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/Analysis/InstructionSimplify.h"
17 #include "llvm/Constants.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/GlobalVariable.h"
20 #include "llvm/GlobalAlias.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/Metadata.h"
24 #include "llvm/Operator.h"
25 #include "llvm/DataLayout.h"
26 #include "llvm/Support/ConstantRange.h"
27 #include "llvm/Support/GetElementPtrTypeIterator.h"
28 #include "llvm/Support/MathExtras.h"
29 #include "llvm/Support/PatternMatch.h"
30 #include "llvm/ADT/SmallPtrSet.h"
33 using namespace llvm::PatternMatch;
35 const unsigned MaxDepth = 6;
37 /// getBitWidth - Returns the bitwidth of the given scalar or pointer type (if
38 /// unknown returns 0). For vector types, returns the element type's bitwidth.
39 static unsigned getBitWidth(Type *Ty, const DataLayout *TD) {
40 if (unsigned BitWidth = Ty->getScalarSizeInBits())
42 assert(isa<PointerType>(Ty) && "Expected a pointer type!");
44 TD->getPointerSizeInBits(cast<PointerType>(Ty)->getAddressSpace()) : 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_32(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->hasByValAttr()) {
315 // Get alignment information off byval arguments if specified in the IR.
316 Align = A->getParamAlignment();
317 } else if (TD && A->hasStructRetAttr()) {
318 // An sret parameter has at least the ABI alignment of the return type.
319 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
320 if (EltTy->isSized())
321 Align = TD->getABITypeAlignment(EltTy);
325 KnownZero = APInt::getLowBitsSet(BitWidth, CountTrailingZeros_32(Align));
329 // Start out not knowing anything.
330 KnownZero.clearAllBits(); KnownOne.clearAllBits();
332 if (Depth == MaxDepth)
333 return; // Limit search depth.
335 Operator *I = dyn_cast<Operator>(V);
338 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
339 switch (I->getOpcode()) {
341 case Instruction::Load:
342 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
343 computeMaskedBitsLoad(*MD, KnownZero);
345 case Instruction::And: {
346 // If either the LHS or the RHS are Zero, the result is zero.
347 ComputeMaskedBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
348 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
349 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
350 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
352 // Output known-1 bits are only known if set in both the LHS & RHS.
353 KnownOne &= KnownOne2;
354 // Output known-0 are known to be clear if zero in either the LHS | RHS.
355 KnownZero |= KnownZero2;
358 case Instruction::Or: {
359 ComputeMaskedBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
360 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
361 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
362 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
364 // Output known-0 bits are only known if clear in both the LHS & RHS.
365 KnownZero &= KnownZero2;
366 // Output known-1 are known to be set if set in either the LHS | RHS.
367 KnownOne |= KnownOne2;
370 case Instruction::Xor: {
371 ComputeMaskedBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
372 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
373 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
374 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
376 // Output known-0 bits are known if clear or set in both the LHS & RHS.
377 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
378 // Output known-1 are known to be set if set in only one of the LHS, RHS.
379 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
380 KnownZero = KnownZeroOut;
383 case Instruction::Mul: {
384 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
385 ComputeMaskedBitsMul(I->getOperand(0), I->getOperand(1), NSW,
386 KnownZero, KnownOne, KnownZero2, KnownOne2, TD, Depth);
389 case Instruction::UDiv: {
390 // For the purposes of computing leading zeros we can conservatively
391 // treat a udiv as a logical right shift by the power of 2 known to
392 // be less than the denominator.
393 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
394 unsigned LeadZ = KnownZero2.countLeadingOnes();
396 KnownOne2.clearAllBits();
397 KnownZero2.clearAllBits();
398 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
399 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
400 if (RHSUnknownLeadingOnes != BitWidth)
401 LeadZ = std::min(BitWidth,
402 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
404 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
407 case Instruction::Select:
408 ComputeMaskedBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1);
409 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD,
411 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
412 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
414 // Only known if known in both the LHS and RHS.
415 KnownOne &= KnownOne2;
416 KnownZero &= KnownZero2;
418 case Instruction::FPTrunc:
419 case Instruction::FPExt:
420 case Instruction::FPToUI:
421 case Instruction::FPToSI:
422 case Instruction::SIToFP:
423 case Instruction::UIToFP:
424 return; // Can't work with floating point.
425 case Instruction::PtrToInt:
426 case Instruction::IntToPtr:
427 // We can't handle these if we don't know the pointer size.
429 // FALL THROUGH and handle them the same as zext/trunc.
430 case Instruction::ZExt:
431 case Instruction::Trunc: {
432 Type *SrcTy = I->getOperand(0)->getType();
434 unsigned SrcBitWidth;
435 // Note that we handle pointer operands here because of inttoptr/ptrtoint
436 // which fall through here.
437 if (SrcTy->isPointerTy())
438 SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
440 SrcBitWidth = SrcTy->getScalarSizeInBits();
442 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
443 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
444 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
445 KnownZero = KnownZero.zextOrTrunc(BitWidth);
446 KnownOne = KnownOne.zextOrTrunc(BitWidth);
447 // Any top bits are known to be zero.
448 if (BitWidth > SrcBitWidth)
449 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
452 case Instruction::BitCast: {
453 Type *SrcTy = I->getOperand(0)->getType();
454 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
455 // TODO: For now, not handling conversions like:
456 // (bitcast i64 %x to <2 x i32>)
457 !I->getType()->isVectorTy()) {
458 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
463 case Instruction::SExt: {
464 // Compute the bits in the result that are not present in the input.
465 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
467 KnownZero = KnownZero.trunc(SrcBitWidth);
468 KnownOne = KnownOne.trunc(SrcBitWidth);
469 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
470 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
471 KnownZero = KnownZero.zext(BitWidth);
472 KnownOne = KnownOne.zext(BitWidth);
474 // If the sign bit of the input is known set or clear, then we know the
475 // top bits of the result.
