1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Constants.h"
19 #include "llvm/DataLayout.h"
20 #include "llvm/GlobalAlias.h"
21 #include "llvm/GlobalVariable.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/IntrinsicInst.h"
24 #include "llvm/LLVMContext.h"
25 #include "llvm/Metadata.h"
26 #include "llvm/Operator.h"
27 #include "llvm/Support/ConstantRange.h"
28 #include "llvm/Support/GetElementPtrTypeIterator.h"
29 #include "llvm/Support/MathExtras.h"
30 #include "llvm/Support/PatternMatch.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!");
43 return TD ? TD->getPointerSizeInBits() : 0;
46 static void ComputeMaskedBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
47 APInt &KnownZero, APInt &KnownOne,
48 APInt &KnownZero2, APInt &KnownOne2,
49 const DataLayout *TD, unsigned Depth) {
51 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
52 // We know that the top bits of C-X are clear if X contains less bits
53 // than C (i.e. no wrap-around can happen). For example, 20-X is
54 // positive if we can prove that X is >= 0 and < 16.
55 if (!CLHS->getValue().isNegative()) {
56 unsigned BitWidth = KnownZero.getBitWidth();
57 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
58 // NLZ can't be BitWidth with no sign bit
59 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
60 llvm::ComputeMaskedBits(Op1, KnownZero2, KnownOne2, TD, Depth+1);
62 // If all of the MaskV bits are known to be zero, then we know the
63 // output top bits are zero, because we now know that the output is
65 if ((KnownZero2 & MaskV) == MaskV) {
66 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
67 // Top bits known zero.
68 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
74 unsigned BitWidth = KnownZero.getBitWidth();
76 // If one of the operands has trailing zeros, then the bits that the
77 // other operand has in those bit positions will be preserved in the
78 // result. For an add, this works with either operand. For a subtract,
79 // this only works if the known zeros are in the right operand.
80 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
81 llvm::ComputeMaskedBits(Op0, LHSKnownZero, LHSKnownOne, TD, Depth+1);
82 assert((LHSKnownZero & LHSKnownOne) == 0 &&
83 "Bits known to be one AND zero?");
84 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
86 llvm::ComputeMaskedBits(Op1, KnownZero2, KnownOne2, TD, Depth+1);
87 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
88 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
90 // Determine which operand has more trailing zeros, and use that
91 // many bits from the other operand.
92 if (LHSKnownZeroOut > RHSKnownZeroOut) {
94 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
95 KnownZero |= KnownZero2 & Mask;
96 KnownOne |= KnownOne2 & Mask;
98 // If the known zeros are in the left operand for a subtract,
99 // fall back to the minimum known zeros in both operands.
100 KnownZero |= APInt::getLowBitsSet(BitWidth,
101 std::min(LHSKnownZeroOut,
104 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
105 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
106 KnownZero |= LHSKnownZero & Mask;
107 KnownOne |= LHSKnownOne & Mask;
110 // Are we still trying to solve for the sign bit?
111 if (!KnownZero.isNegative() && !KnownOne.isNegative()) {
114 // Adding two positive numbers can't wrap into negative
115 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
116 KnownZero |= APInt::getSignBit(BitWidth);
117 // and adding two negative numbers can't wrap into positive.
118 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
119 KnownOne |= APInt::getSignBit(BitWidth);
121 // Subtracting a negative number from a positive one can't wrap
122 if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
123 KnownZero |= APInt::getSignBit(BitWidth);
124 // neither can subtracting a positive number from a negative one.
125 else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
126 KnownOne |= APInt::getSignBit(BitWidth);
132 static void ComputeMaskedBitsMul(Value *Op0, Value *Op1, bool NSW,
133 APInt &KnownZero, APInt &KnownOne,
134 APInt &KnownZero2, APInt &KnownOne2,
135 const DataLayout *TD, unsigned Depth) {
136 unsigned BitWidth = KnownZero.getBitWidth();
137 ComputeMaskedBits(Op1, KnownZero, KnownOne, TD, Depth+1);
138 ComputeMaskedBits(Op0, KnownZero2, KnownOne2, TD, Depth+1);
139 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
140 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
142 bool isKnownNegative = false;
143 bool isKnownNonNegative = false;
144 // If the multiplication is known not to overflow, compute the sign bit.
147 // The product of a number with itself is non-negative.
148 isKnownNonNegative = true;
150 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
151 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
152 bool isKnownNegativeOp1 = KnownOne.isNegative();
153 bool isKnownNegativeOp0 = KnownOne2.isNegative();
154 // The product of two numbers with the same sign is non-negative.
155 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
156 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
157 // The product of a negative number and a non-negative number is either
159 if (!isKnownNonNegative)
160 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
161 isKnownNonZero(Op0, TD, Depth)) ||
162 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
163 isKnownNonZero(Op1, TD, Depth));
167 // If low bits are zero in either operand, output low known-0 bits.
168 // Also compute a conserative estimate for high known-0 bits.
169 // More trickiness is possible, but this is sufficient for the
170 // interesting case of alignment computation.
171 KnownOne.clearAllBits();
172 unsigned TrailZ = KnownZero.countTrailingOnes() +
173 KnownZero2.countTrailingOnes();
174 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
175 KnownZero2.countLeadingOnes(),
176 BitWidth) - BitWidth;
178 TrailZ = std::min(TrailZ, BitWidth);
179 LeadZ = std::min(LeadZ, BitWidth);
180 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
181 APInt::getHighBitsSet(BitWidth, LeadZ);
183 // Only make use of no-wrap flags if we failed to compute the sign bit
184 // directly. This matters if the multiplication always overflows, in
185 // which case we prefer to follow the result of the direct computation,
186 // though as the program is invoking undefined behaviour we can choose
187 // whatever we like here.
188 if (isKnownNonNegative && !KnownOne.isNegative())
189 KnownZero.setBit(BitWidth - 1);
190 else if (isKnownNegative && !KnownZero.isNegative())
191 KnownOne.setBit(BitWidth - 1);
194 void llvm::computeMaskedBitsLoad(const MDNode &Ranges, APInt &KnownZero) {
195 unsigned BitWidth = KnownZero.getBitWidth();
196 unsigned NumRanges = Ranges.getNumOperands() / 2;
197 assert(NumRanges >= 1);
199 // Use the high end of the ranges to find leading zeros.
200 unsigned MinLeadingZeros = BitWidth;
201 for (unsigned i = 0; i < NumRanges; ++i) {
202 ConstantInt *Lower = cast<ConstantInt>(Ranges.getOperand(2*i + 0));
203 ConstantInt *Upper = cast<ConstantInt>(Ranges.getOperand(2*i + 1));
204 ConstantRange Range(Lower->getValue(), Upper->getValue());
205 if (Range.isWrappedSet())
206 MinLeadingZeros = 0; // -1 has no zeros
207 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
208 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
211 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
213 /// ComputeMaskedBits - Determine which of the bits are known to be either zero
214 /// or one and return them in the KnownZero/KnownOne bit sets.
