1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/MemoryBuiltins.h"
19 #include "llvm/IR/ConstantRange.h"
20 #include "llvm/IR/Constants.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/GetElementPtrTypeIterator.h"
23 #include "llvm/IR/GlobalAlias.h"
24 #include "llvm/IR/GlobalVariable.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/LLVMContext.h"
28 #include "llvm/IR/Metadata.h"
29 #include "llvm/IR/Operator.h"
30 #include "llvm/IR/PatternMatch.h"
31 #include "llvm/Support/MathExtras.h"
34 using namespace llvm::PatternMatch;
36 const unsigned MaxDepth = 6;
38 /// getBitWidth - Returns the bitwidth of the given scalar or pointer type (if
39 /// unknown returns 0). For vector types, returns the element type's bitwidth.
40 static unsigned getBitWidth(Type *Ty, const DataLayout *TD) {
41 if (unsigned BitWidth = Ty->getScalarSizeInBits())
44 return TD ? TD->getPointerTypeSizeInBits(Ty) : 0;
47 static void ComputeMaskedBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
48 APInt &KnownZero, APInt &KnownOne,
49 APInt &KnownZero2, APInt &KnownOne2,
50 const DataLayout *TD, unsigned Depth) {
52 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
53 // We know that the top bits of C-X are clear if X contains less bits
54 // than C (i.e. no wrap-around can happen). For example, 20-X is
55 // positive if we can prove that X is >= 0 and < 16.
56 if (!CLHS->getValue().isNegative()) {
57 unsigned BitWidth = KnownZero.getBitWidth();
58 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
59 // NLZ can't be BitWidth with no sign bit
60 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
61 llvm::ComputeMaskedBits(Op1, KnownZero2, KnownOne2, TD, Depth+1);
63 // If all of the MaskV bits are known to be zero, then we know the
64 // output top bits are zero, because we now know that the output is
66 if ((KnownZero2 & MaskV) == MaskV) {
67 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
68 // Top bits known zero.
69 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
75 unsigned BitWidth = KnownZero.getBitWidth();
77 // If one of the operands has trailing zeros, then the bits that the
78 // other operand has in those bit positions will be preserved in the
79 // result. For an add, this works with either operand. For a subtract,
80 // this only works if the known zeros are in the right operand.
81 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
82 llvm::ComputeMaskedBits(Op0, LHSKnownZero, LHSKnownOne, TD, Depth+1);
83 assert((LHSKnownZero & LHSKnownOne) == 0 &&
84 "Bits known to be one AND zero?");
85 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
87 llvm::ComputeMaskedBits(Op1, KnownZero2, KnownOne2, TD, Depth+1);
88 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
89 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
91 // Determine which operand has more trailing zeros, and use that
92 // many bits from the other operand.
93 if (LHSKnownZeroOut > RHSKnownZeroOut) {
95 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
96 KnownZero |= KnownZero2 & Mask;
97 KnownOne |= KnownOne2 & Mask;
99 // If the known zeros are in the left operand for a subtract,
100 // fall back to the minimum known zeros in both operands.
101 KnownZero |= APInt::getLowBitsSet(BitWidth,
102 std::min(LHSKnownZeroOut,
105 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
106 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
107 KnownZero |= LHSKnownZero & Mask;
108 KnownOne |= LHSKnownOne & Mask;
111 // Are we still trying to solve for the sign bit?
112 if (!KnownZero.isNegative() && !KnownOne.isNegative()) {
115 // Adding two positive numbers can't wrap into negative
116 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
117 KnownZero |= APInt::getSignBit(BitWidth);
118 // and adding two negative numbers can't wrap into positive.
119 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
120 KnownOne |= APInt::getSignBit(BitWidth);
122 // Subtracting a negative number from a positive one can't wrap
123 if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
124 KnownZero |= APInt::getSignBit(BitWidth);
125 // neither can subtracting a positive number from a negative one.
126 else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
127 KnownOne |= APInt::getSignBit(BitWidth);
133 static void ComputeMaskedBitsMul(Value *Op0, Value *Op1, bool NSW,
134 APInt &KnownZero, APInt &KnownOne,
135 APInt &KnownZero2, APInt &KnownOne2,
136 const DataLayout *TD, unsigned Depth) {
137 unsigned BitWidth = KnownZero.getBitWidth();
138 ComputeMaskedBits(Op1, KnownZero, KnownOne, TD, Depth+1);
139 ComputeMaskedBits(Op0, KnownZero2, KnownOne2, TD, Depth+1);
140 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
141 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
143 bool isKnownNegative = false;
144 bool isKnownNonNegative = false;
145 // If the multiplication is known not to overflow, compute the sign bit.
148 // The product of a number with itself is non-negative.
149 isKnownNonNegative = true;
151 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
152 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
153 bool isKnownNegativeOp1 = KnownOne.isNegative();
154 bool isKnownNegativeOp0 = KnownOne2.isNegative();
155 // The product of two numbers with the same sign is non-negative.
156 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
157 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
158 // The product of a negative number and a non-negative number is either
160 if (!isKnownNonNegative)
161 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
162 isKnownNonZero(Op0, TD, Depth)) ||
163 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
164 isKnownNonZero(Op1, TD, Depth));
168 // If low bits are zero in either operand, output low known-0 bits.
169 // Also compute a conserative estimate for high known-0 bits.
170 // More trickiness is possible, but this is sufficient for the
171 // interesting case of alignment computation.
172 KnownOne.clearAllBits();
173 unsigned TrailZ = KnownZero.countTrailingOnes() +
174 KnownZero2.countTrailingOnes();
175 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
176 KnownZero2.countLeadingOnes(),
177 BitWidth) - BitWidth;
179 TrailZ = std::min(TrailZ, BitWidth);
180 LeadZ = std::min(LeadZ, BitWidth);
181 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
182 APInt::getHighBitsSet(BitWidth, LeadZ);
184 // Only make use of no-wrap flags if we failed to compute the sign bit
185 // directly. This matters if the multiplication always overflows, in
186 // which case we prefer to follow the result of the direct computation,
187 // though as the program is invoking undefined behaviour we can choose
188 // whatever we like here.
189 if (isKnownNonNegative && !KnownOne.isNegative())
190 KnownZero.setBit(BitWidth - 1);
191 else if (isKnownNegative && !KnownZero.isNegative())
192 KnownOne.setBit(BitWidth - 1);
195 void llvm::computeMaskedBitsLoad(const MDNode &Ranges, APInt &KnownZero) {
196 unsigned BitWidth = KnownZero.getBitWidth();
197 unsigned NumRanges = Ranges.getNumOperands() / 2;
198 assert(NumRanges >= 1);
200 // Use the high end of the ranges to find leading zeros.
201 unsigned MinLeadingZeros = BitWidth;
202 for (unsigned i = 0; i < NumRanges; ++i) {
203 ConstantInt *Lower = cast<ConstantInt>(Ranges.getOperand(2*i + 0));
204 ConstantInt *Upper = cast<ConstantInt>(Ranges.getOperand(2*i + 1));
205 ConstantRange Range(Lower->getValue(), Upper->getValue());
206 if (Range.isWrappedSet())
207 MinLeadingZeros = 0; // -1 has no zeros
208 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
209 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
212 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
214 /// ComputeMaskedBits - Determine which of the bits are known to be either zero
215 /// or one and return them in the KnownZero/KnownOne bit sets.