476 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
477 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
478 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
479 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
482 case Instruction::Shl:
483 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
484 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
485 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
486 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
487 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
488 KnownZero <<= ShiftAmt;
489 KnownOne <<= ShiftAmt;
490 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
494 case Instruction::LShr:
495 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
496 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
497 // Compute the new bits that are at the top now.
498 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
500 // Unsigned shift right.
501 ComputeMaskedBits(I->getOperand(0), KnownZero,KnownOne, TD, Depth+1);
502 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
503 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
504 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
505 // high bits known zero.
506 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
510 case Instruction::AShr:
511 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
512 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
513 // Compute the new bits that are at the top now.
514 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
516 // Signed shift right.
517 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
518 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
519 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
520 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
522 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
523 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
524 KnownZero |= HighBits;
525 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
526 KnownOne |= HighBits;
530 case Instruction::Sub: {
531 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
532 ComputeMaskedBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
533 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
537 case Instruction::Add: {
538 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
539 ComputeMaskedBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
540 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
544 case Instruction::SRem:
545 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
546 APInt RA = Rem->getValue().abs();
547 if (RA.isPowerOf2()) {
548 APInt LowBits = RA - 1;
549 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
551 // The low bits of the first operand are unchanged by the srem.
552 KnownZero = KnownZero2 & LowBits;
553 KnownOne = KnownOne2 & LowBits;
555 // If the first operand is non-negative or has all low bits zero, then
556 // the upper bits are all zero.
557 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
558 KnownZero |= ~LowBits;
560 // If the first operand is negative and not all low bits are zero, then
561 // the upper bits are all one.
562 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
563 KnownOne |= ~LowBits;
565 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
569 // The sign bit is the LHS's sign bit, except when the result of the
570 // remainder is zero.
571 if (KnownZero.isNonNegative()) {
572 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
573 ComputeMaskedBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
575 // If it's known zero, our sign bit is also zero.
576 if (LHSKnownZero.isNegative())
577 KnownZero.setBit(BitWidth - 1);
581 case Instruction::URem: {
582 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
583 APInt RA = Rem->getValue();
584 if (RA.isPowerOf2()) {
585 APInt LowBits = (RA - 1);
586 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD,
588 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
589 KnownZero |= ~LowBits;
595 // Since the result is less than or equal to either operand, any leading
596 // zero bits in either operand must also exist in the result.
597 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
598 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
600 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
601 KnownZero2.countLeadingOnes());
602 KnownOne.clearAllBits();
603 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
607 case Instruction::Alloca: {
608 AllocaInst *AI = cast<AllocaInst>(V);
609 unsigned Align = AI->getAlignment();
610 if (Align == 0 && TD)
611 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
614 KnownZero = APInt::getLowBitsSet(BitWidth, CountTrailingZeros_32(Align));
617 case Instruction::GetElementPtr: {
618 // Analyze all of the subscripts of this getelementptr instruction
619 // to determine if we can prove known low zero bits.
620 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
621 ComputeMaskedBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
623 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
625 gep_type_iterator GTI = gep_type_begin(I);
626 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
627 Value *Index = I->getOperand(i);
628 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
629 // Handle struct member offset arithmetic.
631 const StructLayout *SL = TD->getStructLayout(STy);
632 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
633 uint64_t Offset = SL->getElementOffset(Idx);
634 TrailZ = std::min(TrailZ,
635 CountTrailingZeros_64(Offset));
637 // Handle array index arithmetic.
638 Type *IndexedTy = GTI.getIndexedType();
639 if (!IndexedTy->isSized()) return;
640 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
641 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
642 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
643 ComputeMaskedBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1);
644 TrailZ = std::min(TrailZ,
645 unsigned(CountTrailingZeros_64(TypeSize) +
646 LocalKnownZero.countTrailingOnes()));
650 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
653 case Instruction::PHI: {
654 PHINode *P = cast<PHINode>(I);
655 // Handle the case of a simple two-predecessor recurrence PHI.
656 // There's a lot more that could theoretically be done here, but
657 // this is sufficient to catch some interesting cases.
658 if (P->getNumIncomingValues() == 2) {
659 for (unsigned i = 0; i != 2; ++i) {
660 Value *L = P->getIncomingValue(i);
661 Value *R = P->getIncomingValue(!i);
662 Operator *LU = dyn_cast<Operator>(L);
665 unsigned Opcode = LU->getOpcode();
666 // Check for operations that have the property that if
667 // both their operands have low zero bits, the result
668 // will have low zero bits.
669 if (Opcode == Instruction::Add ||
670 Opcode == Instruction::Sub ||
671 Opcode == Instruction::And ||
672 Opcode == Instruction::Or ||
673 Opcode == Instruction::Mul) {
674 Value *LL = LU->getOperand(0);
675 Value *LR = LU->getOperand(1);
676 // Find a recurrence.
683 // Ok, we have a PHI of the form L op= R. Check for low
685 ComputeMaskedBits(R, KnownZero2, KnownOne2, TD, Depth+1);
687 // We need to take the minimum number of known bits
688 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
689 ComputeMaskedBits(L, KnownZero3, KnownOne3, TD, Depth+1);
691 KnownZero = APInt::getLowBitsSet(BitWidth,
692 std::min(KnownZero2.countTrailingOnes(),
693 KnownZero3.countTrailingOnes()));
699 // Unreachable blocks may have zero-operand PHI nodes.
700 if (P->getNumIncomingValues() == 0)
703 // Otherwise take the unions of the known bit sets of the operands,
704 // taking conservative care to avoid excessive recursion.
705 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
706 // Skip if every incoming value references to ourself.
