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.
436 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
438 assert(SrcBitWidth && "SrcBitWidth can't be zero");
439 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
440 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
441 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
442 KnownZero = KnownZero.zextOrTrunc(BitWidth);
443 KnownOne = KnownOne.zextOrTrunc(BitWidth);
444 // Any top bits are known to be zero.
445 if (BitWidth > SrcBitWidth)
446 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
449 case Instruction::BitCast: {
450 Type *SrcTy = I->getOperand(0)->getType();
451 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
452 // TODO: For now, not handling conversions like:
453 // (bitcast i64 %x to <2 x i32>)
454 !I->getType()->isVectorTy()) {
455 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
460 case Instruction::SExt: {
461 // Compute the bits in the result that are not present in the input.
462 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
464 KnownZero = KnownZero.trunc(SrcBitWidth);
465 KnownOne = KnownOne.trunc(SrcBitWidth);
466 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
467 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
468 KnownZero = KnownZero.zext(BitWidth);
469 KnownOne = KnownOne.zext(BitWidth);
471 // If the sign bit of the input is known set or clear, then we know the
472 // top bits of the result.
473 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
474 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
475 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
476 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
479 case Instruction::Shl:
480 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
481 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
482 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
483 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
484 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
485 KnownZero <<= ShiftAmt;
486 KnownOne <<= ShiftAmt;
487 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
491 case Instruction::LShr:
492 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
493 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
494 // Compute the new bits that are at the top now.
495 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
497 // Unsigned shift right.
498 ComputeMaskedBits(I->getOperand(0), KnownZero,KnownOne, TD, Depth+1);
499 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
500 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
501 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
502 // high bits known zero.
503 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
507 case Instruction::AShr:
508 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
509 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
510 // Compute the new bits that are at the top now.
511 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
513 // Signed shift right.
514 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
515 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
516 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
517 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
519 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
520 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
521 KnownZero |= HighBits;
522 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
523 KnownOne |= HighBits;
527 case Instruction::Sub: {
528 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
529 ComputeMaskedBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
530 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
534 case Instruction::Add: {
535 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
536 ComputeMaskedBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
537 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
541 case Instruction::SRem:
542 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
543 APInt RA = Rem->getValue().abs();
544 if (RA.isPowerOf2()) {
545 APInt LowBits = RA - 1;
546 ComputeMaskedBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
548 // The low bits of the first operand are unchanged by the srem.
549 KnownZero = KnownZero2 & LowBits;
550 KnownOne = KnownOne2 & LowBits;
552 // If the first operand is non-negative or has all low bits zero, then
553 // the upper bits are all zero.
554 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
555 KnownZero |= ~LowBits;
557 // If the first operand is negative and not all low bits are zero, then
558 // the upper bits are all one.
559 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
560 KnownOne |= ~LowBits;
562 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
566 // The sign bit is the LHS's sign bit, except when the result of the
567 // remainder is zero.
568 if (KnownZero.isNonNegative()) {
569 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
570 ComputeMaskedBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
572 // If it's known zero, our sign bit is also zero.
573 if (LHSKnownZero.isNegative())
574 KnownZero.setBit(BitWidth - 1);
578 case Instruction::URem: {
579 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
580 APInt RA = Rem->getValue();
581 if (RA.isPowerOf2()) {
582 APInt LowBits = (RA - 1);
583 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD,
585 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
586 KnownZero |= ~LowBits;
592 // Since the result is less than or equal to either operand, any leading
593 // zero bits in either operand must also exist in the result.
594 ComputeMaskedBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
595 ComputeMaskedBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
597 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
598 KnownZero2.countLeadingOnes());
599 KnownOne.clearAllBits();
600 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
604 case Instruction::Alloca: {
605 AllocaInst *AI = cast<AllocaInst>(V);
606 unsigned Align = AI->getAlignment();
607 if (Align == 0 && TD)
608 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
611 KnownZero = APInt::getLowBitsSet(BitWidth, CountTrailingZeros_32(Align));
614 case Instruction::GetElementPtr: {
615 // Analyze all of the subscripts of this getelementptr instruction
616 // to determine if we can prove known low zero bits.
617 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
618 ComputeMaskedBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
620 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
622 gep_type_iterator GTI = gep_type_begin(I);
623 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
624 Value *Index = I->getOperand(i);
625 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
626 // Handle struct member offset arithmetic.
628 const StructLayout *SL = TD->getStructLayout(STy);
629 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
630 uint64_t Offset = SL->getElementOffset(Idx);
631 TrailZ = std::min(TrailZ,
632 CountTrailingZeros_64(Offset));
634 // Handle array index arithmetic.
635 Type *IndexedTy = GTI.getIndexedType();
636 if (!IndexedTy->isSized()) return;
637 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
638 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
639 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
640 ComputeMaskedBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1);
641 TrailZ = std::min(TrailZ,
642 unsigned(CountTrailingZeros_64(TypeSize) +
643 LocalKnownZero.countTrailingOnes()));
647 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
650 case Instruction::PHI: {
651 PHINode *P = cast<PHINode>(I);
652 // Handle the case of a simple two-predecessor recurrence PHI.
653 // There's a lot more that could theoretically be done here, but
654 // this is sufficient to catch some interesting cases.