217 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
218 /// we cannot optimize based on the assumption that it is zero without changing
219 /// it to be an explicit zero. If we don't change it to zero, other code could
220 /// optimized based on the contradictory assumption that it is non-zero.
221 /// Because instcombine aggressively folds operations with undef args anyway,
222 /// this won't lose us code quality.
224 /// This function is defined on values with integer type, values with pointer
225 /// type (but only if TD is non-null), and vectors of integers. In the case
226 /// where V is a vector, known zero, and known one values are the
227 /// same width as the vector element, and the bit is set only if it is true
228 /// for all of the elements in the vector.
229 void llvm::ComputeMaskedBits(Value *V, APInt &KnownZero, APInt &KnownOne,
230 const DataLayout *TD, unsigned Depth) {
231 assert(V && "No Value?");
232 assert(Depth <= MaxDepth && "Limit Search Depth");
233 unsigned BitWidth = KnownZero.getBitWidth();
235 assert((V->getType()->isIntOrIntVectorTy() ||
236 V->getType()->getScalarType()->isPointerTy()) &&
237 "Not integer or pointer type!");
239 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
240 (!V->getType()->isIntOrIntVectorTy() ||
241 V->getType()->getScalarSizeInBits() == BitWidth) &&
242 KnownZero.getBitWidth() == BitWidth &&
243 KnownOne.getBitWidth() == BitWidth &&
244 "V, Mask, KnownOne and KnownZero should have same BitWidth");
246 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
247 // We know all of the bits for a constant!
248 KnownOne = CI->getValue();
249 KnownZero = ~KnownOne;
252 // Null and aggregate-zero are all-zeros.
253 if (isa<ConstantPointerNull>(V) ||
254 isa<ConstantAggregateZero>(V)) {
255 KnownOne.clearAllBits();
256 KnownZero = APInt::getAllOnesValue(BitWidth);
259 // Handle a constant vector by taking the intersection of the known bits of
260 // each element. There is no real need to handle ConstantVector here, because
261 // we don't handle undef in any particularly useful way.
262 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
263 // We know that CDS must be a vector of integers. Take the intersection of
265 KnownZero.setAllBits(); KnownOne.setAllBits();
266 APInt Elt(KnownZero.getBitWidth(), 0);
267 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
268 Elt = CDS->getElementAsInteger(i);
275 // The address of an aligned GlobalValue has trailing zeros.
276 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
277 unsigned Align = GV->getAlignment();
278 if (Align == 0 && TD) {
279 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
280 Type *ObjectType = GVar->getType()->getElementType();
281 if (ObjectType->isSized()) {
282 // If the object is defined in the current Module, we'll be giving
283 // it the preferred alignment. Otherwise, we have to assume that it
284 // may only have the minimum ABI alignment.
285 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
286 Align = TD->getPreferredAlignment(GVar);
288 Align = TD->getABITypeAlignment(ObjectType);
293 KnownZero = APInt::getLowBitsSet(BitWidth,
294 countTrailingZeros(Align));
296 KnownZero.clearAllBits();
297 KnownOne.clearAllBits();
300 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
301 // the bits of its aliasee.
302 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
303 if (GA->mayBeOverridden()) {
304 KnownZero.clearAllBits(); KnownOne.clearAllBits();
306 ComputeMaskedBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth+1);
311 if (Argument *A = dyn_cast<Argument>(V)) {
314 if (A->hasByValOrInAllocaAttr()) {
315 // Get alignment information off byval/inalloca arguments if specified in
317 Align = A->getParamAlignment();
318 } else if (TD && A->hasStructRetAttr()) {
319 // An sret parameter has at least the ABI alignment of the return type.
320 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
321 if (EltTy->isSized())
322 Align = TD->getABITypeAlignment(EltTy);
326 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
330 // Start out not knowing anything.
331 KnownZero.clearAllBits(); KnownOne.clearAllBits();
333 if (Depth == MaxDepth)
334 return; // Limit search depth.
336 Operator *I = dyn_cast<Operator>(V);
339 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
340 switch (I->getOpcode()) {
342 case Instruction::Load:
343 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
344 computeMaskedBitsLoad(*MD, KnownZero);
346 case Instruction::And: {
347 // If either the LHS or the RHS are Zero, the result is zero.
348 ComputeMaskedBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
349 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
350 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
351 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
353 // Output known-1 bits are only known if set in both the LHS & RHS.
354 KnownOne &= KnownOne2;
355 // Output known-0 are known to be clear if zero in either the LHS | RHS.
356 KnownZero |= KnownZero2;
359 case Instruction::Or: {
360 ComputeMaskedBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
361 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
362 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
363 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
365 // Output known-0 bits are only known if clear in both the LHS & RHS.
366 KnownZero &= KnownZero2;
367 // Output known-1 are known to be set if set in either the LHS | RHS.
368 KnownOne |= KnownOne2;
371 case Instruction::Xor: {
372 ComputeMaskedBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
373 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
374 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
375 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
377 // Output known-0 bits are known if clear or set in both the LHS & RHS.
378 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
379 // Output known-1 are known to be set if set in only one of the LHS, RHS.
380 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
381 KnownZero = KnownZeroOut;
384 case Instruction::Mul: {
385 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
386 ComputeMaskedBitsMul(I->getOperand(0), I->getOperand(1), NSW,
387 KnownZero, KnownOne, KnownZero2, KnownOne2, TD, Depth);
390 case Instruction::UDiv: {
391 // For the purposes of computing leading zeros we can conservatively
392 // treat a udiv as a logical right shift by the power of 2 known to
393 // be less than the denominator.
394 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
395 unsigned LeadZ = KnownZero2.countLeadingOnes();
397 KnownOne2.clearAllBits();
398 KnownZero2.clearAllBits();
399 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
400 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
401 if (RHSUnknownLeadingOnes != BitWidth)
402 LeadZ = std::min(BitWidth,
403 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
405 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
408 case Instruction::Select:
409 ComputeMaskedBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1);
410 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD,
412 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
413 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
415 // Only known if known in both the LHS and RHS.
416 KnownOne &= KnownOne2;
417 KnownZero &= KnownZero2;
419 case Instruction::FPTrunc:
420 case Instruction::FPExt:
421 case Instruction::FPToUI:
422 case Instruction::FPToSI:
423 case Instruction::SIToFP:
424 case Instruction::UIToFP:
425 return; // Can't work with floating point.
426 case Instruction::PtrToInt:
427 case Instruction::IntToPtr:
428 // We can't handle these if we don't know the pointer size.
430 // FALL THROUGH and handle them the same as zext/trunc.
431 case Instruction::ZExt:
432 case Instruction::Trunc: {
433 Type *SrcTy = I->getOperand(0)->getType();
435 unsigned SrcBitWidth;
436 // Note that we handle pointer operands here because of inttoptr/ptrtoint
437 // which fall through here.
439 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
441 SrcBitWidth = SrcTy->getScalarSizeInBits();
442 if (!SrcBitWidth) return;
445 assert(SrcBitWidth && "SrcBitWidth can't be zero");
446 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
447 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
448 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
449 KnownZero = KnownZero.zextOrTrunc(BitWidth);
450 KnownOne = KnownOne.zextOrTrunc(BitWidth);
451 // Any top bits are known to be zero.
