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 = TD->getTypeSizeInBits(SrcTy->getScalarType());
442 assert(SrcBitWidth && "SrcBitWidth can't be zero");
443 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
444 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
445 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
446 KnownZero = KnownZero.zextOrTrunc(BitWidth);
447 KnownOne = KnownOne.zextOrTrunc(BitWidth);
448 // Any top bits are known to be zero.
449 if (BitWidth > SrcBitWidth)
450 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
453 case Instruction::BitCast: {
454 Type *SrcTy = I->getOperand(0)->getType();
455 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
456 // TODO: For now, not handling conversions like:
457 // (bitcast i64 %x to <2 x i32>)
458 !I->getType()->isVectorTy()) {
459 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
464 case Instruction::SExt: {
465 // Compute the bits in the result that are not present in the input.
466 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
468 KnownZero = KnownZero.trunc(SrcBitWidth);
469 KnownOne = KnownOne.trunc(SrcBitWidth);
470 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
471 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
472 KnownZero = KnownZero.zext(BitWidth);
473 KnownOne = KnownOne.zext(BitWidth);
475 // If the sign bit of the input is known set or clear, then we know the
476 // top bits of the result.
477 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
478 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
479 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
480 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
483 case Instruction::Shl:
484 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
485 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
486 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
487 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
488 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
489 KnownZero <<= ShiftAmt;
490 KnownOne <<= ShiftAmt;
491 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
495 case Instruction::LShr:
496 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
497 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
498 // Compute the new bits that are at the top now.
499 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
501 // Unsigned shift right.
502 ComputeMaskedBits(I->getOperand(0), KnownZero,KnownOne, TD, Depth+1);
503 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
504 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
505 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
506 // high bits known zero.
507 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
511 case Instruction::AShr:
512 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
513 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
514 // Compute the new bits that are at the top now.
515 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
517 // Signed shift right.
518 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
519 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
520 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
521 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
523 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
524 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
525 KnownZero |= HighBits;
526 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
527 KnownOne |= HighBits;
531 case Instruction::Sub: {
532 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
533 ComputeMaskedBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
534 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
538 case Instruction::Add: {
539 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
540 ComputeMaskedBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
541 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
545 case Instruction::SRem:
546 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
547 APInt RA = Rem->getValue().abs();
548 if (RA.isPowerOf2()) {
549 APInt LowBits = RA - 1;
550 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
552 // The low bits of the first operand are unchanged by the srem.
553 KnownZero = KnownZero2 & LowBits;
554 KnownOne = KnownOne2 & LowBits;
556 // If the first operand is non-negative or has all low bits zero, then
557 // the upper bits are all zero.
558 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
559 KnownZero |= ~LowBits;
561 // If the first operand is negative and not all low bits are zero, then
562 // the upper bits are all one.
563 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
564 KnownOne |= ~LowBits;
566 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
570 // The sign bit is the LHS's sign bit, except when the result of the
571 // remainder is zero.
572 if (KnownZero.isNonNegative()) {
573 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
574 ComputeMaskedBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
576 // If it's known zero, our sign bit is also zero.
577 if (LHSKnownZero.isNegative())
578 KnownZero.setBit(BitWidth - 1);
582 case Instruction::URem: {
583 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
584 APInt RA = Rem->getValue();
585 if (RA.isPowerOf2()) {
586 APInt LowBits = (RA - 1);
587 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD,
589 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
590 KnownZero |= ~LowBits;
596 // Since the result is less than or equal to either operand, any leading
597 // zero bits in either operand must also exist in the result.
598 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
599 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
601 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
602 KnownZero2.countLeadingOnes());
603 KnownOne.clearAllBits();
604 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
608 case Instruction::Alloca: {
609 AllocaInst *AI = cast<AllocaInst>(V);
610 unsigned Align = AI->getAlignment();
611 if (Align == 0 && TD)
612 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
615 KnownZero = APInt::getLowBitsSet(BitWidth, CountTrailingZeros_32(Align));
618 case Instruction::GetElementPtr: {
619 // Analyze all of the subscripts of this getelementptr instruction
620 // to determine if we can prove known low zero bits.
621 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
622 ComputeMaskedBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
624 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
626 gep_type_iterator GTI = gep_type_begin(I);
627 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
628 Value *Index = I->getOperand(i);
629 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
630 // Handle struct member offset arithmetic.
632 const StructLayout *SL = TD->getStructLayout(STy);
633 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
634 uint64_t Offset = SL->getElementOffset(Idx);
635 TrailZ = std::min(TrailZ,
636 CountTrailingZeros_64(Offset));
638 // Handle array index arithmetic.
639 Type *IndexedTy = GTI.getIndexedType();
640 if (!IndexedTy->isSized()) return;
641 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
642 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
643 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
644 ComputeMaskedBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1);
645 TrailZ = std::min(TrailZ,
646 unsigned(CountTrailingZeros_64(TypeSize) +
647 LocalKnownZero.countTrailingOnes()));
651 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
654 case Instruction::PHI: {
655 PHINode *P = cast<PHINode>(I);
656 // Handle the case of a simple two-predecessor recurrence PHI.