216 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
217 /// we cannot optimize based on the assumption that it is zero without changing
218 /// it to be an explicit zero. If we don't change it to zero, other code could
219 /// optimized based on the contradictory assumption that it is non-zero.
220 /// Because instcombine aggressively folds operations with undef args anyway,
221 /// this won't lose us code quality.
223 /// This function is defined on values with integer type, values with pointer
224 /// type (but only if TD is non-null), and vectors of integers. In the case
225 /// where V is a vector, known zero, and known one values are the
226 /// same width as the vector element, and the bit is set only if it is true
227 /// for all of the elements in the vector.
228 void llvm::ComputeMaskedBits(Value *V, APInt &KnownZero, APInt &KnownOne,
229 const DataLayout *TD, unsigned Depth) {
230 assert(V && "No Value?");
231 assert(Depth <= MaxDepth && "Limit Search Depth");
232 unsigned BitWidth = KnownZero.getBitWidth();
234 assert((V->getType()->isIntOrIntVectorTy() ||
235 V->getType()->getScalarType()->isPointerTy()) &&
236 "Not integer or pointer type!");
238 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
239 (!V->getType()->isIntOrIntVectorTy() ||
240 V->getType()->getScalarSizeInBits() == BitWidth) &&
241 KnownZero.getBitWidth() == BitWidth &&
242 KnownOne.getBitWidth() == BitWidth &&
243 "V, Mask, KnownOne and KnownZero should have same BitWidth");
245 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
246 // We know all of the bits for a constant!
247 KnownOne = CI->getValue();
248 KnownZero = ~KnownOne;
251 // Null and aggregate-zero are all-zeros.
252 if (isa<ConstantPointerNull>(V) ||
253 isa<ConstantAggregateZero>(V)) {
254 KnownOne.clearAllBits();
255 KnownZero = APInt::getAllOnesValue(BitWidth);
258 // Handle a constant vector by taking the intersection of the known bits of
259 // each element. There is no real need to handle ConstantVector here, because
260 // we don't handle undef in any particularly useful way.
261 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
262 // We know that CDS must be a vector of integers. Take the intersection of
264 KnownZero.setAllBits(); KnownOne.setAllBits();
265 APInt Elt(KnownZero.getBitWidth(), 0);
266 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
267 Elt = CDS->getElementAsInteger(i);
274 // The address of an aligned GlobalValue has trailing zeros.
275 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
276 unsigned Align = GV->getAlignment();
277 if (Align == 0 && TD) {
278 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
279 Type *ObjectType = GVar->getType()->getElementType();
280 if (ObjectType->isSized()) {
281 // If the object is defined in the current Module, we'll be giving
282 // it the preferred alignment. Otherwise, we have to assume that it
283 // may only have the minimum ABI alignment.
284 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
285 Align = TD->getPreferredAlignment(GVar);
287 Align = TD->getABITypeAlignment(ObjectType);
292 KnownZero = APInt::getLowBitsSet(BitWidth,
293 CountTrailingZeros_32(Align));
295 KnownZero.clearAllBits();
296 KnownOne.clearAllBits();
299 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
300 // the bits of its aliasee.
301 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
302 if (GA->mayBeOverridden()) {
303 KnownZero.clearAllBits(); KnownOne.clearAllBits();
305 ComputeMaskedBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth+1);
310 if (Argument *A = dyn_cast<Argument>(V)) {
313 if (A->hasByValAttr()) {
314 // Get alignment information off byval arguments if specified in the IR.
315 Align = A->getParamAlignment();
316 } else if (TD && A->hasStructRetAttr()) {
317 // An sret parameter has at least the ABI alignment of the return type.
318 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
319 if (EltTy->isSized())
320 Align = TD->getABITypeAlignment(EltTy);
324 KnownZero = APInt::getLowBitsSet(BitWidth, CountTrailingZeros_32(Align));
328 // Start out not knowing anything.
329 KnownZero.clearAllBits(); KnownOne.clearAllBits();
331 if (Depth == MaxDepth)
332 return; // Limit search depth.
334 Operator *I = dyn_cast<Operator>(V);
337 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
338 switch (I->getOpcode()) {
340 case Instruction::Load:
341 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
342 computeMaskedBitsLoad(*MD, KnownZero);
344 case Instruction::And: {
345 // If either the LHS or the RHS are Zero, the result is zero.
346 ComputeMaskedBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
347 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
348 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
349 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
351 // Output known-1 bits are only known if set in both the LHS & RHS.
352 KnownOne &= KnownOne2;
353 // Output known-0 are known to be clear if zero in either the LHS | RHS.
354 KnownZero |= KnownZero2;
357 case Instruction::Or: {
358 ComputeMaskedBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
359 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
360 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
361 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
363 // Output known-0 bits are only known if clear in both the LHS & RHS.
364 KnownZero &= KnownZero2;
365 // Output known-1 are known to be set if set in either the LHS | RHS.
366 KnownOne |= KnownOne2;
369 case Instruction::Xor: {
370 ComputeMaskedBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
371 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
372 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
373 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
375 // Output known-0 bits are known if clear or set in both the LHS & RHS.
376 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
377 // Output known-1 are known to be set if set in only one of the LHS, RHS.
378 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
379 KnownZero = KnownZeroOut;
382 case Instruction::Mul: {
383 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
384 ComputeMaskedBitsMul(I->getOperand(0), I->getOperand(1), NSW,
385 KnownZero, KnownOne, KnownZero2, KnownOne2, TD, Depth);
388 case Instruction::UDiv: {
389 // For the purposes of computing leading zeros we can conservatively
390 // treat a udiv as a logical right shift by the power of 2 known to
391 // be less than the denominator.
392 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
393 unsigned LeadZ = KnownZero2.countLeadingOnes();
395 KnownOne2.clearAllBits();
396 KnownZero2.clearAllBits();
397 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
398 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
399 if (RHSUnknownLeadingOnes != BitWidth)
400 LeadZ = std::min(BitWidth,
401 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
403 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
406 case Instruction::Select:
407 ComputeMaskedBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1);
408 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD,
410 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
411 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
413 // Only known if known in both the LHS and RHS.
414 KnownOne &= KnownOne2;
415 KnownZero &= KnownZero2;
417 case Instruction::FPTrunc:
418 case Instruction::FPExt:
419 case Instruction::FPToUI:
420 case Instruction::FPToSI:
421 case Instruction::SIToFP:
422 case Instruction::UIToFP:
423 return; // Can't work with floating point.
424 case Instruction::PtrToInt:
425 case Instruction::IntToPtr:
426 // We can't handle these if we don't know the pointer size.
428 // FALL THROUGH and handle them the same as zext/trunc.
