1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Constants.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/GlobalVariable.h"
20 #include "llvm/GlobalAlias.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/Operator.h"
24 #include "llvm/Target/TargetData.h"
25 #include "llvm/Support/GetElementPtrTypeIterator.h"
26 #include "llvm/Support/MathExtras.h"
27 #include "llvm/Support/PatternMatch.h"
28 #include "llvm/ADT/SmallPtrSet.h"
31 using namespace llvm::PatternMatch;
33 const unsigned MaxDepth = 6;
35 /// getBitWidth - Returns the bitwidth of the given scalar or pointer type (if
36 /// unknown returns 0). For vector types, returns the element type's bitwidth.
37 static unsigned getBitWidth(Type *Ty, const TargetData *TD) {
38 if (unsigned BitWidth = Ty->getScalarSizeInBits())
40 assert(isa<PointerType>(Ty) && "Expected a pointer type!");
41 return TD ? TD->getPointerSizeInBits() : 0;
44 static void ComputeMaskedBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
46 APInt &KnownZero, APInt &KnownOne,
47 APInt &KnownZero2, APInt &KnownOne2,
48 const TargetData *TD, unsigned Depth) {
50 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
51 // We know that the top bits of C-X are clear if X contains less bits
52 // than C (i.e. no wrap-around can happen). For example, 20-X is
53 // positive if we can prove that X is >= 0 and < 16.
54 if (!CLHS->getValue().isNegative()) {
55 unsigned BitWidth = Mask.getBitWidth();
56 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
57 // NLZ can't be BitWidth with no sign bit
58 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
59 llvm::ComputeMaskedBits(Op1, MaskV, KnownZero2, KnownOne2, TD, Depth+1);
61 // If all of the MaskV bits are known to be zero, then we know the
62 // output top bits are zero, because we now know that the output is
64 if ((KnownZero2 & MaskV) == MaskV) {
65 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
66 // Top bits known zero.
67 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
73 unsigned BitWidth = Mask.getBitWidth();
75 // If one of the operands has trailing zeros, then the bits that the
76 // other operand has in those bit positions will be preserved in the
77 // result. For an add, this works with either operand. For a subtract,
78 // this only works if the known zeros are in the right operand.
79 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
80 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
81 BitWidth - Mask.countLeadingZeros());
82 llvm::ComputeMaskedBits(Op0, Mask2, LHSKnownZero, LHSKnownOne, TD, Depth+1);
83 assert((LHSKnownZero & LHSKnownOne) == 0 &&
84 "Bits known to be one AND zero?");
85 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
87 llvm::ComputeMaskedBits(Op1, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
88 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
89 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
91 // Determine which operand has more trailing zeros, and use that
92 // many bits from the other operand.
93 if (LHSKnownZeroOut > RHSKnownZeroOut) {
95 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
96 KnownZero |= KnownZero2 & Mask;
97 KnownOne |= KnownOne2 & Mask;
99 // If the known zeros are in the left operand for a subtract,
100 // fall back to the minimum known zeros in both operands.
101 KnownZero |= APInt::getLowBitsSet(BitWidth,
102 std::min(LHSKnownZeroOut,
105 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
106 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
107 KnownZero |= LHSKnownZero & Mask;
108 KnownOne |= LHSKnownOne & Mask;
111 // Are we still trying to solve for the sign bit?
112 if (Mask.isNegative() && !KnownZero.isNegative() && !KnownOne.isNegative()) {
115 // Adding two positive numbers can't wrap into negative
116 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
117 KnownZero |= APInt::getSignBit(BitWidth);
118 // and adding two negative numbers can't wrap into positive.
119 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
120 KnownOne |= APInt::getSignBit(BitWidth);
122 // Subtracting a negative number from a positive one can't wrap
123 if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
124 KnownZero |= APInt::getSignBit(BitWidth);
125 // neither can subtracting a positive number from a negative one.
126 else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
127 KnownOne |= APInt::getSignBit(BitWidth);
133 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
134 /// known to be either zero or one and return them in the KnownZero/KnownOne
135 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
137 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
138 /// we cannot optimize based on the assumption that it is zero without changing
139 /// it to be an explicit zero. If we don't change it to zero, other code could
140 /// optimized based on the contradictory assumption that it is non-zero.
141 /// Because instcombine aggressively folds operations with undef args anyway,
142 /// this won't lose us code quality.
144 /// This function is defined on values with integer type, values with pointer
145 /// type (but only if TD is non-null), and vectors of integers. In the case
146 /// where V is a vector, the mask, known zero, and known one values are the
147 /// same width as the vector element, and the bit is set only if it is true
148 /// for all of the elements in the vector.
149 void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
150 APInt &KnownZero, APInt &KnownOne,
151 const TargetData *TD, unsigned Depth) {
152 assert(V && "No Value?");
153 assert(Depth <= MaxDepth && "Limit Search Depth");
154 unsigned BitWidth = Mask.getBitWidth();
155 assert((V->getType()->isIntOrIntVectorTy() ||
156 V->getType()->getScalarType()->isPointerTy()) &&
157 "Not integer or pointer type!");
159 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
160 (!V->getType()->isIntOrIntVectorTy() ||
161 V->getType()->getScalarSizeInBits() == BitWidth) &&
162 KnownZero.getBitWidth() == BitWidth &&
163 KnownOne.getBitWidth() == BitWidth &&
164 "V, Mask, KnownOne and KnownZero should have same BitWidth");
166 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
167 // We know all of the bits for a constant!
168 KnownOne = CI->getValue() & Mask;
169 KnownZero = ~KnownOne & Mask;
172 // Null and aggregate-zero are all-zeros.
173 if (isa<ConstantPointerNull>(V) ||
174 isa<ConstantAggregateZero>(V)) {
175 KnownOne.clearAllBits();
179 // Handle a constant vector by taking the intersection of the known bits of
180 // each element. There is no real need to handle ConstantVector here, because
181 // we don't handle undef in any particularly useful way.
182 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
183 // We know that CDS must be a vector of integers. Take the intersection of
185 KnownZero.setAllBits(); KnownOne.setAllBits();
186 APInt Elt(KnownZero.getBitWidth(), 0);
187 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
188 Elt = CDS->getElementAsInteger(i);
195 // The address of an aligned GlobalValue has trailing zeros.
196 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
197 unsigned Align = GV->getAlignment();
198 if (Align == 0 && TD) {
199 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
200 Type *ObjectType = GVar->getType()->getElementType();
201 if (ObjectType->isSized()) {
202 // If the object is defined in the current Module, we'll be giving
203 // it the preferred alignment. Otherwise, we have to assume that it
204 // may only have the minimum ABI alignment.