657 // There's a lot more that could theoretically be done here, but
658 // this is sufficient to catch some interesting cases.
659 if (P->getNumIncomingValues() == 2) {
660 for (unsigned i = 0; i != 2; ++i) {
661 Value *L = P->getIncomingValue(i);
662 Value *R = P->getIncomingValue(!i);
663 Operator *LU = dyn_cast<Operator>(L);
666 unsigned Opcode = LU->getOpcode();
667 // Check for operations that have the property that if
668 // both their operands have low zero bits, the result
669 // will have low zero bits.
670 if (Opcode == Instruction::Add ||
671 Opcode == Instruction::Sub ||
672 Opcode == Instruction::And ||
673 Opcode == Instruction::Or ||
674 Opcode == Instruction::Mul) {
675 Value *LL = LU->getOperand(0);
676 Value *LR = LU->getOperand(1);
677 // Find a recurrence.
684 // Ok, we have a PHI of the form L op= R. Check for low
686 ComputeMaskedBits(R, KnownZero2, KnownOne2, TD, Depth+1);
688 // We need to take the minimum number of known bits
689 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
690 ComputeMaskedBits(L, KnownZero3, KnownOne3, TD, Depth+1);
692 KnownZero = APInt::getLowBitsSet(BitWidth,
693 std::min(KnownZero2.countTrailingOnes(),
694 KnownZero3.countTrailingOnes()));
700 // Unreachable blocks may have zero-operand PHI nodes.
701 if (P->getNumIncomingValues() == 0)
704 // Otherwise take the unions of the known bit sets of the operands,
705 // taking conservative care to avoid excessive recursion.
706 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
707 // Skip if every incoming value references to ourself.
708 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
711 KnownZero = APInt::getAllOnesValue(BitWidth);
712 KnownOne = APInt::getAllOnesValue(BitWidth);
713 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
714 // Skip direct self references.
715 if (P->getIncomingValue(i) == P) continue;
717 KnownZero2 = APInt(BitWidth, 0);
718 KnownOne2 = APInt(BitWidth, 0);
719 // Recurse, but cap the recursion to one level, because we don't
720 // want to waste time spinning around in loops.
721 ComputeMaskedBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
723 KnownZero &= KnownZero2;
724 KnownOne &= KnownOne2;
725 // If all bits have been ruled out, there's no need to check
727 if (!KnownZero && !KnownOne)
733 case Instruction::Call:
734 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
735 switch (II->getIntrinsicID()) {
737 case Intrinsic::ctlz:
738 case Intrinsic::cttz: {
739 unsigned LowBits = Log2_32(BitWidth)+1;
740 // If this call is undefined for 0, the result will be less than 2^n.
741 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
743 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
746 case Intrinsic::ctpop: {
747 unsigned LowBits = Log2_32(BitWidth)+1;
748 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
751 case Intrinsic::x86_sse42_crc32_64_8:
752 case Intrinsic::x86_sse42_crc32_64_64:
753 KnownZero = APInt::getHighBitsSet(64, 32);
758 case Instruction::ExtractValue:
759 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
760 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
761 if (EVI->getNumIndices() != 1) break;
762 if (EVI->getIndices()[0] == 0) {
763 switch (II->getIntrinsicID()) {
765 case Intrinsic::uadd_with_overflow:
766 case Intrinsic::sadd_with_overflow:
767 ComputeMaskedBitsAddSub(true, II->getArgOperand(0),
768 II->getArgOperand(1), false, KnownZero,
769 KnownOne, KnownZero2, KnownOne2, TD, Depth);
771 case Intrinsic::usub_with_overflow:
772 case Intrinsic::ssub_with_overflow:
773 ComputeMaskedBitsAddSub(false, II->getArgOperand(0),
774 II->getArgOperand(1), false, KnownZero,
775 KnownOne, KnownZero2, KnownOne2, TD, Depth);
777 case Intrinsic::umul_with_overflow:
778 case Intrinsic::smul_with_overflow:
779 ComputeMaskedBitsMul(II->getArgOperand(0), II->getArgOperand(1),
780 false, KnownZero, KnownOne,
781 KnownZero2, KnownOne2, TD, Depth);
789 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
790 /// one. Convenience wrapper around ComputeMaskedBits.
791 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
792 const DataLayout *TD, unsigned Depth) {
793 unsigned BitWidth = getBitWidth(V->getType(), TD);
799 APInt ZeroBits(BitWidth, 0);
800 APInt OneBits(BitWidth, 0);
801 ComputeMaskedBits(V, ZeroBits, OneBits, TD, Depth);
802 KnownOne = OneBits[BitWidth - 1];
803 KnownZero = ZeroBits[BitWidth - 1];
806 /// isPowerOfTwo - Return true if the given value is known to have exactly one
807 /// bit set when defined. For vectors return true if every element is known to
808 /// be a power of two when defined. Supports values with integer or pointer
809 /// types and vectors of integers.
810 bool llvm::isPowerOfTwo(Value *V, const DataLayout *TD, bool OrZero,
812 if (Constant *C = dyn_cast<Constant>(V)) {
813 if (C->isNullValue())
815 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
816 return CI->getValue().isPowerOf2();
817 // TODO: Handle vector constants.