655 if (P->getNumIncomingValues() == 2) {
656 for (unsigned i = 0; i != 2; ++i) {
657 Value *L = P->getIncomingValue(i);
658 Value *R = P->getIncomingValue(!i);
659 Operator *LU = dyn_cast<Operator>(L);
662 unsigned Opcode = LU->getOpcode();
663 // Check for operations that have the property that if
664 // both their operands have low zero bits, the result
665 // will have low zero bits.
666 if (Opcode == Instruction::Add ||
667 Opcode == Instruction::Sub ||
668 Opcode == Instruction::And ||
669 Opcode == Instruction::Or ||
670 Opcode == Instruction::Mul) {
671 Value *LL = LU->getOperand(0);
672 Value *LR = LU->getOperand(1);
673 // Find a recurrence.
680 // Ok, we have a PHI of the form L op= R. Check for low
682 ComputeMaskedBits(R, KnownZero2, KnownOne2, TD, Depth+1);
684 // We need to take the minimum number of known bits
685 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
686 ComputeMaskedBits(L, KnownZero3, KnownOne3, TD, Depth+1);
688 KnownZero = APInt::getLowBitsSet(BitWidth,
689 std::min(KnownZero2.countTrailingOnes(),
690 KnownZero3.countTrailingOnes()));
696 // Unreachable blocks may have zero-operand PHI nodes.
697 if (P->getNumIncomingValues() == 0)
700 // Otherwise take the unions of the known bit sets of the operands,
701 // taking conservative care to avoid excessive recursion.
702 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
703 // Skip if every incoming value references to ourself.
704 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
707 KnownZero = APInt::getAllOnesValue(BitWidth);
708 KnownOne = APInt::getAllOnesValue(BitWidth);
709 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
710 // Skip direct self references.
711 if (P->getIncomingValue(i) == P) continue;
713 KnownZero2 = APInt(BitWidth, 0);
714 KnownOne2 = APInt(BitWidth, 0);
715 // Recurse, but cap the recursion to one level, because we don't
716 // want to waste time spinning around in loops.
717 ComputeMaskedBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
719 KnownZero &= KnownZero2;
720 KnownOne &= KnownOne2;
721 // If all bits have been ruled out, there's no need to check
723 if (!KnownZero && !KnownOne)
729 case Instruction::Call:
730 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
731 switch (II->getIntrinsicID()) {
733 case Intrinsic::ctlz:
734 case Intrinsic::cttz: {
735 unsigned LowBits = Log2_32(BitWidth)+1;
736 // If this call is undefined for 0, the result will be less than 2^n.
737 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
739 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
742 case Intrinsic::ctpop: {
743 unsigned LowBits = Log2_32(BitWidth)+1;
744 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
747 case Intrinsic::x86_sse42_crc32_64_8:
748 case Intrinsic::x86_sse42_crc32_64_64:
749 KnownZero = APInt::getHighBitsSet(64, 32);
754 case Instruction::ExtractValue:
755 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
756 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
757 if (EVI->getNumIndices() != 1) break;
758 if (EVI->getIndices()[0] == 0) {
759 switch (II->getIntrinsicID()) {
761 case Intrinsic::uadd_with_overflow:
762 case Intrinsic::sadd_with_overflow:
763 ComputeMaskedBitsAddSub(true, II->getArgOperand(0),
764 II->getArgOperand(1), false, KnownZero,
765 KnownOne, KnownZero2, KnownOne2, TD, Depth);
767 case Intrinsic::usub_with_overflow:
768 case Intrinsic::ssub_with_overflow:
769 ComputeMaskedBitsAddSub(false, II->getArgOperand(0),
770 II->getArgOperand(1), false, KnownZero,
771 KnownOne, KnownZero2, KnownOne2, TD, Depth);
773 case Intrinsic::umul_with_overflow:
774 case Intrinsic::smul_with_overflow:
775 ComputeMaskedBitsMul(II->getArgOperand(0), II->getArgOperand(1),
776 false, KnownZero, KnownOne,
777 KnownZero2, KnownOne2, TD, Depth);
785 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
786 /// one. Convenience wrapper around ComputeMaskedBits.
787 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
788 const DataLayout *TD, unsigned Depth) {
789 unsigned BitWidth = getBitWidth(V->getType(), TD);
795 APInt ZeroBits(BitWidth, 0);
796 APInt OneBits(BitWidth, 0);
797 ComputeMaskedBits(V, ZeroBits, OneBits, TD, Depth);
798 KnownOne = OneBits[BitWidth - 1];
799 KnownZero = ZeroBits[BitWidth - 1];
802 /// isKnownToBeAPowerOfTwo - Return true if the given value is known to have exactly one
803 /// bit set when defined. For vectors return true if every element is known to
804 /// be a power of two when defined. Supports values with integer or pointer
805 /// types and vectors of integers.
806 bool llvm::isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth) {
807 if (Constant *C = dyn_cast<Constant>(V)) {
808 if (C->isNullValue())
810 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
811 return CI->getValue().isPowerOf2();
812 // TODO: Handle vector constants.
815 // 1 << X is clearly a power of two if the one is not shifted off the end. If
816 // it is shifted off the end then the result is undefined.
817 if (match(V, m_Shl(m_One(), m_Value())))
820 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
821 // bottom. If it is shifted off the bottom then the result is undefined.
