1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
12 //
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
15 //
16 //===----------------------------------------------------------------------===//
17
18 #include "llvm/Transforms/Scalar.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/Hashing.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/PostOrderIterator.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AssumptionCache.h"
29 #include "llvm/Analysis/CFG.h"
30 #include "llvm/Analysis/ConstantFolding.h"
31 #include "llvm/Analysis/GlobalsModRef.h"
32 #include "llvm/Analysis/InstructionSimplify.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/MemoryBuiltins.h"
35 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
36 #include "llvm/Analysis/PHITransAddr.h"
37 #include "llvm/Analysis/TargetLibraryInfo.h"
38 #include "llvm/Analysis/ValueTracking.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/IntrinsicInst.h"
44 #include "llvm/IR/LLVMContext.h"
45 #include "llvm/IR/Metadata.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/Support/Allocator.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include "llvm/Transforms/Utils/SSAUpdater.h"
54 #include <vector>
55 using namespace llvm;
56 using namespace PatternMatch;
57
58 #define DEBUG_TYPE "gvn"
59
60 STATISTIC(NumGVNInstr, "Number of instructions deleted");
61 STATISTIC(NumGVNLoad, "Number of loads deleted");
62 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
63 STATISTIC(NumGVNBlocks, "Number of blocks merged");
64 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
65 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
66 STATISTIC(NumPRELoad, "Number of loads PRE'd");
67
68 static cl::opt<bool> EnablePRE("enable-pre",
69 cl::init(true), cl::Hidden);
70 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
71
72 // Maximum allowed recursion depth.
73 static cl::opt<uint32_t>
74 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
75 cl::desc("Max recurse depth (default = 1000)"));
76
77 //===----------------------------------------------------------------------===//
78 // ValueTable Class
79 //===----------------------------------------------------------------------===//
80
81 /// This class holds the mapping between values and value numbers. It is used
82 /// as an efficient mechanism to determine the expression-wise equivalence of
83 /// two values.
84 namespace {
85 struct Expression {
86 uint32_t opcode;
87 Type *type;
88 SmallVector<uint32_t, 4> varargs;
89
Expression__anon4436d4200111::Expression90 Expression(uint32_t o = ~2U) : opcode(o) { }
91
operator ==__anon4436d4200111::Expression92 bool operator==(const Expression &other) const {
93 if (opcode != other.opcode)
94 return false;
95 if (opcode == ~0U || opcode == ~1U)
96 return true;
97 if (type != other.type)
98 return false;
99 if (varargs != other.varargs)
100 return false;
101 return true;
102 }
103
hash_value(const Expression & Value)104 friend hash_code hash_value(const Expression &Value) {
105 return hash_combine(Value.opcode, Value.type,
106 hash_combine_range(Value.varargs.begin(),
107 Value.varargs.end()));
108 }
109 };
110
111 class ValueTable {
112 DenseMap<Value*, uint32_t> valueNumbering;
113 DenseMap<Expression, uint32_t> expressionNumbering;
114 AliasAnalysis *AA;
115 MemoryDependenceAnalysis *MD;
116 DominatorTree *DT;
117
118 uint32_t nextValueNumber;
119
120 Expression create_expression(Instruction* I);
121 Expression create_cmp_expression(unsigned Opcode,
122 CmpInst::Predicate Predicate,
123 Value *LHS, Value *RHS);
124 Expression create_extractvalue_expression(ExtractValueInst* EI);
125 uint32_t lookup_or_add_call(CallInst* C);
126 public:
ValueTable()127 ValueTable() : nextValueNumber(1) { }
128 uint32_t lookup_or_add(Value *V);
129 uint32_t lookup(Value *V) const;
130 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
131 Value *LHS, Value *RHS);
132 bool exists(Value *V) const;
133 void add(Value *V, uint32_t num);
134 void clear();
135 void erase(Value *v);
setAliasAnalysis(AliasAnalysis * A)136 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
getAliasAnalysis() const137 AliasAnalysis *getAliasAnalysis() const { return AA; }
setMemDep(MemoryDependenceAnalysis * M)138 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
setDomTree(DominatorTree * D)139 void setDomTree(DominatorTree* D) { DT = D; }
getNextUnusedValueNumber()140 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
141 void verifyRemoved(const Value *) const;
142 };
143 }
144
145 namespace llvm {
146 template <> struct DenseMapInfo<Expression> {
getEmptyKeyllvm::DenseMapInfo147 static inline Expression getEmptyKey() {
148 return ~0U;
149 }
150
getTombstoneKeyllvm::DenseMapInfo151 static inline Expression getTombstoneKey() {
152 return ~1U;
153 }
154
getHashValuellvm::DenseMapInfo155 static unsigned getHashValue(const Expression e) {
156 using llvm::hash_value;
157 return static_cast<unsigned>(hash_value(e));
158 }
isEqualllvm::DenseMapInfo159 static bool isEqual(const Expression &LHS, const Expression &RHS) {
160 return LHS == RHS;
161 }
162 };
163
164 }
165
166 //===----------------------------------------------------------------------===//
167 // ValueTable Internal Functions
168 //===----------------------------------------------------------------------===//
169
create_expression(Instruction * I)170 Expression ValueTable::create_expression(Instruction *I) {
171 Expression e;
172 e.type = I->getType();
173 e.opcode = I->getOpcode();
174 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
175 OI != OE; ++OI)
176 e.varargs.push_back(lookup_or_add(*OI));
177 if (I->isCommutative()) {
178 // Ensure that commutative instructions that only differ by a permutation
179 // of their operands get the same value number by sorting the operand value
180 // numbers. Since all commutative instructions have two operands it is more
181 // efficient to sort by hand rather than using, say, std::sort.
182 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
183 if (e.varargs[0] > e.varargs[1])
184 std::swap(e.varargs[0], e.varargs[1]);
185 }
186
187 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
188 // Sort the operand value numbers so x<y and y>x get the same value number.
189 CmpInst::Predicate Predicate = C->getPredicate();
190 if (e.varargs[0] > e.varargs[1]) {
191 std::swap(e.varargs[0], e.varargs[1]);
192 Predicate = CmpInst::getSwappedPredicate(Predicate);
193 }
194 e.opcode = (C->getOpcode() << 8) | Predicate;
195 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
196 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
197 II != IE; ++II)
198 e.varargs.push_back(*II);
199 }
200
201 return e;
202 }
203
create_cmp_expression(unsigned Opcode,CmpInst::Predicate Predicate,Value * LHS,Value * RHS)204 Expression ValueTable::create_cmp_expression(unsigned Opcode,
205 CmpInst::Predicate Predicate,
206 Value *LHS, Value *RHS) {
207 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
208 "Not a comparison!");
209 Expression e;
210 e.type = CmpInst::makeCmpResultType(LHS->getType());
211 e.varargs.push_back(lookup_or_add(LHS));
212 e.varargs.push_back(lookup_or_add(RHS));
213
214 // Sort the operand value numbers so x<y and y>x get the same value number.
215 if (e.varargs[0] > e.varargs[1]) {
216 std::swap(e.varargs[0], e.varargs[1]);
217 Predicate = CmpInst::getSwappedPredicate(Predicate);
218 }
219 e.opcode = (Opcode << 8) | Predicate;
220 return e;
221 }
222
create_extractvalue_expression(ExtractValueInst * EI)223 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
224 assert(EI && "Not an ExtractValueInst?");
225 Expression e;
226 e.type = EI->getType();
227 e.opcode = 0;
228
229 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
230 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
231 // EI might be an extract from one of our recognised intrinsics. If it
232 // is we'll synthesize a semantically equivalent expression instead on
233 // an extract value expression.
234 switch (I->getIntrinsicID()) {
235 case Intrinsic::sadd_with_overflow:
236 case Intrinsic::uadd_with_overflow:
237 e.opcode = Instruction::Add;
238 break;
239 case Intrinsic::ssub_with_overflow:
240 case Intrinsic::usub_with_overflow:
241 e.opcode = Instruction::Sub;
242 break;
243 case Intrinsic::smul_with_overflow:
244 case Intrinsic::umul_with_overflow:
245 e.opcode = Instruction::Mul;
246 break;
247 default:
248 break;
249 }
250
251 if (e.opcode != 0) {
252 // Intrinsic recognized. Grab its args to finish building the expression.
253 assert(I->getNumArgOperands() == 2 &&
254 "Expect two args for recognised intrinsics.");
255 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
256 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
257 return e;
258 }
259 }
260
261 // Not a recognised intrinsic. Fall back to producing an extract value
262 // expression.
263 e.opcode = EI->getOpcode();
264 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
265 OI != OE; ++OI)
266 e.varargs.push_back(lookup_or_add(*OI));
267
268 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
269 II != IE; ++II)
270 e.varargs.push_back(*II);
271
272 return e;
273 }
274
275 //===----------------------------------------------------------------------===//
276 // ValueTable External Functions
277 //===----------------------------------------------------------------------===//
278
279 /// add - Insert a value into the table with a specified value number.
add(Value * V,uint32_t num)280 void ValueTable::add(Value *V, uint32_t num) {
281 valueNumbering.insert(std::make_pair(V, num));
282 }
283
lookup_or_add_call(CallInst * C)284 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
285 if (AA->doesNotAccessMemory(C)) {
286 Expression exp = create_expression(C);
287 uint32_t &e = expressionNumbering[exp];
288 if (!e) e = nextValueNumber++;
289 valueNumbering[C] = e;
290 return e;
291 } else if (AA->onlyReadsMemory(C)) {
292 Expression exp = create_expression(C);
293 uint32_t &e = expressionNumbering[exp];
294 if (!e) {
295 e = nextValueNumber++;
296 valueNumbering[C] = e;
297 return e;
298 }
299 if (!MD) {
300 e = nextValueNumber++;
301 valueNumbering[C] = e;
302 return e;
303 }
304
305 MemDepResult local_dep = MD->getDependency(C);
306
307 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
308 valueNumbering[C] = nextValueNumber;
309 return nextValueNumber++;
310 }
311
312 if (local_dep.isDef()) {
313 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
314
315 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
316 valueNumbering[C] = nextValueNumber;
317 return nextValueNumber++;
318 }
319
320 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
321 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
322 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
323 if (c_vn != cd_vn) {
324 valueNumbering[C] = nextValueNumber;
325 return nextValueNumber++;
326 }
327 }
328
329 uint32_t v = lookup_or_add(local_cdep);
330 valueNumbering[C] = v;
331 return v;
332 }
333
334 // Non-local case.
335 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
336 MD->getNonLocalCallDependency(CallSite(C));
337 // FIXME: Move the checking logic to MemDep!
338 CallInst* cdep = nullptr;
339
340 // Check to see if we have a single dominating call instruction that is
341 // identical to C.
342 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
343 const NonLocalDepEntry *I = &deps[i];
344 if (I->getResult().isNonLocal())
345 continue;
346
347 // We don't handle non-definitions. If we already have a call, reject
348 // instruction dependencies.
349 if (!I->getResult().isDef() || cdep != nullptr) {
350 cdep = nullptr;
351 break;
352 }
353
354 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
355 // FIXME: All duplicated with non-local case.