820 // 1 << X is clearly a power of two if the one is not shifted off the end. If
821 // it is shifted off the end then the result is undefined.
822 if (match(V, m_Shl(m_One(), m_Value())))
825 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
826 // bottom. If it is shifted off the bottom then the result is undefined.
827 if (match(V, m_LShr(m_SignBit(), m_Value())))
830 // The remaining tests are all recursive, so bail out if we hit the limit.
831 if (Depth++ == MaxDepth)
834 Value *X = 0, *Y = 0;
835 // A shift of a power of two is a power of two or zero.
836 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
837 match(V, m_Shr(m_Value(X), m_Value()))))
838 return isPowerOfTwo(X, TD, /*OrZero*/true, Depth);
840 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
841 return isPowerOfTwo(ZI->getOperand(0), TD, OrZero, Depth);
843 if (SelectInst *SI = dyn_cast<SelectInst>(V))
844 return isPowerOfTwo(SI->getTrueValue(), TD, OrZero, Depth) &&
845 isPowerOfTwo(SI->getFalseValue(), TD, OrZero, Depth);
847 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
848 // A power of two and'd with anything is a power of two or zero.
849 if (isPowerOfTwo(X, TD, /*OrZero*/true, Depth) ||
850 isPowerOfTwo(Y, TD, /*OrZero*/true, Depth))
852 // X & (-X) is always a power of two or zero.
853 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
858 // An exact divide or right shift can only shift off zero bits, so the result
859 // is a power of two only if the first operand is a power of two and not
860 // copying a sign bit (sdiv int_min, 2).
861 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
862 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
863 return isPowerOfTwo(cast<Operator>(V)->getOperand(0), TD, OrZero, Depth);
869 /// isKnownNonZero - Return true if the given value is known to be non-zero
870 /// when defined. For vectors return true if every element is known to be
871 /// non-zero when defined. Supports values with integer or pointer type and
872 /// vectors of integers.
873 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth) {
874 if (Constant *C = dyn_cast<Constant>(V)) {
875 if (C->isNullValue())
877 if (isa<ConstantInt>(C))
878 // Must be non-zero due to null test above.
880 // TODO: Handle vectors
884 // The remaining tests are all recursive, so bail out if we hit the limit.
885 if (Depth++ >= MaxDepth)
888 unsigned BitWidth = getBitWidth(V->getType(), TD);
890 // X | Y != 0 if X != 0 or Y != 0.
891 Value *X = 0, *Y = 0;
892 if (match(V, m_Or(m_Value(X), m_Value(Y))))
893 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
895 // ext X != 0 if X != 0.
896 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
897 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
899 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
900 // if the lowest bit is shifted off the end.
901 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
902 // shl nuw can't remove any non-zero bits.
903 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
904 if (BO->hasNoUnsignedWrap())
905 return isKnownNonZero(X, TD, Depth);
907 APInt KnownZero(BitWidth, 0);
908 APInt KnownOne(BitWidth, 0);
909 ComputeMaskedBits(X, KnownZero, KnownOne, TD, Depth);
913 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
914 // defined if the sign bit is shifted off the end.
915 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
916 // shr exact can only shift out zero bits.
917 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
919 return isKnownNonZero(X, TD, Depth);
921 bool XKnownNonNegative, XKnownNegative;
922 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
926 // div exact can only produce a zero if the dividend is zero.
927 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
928 return isKnownNonZero(X, TD, Depth);
931 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
932 bool XKnownNonNegative, XKnownNegative;
933 bool YKnownNonNegative, YKnownNegative;
934 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
935 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
937 // If X and Y are both non-negative (as signed values) then their sum is not
938 // zero unless both X and Y are zero.
939 if (XKnownNonNegative && YKnownNonNegative)
940 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
943 // If X and Y are both negative (as signed values) then their sum is not
944 // zero unless both X and Y equal INT_MIN.
945 if (BitWidth && XKnownNegative && YKnownNegative) {
946 APInt KnownZero(BitWidth, 0);
947 APInt KnownOne(BitWidth, 0);
948 APInt Mask = APInt::getSignedMaxValue(BitWidth);
949 // The sign bit of X is set. If some other bit is set then X is not equal
951 ComputeMaskedBits(X, KnownZero, KnownOne, TD, Depth);
952 if ((KnownOne & Mask) != 0)
954 // The sign bit of Y is set. If some other bit is set then Y is not equal
956 ComputeMaskedBits(Y, KnownZero, KnownOne, TD, Depth);
957 if ((KnownOne & Mask) != 0)
961 // The sum of a non-negative number and a power of two is not zero.
962 if (XKnownNonNegative && isPowerOfTwo(Y, TD, /*OrZero*/false, Depth))
964 if (YKnownNonNegative && isPowerOfTwo(X, TD, /*OrZero*/false, Depth))
968 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
969 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
970 // If X and Y are non-zero then so is X * Y as long as the multiplication
971 // does not overflow.
972 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
973 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
976 // (C ? X : Y) != 0 if X != 0 and Y != 0.
