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 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
45 /// known to be either zero or one and return them in the KnownZero/KnownOne
46 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
48 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
49 /// we cannot optimize based on the assumption that it is zero without changing
50 /// it to be an explicit zero. If we don't change it to zero, other code could
51 /// optimized based on the contradictory assumption that it is non-zero.
52 /// Because instcombine aggressively folds operations with undef args anyway,
53 /// this won't lose us code quality.
55 /// This function is defined on values with integer type, values with pointer
56 /// type (but only if TD is non-null), and vectors of integers. In the case
57 /// where V is a vector, the mask, known zero, and known one values are the
58 /// same width as the vector element, and the bit is set only if it is true
59 /// for all of the elements in the vector.
60 void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
61 APInt &KnownZero, APInt &KnownOne,
62 const TargetData *TD, unsigned Depth) {
63 assert(V && "No Value?");
64 assert(Depth <= MaxDepth && "Limit Search Depth");
65 unsigned BitWidth = Mask.getBitWidth();
66 assert((V->getType()->isIntOrIntVectorTy() || V->getType()->isPointerTy())
67 && "Not integer or pointer type!");
69 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
70 (!V->getType()->isIntOrIntVectorTy() ||
71 V->getType()->getScalarSizeInBits() == BitWidth) &&
72 KnownZero.getBitWidth() == BitWidth &&
73 KnownOne.getBitWidth() == BitWidth &&
74 "V, Mask, KnownOne and KnownZero should have same BitWidth");
76 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
77 // We know all of the bits for a constant!
78 KnownOne = CI->getValue() & Mask;
79 KnownZero = ~KnownOne & Mask;
82 // Null and aggregate-zero are all-zeros.
83 if (isa<ConstantPointerNull>(V) ||
84 isa<ConstantAggregateZero>(V)) {
85 KnownOne.clearAllBits();
89 // Handle a constant vector by taking the intersection of the known bits of
91 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
92 KnownZero.setAllBits(); KnownOne.setAllBits();
93 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
94 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
95 ComputeMaskedBits(CV->getOperand(i), Mask, KnownZero2, KnownOne2,
97 KnownZero &= KnownZero2;
98 KnownOne &= KnownOne2;
102 // The address of an aligned GlobalValue has trailing zeros.
103 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
104 unsigned Align = GV->getAlignment();
105 if (Align == 0 && TD && GV->getType()->getElementType()->isSized()) {
106 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
107 Type *ObjectType = GVar->getType()->getElementType();
108 // If the object is defined in the current Module, we'll be giving
109 // it the preferred alignment. Otherwise, we have to assume that it
110 // may only have the minimum ABI alignment.
111 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
112 Align = TD->getPreferredAlignment(GVar);
114 Align = TD->getABITypeAlignment(ObjectType);
118 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
119 CountTrailingZeros_32(Align));
121 KnownZero.clearAllBits();
122 KnownOne.clearAllBits();
125 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
126 // the bits of its aliasee.
127 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
128 if (GA->mayBeOverridden()) {
129 KnownZero.clearAllBits(); KnownOne.clearAllBits();
131 ComputeMaskedBits(GA->getAliasee(), Mask, KnownZero, KnownOne,
137 if (Argument *A = dyn_cast<Argument>(V)) {
138 // Get alignment information off byval arguments if specified in the IR.
139 if (A->hasByValAttr())
140 if (unsigned Align = A->getParamAlignment())
141 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
142 CountTrailingZeros_32(Align));
146 // Start out not knowing anything.
147 KnownZero.clearAllBits(); KnownOne.clearAllBits();
149 if (Depth == MaxDepth || Mask == 0)
150 return; // Limit search depth.
152 Operator *I = dyn_cast<Operator>(V);
155 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
156 switch (I->getOpcode()) {
158 case Instruction::And: {
159 // If either the LHS or the RHS are Zero, the result is zero.
160 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
161 APInt Mask2(Mask & ~KnownZero);
162 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
164 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
165 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
167 // Output known-1 bits are only known if set in both the LHS & RHS.
168 KnownOne &= KnownOne2;
169 // Output known-0 are known to be clear if zero in either the LHS | RHS.
170 KnownZero |= KnownZero2;
173 case Instruction::Or: {
174 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
175 APInt Mask2(Mask & ~KnownOne);
176 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
178 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
179 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
181 // Output known-0 bits are only known if clear in both the LHS & RHS.
182 KnownZero &= KnownZero2;
183 // Output known-1 are known to be set if set in either the LHS | RHS.
184 KnownOne |= KnownOne2;
187 case Instruction::Xor: {
188 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
189 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
191 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
192 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
194 // Output known-0 bits are known if clear or set in both the LHS & RHS.
195 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
196 // Output known-1 are known to be set if set in only one of the LHS, RHS.
197 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
198 KnownZero = KnownZeroOut;
201 case Instruction::Mul: {
202 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
203 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
204 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
206 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
207 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
209 bool isKnownNegative = false;
210 bool isKnownNonNegative = false;
211 // If the multiplication is known not to overflow, compute the sign bit.
212 if (Mask.isNegative() &&
213 cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap()) {
214 Value *Op1 = I->getOperand(1), *Op2 = I->getOperand(0);
216 // The product of a number with itself is non-negative.
217 isKnownNonNegative = true;
219 bool isKnownNonNegative1 = KnownZero.isNegative();
220 bool isKnownNonNegative2 = KnownZero2.isNegative();
221 bool isKnownNegative1 = KnownOne.isNegative();
222 bool isKnownNegative2 = KnownOne2.isNegative();
223 // The product of two numbers with the same sign is non-negative.
224 isKnownNonNegative = (isKnownNegative1 && isKnownNegative2) ||
225 (isKnownNonNegative1 && isKnownNonNegative2);
226 // The product of a negative number and a non-negative number is either
228 if (!isKnownNonNegative)
229 isKnownNegative = (isKnownNegative1 && isKnownNonNegative2 &&
230 isKnownNonZero(Op2, TD, Depth)) ||
231 (isKnownNegative2 && isKnownNonNegative1 &&
232 isKnownNonZero(Op1, TD, Depth));
236 // If low bits are zero in either operand, output low known-0 bits.