429 case Instruction::ZExt:
430 case Instruction::Trunc: {
431 Type *SrcTy = I->getOperand(0)->getType();
433 unsigned SrcBitWidth;
434 // Note that we handle pointer operands here because of inttoptr/ptrtoint
435 // which fall through here.
437 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
439 SrcBitWidth = SrcTy->getScalarSizeInBits();
440 if (!SrcBitWidth) return;
443 assert(SrcBitWidth && "SrcBitWidth can't be zero");
444 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
445 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
446 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
447 KnownZero = KnownZero.zextOrTrunc(BitWidth);
448 KnownOne = KnownOne.zextOrTrunc(BitWidth);
449 // Any top bits are known to be zero.
450 if (BitWidth > SrcBitWidth)
451 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
454 case Instruction::BitCast: {
455 Type *SrcTy = I->getOperand(0)->getType();
456 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
457 // TODO: For now, not handling conversions like:
458 // (bitcast i64 %x to <2 x i32>)
459 !I->getType()->isVectorTy()) {
460 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
465 case Instruction::SExt: {
466 // Compute the bits in the result that are not present in the input.
467 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
469 KnownZero = KnownZero.trunc(SrcBitWidth);
470 KnownOne = KnownOne.trunc(SrcBitWidth);
471 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
472 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
473 KnownZero = KnownZero.zext(BitWidth);
474 KnownOne = KnownOne.zext(BitWidth);
476 // If the sign bit of the input is known set or clear, then we know the
477 // top bits of the result.
478 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
479 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
480 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
481 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
484 case Instruction::Shl:
485 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
486 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
487 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
488 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
489 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
490 KnownZero <<= ShiftAmt;
491 KnownOne <<= ShiftAmt;
492 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
496 case Instruction::LShr:
497 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
498 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
499 // Compute the new bits that are at the top now.
500 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
502 // Unsigned shift right.
503 ComputeMaskedBits(I->getOperand(0), KnownZero,KnownOne, TD, Depth+1);
504 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
505 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
506 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
507 // high bits known zero.
508 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
512 case Instruction::AShr:
513 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
514 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
515 // Compute the new bits that are at the top now.
516 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
518 // Signed shift right.
519 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
520 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
521 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
522 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
524 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
525 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
526 KnownZero |= HighBits;
527 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
528 KnownOne |= HighBits;
532 case Instruction::Sub: {
533 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
534 ComputeMaskedBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
535 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
539 case Instruction::Add: {
540 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
541 ComputeMaskedBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
542 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
546 case Instruction::SRem:
547 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
548 APInt RA = Rem->getValue().abs();
549 if (RA.isPowerOf2()) {
550 APInt LowBits = RA - 1;
551 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
553 // The low bits of the first operand are unchanged by the srem.
554 KnownZero = KnownZero2 & LowBits;
555 KnownOne = KnownOne2 & LowBits;
557 // If the first operand is non-negative or has all low bits zero, then
558 // the upper bits are all zero.
559 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
560 KnownZero |= ~LowBits;
562 // If the first operand is negative and not all low bits are zero, then
563 // the upper bits are all one.
564 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
565 KnownOne |= ~LowBits;
567 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
571 // The sign bit is the LHS's sign bit, except when the result of the
572 // remainder is zero.
573 if (KnownZero.isNonNegative()) {
574 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
575 ComputeMaskedBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
577 // If it's known zero, our sign bit is also zero.
578 if (LHSKnownZero.isNegative())
579 KnownZero.setBit(BitWidth - 1);
583 case Instruction::URem: {
584 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
585 APInt RA = Rem->getValue();
586 if (RA.isPowerOf2()) {
587 APInt LowBits = (RA - 1);
588 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD,
590 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
591 KnownZero |= ~LowBits;
597 // Since the result is less than or equal to either operand, any leading
598 // zero bits in either operand must also exist in the result.
599 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
600 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
602 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
603 KnownZero2.countLeadingOnes());
604 KnownOne.clearAllBits();
605 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
609 case Instruction::Alloca: {
610 AllocaInst *AI = cast<AllocaInst>(V);
611 unsigned Align = AI->getAlignment();
612 if (Align == 0 && TD)
613 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
616 KnownZero = APInt::getLowBitsSet(BitWidth, CountTrailingZeros_32(Align));
619 case Instruction::GetElementPtr: {
620 // Analyze all of the subscripts of this getelementptr instruction
621 // to determine if we can prove known low zero bits.
622 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
623 ComputeMaskedBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
625 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
627 gep_type_iterator GTI = gep_type_begin(I);
628 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
629 Value *Index = I->getOperand(i);
630 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
631 // Handle struct member offset arithmetic.
633 const StructLayout *SL = TD->getStructLayout(STy);
634 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
635 uint64_t Offset = SL->getElementOffset(Idx);
636 TrailZ = std::min(TrailZ,
637 CountTrailingZeros_64(Offset));
639 // Handle array index arithmetic.
640 Type *IndexedTy = GTI.getIndexedType();
641 if (!IndexedTy->isSized()) return;
642 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
643 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
644 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
645 ComputeMaskedBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1);
646 TrailZ = std::min(TrailZ,
647 unsigned(CountTrailingZeros_64(TypeSize) +
648 LocalKnownZero.countTrailingOnes()));
652 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
655 case Instruction::PHI: {
656 PHINode *P = cast<PHINode>(I);
657 // Handle the case of a simple two-predecessor recurrence PHI.
658 // There's a lot more that could theoretically be done here, but
659 // this is sufficient to catch some interesting cases.
660 if (P->getNumIncomingValues() == 2) {
661 for (unsigned i = 0; i != 2; ++i) {
662 Value *L = P->getIncomingValue(i);
663 Value *R = P->getIncomingValue(!i);
664 Operator *LU = dyn_cast<Operator>(L);
667 unsigned Opcode = LU->getOpcode();
668 // Check for operations that have the property that if
669 // both their operands have low zero bits, the result
670 // will have low zero bits.
671 if (Opcode == Instruction::Add ||
672 Opcode == Instruction::Sub ||
673 Opcode == Instruction::And ||
674 Opcode == Instruction::Or ||
675 Opcode == Instruction::Mul) {
676 Value *LL = LU->getOperand(0);
677 Value *LR = LU->getOperand(1);
678 // Find a recurrence.
685 // Ok, we have a PHI of the form L op= R. Check for low
687 ComputeMaskedBits(R, KnownZero2, KnownOne2, TD, Depth+1);
689 // We need to take the minimum number of known bits
690 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
691 ComputeMaskedBits(L, KnownZero3, KnownOne3, TD, Depth+1);
693 KnownZero = APInt::getLowBitsSet(BitWidth,
694 std::min(KnownZero2.countTrailingOnes(),
695 KnownZero3.countTrailingOnes()));
701 // Unreachable blocks may have zero-operand PHI nodes.
702 if (P->getNumIncomingValues() == 0)
705 // Otherwise take the unions of the known bit sets of the operands,
706 // taking conservative care to avoid excessive recursion.