356 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
357 cdep = NonLocalDepCall;
358 continue;
359 }
360
361 cdep = nullptr;
362 break;
363 }
364
365 if (!cdep) {
366 valueNumbering[C] = nextValueNumber;
367 return nextValueNumber++;
368 }
369
370 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
371 valueNumbering[C] = nextValueNumber;
372 return nextValueNumber++;
373 }
374 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
375 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
376 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
377 if (c_vn != cd_vn) {
378 valueNumbering[C] = nextValueNumber;
379 return nextValueNumber++;
380 }
381 }
382
383 uint32_t v = lookup_or_add(cdep);
384 valueNumbering[C] = v;
385 return v;
386
387 } else {
388 valueNumbering[C] = nextValueNumber;
389 return nextValueNumber++;
390 }
391 }
392
393 /// Returns true if a value number exists for the specified value.
exists(Value * V) const394 bool ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
395
396 /// lookup_or_add - Returns the value number for the specified value, assigning
397 /// it a new number if it did not have one before.
lookup_or_add(Value * V)398 uint32_t ValueTable::lookup_or_add(Value *V) {
399 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
400 if (VI != valueNumbering.end())
401 return VI->second;
402
403 if (!isa<Instruction>(V)) {
404 valueNumbering[V] = nextValueNumber;
405 return nextValueNumber++;
406 }
407
408 Instruction* I = cast<Instruction>(V);
409 Expression exp;
410 switch (I->getOpcode()) {
411 case Instruction::Call:
412 return lookup_or_add_call(cast<CallInst>(I));
413 case Instruction::Add:
414 case Instruction::FAdd:
415 case Instruction::Sub:
416 case Instruction::FSub:
417 case Instruction::Mul:
418 case Instruction::FMul:
419 case Instruction::UDiv:
420 case Instruction::SDiv:
421 case Instruction::FDiv:
422 case Instruction::URem:
423 case Instruction::SRem:
424 case Instruction::FRem:
425 case Instruction::Shl:
426 case Instruction::LShr:
427 case Instruction::AShr:
428 case Instruction::And:
429 case Instruction::Or:
430 case Instruction::Xor:
431 case Instruction::ICmp:
432 case Instruction::FCmp:
433 case Instruction::Trunc:
434 case Instruction::ZExt:
435 case Instruction::SExt:
436 case Instruction::FPToUI:
437 case Instruction::FPToSI:
438 case Instruction::UIToFP:
439 case Instruction::SIToFP:
440 case Instruction::FPTrunc:
441 case Instruction::FPExt:
442 case Instruction::PtrToInt:
443 case Instruction::IntToPtr:
444 case Instruction::BitCast:
445 case Instruction::Select:
446 case Instruction::ExtractElement:
447 case Instruction::InsertElement:
448 case Instruction::ShuffleVector:
449 case Instruction::InsertValue:
450 case Instruction::GetElementPtr:
451 exp = create_expression(I);
452 break;
453 case Instruction::ExtractValue:
454 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
455 break;
456 default:
457 valueNumbering[V] = nextValueNumber;
458 return nextValueNumber++;
459 }
460
461 uint32_t& e = expressionNumbering[exp];
462 if (!e) e = nextValueNumber++;
463 valueNumbering[V] = e;
464 return e;
465 }
466
467 /// Returns the value number of the specified value. Fails if
468 /// the value has not yet been numbered.
lookup(Value * V) const469 uint32_t ValueTable::lookup(Value *V) const {
470 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
471 assert(VI != valueNumbering.end() && "Value not numbered?");
472 return VI->second;
473 }
474
475 /// Returns the value number of the given comparison,
476 /// assigning it a new number if it did not have one before. Useful when
477 /// we deduced the result of a comparison, but don't immediately have an
478 /// instruction realizing that comparison to hand.
lookup_or_add_cmp(unsigned Opcode,CmpInst::Predicate Predicate,Value * LHS,Value * RHS)479 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
480 CmpInst::Predicate Predicate,
481 Value *LHS, Value *RHS) {
482 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
483 uint32_t& e = expressionNumbering[exp];
484 if (!e) e = nextValueNumber++;
485 return e;
486 }
487
488 /// Remove all entries from the ValueTable.
clear()489 void ValueTable::clear() {
490 valueNumbering.clear();
491 expressionNumbering.clear();
492 nextValueNumber = 1;
493 }
494
495 /// Remove a value from the value numbering.
erase(Value * V)496 void ValueTable::erase(Value *V) {
497 valueNumbering.erase(V);
498 }
499
500 /// verifyRemoved - Verify that the value is removed from all internal data
501 /// structures.
verifyRemoved(const Value * V) const502 void ValueTable::verifyRemoved(const Value *V) const {
503 for (DenseMap<Value*, uint32_t>::const_iterator
504 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
505 assert(I->first != V && "Inst still occurs in value numbering map!");
506 }
507 }
508
509 //===----------------------------------------------------------------------===//
510 // GVN Pass
511 //===----------------------------------------------------------------------===//
512
513 namespace {
514 class GVN;
515 struct AvailableValueInBlock {
516 /// BB - The basic block in question.
517 BasicBlock *BB;
518 enum ValType {
519 SimpleVal, // A simple offsetted value that is accessed.
520 LoadVal, // A value produced by a load.
521 MemIntrin, // A memory intrinsic which is loaded from.
522 UndefVal // A UndefValue representing a value from dead block (which
523 // is not yet physically removed from the CFG).
524 };
525
526 /// V - The value that is live out of the block.
527 PointerIntPair<Value *, 2, ValType> Val;
528
529 /// Offset - The byte offset in Val that is interesting for the load query.
530 unsigned Offset;
531
get__anon4436d4200211::AvailableValueInBlock532 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
533 unsigned Offset = 0) {
534 AvailableValueInBlock Res;
535 Res.BB = BB;
536 Res.Val.setPointer(V);
537 Res.Val.setInt(SimpleVal);
538 Res.Offset = Offset;
539 return Res;
540 }
541
getMI__anon4436d4200211::AvailableValueInBlock542 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
543 unsigned Offset = 0) {
544 AvailableValueInBlock Res;
545 Res.BB = BB;
546 Res.Val.setPointer(MI);
547 Res.Val.setInt(MemIntrin);
548 Res.Offset = Offset;
549 return Res;
550 }
551
getLoad__anon4436d4200211::AvailableValueInBlock552 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
553 unsigned Offset = 0) {
554 AvailableValueInBlock Res;
555 Res.BB = BB;
556 Res.Val.setPointer(LI);
557 Res.Val.setInt(LoadVal);
558 Res.Offset = Offset;
559 return Res;
560 }
561
getUndef__anon4436d4200211::AvailableValueInBlock562 static AvailableValueInBlock getUndef(BasicBlock *BB) {
563 AvailableValueInBlock Res;
564 Res.BB = BB;
565 Res.Val.setPointer(nullptr);
566 Res.Val.setInt(UndefVal);
567 Res.Offset = 0;
568 return Res;
569 }
570
isSimpleValue__anon4436d4200211::AvailableValueInBlock571 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
isCoercedLoadValue__anon4436d4200211::AvailableValueInBlock572 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
isMemIntrinValue__anon4436d4200211::AvailableValueInBlock573 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
isUndefValue__anon4436d4200211::AvailableValueInBlock574 bool isUndefValue() const { return Val.getInt() == UndefVal; }
575
getSimpleValue__anon4436d4200211::AvailableValueInBlock576 Value *getSimpleValue() const {
577 assert(isSimpleValue() && "Wrong accessor");
578 return Val.getPointer();
579 }
580
getCoercedLoadValue__anon4436d4200211::AvailableValueInBlock581 LoadInst *getCoercedLoadValue() const {
582 assert(isCoercedLoadValue() && "Wrong accessor");
583 return cast<LoadInst>(Val.getPointer());
584 }
585
getMemIntrinValue__anon4436d4200211::AvailableValueInBlock586 MemIntrinsic *getMemIntrinValue() const {
587 assert(isMemIntrinValue() && "Wrong accessor");
588 return cast<MemIntrinsic>(Val.getPointer());
589 }
590
591 /// Emit code into this block to adjust the value defined here to the
592 /// specified type. This handles various coercion cases.
593 Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const;
594 };
595
596 class GVN : public FunctionPass {
597 bool NoLoads;
598 MemoryDependenceAnalysis *MD;
599 DominatorTree *DT;
600 const TargetLibraryInfo *TLI;
601 AssumptionCache *AC;
602 SetVector<BasicBlock *> DeadBlocks;
603
604 ValueTable VN;
605
606 /// A mapping from value numbers to lists of Value*'s that
607 /// have that value number. Use findLeader to query it.
608 struct LeaderTableEntry {
609 Value *Val;
610 const BasicBlock *BB;
611 LeaderTableEntry *Next;
612 };
613 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
614 BumpPtrAllocator TableAllocator;
615
616 // Block-local map of equivalent values to their leader, does not
617 // propagate to any successors. Entries added mid-block are applied
618 // to the remaining instructions in the block.
619 SmallMapVector<llvm::Value *, llvm::Constant *, 4> ReplaceWithConstMap;
620 SmallVector<Instruction*, 8> InstrsToErase;
621
622 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
623 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
624 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
625
626 public:
627 static char ID; // Pass identification, replacement for typeid
GVN(bool noloads=false)628 explicit GVN(bool noloads = false)
629 : FunctionPass(ID), NoLoads(noloads), MD(nullptr) {
630 initializeGVNPass(*PassRegistry::getPassRegistry());
631 }
632
633 bool runOnFunction(Function &F) override;
634
635 /// This removes the specified instruction from
636 /// our various maps and marks it for deletion.
markInstructionForDeletion(Instruction * I)637 void markInstructionForDeletion(Instruction *I) {
638 VN.erase(I);
639 InstrsToErase.push_back(I);
640 }
641
getDominatorTree() const642 DominatorTree &getDominatorTree() const { return *DT; }
getAliasAnalysis() const643 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
getMemDep() const644 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
645 private:
646 /// Push a new Value to the LeaderTable onto the list for its value number.
addToLeaderTable(uint32_t N,Value * V,const BasicBlock * BB)647 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
648 LeaderTableEntry &Curr = LeaderTable[N];
649 if (!Curr.Val) {
650 Curr.Val = V;
651 Curr.BB = BB;
652 return;
653 }
654
655 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
656 Node->Val = V;
657 Node->BB = BB;
658 Node->Next = Curr.Next;
659 Curr.Next = Node;
660 }
661
662 /// Scan the list of values corresponding to a given
663 /// value number, and remove the given instruction if encountered.
removeFromLeaderTable(uint32_t N,Instruction * I,BasicBlock * BB)664 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
665 LeaderTableEntry* Prev = nullptr;
666 LeaderTableEntry* Curr = &LeaderTable[N];
667
668 while (Curr && (Curr->Val != I || Curr->BB != BB)) {
669 Prev = Curr;
670 Curr = Curr->Next;
671 }
672
673 if (!Curr)
674 return;
675
676 if (Prev) {
677 Prev->Next = Curr->Next;
678 } else {
679 if (!Curr->Next) {
680 Curr->Val = nullptr;
681 Curr->BB = nullptr;
682 } else {
683 LeaderTableEntry* Next = Curr->Next;
684 Curr->Val = Next->Val;
685 Curr->BB = Next->BB;
686 Curr->Next = Next->Next;
687 }
688 }
689 }
690
691 // List of critical edges to be split between iterations.
692 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
693
694 // This transformation requires dominator postdominator info
getAnalysisUsage(AnalysisUsage & AU) const695 void getAnalysisUsage(AnalysisUsage &AU) const override {
696 AU.addRequired<AssumptionCacheTracker>();
697 AU.addRequired<DominatorTreeWrapperPass>();
698 AU.addRequired<TargetLibraryInfoWrapperPass>();
699 if (!NoLoads)
700 AU.addRequired<MemoryDependenceAnalysis>();
701 AU.addRequired<AAResultsWrapperPass>();
702
703 AU.addPreserved<DominatorTreeWrapperPass>();
704 AU.addPreserved<GlobalsAAWrapperPass>();
705 }
706
707
708 // Helper functions of redundant load elimination
709 bool processLoad(LoadInst *L);
710 bool processNonLocalLoad(LoadInst *L);
711 bool processAssumeIntrinsic(IntrinsicInst *II);
712 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
713 AvailValInBlkVect &ValuesPerBlock,
714 UnavailBlkVect &UnavailableBlocks);
715 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
716 UnavailBlkVect &UnavailableBlocks);
717
718 // Other helper routines
719 bool processInstruction(Instruction *I);
720 bool processBlock(BasicBlock *BB);
721 void dump(DenseMap<uint32_t, Value*> &d);
722 bool iterateOnFunction(Function &F);
723 bool performPRE(Function &F);
724 bool performScalarPRE(Instruction *I);
725 bool performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
726 unsigned int ValNo);
727 Value *findLeader(const BasicBlock *BB, uint32_t num);
728 void cleanupGlobalSets();
729 void verifyRemoved(const Instruction *I) const;
730 bool splitCriticalEdges();
731 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
732 bool replaceOperandsWithConsts(Instruction *I) const;
733 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
734 bool DominatesByEdge);
735 bool processFoldableCondBr(BranchInst *BI);
736 void addDeadBlock(BasicBlock *BB);
737 void assignValNumForDeadCode();
738 };
739
740 char GVN::ID = 0;
741 }
742
743 // The public interface to this file...
createGVNPass(bool NoLoads)744 FunctionPass *llvm::createGVNPass(bool NoLoads) {
745 return new GVN(NoLoads);
746 }
747
748 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)749 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
750 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
751 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
752 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
753 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
754 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
755 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
756
757 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
758 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
759 errs() << "{\n";
760 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
761 E = d.end(); I != E; ++I) {
762 errs() << I->first << "\n";
763 I->second->dump();
764 }
765 errs() << "}\n";
766 }
767 #endif
768
769 /// Return true if we can prove that the value
770 /// we're analyzing is fully available in the specified block. As we go, keep
771 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
772 /// map is actually a tri-state map with the following values:
773 /// 0) we know the block *is not* fully available.
774 /// 1) we know the block *is* fully available.
