1 //===- BasicAliasAnalysis.cpp - Stateless Alias Analysis Impl -------------===//
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 file defines the primary stateless implementation of the
11 // Alias Analysis interface that implements identities (two different
12 // globals cannot alias, etc), but does no stateful analysis.
13 //
14 //===----------------------------------------------------------------------===//
15
16 #include "llvm/Analysis/BasicAliasAnalysis.h"
17 #include "llvm/ADT/APInt.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/AssumptionCache.h"
23 #include "llvm/Analysis/CFG.h"
24 #include "llvm/Analysis/CaptureTracking.h"
25 #include "llvm/Analysis/InstructionSimplify.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/MemoryBuiltins.h"
28 #include "llvm/Analysis/MemoryLocation.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/Analysis/PhiValues.h"
32 #include "llvm/IR/Argument.h"
33 #include "llvm/IR/Attributes.h"
34 #include "llvm/IR/Constant.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/DerivedTypes.h"
38 #include "llvm/IR/Dominators.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GetElementPtrTypeIterator.h"
41 #include "llvm/IR/GlobalAlias.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/InstrTypes.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Intrinsics.h"
48 #include "llvm/IR/Metadata.h"
49 #include "llvm/IR/Operator.h"
50 #include "llvm/IR/Type.h"
51 #include "llvm/IR/User.h"
52 #include "llvm/IR/Value.h"
53 #include "llvm/Pass.h"
54 #include "llvm/Support/Casting.h"
55 #include "llvm/Support/CommandLine.h"
56 #include "llvm/Support/Compiler.h"
57 #include "llvm/Support/KnownBits.h"
58 #include <cassert>
59 #include <cstdint>
60 #include <cstdlib>
61 #include <utility>
62
63 #define DEBUG_TYPE "basicaa"
64
65 using namespace llvm;
66
67 /// Enable analysis of recursive PHI nodes.
68 static cl::opt<bool> EnableRecPhiAnalysis("basicaa-recphi", cl::Hidden,
69 cl::init(false));
70
71 /// By default, even on 32-bit architectures we use 64-bit integers for
72 /// calculations. This will allow us to more-aggressively decompose indexing
73 /// expressions calculated using i64 values (e.g., long long in C) which is
74 /// common enough to worry about.
75 static cl::opt<bool> ForceAtLeast64Bits("basicaa-force-at-least-64b",
76 cl::Hidden, cl::init(true));
77 static cl::opt<bool> DoubleCalcBits("basicaa-double-calc-bits",
78 cl::Hidden, cl::init(false));
79
80 /// SearchLimitReached / SearchTimes shows how often the limit of
81 /// to decompose GEPs is reached. It will affect the precision
82 /// of basic alias analysis.
83 STATISTIC(SearchLimitReached, "Number of times the limit to "
84 "decompose GEPs is reached");
85 STATISTIC(SearchTimes, "Number of times a GEP is decomposed");
86
87 /// Cutoff after which to stop analysing a set of phi nodes potentially involved
88 /// in a cycle. Because we are analysing 'through' phi nodes, we need to be
89 /// careful with value equivalence. We use reachability to make sure a value
90 /// cannot be involved in a cycle.
91 const unsigned MaxNumPhiBBsValueReachabilityCheck = 20;
92
93 // The max limit of the search depth in DecomposeGEPExpression() and
94 // GetUnderlyingObject(), both functions need to use the same search
95 // depth otherwise the algorithm in aliasGEP will assert.
96 static const unsigned MaxLookupSearchDepth = 6;
97
invalidate(Function & Fn,const PreservedAnalyses & PA,FunctionAnalysisManager::Invalidator & Inv)98 bool BasicAAResult::invalidate(Function &Fn, const PreservedAnalyses &PA,
99 FunctionAnalysisManager::Invalidator &Inv) {
100 // We don't care if this analysis itself is preserved, it has no state. But
101 // we need to check that the analyses it depends on have been. Note that we
102 // may be created without handles to some analyses and in that case don't
103 // depend on them.
104 if (Inv.invalidate<AssumptionAnalysis>(Fn, PA) ||
105 (DT && Inv.invalidate<DominatorTreeAnalysis>(Fn, PA)) ||
106 (LI && Inv.invalidate<LoopAnalysis>(Fn, PA)) ||
107 (PV && Inv.invalidate<PhiValuesAnalysis>(Fn, PA)))
108 return true;
109
110 // Otherwise this analysis result remains valid.
111 return false;
112 }
113
114 //===----------------------------------------------------------------------===//
115 // Useful predicates
116 //===----------------------------------------------------------------------===//
117
118 /// Returns true if the pointer is to a function-local object that never
119 /// escapes from the function.
isNonEscapingLocalObject(const Value * V)120 static bool isNonEscapingLocalObject(const Value *V) {
121 // If this is a local allocation, check to see if it escapes.
122 if (isa<AllocaInst>(V) || isNoAliasCall(V))
123 // Set StoreCaptures to True so that we can assume in our callers that the
124 // pointer is not the result of a load instruction. Currently
125 // PointerMayBeCaptured doesn't have any special analysis for the
126 // StoreCaptures=false case; if it did, our callers could be refined to be
127 // more precise.
128 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
129
130 // If this is an argument that corresponds to a byval or noalias argument,
131 // then it has not escaped before entering the function. Check if it escapes
132 // inside the function.
133 if (const Argument *A = dyn_cast<Argument>(V))
134 if (A->hasByValAttr() || A->hasNoAliasAttr())
135 // Note even if the argument is marked nocapture, we still need to check
136 // for copies made inside the function. The nocapture attribute only
137 // specifies that there are no copies made that outlive the function.
138 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
139
140 return false;
141 }
142
143 /// Returns true if the pointer is one which would have been considered an
144 /// escape by isNonEscapingLocalObject.
isEscapeSource(const Value * V)145 static bool isEscapeSource(const Value *V) {
146 if (isa<CallBase>(V))
147 return true;
148
149 if (isa<Argument>(V))
150 return true;
151
152 // The load case works because isNonEscapingLocalObject considers all
153 // stores to be escapes (it passes true for the StoreCaptures argument
154 // to PointerMayBeCaptured).
155 if (isa<LoadInst>(V))
156 return true;
157
158 return false;
159 }
160
161 /// Returns the size of the object specified by V or UnknownSize if unknown.
getObjectSize(const Value * V,const DataLayout & DL,const TargetLibraryInfo & TLI,bool NullIsValidLoc,bool RoundToAlign=false)162 static uint64_t getObjectSize(const Value *V, const DataLayout &DL,
163 const TargetLibraryInfo &TLI,
164 bool NullIsValidLoc,
165 bool RoundToAlign = false) {
166 uint64_t Size;
167 ObjectSizeOpts Opts;
168 Opts.RoundToAlign = RoundToAlign;
169 Opts.NullIsUnknownSize = NullIsValidLoc;
170 if (getObjectSize(V, Size, DL, &TLI, Opts))
171 return Size;
172 return MemoryLocation::UnknownSize;
173 }
174
175 /// Returns true if we can prove that the object specified by V is smaller than
176 /// Size.
isObjectSmallerThan(const Value * V,uint64_t Size,const DataLayout & DL,const TargetLibraryInfo & TLI,bool NullIsValidLoc)177 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
178 const DataLayout &DL,
179 const TargetLibraryInfo &TLI,
180 bool NullIsValidLoc) {
181 // Note that the meanings of the "object" are slightly different in the
182 // following contexts:
183 // c1: llvm::getObjectSize()
184 // c2: llvm.objectsize() intrinsic
185 // c3: isObjectSmallerThan()
186 // c1 and c2 share the same meaning; however, the meaning of "object" in c3
187 // refers to the "entire object".
188 //
189 // Consider this example:
190 // char *p = (char*)malloc(100)
191 // char *q = p+80;
192 //
193 // In the context of c1 and c2, the "object" pointed by q refers to the
194 // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20.
195 //
196 // However, in the context of c3, the "object" refers to the chunk of memory
197 // being allocated. So, the "object" has 100 bytes, and q points to the middle
198 // the "object". In case q is passed to isObjectSmallerThan() as the 1st
199 // parameter, before the llvm::getObjectSize() is called to get the size of
200 // entire object, we should:
201 // - either rewind the pointer q to the base-address of the object in
202 // question (in this case rewind to p), or
203 // - just give up. It is up to caller to make sure the pointer is pointing
204 // to the base address the object.
205 //
206 // We go for 2nd option for simplicity.
207 if (!isIdentifiedObject(V))
208 return false;
209
210 // This function needs to use the aligned object size because we allow
211 // reads a bit past the end given sufficient alignment.
212 uint64_t ObjectSize = getObjectSize(V, DL, TLI, NullIsValidLoc,
213 /*RoundToAlign*/ true);
214
215 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize < Size;
216 }
217
218 /// Returns true if we can prove that the object specified by V has size Size.
isObjectSize(const Value * V,uint64_t Size,const DataLayout & DL,const TargetLibraryInfo & TLI,bool NullIsValidLoc)219 static bool isObjectSize(const Value *V, uint64_t Size, const DataLayout &DL,
220 const TargetLibraryInfo &TLI, bool NullIsValidLoc) {
221 uint64_t ObjectSize = getObjectSize(V, DL, TLI, NullIsValidLoc);
222 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize == Size;
223 }
224
225 //===----------------------------------------------------------------------===//
226 // GetElementPtr Instruction Decomposition and Analysis
227 //===----------------------------------------------------------------------===//
228
229 /// Analyzes the specified value as a linear expression: "A*V + B", where A and
230 /// B are constant integers.
231 ///
232 /// Returns the scale and offset values as APInts and return V as a Value*, and
233 /// return whether we looked through any sign or zero extends. The incoming
234 /// Value is known to have IntegerType, and it may already be sign or zero
235 /// extended.
236 ///
237 /// Note that this looks through extends, so the high bits may not be
238 /// represented in the result.
GetLinearExpression(const Value * V,APInt & Scale,APInt & Offset,unsigned & ZExtBits,unsigned & SExtBits,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,DominatorTree * DT,bool & NSW,bool & NUW)239 /*static*/ const Value *BasicAAResult::GetLinearExpression(
240 const Value *V, APInt &Scale, APInt &Offset, unsigned &ZExtBits,
241 unsigned &SExtBits, const DataLayout &DL, unsigned Depth,
242 AssumptionCache *AC, DominatorTree *DT, bool &NSW, bool &NUW) {
243 assert(V->getType()->isIntegerTy() && "Not an integer value");
244
245 // Limit our recursion depth.
246 if (Depth == 6) {
247 Scale = 1;
248 Offset = 0;
249 return V;
250 }
251
252 if (const ConstantInt *Const = dyn_cast<ConstantInt>(V)) {
253 // If it's a constant, just convert it to an offset and remove the variable.
254 // If we've been called recursively, the Offset bit width will be greater
255 // than the constant's (the Offset's always as wide as the outermost call),
256 // so we'll zext here and process any extension in the isa<SExtInst> &
257 // isa<ZExtInst> cases below.
258 Offset += Const->getValue().zextOrSelf(Offset.getBitWidth());
259 assert(Scale == 0 && "Constant values don't have a scale");
260 return V;
261 }
262
263 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
264 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
265 // If we've been called recursively, then Offset and Scale will be wider
266 // than the BOp operands. We'll always zext it here as we'll process sign
267 // extensions below (see the isa<SExtInst> / isa<ZExtInst> cases).
268 APInt RHS = RHSC->getValue().zextOrSelf(Offset.getBitWidth());
269
270 switch (BOp->getOpcode()) {
271 default:
272 // We don't understand this instruction, so we can't decompose it any
273 // further.