707 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
708 // Skip if every incoming value references to ourself.
709 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
712 KnownZero = APInt::getAllOnesValue(BitWidth);
713 KnownOne = APInt::getAllOnesValue(BitWidth);
714 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
715 // Skip direct self references.
716 if (P->getIncomingValue(i) == P) continue;
718 KnownZero2 = APInt(BitWidth, 0);
719 KnownOne2 = APInt(BitWidth, 0);
720 // Recurse, but cap the recursion to one level, because we don't
721 // want to waste time spinning around in loops.
722 ComputeMaskedBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
724 KnownZero &= KnownZero2;
725 KnownOne &= KnownOne2;
726 // If all bits have been ruled out, there's no need to check
728 if (!KnownZero && !KnownOne)
734 case Instruction::Call:
735 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
736 switch (II->getIntrinsicID()) {
738 case Intrinsic::ctlz:
739 case Intrinsic::cttz: {
740 unsigned LowBits = Log2_32(BitWidth)+1;
741 // If this call is undefined for 0, the result will be less than 2^n.
742 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
744 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
747 case Intrinsic::ctpop: {
748 unsigned LowBits = Log2_32(BitWidth)+1;
749 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
752 case Intrinsic::x86_sse42_crc32_64_8:
753 case Intrinsic::x86_sse42_crc32_64_64:
754 KnownZero = APInt::getHighBitsSet(64, 32);
759 case Instruction::ExtractValue:
760 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
761 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
762 if (EVI->getNumIndices() != 1) break;
763 if (EVI->getIndices()[0] == 0) {
764 switch (II->getIntrinsicID()) {
766 case Intrinsic::uadd_with_overflow:
767 case Intrinsic::sadd_with_overflow:
768 ComputeMaskedBitsAddSub(true, II->getArgOperand(0),
769 II->getArgOperand(1), false, KnownZero,
770 KnownOne, KnownZero2, KnownOne2, TD, Depth);
772 case Intrinsic::usub_with_overflow:
773 case Intrinsic::ssub_with_overflow:
774 ComputeMaskedBitsAddSub(false, II->getArgOperand(0),
775 II->getArgOperand(1), false, KnownZero,
776 KnownOne, KnownZero2, KnownOne2, TD, Depth);
778 case Intrinsic::umul_with_overflow:
779 case Intrinsic::smul_with_overflow:
780 ComputeMaskedBitsMul(II->getArgOperand(0), II->getArgOperand(1),
781 false, KnownZero, KnownOne,
782 KnownZero2, KnownOne2, TD, Depth);
790 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
791 /// one. Convenience wrapper around ComputeMaskedBits.
792 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
793 const DataLayout *TD, unsigned Depth) {
794 unsigned BitWidth = getBitWidth(V->getType(), TD);
800 APInt ZeroBits(BitWidth, 0);
801 APInt OneBits(BitWidth, 0);
802 ComputeMaskedBits(V, ZeroBits, OneBits, TD, Depth);
803 KnownOne = OneBits[BitWidth - 1];
804 KnownZero = ZeroBits[BitWidth - 1];
807 /// isKnownToBeAPowerOfTwo - Return true if the given value is known to have exactly one
808 /// bit set when defined. For vectors return true if every element is known to
809 /// be a power of two when defined. Supports values with integer or pointer
810 /// types and vectors of integers.
811 bool llvm::isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth) {
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 isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth);
840 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
841 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth);
843 if (SelectInst *SI = dyn_cast<SelectInst>(V))
844 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth) &&
845 isKnownToBeAPowerOfTwo(SI->getFalseValue(), 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 (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth) ||
850 isKnownToBeAPowerOfTwo(Y, /*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 isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, Depth);
869 /// \brief Test whether a GEP's result is known to be non-null.
871 /// Uses properties inherent in a GEP to try to determine whether it is known
874 /// Currently this routine does not support vector GEPs.
875 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
877 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
880 // FIXME: Support vector-GEPs.
881 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
883 // If the base pointer is non-null, we cannot walk to a null address with an
884 // inbounds GEP in address space zero.
885 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth))
888 // Past this, if we don't have DataLayout, we can't do much.
892 // Walk the GEP operands and see if any operand introduces a non-zero offset.
893 // If so, then the GEP cannot produce a null pointer, as doing so would
894 // inherently violate the inbounds contract within address space zero.
895 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
897 // Struct types are easy -- they must always be indexed by a constant.
898 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
899 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
900 unsigned ElementIdx = OpC->getZExtValue();
901 const StructLayout *SL = DL->getStructLayout(STy);
902 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
903 if (ElementOffset > 0)
908 // If we have a zero-sized type, the index doesn't matter. Keep looping.
909 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
912 // Fast path the constant operand case both for efficiency and so we don't
913 // increment Depth when just zipping down an all-constant GEP.
914 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
920 // We post-increment Depth here because while isKnownNonZero increments it
921 // as well, when we pop back up that increment won't persist. We don't want
922 // to recurse 10k times just because we have 10k GEP operands. We don't
923 // bail completely out because we want to handle constant GEPs regardless
925 if (Depth++ >= MaxDepth)
928 if (isKnownNonZero(GTI.getOperand(), DL, Depth))
935 /// isKnownNonZero - Return true if the given value is known to be non-zero
936 /// when defined. For vectors return true if every element is known to be
937 /// non-zero when defined. Supports values with integer or pointer type and
938 /// vectors of integers.
939 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth) {
940 if (Constant *C = dyn_cast<Constant>(V)) {
941 if (C->isNullValue())
943 if (isa<ConstantInt>(C))
944 // Must be non-zero due to null test above.
946 // TODO: Handle vectors
950 // The remaining tests are all recursive, so bail out if we hit the limit.
951 if (Depth++ >= MaxDepth)
954 // Check for pointer simplifications.
955 if (V->getType()->isPointerTy()) {
956 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
957 if (isGEPKnownNonNull(GEP, TD, Depth))
961 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
963 // X | Y != 0 if X != 0 or Y != 0.
964 Value *X = 0, *Y = 0;
965 if (match(V, m_Or(m_Value(X), m_Value(Y))))
966 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
968 // ext X != 0 if X != 0.
969 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
970 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
972 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
973 // if the lowest bit is shifted off the end.
974 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
975 // shl nuw can't remove any non-zero bits.
976 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
977 if (BO->hasNoUnsignedWrap())
978 return isKnownNonZero(X, TD, Depth);
980 APInt KnownZero(BitWidth, 0);
981 APInt KnownOne(BitWidth, 0);
982 ComputeMaskedBits(X, KnownZero, KnownOne, TD, Depth);
986 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
987 // defined if the sign bit is shifted off the end.
988 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
989 // shr exact can only shift out zero bits.
990 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
992 return isKnownNonZero(X, TD, Depth);
994 bool XKnownNonNegative, XKnownNegative;
995 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
999 // div exact can only produce a zero if the dividend is zero.