775 /// 2) we do not know whether the block is fully available or not, but we are
776 /// currently speculating that it will be.
777 /// 3) we are speculating for this block and have used that to speculate for
778 /// other blocks.
IsValueFullyAvailableInBlock(BasicBlock * BB,DenseMap<BasicBlock *,char> & FullyAvailableBlocks,uint32_t RecurseDepth)779 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
780 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
781 uint32_t RecurseDepth) {
782 if (RecurseDepth > MaxRecurseDepth)
783 return false;
784
785 // Optimistically assume that the block is fully available and check to see
786 // if we already know about this block in one lookup.
787 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
788 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
789
790 // If the entry already existed for this block, return the precomputed value.
791 if (!IV.second) {
792 // If this is a speculative "available" value, mark it as being used for
793 // speculation of other blocks.
794 if (IV.first->second == 2)
795 IV.first->second = 3;
796 return IV.first->second != 0;
797 }
798
799 // Otherwise, see if it is fully available in all predecessors.
800 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
801
802 // If this block has no predecessors, it isn't live-in here.
803 if (PI == PE)
804 goto SpeculationFailure;
805
806 for (; PI != PE; ++PI)
807 // If the value isn't fully available in one of our predecessors, then it
808 // isn't fully available in this block either. Undo our previous
809 // optimistic assumption and bail out.
810 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
811 goto SpeculationFailure;
812
813 return true;
814
815 // If we get here, we found out that this is not, after
816 // all, a fully-available block. We have a problem if we speculated on this and
817 // used the speculation to mark other blocks as available.
818 SpeculationFailure:
819 char &BBVal = FullyAvailableBlocks[BB];
820
821 // If we didn't speculate on this, just return with it set to false.
822 if (BBVal == 2) {
823 BBVal = 0;
824 return false;
825 }
826
827 // If we did speculate on this value, we could have blocks set to 1 that are
828 // incorrect. Walk the (transitive) successors of this block and mark them as
829 // 0 if set to one.
830 SmallVector<BasicBlock*, 32> BBWorklist;
831 BBWorklist.push_back(BB);
832
833 do {
834 BasicBlock *Entry = BBWorklist.pop_back_val();
835 // Note that this sets blocks to 0 (unavailable) if they happen to not
836 // already be in FullyAvailableBlocks. This is safe.
837 char &EntryVal = FullyAvailableBlocks[Entry];
838 if (EntryVal == 0) continue; // Already unavailable.
839
840 // Mark as unavailable.
841 EntryVal = 0;
842
843 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
844 } while (!BBWorklist.empty());
845
846 return false;
847 }
848
849
850 /// Return true if CoerceAvailableValueToLoadType will succeed.
CanCoerceMustAliasedValueToLoad(Value * StoredVal,Type * LoadTy,const DataLayout & DL)851 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
852 Type *LoadTy,
853 const DataLayout &DL) {
854 // If the loaded or stored value is an first class array or struct, don't try
855 // to transform them. We need to be able to bitcast to integer.
856 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
857 StoredVal->getType()->isStructTy() ||
858 StoredVal->getType()->isArrayTy())
859 return false;
860
861 // The store has to be at least as big as the load.
862 if (DL.getTypeSizeInBits(StoredVal->getType()) <
863 DL.getTypeSizeInBits(LoadTy))
864 return false;
865
866 return true;
867 }
868
869 /// If we saw a store of a value to memory, and
870 /// then a load from a must-aliased pointer of a different type, try to coerce
871 /// the stored value. LoadedTy is the type of the load we want to replace.
872 /// IRB is IRBuilder used to insert new instructions.
873 ///
874 /// If we can't do it, return null.
CoerceAvailableValueToLoadType(Value * StoredVal,Type * LoadedTy,IRBuilder<> & IRB,const DataLayout & DL)875 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy,
876 IRBuilder<> &IRB,
877 const DataLayout &DL) {
878 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
879 return nullptr;
880
881 // If this is already the right type, just return it.
882 Type *StoredValTy = StoredVal->getType();
883
884 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
885 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
886
887 // If the store and reload are the same size, we can always reuse it.
888 if (StoreSize == LoadSize) {
889 // Pointer to Pointer -> use bitcast.
890 if (StoredValTy->getScalarType()->isPointerTy() &&
891 LoadedTy->getScalarType()->isPointerTy())
892 return IRB.CreateBitCast(StoredVal, LoadedTy);
893
894 // Convert source pointers to integers, which can be bitcast.
895 if (StoredValTy->getScalarType()->isPointerTy()) {
896 StoredValTy = DL.getIntPtrType(StoredValTy);
897 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
898 }
899
900 Type *TypeToCastTo = LoadedTy;
901 if (TypeToCastTo->getScalarType()->isPointerTy())
902 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
903
904 if (StoredValTy != TypeToCastTo)
905 StoredVal = IRB.CreateBitCast(StoredVal, TypeToCastTo);
906
907 // Cast to pointer if the load needs a pointer type.
908 if (LoadedTy->getScalarType()->isPointerTy())
909 StoredVal = IRB.CreateIntToPtr(StoredVal, LoadedTy);
910
911 return StoredVal;
912 }
913
914 // If the loaded value is smaller than the available value, then we can
915 // extract out a piece from it. If the available value is too small, then we
916 // can't do anything.
917 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
918
919 // Convert source pointers to integers, which can be manipulated.
920 if (StoredValTy->getScalarType()->isPointerTy()) {
921 StoredValTy = DL.getIntPtrType(StoredValTy);
922 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
923 }
924
925 // Convert vectors and fp to integer, which can be manipulated.
926 if (!StoredValTy->isIntegerTy()) {
927 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
928 StoredVal = IRB.CreateBitCast(StoredVal, StoredValTy);
929 }
930
931 // If this is a big-endian system, we need to shift the value down to the low
932 // bits so that a truncate will work.
933 if (DL.isBigEndian()) {
934 StoredVal = IRB.CreateLShr(StoredVal, StoreSize - LoadSize, "tmp");
935 }
936
937 // Truncate the integer to the right size now.
938 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
939 StoredVal = IRB.CreateTrunc(StoredVal, NewIntTy, "trunc");
940
941 if (LoadedTy == NewIntTy)
942 return StoredVal;
943
944 // If the result is a pointer, inttoptr.
945 if (LoadedTy->getScalarType()->isPointerTy())
946 return IRB.CreateIntToPtr(StoredVal, LoadedTy, "inttoptr");
947
948 // Otherwise, bitcast.
949 return IRB.CreateBitCast(StoredVal, LoadedTy, "bitcast");
950 }
951
952 /// This function is called when we have a
953 /// memdep query of a load that ends up being a clobbering memory write (store,
954 /// memset, memcpy, memmove). This means that the write *may* provide bits used
955 /// by the load but we can't be sure because the pointers don't mustalias.
956 ///
957 /// Check this case to see if there is anything more we can do before we give
958 /// up. This returns -1 if we have to give up, or a byte number in the stored
959 /// value of the piece that feeds the load.
AnalyzeLoadFromClobberingWrite(Type * LoadTy,Value * LoadPtr,Value * WritePtr,uint64_t WriteSizeInBits,const DataLayout & DL)960 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
961 Value *WritePtr,
962 uint64_t WriteSizeInBits,
963 const DataLayout &DL) {
964 // If the loaded or stored value is a first class array or struct, don't try
965 // to transform them. We need to be able to bitcast to integer.
966 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
967 return -1;
968
969 int64_t StoreOffset = 0, LoadOffset = 0;
970 Value *StoreBase =
971 GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL);
972 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL);
973 if (StoreBase != LoadBase)
974 return -1;
975
976 // If the load and store are to the exact same address, they should have been
977 // a must alias. AA must have gotten confused.
978 // FIXME: Study to see if/when this happens. One case is forwarding a memset
979 // to a load from the base of the memset.
980 #if 0
981 if (LoadOffset == StoreOffset) {
982 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
983 << "Base = " << *StoreBase << "\n"
984 << "Store Ptr = " << *WritePtr << "\n"
985 << "Store Offs = " << StoreOffset << "\n"
986 << "Load Ptr = " << *LoadPtr << "\n";
987 abort();
988 }
989 #endif
990
991 // If the load and store don't overlap at all, the store doesn't provide
992 // anything to the load. In this case, they really don't alias at all, AA
993 // must have gotten confused.
994 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
995
996 if ((WriteSizeInBits & 7) | (LoadSize & 7))
997 return -1;
998 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
999 LoadSize >>= 3;
1000
1001
1002 bool isAAFailure = false;
1003 if (StoreOffset < LoadOffset)
1004 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1005 else
1006 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1007
1008 if (isAAFailure) {
1009 #if 0
1010 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1011 << "Base = " << *StoreBase << "\n"
1012 << "Store Ptr = " << *WritePtr << "\n"
1013 << "Store Offs = " << StoreOffset << "\n"
1014 << "Load Ptr = " << *LoadPtr << "\n";
1015 abort();
1016 #endif
1017 return -1;
1018 }
1019
1020 // If the Load isn't completely contained within the stored bits, we don't
1021 // have all the bits to feed it. We could do something crazy in the future
1022 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1023 // valuable.
1024 if (StoreOffset > LoadOffset ||
1025 StoreOffset+StoreSize < LoadOffset+LoadSize)
1026 return -1;
1027
1028 // Okay, we can do this transformation. Return the number of bytes into the
1029 // store that the load is.
1030 return LoadOffset-StoreOffset;
1031 }
1032
1033 /// This function is called when we have a
1034 /// memdep query of a load that ends up being a clobbering store.
AnalyzeLoadFromClobberingStore(Type * LoadTy,Value * LoadPtr,StoreInst * DepSI)1035 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1036 StoreInst *DepSI) {
1037 // Cannot handle reading from store of first-class aggregate yet.
1038 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1039 DepSI->getValueOperand()->getType()->isArrayTy())
1040 return -1;
1041
1042 const DataLayout &DL = DepSI->getModule()->getDataLayout();
1043 Value *StorePtr = DepSI->getPointerOperand();
1044 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1045 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1046 StorePtr, StoreSize, DL);
1047 }
1048
1049 /// This function is called when we have a
1050 /// memdep query of a load that ends up being clobbered by another load. See if
1051 /// the other load can feed into the second load.
AnalyzeLoadFromClobberingLoad(Type * LoadTy,Value * LoadPtr,LoadInst * DepLI,const DataLayout & DL)1052 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1053 LoadInst *DepLI, const DataLayout &DL){
1054 // Cannot handle reading from store of first-class aggregate yet.
1055 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1056 return -1;
1057
1058 Value *DepPtr = DepLI->getPointerOperand();
1059 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
1060 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
1061 if (R != -1) return R;
1062
1063 // If we have a load/load clobber an DepLI can be widened to cover this load,
1064 // then we should widen it!
1065 int64_t LoadOffs = 0;
1066 const Value *LoadBase =
1067 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, DL);
1068 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1069
1070 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
1071 LoadBase, LoadOffs, LoadSize, DepLI);
1072 if (Size == 0) return -1;
1073
1074 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
1075 }
1076
1077
1078
AnalyzeLoadFromClobberingMemInst(Type * LoadTy,Value * LoadPtr,MemIntrinsic * MI,const DataLayout & DL)1079 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1080 MemIntrinsic *MI,
1081 const DataLayout &DL) {
1082 // If the mem operation is a non-constant size, we can't handle it.
1083 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1084 if (!SizeCst) return -1;
1085 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1086
1087 // If this is memset, we just need to see if the offset is valid in the size
1088 // of the memset..
1089 if (MI->getIntrinsicID() == Intrinsic::memset)
1090 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1091 MemSizeInBits, DL);
1092
1093 // If we have a memcpy/memmove, the only case we can handle is if this is a
1094 // copy from constant memory. In that case, we can read directly from the
1095 // constant memory.
1096 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1097
1098 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1099 if (!Src) return -1;
1100
1101 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, DL));
1102 if (!GV || !GV->isConstant()) return -1;
1103
1104 // See if the access is within the bounds of the transfer.
1105 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1106 MI->getDest(), MemSizeInBits, DL);
1107 if (Offset == -1)
1108 return Offset;
1109
1110 unsigned AS = Src->getType()->getPointerAddressSpace();
1111 // Otherwise, see if we can constant fold a load from the constant with the
1112 // offset applied as appropriate.
1113 Src = ConstantExpr::getBitCast(Src,
1114 Type::getInt8PtrTy(Src->getContext(), AS));
1115 Constant *OffsetCst =
1116 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1117 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1118 OffsetCst);
1119 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1120 if (ConstantFoldLoadFromConstPtr(Src, DL))
1121 return Offset;
1122 return -1;
1123 }
1124
1125
1126 /// This function is called when we have a
1127 /// memdep query of a load that ends up being a clobbering store. This means
1128 /// that the store provides bits used by the load but we the pointers don't
1129 /// mustalias. Check this case to see if there is anything more we can do
1130 /// before we give up.