452 if (BitWidth > SrcBitWidth)
453 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
456 case Instruction::BitCast: {
457 Type *SrcTy = I->getOperand(0)->getType();
458 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
459 // TODO: For now, not handling conversions like:
460 // (bitcast i64 %x to <2 x i32>)
461 !I->getType()->isVectorTy()) {
462 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
467 case Instruction::SExt: {
468 // Compute the bits in the result that are not present in the input.
469 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
471 KnownZero = KnownZero.trunc(SrcBitWidth);
472 KnownOne = KnownOne.trunc(SrcBitWidth);
473 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
474 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
475 KnownZero = KnownZero.zext(BitWidth);
476 KnownOne = KnownOne.zext(BitWidth);
478 // If the sign bit of the input is known set or clear, then we know the
479 // top bits of the result.
480 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
481 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
482 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
483 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
486 case Instruction::Shl:
487 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
488 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
489 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
490 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
491 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
492 KnownZero <<= ShiftAmt;
493 KnownOne <<= ShiftAmt;
494 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
498 case Instruction::LShr:
499 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
500 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
501 // Compute the new bits that are at the top now.
502 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
504 // Unsigned shift right.
505 ComputeMaskedBits(I->getOperand(0), KnownZero,KnownOne, TD, Depth+1);
506 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
507 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
508 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
509 // high bits known zero.
510 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
514 case Instruction::AShr:
515 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
516 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
517 // Compute the new bits that are at the top now.
518 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
520 // Signed shift right.
521 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
522 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
523 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
524 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
526 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
527 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
528 KnownZero |= HighBits;
529 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
530 KnownOne |= HighBits;
534 case Instruction::Sub: {
535 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
536 ComputeMaskedBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
537 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
541 case Instruction::Add: {
542 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
543 ComputeMaskedBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
544 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
548 case Instruction::SRem:
549 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
550 APInt RA = Rem->getValue().abs();
551 if (RA.isPowerOf2()) {
552 APInt LowBits = RA - 1;
553 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
555 // The low bits of the first operand are unchanged by the srem.
556 KnownZero = KnownZero2 & LowBits;
557 KnownOne = KnownOne2 & LowBits;
559 // If the first operand is non-negative or has all low bits zero, then
560 // the upper bits are all zero.
561 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
562 KnownZero |= ~LowBits;
564 // If the first operand is negative and not all low bits are zero, then
565 // the upper bits are all one.
566 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
567 KnownOne |= ~LowBits;
569 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
573 // The sign bit is the LHS's sign bit, except when the result of the
574 // remainder is zero.
575 if (KnownZero.isNonNegative()) {
576 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
577 ComputeMaskedBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
579 // If it's known zero, our sign bit is also zero.
580 if (LHSKnownZero.isNegative())
581 KnownZero.setBit(BitWidth - 1);
585 case Instruction::URem: {
586 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
587 APInt RA = Rem->getValue();
588 if (RA.isPowerOf2()) {
589 APInt LowBits = (RA - 1);
590 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD,
592 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
593 KnownZero |= ~LowBits;
599 // Since the result is less than or equal to either operand, any leading
600 // zero bits in either operand must also exist in the result.
601 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
602 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
604 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
605 KnownZero2.countLeadingOnes());
606 KnownOne.clearAllBits();
607 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
611 case Instruction::Alloca: {
612 AllocaInst *AI = cast<AllocaInst>(V);
613 unsigned Align = AI->getAlignment();
614 if (Align == 0 && TD)
615 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
618 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
621 case Instruction::GetElementPtr: {
622 // Analyze all of the subscripts of this getelementptr instruction
623 // to determine if we can prove known low zero bits.
624 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
625 ComputeMaskedBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
627 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
629 gep_type_iterator GTI = gep_type_begin(I);
630 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
631 Value *Index = I->getOperand(i);
632 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
633 // Handle struct member offset arithmetic.
637 // Handle case when index is vector zeroinitializer
638 Constant *CIndex = cast<Constant>(Index);
639 if (CIndex->isZeroValue())
642 if (CIndex->getType()->isVectorTy())
643 Index = CIndex->getSplatValue();
645 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
646 const StructLayout *SL = TD->getStructLayout(STy);
647 uint64_t Offset = SL->getElementOffset(Idx);
648 TrailZ = std::min<unsigned>(TrailZ,
649 countTrailingZeros(Offset));
651 // Handle array index arithmetic.
652 Type *IndexedTy = GTI.getIndexedType();
653 if (!IndexedTy->isSized()) return;
654 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
655 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
656 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
657 ComputeMaskedBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1);
658 TrailZ = std::min(TrailZ,
659 unsigned(countTrailingZeros(TypeSize) +
660 LocalKnownZero.countTrailingOnes()));
664 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
667 case Instruction::PHI: {
668 PHINode *P = cast<PHINode>(I);
669 // Handle the case of a simple two-predecessor recurrence PHI.
670 // There's a lot more that could theoretically be done here, but
671 // this is sufficient to catch some interesting cases.
672 if (P->getNumIncomingValues() == 2) {
673 for (unsigned i = 0; i != 2; ++i) {
674 Value *L = P->getIncomingValue(i);
675 Value *R = P->getIncomingValue(!i);
676 Operator *LU = dyn_cast<Operator>(L);
679 unsigned Opcode = LU->getOpcode();
680 // Check for operations that have the property that if
681 // both their operands have low zero bits, the result
682 // will have low zero bits.
683 if (Opcode == Instruction::Add ||
684 Opcode == Instruction::Sub ||
685 Opcode == Instruction::And ||
686 Opcode == Instruction::Or ||
687 Opcode == Instruction::Mul) {
688 Value *LL = LU->getOperand(0);
689 Value *LR = LU->getOperand(1);
690 // Find a recurrence.
697 // Ok, we have a PHI of the form L op= R. Check for low
699 ComputeMaskedBits(R, KnownZero2, KnownOne2, TD, Depth+1);
701 // We need to take the minimum number of known bits
702 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
703 ComputeMaskedBits(L, KnownZero3, KnownOne3, TD, Depth+1);
705 KnownZero = APInt::getLowBitsSet(BitWidth,
706 std::min(KnownZero2.countTrailingOnes(),
707 KnownZero3.countTrailingOnes()));
713 // Unreachable blocks may have zero-operand PHI nodes.
714 if (P->getNumIncomingValues() == 0)
717 // Otherwise take the unions of the known bit sets of the operands,
718 // taking conservative care to avoid excessive recursion.
719 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
720 // Skip if every incoming value references to ourself.
721 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
724 KnownZero = APInt::getAllOnesValue(BitWidth);
725 KnownOne = APInt::getAllOnesValue(BitWidth);
726 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
727 // Skip direct self references.
728 if (P->getIncomingValue(i) == P) continue;
730 KnownZero2 = APInt(BitWidth, 0);
731 KnownOne2 = APInt(BitWidth, 0);
732 // Recurse, but cap the recursion to one level, because we don't
733 // want to waste time spinning around in loops.
734 ComputeMaskedBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
736 KnownZero &= KnownZero2;
737 KnownOne &= KnownOne2;
738 // If all bits have been ruled out, there's no need to check
740 if (!KnownZero && !KnownOne)
746 case Instruction::Call:
747 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
748 switch (II->getIntrinsicID()) {
750 case Intrinsic::ctlz:
751 case Intrinsic::cttz: {
752 unsigned LowBits = Log2_32(BitWidth)+1;
753 // If this call is undefined for 0, the result will be less than 2^n.