237 // Also compute a conserative estimate for high known-0 bits.
238 // More trickiness is possible, but this is sufficient for the
239 // interesting case of alignment computation.
240 KnownOne.clearAllBits();
241 unsigned TrailZ = KnownZero.countTrailingOnes() +
242 KnownZero2.countTrailingOnes();
243 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
244 KnownZero2.countLeadingOnes(),
245 BitWidth) - BitWidth;
247 TrailZ = std::min(TrailZ, BitWidth);
248 LeadZ = std::min(LeadZ, BitWidth);
249 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
250 APInt::getHighBitsSet(BitWidth, LeadZ);
253 // Only make use of no-wrap flags if we failed to compute the sign bit
254 // directly. This matters if the multiplication always overflows, in
255 // which case we prefer to follow the result of the direct computation,
256 // though as the program is invoking undefined behaviour we can choose
257 // whatever we like here.
258 if (isKnownNonNegative && !KnownOne.isNegative())
259 KnownZero.setBit(BitWidth - 1);
260 else if (isKnownNegative && !KnownZero.isNegative())
261 KnownOne.setBit(BitWidth - 1);
265 case Instruction::UDiv: {
266 // For the purposes of computing leading zeros we can conservatively
267 // treat a udiv as a logical right shift by the power of 2 known to
268 // be less than the denominator.
269 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
270 ComputeMaskedBits(I->getOperand(0),
271 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
272 unsigned LeadZ = KnownZero2.countLeadingOnes();
274 KnownOne2.clearAllBits();
275 KnownZero2.clearAllBits();
276 ComputeMaskedBits(I->getOperand(1),
277 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
278 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
279 if (RHSUnknownLeadingOnes != BitWidth)
280 LeadZ = std::min(BitWidth,
281 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
283 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
286 case Instruction::Select:
287 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
288 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
290 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
291 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
293 // Only known if known in both the LHS and RHS.
294 KnownOne &= KnownOne2;
295 KnownZero &= KnownZero2;
297 case Instruction::FPTrunc:
298 case Instruction::FPExt:
299 case Instruction::FPToUI:
300 case Instruction::FPToSI:
301 case Instruction::SIToFP:
302 case Instruction::UIToFP:
303 return; // Can't work with floating point.
304 case Instruction::PtrToInt:
305 case Instruction::IntToPtr:
306 // We can't handle these if we don't know the pointer size.
308 // FALL THROUGH and handle them the same as zext/trunc.
309 case Instruction::ZExt:
310 case Instruction::Trunc: {
311 Type *SrcTy = I->getOperand(0)->getType();
313 unsigned SrcBitWidth;
314 // Note that we handle pointer operands here because of inttoptr/ptrtoint
315 // which fall through here.
316 if (SrcTy->isPointerTy())
317 SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
319 SrcBitWidth = SrcTy->getScalarSizeInBits();
321 APInt MaskIn = Mask.zextOrTrunc(SrcBitWidth);
322 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
323 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
324 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
326 KnownZero = KnownZero.zextOrTrunc(BitWidth);
327 KnownOne = KnownOne.zextOrTrunc(BitWidth);
328 // Any top bits are known to be zero.
329 if (BitWidth > SrcBitWidth)
330 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
333 case Instruction::BitCast: {
334 Type *SrcTy = I->getOperand(0)->getType();
335 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
336 // TODO: For now, not handling conversions like:
337 // (bitcast i64 %x to <2 x i32>)
338 !I->getType()->isVectorTy()) {
339 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
345 case Instruction::SExt: {
346 // Compute the bits in the result that are not present in the input.
347 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
349 APInt MaskIn = Mask.trunc(SrcBitWidth);
350 KnownZero = KnownZero.trunc(SrcBitWidth);
351 KnownOne = KnownOne.trunc(SrcBitWidth);
352 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
354 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
355 KnownZero = KnownZero.zext(BitWidth);
356 KnownOne = KnownOne.zext(BitWidth);
358 // If the sign bit of the input is known set or clear, then we know the
359 // top bits of the result.
360 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
361 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
362 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
363 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
366 case Instruction::Shl:
367 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
368 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
369 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
370 APInt Mask2(Mask.lshr(ShiftAmt));
371 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
373 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
374 KnownZero <<= ShiftAmt;
375 KnownOne <<= ShiftAmt;
376 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
380 case Instruction::LShr:
381 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
382 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
383 // Compute the new bits that are at the top now.
384 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
386 // Unsigned shift right.
387 APInt Mask2(Mask.shl(ShiftAmt));
388 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
390 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
391 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
392 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
393 // high bits known zero.
394 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
398 case Instruction::AShr:
399 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
400 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
401 // Compute the new bits that are at the top now.
402 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
404 // Signed shift right.
405 APInt Mask2(Mask.shl(ShiftAmt));
406 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
408 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
409 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
410 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
412 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
413 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
414 KnownZero |= HighBits;
415 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
416 KnownOne |= HighBits;
420 case Instruction::Sub: {
421 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
422 // We know that the top bits of C-X are clear if X contains less bits
423 // than C (i.e. no wrap-around can happen). For example, 20-X is
424 // positive if we can prove that X is >= 0 and < 16.
425 if (!CLHS->getValue().isNegative()) {
426 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
427 // NLZ can't be BitWidth with no sign bit
428 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
429 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
432 // If all of the MaskV bits are known to be zero, then we know the
433 // output top bits are zero, because we now know that the output is
435 if ((KnownZero2 & MaskV) == MaskV) {
436 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
437 // Top bits known zero.
438 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
444 case Instruction::Add: {
445 // If one of the operands has trailing zeros, then the bits that the
446 // other operand has in those bit positions will be preserved in the
447 // result. For an add, this works with either operand. For a subtract,
448 // this only works if the known zeros are in the right operand.