274 Scale = 1;
275 Offset = 0;
276 return V;
277 case Instruction::Or:
278 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't
279 // analyze it.
280 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), DL, 0, AC,
281 BOp, DT)) {
282 Scale = 1;
283 Offset = 0;
284 return V;
285 }
286 LLVM_FALLTHROUGH;
287 case Instruction::Add:
288 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
289 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
290 Offset += RHS;
291 break;
292 case Instruction::Sub:
293 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
294 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
295 Offset -= RHS;
296 break;
297 case Instruction::Mul:
298 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
299 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
300 Offset *= RHS;
301 Scale *= RHS;
302 break;
303 case Instruction::Shl:
304 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
305 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
306
307 // We're trying to linearize an expression of the kind:
308 // shl i8 -128, 36
309 // where the shift count exceeds the bitwidth of the type.
310 // We can't decompose this further (the expression would return
311 // a poison value).
312 if (Offset.getBitWidth() < RHS.getLimitedValue() ||
313 Scale.getBitWidth() < RHS.getLimitedValue()) {
314 Scale = 1;
315 Offset = 0;
316 return V;
317 }
318
319 Offset <<= RHS.getLimitedValue();
320 Scale <<= RHS.getLimitedValue();
321 // the semantics of nsw and nuw for left shifts don't match those of
322 // multiplications, so we won't propagate them.
323 NSW = NUW = false;
324 return V;
325 }
326
327 if (isa<OverflowingBinaryOperator>(BOp)) {
328 NUW &= BOp->hasNoUnsignedWrap();
329 NSW &= BOp->hasNoSignedWrap();
330 }
331 return V;
332 }
333 }
334
335 // Since GEP indices are sign extended anyway, we don't care about the high
336 // bits of a sign or zero extended value - just scales and offsets. The
337 // extensions have to be consistent though.
338 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) {
339 Value *CastOp = cast<CastInst>(V)->getOperand(0);
340 unsigned NewWidth = V->getType()->getPrimitiveSizeInBits();
341 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
342 unsigned OldZExtBits = ZExtBits, OldSExtBits = SExtBits;
343 const Value *Result =
344 GetLinearExpression(CastOp, Scale, Offset, ZExtBits, SExtBits, DL,
345 Depth + 1, AC, DT, NSW, NUW);
346
347 // zext(zext(%x)) == zext(%x), and similarly for sext; we'll handle this
348 // by just incrementing the number of bits we've extended by.
349 unsigned ExtendedBy = NewWidth - SmallWidth;
350
351 if (isa<SExtInst>(V) && ZExtBits == 0) {
352 // sext(sext(%x, a), b) == sext(%x, a + b)
353
354 if (NSW) {
355 // We haven't sign-wrapped, so it's valid to decompose sext(%x + c)
356 // into sext(%x) + sext(c). We'll sext the Offset ourselves:
357 unsigned OldWidth = Offset.getBitWidth();
358 Offset = Offset.trunc(SmallWidth).sext(NewWidth).zextOrSelf(OldWidth);
359 } else {
360 // We may have signed-wrapped, so don't decompose sext(%x + c) into
361 // sext(%x) + sext(c)
362 Scale = 1;
363 Offset = 0;
364 Result = CastOp;
365 ZExtBits = OldZExtBits;
366 SExtBits = OldSExtBits;
367 }
368 SExtBits += ExtendedBy;
369 } else {
370 // sext(zext(%x, a), b) = zext(zext(%x, a), b) = zext(%x, a + b)
371
372 if (!NUW) {
373 // We may have unsigned-wrapped, so don't decompose zext(%x + c) into
374 // zext(%x) + zext(c)
375 Scale = 1;
376 Offset = 0;
377 Result = CastOp;
378 ZExtBits = OldZExtBits;
379 SExtBits = OldSExtBits;
380 }
381 ZExtBits += ExtendedBy;
382 }
383
384 return Result;
385 }
386
387 Scale = 1;
388 Offset = 0;
389 return V;
390 }
391
392 /// To ensure a pointer offset fits in an integer of size PointerSize
393 /// (in bits) when that size is smaller than the maximum pointer size. This is
394 /// an issue, for example, in particular for 32b pointers with negative indices
395 /// that rely on two's complement wrap-arounds for precise alias information
396 /// where the maximum pointer size is 64b.
adjustToPointerSize(APInt Offset,unsigned PointerSize)397 static APInt adjustToPointerSize(APInt Offset, unsigned PointerSize) {
398 assert(PointerSize <= Offset.getBitWidth() && "Invalid PointerSize!");
399 unsigned ShiftBits = Offset.getBitWidth() - PointerSize;
400 return (Offset << ShiftBits).ashr(ShiftBits);
401 }
402
getMaxPointerSize(const DataLayout & DL)403 static unsigned getMaxPointerSize(const DataLayout &DL) {
404 unsigned MaxPointerSize = DL.getMaxPointerSizeInBits();
405 if (MaxPointerSize < 64 && ForceAtLeast64Bits) MaxPointerSize = 64;
406 if (DoubleCalcBits) MaxPointerSize *= 2;
407
408 return MaxPointerSize;
409 }
410
411 /// If V is a symbolic pointer expression, decompose it into a base pointer
412 /// with a constant offset and a number of scaled symbolic offsets.
413 ///
414 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale
415 /// in the VarIndices vector) are Value*'s that are known to be scaled by the
416 /// specified amount, but which may have other unrepresented high bits. As
417 /// such, the gep cannot necessarily be reconstructed from its decomposed form.
418 ///
419 /// When DataLayout is around, this function is capable of analyzing everything
420 /// that GetUnderlyingObject can look through. To be able to do that
421 /// GetUnderlyingObject and DecomposeGEPExpression must use the same search
422 /// depth (MaxLookupSearchDepth). When DataLayout not is around, it just looks
423 /// through pointer casts.
DecomposeGEPExpression(const Value * V,DecomposedGEP & Decomposed,const DataLayout & DL,AssumptionCache * AC,DominatorTree * DT)424 bool BasicAAResult::DecomposeGEPExpression(const Value *V,
425 DecomposedGEP &Decomposed, const DataLayout &DL, AssumptionCache *AC,
426 DominatorTree *DT) {
427 // Limit recursion depth to limit compile time in crazy cases.
428 unsigned MaxLookup = MaxLookupSearchDepth;
429 SearchTimes++;
430
431 unsigned MaxPointerSize = getMaxPointerSize(DL);
432 Decomposed.VarIndices.clear();
433 do {
434 // See if this is a bitcast or GEP.
435 const Operator *Op = dyn_cast<Operator>(V);
436 if (!Op) {
437 // The only non-operator case we can handle are GlobalAliases.
438 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
439 if (!GA->isInterposable()) {
440 V = GA->getAliasee();
441 continue;
442 }
443 }
444 Decomposed.Base = V;
445 return false;
446 }
447
448 if (Op->getOpcode() == Instruction::BitCast ||
449 Op->getOpcode() == Instruction::AddrSpaceCast) {
450 V = Op->getOperand(0);
451 continue;
452 }
453
454 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
455 if (!GEPOp) {
456 if (const auto *Call = dyn_cast<CallBase>(V)) {
457 // CaptureTracking can know about special capturing properties of some
458 // intrinsics like launder.invariant.group, that can't be expressed with
459 // the attributes, but have properties like returning aliasing pointer.
460 // Because some analysis may assume that nocaptured pointer is not
461 // returned from some special intrinsic (because function would have to
462 // be marked with returns attribute), it is crucial to use this function
463 // because it should be in sync with CaptureTracking. Not using it may
464 // cause weird miscompilations where 2 aliasing pointers are assumed to
465 // noalias.
466 if (auto *RP = getArgumentAliasingToReturnedPointer(Call)) {
467 V = RP;
468 continue;
469 }
470 }
471
472 // If it's not a GEP, hand it off to SimplifyInstruction to see if it
473 // can come up with something. This matches what GetUnderlyingObject does.
474 if (const Instruction *I = dyn_cast<Instruction>(V))
475 // TODO: Get a DominatorTree and AssumptionCache and use them here
476 // (these are both now available in this function, but this should be
477 // updated when GetUnderlyingObject is updated). TLI should be
478 // provided also.
479 if (const Value *Simplified =
480 SimplifyInstruction(const_cast<Instruction *>(I), DL)) {
481 V = Simplified;
482 continue;
483 }
484
485 Decomposed.Base = V;
486 return false;
487 }
488
489 // Don't attempt to analyze GEPs over unsized objects.
490 if (!GEPOp->getSourceElementType()->isSized()) {
491 Decomposed.Base = V;
492 return false;
493 }
494
495 unsigned AS = GEPOp->getPointerAddressSpace();
496 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
497 gep_type_iterator GTI = gep_type_begin(GEPOp);
498 unsigned PointerSize = DL.getPointerSizeInBits(AS);
499 // Assume all GEP operands are constants until proven otherwise.
500 bool GepHasConstantOffset = true;
501 for (User::const_op_iterator I = GEPOp->op_begin() + 1, E = GEPOp->op_end();
502 I != E; ++I, ++GTI) {
503 const Value *Index = *I;
504 // Compute the (potentially symbolic) offset in bytes for this index.
505 if (StructType *STy = GTI.getStructTypeOrNull()) {
506 // For a struct, add the member offset.
507 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
508 if (FieldNo == 0)
509 continue;
510
511 Decomposed.StructOffset +=
512 DL.getStructLayout(STy)->getElementOffset(FieldNo);
513 continue;
514 }
515
516 // For an array/pointer, add the element offset, explicitly scaled.
517 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
518 if (CIdx->isZero())
519 continue;
520 Decomposed.OtherOffset +=
521 (DL.getTypeAllocSize(GTI.getIndexedType()) *
522 CIdx->getValue().sextOrSelf(MaxPointerSize))
523 .sextOrTrunc(MaxPointerSize);
524 continue;
525 }
526
527 GepHasConstantOffset = false;
528
529 APInt Scale(MaxPointerSize, DL.getTypeAllocSize(GTI.getIndexedType()));
530 unsigned ZExtBits = 0, SExtBits = 0;
531
532 // If the integer type is smaller than the pointer size, it is implicitly
533 // sign extended to pointer size.
534 unsigned Width = Index->getType()->getIntegerBitWidth();
535 if (PointerSize > Width)
536 SExtBits += PointerSize - Width;
537
538 // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
539 APInt IndexScale(Width, 0), IndexOffset(Width, 0);
540 bool NSW = true, NUW = true;
541 const Value *OrigIndex = Index;
542 Index = GetLinearExpression(Index, IndexScale, IndexOffset, ZExtBits,
543 SExtBits, DL, 0, AC, DT, NSW, NUW);
544
545 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
546 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
547
548 // It can be the case that, even through C1*V+C2 does not overflow for
549 // relevant values of V, (C2*Scale) can overflow. In that case, we cannot
550 // decompose the expression in this way.
551 //
552 // FIXME: C1*Scale and the other operations in the decomposed
553 // (C1*Scale)*V+C2*Scale can also overflow. We should check for this
554 // possibility.