205 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
206 Align = TD->getPreferredAlignment(GVar);
208 Align = TD->getABITypeAlignment(ObjectType);
213 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
214 CountTrailingZeros_32(Align));
216 KnownZero.clearAllBits();
217 KnownOne.clearAllBits();
220 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
221 // the bits of its aliasee.
222 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
223 if (GA->mayBeOverridden()) {
224 KnownZero.clearAllBits(); KnownOne.clearAllBits();
226 ComputeMaskedBits(GA->getAliasee(), Mask, KnownZero, KnownOne,
232 if (Argument *A = dyn_cast<Argument>(V)) {
233 // Get alignment information off byval arguments if specified in the IR.
234 if (A->hasByValAttr())
235 if (unsigned Align = A->getParamAlignment())
236 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
237 CountTrailingZeros_32(Align));
241 // Start out not knowing anything.
242 KnownZero.clearAllBits(); KnownOne.clearAllBits();
244 if (Depth == MaxDepth || Mask == 0)
245 return; // Limit search depth.
247 Operator *I = dyn_cast<Operator>(V);
250 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
251 switch (I->getOpcode()) {
253 case Instruction::And: {
254 // If either the LHS or the RHS are Zero, the result is zero.
255 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
256 APInt Mask2(Mask & ~KnownZero);
257 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
259 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
260 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
262 // Output known-1 bits are only known if set in both the LHS & RHS.
263 KnownOne &= KnownOne2;
264 // Output known-0 are known to be clear if zero in either the LHS | RHS.
265 KnownZero |= KnownZero2;
268 case Instruction::Or: {
269 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
270 APInt Mask2(Mask & ~KnownOne);
271 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
273 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
274 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
276 // Output known-0 bits are only known if clear in both the LHS & RHS.
277 KnownZero &= KnownZero2;
278 // Output known-1 are known to be set if set in either the LHS | RHS.
279 KnownOne |= KnownOne2;
282 case Instruction::Xor: {
283 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
284 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
286 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
287 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
289 // Output known-0 bits are known if clear or set in both the LHS & RHS.
290 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
291 // Output known-1 are known to be set if set in only one of the LHS, RHS.
292 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
293 KnownZero = KnownZeroOut;
296 case Instruction::Mul: {
297 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
298 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
299 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
301 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
302 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
304 bool isKnownNegative = false;
305 bool isKnownNonNegative = false;
306 // If the multiplication is known not to overflow, compute the sign bit.
307 if (Mask.isNegative() &&
308 cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap()) {
309 Value *Op1 = I->getOperand(1), *Op2 = I->getOperand(0);
311 // The product of a number with itself is non-negative.
312 isKnownNonNegative = true;
314 bool isKnownNonNegative1 = KnownZero.isNegative();
315 bool isKnownNonNegative2 = KnownZero2.isNegative();
316 bool isKnownNegative1 = KnownOne.isNegative();
317 bool isKnownNegative2 = KnownOne2.isNegative();
318 // The product of two numbers with the same sign is non-negative.
319 isKnownNonNegative = (isKnownNegative1 && isKnownNegative2) ||
320 (isKnownNonNegative1 && isKnownNonNegative2);
321 // The product of a negative number and a non-negative number is either
323 if (!isKnownNonNegative)
324 isKnownNegative = (isKnownNegative1 && isKnownNonNegative2 &&
325 isKnownNonZero(Op2, TD, Depth)) ||
326 (isKnownNegative2 && isKnownNonNegative1 &&
327 isKnownNonZero(Op1, TD, Depth));
331 // If low bits are zero in either operand, output low known-0 bits.
332 // Also compute a conserative estimate for high known-0 bits.
333 // More trickiness is possible, but this is sufficient for the
334 // interesting case of alignment computation.
335 KnownOne.clearAllBits();
336 unsigned TrailZ = KnownZero.countTrailingOnes() +
337 KnownZero2.countTrailingOnes();
338 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
339 KnownZero2.countLeadingOnes(),
340 BitWidth) - BitWidth;
342 TrailZ = std::min(TrailZ, BitWidth);
343 LeadZ = std::min(LeadZ, BitWidth);
344 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
345 APInt::getHighBitsSet(BitWidth, LeadZ);
348 // Only make use of no-wrap flags if we failed to compute the sign bit
349 // directly. This matters if the multiplication always overflows, in
350 // which case we prefer to follow the result of the direct computation,
351 // though as the program is invoking undefined behaviour we can choose
352 // whatever we like here.
353 if (isKnownNonNegative && !KnownOne.isNegative())
354 KnownZero.setBit(BitWidth - 1);
355 else if (isKnownNegative && !KnownZero.isNegative())
356 KnownOne.setBit(BitWidth - 1);
360 case Instruction::UDiv: {
361 // For the purposes of computing leading zeros we can conservatively
362 // treat a udiv as a logical right shift by the power of 2 known to
363 // be less than the denominator.
364 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
365 ComputeMaskedBits(I->getOperand(0),
366 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
367 unsigned LeadZ = KnownZero2.countLeadingOnes();
369 KnownOne2.clearAllBits();
370 KnownZero2.clearAllBits();
371 ComputeMaskedBits(I->getOperand(1),
372 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
373 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
374 if (RHSUnknownLeadingOnes != BitWidth)
375 LeadZ = std::min(BitWidth,
376 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
378 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
381 case Instruction::Select:
382 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
383 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
385 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
386 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
388 // Only known if known in both the LHS and RHS.
389 KnownOne &= KnownOne2;
390 KnownZero &= KnownZero2;
392 case Instruction::FPTrunc:
393 case Instruction::FPExt:
394 case Instruction::FPToUI:
395 case Instruction::FPToSI:
396 case Instruction::SIToFP:
397 case Instruction::UIToFP:
398 return; // Can't work with floating point.
399 case Instruction::PtrToInt:
400 case Instruction::IntToPtr:
401 // We can't handle these if we don't know the pointer size.
403 // FALL THROUGH and handle them the same as zext/trunc.
404 case Instruction::ZExt:
405 case Instruction::Trunc: {
406 Type *SrcTy = I->getOperand(0)->getType();
408 unsigned SrcBitWidth;
409 // Note that we handle pointer operands here because of inttoptr/ptrtoint
410 // which fall through here.