449 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
450 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
451 BitWidth - Mask.countLeadingZeros());
452 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
454 assert((LHSKnownZero & LHSKnownOne) == 0 &&
455 "Bits known to be one AND zero?");
456 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
458 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD,
460 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
461 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
463 // Determine which operand has more trailing zeros, and use that
464 // many bits from the other operand.
465 if (LHSKnownZeroOut > RHSKnownZeroOut) {
466 if (I->getOpcode() == Instruction::Add) {
467 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
468 KnownZero |= KnownZero2 & Mask;
469 KnownOne |= KnownOne2 & Mask;
471 // If the known zeros are in the left operand for a subtract,
472 // fall back to the minimum known zeros in both operands.
473 KnownZero |= APInt::getLowBitsSet(BitWidth,
474 std::min(LHSKnownZeroOut,
477 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
478 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
479 KnownZero |= LHSKnownZero & Mask;
480 KnownOne |= LHSKnownOne & Mask;
483 // Are we still trying to solve for the sign bit?
484 if (Mask.isNegative() && !KnownZero.isNegative() && !KnownOne.isNegative()){
485 OverflowingBinaryOperator *OBO = cast<OverflowingBinaryOperator>(I);
486 if (OBO->hasNoSignedWrap()) {
487 if (I->getOpcode() == Instruction::Add) {
488 // Adding two positive numbers can't wrap into negative
489 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
490 KnownZero |= APInt::getSignBit(BitWidth);
491 // and adding two negative numbers can't wrap into positive.
492 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
493 KnownOne |= APInt::getSignBit(BitWidth);
495 // Subtracting a negative number from a positive one can't wrap
496 if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
497 KnownZero |= APInt::getSignBit(BitWidth);
498 // neither can subtracting a positive number from a negative one.
499 else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
500 KnownOne |= APInt::getSignBit(BitWidth);
507 case Instruction::SRem:
508 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
509 APInt RA = Rem->getValue().abs();
510 if (RA.isPowerOf2()) {
511 APInt LowBits = RA - 1;
512 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
513 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
516 // The low bits of the first operand are unchanged by the srem.
517 KnownZero = KnownZero2 & LowBits;
518 KnownOne = KnownOne2 & LowBits;
520 // If the first operand is non-negative or has all low bits zero, then
521 // the upper bits are all zero.
522 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
523 KnownZero |= ~LowBits;
525 // If the first operand is negative and not all low bits are zero, then
526 // the upper bits are all one.
527 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
528 KnownOne |= ~LowBits;
533 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
537 // The sign bit is the LHS's sign bit, except when the result of the
538 // remainder is zero.
539 if (Mask.isNegative() && KnownZero.isNonNegative()) {
540 APInt Mask2 = APInt::getSignBit(BitWidth);
541 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
542 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
544 // If it's known zero, our sign bit is also zero.
545 if (LHSKnownZero.isNegative())
546 KnownZero |= LHSKnownZero;
550 case Instruction::URem: {
551 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
552 APInt RA = Rem->getValue();
553 if (RA.isPowerOf2()) {
554 APInt LowBits = (RA - 1);
555 APInt Mask2 = LowBits & Mask;
556 KnownZero |= ~LowBits & Mask;
557 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
559 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
564 // Since the result is less than or equal to either operand, any leading
565 // zero bits in either operand must also exist in the result.
566 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
567 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
569 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
572 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
573 KnownZero2.countLeadingOnes());
574 KnownOne.clearAllBits();
575 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
579 case Instruction::Alloca: {
580 AllocaInst *AI = cast<AllocaInst>(V);
581 unsigned Align = AI->getAlignment();
582 if (Align == 0 && TD)
583 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
586 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
587 CountTrailingZeros_32(Align));
590 case Instruction::GetElementPtr: {
591 // Analyze all of the subscripts of this getelementptr instruction
592 // to determine if we can prove known low zero bits.
593 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
594 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
595 ComputeMaskedBits(I->getOperand(0), LocalMask,
596 LocalKnownZero, LocalKnownOne, TD, Depth+1);
597 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
599 gep_type_iterator GTI = gep_type_begin(I);
600 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
601 Value *Index = I->getOperand(i);
602 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
603 // Handle struct member offset arithmetic.
605 const StructLayout *SL = TD->getStructLayout(STy);
606 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
607 uint64_t Offset = SL->getElementOffset(Idx);
608 TrailZ = std::min(TrailZ,
609 CountTrailingZeros_64(Offset));
611 // Handle array index arithmetic.
612 Type *IndexedTy = GTI.getIndexedType();
613 if (!IndexedTy->isSized()) return;
614 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
615 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
616 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
617 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
618 ComputeMaskedBits(Index, LocalMask,
619 LocalKnownZero, LocalKnownOne, TD, Depth+1);
620 TrailZ = std::min(TrailZ,
621 unsigned(CountTrailingZeros_64(TypeSize) +
622 LocalKnownZero.countTrailingOnes()));
626 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
629 case Instruction::PHI: {
630 PHINode *P = cast<PHINode>(I);
631 // Handle the case of a simple two-predecessor recurrence PHI.
632 // There's a lot more that could theoretically be done here, but
633 // this is sufficient to catch some interesting cases.
634 if (P->getNumIncomingValues() == 2) {
635 for (unsigned i = 0; i != 2; ++i) {
636 Value *L = P->getIncomingValue(i);
637 Value *R = P->getIncomingValue(!i);
638 Operator *LU = dyn_cast<Operator>(L);
641 unsigned Opcode = LU->getOpcode();
642 // Check for operations that have the property that if
643 // both their operands have low zero bits, the result
644 // will have low zero bits.