555 APInt WideScaledOffset = IndexOffset.sextOrTrunc(MaxPointerSize*2) *
556 Scale.sext(MaxPointerSize*2);
557 if (WideScaledOffset.getMinSignedBits() > MaxPointerSize) {
558 Index = OrigIndex;
559 IndexScale = 1;
560 IndexOffset = 0;
561
562 ZExtBits = SExtBits = 0;
563 if (PointerSize > Width)
564 SExtBits += PointerSize - Width;
565 } else {
566 Decomposed.OtherOffset += IndexOffset.sextOrTrunc(MaxPointerSize) * Scale;
567 Scale *= IndexScale.sextOrTrunc(MaxPointerSize);
568 }
569
570 // If we already had an occurrence of this index variable, merge this
571 // scale into it. For example, we want to handle:
572 // A[x][x] -> x*16 + x*4 -> x*20
573 // This also ensures that 'x' only appears in the index list once.
574 for (unsigned i = 0, e = Decomposed.VarIndices.size(); i != e; ++i) {
575 if (Decomposed.VarIndices[i].V == Index &&
576 Decomposed.VarIndices[i].ZExtBits == ZExtBits &&
577 Decomposed.VarIndices[i].SExtBits == SExtBits) {
578 Scale += Decomposed.VarIndices[i].Scale;
579 Decomposed.VarIndices.erase(Decomposed.VarIndices.begin() + i);
580 break;
581 }
582 }
583
584 // Make sure that we have a scale that makes sense for this target's
585 // pointer size.
586 Scale = adjustToPointerSize(Scale, PointerSize);
587
588 if (!!Scale) {
589 VariableGEPIndex Entry = {Index, ZExtBits, SExtBits, Scale};
590 Decomposed.VarIndices.push_back(Entry);
591 }
592 }
593
594 // Take care of wrap-arounds
595 if (GepHasConstantOffset) {
596 Decomposed.StructOffset =
597 adjustToPointerSize(Decomposed.StructOffset, PointerSize);
598 Decomposed.OtherOffset =
599 adjustToPointerSize(Decomposed.OtherOffset, PointerSize);
600 }
601
602 // Analyze the base pointer next.
603 V = GEPOp->getOperand(0);
604 } while (--MaxLookup);
605
606 // If the chain of expressions is too deep, just return early.
607 Decomposed.Base = V;
608 SearchLimitReached++;
609 return true;
610 }
611
612 /// Returns whether the given pointer value points to memory that is local to
613 /// the function, with global constants being considered local to all
614 /// functions.
pointsToConstantMemory(const MemoryLocation & Loc,bool OrLocal)615 bool BasicAAResult::pointsToConstantMemory(const MemoryLocation &Loc,
616 bool OrLocal) {
617 assert(Visited.empty() && "Visited must be cleared after use!");
618
619 unsigned MaxLookup = 8;
620 SmallVector<const Value *, 16> Worklist;
621 Worklist.push_back(Loc.Ptr);
622 do {
623 const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), DL);
624 if (!Visited.insert(V).second) {
625 Visited.clear();
626 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
627 }
628
629 // An alloca instruction defines local memory.
630 if (OrLocal && isa<AllocaInst>(V))
631 continue;
632
633 // A global constant counts as local memory for our purposes.
634 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
635 // Note: this doesn't require GV to be "ODR" because it isn't legal for a
636 // global to be marked constant in some modules and non-constant in
637 // others. GV may even be a declaration, not a definition.
638 if (!GV->isConstant()) {
639 Visited.clear();
640 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
641 }
642 continue;
643 }
644
645 // If both select values point to local memory, then so does the select.
646 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
647 Worklist.push_back(SI->getTrueValue());
648 Worklist.push_back(SI->getFalseValue());
649 continue;
650 }
651
652 // If all values incoming to a phi node point to local memory, then so does
653 // the phi.
654 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
655 // Don't bother inspecting phi nodes with many operands.
656 if (PN->getNumIncomingValues() > MaxLookup) {
657 Visited.clear();
658 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
659 }
660 for (Value *IncValue : PN->incoming_values())
661 Worklist.push_back(IncValue);
662 continue;
663 }
664
665 // Otherwise be conservative.
666 Visited.clear();
667 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
668 } while (!Worklist.empty() && --MaxLookup);
669
670 Visited.clear();
671 return Worklist.empty();
672 }
673
674 /// Returns the behavior when calling the given call site.
getModRefBehavior(const CallBase * Call)675 FunctionModRefBehavior BasicAAResult::getModRefBehavior(const CallBase *Call) {
676 if (Call->doesNotAccessMemory())
677 // Can't do better than this.
678 return FMRB_DoesNotAccessMemory;
679
680 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
681
682 // If the callsite knows it only reads memory, don't return worse
683 // than that.
684 if (Call->onlyReadsMemory())
685 Min = FMRB_OnlyReadsMemory;
686 else if (Call->doesNotReadMemory())
687 Min = FMRB_DoesNotReadMemory;
688
689 if (Call->onlyAccessesArgMemory())
690 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
691 else if (Call->onlyAccessesInaccessibleMemory())
692 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleMem);
693 else if (Call->onlyAccessesInaccessibleMemOrArgMem())
694 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleOrArgMem);
695
696 // If the call has operand bundles then aliasing attributes from the function
697 // it calls do not directly apply to the call. This can be made more precise
698 // in the future.
699 if (!Call->hasOperandBundles())
700 if (const Function *F = Call->getCalledFunction())
701 Min =
702 FunctionModRefBehavior(Min & getBestAAResults().getModRefBehavior(F));
703
704 return Min;
705 }
706
707 /// Returns the behavior when calling the given function. For use when the call
708 /// site is not known.
getModRefBehavior(const Function * F)709 FunctionModRefBehavior BasicAAResult::getModRefBehavior(const Function *F) {
710 // If the function declares it doesn't access memory, we can't do better.
711 if (F->doesNotAccessMemory())
712 return FMRB_DoesNotAccessMemory;
713
714 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
715
716 // If the function declares it only reads memory, go with that.
717 if (F->onlyReadsMemory())
718 Min = FMRB_OnlyReadsMemory;
719 else if (F->doesNotReadMemory())
720 Min = FMRB_DoesNotReadMemory;
721
722 if (F->onlyAccessesArgMemory())
723 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
724 else if (F->onlyAccessesInaccessibleMemory())
725 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleMem);
726 else if (F->onlyAccessesInaccessibleMemOrArgMem())
727 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleOrArgMem);
728
729 return Min;
730 }
731
732 /// Returns true if this is a writeonly (i.e Mod only) parameter.
isWriteOnlyParam(const CallBase * Call,unsigned ArgIdx,const TargetLibraryInfo & TLI)733 static bool isWriteOnlyParam(const CallBase *Call, unsigned ArgIdx,
734 const TargetLibraryInfo &TLI) {
735 if (Call->paramHasAttr(ArgIdx, Attribute::WriteOnly))
736 return true;
737
738 // We can bound the aliasing properties of memset_pattern16 just as we can
739 // for memcpy/memset. This is particularly important because the
740 // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
741 // whenever possible.
742 // FIXME Consider handling this in InferFunctionAttr.cpp together with other
743 // attributes.
744 LibFunc F;
745 if (Call->getCalledFunction() &&
746 TLI.getLibFunc(*Call->getCalledFunction(), F) &&
747 F == LibFunc_memset_pattern16 && TLI.has(F))
748 if (ArgIdx == 0)
749 return true;
750
751 // TODO: memset_pattern4, memset_pattern8
752 // TODO: _chk variants
753 // TODO: strcmp, strcpy
754
755 return false;
756 }
757
getArgModRefInfo(const CallBase * Call,unsigned ArgIdx)758 ModRefInfo BasicAAResult::getArgModRefInfo(const CallBase *Call,
759 unsigned ArgIdx) {
760 // Checking for known builtin intrinsics and target library functions.
761 if (isWriteOnlyParam(Call, ArgIdx, TLI))
762 return ModRefInfo::Mod;
763
764 if (Call->paramHasAttr(ArgIdx, Attribute::ReadOnly))
765 return ModRefInfo::Ref;
766
767 if (Call->paramHasAttr(ArgIdx, Attribute::ReadNone))
768 return ModRefInfo::NoModRef;
769
770 return AAResultBase::getArgModRefInfo(Call, ArgIdx);
771 }
772
isIntrinsicCall(const CallBase * Call,Intrinsic::ID IID)773 static bool isIntrinsicCall(const CallBase *Call, Intrinsic::ID IID) {
774 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Call);
775 return II && II->getIntrinsicID() == IID;
776 }
777
778 #ifndef NDEBUG
getParent(const Value * V)779 static const Function *getParent(const Value *V) {
780 if (const Instruction *inst = dyn_cast<Instruction>(V)) {
781 if (!inst->getParent())
782 return nullptr;
783 return inst->getParent()->getParent();
784 }
785
786 if (const Argument *arg = dyn_cast<Argument>(V))
787 return arg->getParent();
788
789 return nullptr;
790 }
791
notDifferentParent(const Value * O1,const Value * O2)792 static bool notDifferentParent(const Value *O1, const Value *O2) {
793
794 const Function *F1 = getParent(O1);
795 const Function *F2 = getParent(O2);
796
797 return !F1 || !F2 || F1 == F2;
798 }
799 #endif
800
alias(const MemoryLocation & LocA,const MemoryLocation & LocB)801 AliasResult BasicAAResult::alias(const MemoryLocation &LocA,
802 const MemoryLocation &LocB) {
803 assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
804 "BasicAliasAnalysis doesn't support interprocedural queries.");
805
806 // If we have a directly cached entry for these locations, we have recursed
807 // through this once, so just return the cached results. Notably, when this
808 // happens, we don't clear the cache.
809 auto CacheIt = AliasCache.find(LocPair(LocA, LocB));
810 if (CacheIt != AliasCache.end())
811 return CacheIt->second;
812
813 AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.AATags, LocB.Ptr,
814 LocB.Size, LocB.AATags);
815 // AliasCache rarely has more than 1 or 2 elements, always use
816 // shrink_and_clear so it quickly returns to the inline capacity of the
817 // SmallDenseMap if it ever grows larger.
818 // FIXME: This should really be shrink_to_inline_capacity_and_clear().
819 AliasCache.shrink_and_clear();
820 VisitedPhiBBs.clear();
821 return Alias;
822 }
823
824 /// Checks to see if the specified callsite can clobber the specified memory
825 /// object.
826 ///
827 /// Since we only look at local properties of this function, we really can't
828 /// say much about this query. We do, however, use simple "address taken"
829 /// analysis on local objects.
getModRefInfo(const CallBase * Call,const MemoryLocation & Loc)830 ModRefInfo BasicAAResult::getModRefInfo(const CallBase *Call,
831 const MemoryLocation &Loc) {
832 assert(notDifferentParent(Call, Loc.Ptr) &&
833 "AliasAnalysis query involving multiple functions!");
834
835 const Value *Object = GetUnderlyingObject(Loc.Ptr, DL);
836
837 // Calls marked 'tail' cannot read or write allocas from the current frame
838 // because the current frame might be destroyed by the time they run. However,
839 // a tail call may use an alloca with byval. Calling with byval copies the
840 // contents of the alloca into argument registers or stack slots, so there is
841 // no lifetime issue.
842 if (isa<AllocaInst>(Object))
843 if (const CallInst *CI = dyn_cast<CallInst>(Call))
844 if (CI->isTailCall() &&
845 !CI->getAttributes().hasAttrSomewhere(Attribute::ByVal))
846 return ModRefInfo::NoModRef;
847
848 // Stack restore is able to modify unescaped dynamic allocas. Assume it may
849 // modify them even though the alloca is not escaped.
850 if (auto *AI = dyn_cast<AllocaInst>(Object))
851 if (!AI->isStaticAlloca() && isIntrinsicCall(Call, Intrinsic::stackrestore))
852 return ModRefInfo::Mod;
853
854 // If the pointer is to a locally allocated object that does not escape,
855 // then the call can not mod/ref the pointer unless the call takes the pointer
856 // as an argument, and itself doesn't capture it.