GetStoreValueForLoad(Value * SrcVal,unsigned Offset,Type * LoadTy,Instruction * InsertPt,const DataLayout & DL)1131 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1132 Type *LoadTy,
1133 Instruction *InsertPt, const DataLayout &DL){
1134 LLVMContext &Ctx = SrcVal->getType()->getContext();
1135
1136 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1137 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1138
1139 IRBuilder<> Builder(InsertPt);
1140
1141 // Compute which bits of the stored value are being used by the load. Convert
1142 // to an integer type to start with.
1143 if (SrcVal->getType()->getScalarType()->isPointerTy())
1144 SrcVal = Builder.CreatePtrToInt(SrcVal,
1145 DL.getIntPtrType(SrcVal->getType()));
1146 if (!SrcVal->getType()->isIntegerTy())
1147 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1148
1149 // Shift the bits to the least significant depending on endianness.
1150 unsigned ShiftAmt;
1151 if (DL.isLittleEndian())
1152 ShiftAmt = Offset*8;
1153 else
1154 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1155
1156 if (ShiftAmt)
1157 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1158
1159 if (LoadSize != StoreSize)
1160 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1161
1162 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL);
1163 }
1164
1165 /// This function is called when we have a
1166 /// memdep query of a load that ends up being a clobbering load. This means
1167 /// that the load *may* provide bits used by the load but we can't be sure
1168 /// because the pointers don't mustalias. Check this case to see if there is
1169 /// anything more we can do before we give up.
GetLoadValueForLoad(LoadInst * SrcVal,unsigned Offset,Type * LoadTy,Instruction * InsertPt,GVN & gvn)1170 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1171 Type *LoadTy, Instruction *InsertPt,
1172 GVN &gvn) {
1173 const DataLayout &DL = SrcVal->getModule()->getDataLayout();
1174 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1175 // widen SrcVal out to a larger load.
1176 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
1177 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1178 if (Offset+LoadSize > SrcValSize) {
1179 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1180 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1181 // If we have a load/load clobber an DepLI can be widened to cover this
1182 // load, then we should widen it to the next power of 2 size big enough!
1183 unsigned NewLoadSize = Offset+LoadSize;
1184 if (!isPowerOf2_32(NewLoadSize))
1185 NewLoadSize = NextPowerOf2(NewLoadSize);
1186
1187 Value *PtrVal = SrcVal->getPointerOperand();
1188
1189 // Insert the new load after the old load. This ensures that subsequent
1190 // memdep queries will find the new load. We can't easily remove the old
1191 // load completely because it is already in the value numbering table.
1192 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1193 Type *DestPTy =
1194 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1195 DestPTy = PointerType::get(DestPTy,
1196 PtrVal->getType()->getPointerAddressSpace());
1197 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1198 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1199 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1200 NewLoad->takeName(SrcVal);
1201 NewLoad->setAlignment(SrcVal->getAlignment());
1202
1203 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1204 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1205
1206 // Replace uses of the original load with the wider load. On a big endian
1207 // system, we need to shift down to get the relevant bits.
1208 Value *RV = NewLoad;
1209 if (DL.isBigEndian())
1210 RV = Builder.CreateLShr(RV,
1211 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1212 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1213 SrcVal->replaceAllUsesWith(RV);
1214
1215 // We would like to use gvn.markInstructionForDeletion here, but we can't
1216 // because the load is already memoized into the leader map table that GVN
1217 // tracks. It is potentially possible to remove the load from the table,
1218 // but then there all of the operations based on it would need to be
1219 // rehashed. Just leave the dead load around.
1220 gvn.getMemDep().removeInstruction(SrcVal);
1221 SrcVal = NewLoad;
1222 }
1223
1224 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1225 }
1226
1227
1228 /// This function is called when we have a
1229 /// memdep query of a load that ends up being a clobbering mem intrinsic.
GetMemInstValueForLoad(MemIntrinsic * SrcInst,unsigned Offset,Type * LoadTy,Instruction * InsertPt,const DataLayout & DL)1230 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1231 Type *LoadTy, Instruction *InsertPt,
1232 const DataLayout &DL){
1233 LLVMContext &Ctx = LoadTy->getContext();
1234 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1235
1236 IRBuilder<> Builder(InsertPt);
1237
1238 // We know that this method is only called when the mem transfer fully
1239 // provides the bits for the load.
1240 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1241 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1242 // independently of what the offset is.
1243 Value *Val = MSI->getValue();
1244 if (LoadSize != 1)
1245 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1246
1247 Value *OneElt = Val;
1248
1249 // Splat the value out to the right number of bits.
1250 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1251 // If we can double the number of bytes set, do it.
1252 if (NumBytesSet*2 <= LoadSize) {
1253 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1254 Val = Builder.CreateOr(Val, ShVal);
1255 NumBytesSet <<= 1;
1256 continue;
1257 }
1258
1259 // Otherwise insert one byte at a time.
1260 Value *ShVal = Builder.CreateShl(Val, 1*8);
1261 Val = Builder.CreateOr(OneElt, ShVal);
1262 ++NumBytesSet;
1263 }
1264
1265 return CoerceAvailableValueToLoadType(Val, LoadTy, Builder, DL);
1266 }
1267
1268 // Otherwise, this is a memcpy/memmove from a constant global.
1269 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1270 Constant *Src = cast<Constant>(MTI->getSource());
1271 unsigned AS = Src->getType()->getPointerAddressSpace();
1272
1273 // Otherwise, see if we can constant fold a load from the constant with the
1274 // offset applied as appropriate.
1275 Src = ConstantExpr::getBitCast(Src,
1276 Type::getInt8PtrTy(Src->getContext(), AS));
1277 Constant *OffsetCst =
1278 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1279 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1280 OffsetCst);
1281 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1282 return ConstantFoldLoadFromConstPtr(Src, DL);
1283 }
1284
1285
1286 /// Given a set of loads specified by ValuesPerBlock,
1287 /// construct SSA form, allowing us to eliminate LI. This returns the value
1288 /// that should be used at LI's definition site.
ConstructSSAForLoadSet(LoadInst * LI,SmallVectorImpl<AvailableValueInBlock> & ValuesPerBlock,GVN & gvn)1289 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1290 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1291 GVN &gvn) {
1292 // Check for the fully redundant, dominating load case. In this case, we can
1293 // just use the dominating value directly.
1294 if (ValuesPerBlock.size() == 1 &&
1295 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1296 LI->getParent())) {
1297 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1298 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
1299 }
1300
1301 // Otherwise, we have to construct SSA form.
1302 SmallVector<PHINode*, 8> NewPHIs;
1303 SSAUpdater SSAUpdate(&NewPHIs);
1304 SSAUpdate.Initialize(LI->getType(), LI->getName());
1305
1306 for (const AvailableValueInBlock &AV : ValuesPerBlock) {
1307 BasicBlock *BB = AV.BB;
1308
1309 if (SSAUpdate.HasValueForBlock(BB))
1310 continue;
1311
1312 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
1313 }
1314
1315 // Perform PHI construction.
1316 return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1317 }
1318
MaterializeAdjustedValue(LoadInst * LI,GVN & gvn) const1319 Value *AvailableValueInBlock::MaterializeAdjustedValue(LoadInst *LI,
1320 GVN &gvn) const {
1321 Value *Res;
1322 Type *LoadTy = LI->getType();
1323 const DataLayout &DL = LI->getModule()->getDataLayout();
1324 if (isSimpleValue()) {
1325 Res = getSimpleValue();
1326 if (Res->getType() != LoadTy) {
1327 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), DL);
1328
1329 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1330 << *getSimpleValue() << '\n'
1331 << *Res << '\n' << "\n\n\n");
1332 }
1333 } else if (isCoercedLoadValue()) {
1334 LoadInst *Load = getCoercedLoadValue();
1335 if (Load->getType() == LoadTy && Offset == 0) {
1336 Res = Load;
1337 } else {
1338 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1339 gvn);
1340
1341 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1342 << *getCoercedLoadValue() << '\n'
1343 << *Res << '\n' << "\n\n\n");
1344 }
1345 } else if (isMemIntrinValue()) {
1346 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
1347 BB->getTerminator(), DL);
1348 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1349 << " " << *getMemIntrinValue() << '\n'
1350 << *Res << '\n' << "\n\n\n");
1351 } else {
1352 assert(isUndefValue() && "Should be UndefVal");
1353 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1354 return UndefValue::get(LoadTy);
1355 }
1356 return Res;
1357 }
1358
isLifetimeStart(const Instruction * Inst)1359 static bool isLifetimeStart(const Instruction *Inst) {
1360 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1361 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1362 return false;
1363 }
1364
AnalyzeLoadAvailability(LoadInst * LI,LoadDepVect & Deps,AvailValInBlkVect & ValuesPerBlock,UnavailBlkVect & UnavailableBlocks)1365 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1366 AvailValInBlkVect &ValuesPerBlock,
1367 UnavailBlkVect &UnavailableBlocks) {
1368
1369 // Filter out useless results (non-locals, etc). Keep track of the blocks
1370 // where we have a value available in repl, also keep track of whether we see
1371 // dependencies that produce an unknown value for the load (such as a call
1372 // that could potentially clobber the load).
1373 unsigned NumDeps = Deps.size();
1374 const DataLayout &DL = LI->getModule()->getDataLayout();
1375 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1376 BasicBlock *DepBB = Deps[i].getBB();
1377 MemDepResult DepInfo = Deps[i].getResult();
1378
1379 if (DeadBlocks.count(DepBB)) {
1380 // Dead dependent mem-op disguise as a load evaluating the same value
1381 // as the load in question.
1382 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1383 continue;
1384 }
1385
1386 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1387 UnavailableBlocks.push_back(DepBB);
1388 continue;
1389 }
1390
1391 if (DepInfo.isClobber()) {
1392 // The address being loaded in this non-local block may not be the same as
1393 // the pointer operand of the load if PHI translation occurs. Make sure
1394 // to consider the right address.
1395 Value *Address = Deps[i].getAddress();
1396
1397 // If the dependence is to a store that writes to a superset of the bits
1398 // read by the load, we can extract the bits we need for the load from the
1399 // stored value.
1400 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1401 if (Address) {
1402 int Offset =
1403 AnalyzeLoadFromClobberingStore(LI->getType(), Address, DepSI);
1404 if (Offset != -1) {
1405 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1406 DepSI->getValueOperand(),
1407 Offset));
1408 continue;
1409 }
1410 }
1411 }
1412
1413 // Check to see if we have something like this:
1414 // load i32* P
1415 // load i8* (P+1)
1416 // if we have this, replace the later with an extraction from the former.
1417 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1418 // If this is a clobber and L is the first instruction in its block, then
1419 // we have the first instruction in the entry block.
1420 if (DepLI != LI && Address) {
1421 int Offset =
1422 AnalyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
1423
1424 if (Offset != -1) {
1425 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1426 Offset));
1427 continue;
1428 }
1429 }
1430 }
1431
1432 // If the clobbering value is a memset/memcpy/memmove, see if we can
1433 // forward a value on from it.
1434 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1435 if (Address) {
1436 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1437 DepMI, DL);
1438 if (Offset != -1) {
1439 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1440 Offset));
1441 continue;
1442 }
1443 }
1444 }
1445
1446 UnavailableBlocks.push_back(DepBB);
1447 continue;
1448 }
1449
1450 // DepInfo.isDef() here
1451
1452 Instruction *DepInst = DepInfo.getInst();
1453
1454 // Loading the allocation -> undef.
1455 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1456 // Loading immediately after lifetime begin -> undef.
1457 isLifetimeStart(DepInst)) {
1458 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1459 UndefValue::get(LI->getType())));
1460 continue;
1461 }
1462
1463 // Loading from calloc (which zero initializes memory) -> zero
1464 if (isCallocLikeFn(DepInst, TLI)) {
1465 ValuesPerBlock.push_back(AvailableValueInBlock::get(
1466 DepBB, Constant::getNullValue(LI->getType())));
1467 continue;
1468 }
1469
1470 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1471 // Reject loads and stores that are to the same address but are of
1472 // different types if we have to.
1473 if (S->getValueOperand()->getType() != LI->getType()) {
1474 // If the stored value is larger or equal to the loaded value, we can
1475 // reuse it.