822 if (match(V, m_LShr(m_SignBit(), m_Value())))
825 // The remaining tests are all recursive, so bail out if we hit the limit.
826 if (Depth++ == MaxDepth)
829 Value *X = 0, *Y = 0;
830 // A shift of a power of two is a power of two or zero.
831 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
832 match(V, m_Shr(m_Value(X), m_Value()))))
833 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth);
835 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
836 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth);
838 if (SelectInst *SI = dyn_cast<SelectInst>(V))
839 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth) &&
840 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth);
842 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
843 // A power of two and'd with anything is a power of two or zero.
844 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth) ||
845 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth))
847 // X & (-X) is always a power of two or zero.
848 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
853 // An exact divide or right shift can only shift off zero bits, so the result
854 // is a power of two only if the first operand is a power of two and not
855 // copying a sign bit (sdiv int_min, 2).
856 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
857 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
858 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, Depth);
864 /// \brief Test whether a GEP's result is known to be non-null.
866 /// Uses properties inherent in a GEP to try to determine whether it is known
869 /// Currently this routine does not support vector GEPs.
870 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
872 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
875 // FIXME: Support vector-GEPs.
876 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
878 // If the base pointer is non-null, we cannot walk to a null address with an
879 // inbounds GEP in address space zero.
880 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth))
883 // Past this, if we don't have DataLayout, we can't do much.
887 // Walk the GEP operands and see if any operand introduces a non-zero offset.
888 // If so, then the GEP cannot produce a null pointer, as doing so would
889 // inherently violate the inbounds contract within address space zero.
890 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
892 // Struct types are easy -- they must always be indexed by a constant.
893 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
894 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
895 unsigned ElementIdx = OpC->getZExtValue();
896 const StructLayout *SL = DL->getStructLayout(STy);
897 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
898 if (ElementOffset > 0)
903 // If we have a zero-sized type, the index doesn't matter. Keep looping.
904 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
907 // Fast path the constant operand case both for efficiency and so we don't
908 // increment Depth when just zipping down an all-constant GEP.
909 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
915 // We post-increment Depth here because while isKnownNonZero increments it
916 // as well, when we pop back up that increment won't persist. We don't want
917 // to recurse 10k times just because we have 10k GEP operands. We don't
918 // bail completely out because we want to handle constant GEPs regardless
920 if (Depth++ >= MaxDepth)
923 if (isKnownNonZero(GTI.getOperand(), DL, Depth))
930 /// isKnownNonZero - Return true if the given value is known to be non-zero
931 /// when defined. For vectors return true if every element is known to be
932 /// non-zero when defined. Supports values with integer or pointer type and
933 /// vectors of integers.
934 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth) {
935 if (Constant *C = dyn_cast<Constant>(V)) {
936 if (C->isNullValue())
938 if (isa<ConstantInt>(C))
939 // Must be non-zero due to null test above.
941 // TODO: Handle vectors
945 // The remaining tests are all recursive, so bail out if we hit the limit.
946 if (Depth++ >= MaxDepth)
949 // Check for pointer simplifications.
950 if (V->getType()->isPointerTy()) {
951 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
952 if (isGEPKnownNonNull(GEP, TD, Depth))
956 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
958 // X | Y != 0 if X != 0 or Y != 0.
959 Value *X = 0, *Y = 0;
960 if (match(V, m_Or(m_Value(X), m_Value(Y))))
961 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
963 // ext X != 0 if X != 0.
964 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
965 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
967 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
968 // if the lowest bit is shifted off the end.
969 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
970 // shl nuw can't remove any non-zero bits.
971 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
972 if (BO->hasNoUnsignedWrap())
973 return isKnownNonZero(X, TD, Depth);
975 APInt KnownZero(BitWidth, 0);
976 APInt KnownOne(BitWidth, 0);
977 ComputeMaskedBits(X, KnownZero, KnownOne, TD, Depth);
981 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
982 // defined if the sign bit is shifted off the end.
983 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
984 // shr exact can only shift out zero bits.
985 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
987 return isKnownNonZero(X, TD, Depth);
989 bool XKnownNonNegative, XKnownNegative;
990 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
994 // div exact can only produce a zero if the dividend is zero.
995 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
996 return isKnownNonZero(X, TD, Depth);
999 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1000 bool XKnownNonNegative, XKnownNegative;
1001 bool YKnownNonNegative, YKnownNegative;
1002 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
1003 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
1005 // If X and Y are both non-negative (as signed values) then their sum is not
1006 // zero unless both X and Y are zero.
1007 if (XKnownNonNegative && YKnownNonNegative)
1008 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
1011 // If X and Y are both negative (as signed values) then their sum is not
1012 // zero unless both X and Y equal INT_MIN.
1013 if (BitWidth && XKnownNegative && YKnownNegative) {
1014 APInt KnownZero(BitWidth, 0);
1015 APInt KnownOne(BitWidth, 0);
1016 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1017 // The sign bit of X is set. If some other bit is set then X is not equal
1019 ComputeMaskedBits(X, KnownZero, KnownOne, TD, Depth);
1020 if ((KnownOne & Mask) != 0)
1022 // The sign bit of Y is set. If some other bit is set then Y is not equal
1024 ComputeMaskedBits(Y, KnownZero, KnownOne, TD, Depth);
1025 if ((KnownOne & Mask) != 0)
1029 // The sum of a non-negative number and a power of two is not zero.