754 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
756 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
759 case Intrinsic::ctpop: {
760 unsigned LowBits = Log2_32(BitWidth)+1;
761 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
764 case Intrinsic::x86_sse42_crc32_64_64:
765 KnownZero = APInt::getHighBitsSet(64, 32);
770 case Instruction::ExtractValue:
771 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
772 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
773 if (EVI->getNumIndices() != 1) break;
774 if (EVI->getIndices()[0] == 0) {
775 switch (II->getIntrinsicID()) {
777 case Intrinsic::uadd_with_overflow:
778 case Intrinsic::sadd_with_overflow:
779 ComputeMaskedBitsAddSub(true, II->getArgOperand(0),
780 II->getArgOperand(1), false, KnownZero,
781 KnownOne, KnownZero2, KnownOne2, TD, Depth);
783 case Intrinsic::usub_with_overflow:
784 case Intrinsic::ssub_with_overflow:
785 ComputeMaskedBitsAddSub(false, II->getArgOperand(0),
786 II->getArgOperand(1), false, KnownZero,
787 KnownOne, KnownZero2, KnownOne2, TD, Depth);
789 case Intrinsic::umul_with_overflow:
790 case Intrinsic::smul_with_overflow:
791 ComputeMaskedBitsMul(II->getArgOperand(0), II->getArgOperand(1),
792 false, KnownZero, KnownOne,
793 KnownZero2, KnownOne2, TD, Depth);
801 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
802 /// one. Convenience wrapper around ComputeMaskedBits.
803 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
804 const DataLayout *TD, unsigned Depth) {
805 unsigned BitWidth = getBitWidth(V->getType(), TD);
811 APInt ZeroBits(BitWidth, 0);
812 APInt OneBits(BitWidth, 0);
813 ComputeMaskedBits(V, ZeroBits, OneBits, TD, Depth);
814 KnownOne = OneBits[BitWidth - 1];
815 KnownZero = ZeroBits[BitWidth - 1];
818 /// isKnownToBeAPowerOfTwo - Return true if the given value is known to have exactly one
819 /// bit set when defined. For vectors return true if every element is known to
820 /// be a power of two when defined. Supports values with integer or pointer
821 /// types and vectors of integers.
822 bool llvm::isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth) {
823 if (Constant *C = dyn_cast<Constant>(V)) {
824 if (C->isNullValue())
826 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
827 return CI->getValue().isPowerOf2();
828 // TODO: Handle vector constants.
831 // 1 << X is clearly a power of two if the one is not shifted off the end. If
832 // it is shifted off the end then the result is undefined.
833 if (match(V, m_Shl(m_One(), m_Value())))
836 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
837 // bottom. If it is shifted off the bottom then the result is undefined.
838 if (match(V, m_LShr(m_SignBit(), m_Value())))
841 // The remaining tests are all recursive, so bail out if we hit the limit.
842 if (Depth++ == MaxDepth)
845 Value *X = nullptr, *Y = nullptr;
846 // A shift of a power of two is a power of two or zero.
847 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
848 match(V, m_Shr(m_Value(X), m_Value()))))
849 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth);
851 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
852 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth);
854 if (SelectInst *SI = dyn_cast<SelectInst>(V))
855 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth) &&
856 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth);
858 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
859 // A power of two and'd with anything is a power of two or zero.
860 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth) ||
861 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth))
863 // X & (-X) is always a power of two or zero.
864 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
869 // Adding a power-of-two or zero to the same power-of-two or zero yields
870 // either the original power-of-two, a larger power-of-two or zero.
871 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
872 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
873 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
874 if (match(X, m_And(m_Specific(Y), m_Value())) ||
875 match(X, m_And(m_Value(), m_Specific(Y))))
876 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth))
878 if (match(Y, m_And(m_Specific(X), m_Value())) ||
879 match(Y, m_And(m_Value(), m_Specific(X))))
880 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth))
883 unsigned BitWidth = V->getType()->getScalarSizeInBits();
884 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
885 ComputeMaskedBits(X, LHSZeroBits, LHSOneBits, nullptr, Depth);
887 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
888 ComputeMaskedBits(Y, RHSZeroBits, RHSOneBits, nullptr, Depth);
889 // If i8 V is a power of two or zero:
890 // ZeroBits: 1 1 1 0 1 1 1 1
891 // ~ZeroBits: 0 0 0 1 0 0 0 0
892 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
893 // If OrZero isn't set, we cannot give back a zero result.
894 // Make sure either the LHS or RHS has a bit set.
895 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
900 // An exact divide or right shift can only shift off zero bits, so the result
901 // is a power of two only if the first operand is a power of two and not
902 // copying a sign bit (sdiv int_min, 2).
903 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
904 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
905 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, Depth);
911 /// \brief Test whether a GEP's result is known to be non-null.
913 /// Uses properties inherent in a GEP to try to determine whether it is known
916 /// Currently this routine does not support vector GEPs.
917 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
919 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
922 // FIXME: Support vector-GEPs.
923 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
925 // If the base pointer is non-null, we cannot walk to a null address with an
926 // inbounds GEP in address space zero.
927 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth))
930 // Past this, if we don't have DataLayout, we can't do much.
934 // Walk the GEP operands and see if any operand introduces a non-zero offset.
935 // If so, then the GEP cannot produce a null pointer, as doing so would
936 // inherently violate the inbounds contract within address space zero.
937 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
939 // Struct types are easy -- they must always be indexed by a constant.
940 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
941 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
942 unsigned ElementIdx = OpC->getZExtValue();
943 const StructLayout *SL = DL->getStructLayout(STy);
944 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
945 if (ElementOffset > 0)
950 // If we have a zero-sized type, the index doesn't matter. Keep looping.
951 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
954 // Fast path the constant operand case both for efficiency and so we don't
955 // increment Depth when just zipping down an all-constant GEP.
956 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
962 // We post-increment Depth here because while isKnownNonZero increments it
963 // as well, when we pop back up that increment won't persist. We don't want
964 // to recurse 10k times just because we have 10k GEP operands. We don't
965 // bail completely out because we want to handle constant GEPs regardless
967 if (Depth++ >= MaxDepth)
970 if (isKnownNonZero(GTI.getOperand(), DL, Depth))
977 /// isKnownNonZero - Return true if the given value is known to be non-zero
978 /// when defined. For vectors return true if every element is known to be
979 /// non-zero when defined. Supports values with integer or pointer type and
980 /// vectors of integers.
981 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth) {
982 if (Constant *C = dyn_cast<Constant>(V)) {
983 if (C->isNullValue())
985 if (isa<ConstantInt>(C))
986 // Must be non-zero due to null test above.
988 // TODO: Handle vectors
992 // The remaining tests are all recursive, so bail out if we hit the limit.
993 if (Depth++ >= MaxDepth)
996 // Check for pointer simplifications.
997 if (V->getType()->isPointerTy()) {
998 if (isKnownNonNull(V))
1000 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1001 if (isGEPKnownNonNull(GEP, TD, Depth))
1005 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
1007 // X | Y != 0 if X != 0 or Y != 0.