411 if (SrcTy->isPointerTy())
412 SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
414 SrcBitWidth = SrcTy->getScalarSizeInBits();
416 APInt MaskIn = Mask.zextOrTrunc(SrcBitWidth);
417 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
418 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
419 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
421 KnownZero = KnownZero.zextOrTrunc(BitWidth);
422 KnownOne = KnownOne.zextOrTrunc(BitWidth);
423 // Any top bits are known to be zero.
424 if (BitWidth > SrcBitWidth)
425 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
428 case Instruction::BitCast: {
429 Type *SrcTy = I->getOperand(0)->getType();
430 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
431 // TODO: For now, not handling conversions like:
432 // (bitcast i64 %x to <2 x i32>)
433 !I->getType()->isVectorTy()) {
434 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
440 case Instruction::SExt: {
441 // Compute the bits in the result that are not present in the input.
442 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
444 APInt MaskIn = Mask.trunc(SrcBitWidth);
445 KnownZero = KnownZero.trunc(SrcBitWidth);
446 KnownOne = KnownOne.trunc(SrcBitWidth);
447 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
449 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
450 KnownZero = KnownZero.zext(BitWidth);
451 KnownOne = KnownOne.zext(BitWidth);
453 // If the sign bit of the input is known set or clear, then we know the
454 // top bits of the result.
455 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
456 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
457 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
458 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
461 case Instruction::Shl:
462 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
463 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
464 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
465 APInt Mask2(Mask.lshr(ShiftAmt));
466 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
468 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
469 KnownZero <<= ShiftAmt;
470 KnownOne <<= ShiftAmt;
471 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
475 case Instruction::LShr:
476 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
477 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
478 // Compute the new bits that are at the top now.
479 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
481 // Unsigned shift right.
482 APInt Mask2(Mask.shl(ShiftAmt));
483 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
485 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
486 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
487 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
488 // high bits known zero.
489 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
493 case Instruction::AShr:
494 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
495 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
496 // Compute the new bits that are at the top now.
497 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
499 // Signed shift right.
500 APInt Mask2(Mask.shl(ShiftAmt));
501 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
503 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
504 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
505 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
507 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
508 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
509 KnownZero |= HighBits;
510 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
511 KnownOne |= HighBits;
515 case Instruction::Sub: {
516 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
517 ComputeMaskedBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
518 Mask, KnownZero, KnownOne, KnownZero2, KnownOne2,
522 case Instruction::Add: {
523 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
524 ComputeMaskedBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
525 Mask, KnownZero, KnownOne, KnownZero2, KnownOne2,
529 case Instruction::SRem:
530 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
531 APInt RA = Rem->getValue().abs();
532 if (RA.isPowerOf2()) {
533 APInt LowBits = RA - 1;
534 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
535 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
538 // The low bits of the first operand are unchanged by the srem.
539 KnownZero = KnownZero2 & LowBits;
540 KnownOne = KnownOne2 & LowBits;
542 // If the first operand is non-negative or has all low bits zero, then
543 // the upper bits are all zero.
544 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
545 KnownZero |= ~LowBits;
547 // If the first operand is negative and not all low bits are zero, then
548 // the upper bits are all one.
549 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
550 KnownOne |= ~LowBits;
555 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
559 // The sign bit is the LHS's sign bit, except when the result of the
560 // remainder is zero.
561 if (Mask.isNegative() && KnownZero.isNonNegative()) {
562 APInt Mask2 = APInt::getSignBit(BitWidth);
563 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
564 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
566 // If it's known zero, our sign bit is also zero.
567 if (LHSKnownZero.isNegative())
568 KnownZero |= LHSKnownZero;
572 case Instruction::URem: {
573 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
574 APInt RA = Rem->getValue();
575 if (RA.isPowerOf2()) {
576 APInt LowBits = (RA - 1);
577 APInt Mask2 = LowBits & Mask;
578 KnownZero |= ~LowBits & Mask;
579 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
581 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
586 // Since the result is less than or equal to either operand, any leading
587 // zero bits in either operand must also exist in the result.
588 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
589 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
591 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
594 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
595 KnownZero2.countLeadingOnes());
596 KnownOne.clearAllBits();
597 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
601 case Instruction::Alloca: {
602 AllocaInst *AI = cast<AllocaInst>(V);
603 unsigned Align = AI->getAlignment();
604 if (Align == 0 && TD)
605 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
608 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
609 CountTrailingZeros_32(Align));
612 case Instruction::GetElementPtr: {
613 // Analyze all of the subscripts of this getelementptr instruction
614 // to determine if we can prove known low zero bits.
615 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
616 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
617 ComputeMaskedBits(I->getOperand(0), LocalMask,
618 LocalKnownZero, LocalKnownOne, TD, Depth+1);
619 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
621 gep_type_iterator GTI = gep_type_begin(I);
622 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
623 Value *Index = I->getOperand(i);
624 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
625 // Handle struct member offset arithmetic.
627 const StructLayout *SL = TD->getStructLayout(STy);
628 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
629 uint64_t Offset = SL->getElementOffset(Idx);
630 TrailZ = std::min(TrailZ,
631 CountTrailingZeros_64(Offset));
633 // Handle array index arithmetic.
634 Type *IndexedTy = GTI.getIndexedType();
635 if (!IndexedTy->isSized()) return;
636 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
637 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
638 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
639 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
640 ComputeMaskedBits(Index, LocalMask,
641 LocalKnownZero, LocalKnownOne, TD, Depth+1);
642 TrailZ = std::min(TrailZ,
643 unsigned(CountTrailingZeros_64(TypeSize) +
644 LocalKnownZero.countTrailingOnes()));
648 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
651 case Instruction::PHI: {
652 PHINode *P = cast<PHINode>(I);
653 // Handle the case of a simple two-predecessor recurrence PHI.
654 // There's a lot more that could theoretically be done here, but
655 // this is sufficient to catch some interesting cases.
656 if (P->getNumIncomingValues() == 2) {
657 for (unsigned i = 0; i != 2; ++i) {
658 Value *L = P->getIncomingValue(i);
659 Value *R = P->getIncomingValue(!i);
660 Operator *LU = dyn_cast<Operator>(L);
663 unsigned Opcode = LU->getOpcode();
664 // Check for operations that have the property that if
665 // both their operands have low zero bits, the result
666 // will have low zero bits.
667 if (Opcode == Instruction::Add ||
668 Opcode == Instruction::Sub ||
669 Opcode == Instruction::And ||
670 Opcode == Instruction::Or ||
671 Opcode == Instruction::Mul) {
672 Value *LL = LU->getOperand(0);
673 Value *LR = LU->getOperand(1);
674 // Find a recurrence.