1030 if (XKnownNonNegative && isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth))
1032 if (YKnownNonNegative && isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth))
1036 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1037 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1038 // If X and Y are non-zero then so is X * Y as long as the multiplication
1039 // does not overflow.
1040 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1041 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
1044 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1045 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1046 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
1047 isKnownNonZero(SI->getFalseValue(), TD, Depth))
1051 if (!BitWidth) return false;
1052 APInt KnownZero(BitWidth, 0);
1053 APInt KnownOne(BitWidth, 0);
1054 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1055 return KnownOne != 0;
1058 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1059 /// this predicate to simplify operations downstream. Mask is known to be zero
1060 /// for bits that V cannot have.
1062 /// This function is defined on values with integer type, values with pointer
1063 /// type (but only if TD is non-null), and vectors of integers. In the case
1064 /// where V is a vector, the mask, known zero, and known one values are the
1065 /// same width as the vector element, and the bit is set only if it is true
1066 /// for all of the elements in the vector.
1067 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
1068 const DataLayout *TD, unsigned Depth) {
1069 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1070 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1071 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1072 return (KnownZero & Mask) == Mask;
1077 /// ComputeNumSignBits - Return the number of times the sign bit of the
1078 /// register is replicated into the other bits. We know that at least 1 bit
1079 /// is always equal to the sign bit (itself), but other cases can give us
1080 /// information. For example, immediately after an "ashr X, 2", we know that
1081 /// the top 3 bits are all equal to each other, so we return 3.
1083 /// 'Op' must have a scalar integer type.
1085 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
1087 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1088 "ComputeNumSignBits requires a DataLayout object to operate "
1089 "on non-integer values!");
1090 Type *Ty = V->getType();
1091 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1092 Ty->getScalarSizeInBits();
1094 unsigned FirstAnswer = 1;
1096 // Note that ConstantInt is handled by the general ComputeMaskedBits case
1100 return 1; // Limit search depth.
1102 Operator *U = dyn_cast<Operator>(V);
1103 switch (Operator::getOpcode(V)) {
1105 case Instruction::SExt:
1106 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1107 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
1109 case Instruction::AShr: {
1110 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1111 // ashr X, C -> adds C sign bits. Vectors too.
1113 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1114 Tmp += ShAmt->getZExtValue();
1115 if (Tmp > TyBits) Tmp = TyBits;
1119 case Instruction::Shl: {
1121 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1122 // shl destroys sign bits.
1123 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1124 Tmp2 = ShAmt->getZExtValue();
1125 if (Tmp2 >= TyBits || // Bad shift.
1126 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1131 case Instruction::And:
1132 case Instruction::Or:
1133 case Instruction::Xor: // NOT is handled here.
1134 // Logical binary ops preserve the number of sign bits at the worst.
1135 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1137 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1138 FirstAnswer = std::min(Tmp, Tmp2);
1139 // We computed what we know about the sign bits as our first
1140 // answer. Now proceed to the generic code that uses
1141 // ComputeMaskedBits, and pick whichever answer is better.
1145 case Instruction::Select:
1146 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1147 if (Tmp == 1) return 1; // Early out.
1148 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1149 return std::min(Tmp, Tmp2);
1151 case Instruction::Add:
1152 // Add can have at most one carry bit. Thus we know that the output
1153 // is, at worst, one more bit than the inputs.
1154 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1155 if (Tmp == 1) return 1; // Early out.
1157 // Special case decrementing a value (ADD X, -1):
1158 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1159 if (CRHS->isAllOnesValue()) {
1160 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1161 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
1163 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1165 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1168 // If we are subtracting one from a positive number, there is no carry
1169 // out of the result.
1170 if (KnownZero.isNegative())
1174 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1175 if (Tmp2 == 1) return 1;
1176 return std::min(Tmp, Tmp2)-1;
1178 case Instruction::Sub:
1179 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1180 if (Tmp2 == 1) return 1;
1183 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1184 if (CLHS->isNullValue()) {
1185 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1186 ComputeMaskedBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
1187 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1189 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1192 // If the input is known to be positive (the sign bit is known clear),
1193 // the output of the NEG has the same number of sign bits as the input.
1194 if (KnownZero.isNegative())
1197 // Otherwise, we treat this like a SUB.
1200 // Sub can have at most one carry bit. Thus we know that the output
1201 // is, at worst, one more bit than the inputs.
1202 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1203 if (Tmp == 1) return 1; // Early out.
1204 return std::min(Tmp, Tmp2)-1;
1206 case Instruction::PHI: {
1207 PHINode *PN = cast<PHINode>(U);
1208 // Don't analyze large in-degree PHIs.
1209 if (PN->getNumIncomingValues() > 4) break;
1211 // Take the minimum of all incoming values. This can't infinitely loop
1212 // because of our depth threshold.
1213 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1214 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1215 if (Tmp == 1) return Tmp;
1217 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1222 case Instruction::Trunc:
1223 // FIXME: it's tricky to do anything useful for this, but it is an important
1224 // case for targets like X86.
1228 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1229 // use this information.
1230 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1232 ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
1234 if (KnownZero.isNegative()) { // sign bit is 0
1236 } else if (KnownOne.isNegative()) { // sign bit is 1;
1243 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1244 // the number of identical bits in the top of the input value.