977 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
978 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
979 isKnownNonZero(SI->getFalseValue(), TD, Depth))
983 if (!BitWidth) return false;
984 APInt KnownZero(BitWidth, 0);
985 APInt KnownOne(BitWidth, 0);
986 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
987 return KnownOne != 0;
990 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
991 /// this predicate to simplify operations downstream. Mask is known to be zero
992 /// for bits that V cannot have.
994 /// This function is defined on values with integer type, values with pointer
995 /// type (but only if TD is non-null), and vectors of integers. In the case
996 /// where V is a vector, the mask, known zero, and known one values are the
997 /// same width as the vector element, and the bit is set only if it is true
998 /// for all of the elements in the vector.
999 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
1000 const DataLayout *TD, unsigned Depth) {
1001 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1002 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1003 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1004 return (KnownZero & Mask) == Mask;
1009 /// ComputeNumSignBits - Return the number of times the sign bit of the
1010 /// register is replicated into the other bits. We know that at least 1 bit
1011 /// is always equal to the sign bit (itself), but other cases can give us
1012 /// information. For example, immediately after an "ashr X, 2", we know that
1013 /// the top 3 bits are all equal to each other, so we return 3.
1015 /// 'Op' must have a scalar integer type.
1017 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
1019 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1020 "ComputeNumSignBits requires a DataLayout object to operate "
1021 "on non-integer values!");
1022 Type *Ty = V->getType();
1023 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1024 Ty->getScalarSizeInBits();
1026 unsigned FirstAnswer = 1;
1028 // Note that ConstantInt is handled by the general ComputeMaskedBits case
1032 return 1; // Limit search depth.
1034 Operator *U = dyn_cast<Operator>(V);
1035 switch (Operator::getOpcode(V)) {
1037 case Instruction::SExt:
1038 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1039 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
1041 case Instruction::AShr: {
1042 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1043 // ashr X, C -> adds C sign bits. Vectors too.
1045 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1046 Tmp += ShAmt->getZExtValue();
1047 if (Tmp > TyBits) Tmp = TyBits;
1051 case Instruction::Shl: {
1053 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1054 // shl destroys sign bits.
1055 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1056 Tmp2 = ShAmt->getZExtValue();
1057 if (Tmp2 >= TyBits || // Bad shift.
1058 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1063 case Instruction::And:
1064 case Instruction::Or:
1065 case Instruction::Xor: // NOT is handled here.
1066 // Logical binary ops preserve the number of sign bits at the worst.
1067 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1069 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1070 FirstAnswer = std::min(Tmp, Tmp2);
1071 // We computed what we know about the sign bits as our first
1072 // answer. Now proceed to the generic code that uses
1073 // ComputeMaskedBits, and pick whichever answer is better.
1077 case Instruction::Select:
1078 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1079 if (Tmp == 1) return 1; // Early out.
1080 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1081 return std::min(Tmp, Tmp2);
1083 case Instruction::Add:
1084 // Add can have at most one carry bit. Thus we know that the output
1085 // is, at worst, one more bit than the inputs.
1086 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1087 if (Tmp == 1) return 1; // Early out.
1089 // Special case decrementing a value (ADD X, -1):
1090 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1091 if (CRHS->isAllOnesValue()) {
1092 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1093 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
1095 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1097 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1100 // If we are subtracting one from a positive number, there is no carry
1101 // out of the result.
1102 if (KnownZero.isNegative())
1106 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1107 if (Tmp2 == 1) return 1;
1108 return std::min(Tmp, Tmp2)-1;
1110 case Instruction::Sub:
1111 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1112 if (Tmp2 == 1) return 1;
1115 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1116 if (CLHS->isNullValue()) {
1117 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1118 ComputeMaskedBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
1119 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1121 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1124 // If the input is known to be positive (the sign bit is known clear),
1125 // the output of the NEG has the same number of sign bits as the input.
1126 if (KnownZero.isNegative())
1129 // Otherwise, we treat this like a SUB.
1132 // Sub can have at most one carry bit. Thus we know that the output
1133 // is, at worst, one more bit than the inputs.
1134 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1135 if (Tmp == 1) return 1; // Early out.
1136 return std::min(Tmp, Tmp2)-1;
1138 case Instruction::PHI: {
1139 PHINode *PN = cast<PHINode>(U);
1140 // Don't analyze large in-degree PHIs.
1141 if (PN->getNumIncomingValues() > 4) break;
1143 // Take the minimum of all incoming values. This can't infinitely loop
1144 // because of our depth threshold.
1145 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1146 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1147 if (Tmp == 1) return Tmp;
1149 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1154 case Instruction::Trunc:
1155 // FIXME: it's tricky to do anything useful for this, but it is an important
1156 // case for targets like X86.
1160 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1161 // use this information.
1162 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1164 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1166 if (KnownZero.isNegative()) { // sign bit is 0
1168 } else if (KnownOne.isNegative()) { // sign bit is 1;
1175 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1176 // the number of identical bits in the top of the input value.
1178 Mask <<= Mask.getBitWidth()-TyBits;
1179 // Return # leading zeros. We use 'min' here in case Val was zero before
1180 // shifting. We don't want to return '64' as for an i32 "0".
1181 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1184 /// ComputeMultiple - This function computes the integer multiple of Base that
1185 /// equals V. If successful, it returns true and returns the multiple in
1186 /// Multiple. If unsuccessful, it returns false. It looks
1187 /// through SExt instructions only if LookThroughSExt is true.