707 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
710 KnownZero = APInt::getAllOnesValue(BitWidth);
711 KnownOne = APInt::getAllOnesValue(BitWidth);
712 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
713 // Skip direct self references.
714 if (P->getIncomingValue(i) == P) continue;
716 KnownZero2 = APInt(BitWidth, 0);
717 KnownOne2 = APInt(BitWidth, 0);
718 // Recurse, but cap the recursion to one level, because we don't
719 // want to waste time spinning around in loops.
720 ComputeMaskedBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
722 KnownZero &= KnownZero2;
723 KnownOne &= KnownOne2;
724 // If all bits have been ruled out, there's no need to check
726 if (!KnownZero && !KnownOne)
732 case Instruction::Call:
733 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
734 switch (II->getIntrinsicID()) {
736 case Intrinsic::ctlz:
737 case Intrinsic::cttz: {
738 unsigned LowBits = Log2_32(BitWidth)+1;
739 // If this call is undefined for 0, the result will be less than 2^n.
740 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
742 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
745 case Intrinsic::ctpop: {
746 unsigned LowBits = Log2_32(BitWidth)+1;
747 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
750 case Intrinsic::x86_sse42_crc32_64_8:
751 case Intrinsic::x86_sse42_crc32_64_64:
752 KnownZero = APInt::getHighBitsSet(64, 32);
757 case Instruction::ExtractValue:
758 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
759 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
760 if (EVI->getNumIndices() != 1) break;
761 if (EVI->getIndices()[0] == 0) {
762 switch (II->getIntrinsicID()) {
764 case Intrinsic::uadd_with_overflow:
765 case Intrinsic::sadd_with_overflow:
766 ComputeMaskedBitsAddSub(true, II->getArgOperand(0),
767 II->getArgOperand(1), false, KnownZero,
768 KnownOne, KnownZero2, KnownOne2, TD, Depth);
770 case Intrinsic::usub_with_overflow:
771 case Intrinsic::ssub_with_overflow:
772 ComputeMaskedBitsAddSub(false, II->getArgOperand(0),
773 II->getArgOperand(1), false, KnownZero,
774 KnownOne, KnownZero2, KnownOne2, TD, Depth);
776 case Intrinsic::umul_with_overflow:
777 case Intrinsic::smul_with_overflow:
778 ComputeMaskedBitsMul(II->getArgOperand(0), II->getArgOperand(1),
779 false, KnownZero, KnownOne,
780 KnownZero2, KnownOne2, TD, Depth);
788 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
789 /// one. Convenience wrapper around ComputeMaskedBits.
790 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
791 const DataLayout *TD, unsigned Depth) {
792 unsigned BitWidth = getBitWidth(V->getType(), TD);
798 APInt ZeroBits(BitWidth, 0);
799 APInt OneBits(BitWidth, 0);
800 ComputeMaskedBits(V, ZeroBits, OneBits, TD, Depth);
801 KnownOne = OneBits[BitWidth - 1];
802 KnownZero = ZeroBits[BitWidth - 1];
805 /// isPowerOfTwo - Return true if the given value is known to have exactly one
806 /// bit set when defined. For vectors return true if every element is known to
807 /// be a power of two when defined. Supports values with integer or pointer
808 /// types and vectors of integers.
809 bool llvm::isPowerOfTwo(Value *V, const DataLayout *TD, bool OrZero,
811 if (Constant *C = dyn_cast<Constant>(V)) {
812 if (C->isNullValue())
814 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
815 return CI->getValue().isPowerOf2();
816 // TODO: Handle vector constants.
819 // 1 << X is clearly a power of two if the one is not shifted off the end. If
820 // it is shifted off the end then the result is undefined.
821 if (match(V, m_Shl(m_One(), m_Value())))
824 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
825 // bottom. If it is shifted off the bottom then the result is undefined.
826 if (match(V, m_LShr(m_SignBit(), m_Value())))
829 // The remaining tests are all recursive, so bail out if we hit the limit.
830 if (Depth++ == MaxDepth)
833 Value *X = 0, *Y = 0;
834 // A shift of a power of two is a power of two or zero.
835 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
836 match(V, m_Shr(m_Value(X), m_Value()))))
837 return isPowerOfTwo(X, TD, /*OrZero*/true, Depth);
839 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
840 return isPowerOfTwo(ZI->getOperand(0), TD, OrZero, Depth);
842 if (SelectInst *SI = dyn_cast<SelectInst>(V))
843 return isPowerOfTwo(SI->getTrueValue(), TD, OrZero, Depth) &&
844 isPowerOfTwo(SI->getFalseValue(), TD, OrZero, Depth);
846 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
847 // A power of two and'd with anything is a power of two or zero.
848 if (isPowerOfTwo(X, TD, /*OrZero*/true, Depth) ||
849 isPowerOfTwo(Y, TD, /*OrZero*/true, Depth))
851 // X & (-X) is always a power of two or zero.
852 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
857 // An exact divide or right shift can only shift off zero bits, so the result
858 // is a power of two only if the first operand is a power of two and not
859 // copying a sign bit (sdiv int_min, 2).
860 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
861 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
862 return isPowerOfTwo(cast<Operator>(V)->getOperand(0), TD, OrZero, Depth);
868 /// isKnownNonZero - Return true if the given value is known to be non-zero
869 /// when defined. For vectors return true if every element is known to be
870 /// non-zero when defined. Supports values with integer or pointer type and
871 /// vectors of integers.
872 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth) {
873 if (Constant *C = dyn_cast<Constant>(V)) {
874 if (C->isNullValue())
876 if (isa<ConstantInt>(C))
877 // Must be non-zero due to null test above.
879 // TODO: Handle vectors
883 // The remaining tests are all recursive, so bail out if we hit the limit.