1476 if (!CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1477 LI->getType(), DL)) {
1478 UnavailableBlocks.push_back(DepBB);
1479 continue;
1480 }
1481 }
1482
1483 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1484 S->getValueOperand()));
1485 continue;
1486 }
1487
1488 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1489 // If the types mismatch and we can't handle it, reject reuse of the load.
1490 if (LD->getType() != LI->getType()) {
1491 // If the stored value is larger or equal to the loaded value, we can
1492 // reuse it.
1493 if (!CanCoerceMustAliasedValueToLoad(LD, LI->getType(), DL)) {
1494 UnavailableBlocks.push_back(DepBB);
1495 continue;
1496 }
1497 }
1498 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1499 continue;
1500 }
1501
1502 UnavailableBlocks.push_back(DepBB);
1503 }
1504 }
1505
PerformLoadPRE(LoadInst * LI,AvailValInBlkVect & ValuesPerBlock,UnavailBlkVect & UnavailableBlocks)1506 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1507 UnavailBlkVect &UnavailableBlocks) {
1508 // Okay, we have *some* definitions of the value. This means that the value
1509 // is available in some of our (transitive) predecessors. Lets think about
1510 // doing PRE of this load. This will involve inserting a new load into the
1511 // predecessor when it's not available. We could do this in general, but
1512 // prefer to not increase code size. As such, we only do this when we know
1513 // that we only have to insert *one* load (which means we're basically moving
1514 // the load, not inserting a new one).
1515
1516 SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1517 UnavailableBlocks.end());
1518
1519 // Let's find the first basic block with more than one predecessor. Walk
1520 // backwards through predecessors if needed.
1521 BasicBlock *LoadBB = LI->getParent();
1522 BasicBlock *TmpBB = LoadBB;
1523
1524 while (TmpBB->getSinglePredecessor()) {
1525 TmpBB = TmpBB->getSinglePredecessor();
1526 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1527 return false;
1528 if (Blockers.count(TmpBB))
1529 return false;
1530
1531 // If any of these blocks has more than one successor (i.e. if the edge we
1532 // just traversed was critical), then there are other paths through this
1533 // block along which the load may not be anticipated. Hoisting the load
1534 // above this block would be adding the load to execution paths along
1535 // which it was not previously executed.
1536 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1537 return false;
1538 }
1539
1540 assert(TmpBB);
1541 LoadBB = TmpBB;
1542
1543 // Check to see how many predecessors have the loaded value fully
1544 // available.
1545 MapVector<BasicBlock *, Value *> PredLoads;
1546 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1547 for (const AvailableValueInBlock &AV : ValuesPerBlock)
1548 FullyAvailableBlocks[AV.BB] = true;
1549 for (BasicBlock *UnavailableBB : UnavailableBlocks)
1550 FullyAvailableBlocks[UnavailableBB] = false;
1551
1552 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1553 for (BasicBlock *Pred : predecessors(LoadBB)) {
1554 // If any predecessor block is an EH pad that does not allow non-PHI
1555 // instructions before the terminator, we can't PRE the load.
1556 if (Pred->getTerminator()->isEHPad()) {
1557 DEBUG(dbgs()
1558 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1559 << Pred->getName() << "': " << *LI << '\n');
1560 return false;
1561 }
1562
1563 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1564 continue;
1565 }
1566
1567 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1568 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1569 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1570 << Pred->getName() << "': " << *LI << '\n');
1571 return false;
1572 }
1573
1574 if (LoadBB->isEHPad()) {
1575 DEBUG(dbgs()
1576 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1577 << Pred->getName() << "': " << *LI << '\n');
1578 return false;
1579 }
1580
1581 CriticalEdgePred.push_back(Pred);
1582 } else {
1583 // Only add the predecessors that will not be split for now.
1584 PredLoads[Pred] = nullptr;
1585 }
1586 }
1587
1588 // Decide whether PRE is profitable for this load.
1589 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1590 assert(NumUnavailablePreds != 0 &&
1591 "Fully available value should already be eliminated!");
1592
1593 // If this load is unavailable in multiple predecessors, reject it.
1594 // FIXME: If we could restructure the CFG, we could make a common pred with
1595 // all the preds that don't have an available LI and insert a new load into
1596 // that one block.
1597 if (NumUnavailablePreds != 1)
1598 return false;
1599
1600 // Split critical edges, and update the unavailable predecessors accordingly.
1601 for (BasicBlock *OrigPred : CriticalEdgePred) {
1602 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1603 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1604 PredLoads[NewPred] = nullptr;
1605 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1606 << LoadBB->getName() << '\n');
1607 }
1608
1609 // Check if the load can safely be moved to all the unavailable predecessors.
1610 bool CanDoPRE = true;
1611 const DataLayout &DL = LI->getModule()->getDataLayout();
1612 SmallVector<Instruction*, 8> NewInsts;
1613 for (auto &PredLoad : PredLoads) {
1614 BasicBlock *UnavailablePred = PredLoad.first;
1615
1616 // Do PHI translation to get its value in the predecessor if necessary. The
1617 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1618
1619 // If all preds have a single successor, then we know it is safe to insert
1620 // the load on the pred (?!?), so we can insert code to materialize the
1621 // pointer if it is not available.
1622 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1623 Value *LoadPtr = nullptr;
1624 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1625 *DT, NewInsts);
1626
1627 // If we couldn't find or insert a computation of this phi translated value,
1628 // we fail PRE.
1629 if (!LoadPtr) {
1630 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1631 << *LI->getPointerOperand() << "\n");
1632 CanDoPRE = false;
1633 break;
1634 }
1635
1636 PredLoad.second = LoadPtr;
1637 }
1638
1639 if (!CanDoPRE) {
1640 while (!NewInsts.empty()) {
1641 Instruction *I = NewInsts.pop_back_val();
1642 if (MD) MD->removeInstruction(I);
1643 I->eraseFromParent();
1644 }
1645 // HINT: Don't revert the edge-splitting as following transformation may
1646 // also need to split these critical edges.
1647 return !CriticalEdgePred.empty();
1648 }
1649
1650 // Okay, we can eliminate this load by inserting a reload in the predecessor
1651 // and using PHI construction to get the value in the other predecessors, do
1652 // it.
1653 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1654 DEBUG(if (!NewInsts.empty())
1655 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1656 << *NewInsts.back() << '\n');
1657
1658 // Assign value numbers to the new instructions.
1659 for (Instruction *I : NewInsts) {
1660 // FIXME: We really _ought_ to insert these value numbers into their
1661 // parent's availability map. However, in doing so, we risk getting into
1662 // ordering issues. If a block hasn't been processed yet, we would be
1663 // marking a value as AVAIL-IN, which isn't what we intend.
1664 VN.lookup_or_add(I);
1665 }
1666
1667 for (const auto &PredLoad : PredLoads) {
1668 BasicBlock *UnavailablePred = PredLoad.first;
1669 Value *LoadPtr = PredLoad.second;
1670
1671 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1672 LI->getAlignment(),
1673 UnavailablePred->getTerminator());
1674
1675 // Transfer the old load's AA tags to the new load.
1676 AAMDNodes Tags;
1677 LI->getAAMetadata(Tags);
1678 if (Tags)
1679 NewLoad->setAAMetadata(Tags);
1680
1681 if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1682 NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1683 if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1684 NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1685
1686 // Transfer DebugLoc.
1687 NewLoad->setDebugLoc(LI->getDebugLoc());
1688
1689 // Add the newly created load.
1690 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1691 NewLoad));
1692 MD->invalidateCachedPointerInfo(LoadPtr);
1693 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1694 }
1695
1696 // Perform PHI construction.
1697 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1698 LI->replaceAllUsesWith(V);
1699 if (isa<PHINode>(V))
1700 V->takeName(LI);
1701 if (Instruction *I = dyn_cast<Instruction>(V))
1702 I->setDebugLoc(LI->getDebugLoc());
1703 if (V->getType()->getScalarType()->isPointerTy())
1704 MD->invalidateCachedPointerInfo(V);
1705 markInstructionForDeletion(LI);
1706 ++NumPRELoad;
1707 return true;
1708 }
1709
1710 /// Attempt to eliminate a load whose dependencies are
1711 /// non-local by performing PHI construction.
processNonLocalLoad(LoadInst * LI)1712 bool GVN::processNonLocalLoad(LoadInst *LI) {
1713 // non-local speculations are not allowed under asan.
1714 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeAddress))
1715 return false;
1716
1717 // Step 1: Find the non-local dependencies of the load.
1718 LoadDepVect Deps;
1719 MD->getNonLocalPointerDependency(LI, Deps);
1720
1721 // If we had to process more than one hundred blocks to find the
1722 // dependencies, this load isn't worth worrying about. Optimizing
1723 // it will be too expensive.
1724 unsigned NumDeps = Deps.size();
1725 if (NumDeps > 100)
1726 return false;
1727
1728 // If we had a phi translation failure, we'll have a single entry which is a
1729 // clobber in the current block. Reject this early.
1730 if (NumDeps == 1 &&
1731 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1732 DEBUG(
1733 dbgs() << "GVN: non-local load ";
1734 LI->printAsOperand(dbgs());
1735 dbgs() << " has unknown dependencies\n";
1736 );
1737 return false;
1738 }
1739
1740 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1741 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1742 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1743 OE = GEP->idx_end();
1744 OI != OE; ++OI)
1745 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1746 performScalarPRE(I);
1747 }
1748
1749 // Step 2: Analyze the availability of the load
1750 AvailValInBlkVect ValuesPerBlock;
1751 UnavailBlkVect UnavailableBlocks;
1752 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1753
1754 // If we have no predecessors that produce a known value for this load, exit
1755 // early.
1756 if (ValuesPerBlock.empty())
1757 return false;
1758
1759 // Step 3: Eliminate fully redundancy.
1760 //
1761 // If all of the instructions we depend on produce a known value for this
1762 // load, then it is fully redundant and we can use PHI insertion to compute
1763 // its value. Insert PHIs and remove the fully redundant value now.
1764 if (UnavailableBlocks.empty()) {
1765 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1766
1767 // Perform PHI construction.
1768 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1769 LI->replaceAllUsesWith(V);
1770
1771 if (isa<PHINode>(V))
1772 V->takeName(LI);
1773 if (Instruction *I = dyn_cast<Instruction>(V))
1774 if (LI->getDebugLoc())
1775 I->setDebugLoc(LI->getDebugLoc());
1776 if (V->getType()->getScalarType()->isPointerTy())
1777 MD->invalidateCachedPointerInfo(V);
1778 markInstructionForDeletion(LI);
1779 ++NumGVNLoad;
1780 return true;
1781 }
1782
1783 // Step 4: Eliminate partial redundancy.
1784 if (!EnablePRE || !EnableLoadPRE)
1785 return false;
1786
1787 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1788 }
1789
processAssumeIntrinsic(IntrinsicInst * IntrinsicI)1790 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1791 assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1792 "This function can only be called with llvm.assume intrinsic");
1793 Value *V = IntrinsicI->getArgOperand(0);
1794
1795 if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1796 if (Cond->isZero()) {
1797 Type *Int8Ty = Type::getInt8Ty(V->getContext());
1798 // Insert a new store to null instruction before the load to indicate that
1799 // this code is not reachable. FIXME: We could insert unreachable
1800 // instruction directly because we can modify the CFG.
1801 new StoreInst(UndefValue::get(Int8Ty),
1802 Constant::getNullValue(Int8Ty->getPointerTo()),
1803 IntrinsicI);
1804 }
1805 markInstructionForDeletion(IntrinsicI);
1806 return false;
1807 }
1808
1809 Constant *True = ConstantInt::getTrue(V->getContext());
1810 bool Changed = false;
1811
1812 for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1813 BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1814
1815 // This property is only true in dominated successors, propagateEquality
1816 // will check dominance for us.
1817 Changed |= propagateEquality(V, True, Edge, false);
1818 }
1819
1820 // We can replace assume value with true, which covers cases like this:
1821 // call void @llvm.assume(i1 %cmp)
1822 // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1823 ReplaceWithConstMap[V] = True;
1824
1825 // If one of *cmp *eq operand is const, adding it to map will cover this:
1826 // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1827 // call void @llvm.assume(i1 %cmp)
1828 // ret float %0 ; will change it to ret float 3.000000e+00
1829 if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1830 if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1831 CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1832 (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1833 CmpI->getFastMathFlags().noNaNs())) {
1834 Value *CmpLHS = CmpI->getOperand(0);
1835 Value *CmpRHS = CmpI->getOperand(1);
1836 if (isa<Constant>(CmpLHS))
1837 std::swap(CmpLHS, CmpRHS);
1838 auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1839
1840 // If only one operand is constant.
