction>(U.get()))
2230 return Inst->getParent() == CurrentBlock;
2231 return false;
2232 })) {
2233 NumWithout = 2;
2234 break;
2235 }
2236
2237 uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2238 Value *predV = findLeader(P, TValNo);
2239 if (!predV) {
2240 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2241 PREPred = P;
2242 ++NumWithout;
2243 } else if (predV == CurInst) {
2244 /* CurInst dominates this predecessor. */
2245 NumWithout = 2;
2246 break;
2247 } else {
2248 predMap.push_back(std::make_pair(predV, P));
2249 ++NumWith;
2250 }
2251 }
2252
2253 // Don't do PRE when it might increase code size, i.e. when
2254 // we would need to insert instructions in more than one pred.
2255 if (NumWithout > 1 || NumWith == 0)
2256 return false;
2257
2258 // We may have a case where all predecessors have the instruction,
2259 // and we just need to insert a phi node. Otherwise, perform
2260 // insertion.
2261 Instruction *PREInstr = nullptr;
2262
2263 if (NumWithout != 0) {
2264 if (!isSafeToSpeculativelyExecute(CurInst)) {
2265 // It is only valid to insert a new instruction if the current instruction
2266 // is always executed. An instruction with implicit control flow could
2267 // prevent us from doing it. If we cannot speculate the execution, then
2268 // PRE should be prohibited.
2269 if (ICF->isDominatedByICFIFromSameBlock(CurInst))
2270 return false;
2271 }
2272
2273 // Don't do PRE across indirect branch.
2274 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2275 return false;
2276
2277 // Don't do PRE across callbr.
2278 // FIXME: Can we do this across the fallthrough edge?
2279 if (isa<CallBrInst>(PREPred->getTerminator()))
2280 return false;
2281
2282 // We can't do PRE safely on a critical edge, so instead we schedule
2283 // the edge to be split and perform the PRE the next time we iterate
2284 // on the function.
2285 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2286 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2287 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2288 return false;
2289 }
2290 // We need to insert somewhere, so let's give it a shot
2291 PREInstr = CurInst->clone();
2292 if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2293 // If we failed insertion, make sure we remove the instruction.
2294 LLVM_DEBUG(verifyRemoved(PREInstr));
2295 PREInstr->deleteValue();
2296 return false;
2297 }
2298 }
2299
2300 // Either we should have filled in the PRE instruction, or we should
2301 // not have needed insertions.
2302 assert(PREInstr != nullptr || NumWithout == 0);
2303
2304 ++NumGVNPRE;
2305
2306 // Create a PHI to make the value available in this block.
2307 PHINode *Phi =
2308 PHINode::Create(CurInst->getType(), predMap.size(),
2309 CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2310 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2311 if (Value *V = predMap[i].first) {
2312 // If we use an existing value in this phi, we have to patch the original
2313 // value because the phi will be used to replace a later value.
2314 patchReplacementInstruction(CurInst, V);
2315 Phi->addIncoming(V, predMap[i].second);
2316 } else
2317 Phi->addIncoming(PREInstr, PREPred);
2318 }
2319
2320 VN.add(Phi, ValNo);
2321 // After creating a new PHI for ValNo, the phi translate result for ValNo will
2322 // be changed, so erase the related stale entries in phi translate cache.
2323 VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2324 addToLeaderTable(ValNo, Phi, CurrentBlock);
2325 Phi->setDebugLoc(CurInst->getDebugLoc());
2326 CurInst->replaceAllUsesWith(Phi);
2327 if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2328 MD->invalidateCachedPointerInfo(Phi);
2329 VN.erase(CurInst);
2330 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2331
2332 LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2333 if (MD)
2334 MD->removeInstruction(CurInst);
2335 LLVM_DEBUG(verifyRemoved(CurInst));
2336 // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes
2337 // some assertion failures.
2338 ICF->removeInstruction(CurInst);
2339 CurInst->eraseFromParent();
2340 ++NumGVNInstr;
2341
2342 return true;
2343 }
2344
2345 /// Perform a purely local form of PRE that looks for diamond
2346 /// control flow patterns and attempts to perform simple PRE at the join point.
performPRE(Function & F)2347 bool GVN::performPRE(Function &F) {
2348 bool Changed = false;
2349 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2350 // Nothing to PRE in the entry block.
2351 if (CurrentBlock == &F.getEntryBlock())
2352 continue;
2353
2354 // Don't perform PRE on an EH pad.