884 if (Depth++ >= MaxDepth)
887 unsigned BitWidth = getBitWidth(V->getType(), TD);
889 // X | Y != 0 if X != 0 or Y != 0.
890 Value *X = 0, *Y = 0;
891 if (match(V, m_Or(m_Value(X), m_Value(Y))))
892 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
894 // ext X != 0 if X != 0.
895 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
896 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
898 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
899 // if the lowest bit is shifted off the end.
900 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
901 // shl nuw can't remove any non-zero bits.
902 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
903 if (BO->hasNoUnsignedWrap())
904 return isKnownNonZero(X, TD, Depth);
906 APInt KnownZero(BitWidth, 0);
907 APInt KnownOne(BitWidth, 0);
908 ComputeMaskedBits(X, KnownZero, KnownOne, TD, Depth);
912 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
913 // defined if the sign bit is shifted off the end.
914 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
915 // shr exact can only shift out zero bits.
916 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
918 return isKnownNonZero(X, TD, Depth);
920 bool XKnownNonNegative, XKnownNegative;
921 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
925 // div exact can only produce a zero if the dividend is zero.
926 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
927 return isKnownNonZero(X, TD, Depth);
930 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
931 bool XKnownNonNegative, XKnownNegative;
932 bool YKnownNonNegative, YKnownNegative;
933 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
934 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
936 // If X and Y are both non-negative (as signed values) then their sum is not
937 // zero unless both X and Y are zero.
938 if (XKnownNonNegative && YKnownNonNegative)
939 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
942 // If X and Y are both negative (as signed values) then their sum is not
943 // zero unless both X and Y equal INT_MIN.
944 if (BitWidth && XKnownNegative && YKnownNegative) {
945 APInt KnownZero(BitWidth, 0);
946 APInt KnownOne(BitWidth, 0);
947 APInt Mask = APInt::getSignedMaxValue(BitWidth);
948 // The sign bit of X is set. If some other bit is set then X is not equal
950 ComputeMaskedBits(X, KnownZero, KnownOne, TD, Depth);
951 if ((KnownOne & Mask) != 0)
953 // The sign bit of Y is set. If some other bit is set then Y is not equal
955 ComputeMaskedBits(Y, KnownZero, KnownOne, TD, Depth);
956 if ((KnownOne & Mask) != 0)
960 // The sum of a non-negative number and a power of two is not zero.
961 if (XKnownNonNegative && isPowerOfTwo(Y, TD, /*OrZero*/false, Depth))
963 if (YKnownNonNegative && isPowerOfTwo(X, TD, /*OrZero*/false, Depth))
967 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
968 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
969 // If X and Y are non-zero then so is X * Y as long as the multiplication
970 // does not overflow.
971 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
972 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
975 // (C ? X : Y) != 0 if X != 0 and Y != 0.
976 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
977 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
978 isKnownNonZero(SI->getFalseValue(), TD, Depth))
982 if (!BitWidth) return false;
983 APInt KnownZero(BitWidth, 0);
984 APInt KnownOne(BitWidth, 0);
985 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
986 return KnownOne != 0;
989 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
990 /// this predicate to simplify operations downstream. Mask is known to be zero
991 /// for bits that V cannot have.
993 /// This function is defined on values with integer type, values with pointer
994 /// type (but only if TD is non-null), and vectors of integers. In the case
995 /// where V is a vector, the mask, known zero, and known one values are the
996 /// same width as the vector element, and the bit is set only if it is true
997 /// for all of the elements in the vector.
998 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
999 const DataLayout *TD, unsigned Depth) {
1000 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1001 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1002 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1003 return (KnownZero & Mask) == Mask;
1008 /// ComputeNumSignBits - Return the number of times the sign bit of the
1009 /// register is replicated into the other bits. We know that at least 1 bit
1010 /// is always equal to the sign bit (itself), but other cases can give us
1011 /// information. For example, immediately after an "ashr X, 2", we know that
1012 /// the top 3 bits are all equal to each other, so we return 3.
1014 /// 'Op' must have a scalar integer type.
1016 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
1018 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1019 "ComputeNumSignBits requires a DataLayout object to operate "
1020 "on non-integer values!");
1021 Type *Ty = V->getType();
1022 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1023 Ty->getScalarSizeInBits();
1025 unsigned FirstAnswer = 1;
1027 // Note that ConstantInt is handled by the general ComputeMaskedBits case
1031 return 1; // Limit search depth.
1033 Operator *U = dyn_cast<Operator>(V);
1034 switch (Operator::getOpcode(V)) {
1036 case Instruction::SExt:
1037 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1038 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
1040 case Instruction::AShr: {
1041 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1042 // ashr X, C -> adds C sign bits. Vectors too.
1044 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1045 Tmp += ShAmt->getZExtValue();
1046 if (Tmp > TyBits) Tmp = TyBits;
1050 case Instruction::Shl: {
1052 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1053 // shl destroys sign bits.
1054 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1055 Tmp2 = ShAmt->getZExtValue();
1056 if (Tmp2 >= TyBits || // Bad shift.
1057 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1062 case Instruction::And:
1063 case Instruction::Or:
1064 case Instruction::Xor: // NOT is handled here.
1065 // Logical binary ops preserve the number of sign bits at the worst.
1066 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1068 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1069 FirstAnswer = std::min(Tmp, Tmp2);
1070 // We computed what we know about the sign bits as our first
1071 // answer. Now proceed to the generic code that uses
1072 // ComputeMaskedBits, and pick whichever answer is better.
1076 case Instruction::Select:
1077 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1078 if (Tmp == 1) return 1; // Early out.
1079 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1080 return std::min(Tmp, Tmp2);
1082 case Instruction::Add:
1083 // Add can have at most one carry bit. Thus we know that the output
1084 // is, at worst, one more bit than the inputs.