645 if (Opcode == Instruction::Add ||
646 Opcode == Instruction::Sub ||
647 Opcode == Instruction::And ||
648 Opcode == Instruction::Or ||
649 Opcode == Instruction::Mul) {
650 Value *LL = LU->getOperand(0);
651 Value *LR = LU->getOperand(1);
652 // Find a recurrence.
659 // Ok, we have a PHI of the form L op= R. Check for low
661 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
662 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
663 Mask2 = APInt::getLowBitsSet(BitWidth,
664 KnownZero2.countTrailingOnes());
666 // We need to take the minimum number of known bits
667 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
668 ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
671 APInt::getLowBitsSet(BitWidth,
672 std::min(KnownZero2.countTrailingOnes(),
673 KnownZero3.countTrailingOnes()));
679 // Unreachable blocks may have zero-operand PHI nodes.
680 if (P->getNumIncomingValues() == 0)
683 // Otherwise take the unions of the known bit sets of the operands,
684 // taking conservative care to avoid excessive recursion.
685 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
686 // Skip if every incoming value references to ourself.
687 if (P->hasConstantValue() == P)
690 KnownZero = APInt::getAllOnesValue(BitWidth);
691 KnownOne = APInt::getAllOnesValue(BitWidth);
692 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
693 // Skip direct self references.
694 if (P->getIncomingValue(i) == P) continue;
696 KnownZero2 = APInt(BitWidth, 0);
697 KnownOne2 = APInt(BitWidth, 0);
698 // Recurse, but cap the recursion to one level, because we don't
699 // want to waste time spinning around in loops.
700 ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
701 KnownZero2, KnownOne2, TD, MaxDepth-1);
702 KnownZero &= KnownZero2;
703 KnownOne &= KnownOne2;
704 // If all bits have been ruled out, there's no need to check
706 if (!KnownZero && !KnownOne)
712 case Instruction::Call:
713 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
714 switch (II->getIntrinsicID()) {
716 case Intrinsic::ctpop:
717 case Intrinsic::ctlz:
718 case Intrinsic::cttz: {
719 unsigned LowBits = Log2_32(BitWidth)+1;
720 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
723 case Intrinsic::x86_sse42_crc32_64_8:
724 case Intrinsic::x86_sse42_crc32_64_64:
725 KnownZero = APInt::getHighBitsSet(64, 32);
733 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
734 /// one. Convenience wrapper around ComputeMaskedBits.
735 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
736 const TargetData *TD, unsigned Depth) {
737 unsigned BitWidth = getBitWidth(V->getType(), TD);
743 APInt ZeroBits(BitWidth, 0);
744 APInt OneBits(BitWidth, 0);
745 ComputeMaskedBits(V, APInt::getSignBit(BitWidth), ZeroBits, OneBits, TD,
747 KnownOne = OneBits[BitWidth - 1];
748 KnownZero = ZeroBits[BitWidth - 1];
751 /// isPowerOfTwo - Return true if the given value is known to have exactly one
752 /// bit set when defined. For vectors return true if every element is known to
753 /// be a power of two when defined. Supports values with integer or pointer
754 /// types and vectors of integers.
755 bool llvm::isPowerOfTwo(Value *V, const TargetData *TD, bool OrZero,
757 if (Constant *C = dyn_cast<Constant>(V)) {
758 if (C->isNullValue())
760 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
761 return CI->getValue().isPowerOf2();
762 // TODO: Handle vector constants.
765 // 1 << X is clearly a power of two if the one is not shifted off the end. If
766 // it is shifted off the end then the result is undefined.
767 if (match(V, m_Shl(m_One(), m_Value())))
770 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
771 // bottom. If it is shifted off the bottom then the result is undefined.
772 if (match(V, m_LShr(m_SignBit(), m_Value())))
775 // The remaining tests are all recursive, so bail out if we hit the limit.
776 if (Depth++ == MaxDepth)
779 Value *X = 0, *Y = 0;
780 // A shift of a power of two is a power of two or zero.
781 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
782 match(V, m_Shr(m_Value(X), m_Value()))))
783 return isPowerOfTwo(X, TD, /*OrZero*/true, Depth);
785 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
786 return isPowerOfTwo(ZI->getOperand(0), TD, OrZero, Depth);
788 if (SelectInst *SI = dyn_cast<SelectInst>(V))
789 return isPowerOfTwo(SI->getTrueValue(), TD, OrZero, Depth) &&
790 isPowerOfTwo(SI->getFalseValue(), TD, OrZero, Depth);
792 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
793 // A power of two and'd with anything is a power of two or zero.
794 if (isPowerOfTwo(X, TD, /*OrZero*/true, Depth) ||
795 isPowerOfTwo(Y, TD, /*OrZero*/true, Depth))
797 // X & (-X) is always a power of two or zero.
798 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
803 // An exact divide or right shift can only shift off zero bits, so the result
804 // is a power of two only if the first operand is a power of two and not
805 // copying a sign bit (sdiv int_min, 2).
806 if (match(V, m_LShr(m_Value(), m_Value())) ||
807 match(V, m_UDiv(m_Value(), m_Value()))) {
808 PossiblyExactOperator *PEO = cast<PossiblyExactOperator>(V);
810 return isPowerOfTwo(PEO->getOperand(0), TD, OrZero, Depth);
816 /// isKnownNonZero - Return true if the given value is known to be non-zero
817 /// when defined. For vectors return true if every element is known to be
818 /// non-zero when defined. Supports values with integer or pointer type and
819 /// vectors of integers.
820 bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
821 if (Constant *C = dyn_cast<Constant>(V)) {
822 if (C->isNullValue())
824 if (isa<ConstantInt>(C))
825 // Must be non-zero due to null test above.
827 // TODO: Handle vectors
831 // The remaining tests are all recursive, so bail out if we hit the limit.
832 if (Depth++ >= MaxDepth)
835 unsigned BitWidth = getBitWidth(V->getType(), TD);
837 // X | Y != 0 if X != 0 or Y != 0.
838 Value *X = 0, *Y = 0;
839 if (match(V, m_Or(m_Value(X), m_Value(Y))))
840 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
842 // ext X != 0 if X != 0.