1188 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1189 bool LookThroughSExt, unsigned Depth) {
1190 const unsigned MaxDepth = 6;
1192 assert(V && "No Value?");
1193 assert(Depth <= MaxDepth && "Limit Search Depth");
1194 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1196 Type *T = V->getType();
1198 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1208 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1209 Constant *BaseVal = ConstantInt::get(T, Base);
1210 if (CO && CO == BaseVal) {
1212 Multiple = ConstantInt::get(T, 1);
1216 if (CI && CI->getZExtValue() % Base == 0) {
1217 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1221 if (Depth == MaxDepth) return false; // Limit search depth.
1223 Operator *I = dyn_cast<Operator>(V);
1224 if (!I) return false;
1226 switch (I->getOpcode()) {
1228 case Instruction::SExt:
1229 if (!LookThroughSExt) return false;
1230 // otherwise fall through to ZExt
1231 case Instruction::ZExt:
1232 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1233 LookThroughSExt, Depth+1);
1234 case Instruction::Shl:
1235 case Instruction::Mul: {
1236 Value *Op0 = I->getOperand(0);
1237 Value *Op1 = I->getOperand(1);
1239 if (I->getOpcode() == Instruction::Shl) {
1240 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1241 if (!Op1CI) return false;
1242 // Turn Op0 << Op1 into Op0 * 2^Op1
1243 APInt Op1Int = Op1CI->getValue();
1244 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1245 APInt API(Op1Int.getBitWidth(), 0);
1246 API.setBit(BitToSet);
1247 Op1 = ConstantInt::get(V->getContext(), API);
1251 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1252 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1253 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1254 if (Op1C->getType()->getPrimitiveSizeInBits() <
1255 MulC->getType()->getPrimitiveSizeInBits())
1256 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1257 if (Op1C->getType()->getPrimitiveSizeInBits() >
1258 MulC->getType()->getPrimitiveSizeInBits())
1259 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1261 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1262 Multiple = ConstantExpr::getMul(MulC, Op1C);
1266 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1267 if (Mul0CI->getValue() == 1) {
1268 // V == Base * Op1, so return Op1
1275 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1276 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1277 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1278 if (Op0C->getType()->getPrimitiveSizeInBits() <
1279 MulC->getType()->getPrimitiveSizeInBits())
1280 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1281 if (Op0C->getType()->getPrimitiveSizeInBits() >
1282 MulC->getType()->getPrimitiveSizeInBits())
1283 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1285 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1286 Multiple = ConstantExpr::getMul(MulC, Op0C);
1290 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1291 if (Mul1CI->getValue() == 1) {
1292 // V == Base * Op0, so return Op0
1300 // We could not determine if V is a multiple of Base.
1304 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1305 /// value is never equal to -0.0.
1307 /// NOTE: this function will need to be revisited when we support non-default
1310 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1311 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1312 return !CFP->getValueAPF().isNegZero();
1315 return 1; // Limit search depth.
1317 const Operator *I = dyn_cast<Operator>(V);
1318 if (I == 0) return false;
1320 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1321 if (I->getOpcode() == Instruction::FAdd &&
1322 isa<ConstantFP>(I->getOperand(1)) &&
1323 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1326 // sitofp and uitofp turn into +0.0 for zero.
1327 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1330 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1331 // sqrt(-0.0) = -0.0, no other negative results are possible.
1332 if (II->getIntrinsicID() == Intrinsic::sqrt)
1333 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1335 if (const CallInst *CI = dyn_cast<CallInst>(I))
1336 if (const Function *F = CI->getCalledFunction()) {
1337 if (F->isDeclaration()) {
1339 if (F->getName() == "abs") return true;
1340 // fabs[lf](x) != -0.0
1341 if (F->getName() == "fabs") return true;
1342 if (F->getName() == "fabsf") return true;
1343 if (F->getName() == "fabsl") return true;
1344 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1345 F->getName() == "sqrtl")
1346 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1353 /// isBytewiseValue - If the specified value can be set by repeating the same
1354 /// byte in memory, return the i8 value that it is represented with. This is
1355 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1356 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1357 /// byte store (e.g. i16 0x1234), return null.
1358 Value *llvm::isBytewiseValue(Value *V) {
1359 // All byte-wide stores are splatable, even of arbitrary variables.
1360 if (V->getType()->isIntegerTy(8)) return V;
1362 // Handle 'null' ConstantArrayZero etc.
1363 if (Constant *C = dyn_cast<Constant>(V))
1364 if (C->isNullValue())
1365 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1367 // Constant float and double values can be handled as integer values if the
1368 // corresponding integer value is "byteable". An important case is 0.0.
1369 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1370 if (CFP->getType()->isFloatTy())
1371 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1372 if (CFP->getType()->isDoubleTy())
1373 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1374 // Don't handle long double formats, which have strange constraints.
1377 // We can handle constant integers that are power of two in size and a
1378 // multiple of 8 bits.
1379 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1380 unsigned Width = CI->getBitWidth();
1381 if (isPowerOf2_32(Width) && Width > 8) {
1382 // We can handle this value if the recursive binary decomposition is the
1383 // same at all levels.