1085 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1086 if (Tmp == 1) return 1; // Early out.
1088 // Special case decrementing a value (ADD X, -1):
1089 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1090 if (CRHS->isAllOnesValue()) {
1091 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1092 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
1094 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1096 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1099 // If we are subtracting one from a positive number, there is no carry
1100 // out of the result.
1101 if (KnownZero.isNegative())
1105 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1106 if (Tmp2 == 1) return 1;
1107 return std::min(Tmp, Tmp2)-1;
1109 case Instruction::Sub:
1110 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1111 if (Tmp2 == 1) return 1;
1114 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1115 if (CLHS->isNullValue()) {
1116 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1117 ComputeMaskedBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
1118 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1120 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1123 // If the input is known to be positive (the sign bit is known clear),
1124 // the output of the NEG has the same number of sign bits as the input.
1125 if (KnownZero.isNegative())
1128 // Otherwise, we treat this like a SUB.
1131 // Sub can have at most one carry bit. Thus we know that the output
1132 // is, at worst, one more bit than the inputs.
1133 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1134 if (Tmp == 1) return 1; // Early out.
1135 return std::min(Tmp, Tmp2)-1;
1137 case Instruction::PHI: {
1138 PHINode *PN = cast<PHINode>(U);
1139 // Don't analyze large in-degree PHIs.
1140 if (PN->getNumIncomingValues() > 4) break;
1142 // Take the minimum of all incoming values. This can't infinitely loop
1143 // because of our depth threshold.
1144 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1145 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1146 if (Tmp == 1) return Tmp;
1148 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1153 case Instruction::Trunc:
1154 // FIXME: it's tricky to do anything useful for this, but it is an important
1155 // case for targets like X86.
1159 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1160 // use this information.
1161 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1163 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1165 if (KnownZero.isNegative()) { // sign bit is 0
1167 } else if (KnownOne.isNegative()) { // sign bit is 1;
1174 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1175 // the number of identical bits in the top of the input value.
1177 Mask <<= Mask.getBitWidth()-TyBits;
1178 // Return # leading zeros. We use 'min' here in case Val was zero before
1179 // shifting. We don't want to return '64' as for an i32 "0".
1180 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1183 /// ComputeMultiple - This function computes the integer multiple of Base that
1184 /// equals V. If successful, it returns true and returns the multiple in
1185 /// Multiple. If unsuccessful, it returns false. It looks
1186 /// through SExt instructions only if LookThroughSExt is true.
1187 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1188 bool LookThroughSExt, unsigned Depth) {
1189 const unsigned MaxDepth = 6;
1191 assert(V && "No Value?");
1192 assert(Depth <= MaxDepth && "Limit Search Depth");
1193 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1195 Type *T = V->getType();
1197 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1207 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1208 Constant *BaseVal = ConstantInt::get(T, Base);
1209 if (CO && CO == BaseVal) {
1211 Multiple = ConstantInt::get(T, 1);
1215 if (CI && CI->getZExtValue() % Base == 0) {
1216 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1220 if (Depth == MaxDepth) return false; // Limit search depth.
1222 Operator *I = dyn_cast<Operator>(V);
1223 if (!I) return false;
1225 switch (I->getOpcode()) {
1227 case Instruction::SExt:
1228 if (!LookThroughSExt) return false;
1229 // otherwise fall through to ZExt
1230 case Instruction::ZExt:
1231 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1232 LookThroughSExt, Depth+1);
1233 case Instruction::Shl:
1234 case Instruction::Mul: {
1235 Value *Op0 = I->getOperand(0);
1236 Value *Op1 = I->getOperand(1);
1238 if (I->getOpcode() == Instruction::Shl) {
1239 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1240 if (!Op1CI) return false;
1241 // Turn Op0 << Op1 into Op0 * 2^Op1
1242 APInt Op1Int = Op1CI->getValue();
1243 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1244 APInt API(Op1Int.getBitWidth(), 0);
1245 API.setBit(BitToSet);
1246 Op1 = ConstantInt::get(V->getContext(), API);
1250 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1251 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1252 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1253 if (Op1C->getType()->getPrimitiveSizeInBits() <
1254 MulC->getType()->getPrimitiveSizeInBits())
1255 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1256 if (Op1C->getType()->getPrimitiveSizeInBits() >
1257 MulC->getType()->getPrimitiveSizeInBits())
1258 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1260 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1261 Multiple = ConstantExpr::getMul(MulC, Op1C);
1265 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1266 if (Mul0CI->getValue() == 1) {
1267 // V == Base * Op1, so return Op1
1274 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1275 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1276 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1277 if (Op0C->getType()->getPrimitiveSizeInBits() <
1278 MulC->getType()->getPrimitiveSizeInBits())
1279 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1280 if (Op0C->getType()->getPrimitiveSizeInBits() >
1281 MulC->getType()->getPrimitiveSizeInBits())
1282 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1284 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1285 Multiple = ConstantExpr::getMul(MulC, Op0C);
1289 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1290 if (Mul1CI->getValue() == 1) {
1291 // V == Base * Op0, so return Op0
1299 // We could not determine if V is a multiple of Base.
1303 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1304 /// value is never equal to -0.0.
1306 /// NOTE: this function will need to be revisited when we support non-default
1309 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1310 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1311 return !CFP->getValueAPF().isNegZero();
1314 return 1; // Limit search depth.
1316 const Operator *I = dyn_cast<Operator>(V);
1317 if (I == 0) return false;
1319 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1320 if (I->getOpcode() == Instruction::FAdd &&
1321 isa<ConstantFP>(I->getOperand(1)) &&
1322 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1325 // sitofp and uitofp turn into +0.0 for zero.