843 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
844 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
846 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
847 // if the lowest bit is shifted off the end.
848 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
849 // shl nuw can't remove any non-zero bits.
850 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
851 if (BO->hasNoUnsignedWrap())
852 return isKnownNonZero(X, TD, Depth);
854 APInt KnownZero(BitWidth, 0);
855 APInt KnownOne(BitWidth, 0);
856 ComputeMaskedBits(X, APInt(BitWidth, 1), KnownZero, KnownOne, TD, Depth);
860 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
861 // defined if the sign bit is shifted off the end.
862 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
863 // shr exact can only shift out zero bits.
864 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
866 return isKnownNonZero(X, TD, Depth);
868 bool XKnownNonNegative, XKnownNegative;
869 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
873 // div exact can only produce a zero if the dividend is zero.
874 else if (match(V, m_IDiv(m_Value(X), m_Value()))) {
875 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
877 return isKnownNonZero(X, TD, Depth);
880 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
881 bool XKnownNonNegative, XKnownNegative;
882 bool YKnownNonNegative, YKnownNegative;
883 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
884 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
886 // If X and Y are both non-negative (as signed values) then their sum is not
887 // zero unless both X and Y are zero.
888 if (XKnownNonNegative && YKnownNonNegative)
889 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
892 // If X and Y are both negative (as signed values) then their sum is not
893 // zero unless both X and Y equal INT_MIN.
894 if (BitWidth && XKnownNegative && YKnownNegative) {
895 APInt KnownZero(BitWidth, 0);
896 APInt KnownOne(BitWidth, 0);
897 APInt Mask = APInt::getSignedMaxValue(BitWidth);
898 // The sign bit of X is set. If some other bit is set then X is not equal
900 ComputeMaskedBits(X, Mask, KnownZero, KnownOne, TD, Depth);
901 if ((KnownOne & Mask) != 0)
903 // The sign bit of Y is set. If some other bit is set then Y is not equal
905 ComputeMaskedBits(Y, Mask, KnownZero, KnownOne, TD, Depth);
906 if ((KnownOne & Mask) != 0)
910 // The sum of a non-negative number and a power of two is not zero.
911 if (XKnownNonNegative && isPowerOfTwo(Y, TD, /*OrZero*/false, Depth))
913 if (YKnownNonNegative && isPowerOfTwo(X, TD, /*OrZero*/false, Depth))
917 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
918 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
919 // If X and Y are non-zero then so is X * Y as long as the multiplication
920 // does not overflow.
921 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
922 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
925 // (C ? X : Y) != 0 if X != 0 and Y != 0.
926 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
927 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
928 isKnownNonZero(SI->getFalseValue(), TD, Depth))
932 if (!BitWidth) return false;
933 APInt KnownZero(BitWidth, 0);
934 APInt KnownOne(BitWidth, 0);
935 ComputeMaskedBits(V, APInt::getAllOnesValue(BitWidth), KnownZero, KnownOne,
937 return KnownOne != 0;
940 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
941 /// this predicate to simplify operations downstream. Mask is known to be zero
942 /// for bits that V cannot have.
944 /// This function is defined on values with integer type, values with pointer
945 /// type (but only if TD is non-null), and vectors of integers. In the case
946 /// where V is a vector, the mask, known zero, and known one values are the
947 /// same width as the vector element, and the bit is set only if it is true
948 /// for all of the elements in the vector.
949 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
950 const TargetData *TD, unsigned Depth) {
951 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
952 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
953 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
954 return (KnownZero & Mask) == Mask;
959 /// ComputeNumSignBits - Return the number of times the sign bit of the
960 /// register is replicated into the other bits. We know that at least 1 bit
961 /// is always equal to the sign bit (itself), but other cases can give us
962 /// information. For example, immediately after an "ashr X, 2", we know that
963 /// the top 3 bits are all equal to each other, so we return 3.
965 /// 'Op' must have a scalar integer type.
967 unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
969 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
970 "ComputeNumSignBits requires a TargetData object to operate "
971 "on non-integer values!");
972 Type *Ty = V->getType();
973 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
974 Ty->getScalarSizeInBits();
976 unsigned FirstAnswer = 1;
978 // Note that ConstantInt is handled by the general ComputeMaskedBits case
982 return 1; // Limit search depth.
984 Operator *U = dyn_cast<Operator>(V);
985 switch (Operator::getOpcode(V)) {
987 case Instruction::SExt:
988 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
989 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
991 case Instruction::AShr:
992 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
993 // ashr X, C -> adds C sign bits.
994 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
995 Tmp += C->getZExtValue();
996 if (Tmp > TyBits) Tmp = TyBits;
998 // vector ashr X, <C, C, C, C> -> adds C sign bits
999 if (ConstantVector *C = dyn_cast<ConstantVector>(U->getOperand(1))) {
1000 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) {
1001 Tmp += CI->getZExtValue();
1002 if (Tmp > TyBits) Tmp = TyBits;
1006 case Instruction::Shl:
1007 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
1008 // shl destroys sign bits.
1009 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1010 if (C->getZExtValue() >= TyBits || // Bad shift.
1011 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
1012 return Tmp - C->getZExtValue();
1015 case Instruction::And:
1016 case Instruction::Or:
1017 case Instruction::Xor: // NOT is handled here.
1018 // Logical binary ops preserve the number of sign bits at the worst.
1019 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1021 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1022 FirstAnswer = std::min(Tmp, Tmp2);
1023 // We computed what we know about the sign bits as our first
1024 // answer. Now proceed to the generic code that uses
1025 // ComputeMaskedBits, and pick whichever answer is better.
1029 case Instruction::Select:
1030 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1031 if (Tmp == 1) return 1; // Early out.
1032 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1033 return std::min(Tmp, Tmp2);
1035 case Instruction::Add:
1036 // Add can have at most one carry bit. Thus we know that the output
1037 // is, at worst, one more bit than the inputs.