857 if (!isa<Constant>(Object) && Call != Object &&
858 isNonEscapingLocalObject(Object)) {
859
860 // Optimistically assume that call doesn't touch Object and check this
861 // assumption in the following loop.
862 ModRefInfo Result = ModRefInfo::NoModRef;
863 bool IsMustAlias = true;
864
865 unsigned OperandNo = 0;
866 for (auto CI = Call->data_operands_begin(), CE = Call->data_operands_end();
867 CI != CE; ++CI, ++OperandNo) {
868 // Only look at the no-capture or byval pointer arguments. If this
869 // pointer were passed to arguments that were neither of these, then it
870 // couldn't be no-capture.
871 if (!(*CI)->getType()->isPointerTy() ||
872 (!Call->doesNotCapture(OperandNo) &&
873 OperandNo < Call->getNumArgOperands() &&
874 !Call->isByValArgument(OperandNo)))
875 continue;
876
877 // Call doesn't access memory through this operand, so we don't care
878 // if it aliases with Object.
879 if (Call->doesNotAccessMemory(OperandNo))
880 continue;
881
882 // If this is a no-capture pointer argument, see if we can tell that it
883 // is impossible to alias the pointer we're checking.
884 AliasResult AR =
885 getBestAAResults().alias(MemoryLocation(*CI), MemoryLocation(Object));
886 if (AR != MustAlias)
887 IsMustAlias = false;
888 // Operand doesnt alias 'Object', continue looking for other aliases
889 if (AR == NoAlias)
890 continue;
891 // Operand aliases 'Object', but call doesn't modify it. Strengthen
892 // initial assumption and keep looking in case if there are more aliases.
893 if (Call->onlyReadsMemory(OperandNo)) {
894 Result = setRef(Result);
895 continue;
896 }
897 // Operand aliases 'Object' but call only writes into it.
898 if (Call->doesNotReadMemory(OperandNo)) {
899 Result = setMod(Result);
900 continue;
901 }
902 // This operand aliases 'Object' and call reads and writes into it.
903 // Setting ModRef will not yield an early return below, MustAlias is not
904 // used further.
905 Result = ModRefInfo::ModRef;
906 break;
907 }
908
909 // No operand aliases, reset Must bit. Add below if at least one aliases
910 // and all aliases found are MustAlias.
911 if (isNoModRef(Result))
912 IsMustAlias = false;
913
914 // Early return if we improved mod ref information
915 if (!isModAndRefSet(Result)) {
916 if (isNoModRef(Result))
917 return ModRefInfo::NoModRef;
918 return IsMustAlias ? setMust(Result) : clearMust(Result);
919 }
920 }
921
922 // If the call is to malloc or calloc, we can assume that it doesn't
923 // modify any IR visible value. This is only valid because we assume these
924 // routines do not read values visible in the IR. TODO: Consider special
925 // casing realloc and strdup routines which access only their arguments as
926 // well. Or alternatively, replace all of this with inaccessiblememonly once
927 // that's implemented fully.
928 if (isMallocOrCallocLikeFn(Call, &TLI)) {
929 // Be conservative if the accessed pointer may alias the allocation -
930 // fallback to the generic handling below.
931 if (getBestAAResults().alias(MemoryLocation(Call), Loc) == NoAlias)
932 return ModRefInfo::NoModRef;
933 }
934
935 // The semantics of memcpy intrinsics forbid overlap between their respective
936 // operands, i.e., source and destination of any given memcpy must no-alias.
937 // If Loc must-aliases either one of these two locations, then it necessarily
938 // no-aliases the other.
939 if (auto *Inst = dyn_cast<AnyMemCpyInst>(Call)) {
940 AliasResult SrcAA, DestAA;
941
942 if ((SrcAA = getBestAAResults().alias(MemoryLocation::getForSource(Inst),
943 Loc)) == MustAlias)
944 // Loc is exactly the memcpy source thus disjoint from memcpy dest.
945 return ModRefInfo::Ref;
946 if ((DestAA = getBestAAResults().alias(MemoryLocation::getForDest(Inst),
947 Loc)) == MustAlias)
948 // The converse case.
949 return ModRefInfo::Mod;
950
951 // It's also possible for Loc to alias both src and dest, or neither.
952 ModRefInfo rv = ModRefInfo::NoModRef;
953 if (SrcAA != NoAlias)
954 rv = setRef(rv);
955 if (DestAA != NoAlias)
956 rv = setMod(rv);
957 return rv;
958 }
959
960 // While the assume intrinsic is marked as arbitrarily writing so that
961 // proper control dependencies will be maintained, it never aliases any
962 // particular memory location.
963 if (isIntrinsicCall(Call, Intrinsic::assume))
964 return ModRefInfo::NoModRef;
965
966 // Like assumes, guard intrinsics are also marked as arbitrarily writing so
967 // that proper control dependencies are maintained but they never mods any
968 // particular memory location.
969 //
970 // *Unlike* assumes, guard intrinsics are modeled as reading memory since the
971 // heap state at the point the guard is issued needs to be consistent in case
972 // the guard invokes the "deopt" continuation.
973 if (isIntrinsicCall(Call, Intrinsic::experimental_guard))
974 return ModRefInfo::Ref;
975
976 // Like assumes, invariant.start intrinsics were also marked as arbitrarily
977 // writing so that proper control dependencies are maintained but they never
978 // mod any particular memory location visible to the IR.
979 // *Unlike* assumes (which are now modeled as NoModRef), invariant.start
980 // intrinsic is now modeled as reading memory. This prevents hoisting the
981 // invariant.start intrinsic over stores. Consider:
982 // *ptr = 40;
983 // *ptr = 50;
984 // invariant_start(ptr)
985 // int val = *ptr;
986 // print(val);
987 //
988 // This cannot be transformed to:
989 //
990 // *ptr = 40;
991 // invariant_start(ptr)
992 // *ptr = 50;
993 // int val = *ptr;
994 // print(val);
995 //
996 // The transformation will cause the second store to be ignored (based on
997 // rules of invariant.start) and print 40, while the first program always
998 // prints 50.
999 if (isIntrinsicCall(Call, Intrinsic::invariant_start))
1000 return ModRefInfo::Ref;
1001
1002 // The AAResultBase base class has some smarts, lets use them.
1003 return AAResultBase::getModRefInfo(Call, Loc);
1004 }
1005
getModRefInfo(const CallBase * Call1,const CallBase * Call2)1006 ModRefInfo BasicAAResult::getModRefInfo(const CallBase *Call1,
1007 const CallBase *Call2) {
1008 // While the assume intrinsic is marked as arbitrarily writing so that
1009 // proper control dependencies will be maintained, it never aliases any
1010 // particular memory location.
1011 if (isIntrinsicCall(Call1, Intrinsic::assume) ||
1012 isIntrinsicCall(Call2, Intrinsic::assume))
1013 return ModRefInfo::NoModRef;
1014
1015 // Like assumes, guard intrinsics are also marked as arbitrarily writing so
1016 // that proper control dependencies are maintained but they never mod any
1017 // particular memory location.
1018 //
1019 // *Unlike* assumes, guard intrinsics are modeled as reading memory since the
1020 // heap state at the point the guard is issued needs to be consistent in case
1021 // the guard invokes the "deopt" continuation.
1022
1023 // NB! This function is *not* commutative, so we specical case two
1024 // possibilities for guard intrinsics.
1025
1026 if (isIntrinsicCall(Call1, Intrinsic::experimental_guard))
1027 return isModSet(createModRefInfo(getModRefBehavior(Call2)))
1028 ? ModRefInfo::Ref
1029 : ModRefInfo::NoModRef;
1030
1031 if (isIntrinsicCall(Call2, Intrinsic::experimental_guard))
1032 return isModSet(createModRefInfo(getModRefBehavior(Call1)))
1033 ? ModRefInfo::Mod
1034 : ModRefInfo::NoModRef;
1035
1036 // The AAResultBase base class has some smarts, lets use them.
1037 return AAResultBase::getModRefInfo(Call1, Call2);
1038 }
1039
1040 /// Provide ad-hoc rules to disambiguate accesses through two GEP operators,
1041 /// both having the exact same pointer operand.
aliasSameBasePointerGEPs(const GEPOperator * GEP1,LocationSize MaybeV1Size,const GEPOperator * GEP2,LocationSize MaybeV2Size,const DataLayout & DL)1042 static AliasResult aliasSameBasePointerGEPs(const GEPOperator *GEP1,
1043 LocationSize MaybeV1Size,
1044 const GEPOperator *GEP2,
1045 LocationSize MaybeV2Size,
1046 const DataLayout &DL) {
1047 assert(GEP1->getPointerOperand()->stripPointerCastsAndInvariantGroups() ==
1048 GEP2->getPointerOperand()->stripPointerCastsAndInvariantGroups() &&
1049 GEP1->getPointerOperandType() == GEP2->getPointerOperandType() &&
1050 "Expected GEPs with the same pointer operand");
1051
1052 // Try to determine whether GEP1 and GEP2 index through arrays, into structs,
1053 // such that the struct field accesses provably cannot alias.
1054 // We also need at least two indices (the pointer, and the struct field).
1055 if (GEP1->getNumIndices() != GEP2->getNumIndices() ||
1056 GEP1->getNumIndices() < 2)
1057 return MayAlias;
1058
1059 // If we don't know the size of the accesses through both GEPs, we can't
1060 // determine whether the struct fields accessed can't alias.
1061 if (MaybeV1Size == LocationSize::unknown() ||
1062 MaybeV2Size == LocationSize::unknown())
1063 return MayAlias;
1064
1065 const uint64_t V1Size = MaybeV1Size.getValue();
1066 const uint64_t V2Size = MaybeV2Size.getValue();
1067
1068 ConstantInt *C1 =
1069 dyn_cast<ConstantInt>(GEP1->getOperand(GEP1->getNumOperands() - 1));
1070 ConstantInt *C2 =
1071 dyn_cast<ConstantInt>(GEP2->getOperand(GEP2->getNumOperands() - 1));
1072
1073 // If the last (struct) indices are constants and are equal, the other indices
1074 // might be also be dynamically equal, so the GEPs can alias.
1075 if (C1 && C2) {
1076 unsigned BitWidth = std::max(C1->getBitWidth(), C2->getBitWidth());
1077 if (C1->getValue().sextOrSelf(BitWidth) ==
1078 C2->getValue().sextOrSelf(BitWidth))
1079 return MayAlias;
1080 }
1081
1082 // Find the last-indexed type of the GEP, i.e., the type you'd get if
1083 // you stripped the last index.
1084 // On the way, look at each indexed type. If there's something other
1085 // than an array, different indices can lead to different final types.
1086 SmallVector<Value *, 8> IntermediateIndices;
1087
1088 // Insert the first index; we don't need to check the type indexed
1089 // through it as it only drops the pointer indirection.
1090 assert(GEP1->getNumIndices() > 1 && "Not enough GEP indices to examine");
1091 IntermediateIndices.push_back(GEP1->getOperand(1));
1092
1093 // Insert all the remaining indices but the last one.
1094 // Also, check that they all index through arrays.