1000 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1001 return isKnownNonZero(X, TD, Depth);
1004 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1005 bool XKnownNonNegative, XKnownNegative;
1006 bool YKnownNonNegative, YKnownNegative;
1007 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
1008 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
1010 // If X and Y are both non-negative (as signed values) then their sum is not
1011 // zero unless both X and Y are zero.
1012 if (XKnownNonNegative && YKnownNonNegative)
1013 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
1016 // If X and Y are both negative (as signed values) then their sum is not
1017 // zero unless both X and Y equal INT_MIN.
1018 if (BitWidth && XKnownNegative && YKnownNegative) {
1019 APInt KnownZero(BitWidth, 0);
1020 APInt KnownOne(BitWidth, 0);
1021 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1022 // The sign bit of X is set. If some other bit is set then X is not equal
1024 ComputeMaskedBits(X, KnownZero, KnownOne, TD, Depth);
1025 if ((KnownOne & Mask) != 0)
1027 // The sign bit of Y is set. If some other bit is set then Y is not equal
1029 ComputeMaskedBits(Y, KnownZero, KnownOne, TD, Depth);
1030 if ((KnownOne & Mask) != 0)
1034 // The sum of a non-negative number and a power of two is not zero.
1035 if (XKnownNonNegative && isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth))
1037 if (YKnownNonNegative && isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth))
1041 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1042 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1043 // If X and Y are non-zero then so is X * Y as long as the multiplication
1044 // does not overflow.
1045 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1046 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
1049 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1050 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1051 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
1052 isKnownNonZero(SI->getFalseValue(), TD, Depth))
1056 if (!BitWidth) return false;
1057 APInt KnownZero(BitWidth, 0);
1058 APInt KnownOne(BitWidth, 0);
1059 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1060 return KnownOne != 0;
1063 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1064 /// this predicate to simplify operations downstream. Mask is known to be zero
1065 /// for bits that V cannot have.
1067 /// This function is defined on values with integer type, values with pointer
1068 /// type (but only if TD is non-null), and vectors of integers. In the case
1069 /// where V is a vector, the mask, known zero, and known one values are the
1070 /// same width as the vector element, and the bit is set only if it is true
1071 /// for all of the elements in the vector.
1072 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
1073 const DataLayout *TD, unsigned Depth) {
1074 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1075 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1076 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1077 return (KnownZero & Mask) == Mask;
1082 /// ComputeNumSignBits - Return the number of times the sign bit of the
1083 /// register is replicated into the other bits. We know that at least 1 bit
1084 /// is always equal to the sign bit (itself), but other cases can give us
1085 /// information. For example, immediately after an "ashr X, 2", we know that
1086 /// the top 3 bits are all equal to each other, so we return 3.
1088 /// 'Op' must have a scalar integer type.
1090 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
1092 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1093 "ComputeNumSignBits requires a DataLayout object to operate "
1094 "on non-integer values!");
1095 Type *Ty = V->getType();
1096 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1097 Ty->getScalarSizeInBits();
1099 unsigned FirstAnswer = 1;
1101 // Note that ConstantInt is handled by the general ComputeMaskedBits case
1105 return 1; // Limit search depth.
1107 Operator *U = dyn_cast<Operator>(V);
1108 switch (Operator::getOpcode(V)) {
1110 case Instruction::SExt:
1111 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1112 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
1114 case Instruction::AShr: {
1115 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1116 // ashr X, C -> adds C sign bits. Vectors too.
1118 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1119 Tmp += ShAmt->getZExtValue();
1120 if (Tmp > TyBits) Tmp = TyBits;
1124 case Instruction::Shl: {
1126 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1127 // shl destroys sign bits.
1128 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1129 Tmp2 = ShAmt->getZExtValue();
1130 if (Tmp2 >= TyBits || // Bad shift.
1131 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1136 case Instruction::And:
1137 case Instruction::Or:
1138 case Instruction::Xor: // NOT is handled here.
1139 // Logical binary ops preserve the number of sign bits at the worst.
1140 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1142 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1143 FirstAnswer = std::min(Tmp, Tmp2);
1144 // We computed what we know about the sign bits as our first
1145 // answer. Now proceed to the generic code that uses
1146 // ComputeMaskedBits, and pick whichever answer is better.
1150 case Instruction::Select:
1151 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1152 if (Tmp == 1) return 1; // Early out.
1153 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1154 return std::min(Tmp, Tmp2);
1156 case Instruction::Add:
1157 // Add can have at most one carry bit. Thus we know that the output
1158 // is, at worst, one more bit than the inputs.
1159 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1160 if (Tmp == 1) return 1; // Early out.
1162 // Special case decrementing a value (ADD X, -1):
1163 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1164 if (CRHS->isAllOnesValue()) {
1165 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1166 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
1168 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1170 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1173 // If we are subtracting one from a positive number, there is no carry
1174 // out of the result.
1175 if (KnownZero.isNegative())
1179 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1180 if (Tmp2 == 1) return 1;
1181 return std::min(Tmp, Tmp2)-1;
1183 case Instruction::Sub:
1184 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1185 if (Tmp2 == 1) return 1;
1188 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1189 if (CLHS->isNullValue()) {
1190 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1191 ComputeMaskedBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
1192 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1194 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1197 // If the input is known to be positive (the sign bit is known clear),
1198 // the output of the NEG has the same number of sign bits as the input.
1199 if (KnownZero.isNegative())
1202 // Otherwise, we treat this like a SUB.
1205 // Sub can have at most one carry bit. Thus we know that the output
1206 // is, at worst, one more bit than the inputs.
1207 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1208 if (Tmp == 1) return 1; // Early out.
1209 return std::min(Tmp, Tmp2)-1;
1211 case Instruction::PHI: {
1212 PHINode *PN = cast<PHINode>(U);
1213 // Don't analyze large in-degree PHIs.
1214 if (PN->getNumIncomingValues() > 4) break;
1216 // Take the minimum of all incoming values. This can't infinitely loop
1217 // because of our depth threshold.
1218 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1219 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1220 if (Tmp == 1) return Tmp;
1222 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1227 case Instruction::Trunc:
1228 // FIXME: it's tricky to do anything useful for this, but it is an important
1229 // case for targets like X86.
1233 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1234 // use this information.
1235 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1237 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1239 if (KnownZero.isNegative()) { // sign bit is 0
1241 } else if (KnownOne.isNegative()) { // sign bit is 1;
1248 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1249 // the number of identical bits in the top of the input value.
1251 Mask <<= Mask.getBitWidth()-TyBits;
1252 // Return # leading zeros. We use 'min' here in case Val was zero before
1253 // shifting. We don't want to return '64' as for an i32 "0".