1326 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1329 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1330 // sqrt(-0.0) = -0.0, no other negative results are possible.
1331 if (II->getIntrinsicID() == Intrinsic::sqrt)
1332 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1334 if (const CallInst *CI = dyn_cast<CallInst>(I))
1335 if (const Function *F = CI->getCalledFunction()) {
1336 if (F->isDeclaration()) {
1338 if (F->getName() == "abs") return true;
1339 // fabs[lf](x) != -0.0
1340 if (F->getName() == "fabs") return true;
1341 if (F->getName() == "fabsf") return true;
1342 if (F->getName() == "fabsl") return true;
1343 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1344 F->getName() == "sqrtl")
1345 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1352 /// isBytewiseValue - If the specified value can be set by repeating the same
1353 /// byte in memory, return the i8 value that it is represented with. This is
1354 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1355 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1356 /// byte store (e.g. i16 0x1234), return null.
1357 Value *llvm::isBytewiseValue(Value *V) {
1358 // All byte-wide stores are splatable, even of arbitrary variables.
1359 if (V->getType()->isIntegerTy(8)) return V;
1361 // Handle 'null' ConstantArrayZero etc.
1362 if (Constant *C = dyn_cast<Constant>(V))
1363 if (C->isNullValue())
1364 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1366 // Constant float and double values can be handled as integer values if the
1367 // corresponding integer value is "byteable". An important case is 0.0.
1368 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1369 if (CFP->getType()->isFloatTy())
1370 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1371 if (CFP->getType()->isDoubleTy())
1372 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1373 // Don't handle long double formats, which have strange constraints.
1376 // We can handle constant integers that are power of two in size and a
1377 // multiple of 8 bits.
1378 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1379 unsigned Width = CI->getBitWidth();
1380 if (isPowerOf2_32(Width) && Width > 8) {
1381 // We can handle this value if the recursive binary decomposition is the
1382 // same at all levels.
1383 APInt Val = CI->getValue();
1385 while (Val.getBitWidth() != 8) {
1386 unsigned NextWidth = Val.getBitWidth()/2;
1387 Val2 = Val.lshr(NextWidth);
1388 Val2 = Val2.trunc(Val.getBitWidth()/2);
1389 Val = Val.trunc(Val.getBitWidth()/2);
1391 // If the top/bottom halves aren't the same, reject it.
1395 return ConstantInt::get(V->getContext(), Val);
1399 // A ConstantDataArray/Vector is splatable if all its members are equal and
1401 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1402 Value *Elt = CA->getElementAsConstant(0);
1403 Value *Val = isBytewiseValue(Elt);
1407 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1408 if (CA->getElementAsConstant(I) != Elt)
1414 // Conceptually, we could handle things like:
1415 // %a = zext i8 %X to i16
1416 // %b = shl i16 %a, 8
1417 // %c = or i16 %a, %b
1418 // but until there is an example that actually needs this, it doesn't seem
1419 // worth worrying about.
1424 // This is the recursive version of BuildSubAggregate. It takes a few different
1425 // arguments. Idxs is the index within the nested struct From that we are
1426 // looking at now (which is of type IndexedType). IdxSkip is the number of
1427 // indices from Idxs that should be left out when inserting into the resulting
1428 // struct. To is the result struct built so far, new insertvalue instructions
1430 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1431 SmallVector<unsigned, 10> &Idxs,
1433 Instruction *InsertBefore) {
1434 llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1436 // Save the original To argument so we can modify it
1438 // General case, the type indexed by Idxs is a struct
1439 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1440 // Process each struct element recursively
1443 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1447 // Couldn't find any inserted value for this index? Cleanup
1448 while (PrevTo != OrigTo) {
1449 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1450 PrevTo = Del->getAggregateOperand();
1451 Del->eraseFromParent();
1453 // Stop processing elements
1457 // If we successfully found a value for each of our subaggregates
1461 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1462 // the struct's elements had a value that was inserted directly. In the latter
1463 // case, perhaps we can't determine each of the subelements individually, but
1464 // we might be able to find the complete struct somewhere.
1466 // Find the value that is at that particular spot
1467 Value *V = FindInsertedValue(From, Idxs);
1472 // Insert the value in the new (sub) aggregrate
1473 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1474 "tmp", InsertBefore);
1477 // This helper takes a nested struct and extracts a part of it (which is again a
1478 // struct) into a new value. For example, given the struct:
1479 // { a, { b, { c, d }, e } }
1480 // and the indices "1, 1" this returns
1483 // It does this by inserting an insertvalue for each element in the resulting
1484 // struct, as opposed to just inserting a single struct. This will only work if
1485 // each of the elements of the substruct are known (ie, inserted into From by an
1486 // insertvalue instruction somewhere).
1488 // All inserted insertvalue instructions are inserted before InsertBefore
1489 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1490 Instruction *InsertBefore) {
1491 assert(InsertBefore && "Must have someplace to insert!");
1492 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1494 Value *To = UndefValue::get(IndexedType);
1495 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1496 unsigned IdxSkip = Idxs.size();
1498 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1501 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1502 /// the scalar value indexed is already around as a register, for example if it
1503 /// were inserted directly into the aggregrate.
1505 /// If InsertBefore is not null, this function will duplicate (modified)
1506 /// insertvalues when a part of a nested struct is extracted.
1507 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1508 Instruction *InsertBefore) {
1509 // Nothing to index? Just return V then (this is useful at the end of our
1511 if (idx_range.empty())
1513 // We have indices, so V should have an indexable type.