1095 for (unsigned i = 1, e = GEP1->getNumIndices() - 1; i != e; ++i) {
1096 if (!isa<ArrayType>(GetElementPtrInst::getIndexedType(
1097 GEP1->getSourceElementType(), IntermediateIndices)))
1098 return MayAlias;
1099 IntermediateIndices.push_back(GEP1->getOperand(i + 1));
1100 }
1101
1102 auto *Ty = GetElementPtrInst::getIndexedType(
1103 GEP1->getSourceElementType(), IntermediateIndices);
1104 StructType *LastIndexedStruct = dyn_cast<StructType>(Ty);
1105
1106 if (isa<SequentialType>(Ty)) {
1107 // We know that:
1108 // - both GEPs begin indexing from the exact same pointer;
1109 // - the last indices in both GEPs are constants, indexing into a sequential
1110 // type (array or pointer);
1111 // - both GEPs only index through arrays prior to that.
1112 //
1113 // Because array indices greater than the number of elements are valid in
1114 // GEPs, unless we know the intermediate indices are identical between
1115 // GEP1 and GEP2 we cannot guarantee that the last indexed arrays don't
1116 // partially overlap. We also need to check that the loaded size matches
1117 // the element size, otherwise we could still have overlap.
1118 const uint64_t ElementSize =
1119 DL.getTypeStoreSize(cast<SequentialType>(Ty)->getElementType());
1120 if (V1Size != ElementSize || V2Size != ElementSize)
1121 return MayAlias;
1122
1123 for (unsigned i = 0, e = GEP1->getNumIndices() - 1; i != e; ++i)
1124 if (GEP1->getOperand(i + 1) != GEP2->getOperand(i + 1))
1125 return MayAlias;
1126
1127 // Now we know that the array/pointer that GEP1 indexes into and that
1128 // that GEP2 indexes into must either precisely overlap or be disjoint.
1129 // Because they cannot partially overlap and because fields in an array
1130 // cannot overlap, if we can prove the final indices are different between
1131 // GEP1 and GEP2, we can conclude GEP1 and GEP2 don't alias.
1132
1133 // If the last indices are constants, we've already checked they don't
1134 // equal each other so we can exit early.
1135 if (C1 && C2)
1136 return NoAlias;
1137 {
1138 Value *GEP1LastIdx = GEP1->getOperand(GEP1->getNumOperands() - 1);
1139 Value *GEP2LastIdx = GEP2->getOperand(GEP2->getNumOperands() - 1);
1140 if (isa<PHINode>(GEP1LastIdx) || isa<PHINode>(GEP2LastIdx)) {
1141 // If one of the indices is a PHI node, be safe and only use
1142 // computeKnownBits so we don't make any assumptions about the
1143 // relationships between the two indices. This is important if we're
1144 // asking about values from different loop iterations. See PR32314.
1145 // TODO: We may be able to change the check so we only do this when
1146 // we definitely looked through a PHINode.
1147 if (GEP1LastIdx != GEP2LastIdx &&
1148 GEP1LastIdx->getType() == GEP2LastIdx->getType()) {
1149 KnownBits Known1 = computeKnownBits(GEP1LastIdx, DL);
1150 KnownBits Known2 = computeKnownBits(GEP2LastIdx, DL);
1151 if (Known1.Zero.intersects(Known2.One) ||
1152 Known1.One.intersects(Known2.Zero))
1153 return NoAlias;
1154 }
1155 } else if (isKnownNonEqual(GEP1LastIdx, GEP2LastIdx, DL))
1156 return NoAlias;
1157 }
1158 return MayAlias;
1159 } else if (!LastIndexedStruct || !C1 || !C2) {
1160 return MayAlias;
1161 }
1162
1163 if (C1->getValue().getActiveBits() > 64 ||
1164 C2->getValue().getActiveBits() > 64)
1165 return MayAlias;
1166
1167 // We know that:
1168 // - both GEPs begin indexing from the exact same pointer;
1169 // - the last indices in both GEPs are constants, indexing into a struct;
1170 // - said indices are different, hence, the pointed-to fields are different;
1171 // - both GEPs only index through arrays prior to that.
1172 //
1173 // This lets us determine that the struct that GEP1 indexes into and the
1174 // struct that GEP2 indexes into must either precisely overlap or be
1175 // completely disjoint. Because they cannot partially overlap, indexing into
1176 // different non-overlapping fields of the struct will never alias.
1177
1178 // Therefore, the only remaining thing needed to show that both GEPs can't
1179 // alias is that the fields are not overlapping.
1180 const StructLayout *SL = DL.getStructLayout(LastIndexedStruct);
1181 const uint64_t StructSize = SL->getSizeInBytes();
1182 const uint64_t V1Off = SL->getElementOffset(C1->getZExtValue());
1183 const uint64_t V2Off = SL->getElementOffset(C2->getZExtValue());
1184
1185 auto EltsDontOverlap = [StructSize](uint64_t V1Off, uint64_t V1Size,
1186 uint64_t V2Off, uint64_t V2Size) {
1187 return V1Off < V2Off && V1Off + V1Size <= V2Off &&
1188 ((V2Off + V2Size <= StructSize) ||
1189 (V2Off + V2Size - StructSize <= V1Off));
1190 };
1191
1192 if (EltsDontOverlap(V1Off, V1Size, V2Off, V2Size) ||
1193 EltsDontOverlap(V2Off, V2Size, V1Off, V1Size))
1194 return NoAlias;
1195
1196 return MayAlias;
1197 }
1198
1199 // If a we have (a) a GEP and (b) a pointer based on an alloca, and the
1200 // beginning of the object the GEP points would have a negative offset with
1201 // repsect to the alloca, that means the GEP can not alias pointer (b).
1202 // Note that the pointer based on the alloca may not be a GEP. For
1203 // example, it may be the alloca itself.
1204 // The same applies if (b) is based on a GlobalVariable. Note that just being
1205 // based on isIdentifiedObject() is not enough - we need an identified object
1206 // that does not permit access to negative offsets. For example, a negative
1207 // offset from a noalias argument or call can be inbounds w.r.t the actual
1208 // underlying object.
1209 //
1210 // For example, consider:
1211 //
1212 // struct { int f0, int f1, ...} foo;
1213 // foo alloca;
1214 // foo* random = bar(alloca);
1215 // int *f0 = &alloca.f0
1216 // int *f1 = &random->f1;
1217 //
1218 // Which is lowered, approximately, to:
1219 //
1220 // %alloca = alloca %struct.foo
1221 // %random = call %struct.foo* @random(%struct.foo* %alloca)
1222 // %f0 = getelementptr inbounds %struct, %struct.foo* %alloca, i32 0, i32 0
1223 // %f1 = getelementptr inbounds %struct, %struct.foo* %random, i32 0, i32 1
1224 //
1225 // Assume %f1 and %f0 alias. Then %f1 would point into the object allocated
1226 // by %alloca. Since the %f1 GEP is inbounds, that means %random must also
1227 // point into the same object. But since %f0 points to the beginning of %alloca,
1228 // the highest %f1 can be is (%alloca + 3). This means %random can not be higher
1229 // than (%alloca - 1), and so is not inbounds, a contradiction.
isGEPBaseAtNegativeOffset(const GEPOperator * GEPOp,const DecomposedGEP & DecompGEP,const DecomposedGEP & DecompObject,LocationSize MaybeObjectAccessSize)1230 bool BasicAAResult::isGEPBaseAtNegativeOffset(const GEPOperator *GEPOp,
1231 const DecomposedGEP &DecompGEP, const DecomposedGEP &DecompObject,
1232 LocationSize MaybeObjectAccessSize) {
1233 // If the object access size is unknown, or the GEP isn't inbounds, bail.
1234 if (MaybeObjectAccessSize == LocationSize::unknown() || !GEPOp->isInBounds())
1235 return false;
1236
1237 const uint64_t ObjectAccessSize = MaybeObjectAccessSize.getValue();
1238
1239 // We need the object to be an alloca or a globalvariable, and want to know
1240 // the offset of the pointer from the object precisely, so no variable
1241 // indices are allowed.
1242 if (!(isa<AllocaInst>(DecompObject.Base) ||
1243 isa<GlobalVariable>(DecompObject.Base)) ||
1244 !DecompObject.VarIndices.empty())
1245 return false;
1246
1247 APInt ObjectBaseOffset = DecompObject.StructOffset +
1248 DecompObject.OtherOffset;
1249
1250 // If the GEP has no variable indices, we know the precise offset
1251 // from the base, then use it. If the GEP has variable indices,
1252 // we can't get exact GEP offset to identify pointer alias. So return
1253 // false in that case.
1254 if (!DecompGEP.VarIndices.empty())
1255 return false;
1256
1257 APInt GEPBaseOffset = DecompGEP.StructOffset;
1258 GEPBaseOffset += DecompGEP.OtherOffset;
1259
1260 return GEPBaseOffset.sge(ObjectBaseOffset + (int64_t)ObjectAccessSize);
1261 }
1262
1263 /// Provides a bunch of ad-hoc rules to disambiguate a GEP instruction against
1264 /// another pointer.
1265 ///
1266 /// We know that V1 is a GEP, but we don't know anything about V2.
1267 /// UnderlyingV1 is GetUnderlyingObject(GEP1, DL), UnderlyingV2 is the same for
1268 /// V2.
1269 AliasResult
aliasGEP(const GEPOperator * GEP1,LocationSize V1Size,const AAMDNodes & V1AAInfo,const Value * V2,LocationSize V2Size,const AAMDNodes & V2AAInfo,const Value * UnderlyingV1,const Value * UnderlyingV2)1270 BasicAAResult::aliasGEP(const GEPOperator *GEP1, LocationSize V1Size,
1271 const AAMDNodes &V1AAInfo, const Value *V2,
1272 LocationSize V2Size, const AAMDNodes &V2AAInfo,
1273 const Value *UnderlyingV1, const Value *UnderlyingV2) {
1274 DecomposedGEP DecompGEP1, DecompGEP2;
1275 unsigned MaxPointerSize = getMaxPointerSize(DL);
1276 DecompGEP1.StructOffset = DecompGEP1.OtherOffset = APInt(MaxPointerSize, 0);
1277 DecompGEP2.StructOffset = DecompGEP2.OtherOffset = APInt(MaxPointerSize, 0);
1278
1279 bool GEP1MaxLookupReached =
1280 DecomposeGEPExpression(GEP1, DecompGEP1, DL, &AC, DT);
1281 bool GEP2MaxLookupReached =
1282 DecomposeGEPExpression(V2, DecompGEP2, DL, &AC, DT);
1283
1284 APInt GEP1BaseOffset = DecompGEP1.StructOffset + DecompGEP1.OtherOffset;
1285 APInt GEP2BaseOffset = DecompGEP2.StructOffset + DecompGEP2.OtherOffset;
1286
1287 assert(DecompGEP1.Base == UnderlyingV1 && DecompGEP2.Base == UnderlyingV2 &&
1288 "DecomposeGEPExpression returned a result different from "
1289 "GetUnderlyingObject");
1290
1291 // If the GEP's offset relative to its base is such that the base would
1292 // fall below the start of the object underlying V2, then the GEP and V2
1293 // cannot alias.
1294 if (!GEP1MaxLookupReached && !GEP2MaxLookupReached &&
1295 isGEPBaseAtNegativeOffset(GEP1, DecompGEP1, DecompGEP2, V2Size))
1296 return NoAlias;
1297 // If we have two gep instructions with must-alias or not-alias'ing base
1298 // pointers, figure out if the indexes to the GEP tell us anything about the
1299 // derived pointer.
1300 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
1301 // Check for the GEP base being at a negative offset, this time in the other
1302 // direction.
1303 if (!GEP1MaxLookupReached && !GEP2MaxLookupReached &&
1304 isGEPBaseAtNegativeOffset(GEP2, DecompGEP2, DecompGEP1, V1Size))
1305 return NoAlias;
1306 // Do the base pointers alias?