1246 Mask <<= Mask.getBitWidth()-TyBits;
1247 // Return # leading zeros. We use 'min' here in case Val was zero before
1248 // shifting. We don't want to return '64' as for an i32 "0".
1249 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1252 /// ComputeMultiple - This function computes the integer multiple of Base that
1253 /// equals V. If successful, it returns true and returns the multiple in
1254 /// Multiple. If unsuccessful, it returns false. It looks
1255 /// through SExt instructions only if LookThroughSExt is true.
1256 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1257 bool LookThroughSExt, unsigned Depth) {
1258 const unsigned MaxDepth = 6;
1260 assert(V && "No Value?");
1261 assert(Depth <= MaxDepth && "Limit Search Depth");
1262 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1264 Type *T = V->getType();
1266 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1276 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1277 Constant *BaseVal = ConstantInt::get(T, Base);
1278 if (CO && CO == BaseVal) {
1280 Multiple = ConstantInt::get(T, 1);
1284 if (CI && CI->getZExtValue() % Base == 0) {
1285 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1289 if (Depth == MaxDepth) return false; // Limit search depth.
1291 Operator *I = dyn_cast<Operator>(V);
1292 if (!I) return false;
1294 switch (I->getOpcode()) {
1296 case Instruction::SExt:
1297 if (!LookThroughSExt) return false;
1298 // otherwise fall through to ZExt
1299 case Instruction::ZExt:
1300 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1301 LookThroughSExt, Depth+1);
1302 case Instruction::Shl:
1303 case Instruction::Mul: {
1304 Value *Op0 = I->getOperand(0);
1305 Value *Op1 = I->getOperand(1);
1307 if (I->getOpcode() == Instruction::Shl) {
1308 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1309 if (!Op1CI) return false;
1310 // Turn Op0 << Op1 into Op0 * 2^Op1
1311 APInt Op1Int = Op1CI->getValue();
1312 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1313 APInt API(Op1Int.getBitWidth(), 0);
1314 API.setBit(BitToSet);
1315 Op1 = ConstantInt::get(V->getContext(), API);
1319 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1320 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1321 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1322 if (Op1C->getType()->getPrimitiveSizeInBits() <
1323 MulC->getType()->getPrimitiveSizeInBits())
1324 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1325 if (Op1C->getType()->getPrimitiveSizeInBits() >
1326 MulC->getType()->getPrimitiveSizeInBits())
1327 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1329 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1330 Multiple = ConstantExpr::getMul(MulC, Op1C);
1334 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1335 if (Mul0CI->getValue() == 1) {
1336 // V == Base * Op1, so return Op1
1343 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1344 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1345 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1346 if (Op0C->getType()->getPrimitiveSizeInBits() <
1347 MulC->getType()->getPrimitiveSizeInBits())
1348 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1349 if (Op0C->getType()->getPrimitiveSizeInBits() >
1350 MulC->getType()->getPrimitiveSizeInBits())
1351 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1353 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1354 Multiple = ConstantExpr::getMul(MulC, Op0C);
1358 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1359 if (Mul1CI->getValue() == 1) {
1360 // V == Base * Op0, so return Op0
1368 // We could not determine if V is a multiple of Base.
1372 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1373 /// value is never equal to -0.0.
1375 /// NOTE: this function will need to be revisited when we support non-default
1378 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1379 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1380 return !CFP->getValueAPF().isNegZero();
1383 return 1; // Limit search depth.
1385 const Operator *I = dyn_cast<Operator>(V);
1386 if (I == 0) return false;
1388 // Check if the nsz fast-math flag is set
1389 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
1390 if (FPO->hasNoSignedZeros())
1393 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1394 if (I->getOpcode() == Instruction::FAdd &&
1395 isa<ConstantFP>(I->getOperand(1)) &&
1396 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1399 // sitofp and uitofp turn into +0.0 for zero.
1400 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1403 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1404 // sqrt(-0.0) = -0.0, no other negative results are possible.
1405 if (II->getIntrinsicID() == Intrinsic::sqrt)
1406 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1408 if (const CallInst *CI = dyn_cast<CallInst>(I))
1409 if (const Function *F = CI->getCalledFunction()) {
1410 if (F->isDeclaration()) {
1412 if (F->getName() == "abs") return true;
1413 // fabs[lf](x) != -0.0
1414 if (F->getName() == "fabs") return true;
1415 if (F->getName() == "fabsf") return true;
1416 if (F->getName() == "fabsl") return true;
1417 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1418 F->getName() == "sqrtl")
1419 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1426 /// isBytewiseValue - If the specified value can be set by repeating the same
1427 /// byte in memory, return the i8 value that it is represented with. This is
1428 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1429 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1430 /// byte store (e.g. i16 0x1234), return null.
1431 Value *llvm::isBytewiseValue(Value *V) {
1432 // All byte-wide stores are splatable, even of arbitrary variables.
1433 if (V->getType()->isIntegerTy(8)) return V;
1435 // Handle 'null' ConstantArrayZero etc.
1436 if (Constant *C = dyn_cast<Constant>(V))
1437 if (C->isNullValue())
1438 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1440 // Constant float and double values can be handled as integer values if the
1441 // corresponding integer value is "byteable". An important case is 0.0.