681 // Ok, we have a PHI of the form L op= R. Check for low
683 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
684 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
685 Mask2 = APInt::getLowBitsSet(BitWidth,
686 KnownZero2.countTrailingOnes());
688 // We need to take the minimum number of known bits
689 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
690 ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
693 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 (P->hasConstantValue() == P)
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), KnownZero | KnownOne,
723 KnownZero2, KnownOne2, TD, MaxDepth-1);
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 = Mask & APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
747 case Intrinsic::ctpop: {
748 unsigned LowBits = Log2_32(BitWidth)+1;
749 KnownZero = Mask & APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
752 case Intrinsic::x86_sse42_crc32_64_8:
753 case Intrinsic::x86_sse42_crc32_64_64:
754 KnownZero = Mask & 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, Mask,
770 KnownZero, KnownOne, KnownZero2, KnownOne2,
773 case Intrinsic::usub_with_overflow:
774 case Intrinsic::ssub_with_overflow:
775 ComputeMaskedBitsAddSub(false, II->getArgOperand(0),
776 II->getArgOperand(1), false, Mask,
777 KnownZero, KnownOne, KnownZero2, KnownOne2,
786 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
787 /// one. Convenience wrapper around ComputeMaskedBits.
788 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
789 const TargetData *TD, unsigned Depth) {
790 unsigned BitWidth = getBitWidth(V->getType(), TD);
796 APInt ZeroBits(BitWidth, 0);
797 APInt OneBits(BitWidth, 0);
798 ComputeMaskedBits(V, APInt::getSignBit(BitWidth), ZeroBits, OneBits, TD,
800 KnownOne = OneBits[BitWidth - 1];
801 KnownZero = ZeroBits[BitWidth - 1];
804 /// isPowerOfTwo - Return true if the given value is known to have exactly one
805 /// bit set when defined. For vectors return true if every element is known to
806 /// be a power of two when defined. Supports values with integer or pointer
807 /// types and vectors of integers.
808 bool llvm::isPowerOfTwo(Value *V, const TargetData *TD, bool OrZero,
810 if (Constant *C = dyn_cast<Constant>(V)) {
811 if (C->isNullValue())
813 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
814 return CI->getValue().isPowerOf2();
815 // TODO: Handle vector constants.
818 // 1 << X is clearly a power of two if the one is not shifted off the end. If
819 // it is shifted off the end then the result is undefined.
820 if (match(V, m_Shl(m_One(), m_Value())))
823 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
824 // bottom. If it is shifted off the bottom then the result is undefined.
825 if (match(V, m_LShr(m_SignBit(), m_Value())))
828 // The remaining tests are all recursive, so bail out if we hit the limit.
829 if (Depth++ == MaxDepth)
832 Value *X = 0, *Y = 0;
833 // A shift of a power of two is a power of two or zero.
834 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
835 match(V, m_Shr(m_Value(X), m_Value()))))
836 return isPowerOfTwo(X, TD, /*OrZero*/true, Depth);
838 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
839 return isPowerOfTwo(ZI->getOperand(0), TD, OrZero, Depth);
841 if (SelectInst *SI = dyn_cast<SelectInst>(V))
842 return isPowerOfTwo(SI->getTrueValue(), TD, OrZero, Depth) &&
843 isPowerOfTwo(SI->getFalseValue(), TD, OrZero, Depth);
845 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
846 // A power of two and'd with anything is a power of two or zero.
847 if (isPowerOfTwo(X, TD, /*OrZero*/true, Depth) ||
848 isPowerOfTwo(Y, TD, /*OrZero*/true, Depth))
850 // X & (-X) is always a power of two or zero.
851 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
856 // An exact divide or right shift can only shift off zero bits, so the result
857 // is a power of two only if the first operand is a power of two and not
858 // copying a sign bit (sdiv int_min, 2).
859 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
860 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
861 return isPowerOfTwo(cast<Operator>(V)->getOperand(0), TD, OrZero, Depth);
867 /// isKnownNonZero - Return true if the given value is known to be non-zero
868 /// when defined. For vectors return true if every element is known to be
869 /// non-zero when defined. Supports values with integer or pointer type and
870 /// vectors of integers.
871 bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
872 if (Constant *C = dyn_cast<Constant>(V)) {
873 if (C->isNullValue())
875 if (isa<ConstantInt>(C))
876 // Must be non-zero due to null test above.
878 // TODO: Handle vectors
882 // The remaining tests are all recursive, so bail out if we hit the limit.
883 if (Depth++ >= MaxDepth)
886 unsigned BitWidth = getBitWidth(V->getType(), TD);
888 // X | Y != 0 if X != 0 or Y != 0.
889 Value *X = 0, *Y = 0;
890 if (match(V, m_Or(m_Value(X), m_Value(Y))))
891 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
893 // ext X != 0 if X != 0.
894 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
895 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
897 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
898 // if the lowest bit is shifted off the end.
899 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
900 // shl nuw can't remove any non-zero bits.
901 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
902 if (BO->hasNoUnsignedWrap())
903 return isKnownNonZero(X, TD, Depth);
905 APInt KnownZero(BitWidth, 0);
906 APInt KnownOne(BitWidth, 0);
907 ComputeMaskedBits(X, APInt(BitWidth, 1), KnownZero, KnownOne, TD, Depth);
911 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
912 // defined if the sign bit is shifted off the end.
913 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
914 // shr exact can only shift out zero bits.
915 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
917 return isKnownNonZero(X, TD, Depth);
919 bool XKnownNonNegative, XKnownNegative;
920 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
924 // div exact can only produce a zero if the dividend is zero.
925 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
926 return isKnownNonZero(X, TD, Depth);
929 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
930 bool XKnownNonNegative, XKnownNegative;
931 bool YKnownNonNegative, YKnownNegative;
932 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
933 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
935 // If X and Y are both non-negative (as signed values) then their sum is not
936 // zero unless both X and Y are zero.
937 if (XKnownNonNegative && YKnownNonNegative)
938 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
941 // If X and Y are both negative (as signed values) then their sum is not
942 // zero unless both X and Y equal INT_MIN.