1514 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1515 "Not looking at a struct or array?");
1516 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1517 "Invalid indices for type?");
1519 if (Constant *C = dyn_cast<Constant>(V)) {
1520 C = C->getAggregateElement(idx_range[0]);
1521 if (C == 0) return 0;
1522 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1525 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1526 // Loop the indices for the insertvalue instruction in parallel with the
1527 // requested indices
1528 const unsigned *req_idx = idx_range.begin();
1529 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1530 i != e; ++i, ++req_idx) {
1531 if (req_idx == idx_range.end()) {
1532 // We can't handle this without inserting insertvalues
1536 // The requested index identifies a part of a nested aggregate. Handle
1537 // this specially. For example,
1538 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1539 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1540 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1541 // This can be changed into
1542 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1543 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1544 // which allows the unused 0,0 element from the nested struct to be
1546 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1550 // This insert value inserts something else than what we are looking for.
1551 // See if the (aggregrate) value inserted into has the value we are
1552 // looking for, then.
1554 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1557 // If we end up here, the indices of the insertvalue match with those
1558 // requested (though possibly only partially). Now we recursively look at
1559 // the inserted value, passing any remaining indices.
1560 return FindInsertedValue(I->getInsertedValueOperand(),
1561 makeArrayRef(req_idx, idx_range.end()),
1565 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1566 // If we're extracting a value from an aggregrate that was extracted from
1567 // something else, we can extract from that something else directly instead.
1568 // However, we will need to chain I's indices with the requested indices.
1570 // Calculate the number of indices required
1571 unsigned size = I->getNumIndices() + idx_range.size();
1572 // Allocate some space to put the new indices in
1573 SmallVector<unsigned, 5> Idxs;
1575 // Add indices from the extract value instruction
1576 Idxs.append(I->idx_begin(), I->idx_end());
1578 // Add requested indices
1579 Idxs.append(idx_range.begin(), idx_range.end());
1581 assert(Idxs.size() == size
1582 && "Number of indices added not correct?");
1584 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1586 // Otherwise, we don't know (such as, extracting from a function return value
1587 // or load instruction)
1591 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1592 /// it can be expressed as a base pointer plus a constant offset. Return the
1593 /// base and offset to the caller.
1594 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1595 const DataLayout &TD) {
1596 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1597 if (PtrOp == 0 || Ptr->getType()->isVectorTy())
1600 // Just look through bitcasts.
1601 if (PtrOp->getOpcode() == Instruction::BitCast)
1602 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1604 // If this is a GEP with constant indices, we can look through it.
1605 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1606 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1608 gep_type_iterator GTI = gep_type_begin(GEP);
1609 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1611 ConstantInt *OpC = cast<ConstantInt>(*I);
1612 if (OpC->isZero()) continue;
1614 // Handle a struct and array indices which add their offset to the pointer.
1615 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1616 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1618 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1619 Offset += OpC->getSExtValue()*Size;
1623 // Re-sign extend from the pointer size if needed to get overflow edge cases
1625 unsigned AS = GEP->getPointerAddressSpace();
1626 unsigned PtrSize = TD.getPointerSizeInBits(AS);
1628 Offset = SignExtend64(Offset, PtrSize);
1630 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1634 /// getConstantStringInfo - This function computes the length of a
1635 /// null-terminated C string pointed to by V. If successful, it returns true
1636 /// and returns the string in Str. If unsuccessful, it returns false.
1637 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
1638 uint64_t Offset, bool TrimAtNul) {
1641 // Look through bitcast instructions and geps.
1642 V = V->stripPointerCasts();
1644 // If the value is a GEP instructionor constant expression, treat it as an
1646 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1647 // Make sure the GEP has exactly three arguments.
1648 if (GEP->getNumOperands() != 3)
1651 // Make sure the index-ee is a pointer to array of i8.
1652 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1653 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1654 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1657 // Check to make sure that the first operand of the GEP is an integer and
1658 // has value 0 so that we are sure we're indexing into the initializer.
1659 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1660 if (FirstIdx == 0 || !FirstIdx->isZero())
1663 // If the second index isn't a ConstantInt, then this is a variable index
1664 // into the array. If this occurs, we can't say anything meaningful about
1666 uint64_t StartIdx = 0;
1667 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1668 StartIdx = CI->getZExtValue();
1671 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
1674 // The GEP instruction, constant or instruction, must reference a global
1675 // variable that is a constant and is initialized. The referenced constant
1676 // initializer is the array that we'll use for optimization.
1677 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
1678 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1681 // Handle the all-zeros case
1682 if (GV->getInitializer()->isNullValue()) {
1683 // This is a degenerate case. The initializer is constant zero so the
1684 // length of the string must be zero.
1689 // Must be a Constant Array
1690 const ConstantDataArray *Array =
1691 dyn_cast<ConstantDataArray>(GV->getInitializer());
1692 if (Array == 0 || !Array->isString())
1695 // Get the number of elements in the array
1696 uint64_t NumElts = Array->getType()->getArrayNumElements();
1698 // Start out with the entire array in the StringRef.
1699 Str = Array->getAsString();
1701 if (Offset > NumElts)
1704 // Skip over 'offset' bytes.
1705 Str = Str.substr(Offset);
1708 // Trim off the \0 and anything after it. If the array is not nul
1709 // terminated, we just return the whole end of string. The client may know
1710 // some other way that the string is length-bound.
1711 Str = Str.substr(0, Str.find('\0'));
1716 // These next two are very similar to the above, but also look through PHI
1718 // TODO: See if we can integrate these two together.
1720 /// GetStringLengthH - If we can compute the length of the string pointed to by
1721 /// the specified pointer, return 'len+1'. If we can't, return 0.
1722 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1723 // Look through noop bitcast instructions.
1724 V = V->stripPointerCasts();
1726 // If this is a PHI node, there are two cases: either we have already seen it
1728 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1729 if (!PHIs.insert(PN))
1730 return ~0ULL; // already in the set.
1732 // If it was new, see if all the input strings are the same length.