1038 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1039 if (Tmp == 1) return 1; // Early out.
1041 // Special case decrementing a value (ADD X, -1):
1042 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1043 if (CRHS->isAllOnesValue()) {
1044 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1045 APInt Mask = APInt::getAllOnesValue(TyBits);
1046 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
1049 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1051 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1054 // If we are subtracting one from a positive number, there is no carry
1055 // out of the result.
1056 if (KnownZero.isNegative())
1060 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1061 if (Tmp2 == 1) return 1;
1062 return std::min(Tmp, Tmp2)-1;
1064 case Instruction::Sub:
1065 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1066 if (Tmp2 == 1) return 1;
1069 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1070 if (CLHS->isNullValue()) {
1071 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1072 APInt Mask = APInt::getAllOnesValue(TyBits);
1073 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
1075 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1077 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1080 // If the input is known to be positive (the sign bit is known clear),
1081 // the output of the NEG has the same number of sign bits as the input.
1082 if (KnownZero.isNegative())
1085 // Otherwise, we treat this like a SUB.
1088 // Sub can have at most one carry bit. Thus we know that the output
1089 // is, at worst, one more bit than the inputs.
1090 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1091 if (Tmp == 1) return 1; // Early out.
1092 return std::min(Tmp, Tmp2)-1;
1094 case Instruction::PHI: {
1095 PHINode *PN = cast<PHINode>(U);
1096 // Don't analyze large in-degree PHIs.
1097 if (PN->getNumIncomingValues() > 4) break;
1099 // Take the minimum of all incoming values. This can't infinitely loop
1100 // because of our depth threshold.
1101 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1102 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1103 if (Tmp == 1) return Tmp;
1105 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1110 case Instruction::Trunc:
1111 // FIXME: it's tricky to do anything useful for this, but it is an important
1112 // case for targets like X86.
1116 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1117 // use this information.
1118 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1119 APInt Mask = APInt::getAllOnesValue(TyBits);
1120 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
1122 if (KnownZero.isNegative()) { // sign bit is 0
1124 } else if (KnownOne.isNegative()) { // sign bit is 1;
1131 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1132 // the number of identical bits in the top of the input value.
1134 Mask <<= Mask.getBitWidth()-TyBits;
1135 // Return # leading zeros. We use 'min' here in case Val was zero before
1136 // shifting. We don't want to return '64' as for an i32 "0".
1137 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1140 /// ComputeMultiple - This function computes the integer multiple of Base that
1141 /// equals V. If successful, it returns true and returns the multiple in
1142 /// Multiple. If unsuccessful, it returns false. It looks
1143 /// through SExt instructions only if LookThroughSExt is true.
1144 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1145 bool LookThroughSExt, unsigned Depth) {
1146 const unsigned MaxDepth = 6;
1148 assert(V && "No Value?");
1149 assert(Depth <= MaxDepth && "Limit Search Depth");
1150 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1152 Type *T = V->getType();
1154 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1164 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1165 Constant *BaseVal = ConstantInt::get(T, Base);
1166 if (CO && CO == BaseVal) {
1168 Multiple = ConstantInt::get(T, 1);
1172 if (CI && CI->getZExtValue() % Base == 0) {
1173 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1177 if (Depth == MaxDepth) return false; // Limit search depth.
1179 Operator *I = dyn_cast<Operator>(V);
1180 if (!I) return false;
1182 switch (I->getOpcode()) {
1184 case Instruction::SExt:
1185 if (!LookThroughSExt) return false;
1186 // otherwise fall through to ZExt
1187 case Instruction::ZExt:
1188 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1189 LookThroughSExt, Depth+1);
1190 case Instruction::Shl:
1191 case Instruction::Mul: {
1192 Value *Op0 = I->getOperand(0);
1193 Value *Op1 = I->getOperand(1);
1195 if (I->getOpcode() == Instruction::Shl) {
1196 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1197 if (!Op1CI) return false;
1198 // Turn Op0 << Op1 into Op0 * 2^Op1
1199 APInt Op1Int = Op1CI->getValue();
1200 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1201 APInt API(Op1Int.getBitWidth(), 0);
1202 API.setBit(BitToSet);
1203 Op1 = ConstantInt::get(V->getContext(), API);
1207 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1208 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1209 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1210 if (Op1C->getType()->getPrimitiveSizeInBits() <
1211 MulC->getType()->getPrimitiveSizeInBits())
1212 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1213 if (Op1C->getType()->getPrimitiveSizeInBits() >
1214 MulC->getType()->getPrimitiveSizeInBits())
1215 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1217 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1218 Multiple = ConstantExpr::getMul(MulC, Op1C);
1222 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1223 if (Mul0CI->getValue() == 1) {
1224 // V == Base * Op1, so return Op1
1231 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1232 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1233 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1234 if (Op0C->getType()->getPrimitiveSizeInBits() <
1235 MulC->getType()->getPrimitiveSizeInBits())
1236 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1237 if (Op0C->getType()->getPrimitiveSizeInBits() >
1238 MulC->getType()->getPrimitiveSizeInBits())
1239 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1241 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1242 Multiple = ConstantExpr::getMul(MulC, Op0C);
1246 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1247 if (Mul1CI->getValue() == 1) {
1248 // V == Base * Op0, so return Op0
1256 // We could not determine if V is a multiple of Base.
1260 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1261 /// value is never equal to -0.0.
1263 /// NOTE: this function will need to be revisited when we support non-default
1266 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1267 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1268 return !CFP->getValueAPF().isNegZero();
1271 return 1; // Limit search depth.
1273 const Operator *I = dyn_cast<Operator>(V);
1274 if (I == 0) return false;
1276 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1277 if (I->getOpcode() == Instruction::FAdd &&
1278 isa<ConstantFP>(I->getOperand(1)) &&
1279 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1282 // sitofp and uitofp turn into +0.0 for zero.