943 if (BitWidth && XKnownNegative && YKnownNegative) {
944 APInt KnownZero(BitWidth, 0);
945 APInt KnownOne(BitWidth, 0);
946 APInt Mask = APInt::getSignedMaxValue(BitWidth);
947 // The sign bit of X is set. If some other bit is set then X is not equal
949 ComputeMaskedBits(X, Mask, KnownZero, KnownOne, TD, Depth);
950 if ((KnownOne & Mask) != 0)
952 // The sign bit of Y is set. If some other bit is set then Y is not equal
954 ComputeMaskedBits(Y, Mask, KnownZero, KnownOne, TD, Depth);
955 if ((KnownOne & Mask) != 0)
959 // The sum of a non-negative number and a power of two is not zero.
960 if (XKnownNonNegative && isPowerOfTwo(Y, TD, /*OrZero*/false, Depth))
962 if (YKnownNonNegative && isPowerOfTwo(X, TD, /*OrZero*/false, Depth))
966 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
967 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
968 // If X and Y are non-zero then so is X * Y as long as the multiplication
969 // does not overflow.
970 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
971 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
974 // (C ? X : Y) != 0 if X != 0 and Y != 0.
975 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
976 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
977 isKnownNonZero(SI->getFalseValue(), TD, Depth))
981 if (!BitWidth) return false;
982 APInt KnownZero(BitWidth, 0);
983 APInt KnownOne(BitWidth, 0);
984 ComputeMaskedBits(V, APInt::getAllOnesValue(BitWidth), KnownZero, KnownOne,
986 return KnownOne != 0;
989 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
990 /// this predicate to simplify operations downstream. Mask is known to be zero
991 /// for bits that V cannot have.
993 /// This function is defined on values with integer type, values with pointer
994 /// type (but only if TD is non-null), and vectors of integers. In the case
995 /// where V is a vector, the mask, known zero, and known one values are the
996 /// same width as the vector element, and the bit is set only if it is true
997 /// for all of the elements in the vector.
998 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
999 const TargetData *TD, unsigned Depth) {
1000 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1001 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
1002 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1003 return (KnownZero & Mask) == Mask;
1008 /// ComputeNumSignBits - Return the number of times the sign bit of the
1009 /// register is replicated into the other bits. We know that at least 1 bit
1010 /// is always equal to the sign bit (itself), but other cases can give us
1011 /// information. For example, immediately after an "ashr X, 2", we know that
1012 /// the top 3 bits are all equal to each other, so we return 3.
1014 /// 'Op' must have a scalar integer type.
1016 unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
1018 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1019 "ComputeNumSignBits requires a TargetData object to operate "
1020 "on non-integer values!");
1021 Type *Ty = V->getType();
1022 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1023 Ty->getScalarSizeInBits();
1025 unsigned FirstAnswer = 1;
1027 // Note that ConstantInt is handled by the general ComputeMaskedBits case
1031 return 1; // Limit search depth.
1033 Operator *U = dyn_cast<Operator>(V);
1034 switch (Operator::getOpcode(V)) {
1036 case Instruction::SExt:
1037 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1038 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
1040 case Instruction::AShr: {
1041 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1042 // ashr X, C -> adds C sign bits. Vectors too.
1044 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1045 Tmp += ShAmt->getZExtValue();
1046 if (Tmp > TyBits) Tmp = TyBits;
1050 case Instruction::Shl: {
1052 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1053 // shl destroys sign bits.
1054 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1055 Tmp2 = ShAmt->getZExtValue();
1056 if (Tmp2 >= TyBits || // Bad shift.
1057 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1062 case Instruction::And:
1063 case Instruction::Or:
1064 case Instruction::Xor: // NOT is handled here.
1065 // Logical binary ops preserve the number of sign bits at the worst.
1066 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1068 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1069 FirstAnswer = std::min(Tmp, Tmp2);
1070 // We computed what we know about the sign bits as our first
1071 // answer. Now proceed to the generic code that uses
1072 // ComputeMaskedBits, and pick whichever answer is better.
1076 case Instruction::Select:
1077 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1078 if (Tmp == 1) return 1; // Early out.
1079 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1080 return std::min(Tmp, Tmp2);
1082 case Instruction::Add:
1083 // Add can have at most one carry bit. Thus we know that the output
1084 // is, at worst, one more bit than the inputs.
1085 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1086 if (Tmp == 1) return 1; // Early out.
1088 // Special case decrementing a value (ADD X, -1):
1089 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1090 if (CRHS->isAllOnesValue()) {
1091 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1092 APInt Mask = APInt::getAllOnesValue(TyBits);
1093 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
1096 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1098 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1101 // If we are subtracting one from a positive number, there is no carry
1102 // out of the result.
1103 if (KnownZero.isNegative())
1107 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1108 if (Tmp2 == 1) return 1;
1109 return std::min(Tmp, Tmp2)-1;
1111 case Instruction::Sub:
1112 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1113 if (Tmp2 == 1) return 1;
1116 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1117 if (CLHS->isNullValue()) {
1118 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1119 APInt Mask = APInt::getAllOnesValue(TyBits);
1120 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
1122 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1124 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1127 // If the input is known to be positive (the sign bit is known clear),
1128 // the output of the NEG has the same number of sign bits as the input.
1129 if (KnownZero.isNegative())
1132 // Otherwise, we treat this like a SUB.
1135 // Sub can have at most one carry bit. Thus we know that the output
1136 // is, at worst, one more bit than the inputs.
1137 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1138 if (Tmp == 1) return 1; // Early out.
1139 return std::min(Tmp, Tmp2)-1;
1141 case Instruction::PHI: {
1142 PHINode *PN = cast<PHINode>(U);
1143 // Don't analyze large in-degree PHIs.
1144 if (PN->getNumIncomingValues() > 4) break;
1146 // Take the minimum of all incoming values. This can't infinitely loop
1147 // because of our depth threshold.
1148 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1149 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1150 if (Tmp == 1) return Tmp;
1152 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1157 case Instruction::Trunc:
1158 // FIXME: it's tricky to do anything useful for this, but it is an important
1159 // case for targets like X86.
1163 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1164 // use this information.
1165 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1166 APInt Mask = APInt::getAllOnesValue(TyBits);
1167 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
1169 if (KnownZero.isNegative()) { // sign bit is 0
1171 } else if (KnownOne.isNegative()) { // sign bit is 1;
1178 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1179 // the number of identical bits in the top of the input value.
1181 Mask <<= Mask.getBitWidth()-TyBits;
1182 // Return # leading zeros. We use 'min' here in case Val was zero before
1183 // shifting. We don't want to return '64' as for an i32 "0".