1733 uint64_t LenSoFar = ~0ULL;
1734 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1735 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1736 if (Len == 0) return 0; // Unknown length -> unknown.
1738 if (Len == ~0ULL) continue;
1740 if (Len != LenSoFar && LenSoFar != ~0ULL)
1741 return 0; // Disagree -> unknown.
1745 // Success, all agree.
1749 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1750 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1751 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1752 if (Len1 == 0) return 0;
1753 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1754 if (Len2 == 0) return 0;
1755 if (Len1 == ~0ULL) return Len2;
1756 if (Len2 == ~0ULL) return Len1;
1757 if (Len1 != Len2) return 0;
1761 // Otherwise, see if we can read the string.
1763 if (!getConstantStringInfo(V, StrData))
1766 return StrData.size()+1;
1769 /// GetStringLength - If we can compute the length of the string pointed to by
1770 /// the specified pointer, return 'len+1'. If we can't, return 0.
1771 uint64_t llvm::GetStringLength(Value *V) {
1772 if (!V->getType()->isPointerTy()) return 0;
1774 SmallPtrSet<PHINode*, 32> PHIs;
1775 uint64_t Len = GetStringLengthH(V, PHIs);
1776 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1777 // an empty string as a length.
1778 return Len == ~0ULL ? 1 : Len;
1782 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
1783 if (!V->getType()->isPointerTy())
1785 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1786 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1787 V = GEP->getPointerOperand();
1788 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1789 V = cast<Operator>(V)->getOperand(0);
1790 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1791 if (GA->mayBeOverridden())
1793 V = GA->getAliasee();
1795 // See if InstructionSimplify knows any relevant tricks.
1796 if (Instruction *I = dyn_cast<Instruction>(V))
1797 // TODO: Acquire a DominatorTree and use it.
1798 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1805 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1811 llvm::GetUnderlyingObjects(Value *V,
1812 SmallVectorImpl<Value *> &Objects,
1813 const DataLayout *TD,
1814 unsigned MaxLookup) {
1815 SmallPtrSet<Value *, 4> Visited;
1816 SmallVector<Value *, 4> Worklist;
1817 Worklist.push_back(V);
1819 Value *P = Worklist.pop_back_val();
1820 P = GetUnderlyingObject(P, TD, MaxLookup);
1822 if (!Visited.insert(P))
1825 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
1826 Worklist.push_back(SI->getTrueValue());
1827 Worklist.push_back(SI->getFalseValue());
1831 if (PHINode *PN = dyn_cast<PHINode>(P)) {
1832 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1833 Worklist.push_back(PN->getIncomingValue(i));
1837 Objects.push_back(P);
1838 } while (!Worklist.empty());
1841 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1842 /// are lifetime markers.
1844 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1845 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1847 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1848 if (!II) return false;
1850 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1851 II->getIntrinsicID() != Intrinsic::lifetime_end)
1857 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1858 const DataLayout *TD) {
1859 const Operator *Inst = dyn_cast<Operator>(V);
1863 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1864 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1868 switch (Inst->getOpcode()) {
1871 case Instruction::UDiv:
1872 case Instruction::URem:
1873 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1874 return isKnownNonZero(Inst->getOperand(1), TD);
1875 case Instruction::SDiv:
1876 case Instruction::SRem: {
1877 Value *Op = Inst->getOperand(1);
1878 // x / y is undefined if y == 0
1879 if (!isKnownNonZero(Op, TD))
1881 // x / y might be undefined if y == -1
1882 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1885 APInt KnownZero(BitWidth, 0);
1886 APInt KnownOne(BitWidth, 0);
1887 ComputeMaskedBits(Op, KnownZero, KnownOne, TD);
1890 case Instruction::Load: {
1891 const LoadInst *LI = cast<LoadInst>(Inst);
1892 if (!LI->isUnordered())
1894 return LI->getPointerOperand()->isDereferenceablePointer();
1896 case Instruction::Call: {
1897 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1898 switch (II->getIntrinsicID()) {
1899 // These synthetic intrinsics have no side-effects, and just mark
1900 // information about their operands.
1901 // FIXME: There are other no-op synthetic instructions that potentially
1902 // should be considered at least *safe* to speculate...
1903 case Intrinsic::dbg_declare:
1904 case Intrinsic::dbg_value:
1907 case Intrinsic::bswap:
1908 case Intrinsic::ctlz:
1909 case Intrinsic::ctpop:
1910 case Intrinsic::cttz:
1911 case Intrinsic::objectsize:
1912 case Intrinsic::sadd_with_overflow:
1913 case Intrinsic::smul_with_overflow:
1914 case Intrinsic::ssub_with_overflow:
1915 case Intrinsic::uadd_with_overflow:
1916 case Intrinsic::umul_with_overflow:
1917 case Intrinsic::usub_with_overflow:
1919 // TODO: some fp intrinsics are marked as having the same error handling
1920 // as libm. They're safe to speculate when they won't error.
1921 // TODO: are convert_{from,to}_fp16 safe?
1922 // TODO: can we list target-specific intrinsics here?
1926 return false; // The called function could have undefined behavior or
1927 // side-effects, even if marked readnone nounwind.
1929 case Instruction::VAArg:
1930 case Instruction::Alloca:
1931 case Instruction::Invoke:
1932 case Instruction::PHI:
1933 case Instruction::Store:
1934 case Instruction::Ret:
1935 case Instruction::Br:
1936 case Instruction::IndirectBr:
1937 case Instruction::Switch:
1938 case Instruction::Unreachable:
1939 case Instruction::Fence:
1940 case Instruction::LandingPad:
1941 case Instruction::AtomicRMW:
1942 case Instruction::AtomicCmpXchg:
1943 case Instruction::Resume:
1944 return false; // Misc instructions which have effects