1307 AliasResult BaseAlias =
1308 aliasCheck(UnderlyingV1, LocationSize::unknown(), AAMDNodes(),
1309 UnderlyingV2, LocationSize::unknown(), AAMDNodes());
1310
1311 // Check for geps of non-aliasing underlying pointers where the offsets are
1312 // identical.
1313 if ((BaseAlias == MayAlias) && V1Size == V2Size) {
1314 // Do the base pointers alias assuming type and size.
1315 AliasResult PreciseBaseAlias = aliasCheck(UnderlyingV1, V1Size, V1AAInfo,
1316 UnderlyingV2, V2Size, V2AAInfo);
1317 if (PreciseBaseAlias == NoAlias) {
1318 // See if the computed offset from the common pointer tells us about the
1319 // relation of the resulting pointer.
1320 // If the max search depth is reached the result is undefined
1321 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
1322 return MayAlias;
1323
1324 // Same offsets.
1325 if (GEP1BaseOffset == GEP2BaseOffset &&
1326 DecompGEP1.VarIndices == DecompGEP2.VarIndices)
1327 return NoAlias;
1328 }
1329 }
1330
1331 // If we get a No or May, then return it immediately, no amount of analysis
1332 // will improve this situation.
1333 if (BaseAlias != MustAlias) {
1334 assert(BaseAlias == NoAlias || BaseAlias == MayAlias);
1335 return BaseAlias;
1336 }
1337
1338 // Otherwise, we have a MustAlias. Since the base pointers alias each other
1339 // exactly, see if the computed offset from the common pointer tells us
1340 // about the relation of the resulting pointer.
1341 // If we know the two GEPs are based off of the exact same pointer (and not
1342 // just the same underlying object), see if that tells us anything about
1343 // the resulting pointers.
1344 if (GEP1->getPointerOperand()->stripPointerCastsAndInvariantGroups() ==
1345 GEP2->getPointerOperand()->stripPointerCastsAndInvariantGroups() &&
1346 GEP1->getPointerOperandType() == GEP2->getPointerOperandType()) {
1347 AliasResult R = aliasSameBasePointerGEPs(GEP1, V1Size, GEP2, V2Size, DL);
1348 // If we couldn't find anything interesting, don't abandon just yet.
1349 if (R != MayAlias)
1350 return R;
1351 }
1352
1353 // If the max search depth is reached, the result is undefined
1354 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
1355 return MayAlias;
1356
1357 // Subtract the GEP2 pointer from the GEP1 pointer to find out their
1358 // symbolic difference.
1359 GEP1BaseOffset -= GEP2BaseOffset;
1360 GetIndexDifference(DecompGEP1.VarIndices, DecompGEP2.VarIndices);
1361
1362 } else {
1363 // Check to see if these two pointers are related by the getelementptr
1364 // instruction. If one pointer is a GEP with a non-zero index of the other
1365 // pointer, we know they cannot alias.
1366
1367 // If both accesses are unknown size, we can't do anything useful here.
1368 if (V1Size == LocationSize::unknown() && V2Size == LocationSize::unknown())
1369 return MayAlias;
1370
1371 AliasResult R =
1372 aliasCheck(UnderlyingV1, LocationSize::unknown(), AAMDNodes(), V2,
1373 LocationSize::unknown(), V2AAInfo, nullptr, UnderlyingV2);
1374 if (R != MustAlias) {
1375 // If V2 may alias GEP base pointer, conservatively returns MayAlias.
1376 // If V2 is known not to alias GEP base pointer, then the two values
1377 // cannot alias per GEP semantics: "Any memory access must be done through
1378 // a pointer value associated with an address range of the memory access,
1379 // otherwise the behavior is undefined.".
1380 assert(R == NoAlias || R == MayAlias);
1381 return R;
1382 }
1383
1384 // If the max search depth is reached the result is undefined
1385 if (GEP1MaxLookupReached)
1386 return MayAlias;
1387 }
1388
1389 // In the two GEP Case, if there is no difference in the offsets of the
1390 // computed pointers, the resultant pointers are a must alias. This
1391 // happens when we have two lexically identical GEP's (for example).
1392 //
1393 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
1394 // must aliases the GEP, the end result is a must alias also.
1395 if (GEP1BaseOffset == 0 && DecompGEP1.VarIndices.empty())
1396 return MustAlias;
1397
1398 // If there is a constant difference between the pointers, but the difference
1399 // is less than the size of the associated memory object, then we know
1400 // that the objects are partially overlapping. If the difference is
1401 // greater, we know they do not overlap.
1402 if (GEP1BaseOffset != 0 && DecompGEP1.VarIndices.empty()) {
1403 if (GEP1BaseOffset.sge(0)) {
1404 if (V2Size != LocationSize::unknown()) {
1405 if (GEP1BaseOffset.ult(V2Size.getValue()))
1406 return PartialAlias;
1407 return NoAlias;
1408 }
1409 } else {
1410 // We have the situation where:
1411 // + +
1412 // | BaseOffset |
1413 // ---------------->|
1414 // |-->V1Size |-------> V2Size
1415 // GEP1 V2
1416 // We need to know that V2Size is not unknown, otherwise we might have
1417 // stripped a gep with negative index ('gep <ptr>, -1, ...).
1418 if (V1Size != LocationSize::unknown() &&
1419 V2Size != LocationSize::unknown()) {
1420 if ((-GEP1BaseOffset).ult(V1Size.getValue()))
1421 return PartialAlias;
1422 return NoAlias;
1423 }
1424 }
1425 }
1426
1427 if (!DecompGEP1.VarIndices.empty()) {
1428 APInt Modulo(MaxPointerSize, 0);
1429 bool AllPositive = true;
1430 for (unsigned i = 0, e = DecompGEP1.VarIndices.size(); i != e; ++i) {
1431
1432 // Try to distinguish something like &A[i][1] against &A[42][0].
1433 // Grab the least significant bit set in any of the scales. We
1434 // don't need std::abs here (even if the scale's negative) as we'll
1435 // be ^'ing Modulo with itself later.
1436 Modulo |= DecompGEP1.VarIndices[i].Scale;
1437
1438 if (AllPositive) {
1439 // If the Value could change between cycles, then any reasoning about
1440 // the Value this cycle may not hold in the next cycle. We'll just
1441 // give up if we can't determine conditions that hold for every cycle:
1442 const Value *V = DecompGEP1.VarIndices[i].V;
1443
1444 KnownBits Known = computeKnownBits(V, DL, 0, &AC, nullptr, DT);
1445 bool SignKnownZero = Known.isNonNegative();
1446 bool SignKnownOne = Known.isNegative();
1447
1448 // Zero-extension widens the variable, and so forces the sign
1449 // bit to zero.
1450 bool IsZExt = DecompGEP1.VarIndices[i].ZExtBits > 0 || isa<ZExtInst>(V);
1451 SignKnownZero |= IsZExt;
1452 SignKnownOne &= !IsZExt;
1453
1454 // If the variable begins with a zero then we know it's
1455 // positive, regardless of whether the value is signed or
1456 // unsigned.
1457 APInt Scale = DecompGEP1.VarIndices[i].Scale;
1458 AllPositive =
1459 (SignKnownZero && Scale.sge(0)) || (SignKnownOne && Scale.slt(0));
1460 }
1461 }
1462
1463 Modulo = Modulo ^ (Modulo & (Modulo - 1));
1464
1465 // We can compute the difference between the two addresses
1466 // mod Modulo. Check whether that difference guarantees that the
1467 // two locations do not alias.
1468 APInt ModOffset = GEP1BaseOffset & (Modulo - 1);
1469 if (V1Size != LocationSize::unknown() &&
1470 V2Size != LocationSize::unknown() && ModOffset.uge(V2Size.getValue()) &&
1471 (Modulo - ModOffset).uge(V1Size.getValue()))
1472 return NoAlias;
1473
1474 // If we know all the variables are positive, then GEP1 >= GEP1BasePtr.
1475 // If GEP1BasePtr > V2 (GEP1BaseOffset > 0) then we know the pointers
1476 // don't alias if V2Size can fit in the gap between V2 and GEP1BasePtr.
1477 if (AllPositive && GEP1BaseOffset.sgt(0) &&
1478 V2Size != LocationSize::unknown() &&
1479 GEP1BaseOffset.uge(V2Size.getValue()))
1480 return NoAlias;
1481
1482 if (constantOffsetHeuristic(DecompGEP1.VarIndices, V1Size, V2Size,
1483 GEP1BaseOffset, &AC, DT))
1484 return NoAlias;
1485 }
1486
1487 // Statically, we can see that the base objects are the same, but the
1488 // pointers have dynamic offsets which we can't resolve. And none of our
1489 // little tricks above worked.
1490 return MayAlias;
1491 }
1492
MergeAliasResults(AliasResult A,AliasResult B)1493 static AliasResult MergeAliasResults(AliasResult A, AliasResult B) {
1494 // If the results agree, take it.
1495 if (A == B)
1496 return A;
1497 // A mix of PartialAlias and MustAlias is PartialAlias.
1498 if ((A == PartialAlias && B == MustAlias) ||
1499 (B == PartialAlias && A == MustAlias))
1500 return PartialAlias;
1501 // Otherwise, we don't know anything.
1502 return MayAlias;
1503 }
1504
1505 /// Provides a bunch of ad-hoc rules to disambiguate a Select instruction
1506 /// against another.
aliasSelect(const SelectInst * SI,LocationSize SISize,const AAMDNodes & SIAAInfo,const Value * V2,LocationSize V2Size,const AAMDNodes & V2AAInfo,const Value * UnderV2)1507 AliasResult BasicAAResult::aliasSelect(const SelectInst *SI,
1508 LocationSize SISize,
1509 const AAMDNodes &SIAAInfo,
1510 const Value *V2, LocationSize V2Size,
1511 const AAMDNodes &V2AAInfo,
1512 const Value *UnderV2) {
1513 // If the values are Selects with the same condition, we can do a more precise
1514 // check: just check for aliases between the values on corresponding arms.
1515 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1516 if (SI->getCondition() == SI2->getCondition()) {
1517 AliasResult Alias = aliasCheck(SI->getTrueValue(), SISize, SIAAInfo,
1518 SI2->getTrueValue(), V2Size, V2AAInfo);
1519 if (Alias == MayAlias)
1520 return MayAlias;
1521 AliasResult ThisAlias =
1522 aliasCheck(SI->getFalseValue(), SISize, SIAAInfo,
1523 SI2->getFalseValue(), V2Size, V2AAInfo);
1524 return MergeAliasResults(ThisAlias, Alias);
1525 }
1526
1527 // If both arms of the Select node NoAlias or MustAlias V2, then returns
1528 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1529 AliasResult Alias =
1530 aliasCheck(V2, V2Size, V2AAInfo, SI->getTrueValue(),
1531 SISize, SIAAInfo, UnderV2);
1532 if (Alias == MayAlias)
1533 return MayAlias;
1534
1535 AliasResult ThisAlias =
1536 aliasCheck(V2, V2Size, V2AAInfo, SI->getFalseValue(), SISize, SIAAInfo,
1537 UnderV2);
1538 return MergeAliasResults(ThisAlias, Alias);
1539 }
1540
1541 /// Provide a bunch of ad-hoc rules to disambiguate a PHI instruction against
1542 /// another.
aliasPHI(const PHINode * PN,LocationSize PNSize,const AAMDNodes & PNAAInfo,const Value * V2,LocationSize V2Size,const AAMDNodes & V2AAInfo,const Value * UnderV2)1543 AliasResult BasicAAResult::aliasPHI(const PHINode *PN, LocationSize PNSize,
1544 const AAMDNodes &PNAAInfo, const Value *V2,
1545 LocationSize V2Size,
1546 const AAMDNodes &V2AAInfo,
1547 const Value *UnderV2) {
1548 // Track phi nodes we have visited. We use this information when we determine
1549 // value equivalence.