1184 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1187 /// ComputeMultiple - This function computes the integer multiple of Base that
1188 /// equals V. If successful, it returns true and returns the multiple in
1189 /// Multiple. If unsuccessful, it returns false. It looks
1190 /// through SExt instructions only if LookThroughSExt is true.
1191 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1192 bool LookThroughSExt, unsigned Depth) {
1193 const unsigned MaxDepth = 6;
1195 assert(V && "No Value?");
1196 assert(Depth <= MaxDepth && "Limit Search Depth");
1197 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1199 Type *T = V->getType();
1201 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1211 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1212 Constant *BaseVal = ConstantInt::get(T, Base);
1213 if (CO && CO == BaseVal) {
1215 Multiple = ConstantInt::get(T, 1);
1219 if (CI && CI->getZExtValue() % Base == 0) {
1220 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1224 if (Depth == MaxDepth) return false; // Limit search depth.
1226 Operator *I = dyn_cast<Operator>(V);
1227 if (!I) return false;
1229 switch (I->getOpcode()) {
1231 case Instruction::SExt:
1232 if (!LookThroughSExt) return false;
1233 // otherwise fall through to ZExt
1234 case Instruction::ZExt:
1235 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1236 LookThroughSExt, Depth+1);
1237 case Instruction::Shl:
1238 case Instruction::Mul: {
1239 Value *Op0 = I->getOperand(0);
1240 Value *Op1 = I->getOperand(1);
1242 if (I->getOpcode() == Instruction::Shl) {
1243 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1244 if (!Op1CI) return false;
1245 // Turn Op0 << Op1 into Op0 * 2^Op1
1246 APInt Op1Int = Op1CI->getValue();
1247 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1248 APInt API(Op1Int.getBitWidth(), 0);
1249 API.setBit(BitToSet);
1250 Op1 = ConstantInt::get(V->getContext(), API);
1254 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1255 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1256 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1257 if (Op1C->getType()->getPrimitiveSizeInBits() <
1258 MulC->getType()->getPrimitiveSizeInBits())
1259 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1260 if (Op1C->getType()->getPrimitiveSizeInBits() >
1261 MulC->getType()->getPrimitiveSizeInBits())
1262 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1264 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1265 Multiple = ConstantExpr::getMul(MulC, Op1C);
1269 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1270 if (Mul0CI->getValue() == 1) {
1271 // V == Base * Op1, so return Op1
1278 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1279 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1280 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1281 if (Op0C->getType()->getPrimitiveSizeInBits() <
1282 MulC->getType()->getPrimitiveSizeInBits())
1283 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1284 if (Op0C->getType()->getPrimitiveSizeInBits() >
1285 MulC->getType()->getPrimitiveSizeInBits())
1286 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1288 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1289 Multiple = ConstantExpr::getMul(MulC, Op0C);
1293 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1294 if (Mul1CI->getValue() == 1) {
1295 // V == Base * Op0, so return Op0
1303 // We could not determine if V is a multiple of Base.
1307 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1308 /// value is never equal to -0.0.
1310 /// NOTE: this function will need to be revisited when we support non-default
1313 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1314 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1315 return !CFP->getValueAPF().isNegZero();
1318 return 1; // Limit search depth.
1320 const Operator *I = dyn_cast<Operator>(V);
1321 if (I == 0) return false;
1323 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1324 if (I->getOpcode() == Instruction::FAdd &&
1325 isa<ConstantFP>(I->getOperand(1)) &&
1326 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1329 // sitofp and uitofp turn into +0.0 for zero.
1330 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1333 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1334 // sqrt(-0.0) = -0.0, no other negative results are possible.
1335 if (II->getIntrinsicID() == Intrinsic::sqrt)
1336 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1338 if (const CallInst *CI = dyn_cast<CallInst>(I))
1339 if (const Function *F = CI->getCalledFunction()) {
1340 if (F->isDeclaration()) {
1342 if (F->getName() == "abs") return true;
1343 // fabs[lf](x) != -0.0
1344 if (F->getName() == "fabs") return true;
1345 if (F->getName() == "fabsf") return true;
1346 if (F->getName() == "fabsl") return true;
1347 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1348 F->getName() == "sqrtl")
1349 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1356 /// isBytewiseValue - If the specified value can be set by repeating the same
1357 /// byte in memory, return the i8 value that it is represented with. This is
1358 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1359 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1360 /// byte store (e.g. i16 0x1234), return null.
1361 Value *llvm::isBytewiseValue(Value *V) {
1362 // All byte-wide stores are splatable, even of arbitrary variables.
1363 if (V->getType()->isIntegerTy(8)) return V;
1365 // Handle 'null' ConstantArrayZero etc.
1366 if (Constant *C = dyn_cast<Constant>(V))
1367 if (C->isNullValue())
1368 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1370 // Constant float and double values can be handled as integer values if the
1371 // corresponding integer value is "byteable". An important case is 0.0.
1372 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1373 if (CFP->getType()->isFloatTy())
1374 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1375 if (CFP->getType()->isDoubleTy())
1376 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1377 // Don't handle long double formats, which have strange constraints.
1380 // We can handle constant integers that are power of two in size and a
1381 // multiple of 8 bits.
1382 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1383 unsigned Width = CI->getBitWidth();
1384 if (isPowerOf2_32(Width) && Width > 8) {
1385 // We can handle this value if the recursive binary decomposition is the
1386 // same at all levels.
1387 APInt Val = CI->getValue();
1389 while (Val.getBitWidth() != 8) {
1390 unsigned NextWidth = Val.getBitWidth()/2;
1391 Val2 = Val.lshr(NextWidth);
1392 Val2 = Val2.trunc(Val.getBitWidth()/2);
1393 Val = Val.trunc(Val.getBitWidth()/2);
1395 // If the top/bottom halves aren't the same, reject it.
1399 return ConstantInt::get(V->getContext(), Val);
1403 // A ConstantDataArray/Vector is splatable if all its members are equal and
1405 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1406 Value *Elt = CA->getElementAsConstant(0);
1407 Value *Val = isBytewiseValue(Elt);
1411 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1412 if (CA->getElementAsConstant(I) != Elt)
1418 // Conceptually, we could handle things like:
1419 // %a = zext i8 %X to i16
1420 // %b = shl i16 %a, 8
1421 // %c = or i16 %a, %b
1422 // but until there is an example that actually needs this, it doesn't seem
1423 // worth worrying about.