1384 APInt Val = CI->getValue();
1386 while (Val.getBitWidth() != 8) {
1387 unsigned NextWidth = Val.getBitWidth()/2;
1388 Val2 = Val.lshr(NextWidth);
1389 Val2 = Val2.trunc(Val.getBitWidth()/2);
1390 Val = Val.trunc(Val.getBitWidth()/2);
1392 // If the top/bottom halves aren't the same, reject it.
1396 return ConstantInt::get(V->getContext(), Val);
1400 // A ConstantDataArray/Vector is splatable if all its members are equal and
1402 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1403 Value *Elt = CA->getElementAsConstant(0);
1404 Value *Val = isBytewiseValue(Elt);
1408 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1409 if (CA->getElementAsConstant(I) != Elt)
1415 // Conceptually, we could handle things like:
1416 // %a = zext i8 %X to i16
1417 // %b = shl i16 %a, 8
1418 // %c = or i16 %a, %b
1419 // but until there is an example that actually needs this, it doesn't seem
1420 // worth worrying about.
1425 // This is the recursive version of BuildSubAggregate. It takes a few different
1426 // arguments. Idxs is the index within the nested struct From that we are
1427 // looking at now (which is of type IndexedType). IdxSkip is the number of
1428 // indices from Idxs that should be left out when inserting into the resulting
1429 // struct. To is the result struct built so far, new insertvalue instructions
1431 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1432 SmallVector<unsigned, 10> &Idxs,
1434 Instruction *InsertBefore) {
1435 llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1437 // Save the original To argument so we can modify it
1439 // General case, the type indexed by Idxs is a struct
1440 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1441 // Process each struct element recursively
1444 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1448 // Couldn't find any inserted value for this index? Cleanup
1449 while (PrevTo != OrigTo) {
1450 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1451 PrevTo = Del->getAggregateOperand();
1452 Del->eraseFromParent();
1454 // Stop processing elements
1458 // If we successfully found a value for each of our subaggregates
1462 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1463 // the struct's elements had a value that was inserted directly. In the latter
1464 // case, perhaps we can't determine each of the subelements individually, but
1465 // we might be able to find the complete struct somewhere.
1467 // Find the value that is at that particular spot
1468 Value *V = FindInsertedValue(From, Idxs);
1473 // Insert the value in the new (sub) aggregrate
1474 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1475 "tmp", InsertBefore);
1478 // This helper takes a nested struct and extracts a part of it (which is again a
1479 // struct) into a new value. For example, given the struct:
1480 // { a, { b, { c, d }, e } }
1481 // and the indices "1, 1" this returns
1484 // It does this by inserting an insertvalue for each element in the resulting
1485 // struct, as opposed to just inserting a single struct. This will only work if
1486 // each of the elements of the substruct are known (ie, inserted into From by an
1487 // insertvalue instruction somewhere).
1489 // All inserted insertvalue instructions are inserted before InsertBefore
1490 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1491 Instruction *InsertBefore) {
1492 assert(InsertBefore && "Must have someplace to insert!");
1493 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1495 Value *To = UndefValue::get(IndexedType);
1496 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1497 unsigned IdxSkip = Idxs.size();
1499 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1502 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1503 /// the scalar value indexed is already around as a register, for example if it
1504 /// were inserted directly into the aggregrate.
1506 /// If InsertBefore is not null, this function will duplicate (modified)
1507 /// insertvalues when a part of a nested struct is extracted.
1508 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1509 Instruction *InsertBefore) {
1510 // Nothing to index? Just return V then (this is useful at the end of our
1512 if (idx_range.empty())
1514 // We have indices, so V should have an indexable type.
1515 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1516 "Not looking at a struct or array?");
1517 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1518 "Invalid indices for type?");
1520 if (Constant *C = dyn_cast<Constant>(V)) {
1521 C = C->getAggregateElement(idx_range[0]);
1522 if (C == 0) return 0;
1523 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1526 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1527 // Loop the indices for the insertvalue instruction in parallel with the
1528 // requested indices
1529 const unsigned *req_idx = idx_range.begin();
1530 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1531 i != e; ++i, ++req_idx) {
1532 if (req_idx == idx_range.end()) {
1533 // We can't handle this without inserting insertvalues
1537 // The requested index identifies a part of a nested aggregate. Handle
1538 // this specially. For example,
1539 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1540 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1541 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1542 // This can be changed into
1543 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1544 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1545 // which allows the unused 0,0 element from the nested struct to be
1547 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1551 // This insert value inserts something else than what we are looking for.
1552 // See if the (aggregrate) value inserted into has the value we are
1553 // looking for, then.
1555 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1558 // If we end up here, the indices of the insertvalue match with those
1559 // requested (though possibly only partially). Now we recursively look at
1560 // the inserted value, passing any remaining indices.
1561 return FindInsertedValue(I->getInsertedValueOperand(),
1562 makeArrayRef(req_idx, idx_range.end()),
1566 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1567 // If we're extracting a value from an aggregrate that was extracted from
1568 // something else, we can extract from that something else directly instead.
1569 // However, we will need to chain I's indices with the requested indices.