1550 VisitedPhiBBs.insert(PN->getParent());
1551
1552 // If the values are PHIs in the same block, we can do a more precise
1553 // as well as efficient check: just check for aliases between the values
1554 // on corresponding edges.
1555 if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1556 if (PN2->getParent() == PN->getParent()) {
1557 LocPair Locs(MemoryLocation(PN, PNSize, PNAAInfo),
1558 MemoryLocation(V2, V2Size, V2AAInfo));
1559 if (PN > V2)
1560 std::swap(Locs.first, Locs.second);
1561 // Analyse the PHIs' inputs under the assumption that the PHIs are
1562 // NoAlias.
1563 // If the PHIs are May/MustAlias there must be (recursively) an input
1564 // operand from outside the PHIs' cycle that is MayAlias/MustAlias or
1565 // there must be an operation on the PHIs within the PHIs' value cycle
1566 // that causes a MayAlias.
1567 // Pretend the phis do not alias.
1568 AliasResult Alias = NoAlias;
1569 assert(AliasCache.count(Locs) &&
1570 "There must exist an entry for the phi node");
1571 AliasResult OrigAliasResult = AliasCache[Locs];
1572 AliasCache[Locs] = NoAlias;
1573
1574 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1575 AliasResult ThisAlias =
1576 aliasCheck(PN->getIncomingValue(i), PNSize, PNAAInfo,
1577 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1578 V2Size, V2AAInfo);
1579 Alias = MergeAliasResults(ThisAlias, Alias);
1580 if (Alias == MayAlias)
1581 break;
1582 }
1583
1584 // Reset if speculation failed.
1585 if (Alias != NoAlias)
1586 AliasCache[Locs] = OrigAliasResult;
1587
1588 return Alias;
1589 }
1590
1591 SmallVector<Value *, 4> V1Srcs;
1592 bool isRecursive = false;
1593 if (PV) {
1594 // If we have PhiValues then use it to get the underlying phi values.
1595 const PhiValues::ValueSet &PhiValueSet = PV->getValuesForPhi(PN);
1596 // If we have more phi values than the search depth then return MayAlias
1597 // conservatively to avoid compile time explosion. The worst possible case
1598 // is if both sides are PHI nodes. In which case, this is O(m x n) time
1599 // where 'm' and 'n' are the number of PHI sources.
1600 if (PhiValueSet.size() > MaxLookupSearchDepth)
1601 return MayAlias;
1602 // Add the values to V1Srcs
1603 for (Value *PV1 : PhiValueSet) {
1604 if (EnableRecPhiAnalysis) {
1605 if (GEPOperator *PV1GEP = dyn_cast<GEPOperator>(PV1)) {
1606 // Check whether the incoming value is a GEP that advances the pointer
1607 // result of this PHI node (e.g. in a loop). If this is the case, we
1608 // would recurse and always get a MayAlias. Handle this case specially
1609 // below.
1610 if (PV1GEP->getPointerOperand() == PN && PV1GEP->getNumIndices() == 1 &&
1611 isa<ConstantInt>(PV1GEP->idx_begin())) {
1612 isRecursive = true;
1613 continue;
1614 }
1615 }
1616 }
1617 V1Srcs.push_back(PV1);
1618 }
1619 } else {
1620 // If we don't have PhiInfo then just look at the operands of the phi itself
1621 // FIXME: Remove this once we can guarantee that we have PhiInfo always
1622 SmallPtrSet<Value *, 4> UniqueSrc;
1623 for (Value *PV1 : PN->incoming_values()) {
1624 if (isa<PHINode>(PV1))
1625 // If any of the source itself is a PHI, return MayAlias conservatively
1626 // to avoid compile time explosion. The worst possible case is if both
1627 // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1628 // and 'n' are the number of PHI sources.
1629 return MayAlias;
1630
1631 if (EnableRecPhiAnalysis)
1632 if (GEPOperator *PV1GEP = dyn_cast<GEPOperator>(PV1)) {
1633 // Check whether the incoming value is a GEP that advances the pointer
1634 // result of this PHI node (e.g. in a loop). If this is the case, we
1635 // would recurse and always get a MayAlias. Handle this case specially
1636 // below.
1637 if (PV1GEP->getPointerOperand() == PN && PV1GEP->getNumIndices() == 1 &&
1638 isa<ConstantInt>(PV1GEP->idx_begin())) {
1639 isRecursive = true;
1640 continue;
1641 }
1642 }
1643
1644 if (UniqueSrc.insert(PV1).second)
1645 V1Srcs.push_back(PV1);
1646 }
1647 }
1648
1649 // If V1Srcs is empty then that means that the phi has no underlying non-phi
1650 // value. This should only be possible in blocks unreachable from the entry
1651 // block, but return MayAlias just in case.
1652 if (V1Srcs.empty())
1653 return MayAlias;
1654
1655 // If this PHI node is recursive, set the size of the accessed memory to
1656 // unknown to represent all the possible values the GEP could advance the
1657 // pointer to.
1658 if (isRecursive)
1659 PNSize = LocationSize::unknown();
1660
1661 AliasResult Alias =
1662 aliasCheck(V2, V2Size, V2AAInfo, V1Srcs[0],
1663 PNSize, PNAAInfo, UnderV2);
1664
1665 // Early exit if the check of the first PHI source against V2 is MayAlias.
1666 // Other results are not possible.
1667 if (Alias == MayAlias)
1668 return MayAlias;
1669
1670 // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1671 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1672 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1673 Value *V = V1Srcs[i];
1674
1675 AliasResult ThisAlias =
1676 aliasCheck(V2, V2Size, V2AAInfo, V, PNSize, PNAAInfo, UnderV2);
1677 Alias = MergeAliasResults(ThisAlias, Alias);
1678 if (Alias == MayAlias)
1679 break;
1680 }
1681
1682 return Alias;
1683 }
1684
1685 /// Provides a bunch of ad-hoc rules to disambiguate in common cases, such as
1686 /// array references.
aliasCheck(const Value * V1,LocationSize V1Size,AAMDNodes V1AAInfo,const Value * V2,LocationSize V2Size,AAMDNodes V2AAInfo,const Value * O1,const Value * O2)1687 AliasResult BasicAAResult::aliasCheck(const Value *V1, LocationSize V1Size,
1688 AAMDNodes V1AAInfo, const Value *V2,
1689 LocationSize V2Size, AAMDNodes V2AAInfo,
1690 const Value *O1, const Value *O2) {
1691 // If either of the memory references is empty, it doesn't matter what the
1692 // pointer values are.
1693 if (V1Size.isZero() || V2Size.isZero())
1694 return NoAlias;
1695
1696 // Strip off any casts if they exist.
1697 V1 = V1->stripPointerCastsAndInvariantGroups();
1698 V2 = V2->stripPointerCastsAndInvariantGroups();
1699
1700 // If V1 or V2 is undef, the result is NoAlias because we can always pick a
1701 // value for undef that aliases nothing in the program.
1702 if (isa<UndefValue>(V1) || isa<UndefValue>(V2))
1703 return NoAlias;
1704
1705 // Are we checking for alias of the same value?
1706 // Because we look 'through' phi nodes, we could look at "Value" pointers from
1707 // different iterations. We must therefore make sure that this is not the
1708 // case. The function isValueEqualInPotentialCycles ensures that this cannot
1709 // happen by looking at the visited phi nodes and making sure they cannot
1710 // reach the value.
1711 if (isValueEqualInPotentialCycles(V1, V2))
1712 return MustAlias;
1713
1714 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1715 return NoAlias; // Scalars cannot alias each other
1716
1717 // Figure out what objects these things are pointing to if we can.
1718 if (O1 == nullptr)
1719 O1 = GetUnderlyingObject(V1, DL, MaxLookupSearchDepth);
1720
1721 if (O2 == nullptr)
1722 O2 = GetUnderlyingObject(V2, DL, MaxLookupSearchDepth);
1723
1724 // Null values in the default address space don't point to any object, so they
1725 // don't alias any other pointer.
1726 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1727 if (!NullPointerIsDefined(&F, CPN->getType()->getAddressSpace()))
1728 return NoAlias;
1729 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1730 if (!NullPointerIsDefined(&F, CPN->getType()->getAddressSpace()))
1731 return NoAlias;
1732
1733 if (O1 != O2) {
1734 // If V1/V2 point to two different objects, we know that we have no alias.
1735 if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1736 return NoAlias;
1737
1738 // Constant pointers can't alias with non-const isIdentifiedObject objects.
1739 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1740 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1741 return NoAlias;
1742
1743 // Function arguments can't alias with things that are known to be
1744 // unambigously identified at the function level.
1745 if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
1746 (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
1747 return NoAlias;
1748
1749 // If one pointer is the result of a call/invoke or load and the other is a
1750 // non-escaping local object within the same function, then we know the
1751 // object couldn't escape to a point where the call could return it.
1752 //
1753 // Note that if the pointers are in different functions, there are a
1754 // variety of complications. A call with a nocapture argument may still
1755 // temporary store the nocapture argument's value in a temporary memory
1756 // location if that memory location doesn't escape. Or it may pass a
1757 // nocapture value to other functions as long as they don't capture it.
1758 if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1759 return NoAlias;
1760 if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1761 return NoAlias;
1762 }
1763
1764 // If the size of one access is larger than the entire object on the other
1765 // side, then we know such behavior is undefined and can assume no alias.
1766 bool NullIsValidLocation = NullPointerIsDefined(&F);
1767 if ((V1Size.isPrecise() && isObjectSmallerThan(O2, V1Size.getValue(), DL, TLI,
1768 NullIsValidLocation)) ||
1769 (V2Size.isPrecise() && isObjectSmallerThan(O1, V2Size.getValue(), DL, TLI,
1770 NullIsValidLocation)))
1771 return NoAlias;
1772
1773 // Check the cache before climbing up use-def chains. This also terminates
1774 // otherwise infinitely recursive queries.
1775 LocPair Locs(MemoryLocation(V1, V1Size, V1AAInfo),
1776 MemoryLocation(V2, V2Size, V2AAInfo));
1777 if (V1 > V2)
1778 std::swap(Locs.first, Locs.second);
1779 std::pair<AliasCacheTy::iterator, bool> Pair =
1780 AliasCache.insert(std::make_pair(Locs, MayAlias));
1781 if (!Pair.second)
1782 return Pair.first->second;
1783
1784 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1785 // GEP can't simplify, we don't even look at the PHI cases.
1786 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1787 std::swap(V1, V2);
1788 std::swap(V1Size, V2Size);
1789 std::swap(O1, O2);
1790 std::swap(V1AAInfo, V2AAInfo);
1791 }
1792 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1793 AliasResult Result =
1794 aliasGEP(GV1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O1, O2);
1795 if (Result != MayAlias)
1796 return AliasCache[Locs] = Result;
1797 }
1798
1799 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1800 std::swap(V1, V2);
1801 std::swap(O1, O2);
1802 std::swap(V1Size, V2Size);
1803 std::swap(V1AAInfo, V2AAInfo);
1804 }
1805 if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1806 AliasResult Result = aliasPHI(PN, V1Size, V1AAInfo,
1807 V2, V2Size, V2AAInfo, O2);
1808 if (Result != MayAlias)
1809 return AliasCache[Locs] = Result;
1810 }
1811
1812 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1813 std::swap(V1, V2);
1814 std::swap(O1, O2);
1815 std::swap(V1Size, V2Size);
1816 std::swap(V1AAInfo, V2AAInfo);
1817 }
1818 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1819 AliasResult Result =
1820 aliasSelect(S1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O2);
1821 if (Result != MayAlias)
1822 return AliasCache[Locs] = Result;
1823 }
1824
1825 // If both pointers are pointing into the same object and one of them
1826 // accesses the entire object, then the accesses must overlap in some way.