1428 // This is the recursive version of BuildSubAggregate. It takes a few different
1429 // arguments. Idxs is the index within the nested struct From that we are
1430 // looking at now (which is of type IndexedType). IdxSkip is the number of
1431 // indices from Idxs that should be left out when inserting into the resulting
1432 // struct. To is the result struct built so far, new insertvalue instructions
1434 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1435 SmallVector<unsigned, 10> &Idxs,
1437 Instruction *InsertBefore) {
1438 llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1440 // Save the original To argument so we can modify it
1442 // General case, the type indexed by Idxs is a struct
1443 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1444 // Process each struct element recursively
1447 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1451 // Couldn't find any inserted value for this index? Cleanup
1452 while (PrevTo != OrigTo) {
1453 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1454 PrevTo = Del->getAggregateOperand();
1455 Del->eraseFromParent();
1457 // Stop processing elements
1461 // If we successfully found a value for each of our subaggregates
1465 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1466 // the struct's elements had a value that was inserted directly. In the latter
1467 // case, perhaps we can't determine each of the subelements individually, but
1468 // we might be able to find the complete struct somewhere.
1470 // Find the value that is at that particular spot
1471 Value *V = FindInsertedValue(From, Idxs);
1476 // Insert the value in the new (sub) aggregrate
1477 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1478 "tmp", InsertBefore);
1481 // This helper takes a nested struct and extracts a part of it (which is again a
1482 // struct) into a new value. For example, given the struct:
1483 // { a, { b, { c, d }, e } }
1484 // and the indices "1, 1" this returns
1487 // It does this by inserting an insertvalue for each element in the resulting
1488 // struct, as opposed to just inserting a single struct. This will only work if
1489 // each of the elements of the substruct are known (ie, inserted into From by an
1490 // insertvalue instruction somewhere).
1492 // All inserted insertvalue instructions are inserted before InsertBefore
1493 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1494 Instruction *InsertBefore) {
1495 assert(InsertBefore && "Must have someplace to insert!");
1496 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1498 Value *To = UndefValue::get(IndexedType);
1499 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1500 unsigned IdxSkip = Idxs.size();
1502 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1505 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1506 /// the scalar value indexed is already around as a register, for example if it
1507 /// were inserted directly into the aggregrate.
1509 /// If InsertBefore is not null, this function will duplicate (modified)
1510 /// insertvalues when a part of a nested struct is extracted.
1511 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1512 Instruction *InsertBefore) {
1513 // Nothing to index? Just return V then (this is useful at the end of our
1515 if (idx_range.empty())
1517 // We have indices, so V should have an indexable type.
1518 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1519 "Not looking at a struct or array?");
1520 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1521 "Invalid indices for type?");
1523 if (Constant *C = dyn_cast<Constant>(V)) {
1524 C = C->getAggregateElement(idx_range[0]);
1525 if (C == 0) return 0;
1526 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1529 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1530 // Loop the indices for the insertvalue instruction in parallel with the
1531 // requested indices
1532 const unsigned *req_idx = idx_range.begin();
1533 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1534 i != e; ++i, ++req_idx) {
1535 if (req_idx == idx_range.end()) {
1536 // We can't handle this without inserting insertvalues
1540 // The requested index identifies a part of a nested aggregate. Handle
1541 // this specially. For example,
1542 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1543 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1544 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1545 // This can be changed into
1546 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1547 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1548 // which allows the unused 0,0 element from the nested struct to be
1550 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1554 // This insert value inserts something else than what we are looking for.
1555 // See if the (aggregrate) value inserted into has the value we are
1556 // looking for, then.
1558 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1561 // If we end up here, the indices of the insertvalue match with those
1562 // requested (though possibly only partially). Now we recursively look at
1563 // the inserted value, passing any remaining indices.
1564 return FindInsertedValue(I->getInsertedValueOperand(),
1565 makeArrayRef(req_idx, idx_range.end()),
1569 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1570 // If we're extracting a value from an aggregrate that was extracted from
1571 // something else, we can extract from that something else directly instead.
1572 // However, we will need to chain I's indices with the requested indices.
1574 // Calculate the number of indices required
1575 unsigned size = I->getNumIndices() + idx_range.size();
1576 // Allocate some space to put the new indices in
1577 SmallVector<unsigned, 5> Idxs;
1579 // Add indices from the extract value instruction
1580 Idxs.append(I->idx_begin(), I->idx_end());
1582 // Add requested indices
1583 Idxs.append(idx_range.begin(), idx_range.end());
1585 assert(Idxs.size() == size
1586 && "Number of indices added not correct?");
1588 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1590 // Otherwise, we don't know (such as, extracting from a function return value
1591 // or load instruction)
1595 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1596 /// it can be expressed as a base pointer plus a constant offset. Return the
1597 /// base and offset to the caller.
1598 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1599 const TargetData &TD) {
1600 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1601 if (PtrOp == 0 || Ptr->getType()->isVectorTy())
1604 // Just look through bitcasts.
1605 if (PtrOp->getOpcode() == Instruction::BitCast)
1606 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1608 // If this is a GEP with constant indices, we can look through it.
1609 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1610 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1612 gep_type_iterator GTI = gep_type_begin(GEP);
1613 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1615 ConstantInt *OpC = cast<ConstantInt>(*I);
1616 if (OpC->isZero()) continue;
1618 // Handle a struct and array indices which add their offset to the pointer.
1619 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1620 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1622 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1623 Offset += OpC->getSExtValue()*Size;
1627 // Re-sign extend from the pointer size if needed to get overflow edge cases
1629 unsigned PtrSize = TD.getPointerSizeInBits();
1631 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1633 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1637 /// getConstantStringInfo - This function computes the length of a
1638 /// null-terminated C string pointed to by V. If successful, it returns true
1639 /// and returns the string in Str. If unsuccessful, it returns false.
1640 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
1641 uint64_t Offset, bool TrimAtNul) {
1644 // Look through bitcast instructions and geps.
1645 V = V->stripPointerCasts();
1647 // If the value is a GEP instructionor constant expression, treat it as an
1649 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1650 // Make sure the GEP has exactly three arguments.
1651 if (GEP->getNumOperands() != 3)
1654 // Make sure the index-ee is a pointer to array of i8.
1655 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1656 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1657 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1660 // Check to make sure that the first operand of the GEP is an integer and
1661 // has value 0 so that we are sure we're indexing into the initializer.
1662 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1663 if (FirstIdx == 0 || !FirstIdx->isZero())
1666 // If the second index isn't a ConstantInt, then this is a variable index
1667 // into the array. If this occurs, we can't say anything meaningful about
1669 uint64_t StartIdx = 0;
1670 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1671 StartIdx = CI->getZExtValue();
1674 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
1677 // The GEP instruction, constant or instruction, must reference a global
1678 // variable that is a constant and is initialized. The referenced constant
1679 // initializer is the array that we'll use for optimization.