2355 if (CurrentBlock->isEHPad())
2356 continue;
2357
2358 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2359 BE = CurrentBlock->end();
2360 BI != BE;) {
2361 Instruction *CurInst = &*BI++;
2362 Changed |= performScalarPRE(CurInst);
2363 }
2364 }
2365
2366 if (splitCriticalEdges())
2367 Changed = true;
2368
2369 return Changed;
2370 }
2371
2372 /// Split the critical edge connecting the given two blocks, and return
2373 /// the block inserted to the critical edge.
splitCriticalEdges(BasicBlock * Pred,BasicBlock * Succ)2374 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2375 BasicBlock *BB =
2376 SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT));
2377 if (MD)
2378 MD->invalidateCachedPredecessors();
2379 InvalidBlockRPONumbers = true;
2380 return BB;
2381 }
2382
2383 /// Split critical edges found during the previous
2384 /// iteration that may enable further optimization.
splitCriticalEdges()2385 bool GVN::splitCriticalEdges() {
2386 if (toSplit.empty())
2387 return false;
2388 do {
2389 std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val();
2390 SplitCriticalEdge(Edge.first, Edge.second,
2391 CriticalEdgeSplittingOptions(DT));
2392 } while (!toSplit.empty());
2393 if (MD) MD->invalidateCachedPredecessors();
2394 InvalidBlockRPONumbers = true;
2395 return true;
2396 }
2397
2398 /// Executes one iteration of GVN
iterateOnFunction(Function & F)2399 bool GVN::iterateOnFunction(Function &F) {
2400 cleanupGlobalSets();
2401
2402 // Top-down walk of the dominator tree
2403 bool Changed = false;
2404 // Needed for value numbering with phi construction to work.
2405 // RPOT walks the graph in its constructor and will not be invalidated during
2406 // processBlock.
2407 ReversePostOrderTraversal<Function *> RPOT(&F);
2408
2409 for (BasicBlock *BB : RPOT)
2410 Changed |= processBlock(BB);
2411
2412 return Changed;
2413 }
2414
cleanupGlobalSets()2415 void GVN::cleanupGlobalSets() {
2416 VN.clear();
2417 LeaderTable.clear();
2418 BlockRPONumber.clear();
2419 TableAllocator.Reset();
2420 ICF->clear();
2421 InvalidBlockRPONumbers = true;
2422 }
2423
2424 /// Verify that the specified instruction does not occur in our
2425 /// internal data structures.
verifyRemoved(const Instruction * Inst) const2426 void GVN::verifyRemoved(const Instruction *Inst) const {
2427 VN.verifyRemoved(Inst);
2428
2429 // Walk through the value number scope to make sure the instruction isn't
2430 // ferreted away in it.
2431 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2432 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2433 const LeaderTableEntry *Node = &I->second;
2434 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2435
2436 while (Node->Next) {
2437 Node = Node->Next;
2438 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2439 }
2440 }
2441 }
2442
2443 /// BB is declared dead, which implied other blocks become dead as well. This
2444 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2445 /// live successors, update their phi nodes by replacing the operands
2446 /// corresponding to dead blocks with UndefVal.
addDeadBlock(BasicBlock * BB)2447 void GVN::addDeadBlock(BasicBlock *BB) {
2448 SmallVector<BasicBlock *, 4> NewDead;
2449 SmallSetVector<BasicBlock *, 4> DF;
2450
2451 NewDead.push_back(BB);
2452 while (!NewDead.empty()) {
2453 BasicBlock *D = NewDead.pop_back_val();
2454 if (DeadBlocks.count(D))
2455 continue;
2456
2457 // All blocks dominated by D are dead.
2458 SmallVector<BasicBlock *, 8> Dom;
2459 DT->getDescendants(D, Dom);
2460 DeadBlocks.insert(Dom.begin(), Dom.end());
2461
2462 // Figure out the dominance-frontier(D).
2463 for (BasicBlock *B : Dom) {
2464 for (BasicBlock *S : successors(B)) {
2465 if (DeadBlocks.count(S))
2466 continue;
2467
2468 bool AllPredDead = true;
2469 for (BasicBlock *P : predecessors(S))
2470 if (!DeadBlocks.count(P)) {
2471 AllPredDead = false;
2472 break;
2473 }
2474
2475 if (!AllPredDead) {
2476 // S could be proved dead later on. That is why we don't update phi
2477 // operands at this moment.
2478 DF.insert(S);
2479 } else {
2480 // While S is not dominated by D, it is dead by now. This could take
2481 // place if S already have a dead predecessor before D is declared
2482 // dead.
2483 NewDead.push_back(S);
2484 }
2485 }
2486 }
2487 }
2488
2489 // For the dead blocks' live successors, update their phi nodes by replacing
2490 // the operands corresponding to dead blocks with UndefVal.