1442 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1443 if (CFP->getType()->isFloatTy())
1444 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1445 if (CFP->getType()->isDoubleTy())
1446 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1447 // Don't handle long double formats, which have strange constraints.
1450 // We can handle constant integers that are power of two in size and a
1451 // multiple of 8 bits.
1452 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1453 unsigned Width = CI->getBitWidth();
1454 if (isPowerOf2_32(Width) && Width > 8) {
1455 // We can handle this value if the recursive binary decomposition is the
1456 // same at all levels.
1457 APInt Val = CI->getValue();
1459 while (Val.getBitWidth() != 8) {
1460 unsigned NextWidth = Val.getBitWidth()/2;
1461 Val2 = Val.lshr(NextWidth);
1462 Val2 = Val2.trunc(Val.getBitWidth()/2);
1463 Val = Val.trunc(Val.getBitWidth()/2);
1465 // If the top/bottom halves aren't the same, reject it.
1469 return ConstantInt::get(V->getContext(), Val);
1473 // A ConstantDataArray/Vector is splatable if all its members are equal and
1475 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1476 Value *Elt = CA->getElementAsConstant(0);
1477 Value *Val = isBytewiseValue(Elt);
1481 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1482 if (CA->getElementAsConstant(I) != Elt)
1488 // Conceptually, we could handle things like:
1489 // %a = zext i8 %X to i16
1490 // %b = shl i16 %a, 8
1491 // %c = or i16 %a, %b
1492 // but until there is an example that actually needs this, it doesn't seem
1493 // worth worrying about.
1498 // This is the recursive version of BuildSubAggregate. It takes a few different
1499 // arguments. Idxs is the index within the nested struct From that we are
1500 // looking at now (which is of type IndexedType). IdxSkip is the number of
1501 // indices from Idxs that should be left out when inserting into the resulting
1502 // struct. To is the result struct built so far, new insertvalue instructions
1504 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1505 SmallVector<unsigned, 10> &Idxs,
1507 Instruction *InsertBefore) {
1508 llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1510 // Save the original To argument so we can modify it
1512 // General case, the type indexed by Idxs is a struct
1513 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1514 // Process each struct element recursively
1517 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1521 // Couldn't find any inserted value for this index? Cleanup
1522 while (PrevTo != OrigTo) {
1523 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1524 PrevTo = Del->getAggregateOperand();
1525 Del->eraseFromParent();
1527 // Stop processing elements
1531 // If we successfully found a value for each of our subaggregates
1535 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1536 // the struct's elements had a value that was inserted directly. In the latter
1537 // case, perhaps we can't determine each of the subelements individually, but
1538 // we might be able to find the complete struct somewhere.
1540 // Find the value that is at that particular spot
1541 Value *V = FindInsertedValue(From, Idxs);
1546 // Insert the value in the new (sub) aggregrate
1547 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1548 "tmp", InsertBefore);
1551 // This helper takes a nested struct and extracts a part of it (which is again a
1552 // struct) into a new value. For example, given the struct:
1553 // { a, { b, { c, d }, e } }
1554 // and the indices "1, 1" this returns
1557 // It does this by inserting an insertvalue for each element in the resulting
1558 // struct, as opposed to just inserting a single struct. This will only work if
1559 // each of the elements of the substruct are known (ie, inserted into From by an
1560 // insertvalue instruction somewhere).
1562 // All inserted insertvalue instructions are inserted before InsertBefore
1563 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1564 Instruction *InsertBefore) {
1565 assert(InsertBefore && "Must have someplace to insert!");
1566 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1568 Value *To = UndefValue::get(IndexedType);
1569 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1570 unsigned IdxSkip = Idxs.size();
1572 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1575 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1576 /// the scalar value indexed is already around as a register, for example if it
1577 /// were inserted directly into the aggregrate.
1579 /// If InsertBefore is not null, this function will duplicate (modified)
1580 /// insertvalues when a part of a nested struct is extracted.
1581 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1582 Instruction *InsertBefore) {
1583 // Nothing to index? Just return V then (this is useful at the end of our
1585 if (idx_range.empty())
1587 // We have indices, so V should have an indexable type.
1588 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1589 "Not looking at a struct or array?");
1590 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1591 "Invalid indices for type?");
1593 if (Constant *C = dyn_cast<Constant>(V)) {
1594 C = C->getAggregateElement(idx_range[0]);
1595 if (C == 0) return 0;
1596 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1599 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1600 // Loop the indices for the insertvalue instruction in parallel with the
1601 // requested indices
1602 const unsigned *req_idx = idx_range.begin();
1603 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1604 i != e; ++i, ++req_idx) {
1605 if (req_idx == idx_range.end()) {
1606 // We can't handle this without inserting insertvalues
1610 // The requested index identifies a part of a nested aggregate. Handle
1611 // this specially. For example,
1612 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1613 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1614 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1615 // This can be changed into
1616 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1617 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1618 // which allows the unused 0,0 element from the nested struct to be
1620 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1624 // This insert value inserts something else than what we are looking for.
1625 // See if the (aggregrate) value inserted into has the value we are
1626 // looking for, then.
1628 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1631 // If we end up here, the indices of the insertvalue match with those
1632 // requested (though possibly only partially). Now we recursively look at
1633 // the inserted value, passing any remaining indices.