1841 if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1842 ReplaceWithConstMap[CmpLHS] = RHSConst;
1843 }
1844 }
1845 return Changed;
1846 }
1847
patchReplacementInstruction(Instruction * I,Value * Repl)1848 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1849 // Patch the replacement so that it is not more restrictive than the value
1850 // being replaced.
1851 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1852 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1853 if (Op && ReplOp)
1854 ReplOp->andIRFlags(Op);
1855
1856 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1857 // FIXME: If both the original and replacement value are part of the
1858 // same control-flow region (meaning that the execution of one
1859 // guarantees the execution of the other), then we can combine the
1860 // noalias scopes here and do better than the general conservative
1861 // answer used in combineMetadata().
1862
1863 // In general, GVN unifies expressions over different control-flow
1864 // regions, and so we need a conservative combination of the noalias
1865 // scopes.
1866 static const unsigned KnownIDs[] = {
1867 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
1868 LLVMContext::MD_noalias, LLVMContext::MD_range,
1869 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
1870 LLVMContext::MD_invariant_group};
1871 combineMetadata(ReplInst, I, KnownIDs);
1872 }
1873 }
1874
patchAndReplaceAllUsesWith(Instruction * I,Value * Repl)1875 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1876 patchReplacementInstruction(I, Repl);
1877 I->replaceAllUsesWith(Repl);
1878 }
1879
1880 /// Attempt to eliminate a load, first by eliminating it
1881 /// locally, and then attempting non-local elimination if that fails.
processLoad(LoadInst * L)1882 bool GVN::processLoad(LoadInst *L) {
1883 if (!MD)
1884 return false;
1885
1886 if (!L->isSimple())
1887 return false;
1888
1889 if (L->use_empty()) {
1890 markInstructionForDeletion(L);
1891 return true;
1892 }
1893
1894 // ... to a pointer that has been loaded from before...
1895 MemDepResult Dep = MD->getDependency(L);
1896 const DataLayout &DL = L->getModule()->getDataLayout();
1897
1898 // If we have a clobber and target data is around, see if this is a clobber
1899 // that we can fix up through code synthesis.
1900 if (Dep.isClobber()) {
1901 // Check to see if we have something like this:
1902 // store i32 123, i32* %P
1903 // %A = bitcast i32* %P to i8*
1904 // %B = gep i8* %A, i32 1
1905 // %C = load i8* %B
1906 //
1907 // We could do that by recognizing if the clobber instructions are obviously
1908 // a common base + constant offset, and if the previous store (or memset)
1909 // completely covers this load. This sort of thing can happen in bitfield
1910 // access code.
1911 Value *AvailVal = nullptr;
1912 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1913 int Offset = AnalyzeLoadFromClobberingStore(
1914 L->getType(), L->getPointerOperand(), DepSI);
1915 if (Offset != -1)
1916 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1917 L->getType(), L, DL);
1918 }
1919
1920 // Check to see if we have something like this:
1921 // load i32* P
1922 // load i8* (P+1)
1923 // if we have this, replace the later with an extraction from the former.
1924 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1925 // If this is a clobber and L is the first instruction in its block, then
1926 // we have the first instruction in the entry block.
1927 if (DepLI == L)
1928 return false;
1929
1930 int Offset = AnalyzeLoadFromClobberingLoad(
1931 L->getType(), L->getPointerOperand(), DepLI, DL);
1932 if (Offset != -1)
1933 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1934 }
1935
1936 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1937 // a value on from it.
1938 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1939 int Offset = AnalyzeLoadFromClobberingMemInst(
1940 L->getType(), L->getPointerOperand(), DepMI, DL);
1941 if (Offset != -1)
1942 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, DL);
1943 }
1944
1945 if (AvailVal) {
1946 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1947 << *AvailVal << '\n' << *L << "\n\n\n");
1948
1949 // Replace the load!
1950 L->replaceAllUsesWith(AvailVal);
1951 if (AvailVal->getType()->getScalarType()->isPointerTy())
1952 MD->invalidateCachedPointerInfo(AvailVal);
1953 markInstructionForDeletion(L);
1954 ++NumGVNLoad;
1955 return true;
1956 }
1957
1958 // If the value isn't available, don't do anything!
1959 DEBUG(
1960 // fast print dep, using operator<< on instruction is too slow.
1961 dbgs() << "GVN: load ";
1962 L->printAsOperand(dbgs());
1963 Instruction *I = Dep.getInst();
1964 dbgs() << " is clobbered by " << *I << '\n';
1965 );
1966 return false;
1967 }
1968
1969 // If it is defined in another block, try harder.
1970 if (Dep.isNonLocal())
1971 return processNonLocalLoad(L);
1972
1973 if (!Dep.isDef()) {
1974 DEBUG(
1975 // fast print dep, using operator<< on instruction is too slow.
1976 dbgs() << "GVN: load ";
1977 L->printAsOperand(dbgs());
1978 dbgs() << " has unknown dependence\n";
1979 );
1980 return false;
1981 }
1982
1983 Instruction *DepInst = Dep.getInst();
1984 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1985 Value *StoredVal = DepSI->getValueOperand();
1986
1987 // The store and load are to a must-aliased pointer, but they may not
1988 // actually have the same type. See if we know how to reuse the stored
1989 // value (depending on its type).
1990 if (StoredVal->getType() != L->getType()) {
1991 IRBuilder<> Builder(L);
1992 StoredVal =
1993 CoerceAvailableValueToLoadType(StoredVal, L->getType(), Builder, DL);
1994 if (!StoredVal)
1995 return false;
1996
1997 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1998 << '\n' << *L << "\n\n\n");
1999 }
2000
2001 // Remove it!
2002 L->replaceAllUsesWith(StoredVal);
2003 if (StoredVal->getType()->getScalarType()->isPointerTy())
2004 MD->invalidateCachedPointerInfo(StoredVal);
2005 markInstructionForDeletion(L);
2006 ++NumGVNLoad;
2007 return true;
2008 }
2009
2010 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
2011 Value *AvailableVal = DepLI;
2012
2013 // The loads are of a must-aliased pointer, but they may not actually have
2014 // the same type. See if we know how to reuse the previously loaded value
2015 // (depending on its type).
2016 if (DepLI->getType() != L->getType()) {
2017 IRBuilder<> Builder(L);
2018 AvailableVal =
2019 CoerceAvailableValueToLoadType(DepLI, L->getType(), Builder, DL);
2020 if (!AvailableVal)
2021 return false;
2022
2023 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
2024 << "\n" << *L << "\n\n\n");
2025 }
2026
2027 // Remove it!
2028 patchAndReplaceAllUsesWith(L, AvailableVal);
2029 if (DepLI->getType()->getScalarType()->isPointerTy())
2030 MD->invalidateCachedPointerInfo(DepLI);
2031 markInstructionForDeletion(L);
2032 ++NumGVNLoad;
2033 return true;
2034 }
2035
2036 // If this load really doesn't depend on anything, then we must be loading an
2037 // undef value. This can happen when loading for a fresh allocation with no
2038 // intervening stores, for example.
2039 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
2040 L->replaceAllUsesWith(UndefValue::get(L->getType()));
2041 markInstructionForDeletion(L);
2042 ++NumGVNLoad;
2043 return true;
2044 }
2045
2046 // If this load occurs either right after a lifetime begin,
2047 // then the loaded value is undefined.
2048 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
2049 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
2050 L->replaceAllUsesWith(UndefValue::get(L->getType()));
2051 markInstructionForDeletion(L);
2052 ++NumGVNLoad;
2053 return true;
2054 }
2055 }
2056
2057 // If this load follows a calloc (which zero initializes memory),
2058 // then the loaded value is zero
2059 if (isCallocLikeFn(DepInst, TLI)) {
2060 L->replaceAllUsesWith(Constant::getNullValue(L->getType()));
2061 markInstructionForDeletion(L);
2062 ++NumGVNLoad;
2063 return true;
2064 }
2065
2066 return false;
2067 }
2068
2069 // In order to find a leader for a given value number at a
2070 // specific basic block, we first obtain the list of all Values for that number,
2071 // and then scan the list to find one whose block dominates the block in
2072 // question. This is fast because dominator tree queries consist of only
2073 // a few comparisons of DFS numbers.
findLeader(const BasicBlock * BB,uint32_t num)2074 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2075 LeaderTableEntry Vals = LeaderTable[num];
2076 if (!Vals.Val) return nullptr;
2077
2078 Value *Val = nullptr;
2079 if (DT->dominates(Vals.BB, BB)) {
2080 Val = Vals.Val;
2081 if (isa<Constant>(Val)) return Val;
2082 }
2083
2084 LeaderTableEntry* Next = Vals.Next;
2085 while (Next) {
2086 if (DT->dominates(Next->BB, BB)) {
2087 if (isa<Constant>(Next->Val)) return Next->Val;
2088 if (!Val) Val = Next->Val;
2089 }
2090
2091 Next = Next->Next;
2092 }
2093
2094 return Val;
2095 }
2096
2097 /// There is an edge from 'Src' to 'Dst'. Return
2098 /// true if every path from the entry block to 'Dst' passes via this edge. In
2099 /// particular 'Dst' must not be reachable via another edge from 'Src'.
isOnlyReachableViaThisEdge(const BasicBlockEdge & E,DominatorTree * DT)2100 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2101 DominatorTree *DT) {
2102 // While in theory it is interesting to consider the case in which Dst has
2103 // more than one predecessor, because Dst might be part of a loop which is
2104 // only reachable from Src, in practice it is pointless since at the time
2105 // GVN runs all such loops have preheaders, which means that Dst will have
2106 // been changed to have only one predecessor, namely Src.
2107 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2108 const BasicBlock *Src = E.getStart();
2109 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2110 (void)Src;
2111 return Pred != nullptr;
2112 }
2113
2114 // Tries to replace instruction with const, using information from
2115 // ReplaceWithConstMap.
replaceOperandsWithConsts(Instruction * Instr) const2116 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
2117 bool Changed = false;
2118 for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
2119 Value *Operand = Instr->getOperand(OpNum);
2120 auto it = ReplaceWithConstMap.find(Operand);
2121 if (it != ReplaceWithConstMap.end()) {
2122 assert(!isa<Constant>(Operand) &&
2123 "Replacing constants with constants is invalid");
2124 DEBUG(dbgs() << "GVN replacing: " << *Operand << " with " << *it->second
2125 << " in instruction " << *Instr << '\n');
2126 Instr->setOperand(OpNum, it->second);
2127 Changed = true;
2128 }
2129 }
2130 return Changed;
2131 }
2132
2133 /// The given values are known to be equal in every block
2134 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2135 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2136 /// If DominatesByEdge is false, then it means that it is dominated by Root.End.
propagateEquality(Value * LHS,Value * RHS,const BasicBlockEdge & Root,bool DominatesByEdge)2137 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
2138 bool DominatesByEdge) {
2139 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2140 Worklist.push_back(std::make_pair(LHS, RHS));
2141 bool Changed = false;
2142 // For speed, compute a conservative fast approximation to
2143 // DT->dominates(Root, Root.getEnd());
2144 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2145
2146 while (!Worklist.empty()) {
2147 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2148 LHS = Item.first; RHS = Item.second;
2149
2150 if (LHS == RHS)
2151 continue;
2152 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2153
2154 // Don't try to propagate equalities between constants.
2155 if (isa<Constant>(LHS) && isa<Constant>(RHS))
2156 continue;
2157
2158 // Prefer a constant on the right-hand side, or an Argument if no constants.
2159 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2160 std::swap(LHS, RHS);
2161 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2162
2163 // If there is no obvious reason to prefer the left-hand side over the
2164 // right-hand side, ensure the longest lived term is on the right-hand side,
2165 // so the shortest lived term will be replaced by the longest lived.
2166 // This tends to expose more simplifications.
2167 uint32_t LVN = VN.lookup_or_add(LHS);
2168 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2169 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2170 // Move the 'oldest' value to the right-hand side, using the value number
2171 // as a proxy for age.
2172 uint32_t RVN = VN.lookup_or_add(RHS);
2173 if (LVN < RVN) {
2174 std::swap(LHS, RHS);
2175 LVN = RVN;
2176 }
2177 }
2178
2179 // If value numbering later sees that an instruction in the scope is equal
2180 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2181 // the invariant that instructions only occur in the leader table for their
2182 // own value number (this is used by removeFromLeaderTable), do not do this
2183 // if RHS is an instruction (if an instruction in the scope is morphed into
2184 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2185 // using the leader table is about compiling faster, not optimizing better).
2186 // The leader table only tracks basic blocks, not edges. Only add to if we
2187 // have the simple case where the edge dominates the end.