2491 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2492 I != E; I++) {
2493 BasicBlock *B = *I;
2494 if (DeadBlocks.count(B))
2495 continue;
2496
2497 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2498 for (BasicBlock *P : Preds) {
2499 if (!DeadBlocks.count(P))
2500 continue;
2501
2502 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2503 if (BasicBlock *S = splitCriticalEdges(P, B))
2504 DeadBlocks.insert(P = S);
2505 }
2506
2507 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2508 PHINode &Phi = cast<PHINode>(*II);
2509 Phi.setIncomingValueForBlock(P, UndefValue::get(Phi.getType()));
2510 if (MD)
2511 MD->invalidateCachedPointerInfo(&Phi);
2512 }
2513 }
2514 }
2515 }
2516
2517 // If the given branch is recognized as a foldable branch (i.e. conditional
2518 // branch with constant condition), it will perform following analyses and
2519 // transformation.
2520 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2521 // R be the target of the dead out-coming edge.
2522 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2523 // edge. The result of this step will be {X| X is dominated by R}
2524 // 2) Identify those blocks which haves at least one dead predecessor. The
2525 // result of this step will be dominance-frontier(R).
2526 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2527 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2528 //
2529 // Return true iff *NEW* dead code are found.
processFoldableCondBr(BranchInst * BI)2530 bool GVN::processFoldableCondBr(BranchInst *BI) {
2531 if (!BI || BI->isUnconditional())
2532 return false;
2533
2534 // If a branch has two identical successors, we cannot declare either dead.
2535 if (BI->getSuccessor(0) == BI->getSuccessor(1))
2536 return false;
2537
2538 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2539 if (!Cond)
2540 return false;
2541
2542 BasicBlock *DeadRoot =
2543 Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2544 if (DeadBlocks.count(DeadRoot))
2545 return false;
2546
2547 if (!DeadRoot->getSinglePredecessor())
2548 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2549
2550 addDeadBlock(DeadRoot);
2551 return true;
2552 }
2553
2554 // performPRE() will trigger assert if it comes across an instruction without
2555 // associated val-num. As it normally has far more live instructions than dead
2556 // instructions, it makes more sense just to "fabricate" a val-number for the
2557 // dead code than checking if instruction involved is dead or not.
assignValNumForDeadCode()2558 void GVN::assignValNumForDeadCode() {
2559 for (BasicBlock *BB : DeadBlocks) {
2560 for (Instruction &Inst : *BB) {
2561 unsigned ValNum = VN.lookupOrAdd(&Inst);
2562 addToLeaderTable(ValNum, &Inst, BB);
2563 }
2564 }
2565 }
2566
2567 class llvm::gvn::GVNLegacyPass : public FunctionPass {
2568 public:
2569 static char ID; // Pass identification, replacement for typeid
2570
GVNLegacyPass(bool NoMemDepAnalysis=!EnableMemDep)2571 explicit GVNLegacyPass(bool NoMemDepAnalysis = !EnableMemDep)
2572 : FunctionPass(ID), NoMemDepAnalysis(NoMemDepAnalysis) {
2573 initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry());
2574 }
2575
runOnFunction(Function & F)2576 bool runOnFunction(Function &F) override {
2577 if (skipFunction(F))
2578 return false;
2579
2580 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2581
2582 return Impl.runImpl(
2583 F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2584 getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2585 getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2586 getAnalysis<AAResultsWrapperPass>().getAAResults(),
2587 NoMemDepAnalysis ? nullptr
2588 : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2589 LIWP ? &LIWP->getLoopInfo() : nullptr,
2590 &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2591 }
2592
getAnalysisUsage(AnalysisUsage & AU) const2593 void getAnalysisUsage(AnalysisUsage &AU) const override {
2594 AU.addRequired<AssumptionCacheTracker>();
2595 AU.addRequired<DominatorTreeWrapperPass>();
2596 AU.addRequired<TargetLibraryInfoWrapperPass>();
2597 if (!NoMemDepAnalysis)
2598 AU.addRequired<MemoryDependenceWrapperPass>();
2599 AU.addRequired<AAResultsWrapperPass>();
2600
2601 AU.addPreserved<DominatorTreeWrapperPass>();
2602 AU.addPreserved<GlobalsAAWrapperPass>();
2603 AU.addPreserved<TargetLibraryInfoWrapperPass>();
2604 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2605 }
2606
2607 private:
2608 bool NoMemDepAnalysis;
2609 GVN Impl;
2610 };
2611
2612 char GVNLegacyPass::ID = 0;
2613
2614 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)2615 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2616 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
2617 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2618 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
2619 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2620 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
2621 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
2622 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2623
2624 // The public interface to this file...
2625 FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) {
2626 return new GVNLegacyPass(NoMemDepAnalysis);
2627 }
2628