1634 return FindInsertedValue(I->getInsertedValueOperand(),
1635 makeArrayRef(req_idx, idx_range.end()),
1639 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1640 // If we're extracting a value from an aggregrate that was extracted from
1641 // something else, we can extract from that something else directly instead.
1642 // However, we will need to chain I's indices with the requested indices.
1644 // Calculate the number of indices required
1645 unsigned size = I->getNumIndices() + idx_range.size();
1646 // Allocate some space to put the new indices in
1647 SmallVector<unsigned, 5> Idxs;
1649 // Add indices from the extract value instruction
1650 Idxs.append(I->idx_begin(), I->idx_end());
1652 // Add requested indices
1653 Idxs.append(idx_range.begin(), idx_range.end());
1655 assert(Idxs.size() == size
1656 && "Number of indices added not correct?");
1658 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1660 // Otherwise, we don't know (such as, extracting from a function return value
1661 // or load instruction)
1665 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1666 /// it can be expressed as a base pointer plus a constant offset. Return the
1667 /// base and offset to the caller.
1668 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1669 const DataLayout &TD) {
1670 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1671 if (PtrOp == 0 || Ptr->getType()->isVectorTy())
1674 // Just look through bitcasts.
1675 if (PtrOp->getOpcode() == Instruction::BitCast)
1676 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1678 // If this is a GEP with constant indices, we can look through it.
1679 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1680 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1682 gep_type_iterator GTI = gep_type_begin(GEP);
1683 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1685 ConstantInt *OpC = cast<ConstantInt>(*I);
1686 if (OpC->isZero()) continue;
1688 // Handle a struct and array indices which add their offset to the pointer.
1689 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1690 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1692 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1693 Offset += OpC->getSExtValue()*Size;
1697 // Re-sign extend from the pointer size if needed to get overflow edge cases
1699 unsigned PtrSize = TD.getPointerSizeInBits();
1701 Offset = SignExtend64(Offset, PtrSize);
1703 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1707 /// getConstantStringInfo - This function computes the length of a
1708 /// null-terminated C string pointed to by V. If successful, it returns true
1709 /// and returns the string in Str. If unsuccessful, it returns false.
1710 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
1711 uint64_t Offset, bool TrimAtNul) {
1714 // Look through bitcast instructions and geps.
1715 V = V->stripPointerCasts();
1717 // If the value is a GEP instructionor constant expression, treat it as an
1719 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1720 // Make sure the GEP has exactly three arguments.
1721 if (GEP->getNumOperands() != 3)
1724 // Make sure the index-ee is a pointer to array of i8.
1725 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1726 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1727 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1730 // Check to make sure that the first operand of the GEP is an integer and
1731 // has value 0 so that we are sure we're indexing into the initializer.
1732 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1733 if (FirstIdx == 0 || !FirstIdx->isZero())
1736 // If the second index isn't a ConstantInt, then this is a variable index
1737 // into the array. If this occurs, we can't say anything meaningful about
1739 uint64_t StartIdx = 0;
1740 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1741 StartIdx = CI->getZExtValue();
1744 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
1747 // The GEP instruction, constant or instruction, must reference a global
1748 // variable that is a constant and is initialized. The referenced constant
1749 // initializer is the array that we'll use for optimization.
1750 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
1751 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1754 // Handle the all-zeros case
1755 if (GV->getInitializer()->isNullValue()) {
1756 // This is a degenerate case. The initializer is constant zero so the
1757 // length of the string must be zero.
1762 // Must be a Constant Array
1763 const ConstantDataArray *Array =
1764 dyn_cast<ConstantDataArray>(GV->getInitializer());
1765 if (Array == 0 || !Array->isString())
1768 // Get the number of elements in the array
1769 uint64_t NumElts = Array->getType()->getArrayNumElements();
1771 // Start out with the entire array in the StringRef.
1772 Str = Array->getAsString();
1774 if (Offset > NumElts)
1777 // Skip over 'offset' bytes.
1778 Str = Str.substr(Offset);
1781 // Trim off the \0 and anything after it. If the array is not nul
1782 // terminated, we just return the whole end of string. The client may know
1783 // some other way that the string is length-bound.
1784 Str = Str.substr(0, Str.find('\0'));
1789 // These next two are very similar to the above, but also look through PHI
1791 // TODO: See if we can integrate these two together.
1793 /// GetStringLengthH - If we can compute the length of the string pointed to by
1794 /// the specified pointer, return 'len+1'. If we can't, return 0.
1795 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1796 // Look through noop bitcast instructions.
1797 V = V->stripPointerCasts();
1799 // If this is a PHI node, there are two cases: either we have already seen it
1801 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1802 if (!PHIs.insert(PN))
1803 return ~0ULL; // already in the set.
1805 // If it was new, see if all the input strings are the same length.
1806 uint64_t LenSoFar = ~0ULL;
1807 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1808 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1809 if (Len == 0) return 0; // Unknown length -> unknown.
1811 if (Len == ~0ULL) continue;
1813 if (Len != LenSoFar && LenSoFar != ~0ULL)
1814 return 0; // Disagree -> unknown.
1818 // Success, all agree.