1254 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1257 /// ComputeMultiple - This function computes the integer multiple of Base that
1258 /// equals V. If successful, it returns true and returns the multiple in
1259 /// Multiple. If unsuccessful, it returns false. It looks
1260 /// through SExt instructions only if LookThroughSExt is true.
1261 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1262 bool LookThroughSExt, unsigned Depth) {
1263 const unsigned MaxDepth = 6;
1265 assert(V && "No Value?");
1266 assert(Depth <= MaxDepth && "Limit Search Depth");
1267 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1269 Type *T = V->getType();
1271 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1281 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1282 Constant *BaseVal = ConstantInt::get(T, Base);
1283 if (CO && CO == BaseVal) {
1285 Multiple = ConstantInt::get(T, 1);
1289 if (CI && CI->getZExtValue() % Base == 0) {
1290 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1294 if (Depth == MaxDepth) return false; // Limit search depth.
1296 Operator *I = dyn_cast<Operator>(V);
1297 if (!I) return false;
1299 switch (I->getOpcode()) {
1301 case Instruction::SExt:
1302 if (!LookThroughSExt) return false;
1303 // otherwise fall through to ZExt
1304 case Instruction::ZExt:
1305 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1306 LookThroughSExt, Depth+1);
1307 case Instruction::Shl:
1308 case Instruction::Mul: {
1309 Value *Op0 = I->getOperand(0);
1310 Value *Op1 = I->getOperand(1);
1312 if (I->getOpcode() == Instruction::Shl) {
1313 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1314 if (!Op1CI) return false;
1315 // Turn Op0 << Op1 into Op0 * 2^Op1
1316 APInt Op1Int = Op1CI->getValue();
1317 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1318 APInt API(Op1Int.getBitWidth(), 0);
1319 API.setBit(BitToSet);
1320 Op1 = ConstantInt::get(V->getContext(), API);
1324 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1325 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1326 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1327 if (Op1C->getType()->getPrimitiveSizeInBits() <
1328 MulC->getType()->getPrimitiveSizeInBits())
1329 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1330 if (Op1C->getType()->getPrimitiveSizeInBits() >
1331 MulC->getType()->getPrimitiveSizeInBits())
1332 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1334 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1335 Multiple = ConstantExpr::getMul(MulC, Op1C);
1339 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1340 if (Mul0CI->getValue() == 1) {
1341 // V == Base * Op1, so return Op1
1348 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1349 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1350 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1351 if (Op0C->getType()->getPrimitiveSizeInBits() <
1352 MulC->getType()->getPrimitiveSizeInBits())
1353 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1354 if (Op0C->getType()->getPrimitiveSizeInBits() >
1355 MulC->getType()->getPrimitiveSizeInBits())
1356 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1358 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1359 Multiple = ConstantExpr::getMul(MulC, Op0C);
1363 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1364 if (Mul1CI->getValue() == 1) {
1365 // V == Base * Op0, so return Op0
1373 // We could not determine if V is a multiple of Base.
1377 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1378 /// value is never equal to -0.0.
1380 /// NOTE: this function will need to be revisited when we support non-default
1383 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1384 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1385 return !CFP->getValueAPF().isNegZero();
1388 return 1; // Limit search depth.
1390 const Operator *I = dyn_cast<Operator>(V);
1391 if (I == 0) return false;
1393 // Check if the nsz fast-math flag is set
1394 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
1395 if (FPO->hasNoSignedZeros())
1398 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1399 if (I->getOpcode() == Instruction::FAdd &&
1400 isa<ConstantFP>(I->getOperand(1)) &&
1401 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1404 // sitofp and uitofp turn into +0.0 for zero.
1405 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1408 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1409 // sqrt(-0.0) = -0.0, no other negative results are possible.
1410 if (II->getIntrinsicID() == Intrinsic::sqrt)
1411 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1413 if (const CallInst *CI = dyn_cast<CallInst>(I))
1414 if (const Function *F = CI->getCalledFunction()) {
1415 if (F->isDeclaration()) {
1417 if (F->getName() == "abs") return true;
1418 // fabs[lf](x) != -0.0
1419 if (F->getName() == "fabs") return true;
1420 if (F->getName() == "fabsf") return true;
1421 if (F->getName() == "fabsl") return true;
1422 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1423 F->getName() == "sqrtl")
1424 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1431 /// isBytewiseValue - If the specified value can be set by repeating the same
1432 /// byte in memory, return the i8 value that it is represented with. This is
1433 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1434 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1435 /// byte store (e.g. i16 0x1234), return null.
1436 Value *llvm::isBytewiseValue(Value *V) {
1437 // All byte-wide stores are splatable, even of arbitrary variables.
1438 if (V->getType()->isIntegerTy(8)) return V;
1440 // Handle 'null' ConstantArrayZero etc.
1441 if (Constant *C = dyn_cast<Constant>(V))
1442 if (C->isNullValue())
1443 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1445 // Constant float and double values can be handled as integer values if the
1446 // corresponding integer value is "byteable". An important case is 0.0.
1447 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1448 if (CFP->getType()->isFloatTy())
1449 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1450 if (CFP->getType()->isDoubleTy())
1451 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1452 // Don't handle long double formats, which have strange constraints.
1455 // We can handle constant integers that are power of two in size and a
1456 // multiple of 8 bits.
1457 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1458 unsigned Width = CI->getBitWidth();
1459 if (isPowerOf2_32(Width) && Width > 8) {
1460 // We can handle this value if the recursive binary decomposition is the
1461 // same at all levels.
1462 APInt Val = CI->getValue();
1464 while (Val.getBitWidth() != 8) {
1465 unsigned NextWidth = Val.getBitWidth()/2;
1466 Val2 = Val.lshr(NextWidth);
1467 Val2 = Val2.trunc(Val.getBitWidth()/2);
1468 Val = Val.trunc(Val.getBitWidth()/2);
1470 // If the top/bottom halves aren't the same, reject it.
1474 return ConstantInt::get(V->getContext(), Val);
1478 // A ConstantDataArray/Vector is splatable if all its members are equal and
1480 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1481 Value *Elt = CA->getElementAsConstant(0);
1482 Value *Val = isBytewiseValue(Elt);
1486 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1487 if (CA->getElementAsConstant(I) != Elt)
1493 // Conceptually, we could handle things like:
1494 // %a = zext i8 %X to i16
1495 // %b = shl i16 %a, 8
1496 // %c = or i16 %a, %b
1497 // but until there is an example that actually needs this, it doesn't seem
1498 // worth worrying about.