1283 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1286 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1287 // sqrt(-0.0) = -0.0, no other negative results are possible.
1288 if (II->getIntrinsicID() == Intrinsic::sqrt)
1289 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1291 if (const CallInst *CI = dyn_cast<CallInst>(I))
1292 if (const Function *F = CI->getCalledFunction()) {
1293 if (F->isDeclaration()) {
1295 if (F->getName() == "abs") return true;
1296 // fabs[lf](x) != -0.0
1297 if (F->getName() == "fabs") return true;
1298 if (F->getName() == "fabsf") return true;
1299 if (F->getName() == "fabsl") return true;
1300 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1301 F->getName() == "sqrtl")
1302 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1309 /// isBytewiseValue - If the specified value can be set by repeating the same
1310 /// byte in memory, return the i8 value that it is represented with. This is
1311 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1312 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1313 /// byte store (e.g. i16 0x1234), return null.
1314 Value *llvm::isBytewiseValue(Value *V) {
1315 // All byte-wide stores are splatable, even of arbitrary variables.
1316 if (V->getType()->isIntegerTy(8)) return V;
1318 // Handle 'null' ConstantArrayZero etc.
1319 if (Constant *C = dyn_cast<Constant>(V))
1320 if (C->isNullValue())
1321 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1323 // Constant float and double values can be handled as integer values if the
1324 // corresponding integer value is "byteable". An important case is 0.0.
1325 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1326 if (CFP->getType()->isFloatTy())
1327 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1328 if (CFP->getType()->isDoubleTy())
1329 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1330 // Don't handle long double formats, which have strange constraints.
1333 // We can handle constant integers that are power of two in size and a
1334 // multiple of 8 bits.
1335 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1336 unsigned Width = CI->getBitWidth();
1337 if (isPowerOf2_32(Width) && Width > 8) {
1338 // We can handle this value if the recursive binary decomposition is the
1339 // same at all levels.
1340 APInt Val = CI->getValue();
1342 while (Val.getBitWidth() != 8) {
1343 unsigned NextWidth = Val.getBitWidth()/2;
1344 Val2 = Val.lshr(NextWidth);
1345 Val2 = Val2.trunc(Val.getBitWidth()/2);
1346 Val = Val.trunc(Val.getBitWidth()/2);
1348 // If the top/bottom halves aren't the same, reject it.
1352 return ConstantInt::get(V->getContext(), Val);
1356 // A ConstantArray is splatable if all its members are equal and also
1358 if (ConstantArray *CA = dyn_cast<ConstantArray>(V)) {
1359 if (CA->getNumOperands() == 0)
1362 Value *Val = isBytewiseValue(CA->getOperand(0));
1366 for (unsigned I = 1, E = CA->getNumOperands(); I != E; ++I)
1367 if (CA->getOperand(I-1) != CA->getOperand(I))
1373 // Conceptually, we could handle things like:
1374 // %a = zext i8 %X to i16
1375 // %b = shl i16 %a, 8
1376 // %c = or i16 %a, %b
1377 // but until there is an example that actually needs this, it doesn't seem
1378 // worth worrying about.
1383 // This is the recursive version of BuildSubAggregate. It takes a few different
1384 // arguments. Idxs is the index within the nested struct From that we are
1385 // looking at now (which is of type IndexedType). IdxSkip is the number of
1386 // indices from Idxs that should be left out when inserting into the resulting
1387 // struct. To is the result struct built so far, new insertvalue instructions
1389 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1390 SmallVector<unsigned, 10> &Idxs,
1392 Instruction *InsertBefore) {
1393 llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1395 // Save the original To argument so we can modify it
1397 // General case, the type indexed by Idxs is a struct
1398 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1399 // Process each struct element recursively
1402 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1406 // Couldn't find any inserted value for this index? Cleanup
1407 while (PrevTo != OrigTo) {
1408 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1409 PrevTo = Del->getAggregateOperand();
1410 Del->eraseFromParent();
1412 // Stop processing elements
1416 // If we successfully found a value for each of our subaggregates
1420 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1421 // the struct's elements had a value that was inserted directly. In the latter
1422 // case, perhaps we can't determine each of the subelements individually, but
1423 // we might be able to find the complete struct somewhere.
1425 // Find the value that is at that particular spot
1426 Value *V = FindInsertedValue(From, Idxs);
1431 // Insert the value in the new (sub) aggregrate
1432 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1433 "tmp", InsertBefore);
1436 // This helper takes a nested struct and extracts a part of it (which is again a
1437 // struct) into a new value. For example, given the struct:
1438 // { a, { b, { c, d }, e } }
1439 // and the indices "1, 1" this returns
1442 // It does this by inserting an insertvalue for each element in the resulting
1443 // struct, as opposed to just inserting a single struct. This will only work if
1444 // each of the elements of the substruct are known (ie, inserted into From by an
1445 // insertvalue instruction somewhere).
1447 // All inserted insertvalue instructions are inserted before InsertBefore
1448 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1449 Instruction *InsertBefore) {
1450 assert(InsertBefore && "Must have someplace to insert!");
1451 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1453 Value *To = UndefValue::get(IndexedType);
1454 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1455 unsigned IdxSkip = Idxs.size();
1457 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1460 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1461 /// the scalar value indexed is already around as a register, for example if it
1462 /// were inserted directly into the aggregrate.
1464 /// If InsertBefore is not null, this function will duplicate (modified)
1465 /// insertvalues when a part of a nested struct is extracted.