1571 // Calculate the number of indices required
1572 unsigned size = I->getNumIndices() + idx_range.size();
1573 // Allocate some space to put the new indices in
1574 SmallVector<unsigned, 5> Idxs;
1576 // Add indices from the extract value instruction
1577 Idxs.append(I->idx_begin(), I->idx_end());
1579 // Add requested indices
1580 Idxs.append(idx_range.begin(), idx_range.end());
1582 assert(Idxs.size() == size
1583 && "Number of indices added not correct?");
1585 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1587 // Otherwise, we don't know (such as, extracting from a function return value
1588 // or load instruction)
1592 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1593 /// it can be expressed as a base pointer plus a constant offset. Return the
1594 /// base and offset to the caller.
1595 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1596 const DataLayout &TD) {
1597 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1598 if (PtrOp == 0 || Ptr->getType()->isVectorTy())
1601 // Just look through bitcasts.
1602 if (PtrOp->getOpcode() == Instruction::BitCast)
1603 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1605 // If this is a GEP with constant indices, we can look through it.
1606 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1607 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1609 gep_type_iterator GTI = gep_type_begin(GEP);
1610 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1612 ConstantInt *OpC = cast<ConstantInt>(*I);
1613 if (OpC->isZero()) continue;
1615 // Handle a struct and array indices which add their offset to the pointer.
1616 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1617 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1619 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1620 Offset += OpC->getSExtValue()*Size;
1624 // Re-sign extend from the pointer size if needed to get overflow edge cases
1626 unsigned AS = GEP->getPointerAddressSpace();
1627 unsigned PtrSize = TD.getPointerSizeInBits(AS);
1629 Offset = SignExtend64(Offset, PtrSize);
1631 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1635 /// getConstantStringInfo - This function computes the length of a
1636 /// null-terminated C string pointed to by V. If successful, it returns true
1637 /// and returns the string in Str. If unsuccessful, it returns false.
1638 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
1639 uint64_t Offset, bool TrimAtNul) {
1642 // Look through bitcast instructions and geps.
1643 V = V->stripPointerCasts();
1645 // If the value is a GEP instructionor constant expression, treat it as an
1647 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1648 // Make sure the GEP has exactly three arguments.
1649 if (GEP->getNumOperands() != 3)
1652 // Make sure the index-ee is a pointer to array of i8.
1653 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1654 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1655 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1658 // Check to make sure that the first operand of the GEP is an integer and
1659 // has value 0 so that we are sure we're indexing into the initializer.
1660 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1661 if (FirstIdx == 0 || !FirstIdx->isZero())
1664 // If the second index isn't a ConstantInt, then this is a variable index
1665 // into the array. If this occurs, we can't say anything meaningful about
1667 uint64_t StartIdx = 0;
1668 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1669 StartIdx = CI->getZExtValue();
1672 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
1675 // The GEP instruction, constant or instruction, must reference a global
1676 // variable that is a constant and is initialized. The referenced constant
1677 // initializer is the array that we'll use for optimization.
1678 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
1679 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1682 // Handle the all-zeros case
1683 if (GV->getInitializer()->isNullValue()) {
1684 // This is a degenerate case. The initializer is constant zero so the
1685 // length of the string must be zero.
1690 // Must be a Constant Array
1691 const ConstantDataArray *Array =
1692 dyn_cast<ConstantDataArray>(GV->getInitializer());
1693 if (Array == 0 || !Array->isString())
1696 // Get the number of elements in the array
1697 uint64_t NumElts = Array->getType()->getArrayNumElements();
1699 // Start out with the entire array in the StringRef.
1700 Str = Array->getAsString();
1702 if (Offset > NumElts)
1705 // Skip over 'offset' bytes.
1706 Str = Str.substr(Offset);
1709 // Trim off the \0 and anything after it. If the array is not nul
1710 // terminated, we just return the whole end of string. The client may know
1711 // some other way that the string is length-bound.
1712 Str = Str.substr(0, Str.find('\0'));
1717 // These next two are very similar to the above, but also look through PHI
1719 // TODO: See if we can integrate these two together.
1721 /// GetStringLengthH - If we can compute the length of the string pointed to by
1722 /// the specified pointer, return 'len+1'. If we can't, return 0.
1723 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1724 // Look through noop bitcast instructions.
1725 V = V->stripPointerCasts();
1727 // If this is a PHI node, there are two cases: either we have already seen it
1729 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1730 if (!PHIs.insert(PN))
1731 return ~0ULL; // already in the set.
1733 // If it was new, see if all the input strings are the same length.
1734 uint64_t LenSoFar = ~0ULL;
1735 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1736 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1737 if (Len == 0) return 0; // Unknown length -> unknown.
1739 if (Len == ~0ULL) continue;
1741 if (Len != LenSoFar && LenSoFar != ~0ULL)
1742 return 0; // Disagree -> unknown.
1746 // Success, all agree.
1750 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1751 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1752 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1753 if (Len1 == 0) return 0;
1754 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1755 if (Len2 == 0) return 0;
1756 if (Len1 == ~0ULL) return Len2;
1757 if (Len2 == ~0ULL) return Len1;
1758 if (Len1 != Len2) return 0;
1762 // Otherwise, see if we can read the string.