1822 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1823 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1824 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1825 if (Len1 == 0) return 0;
1826 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1827 if (Len2 == 0) return 0;
1828 if (Len1 == ~0ULL) return Len2;
1829 if (Len2 == ~0ULL) return Len1;
1830 if (Len1 != Len2) return 0;
1834 // Otherwise, see if we can read the string.
1836 if (!getConstantStringInfo(V, StrData))
1839 return StrData.size()+1;
1842 /// GetStringLength - If we can compute the length of the string pointed to by
1843 /// the specified pointer, return 'len+1'. If we can't, return 0.
1844 uint64_t llvm::GetStringLength(Value *V) {
1845 if (!V->getType()->isPointerTy()) return 0;
1847 SmallPtrSet<PHINode*, 32> PHIs;
1848 uint64_t Len = GetStringLengthH(V, PHIs);
1849 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1850 // an empty string as a length.
1851 return Len == ~0ULL ? 1 : Len;
1855 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
1856 if (!V->getType()->isPointerTy())
1858 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1859 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1860 V = GEP->getPointerOperand();
1861 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1862 V = cast<Operator>(V)->getOperand(0);
1863 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1864 if (GA->mayBeOverridden())
1866 V = GA->getAliasee();
1868 // See if InstructionSimplify knows any relevant tricks.
1869 if (Instruction *I = dyn_cast<Instruction>(V))
1870 // TODO: Acquire a DominatorTree and use it.
1871 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1878 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1884 llvm::GetUnderlyingObjects(Value *V,
1885 SmallVectorImpl<Value *> &Objects,
1886 const DataLayout *TD,
1887 unsigned MaxLookup) {
1888 SmallPtrSet<Value *, 4> Visited;
1889 SmallVector<Value *, 4> Worklist;
1890 Worklist.push_back(V);
1892 Value *P = Worklist.pop_back_val();
1893 P = GetUnderlyingObject(P, TD, MaxLookup);
1895 if (!Visited.insert(P))
1898 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
1899 Worklist.push_back(SI->getTrueValue());
1900 Worklist.push_back(SI->getFalseValue());
1904 if (PHINode *PN = dyn_cast<PHINode>(P)) {
1905 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1906 Worklist.push_back(PN->getIncomingValue(i));
1910 Objects.push_back(P);
1911 } while (!Worklist.empty());
1914 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1915 /// are lifetime markers.
1917 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1918 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1920 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1921 if (!II) return false;
1923 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1924 II->getIntrinsicID() != Intrinsic::lifetime_end)
1930 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1931 const DataLayout *TD) {
1932 const Operator *Inst = dyn_cast<Operator>(V);
1936 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1937 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1941 switch (Inst->getOpcode()) {
1944 case Instruction::UDiv:
1945 case Instruction::URem:
1946 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1947 return isKnownNonZero(Inst->getOperand(1), TD);
1948 case Instruction::SDiv:
1949 case Instruction::SRem: {
1950 Value *Op = Inst->getOperand(1);
1951 // x / y is undefined if y == 0
1952 if (!isKnownNonZero(Op, TD))
1954 // x / y might be undefined if y == -1
1955 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1958 APInt KnownZero(BitWidth, 0);
1959 APInt KnownOne(BitWidth, 0);
1960 ComputeMaskedBits(Op, KnownZero, KnownOne, TD);
1963 case Instruction::Load: {
1964 const LoadInst *LI = cast<LoadInst>(Inst);
1965 if (!LI->isUnordered())
1967 return LI->getPointerOperand()->isDereferenceablePointer();
1969 case Instruction::Call: {
1970 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1971 switch (II->getIntrinsicID()) {
1972 // These synthetic intrinsics have no side-effects, and just mark
1973 // information about their operands.
1974 // FIXME: There are other no-op synthetic instructions that potentially
1975 // should be considered at least *safe* to speculate...
1976 case Intrinsic::dbg_declare:
1977 case Intrinsic::dbg_value:
1980 case Intrinsic::bswap:
1981 case Intrinsic::ctlz:
1982 case Intrinsic::ctpop:
1983 case Intrinsic::cttz:
1984 case Intrinsic::objectsize:
1985 case Intrinsic::sadd_with_overflow:
1986 case Intrinsic::smul_with_overflow:
1987 case Intrinsic::ssub_with_overflow:
1988 case Intrinsic::uadd_with_overflow:
1989 case Intrinsic::umul_with_overflow:
1990 case Intrinsic::usub_with_overflow:
1992 // TODO: some fp intrinsics are marked as having the same error handling
1993 // as libm. They're safe to speculate when they won't error.
1994 // TODO: are convert_{from,to}_fp16 safe?
1995 // TODO: can we list target-specific intrinsics here?
1999 return false; // The called function could have undefined behavior or
2000 // side-effects, even if marked readnone nounwind.
2002 case Instruction::VAArg:
2003 case Instruction::Alloca:
2004 case Instruction::Invoke:
2005 case Instruction::PHI:
2006 case Instruction::Store:
2007 case Instruction::Ret:
2008 case Instruction::Br:
2009 case Instruction::IndirectBr:
2010 case Instruction::Switch:
2011 case Instruction::Unreachable:
2012 case Instruction::Fence:
2013 case Instruction::LandingPad:
2014 case Instruction::AtomicRMW:
2015 case Instruction::AtomicCmpXchg:
2016 case Instruction::Resume:
2017 return false; // Misc instructions which have effects