1008 Value *X = nullptr, *Y = nullptr;
1009 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1010 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
1012 // ext X != 0 if X != 0.
1013 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1014 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
1016 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1017 // if the lowest bit is shifted off the end.
1018 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1019 // shl nuw can't remove any non-zero bits.
1020 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1021 if (BO->hasNoUnsignedWrap())
1022 return isKnownNonZero(X, TD, Depth);
1024 APInt KnownZero(BitWidth, 0);
1025 APInt KnownOne(BitWidth, 0);
1026 ComputeMaskedBits(X, KnownZero, KnownOne, TD, Depth);
1030 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1031 // defined if the sign bit is shifted off the end.
1032 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1033 // shr exact can only shift out zero bits.
1034 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1036 return isKnownNonZero(X, TD, Depth);
1038 bool XKnownNonNegative, XKnownNegative;
1039 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
1043 // div exact can only produce a zero if the dividend is zero.
1044 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1045 return isKnownNonZero(X, TD, Depth);
1048 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1049 bool XKnownNonNegative, XKnownNegative;
1050 bool YKnownNonNegative, YKnownNegative;
1051 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
1052 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
1054 // If X and Y are both non-negative (as signed values) then their sum is not
1055 // zero unless both X and Y are zero.
1056 if (XKnownNonNegative && YKnownNonNegative)
1057 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
1060 // If X and Y are both negative (as signed values) then their sum is not
1061 // zero unless both X and Y equal INT_MIN.
1062 if (BitWidth && XKnownNegative && YKnownNegative) {
1063 APInt KnownZero(BitWidth, 0);
1064 APInt KnownOne(BitWidth, 0);
1065 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1066 // The sign bit of X is set. If some other bit is set then X is not equal
1068 ComputeMaskedBits(X, KnownZero, KnownOne, TD, Depth);
1069 if ((KnownOne & Mask) != 0)
1071 // The sign bit of Y is set. If some other bit is set then Y is not equal
1073 ComputeMaskedBits(Y, KnownZero, KnownOne, TD, Depth);
1074 if ((KnownOne & Mask) != 0)
1078 // The sum of a non-negative number and a power of two is not zero.
1079 if (XKnownNonNegative && isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth))
1081 if (YKnownNonNegative && isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth))
1085 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1086 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1087 // If X and Y are non-zero then so is X * Y as long as the multiplication
1088 // does not overflow.
1089 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1090 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
1093 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1094 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1095 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
1096 isKnownNonZero(SI->getFalseValue(), TD, Depth))
1100 if (!BitWidth) return false;
1101 APInt KnownZero(BitWidth, 0);
1102 APInt KnownOne(BitWidth, 0);
1103 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1104 return KnownOne != 0;
1107 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1108 /// this predicate to simplify operations downstream. Mask is known to be zero
1109 /// for bits that V cannot have.
1111 /// This function is defined on values with integer type, values with pointer
1112 /// type (but only if TD is non-null), and vectors of integers. In the case
1113 /// where V is a vector, the mask, known zero, and known one values are the
1114 /// same width as the vector element, and the bit is set only if it is true
1115 /// for all of the elements in the vector.
1116 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
1117 const DataLayout *TD, unsigned Depth) {
1118 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1119 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1120 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1121 return (KnownZero & Mask) == Mask;
1126 /// ComputeNumSignBits - Return the number of times the sign bit of the
1127 /// register is replicated into the other bits. We know that at least 1 bit
1128 /// is always equal to the sign bit (itself), but other cases can give us
1129 /// information. For example, immediately after an "ashr X, 2", we know that
1130 /// the top 3 bits are all equal to each other, so we return 3.
1132 /// 'Op' must have a scalar integer type.
1134 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
1136 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1137 "ComputeNumSignBits requires a DataLayout object to operate "
1138 "on non-integer values!");
1139 Type *Ty = V->getType();
1140 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1141 Ty->getScalarSizeInBits();
1143 unsigned FirstAnswer = 1;
1145 // Note that ConstantInt is handled by the general ComputeMaskedBits case
1149 return 1; // Limit search depth.
1151 Operator *U = dyn_cast<Operator>(V);
1152 switch (Operator::getOpcode(V)) {
1154 case Instruction::SExt:
1155 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1156 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
1158 case Instruction::AShr: {
1159 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1160 // ashr X, C -> adds C sign bits. Vectors too.
1162 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1163 Tmp += ShAmt->getZExtValue();
1164 if (Tmp > TyBits) Tmp = TyBits;
1168 case Instruction::Shl: {
1170 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1171 // shl destroys sign bits.
1172 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1173 Tmp2 = ShAmt->getZExtValue();
1174 if (Tmp2 >= TyBits || // Bad shift.
1175 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1180 case Instruction::And:
1181 case Instruction::Or:
1182 case Instruction::Xor: // NOT is handled here.
1183 // Logical binary ops preserve the number of sign bits at the worst.
1184 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1186 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1187 FirstAnswer = std::min(Tmp, Tmp2);
1188 // We computed what we know about the sign bits as our first
1189 // answer. Now proceed to the generic code that uses
1190 // ComputeMaskedBits, and pick whichever answer is better.
1194 case Instruction::Select:
1195 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1196 if (Tmp == 1) return 1; // Early out.
1197 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1198 return std::min(Tmp, Tmp2);
1200 case Instruction::Add:
1201 // Add can have at most one carry bit. Thus we know that the output
1202 // is, at worst, one more bit than the inputs.
1203 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1204 if (Tmp == 1) return 1; // Early out.
1206 // Special case decrementing a value (ADD X, -1):
1207 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1208 if (CRHS->isAllOnesValue()) {
1209 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1210 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
1212 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1214 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1217 // If we are subtracting one from a positive number, there is no carry
1218 // out of the result.
1219 if (KnownZero.isNegative())
1223 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1224 if (Tmp2 == 1) return 1;
1225 return std::min(Tmp, Tmp2)-1;
1227 case Instruction::Sub:
1228 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1229 if (Tmp2 == 1) return 1;
1232 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1233 if (CLHS->isNullValue()) {
1234 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1235 ComputeMaskedBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
1236 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1238 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1241 // If the input is known to be positive (the sign bit is known clear),
1242 // the output of the NEG has the same number of sign bits as the input.
1243 if (KnownZero.isNegative())
1246 // Otherwise, we treat this like a SUB.
1249 // Sub can have at most one carry bit. Thus we know that the output
1250 // is, at worst, one more bit than the inputs.
1251 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1252 if (Tmp == 1) return 1; // Early out.
1253 return std::min(Tmp, Tmp2)-1;
1255 case Instruction::PHI: {
1256 PHINode *PN = cast<PHINode>(U);
1257 // Don't analyze large in-degree PHIs.
1258 if (PN->getNumIncomingValues() > 4) break;
1260 // Take the minimum of all incoming values. This can't infinitely loop
1261 // because of our depth threshold.
1262 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1263 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1264 if (Tmp == 1) return Tmp;
1266 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1271 case Instruction::Trunc:
1272 // FIXME: it's tricky to do anything useful for this, but it is an important
1273 // case for targets like X86.
1277 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1278 // use this information.
1279 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1281 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1283 if (KnownZero.isNegative()) { // sign bit is 0
1285 } else if (KnownOne.isNegative()) { // sign bit is 1;
1292 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1293 // the number of identical bits in the top of the input value.