1764 if (!getConstantStringInfo(V, StrData))
1767 return StrData.size()+1;
1770 /// GetStringLength - If we can compute the length of the string pointed to by
1771 /// the specified pointer, return 'len+1'. If we can't, return 0.
1772 uint64_t llvm::GetStringLength(Value *V) {
1773 if (!V->getType()->isPointerTy()) return 0;
1775 SmallPtrSet<PHINode*, 32> PHIs;
1776 uint64_t Len = GetStringLengthH(V, PHIs);
1777 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1778 // an empty string as a length.
1779 return Len == ~0ULL ? 1 : Len;
1783 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
1784 if (!V->getType()->isPointerTy())
1786 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1787 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1788 V = GEP->getPointerOperand();
1789 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1790 V = cast<Operator>(V)->getOperand(0);
1791 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1792 if (GA->mayBeOverridden())
1794 V = GA->getAliasee();
1796 // See if InstructionSimplify knows any relevant tricks.
1797 if (Instruction *I = dyn_cast<Instruction>(V))
1798 // TODO: Acquire a DominatorTree and use it.
1799 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1806 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1812 llvm::GetUnderlyingObjects(Value *V,
1813 SmallVectorImpl<Value *> &Objects,
1814 const DataLayout *TD,
1815 unsigned MaxLookup) {
1816 SmallPtrSet<Value *, 4> Visited;
1817 SmallVector<Value *, 4> Worklist;
1818 Worklist.push_back(V);
1820 Value *P = Worklist.pop_back_val();
1821 P = GetUnderlyingObject(P, TD, MaxLookup);
1823 if (!Visited.insert(P))
1826 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
1827 Worklist.push_back(SI->getTrueValue());
1828 Worklist.push_back(SI->getFalseValue());
1832 if (PHINode *PN = dyn_cast<PHINode>(P)) {
1833 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1834 Worklist.push_back(PN->getIncomingValue(i));
1838 Objects.push_back(P);
1839 } while (!Worklist.empty());
1842 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1843 /// are lifetime markers.
1845 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1846 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1848 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1849 if (!II) return false;
1851 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1852 II->getIntrinsicID() != Intrinsic::lifetime_end)
1858 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1859 const DataLayout *TD) {
1860 const Operator *Inst = dyn_cast<Operator>(V);
1864 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1865 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1869 switch (Inst->getOpcode()) {
1872 case Instruction::UDiv:
1873 case Instruction::URem:
1874 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1875 return isKnownNonZero(Inst->getOperand(1), TD);
1876 case Instruction::SDiv:
1877 case Instruction::SRem: {
1878 Value *Op = Inst->getOperand(1);
1879 // x / y is undefined if y == 0
1880 if (!isKnownNonZero(Op, TD))
1882 // x / y might be undefined if y == -1
1883 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1886 APInt KnownZero(BitWidth, 0);
1887 APInt KnownOne(BitWidth, 0);
1888 ComputeMaskedBits(Op, KnownZero, KnownOne, TD);
1891 case Instruction::Load: {
1892 const LoadInst *LI = cast<LoadInst>(Inst);
1893 if (!LI->isUnordered())
1895 return LI->getPointerOperand()->isDereferenceablePointer();
1897 case Instruction::Call: {
1898 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1899 switch (II->getIntrinsicID()) {
1900 // These synthetic intrinsics have no side-effects, and just mark
1901 // information about their operands.
1902 // FIXME: There are other no-op synthetic instructions that potentially
1903 // should be considered at least *safe* to speculate...
1904 case Intrinsic::dbg_declare:
1905 case Intrinsic::dbg_value:
1908 case Intrinsic::bswap:
1909 case Intrinsic::ctlz:
1910 case Intrinsic::ctpop:
1911 case Intrinsic::cttz:
1912 case Intrinsic::objectsize:
1913 case Intrinsic::sadd_with_overflow:
1914 case Intrinsic::smul_with_overflow:
1915 case Intrinsic::ssub_with_overflow:
1916 case Intrinsic::uadd_with_overflow:
1917 case Intrinsic::umul_with_overflow:
1918 case Intrinsic::usub_with_overflow:
1920 // TODO: some fp intrinsics are marked as having the same error handling
1921 // as libm. They're safe to speculate when they won't error.
1922 // TODO: are convert_{from,to}_fp16 safe?
1923 // TODO: can we list target-specific intrinsics here?
1927 return false; // The called function could have undefined behavior or
1928 // side-effects, even if marked readnone nounwind.
1930 case Instruction::VAArg:
1931 case Instruction::Alloca:
1932 case Instruction::Invoke:
1933 case Instruction::PHI:
1934 case Instruction::Store:
1935 case Instruction::Ret:
1936 case Instruction::Br:
1937 case Instruction::IndirectBr:
1938 case Instruction::Switch:
1939 case Instruction::Unreachable:
1940 case Instruction::Fence:
1941 case Instruction::LandingPad:
1942 case Instruction::AtomicRMW:
1943 case Instruction::AtomicCmpXchg:
1944 case Instruction::Resume:
1945 return false; // Misc instructions which have effects