1503 // This is the recursive version of BuildSubAggregate. It takes a few different
1504 // arguments. Idxs is the index within the nested struct From that we are
1505 // looking at now (which is of type IndexedType). IdxSkip is the number of
1506 // indices from Idxs that should be left out when inserting into the resulting
1507 // struct. To is the result struct built so far, new insertvalue instructions
1509 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1510 SmallVector<unsigned, 10> &Idxs,
1512 Instruction *InsertBefore) {
1513 llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1515 // Save the original To argument so we can modify it
1517 // General case, the type indexed by Idxs is a struct
1518 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1519 // Process each struct element recursively
1522 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1526 // Couldn't find any inserted value for this index? Cleanup
1527 while (PrevTo != OrigTo) {
1528 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1529 PrevTo = Del->getAggregateOperand();
1530 Del->eraseFromParent();
1532 // Stop processing elements
1536 // If we successfully found a value for each of our subaggregates
1540 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1541 // the struct's elements had a value that was inserted directly. In the latter
1542 // case, perhaps we can't determine each of the subelements individually, but
1543 // we might be able to find the complete struct somewhere.
1545 // Find the value that is at that particular spot
1546 Value *V = FindInsertedValue(From, Idxs);
1551 // Insert the value in the new (sub) aggregrate
1552 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1553 "tmp", InsertBefore);
1556 // This helper takes a nested struct and extracts a part of it (which is again a
1557 // struct) into a new value. For example, given the struct:
1558 // { a, { b, { c, d }, e } }
1559 // and the indices "1, 1" this returns
1562 // It does this by inserting an insertvalue for each element in the resulting
1563 // struct, as opposed to just inserting a single struct. This will only work if
1564 // each of the elements of the substruct are known (ie, inserted into From by an
1565 // insertvalue instruction somewhere).
1567 // All inserted insertvalue instructions are inserted before InsertBefore
1568 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1569 Instruction *InsertBefore) {
1570 assert(InsertBefore && "Must have someplace to insert!");
1571 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1573 Value *To = UndefValue::get(IndexedType);
1574 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1575 unsigned IdxSkip = Idxs.size();
1577 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1580 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1581 /// the scalar value indexed is already around as a register, for example if it
1582 /// were inserted directly into the aggregrate.
1584 /// If InsertBefore is not null, this function will duplicate (modified)
1585 /// insertvalues when a part of a nested struct is extracted.
1586 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1587 Instruction *InsertBefore) {
1588 // Nothing to index? Just return V then (this is useful at the end of our
1590 if (idx_range.empty())
1592 // We have indices, so V should have an indexable type.
1593 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1594 "Not looking at a struct or array?");
1595 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1596 "Invalid indices for type?");
1598 if (Constant *C = dyn_cast<Constant>(V)) {
1599 C = C->getAggregateElement(idx_range[0]);
1600 if (C == 0) return 0;
1601 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1604 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1605 // Loop the indices for the insertvalue instruction in parallel with the
1606 // requested indices
1607 const unsigned *req_idx = idx_range.begin();
1608 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1609 i != e; ++i, ++req_idx) {
1610 if (req_idx == idx_range.end()) {
1611 // We can't handle this without inserting insertvalues
1615 // The requested index identifies a part of a nested aggregate. Handle
1616 // this specially. For example,
1617 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1618 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1619 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1620 // This can be changed into
1621 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1622 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1623 // which allows the unused 0,0 element from the nested struct to be
1625 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1629 // This insert value inserts something else than what we are looking for.
1630 // See if the (aggregrate) value inserted into has the value we are
1631 // looking for, then.
1633 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1636 // If we end up here, the indices of the insertvalue match with those
1637 // requested (though possibly only partially). Now we recursively look at
1638 // the inserted value, passing any remaining indices.
1639 return FindInsertedValue(I->getInsertedValueOperand(),
1640 makeArrayRef(req_idx, idx_range.end()),
1644 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1645 // If we're extracting a value from an aggregrate that was extracted from
1646 // something else, we can extract from that something else directly instead.
1647 // However, we will need to chain I's indices with the requested indices.
1649 // Calculate the number of indices required
1650 unsigned size = I->getNumIndices() + idx_range.size();
1651 // Allocate some space to put the new indices in
1652 SmallVector<unsigned, 5> Idxs;
1654 // Add indices from the extract value instruction
1655 Idxs.append(I->idx_begin(), I->idx_end());
1657 // Add requested indices
1658 Idxs.append(idx_range.begin(), idx_range.end());
1660 assert(Idxs.size() == size
1661 && "Number of indices added not correct?");
1663 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1665 // Otherwise, we don't know (such as, extracting from a function return value
1666 // or load instruction)
1670 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1671 /// it can be expressed as a base pointer plus a constant offset. Return the
1672 /// base and offset to the caller.
1673 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1674 const DataLayout &TD) {
1675 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1676 if (PtrOp == 0 || Ptr->getType()->isVectorTy())
1679 // Just look through bitcasts.
1680 if (PtrOp->getOpcode() == Instruction::BitCast)
1681 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1683 // If this is a GEP with constant indices, we can look through it.
1684 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1685 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1687 gep_type_iterator GTI = gep_type_begin(GEP);
1688 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1690 ConstantInt *OpC = cast<ConstantInt>(*I);
1691 if (OpC->isZero()) continue;
1693 // Handle a struct and array indices which add their offset to the pointer.
1694 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1695 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1697 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1698 Offset += OpC->getSExtValue()*Size;
1702 // Re-sign extend from the pointer size if needed to get overflow edge cases
1704 unsigned PtrSize = TD.getPointerSizeInBits();
1706 Offset = SignExtend64(Offset, PtrSize);
1708 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1712 /// getConstantStringInfo - This function computes the length of a
1713 /// null-terminated C string pointed to by V. If successful, it returns true
1714 /// and returns the string in Str. If unsuccessful, it returns false.
1715 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
1716 uint64_t Offset, bool TrimAtNul) {
1719 // Look through bitcast instructions and geps.
1720 V = V->stripPointerCasts();
1722 // If the value is a GEP instructionor constant expression, treat it as an
1724 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1725 // Make sure the GEP has exactly three arguments.
1726 if (GEP->getNumOperands() != 3)
1729 // Make sure the index-ee is a pointer to array of i8.
1730 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1731 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1732 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1735 // Check to make sure that the first operand of the GEP is an integer and
1736 // has value 0 so that we are sure we're indexing into the initializer.
1737 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1738 if (FirstIdx == 0 || !FirstIdx->isZero())
1741 // If the second index isn't a ConstantInt, then this is a variable index
1742 // into the array. If this occurs, we can't say anything meaningful about
1744 uint64_t StartIdx = 0;
1745 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1746 StartIdx = CI->getZExtValue();
1749 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
1752 // The GEP instruction, constant or instruction, must reference a global
1753 // variable that is a constant and is initialized. The referenced constant
1754 // initializer is the array that we'll use for optimization.
1755 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
1756 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1759 // Handle the all-zeros case
1760 if (GV->getInitializer()->isNullValue()) {
1761 // This is a degenerate case. The initializer is constant zero so the
1762 // length of the string must be zero.
1767 // Must be a Constant Array
1768 const ConstantDataArray *Array =
1769 dyn_cast<ConstantDataArray>(GV->getInitializer());
1770 if (Array == 0 || !Array->isString())
1773 // Get the number of elements in the array
1774 uint64_t NumElts = Array->getType()->getArrayNumElements();
1776 // Start out with the entire array in the StringRef.