1295 Mask <<= Mask.getBitWidth()-TyBits;
1296 // Return # leading zeros. We use 'min' here in case Val was zero before
1297 // shifting. We don't want to return '64' as for an i32 "0".
1298 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1301 /// ComputeMultiple - This function computes the integer multiple of Base that
1302 /// equals V. If successful, it returns true and returns the multiple in
1303 /// Multiple. If unsuccessful, it returns false. It looks
1304 /// through SExt instructions only if LookThroughSExt is true.
1305 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1306 bool LookThroughSExt, unsigned Depth) {
1307 const unsigned MaxDepth = 6;
1309 assert(V && "No Value?");
1310 assert(Depth <= MaxDepth && "Limit Search Depth");
1311 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1313 Type *T = V->getType();
1315 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1325 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1326 Constant *BaseVal = ConstantInt::get(T, Base);
1327 if (CO && CO == BaseVal) {
1329 Multiple = ConstantInt::get(T, 1);
1333 if (CI && CI->getZExtValue() % Base == 0) {
1334 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1338 if (Depth == MaxDepth) return false; // Limit search depth.
1340 Operator *I = dyn_cast<Operator>(V);
1341 if (!I) return false;
1343 switch (I->getOpcode()) {
1345 case Instruction::SExt:
1346 if (!LookThroughSExt) return false;
1347 // otherwise fall through to ZExt
1348 case Instruction::ZExt:
1349 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1350 LookThroughSExt, Depth+1);
1351 case Instruction::Shl:
1352 case Instruction::Mul: {
1353 Value *Op0 = I->getOperand(0);
1354 Value *Op1 = I->getOperand(1);
1356 if (I->getOpcode() == Instruction::Shl) {
1357 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1358 if (!Op1CI) return false;
1359 // Turn Op0 << Op1 into Op0 * 2^Op1
1360 APInt Op1Int = Op1CI->getValue();
1361 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1362 APInt API(Op1Int.getBitWidth(), 0);
1363 API.setBit(BitToSet);
1364 Op1 = ConstantInt::get(V->getContext(), API);
1367 Value *Mul0 = nullptr;
1368 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1369 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1370 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1371 if (Op1C->getType()->getPrimitiveSizeInBits() <
1372 MulC->getType()->getPrimitiveSizeInBits())
1373 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1374 if (Op1C->getType()->getPrimitiveSizeInBits() >
1375 MulC->getType()->getPrimitiveSizeInBits())
1376 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1378 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1379 Multiple = ConstantExpr::getMul(MulC, Op1C);
1383 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1384 if (Mul0CI->getValue() == 1) {
1385 // V == Base * Op1, so return Op1
1391 Value *Mul1 = nullptr;
1392 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1393 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1394 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1395 if (Op0C->getType()->getPrimitiveSizeInBits() <
1396 MulC->getType()->getPrimitiveSizeInBits())
1397 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1398 if (Op0C->getType()->getPrimitiveSizeInBits() >
1399 MulC->getType()->getPrimitiveSizeInBits())
1400 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1402 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1403 Multiple = ConstantExpr::getMul(MulC, Op0C);
1407 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1408 if (Mul1CI->getValue() == 1) {
1409 // V == Base * Op0, so return Op0
1417 // We could not determine if V is a multiple of Base.
1421 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1422 /// value is never equal to -0.0.
1424 /// NOTE: this function will need to be revisited when we support non-default
1427 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1428 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1429 return !CFP->getValueAPF().isNegZero();
1432 return 1; // Limit search depth.
1434 const Operator *I = dyn_cast<Operator>(V);
1435 if (!I) return false;
1437 // Check if the nsz fast-math flag is set
1438 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
1439 if (FPO->hasNoSignedZeros())
1442 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1443 if (I->getOpcode() == Instruction::FAdd)
1444 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
1445 if (CFP->isNullValue())
1448 // sitofp and uitofp turn into +0.0 for zero.
1449 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1452 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1453 // sqrt(-0.0) = -0.0, no other negative results are possible.
1454 if (II->getIntrinsicID() == Intrinsic::sqrt)
1455 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1457 if (const CallInst *CI = dyn_cast<CallInst>(I))
1458 if (const Function *F = CI->getCalledFunction()) {
1459 if (F->isDeclaration()) {
1461 if (F->getName() == "abs") return true;
1462 // fabs[lf](x) != -0.0
1463 if (F->getName() == "fabs") return true;
1464 if (F->getName() == "fabsf") return true;
1465 if (F->getName() == "fabsl") return true;
1466 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1467 F->getName() == "sqrtl")
1468 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1475 /// isBytewiseValue - If the specified value can be set by repeating the same
1476 /// byte in memory, return the i8 value that it is represented with. This is
1477 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1478 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1479 /// byte store (e.g. i16 0x1234), return null.
1480 Value *llvm::isBytewiseValue(Value *V) {
1481 // All byte-wide stores are splatable, even of arbitrary variables.
1482 if (V->getType()->isIntegerTy(8)) return V;
1484 // Handle 'null' ConstantArrayZero etc.
1485 if (Constant *C = dyn_cast<Constant>(V))
1486 if (C->isNullValue())
1487 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1489 // Constant float and double values can be handled as integer values if the
1490 // corresponding integer value is "byteable". An important case is 0.0.
1491 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1492 if (CFP->getType()->isFloatTy())
1493 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1494 if (CFP->getType()->isDoubleTy())
1495 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1496 // Don't handle long double formats, which have strange constraints.
1499 // We can handle constant integers that are power of two in size and a
1500 // multiple of 8 bits.
1501 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1502 unsigned Width = CI->getBitWidth();
1503 if (isPowerOf2_32(Width) && Width > 8) {
1504 // We can handle this value if the recursive binary decomposition is the
1505 // same at all levels.
1506 APInt Val = CI->getValue();
1508 while (Val.getBitWidth() != 8) {
1509 unsigned NextWidth = Val.getBitWidth()/2;
1510 Val2 = Val.lshr(NextWidth);
1511 Val2 = Val2.trunc(Val.getBitWidth()/2);
1512 Val = Val.trunc(Val.getBitWidth()/2);
1514 // If the top/bottom halves aren't the same, reject it.
1518 return ConstantInt::get(V->getContext(), Val);
1522 // A ConstantDataArray/Vector is splatable if all its members are equal and
1524 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1525 Value *Elt = CA->getElementAsConstant(0);
1526 Value *Val = isBytewiseValue(Elt);
1530 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1531 if (CA->getElementAsConstant(I) != Elt)
1537 // Conceptually, we could handle things like:
1538 // %a = zext i8 %X to i16
1539 // %b = shl i16 %a, 8
1540 // %c = or i16 %a, %b
1541 // but until there is an example that actually needs this, it doesn't seem
1542 // worth worrying about.