2188 if (RootDominatesEnd && !isa<Instruction>(RHS))
2189 addToLeaderTable(LVN, RHS, Root.getEnd());
2190
2191 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2192 // LHS always has at least one use that is not dominated by Root, this will
2193 // never do anything if LHS has only one use.
2194 if (!LHS->hasOneUse()) {
2195 unsigned NumReplacements =
2196 DominatesByEdge
2197 ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
2198 : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getEnd());
2199
2200 Changed |= NumReplacements > 0;
2201 NumGVNEqProp += NumReplacements;
2202 }
2203
2204 // Now try to deduce additional equalities from this one. For example, if
2205 // the known equality was "(A != B)" == "false" then it follows that A and B
2206 // are equal in the scope. Only boolean equalities with an explicit true or
2207 // false RHS are currently supported.
2208 if (!RHS->getType()->isIntegerTy(1))
2209 // Not a boolean equality - bail out.
2210 continue;
2211 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2212 if (!CI)
2213 // RHS neither 'true' nor 'false' - bail out.
2214 continue;
2215 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2216 bool isKnownTrue = CI->isAllOnesValue();
2217 bool isKnownFalse = !isKnownTrue;
2218
2219 // If "A && B" is known true then both A and B are known true. If "A || B"
2220 // is known false then both A and B are known false.
2221 Value *A, *B;
2222 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2223 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2224 Worklist.push_back(std::make_pair(A, RHS));
2225 Worklist.push_back(std::make_pair(B, RHS));
2226 continue;
2227 }
2228
2229 // If we are propagating an equality like "(A == B)" == "true" then also
2230 // propagate the equality A == B. When propagating a comparison such as
2231 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2232 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
2233 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2234
2235 // If "A == B" is known true, or "A != B" is known false, then replace
2236 // A with B everywhere in the scope.
2237 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2238 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2239 Worklist.push_back(std::make_pair(Op0, Op1));
2240
2241 // Handle the floating point versions of equality comparisons too.
2242 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
2243 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
2244
2245 // Floating point -0.0 and 0.0 compare equal, so we can only
2246 // propagate values if we know that we have a constant and that
2247 // its value is non-zero.
2248
2249 // FIXME: We should do this optimization if 'no signed zeros' is
2250 // applicable via an instruction-level fast-math-flag or some other
2251 // indicator that relaxed FP semantics are being used.
2252
2253 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
2254 Worklist.push_back(std::make_pair(Op0, Op1));
2255 }
2256
2257 // If "A >= B" is known true, replace "A < B" with false everywhere.
2258 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2259 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2260 // Since we don't have the instruction "A < B" immediately to hand, work
2261 // out the value number that it would have and use that to find an
2262 // appropriate instruction (if any).
2263 uint32_t NextNum = VN.getNextUnusedValueNumber();
2264 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2265 // If the number we were assigned was brand new then there is no point in
2266 // looking for an instruction realizing it: there cannot be one!
2267 if (Num < NextNum) {
2268 Value *NotCmp = findLeader(Root.getEnd(), Num);
2269 if (NotCmp && isa<Instruction>(NotCmp)) {
2270 unsigned NumReplacements =
2271 DominatesByEdge
2272 ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
2273 : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
2274 Root.getEnd());
2275 Changed |= NumReplacements > 0;
2276 NumGVNEqProp += NumReplacements;
2277 }
2278 }
2279 // Ensure that any instruction in scope that gets the "A < B" value number
2280 // is replaced with false.
2281 // The leader table only tracks basic blocks, not edges. Only add to if we
2282 // have the simple case where the edge dominates the end.
2283 if (RootDominatesEnd)
2284 addToLeaderTable(Num, NotVal, Root.getEnd());
2285
2286 continue;
2287 }
2288 }
2289
2290 return Changed;
2291 }
2292
2293 /// When calculating availability, handle an instruction
2294 /// by inserting it into the appropriate sets
processInstruction(Instruction * I)2295 bool GVN::processInstruction(Instruction *I) {
2296 // Ignore dbg info intrinsics.
2297 if (isa<DbgInfoIntrinsic>(I))
2298 return false;
2299
2300 // If the instruction can be easily simplified then do so now in preference
2301 // to value numbering it. Value numbering often exposes redundancies, for
2302 // example if it determines that %y is equal to %x then the instruction
2303 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2304 const DataLayout &DL = I->getModule()->getDataLayout();
2305 if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
2306 I->replaceAllUsesWith(V);
2307 if (MD && V->getType()->getScalarType()->isPointerTy())
2308 MD->invalidateCachedPointerInfo(V);
2309 markInstructionForDeletion(I);
2310 ++NumGVNSimpl;
2311 return true;
2312 }
2313
2314 if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
2315 if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
2316 return processAssumeIntrinsic(IntrinsicI);
2317
2318 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2319 if (processLoad(LI))
2320 return true;
2321
2322 unsigned Num = VN.lookup_or_add(LI);
2323 addToLeaderTable(Num, LI, LI->getParent());
2324 return false;
2325 }
2326
2327 // For conditional branches, we can perform simple conditional propagation on
2328 // the condition value itself.
2329 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2330 if (!BI->isConditional())
2331 return false;
2332
2333 if (isa<Constant>(BI->getCondition()))
2334 return processFoldableCondBr(BI);
2335
2336 Value *BranchCond = BI->getCondition();
2337 BasicBlock *TrueSucc = BI->getSuccessor(0);
2338 BasicBlock *FalseSucc = BI->getSuccessor(1);
2339 // Avoid multiple edges early.
2340 if (TrueSucc == FalseSucc)
2341 return false;
2342
2343 BasicBlock *Parent = BI->getParent();
2344 bool Changed = false;
2345
2346 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2347 BasicBlockEdge TrueE(Parent, TrueSucc);
2348 Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
2349
2350 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2351 BasicBlockEdge FalseE(Parent, FalseSucc);
2352 Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
2353
2354 return Changed;
2355 }
2356
2357 // For switches, propagate the case values into the case destinations.
2358 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2359 Value *SwitchCond = SI->getCondition();
2360 BasicBlock *Parent = SI->getParent();
2361 bool Changed = false;
2362
2363 // Remember how many outgoing edges there are to every successor.
2364 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2365 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2366 ++SwitchEdges[SI->getSuccessor(i)];
2367
2368 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2369 i != e; ++i) {
2370 BasicBlock *Dst = i.getCaseSuccessor();
2371 // If there is only a single edge, propagate the case value into it.
2372 if (SwitchEdges.lookup(Dst) == 1) {
2373 BasicBlockEdge E(Parent, Dst);
2374 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E, true);
2375 }
2376 }
2377 return Changed;
2378 }
2379
2380 // Instructions with void type don't return a value, so there's
2381 // no point in trying to find redundancies in them.
2382 if (I->getType()->isVoidTy())
2383 return false;
2384
2385 uint32_t NextNum = VN.getNextUnusedValueNumber();
2386 unsigned Num = VN.lookup_or_add(I);
2387
2388 // Allocations are always uniquely numbered, so we can save time and memory
2389 // by fast failing them.
2390 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2391 addToLeaderTable(Num, I, I->getParent());
2392 return false;
2393 }
2394
2395 // If the number we were assigned was a brand new VN, then we don't
2396 // need to do a lookup to see if the number already exists
2397 // somewhere in the domtree: it can't!
2398 if (Num >= NextNum) {
2399 addToLeaderTable(Num, I, I->getParent());
2400 return false;
2401 }
2402
2403 // Perform fast-path value-number based elimination of values inherited from
2404 // dominators.
2405 Value *Repl = findLeader(I->getParent(), Num);
2406 if (!Repl) {
2407 // Failure, just remember this instance for future use.
2408 addToLeaderTable(Num, I, I->getParent());
2409 return false;
2410 } else if (Repl == I) {
2411 // If I was the result of a shortcut PRE, it might already be in the table
2412 // and the best replacement for itself. Nothing to do.
2413 return false;
2414 }
2415
2416 // Remove it!
2417 patchAndReplaceAllUsesWith(I, Repl);
2418 if (MD && Repl->getType()->getScalarType()->isPointerTy())
2419 MD->invalidateCachedPointerInfo(Repl);
2420 markInstructionForDeletion(I);
2421 return true;
2422 }
2423
2424 /// runOnFunction - This is the main transformation entry point for a function.
runOnFunction(Function & F)2425 bool GVN::runOnFunction(Function& F) {
2426 if (skipOptnoneFunction(F))
2427 return false;
2428
2429 if (!NoLoads)
2430 MD = &getAnalysis<MemoryDependenceAnalysis>();
2431 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2432 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2433 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
2434 VN.setAliasAnalysis(&getAnalysis<AAResultsWrapperPass>().getAAResults());
2435 VN.setMemDep(MD);
2436 VN.setDomTree(DT);
2437
2438 bool Changed = false;
2439 bool ShouldContinue = true;
2440
2441 // Merge unconditional branches, allowing PRE to catch more
2442 // optimization opportunities.
2443 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2444 BasicBlock *BB = &*FI++;
2445
2446 bool removedBlock =
2447 MergeBlockIntoPredecessor(BB, DT, /* LoopInfo */ nullptr, MD);
2448 if (removedBlock) ++NumGVNBlocks;
2449
2450 Changed |= removedBlock;
2451 }
2452
2453 unsigned Iteration = 0;
2454 while (ShouldContinue) {
2455 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2456 ShouldContinue = iterateOnFunction(F);
2457 Changed |= ShouldContinue;
2458 ++Iteration;
2459 }
2460
2461 if (EnablePRE) {
2462 // Fabricate val-num for dead-code in order to suppress assertion in
2463 // performPRE().
2464 assignValNumForDeadCode();
2465 bool PREChanged = true;
2466 while (PREChanged) {
2467 PREChanged = performPRE(F);
2468 Changed |= PREChanged;
2469 }
2470 }
2471
2472 // FIXME: Should perform GVN again after PRE does something. PRE can move
2473 // computations into blocks where they become fully redundant. Note that
2474 // we can't do this until PRE's critical edge splitting updates memdep.
2475 // Actually, when this happens, we should just fully integrate PRE into GVN.
2476
2477 cleanupGlobalSets();
2478 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2479 // iteration.
2480 DeadBlocks.clear();
2481
2482 return Changed;
2483 }
2484
processBlock(BasicBlock * BB)2485 bool GVN::processBlock(BasicBlock *BB) {
2486 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2487 // (and incrementing BI before processing an instruction).
2488 assert(InstrsToErase.empty() &&
2489 "We expect InstrsToErase to be empty across iterations");
2490 if (DeadBlocks.count(BB))
2491 return false;
2492
2493 // Clearing map before every BB because it can be used only for single BB.
2494 ReplaceWithConstMap.clear();
2495 bool ChangedFunction = false;
2496
2497 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2498 BI != BE;) {
2499 if (!ReplaceWithConstMap.empty())
2500 ChangedFunction |= replaceOperandsWithConsts(&*BI);
2501 ChangedFunction |= processInstruction(&*BI);
2502
2503 if (InstrsToErase.empty()) {
2504 ++BI;
2505 continue;
2506 }
2507
2508 // If we need some instructions deleted, do it now.
2509 NumGVNInstr += InstrsToErase.size();
2510
2511 // Avoid iterator invalidation.
2512 bool AtStart = BI == BB->begin();
2513 if (!AtStart)
2514 --BI;
2515
2516 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2517 E = InstrsToErase.end(); I != E; ++I) {
2518 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2519 if (MD) MD->removeInstruction(*I);
2520 DEBUG(verifyRemoved(*I));
2521 (*I)->eraseFromParent();
2522 }
2523 InstrsToErase.clear();
2524
2525 if (AtStart)
2526 BI = BB->begin();
2527 else
2528 ++BI;
2529 }
2530
2531 return ChangedFunction;
2532 }
2533
2534 // Instantiate an expression in a predecessor that lacked it.
performScalarPREInsertion(Instruction * Instr,BasicBlock * Pred,unsigned int ValNo)2535 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2536 unsigned int ValNo) {
2537 // Because we are going top-down through the block, all value numbers
2538 // will be available in the predecessor by the time we need them. Any
2539 // that weren't originally present will have been instantiated earlier
2540 // in this loop.
2541 bool success = true;
2542 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2543 Value *Op = Instr->getOperand(i);
2544 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2545 continue;
2546 // This could be a newly inserted instruction, in which case, we won't
2547 // find a value number, and should give up before we hurt ourselves.
2548 // FIXME: Rewrite the infrastructure to let it easier to value number
2549 // and process newly inserted instructions.
2550 if (!VN.exists(Op)) {
2551 success = false;
2552 break;
2553 }
2554 if (Value *V = findLeader(Pred, VN.lookup(Op))) {
2555 Instr->setOperand(i, V);
2556 } else {
2557 success = false;
2558 break;
2559 }
2560 }
2561
2562 // Fail out if we encounter an operand that is not available in
2563 // the PRE predecessor. This is typically because of loads which
2564 // are not value numbered precisely.