1777 Str = Array->getAsString();
1779 if (Offset > NumElts)
1782 // Skip over 'offset' bytes.
1783 Str = Str.substr(Offset);
1786 // Trim off the \0 and anything after it. If the array is not nul
1787 // terminated, we just return the whole end of string. The client may know
1788 // some other way that the string is length-bound.
1789 Str = Str.substr(0, Str.find('\0'));
1794 // These next two are very similar to the above, but also look through PHI
1796 // TODO: See if we can integrate these two together.
1798 /// GetStringLengthH - If we can compute the length of the string pointed to by
1799 /// the specified pointer, return 'len+1'. If we can't, return 0.
1800 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1801 // Look through noop bitcast instructions.
1802 V = V->stripPointerCasts();
1804 // If this is a PHI node, there are two cases: either we have already seen it
1806 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1807 if (!PHIs.insert(PN))
1808 return ~0ULL; // already in the set.
1810 // If it was new, see if all the input strings are the same length.
1811 uint64_t LenSoFar = ~0ULL;
1812 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1813 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1814 if (Len == 0) return 0; // Unknown length -> unknown.
1816 if (Len == ~0ULL) continue;
1818 if (Len != LenSoFar && LenSoFar != ~0ULL)
1819 return 0; // Disagree -> unknown.
1823 // Success, all agree.
1827 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1828 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1829 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1830 if (Len1 == 0) return 0;
1831 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1832 if (Len2 == 0) return 0;
1833 if (Len1 == ~0ULL) return Len2;
1834 if (Len2 == ~0ULL) return Len1;
1835 if (Len1 != Len2) return 0;
1839 // Otherwise, see if we can read the string.
1841 if (!getConstantStringInfo(V, StrData))
1844 return StrData.size()+1;
1847 /// GetStringLength - If we can compute the length of the string pointed to by
1848 /// the specified pointer, return 'len+1'. If we can't, return 0.
1849 uint64_t llvm::GetStringLength(Value *V) {
1850 if (!V->getType()->isPointerTy()) return 0;
1852 SmallPtrSet<PHINode*, 32> PHIs;
1853 uint64_t Len = GetStringLengthH(V, PHIs);
1854 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1855 // an empty string as a length.
1856 return Len == ~0ULL ? 1 : Len;
1860 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
1861 if (!V->getType()->isPointerTy())
1863 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1864 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1865 V = GEP->getPointerOperand();
1866 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1867 V = cast<Operator>(V)->getOperand(0);
1868 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1869 if (GA->mayBeOverridden())
1871 V = GA->getAliasee();
1873 // See if InstructionSimplify knows any relevant tricks.
1874 if (Instruction *I = dyn_cast<Instruction>(V))
1875 // TODO: Acquire a DominatorTree and use it.
1876 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1883 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1889 llvm::GetUnderlyingObjects(Value *V,
1890 SmallVectorImpl<Value *> &Objects,
1891 const DataLayout *TD,
1892 unsigned MaxLookup) {
1893 SmallPtrSet<Value *, 4> Visited;
1894 SmallVector<Value *, 4> Worklist;
1895 Worklist.push_back(V);
1897 Value *P = Worklist.pop_back_val();
1898 P = GetUnderlyingObject(P, TD, MaxLookup);
1900 if (!Visited.insert(P))
1903 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
1904 Worklist.push_back(SI->getTrueValue());
1905 Worklist.push_back(SI->getFalseValue());
1909 if (PHINode *PN = dyn_cast<PHINode>(P)) {
1910 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1911 Worklist.push_back(PN->getIncomingValue(i));
1915 Objects.push_back(P);
1916 } while (!Worklist.empty());
1919 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1920 /// are lifetime markers.
1922 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1923 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1925 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1926 if (!II) return false;
1928 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1929 II->getIntrinsicID() != Intrinsic::lifetime_end)
1935 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1936 const DataLayout *TD) {
1937 const Operator *Inst = dyn_cast<Operator>(V);
1941 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1942 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1946 switch (Inst->getOpcode()) {
1949 case Instruction::UDiv:
1950 case Instruction::URem:
1951 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1952 return isKnownNonZero(Inst->getOperand(1), TD);
1953 case Instruction::SDiv:
1954 case Instruction::SRem: {
1955 Value *Op = Inst->getOperand(1);
1956 // x / y is undefined if y == 0
1957 if (!isKnownNonZero(Op, TD))
1959 // x / y might be undefined if y == -1
1960 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1963 APInt KnownZero(BitWidth, 0);
1964 APInt KnownOne(BitWidth, 0);
1965 ComputeMaskedBits(Op, KnownZero, KnownOne, TD);
1968 case Instruction::Load: {
1969 const LoadInst *LI = cast<LoadInst>(Inst);
1970 if (!LI->isUnordered())
1972 return LI->getPointerOperand()->isDereferenceablePointer();
1974 case Instruction::Call: {
1975 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1976 switch (II->getIntrinsicID()) {
1977 // These synthetic intrinsics have no side-effects, and just mark
1978 // information about their operands.
1979 // FIXME: There are other no-op synthetic instructions that potentially
1980 // should be considered at least *safe* to speculate...
1981 case Intrinsic::dbg_declare:
1982 case Intrinsic::dbg_value:
1985 case Intrinsic::bswap:
1986 case Intrinsic::ctlz:
1987 case Intrinsic::ctpop:
1988 case Intrinsic::cttz:
1989 case Intrinsic::objectsize:
1990 case Intrinsic::sadd_with_overflow:
1991 case Intrinsic::smul_with_overflow:
1992 case Intrinsic::ssub_with_overflow:
1993 case Intrinsic::uadd_with_overflow:
1994 case Intrinsic::umul_with_overflow:
1995 case Intrinsic::usub_with_overflow:
1997 // TODO: some fp intrinsics are marked as having the same error handling
1998 // as libm. They're safe to speculate when they won't error.
1999 // TODO: are convert_{from,to}_fp16 safe?
2000 // TODO: can we list target-specific intrinsics here?
2004 return false; // The called function could have undefined behavior or
2005 // side-effects, even if marked readnone nounwind.
2007 case Instruction::VAArg:
2008 case Instruction::Alloca:
2009 case Instruction::Invoke:
2010 case Instruction::PHI:
2011 case Instruction::Store:
2012 case Instruction::Ret:
2013 case Instruction::Br:
2014 case Instruction::IndirectBr:
2015 case Instruction::Switch:
2016 case Instruction::Unreachable:
2017 case Instruction::Fence:
2018 case Instruction::LandingPad:
2019 case Instruction::AtomicRMW:
2020 case Instruction::AtomicCmpXchg:
2021 case Instruction::Resume:
2022 return false; // Misc instructions which have effects