1547 // This is the recursive version of BuildSubAggregate. It takes a few different
1548 // arguments. Idxs is the index within the nested struct From that we are
1549 // looking at now (which is of type IndexedType). IdxSkip is the number of
1550 // indices from Idxs that should be left out when inserting into the resulting
1551 // struct. To is the result struct built so far, new insertvalue instructions
1553 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1554 SmallVectorImpl<unsigned> &Idxs,
1556 Instruction *InsertBefore) {
1557 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
1559 // Save the original To argument so we can modify it
1561 // General case, the type indexed by Idxs is a struct
1562 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1563 // Process each struct element recursively
1566 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1570 // Couldn't find any inserted value for this index? Cleanup
1571 while (PrevTo != OrigTo) {
1572 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1573 PrevTo = Del->getAggregateOperand();
1574 Del->eraseFromParent();
1576 // Stop processing elements
1580 // If we successfully found a value for each of our subaggregates
1584 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1585 // the struct's elements had a value that was inserted directly. In the latter
1586 // case, perhaps we can't determine each of the subelements individually, but
1587 // we might be able to find the complete struct somewhere.
1589 // Find the value that is at that particular spot
1590 Value *V = FindInsertedValue(From, Idxs);
1595 // Insert the value in the new (sub) aggregrate
1596 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1597 "tmp", InsertBefore);
1600 // This helper takes a nested struct and extracts a part of it (which is again a
1601 // struct) into a new value. For example, given the struct:
1602 // { a, { b, { c, d }, e } }
1603 // and the indices "1, 1" this returns
1606 // It does this by inserting an insertvalue for each element in the resulting
1607 // struct, as opposed to just inserting a single struct. This will only work if
1608 // each of the elements of the substruct are known (ie, inserted into From by an
1609 // insertvalue instruction somewhere).
1611 // All inserted insertvalue instructions are inserted before InsertBefore
1612 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1613 Instruction *InsertBefore) {
1614 assert(InsertBefore && "Must have someplace to insert!");
1615 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1617 Value *To = UndefValue::get(IndexedType);
1618 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1619 unsigned IdxSkip = Idxs.size();
1621 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1624 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1625 /// the scalar value indexed is already around as a register, for example if it
1626 /// were inserted directly into the aggregrate.
1628 /// If InsertBefore is not null, this function will duplicate (modified)
1629 /// insertvalues when a part of a nested struct is extracted.
1630 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1631 Instruction *InsertBefore) {
1632 // Nothing to index? Just return V then (this is useful at the end of our
1634 if (idx_range.empty())
1636 // We have indices, so V should have an indexable type.
1637 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1638 "Not looking at a struct or array?");
1639 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1640 "Invalid indices for type?");
1642 if (Constant *C = dyn_cast<Constant>(V)) {
1643 C = C->getAggregateElement(idx_range[0]);
1644 if (!C) return nullptr;
1645 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1648 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1649 // Loop the indices for the insertvalue instruction in parallel with the
1650 // requested indices
1651 const unsigned *req_idx = idx_range.begin();
1652 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1653 i != e; ++i, ++req_idx) {
1654 if (req_idx == idx_range.end()) {
1655 // We can't handle this without inserting insertvalues
1659 // The requested index identifies a part of a nested aggregate. Handle
1660 // this specially. For example,
1661 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1662 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1663 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1664 // This can be changed into
1665 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1666 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1667 // which allows the unused 0,0 element from the nested struct to be
1669 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1673 // This insert value inserts something else than what we are looking for.
1674 // See if the (aggregrate) value inserted into has the value we are
1675 // looking for, then.
1677 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1680 // If we end up here, the indices of the insertvalue match with those
1681 // requested (though possibly only partially). Now we recursively look at
1682 // the inserted value, passing any remaining indices.
1683 return FindInsertedValue(I->getInsertedValueOperand(),
1684 makeArrayRef(req_idx, idx_range.end()),
1688 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1689 // If we're extracting a value from an aggregrate that was extracted from
1690 // something else, we can extract from that something else directly instead.
1691 // However, we will need to chain I's indices with the requested indices.
1693 // Calculate the number of indices required
1694 unsigned size = I->getNumIndices() + idx_range.size();
1695 // Allocate some space to put the new indices in
1696 SmallVector<unsigned, 5> Idxs;
1698 // Add indices from the extract value instruction
1699 Idxs.append(I->idx_begin(), I->idx_end());
1701 // Add requested indices
1702 Idxs.append(idx_range.begin(), idx_range.end());
1704 assert(Idxs.size() == size
1705 && "Number of indices added not correct?");
1707 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1709 // Otherwise, we don't know (such as, extracting from a function return value
1710 // or load instruction)
1714 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1715 /// it can be expressed as a base pointer plus a constant offset. Return the
1716 /// base and offset to the caller.
1717 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1718 const DataLayout *DL) {
1719 // Without DataLayout, conservatively assume 64-bit offsets, which is
1720 // the widest we support.
1721 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(Ptr->getType()) : 64;
1722 APInt ByteOffset(BitWidth, 0);
1724 if (Ptr->getType()->isVectorTy())
1727 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1729 APInt GEPOffset(BitWidth, 0);
1730 if (!GEP->accumulateConstantOffset(*DL, GEPOffset))
1733 ByteOffset += GEPOffset;
1736 Ptr = GEP->getPointerOperand();
1737 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1738 Ptr = cast<Operator>(Ptr)->getOperand(0);
1739 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1740 if (GA->mayBeOverridden())
1742 Ptr = GA->getAliasee();
1747 Offset = ByteOffset.getSExtValue();
1752 /// getConstantStringInfo - This function computes the length of a
1753 /// null-terminated C string pointed to by V. If successful, it returns true
1754 /// and returns the string in Str. If unsuccessful, it returns false.
1755 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
1756 uint64_t Offset, bool TrimAtNul) {
1759 // Look through bitcast instructions and geps.
1760 V = V->stripPointerCasts();
1762 // If the value is a GEP instructionor constant expression, treat it as an
1764 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1765 // Make sure the GEP has exactly three arguments.
1766 if (GEP->getNumOperands() != 3)
1769 // Make sure the index-ee is a pointer to array of i8.
1770 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1771 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1772 if (!AT || !AT->getElementType()->isIntegerTy(8))
1775 // Check to make sure that the first operand of the GEP is an integer and
1776 // has value 0 so that we are sure we're indexing into the initializer.
1777 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1778 if (!FirstIdx || !FirstIdx->isZero())
1781 // If the second index isn't a ConstantInt, then this is a variable index
1782 // into the array. If this occurs, we can't say anything meaningful about
1784 uint64_t StartIdx = 0;
1785 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1786 StartIdx = CI->getZExtValue();
1789 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
1792 // The GEP instruction, constant or instruction, must reference a global
1793 // variable that is a constant and is initialized. The referenced constant
1794 // initializer is the array that we'll use for optimization.
1795 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
1796 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1799 // Handle the all-zeros case
1800 if (GV->getInitializer()->isNullValue()) {
1801 // This is a degenerate case. The initializer is constant zero so the
1802 // length of the string must be zero.
1807 // Must be a Constant Array
1808 const ConstantDataArray *Array =
1809 dyn_cast<ConstantDataArray>(GV->getInitializer());
1810 if (!Array || !Array->isString())
1813 // Get the number of elements in the array
1814 uint64_t NumElts = Array->getType()->getArrayNumElements();
1816 // Start out with the entire array in the StringRef.
1817 Str = Array->getAsString();
1819 if (Offset > NumElts)
1822 // Skip over 'offset' bytes.
1823 Str = Str.substr(Offset);
1826 // Trim off the \0 and anything after it. If the array is not nul
1827 // terminated, we just return the whole end of string. The client may know
1828 // some other way that the string is length-bound.