1827 if (O1 == O2)
1828 if (V1Size.isPrecise() && V2Size.isPrecise() &&
1829 (isObjectSize(O1, V1Size.getValue(), DL, TLI, NullIsValidLocation) ||
1830 isObjectSize(O2, V2Size.getValue(), DL, TLI, NullIsValidLocation)))
1831 return AliasCache[Locs] = PartialAlias;
1832
1833 // Recurse back into the best AA results we have, potentially with refined
1834 // memory locations. We have already ensured that BasicAA has a MayAlias
1835 // cache result for these, so any recursion back into BasicAA won't loop.
1836 AliasResult Result = getBestAAResults().alias(Locs.first, Locs.second);
1837 return AliasCache[Locs] = Result;
1838 }
1839
1840 /// Check whether two Values can be considered equivalent.
1841 ///
1842 /// In addition to pointer equivalence of \p V1 and \p V2 this checks whether
1843 /// they can not be part of a cycle in the value graph by looking at all
1844 /// visited phi nodes an making sure that the phis cannot reach the value. We
1845 /// have to do this because we are looking through phi nodes (That is we say
1846 /// noalias(V, phi(VA, VB)) if noalias(V, VA) and noalias(V, VB).
isValueEqualInPotentialCycles(const Value * V,const Value * V2)1847 bool BasicAAResult::isValueEqualInPotentialCycles(const Value *V,
1848 const Value *V2) {
1849 if (V != V2)
1850 return false;
1851
1852 const Instruction *Inst = dyn_cast<Instruction>(V);
1853 if (!Inst)
1854 return true;
1855
1856 if (VisitedPhiBBs.empty())
1857 return true;
1858
1859 if (VisitedPhiBBs.size() > MaxNumPhiBBsValueReachabilityCheck)
1860 return false;
1861
1862 // Make sure that the visited phis cannot reach the Value. This ensures that
1863 // the Values cannot come from different iterations of a potential cycle the
1864 // phi nodes could be involved in.
1865 for (auto *P : VisitedPhiBBs)
1866 if (isPotentiallyReachable(&P->front(), Inst, DT, LI))
1867 return false;
1868
1869 return true;
1870 }
1871
1872 /// Computes the symbolic difference between two de-composed GEPs.
1873 ///
1874 /// Dest and Src are the variable indices from two decomposed GetElementPtr
1875 /// instructions GEP1 and GEP2 which have common base pointers.
GetIndexDifference(SmallVectorImpl<VariableGEPIndex> & Dest,const SmallVectorImpl<VariableGEPIndex> & Src)1876 void BasicAAResult::GetIndexDifference(
1877 SmallVectorImpl<VariableGEPIndex> &Dest,
1878 const SmallVectorImpl<VariableGEPIndex> &Src) {
1879 if (Src.empty())
1880 return;
1881
1882 for (unsigned i = 0, e = Src.size(); i != e; ++i) {
1883 const Value *V = Src[i].V;
1884 unsigned ZExtBits = Src[i].ZExtBits, SExtBits = Src[i].SExtBits;
1885 APInt Scale = Src[i].Scale;
1886
1887 // Find V in Dest. This is N^2, but pointer indices almost never have more
1888 // than a few variable indexes.
1889 for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
1890 if (!isValueEqualInPotentialCycles(Dest[j].V, V) ||
1891 Dest[j].ZExtBits != ZExtBits || Dest[j].SExtBits != SExtBits)
1892 continue;
1893
1894 // If we found it, subtract off Scale V's from the entry in Dest. If it
1895 // goes to zero, remove the entry.
1896 if (Dest[j].Scale != Scale)
1897 Dest[j].Scale -= Scale;
1898 else
1899 Dest.erase(Dest.begin() + j);
1900 Scale = 0;
1901 break;
1902 }
1903
1904 // If we didn't consume this entry, add it to the end of the Dest list.
1905 if (!!Scale) {
1906 VariableGEPIndex Entry = {V, ZExtBits, SExtBits, -Scale};
1907 Dest.push_back(Entry);
1908 }
1909 }
1910 }
1911
constantOffsetHeuristic(const SmallVectorImpl<VariableGEPIndex> & VarIndices,LocationSize MaybeV1Size,LocationSize MaybeV2Size,APInt BaseOffset,AssumptionCache * AC,DominatorTree * DT)1912 bool BasicAAResult::constantOffsetHeuristic(
1913 const SmallVectorImpl<VariableGEPIndex> &VarIndices,
1914 LocationSize MaybeV1Size, LocationSize MaybeV2Size, APInt BaseOffset,
1915 AssumptionCache *AC, DominatorTree *DT) {
1916 if (VarIndices.size() != 2 || MaybeV1Size == LocationSize::unknown() ||
1917 MaybeV2Size == LocationSize::unknown())
1918 return false;
1919
1920 const uint64_t V1Size = MaybeV1Size.getValue();
1921 const uint64_t V2Size = MaybeV2Size.getValue();
1922
1923 const VariableGEPIndex &Var0 = VarIndices[0], &Var1 = VarIndices[1];
1924
1925 if (Var0.ZExtBits != Var1.ZExtBits || Var0.SExtBits != Var1.SExtBits ||
1926 Var0.Scale != -Var1.Scale)
1927 return false;
1928
1929 unsigned Width = Var1.V->getType()->getIntegerBitWidth();
1930
1931 // We'll strip off the Extensions of Var0 and Var1 and do another round
1932 // of GetLinearExpression decomposition. In the example above, if Var0
1933 // is zext(%x + 1) we should get V1 == %x and V1Offset == 1.
1934
1935 APInt V0Scale(Width, 0), V0Offset(Width, 0), V1Scale(Width, 0),
1936 V1Offset(Width, 0);
1937 bool NSW = true, NUW = true;
1938 unsigned V0ZExtBits = 0, V0SExtBits = 0, V1ZExtBits = 0, V1SExtBits = 0;
1939 const Value *V0 = GetLinearExpression(Var0.V, V0Scale, V0Offset, V0ZExtBits,
1940 V0SExtBits, DL, 0, AC, DT, NSW, NUW);
1941 NSW = true;
1942 NUW = true;
1943 const Value *V1 = GetLinearExpression(Var1.V, V1Scale, V1Offset, V1ZExtBits,
1944 V1SExtBits, DL, 0, AC, DT, NSW, NUW);
1945
1946 if (V0Scale != V1Scale || V0ZExtBits != V1ZExtBits ||
1947 V0SExtBits != V1SExtBits || !isValueEqualInPotentialCycles(V0, V1))
1948 return false;
1949
1950 // We have a hit - Var0 and Var1 only differ by a constant offset!
1951
1952 // If we've been sext'ed then zext'd the maximum difference between Var0 and
1953 // Var1 is possible to calculate, but we're just interested in the absolute
1954 // minimum difference between the two. The minimum distance may occur due to
1955 // wrapping; consider "add i3 %i, 5": if %i == 7 then 7 + 5 mod 8 == 4, and so
1956 // the minimum distance between %i and %i + 5 is 3.
1957 APInt MinDiff = V0Offset - V1Offset, Wrapped = -MinDiff;
1958 MinDiff = APIntOps::umin(MinDiff, Wrapped);
1959 APInt MinDiffBytes =
1960 MinDiff.zextOrTrunc(Var0.Scale.getBitWidth()) * Var0.Scale.abs();
1961
1962 // We can't definitely say whether GEP1 is before or after V2 due to wrapping
1963 // arithmetic (i.e. for some values of GEP1 and V2 GEP1 < V2, and for other
1964 // values GEP1 > V2). We'll therefore only declare NoAlias if both V1Size and
1965 // V2Size can fit in the MinDiffBytes gap.
1966 return MinDiffBytes.uge(V1Size + BaseOffset.abs()) &&
1967 MinDiffBytes.uge(V2Size + BaseOffset.abs());
1968 }
1969
1970 //===----------------------------------------------------------------------===//
1971 // BasicAliasAnalysis Pass
1972 //===----------------------------------------------------------------------===//
1973
1974 AnalysisKey BasicAA::Key;
1975
run(Function & F,FunctionAnalysisManager & AM)1976 BasicAAResult BasicAA::run(Function &F, FunctionAnalysisManager &AM) {
1977 return BasicAAResult(F.getParent()->getDataLayout(),
1978 F,
1979 AM.getResult<TargetLibraryAnalysis>(F),
1980 AM.getResult<AssumptionAnalysis>(F),
1981 &AM.getResult<DominatorTreeAnalysis>(F),
1982 AM.getCachedResult<LoopAnalysis>(F),
1983 AM.getCachedResult<PhiValuesAnalysis>(F));
1984 }
1985
BasicAAWrapperPass()1986 BasicAAWrapperPass::BasicAAWrapperPass() : FunctionPass(ID) {
1987 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
1988 }
1989
1990 char BasicAAWrapperPass::ID = 0;
1991
anchor()1992 void BasicAAWrapperPass::anchor() {}
1993
1994 INITIALIZE_PASS_BEGIN(BasicAAWrapperPass, "basicaa",
1995 "Basic Alias Analysis (stateless AA impl)", false, true)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)1996 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1997 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1998 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1999 INITIALIZE_PASS_END(BasicAAWrapperPass, "basicaa",
2000 "Basic Alias Analysis (stateless AA impl)", false, true)
2001
2002 FunctionPass *llvm::createBasicAAWrapperPass() {
2003 return new BasicAAWrapperPass();
2004 }
2005
runOnFunction(Function & F)2006 bool BasicAAWrapperPass::runOnFunction(Function &F) {
2007 auto &ACT = getAnalysis<AssumptionCacheTracker>();
2008 auto &TLIWP = getAnalysis<TargetLibraryInfoWrapperPass>();
2009 auto &DTWP = getAnalysis<DominatorTreeWrapperPass>();
2010 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2011 auto *PVWP = getAnalysisIfAvailable<PhiValuesWrapperPass>();
2012
2013 Result.reset(new BasicAAResult(F.getParent()->getDataLayout(), F, TLIWP.getTLI(),
2014 ACT.getAssumptionCache(F), &DTWP.getDomTree(),
2015 LIWP ? &LIWP->getLoopInfo() : nullptr,
2016 PVWP ? &PVWP->getResult() : nullptr));
2017
2018 return false;
2019 }
2020
getAnalysisUsage(AnalysisUsage & AU) const2021 void BasicAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
2022 AU.setPreservesAll();
2023 AU.addRequired<AssumptionCacheTracker>();
2024 AU.addRequired<DominatorTreeWrapperPass>();
2025 AU.addRequired<TargetLibraryInfoWrapperPass>();
2026 AU.addUsedIfAvailable<PhiValuesWrapperPass>();
2027 }
2028
createLegacyPMBasicAAResult(Pass & P,Function & F)2029 BasicAAResult llvm::createLegacyPMBasicAAResult(Pass &P, Function &F) {
2030 return BasicAAResult(
2031 F.getParent()->getDataLayout(),
2032 F,
2033 P.getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2034 P.getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
2035 }
2036