2565 if (!success)
2566 return false;
2567
2568 Instr->insertBefore(Pred->getTerminator());
2569 Instr->setName(Instr->getName() + ".pre");
2570 Instr->setDebugLoc(Instr->getDebugLoc());
2571 VN.add(Instr, ValNo);
2572
2573 // Update the availability map to include the new instruction.
2574 addToLeaderTable(ValNo, Instr, Pred);
2575 return true;
2576 }
2577
performScalarPRE(Instruction * CurInst)2578 bool GVN::performScalarPRE(Instruction *CurInst) {
2579 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2580
2581 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2582 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2583 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2584 isa<DbgInfoIntrinsic>(CurInst))
2585 return false;
2586
2587 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2588 // sinking the compare again, and it would force the code generator to
2589 // move the i1 from processor flags or predicate registers into a general
2590 // purpose register.
2591 if (isa<CmpInst>(CurInst))
2592 return false;
2593
2594 // We don't currently value number ANY inline asm calls.
2595 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2596 if (CallI->isInlineAsm())
2597 return false;
2598
2599 uint32_t ValNo = VN.lookup(CurInst);
2600
2601 // Look for the predecessors for PRE opportunities. We're
2602 // only trying to solve the basic diamond case, where
2603 // a value is computed in the successor and one predecessor,
2604 // but not the other. We also explicitly disallow cases
2605 // where the successor is its own predecessor, because they're
2606 // more complicated to get right.
2607 unsigned NumWith = 0;
2608 unsigned NumWithout = 0;
2609 BasicBlock *PREPred = nullptr;
2610 BasicBlock *CurrentBlock = CurInst->getParent();
2611 predMap.clear();
2612
2613 for (BasicBlock *P : predecessors(CurrentBlock)) {
2614 // We're not interested in PRE where the block is its
2615 // own predecessor, or in blocks with predecessors
2616 // that are not reachable.
2617 if (P == CurrentBlock) {
2618 NumWithout = 2;
2619 break;
2620 } else if (!DT->isReachableFromEntry(P)) {
2621 NumWithout = 2;
2622 break;
2623 }
2624
2625 Value *predV = findLeader(P, ValNo);
2626 if (!predV) {
2627 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2628 PREPred = P;
2629 ++NumWithout;
2630 } else if (predV == CurInst) {
2631 /* CurInst dominates this predecessor. */
2632 NumWithout = 2;
2633 break;
2634 } else {
2635 predMap.push_back(std::make_pair(predV, P));
2636 ++NumWith;
2637 }
2638 }
2639
2640 // Don't do PRE when it might increase code size, i.e. when
2641 // we would need to insert instructions in more than one pred.
2642 if (NumWithout > 1 || NumWith == 0)
2643 return false;
2644
2645 // We may have a case where all predecessors have the instruction,
2646 // and we just need to insert a phi node. Otherwise, perform
2647 // insertion.
2648 Instruction *PREInstr = nullptr;
2649
2650 if (NumWithout != 0) {
2651 // Don't do PRE across indirect branch.
2652 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2653 return false;
2654
2655 // We can't do PRE safely on a critical edge, so instead we schedule
2656 // the edge to be split and perform the PRE the next time we iterate
2657 // on the function.
2658 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2659 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2660 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2661 return false;
2662 }
2663 // We need to insert somewhere, so let's give it a shot
2664 PREInstr = CurInst->clone();
2665 if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2666 // If we failed insertion, make sure we remove the instruction.
2667 DEBUG(verifyRemoved(PREInstr));
2668 delete PREInstr;
2669 return false;
2670 }
2671 }
2672
2673 // Either we should have filled in the PRE instruction, or we should
2674 // not have needed insertions.
2675 assert (PREInstr != nullptr || NumWithout == 0);
2676
2677 ++NumGVNPRE;
2678
2679 // Create a PHI to make the value available in this block.
2680 PHINode *Phi =
2681 PHINode::Create(CurInst->getType(), predMap.size(),
2682 CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2683 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2684 if (Value *V = predMap[i].first)
2685 Phi->addIncoming(V, predMap[i].second);
2686 else
2687 Phi->addIncoming(PREInstr, PREPred);
2688 }
2689
2690 VN.add(Phi, ValNo);
2691 addToLeaderTable(ValNo, Phi, CurrentBlock);
2692 Phi->setDebugLoc(CurInst->getDebugLoc());
2693 CurInst->replaceAllUsesWith(Phi);
2694 if (MD && Phi->getType()->getScalarType()->isPointerTy())
2695 MD->invalidateCachedPointerInfo(Phi);
2696 VN.erase(CurInst);
2697 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2698
2699 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2700 if (MD)
2701 MD->removeInstruction(CurInst);
2702 DEBUG(verifyRemoved(CurInst));
2703 CurInst->eraseFromParent();
2704 ++NumGVNInstr;
2705
2706 return true;
2707 }
2708
2709 /// Perform a purely local form of PRE that looks for diamond
2710 /// control flow patterns and attempts to perform simple PRE at the join point.
performPRE(Function & F)2711 bool GVN::performPRE(Function &F) {
2712 bool Changed = false;
2713 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2714 // Nothing to PRE in the entry block.
2715 if (CurrentBlock == &F.getEntryBlock())
2716 continue;
2717
2718 // Don't perform PRE on an EH pad.
2719 if (CurrentBlock->isEHPad())
2720 continue;
2721
2722 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2723 BE = CurrentBlock->end();
2724 BI != BE;) {
2725 Instruction *CurInst = &*BI++;
2726 Changed |= performScalarPRE(CurInst);
2727 }
2728 }
2729
2730 if (splitCriticalEdges())
2731 Changed = true;
2732
2733 return Changed;
2734 }
2735
2736 /// Split the critical edge connecting the given two blocks, and return
2737 /// the block inserted to the critical edge.
splitCriticalEdges(BasicBlock * Pred,BasicBlock * Succ)2738 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2739 BasicBlock *BB =
2740 SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT));
2741 if (MD)
2742 MD->invalidateCachedPredecessors();
2743 return BB;
2744 }
2745
2746 /// Split critical edges found during the previous
2747 /// iteration that may enable further optimization.
splitCriticalEdges()2748 bool GVN::splitCriticalEdges() {
2749 if (toSplit.empty())
2750 return false;
2751 do {
2752 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2753 SplitCriticalEdge(Edge.first, Edge.second,
2754 CriticalEdgeSplittingOptions(DT));
2755 } while (!toSplit.empty());
2756 if (MD) MD->invalidateCachedPredecessors();
2757 return true;
2758 }
2759
2760 /// Executes one iteration of GVN
iterateOnFunction(Function & F)2761 bool GVN::iterateOnFunction(Function &F) {
2762 cleanupGlobalSets();
2763
2764 // Top-down walk of the dominator tree
2765 bool Changed = false;
2766 // Save the blocks this function have before transformation begins. GVN may
2767 // split critical edge, and hence may invalidate the RPO/DT iterator.
2768 //
2769 std::vector<BasicBlock *> BBVect;
2770 BBVect.reserve(256);
2771 // Needed for value numbering with phi construction to work.
2772 ReversePostOrderTraversal<Function *> RPOT(&F);
2773 for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
2774 RE = RPOT.end();
2775 RI != RE; ++RI)
2776 BBVect.push_back(*RI);
2777
2778 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2779 I != E; I++)
2780 Changed |= processBlock(*I);
2781
2782 return Changed;
2783 }
2784
cleanupGlobalSets()2785 void GVN::cleanupGlobalSets() {
2786 VN.clear();
2787 LeaderTable.clear();
2788 TableAllocator.Reset();
2789 }
2790
2791 /// Verify that the specified instruction does not occur in our
2792 /// internal data structures.
verifyRemoved(const Instruction * Inst) const2793 void GVN::verifyRemoved(const Instruction *Inst) const {
2794 VN.verifyRemoved(Inst);
2795
2796 // Walk through the value number scope to make sure the instruction isn't
2797 // ferreted away in it.
2798 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2799 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2800 const LeaderTableEntry *Node = &I->second;
2801 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2802
2803 while (Node->Next) {
2804 Node = Node->Next;
2805 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2806 }
2807 }
2808 }
2809
2810 /// BB is declared dead, which implied other blocks become dead as well. This
2811 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2812 /// live successors, update their phi nodes by replacing the operands
2813 /// corresponding to dead blocks with UndefVal.
addDeadBlock(BasicBlock * BB)2814 void GVN::addDeadBlock(BasicBlock *BB) {
2815 SmallVector<BasicBlock *, 4> NewDead;
2816 SmallSetVector<BasicBlock *, 4> DF;
2817
2818 NewDead.push_back(BB);
2819 while (!NewDead.empty()) {
2820 BasicBlock *D = NewDead.pop_back_val();
2821 if (DeadBlocks.count(D))
2822 continue;
2823
2824 // All blocks dominated by D are dead.
2825 SmallVector<BasicBlock *, 8> Dom;
2826 DT->getDescendants(D, Dom);
2827 DeadBlocks.insert(Dom.begin(), Dom.end());
2828
2829 // Figure out the dominance-frontier(D).
2830 for (BasicBlock *B : Dom) {
2831 for (BasicBlock *S : successors(B)) {
2832 if (DeadBlocks.count(S))
2833 continue;
2834
2835 bool AllPredDead = true;
2836 for (BasicBlock *P : predecessors(S))
2837 if (!DeadBlocks.count(P)) {
2838 AllPredDead = false;
2839 break;
2840 }
2841
2842 if (!AllPredDead) {
2843 // S could be proved dead later on. That is why we don't update phi
2844 // operands at this moment.
2845 DF.insert(S);
2846 } else {
2847 // While S is not dominated by D, it is dead by now. This could take
2848 // place if S already have a dead predecessor before D is declared
2849 // dead.
2850 NewDead.push_back(S);
2851 }
2852 }
2853 }
2854 }
2855
2856 // For the dead blocks' live successors, update their phi nodes by replacing
2857 // the operands corresponding to dead blocks with UndefVal.
2858 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2859 I != E; I++) {
2860 BasicBlock *B = *I;
2861 if (DeadBlocks.count(B))
2862 continue;
2863
2864 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2865 for (BasicBlock *P : Preds) {
2866 if (!DeadBlocks.count(P))
2867 continue;
2868
2869 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2870 if (BasicBlock *S = splitCriticalEdges(P, B))
2871 DeadBlocks.insert(P = S);
2872 }
2873
2874 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2875 PHINode &Phi = cast<PHINode>(*II);
2876 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2877 UndefValue::get(Phi.getType()));
2878 }
2879 }
2880 }
2881 }
2882
2883 // If the given branch is recognized as a foldable branch (i.e. conditional
2884 // branch with constant condition), it will perform following analyses and
2885 // transformation.
2886 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2887 // R be the target of the dead out-coming edge.
2888 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2889 // edge. The result of this step will be {X| X is dominated by R}
2890 // 2) Identify those blocks which haves at least one dead predecessor. The
2891 // result of this step will be dominance-frontier(R).
2892 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2893 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2894 //
2895 // Return true iff *NEW* dead code are found.
processFoldableCondBr(BranchInst * BI)2896 bool GVN::processFoldableCondBr(BranchInst *BI) {
2897 if (!BI || BI->isUnconditional())
2898 return false;
2899
2900 // If a branch has two identical successors, we cannot declare either dead.
2901 if (BI->getSuccessor(0) == BI->getSuccessor(1))
2902 return false;
2903
2904 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2905 if (!Cond)
2906 return false;
2907
2908 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2909 BI->getSuccessor(1) : BI->getSuccessor(0);
2910 if (DeadBlocks.count(DeadRoot))
2911 return false;
2912
2913 if (!DeadRoot->getSinglePredecessor())
2914 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2915
2916 addDeadBlock(DeadRoot);
2917 return true;
2918 }
2919
2920 // performPRE() will trigger assert if it comes across an instruction without
2921 // associated val-num. As it normally has far more live instructions than dead
2922 // instructions, it makes more sense just to "fabricate" a val-number for the
2923 // dead code than checking if instruction involved is dead or not.
assignValNumForDeadCode()2924 void GVN::assignValNumForDeadCode() {
2925 for (BasicBlock *BB : DeadBlocks) {
2926 for (Instruction &Inst : *BB) {
2927 unsigned ValNum = VN.lookup_or_add(&Inst);
2928 addToLeaderTable(ValNum, &Inst, BB);
2929 }
2930 }
2931 }
2932