1829 Str = Str.substr(0, Str.find('\0'));
1834 // These next two are very similar to the above, but also look through PHI
1836 // TODO: See if we can integrate these two together.
1838 /// GetStringLengthH - If we can compute the length of the string pointed to by
1839 /// the specified pointer, return 'len+1'. If we can't, return 0.
1840 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1841 // Look through noop bitcast instructions.
1842 V = V->stripPointerCasts();
1844 // If this is a PHI node, there are two cases: either we have already seen it
1846 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1847 if (!PHIs.insert(PN))
1848 return ~0ULL; // already in the set.
1850 // If it was new, see if all the input strings are the same length.
1851 uint64_t LenSoFar = ~0ULL;
1852 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1853 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1854 if (Len == 0) return 0; // Unknown length -> unknown.
1856 if (Len == ~0ULL) continue;
1858 if (Len != LenSoFar && LenSoFar != ~0ULL)
1859 return 0; // Disagree -> unknown.
1863 // Success, all agree.
1867 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1868 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1869 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1870 if (Len1 == 0) return 0;
1871 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1872 if (Len2 == 0) return 0;
1873 if (Len1 == ~0ULL) return Len2;
1874 if (Len2 == ~0ULL) return Len1;
1875 if (Len1 != Len2) return 0;
1879 // Otherwise, see if we can read the string.
1881 if (!getConstantStringInfo(V, StrData))
1884 return StrData.size()+1;
1887 /// GetStringLength - If we can compute the length of the string pointed to by
1888 /// the specified pointer, return 'len+1'. If we can't, return 0.
1889 uint64_t llvm::GetStringLength(Value *V) {
1890 if (!V->getType()->isPointerTy()) return 0;
1892 SmallPtrSet<PHINode*, 32> PHIs;
1893 uint64_t Len = GetStringLengthH(V, PHIs);
1894 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1895 // an empty string as a length.
1896 return Len == ~0ULL ? 1 : Len;
1900 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
1901 if (!V->getType()->isPointerTy())
1903 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1904 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1905 V = GEP->getPointerOperand();
1906 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1907 V = cast<Operator>(V)->getOperand(0);
1908 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1909 if (GA->mayBeOverridden())
1911 V = GA->getAliasee();
1913 // See if InstructionSimplify knows any relevant tricks.
1914 if (Instruction *I = dyn_cast<Instruction>(V))
1915 // TODO: Acquire a DominatorTree and use it.
1916 if (Value *Simplified = SimplifyInstruction(I, TD, nullptr)) {
1923 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1929 llvm::GetUnderlyingObjects(Value *V,
1930 SmallVectorImpl<Value *> &Objects,
1931 const DataLayout *TD,
1932 unsigned MaxLookup) {
1933 SmallPtrSet<Value *, 4> Visited;
1934 SmallVector<Value *, 4> Worklist;
1935 Worklist.push_back(V);
1937 Value *P = Worklist.pop_back_val();
1938 P = GetUnderlyingObject(P, TD, MaxLookup);
1940 if (!Visited.insert(P))
1943 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
1944 Worklist.push_back(SI->getTrueValue());
1945 Worklist.push_back(SI->getFalseValue());
1949 if (PHINode *PN = dyn_cast<PHINode>(P)) {
1950 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1951 Worklist.push_back(PN->getIncomingValue(i));
1955 Objects.push_back(P);
1956 } while (!Worklist.empty());
1959 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1960 /// are lifetime markers.
1962 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1963 for (const User *U : V->users()) {
1964 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1965 if (!II) return false;
1967 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1968 II->getIntrinsicID() != Intrinsic::lifetime_end)
1974 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1975 const DataLayout *TD) {
1976 const Operator *Inst = dyn_cast<Operator>(V);
1980 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1981 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1985 switch (Inst->getOpcode()) {
1988 case Instruction::UDiv:
1989 case Instruction::URem:
1990 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1991 return isKnownNonZero(Inst->getOperand(1), TD);
1992 case Instruction::SDiv:
1993 case Instruction::SRem: {
1994 Value *Op = Inst->getOperand(1);
1995 // x / y is undefined if y == 0
1996 if (!isKnownNonZero(Op, TD))
1998 // x / y might be undefined if y == -1
1999 unsigned BitWidth = getBitWidth(Op->getType(), TD);
2002 APInt KnownZero(BitWidth, 0);
2003 APInt KnownOne(BitWidth, 0);
2004 ComputeMaskedBits(Op, KnownZero, KnownOne, TD);
2007 case Instruction::Load: {
2008 const LoadInst *LI = cast<LoadInst>(Inst);
2009 if (!LI->isUnordered() ||
2010 // Speculative load may create a race that did not exist in the source.
2011 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2013 return LI->getPointerOperand()->isDereferenceablePointer();
2015 case Instruction::Call: {
2016 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2017 switch (II->getIntrinsicID()) {
2018 // These synthetic intrinsics have no side-effects, and just mark
2019 // information about their operands.
2020 // FIXME: There are other no-op synthetic instructions that potentially
2021 // should be considered at least *safe* to speculate...
2022 case Intrinsic::dbg_declare:
2023 case Intrinsic::dbg_value:
2026 case Intrinsic::bswap:
2027 case Intrinsic::ctlz:
2028 case Intrinsic::ctpop:
2029 case Intrinsic::cttz:
2030 case Intrinsic::objectsize:
2031 case Intrinsic::sadd_with_overflow:
2032 case Intrinsic::smul_with_overflow:
2033 case Intrinsic::ssub_with_overflow:
2034 case Intrinsic::uadd_with_overflow:
2035 case Intrinsic::umul_with_overflow:
2036 case Intrinsic::usub_with_overflow:
2038 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2039 // errno like libm sqrt would.
2040 case Intrinsic::sqrt:
2041 case Intrinsic::fma:
2042 case Intrinsic::fmuladd:
2044 // TODO: some fp intrinsics are marked as having the same error handling
2045 // as libm. They're safe to speculate when they won't error.
2046 // TODO: are convert_{from,to}_fp16 safe?
2047 // TODO: can we list target-specific intrinsics here?
2051 return false; // The called function could have undefined behavior or
2052 // side-effects, even if marked readnone nounwind.
2054 case Instruction::VAArg:
2055 case Instruction::Alloca:
2056 case Instruction::Invoke:
2057 case Instruction::PHI:
2058 case Instruction::Store:
2059 case Instruction::Ret:
2060 case Instruction::Br:
2061 case Instruction::IndirectBr:
2062 case Instruction::Switch:
2063 case Instruction::Unreachable:
2064 case Instruction::Fence:
2065 case Instruction::LandingPad:
2066 case Instruction::AtomicRMW:
2067 case Instruction::AtomicCmpXchg:
2068 case Instruction::Resume:
2069 return false; // Misc instructions which have effects
2073 /// isKnownNonNull - Return true if we know that the specified value is never
2075 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2076 // Alloca never returns null, malloc might.
2077 if (isa<AllocaInst>(V)) return true;
2079 // A byval or inalloca argument is never null.
2080 if (const Argument *A = dyn_cast<Argument>(V))
2081 return A->hasByValOrInAllocaAttr();
2083 // Global values are not null unless extern weak.
2084 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2085 return !GV->hasExternalWeakLinkage();
2087 // operator new never returns null.
2088 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))