1680 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
1681 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1684 // Handle the all-zeros case
1685 if (GV->getInitializer()->isNullValue()) {
1686 // This is a degenerate case. The initializer is constant zero so the
1687 // length of the string must be zero.
1692 // Must be a Constant Array
1693 const ConstantDataArray *Array =
1694 dyn_cast<ConstantDataArray>(GV->getInitializer());
1695 if (Array == 0 || !Array->isString())
1698 // Get the number of elements in the array
1699 uint64_t NumElts = Array->getType()->getArrayNumElements();
1701 // Start out with the entire array in the StringRef.
1702 Str = Array->getAsString();
1704 if (Offset > NumElts)
1707 // Skip over 'offset' bytes.
1708 Str = Str.substr(Offset);
1711 // Trim off the \0 and anything after it. If the array is not nul
1712 // terminated, we just return the whole end of string. The client may know
1713 // some other way that the string is length-bound.
1714 Str = Str.substr(0, Str.find('\0'));
1719 // These next two are very similar to the above, but also look through PHI
1721 // TODO: See if we can integrate these two together.
1723 /// GetStringLengthH - If we can compute the length of the string pointed to by
1724 /// the specified pointer, return 'len+1'. If we can't, return 0.
1725 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1726 // Look through noop bitcast instructions.
1727 V = V->stripPointerCasts();
1729 // If this is a PHI node, there are two cases: either we have already seen it
1731 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1732 if (!PHIs.insert(PN))
1733 return ~0ULL; // already in the set.
1735 // If it was new, see if all the input strings are the same length.
1736 uint64_t LenSoFar = ~0ULL;
1737 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1738 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1739 if (Len == 0) return 0; // Unknown length -> unknown.
1741 if (Len == ~0ULL) continue;
1743 if (Len != LenSoFar && LenSoFar != ~0ULL)
1744 return 0; // Disagree -> unknown.
1748 // Success, all agree.
1752 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1753 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1754 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1755 if (Len1 == 0) return 0;
1756 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1757 if (Len2 == 0) return 0;
1758 if (Len1 == ~0ULL) return Len2;
1759 if (Len2 == ~0ULL) return Len1;
1760 if (Len1 != Len2) return 0;
1764 // Otherwise, see if we can read the string.
1766 if (!getConstantStringInfo(V, StrData))
1769 return StrData.size()+1;
1772 /// GetStringLength - If we can compute the length of the string pointed to by
1773 /// the specified pointer, return 'len+1'. If we can't, return 0.
1774 uint64_t llvm::GetStringLength(Value *V) {
1775 if (!V->getType()->isPointerTy()) return 0;
1777 SmallPtrSet<PHINode*, 32> PHIs;
1778 uint64_t Len = GetStringLengthH(V, PHIs);
1779 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1780 // an empty string as a length.
1781 return Len == ~0ULL ? 1 : Len;
1785 llvm::GetUnderlyingObject(Value *V, const TargetData *TD, unsigned MaxLookup) {
1786 if (!V->getType()->isPointerTy())
1788 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1789 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1790 V = GEP->getPointerOperand();
1791 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1792 V = cast<Operator>(V)->getOperand(0);
1793 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1794 if (GA->mayBeOverridden())
1796 V = GA->getAliasee();
1798 // See if InstructionSimplify knows any relevant tricks.
1799 if (Instruction *I = dyn_cast<Instruction>(V))
1800 // TODO: Acquire a DominatorTree and use it.
1801 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1808 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1813 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1814 /// are lifetime markers.
1816 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1817 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1819 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1820 if (!II) return false;
1822 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1823 II->getIntrinsicID() != Intrinsic::lifetime_end)
1829 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1830 const TargetData *TD) {
1831 const Operator *Inst = dyn_cast<Operator>(V);
1835 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1836 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1840 switch (Inst->getOpcode()) {
1843 case Instruction::UDiv:
1844 case Instruction::URem:
1845 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1846 return isKnownNonZero(Inst->getOperand(1), TD);
1847 case Instruction::SDiv:
1848 case Instruction::SRem: {
1849 Value *Op = Inst->getOperand(1);
1850 // x / y is undefined if y == 0
1851 if (!isKnownNonZero(Op, TD))
1853 // x / y might be undefined if y == -1
1854 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1857 APInt KnownZero(BitWidth, 0);
1858 APInt KnownOne(BitWidth, 0);
1859 ComputeMaskedBits(Op, APInt::getAllOnesValue(BitWidth),
1860 KnownZero, KnownOne, TD);
1863 case Instruction::Load: {
1864 const LoadInst *LI = cast<LoadInst>(Inst);
1865 if (!LI->isUnordered())
1867 return LI->getPointerOperand()->isDereferenceablePointer();
1869 case Instruction::Call: {
1870 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1871 switch (II->getIntrinsicID()) {
1872 case Intrinsic::bswap:
1873 case Intrinsic::ctlz:
1874 case Intrinsic::ctpop:
1875 case Intrinsic::cttz:
1876 case Intrinsic::objectsize:
1877 case Intrinsic::sadd_with_overflow:
1878 case Intrinsic::smul_with_overflow:
1879 case Intrinsic::ssub_with_overflow:
1880 case Intrinsic::uadd_with_overflow:
1881 case Intrinsic::umul_with_overflow:
1882 case Intrinsic::usub_with_overflow:
1884 // TODO: some fp intrinsics are marked as having the same error handling
1885 // as libm. They're safe to speculate when they won't error.
1886 // TODO: are convert_{from,to}_fp16 safe?
1887 // TODO: can we list target-specific intrinsics here?
1891 return false; // The called function could have undefined behavior or
1892 // side-effects, even if marked readnone nounwind.
1894 case Instruction::VAArg:
1895 case Instruction::Alloca:
1896 case Instruction::Invoke:
1897 case Instruction::PHI:
1898 case Instruction::Store:
1899 case Instruction::Ret:
1900 case Instruction::Br:
1901 case Instruction::IndirectBr:
1902 case Instruction::Switch:
1903 case Instruction::Unreachable:
1904 case Instruction::Fence:
1905 case Instruction::LandingPad:
1906 case Instruction::AtomicRMW:
1907 case Instruction::AtomicCmpXchg:
1908 case Instruction::Resume:
1909 return false; // Misc instructions which have effects