1466 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1467 Instruction *InsertBefore) {
1468 // Nothing to index? Just return V then (this is useful at the end of our
1470 if (idx_range.empty())
1472 // We have indices, so V should have an indexable type
1473 assert((V->getType()->isStructTy() || V->getType()->isArrayTy())
1474 && "Not looking at a struct or array?");
1475 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range)
1476 && "Invalid indices for type?");
1477 CompositeType *PTy = cast<CompositeType>(V->getType());
1479 if (isa<UndefValue>(V))
1480 return UndefValue::get(ExtractValueInst::getIndexedType(PTy,
1482 else if (isa<ConstantAggregateZero>(V))
1483 return Constant::getNullValue(ExtractValueInst::getIndexedType(PTy,
1485 else if (Constant *C = dyn_cast<Constant>(V)) {
1486 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C))
1487 // Recursively process this constant
1488 return FindInsertedValue(C->getOperand(idx_range[0]), idx_range.slice(1),
1490 } else if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1491 // Loop the indices for the insertvalue instruction in parallel with the
1492 // requested indices
1493 const unsigned *req_idx = idx_range.begin();
1494 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1495 i != e; ++i, ++req_idx) {
1496 if (req_idx == idx_range.end()) {
1498 // The requested index identifies a part of a nested aggregate. Handle
1499 // this specially. For example,
1500 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1501 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1502 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1503 // This can be changed into
1504 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1505 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1506 // which allows the unused 0,0 element from the nested struct to be
1508 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1511 // We can't handle this without inserting insertvalues
1515 // This insert value inserts something else than what we are looking for.
1516 // See if the (aggregrate) value inserted into has the value we are
1517 // looking for, then.
1519 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1522 // If we end up here, the indices of the insertvalue match with those
1523 // requested (though possibly only partially). Now we recursively look at
1524 // the inserted value, passing any remaining indices.
1525 return FindInsertedValue(I->getInsertedValueOperand(),
1526 makeArrayRef(req_idx, idx_range.end()),
1528 } else if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1529 // If we're extracting a value from an aggregrate that was extracted from
1530 // something else, we can extract from that something else directly instead.
1531 // However, we will need to chain I's indices with the requested indices.
1533 // Calculate the number of indices required
1534 unsigned size = I->getNumIndices() + idx_range.size();
1535 // Allocate some space to put the new indices in
1536 SmallVector<unsigned, 5> Idxs;
1538 // Add indices from the extract value instruction
1539 Idxs.append(I->idx_begin(), I->idx_end());
1541 // Add requested indices
1542 Idxs.append(idx_range.begin(), idx_range.end());
1544 assert(Idxs.size() == size
1545 && "Number of indices added not correct?");
1547 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1549 // Otherwise, we don't know (such as, extracting from a function return value
1550 // or load instruction)
1554 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1555 /// it can be expressed as a base pointer plus a constant offset. Return the
1556 /// base and offset to the caller.
1557 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1558 const TargetData &TD) {
1559 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1560 if (PtrOp == 0) return Ptr;
1562 // Just look through bitcasts.
1563 if (PtrOp->getOpcode() == Instruction::BitCast)
1564 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1566 // If this is a GEP with constant indices, we can look through it.
1567 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1568 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1570 gep_type_iterator GTI = gep_type_begin(GEP);
1571 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1573 ConstantInt *OpC = cast<ConstantInt>(*I);
1574 if (OpC->isZero()) continue;
1576 // Handle a struct and array indices which add their offset to the pointer.
1577 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1578 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1580 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1581 Offset += OpC->getSExtValue()*Size;
1585 // Re-sign extend from the pointer size if needed to get overflow edge cases
1587 unsigned PtrSize = TD.getPointerSizeInBits();
1589 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1591 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1595 /// GetConstantStringInfo - This function computes the length of a
1596 /// null-terminated C string pointed to by V. If successful, it returns true
1597 /// and returns the string in Str. If unsuccessful, it returns false.
1598 bool llvm::GetConstantStringInfo(const Value *V, std::string &Str,
1599 uint64_t Offset, bool StopAtNul) {
1600 // If V is NULL then return false;
1601 if (V == NULL) return false;
1603 // Look through bitcast instructions.
1604 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1605 return GetConstantStringInfo(BCI->getOperand(0), Str, Offset, StopAtNul);
1607 // If the value is not a GEP instruction nor a constant expression with a
1608 // GEP instruction, then return false because ConstantArray can't occur
1610 const User *GEP = 0;
1611 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1613 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1614 if (CE->getOpcode() == Instruction::BitCast)
1615 return GetConstantStringInfo(CE->getOperand(0), Str, Offset, StopAtNul);
1616 if (CE->getOpcode() != Instruction::GetElementPtr)
1622 // Make sure the GEP has exactly three arguments.
1623 if (GEP->getNumOperands() != 3)
1626 // Make sure the index-ee is a pointer to array of i8.
1627 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1628 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1629 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1632 // Check to make sure that the first operand of the GEP is an integer and
1633 // has value 0 so that we are sure we're indexing into the initializer.
1634 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1635 if (FirstIdx == 0 || !FirstIdx->isZero())
1638 // If the second index isn't a ConstantInt, then this is a variable index
1639 // into the array. If this occurs, we can't say anything meaningful about
1641 uint64_t StartIdx = 0;
1642 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1643 StartIdx = CI->getZExtValue();
1646 return GetConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset,
1650 // The GEP instruction, constant or instruction, must reference a global
1651 // variable that is a constant and is initialized. The referenced constant
1652 // initializer is the array that we'll use for optimization.
1653 const GlobalVariable* GV = dyn_cast<GlobalVariable>(V);
1654 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1656 const Constant *GlobalInit = GV->getInitializer();
1658 // Handle the all-zeros case
1659 if (GlobalInit->isNullValue()) {
1660 // This is a degenerate case. The initializer is constant zero so the
1661 // length of the string must be zero.
1666 // Must be a Constant Array
1667 const ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1668 if (Array == 0 || !Array->getType()->getElementType()->isIntegerTy(8))
1671 // Get the number of elements in the array
1672 uint64_t NumElts = Array->getType()->getNumElements();
1674 if (Offset > NumElts)
1677 // Traverse the constant array from 'Offset' which is the place the GEP refers
1679 Str.reserve(NumElts-Offset);
1680 for (unsigned i = Offset; i != NumElts; ++i) {
1681 const Constant *Elt = Array->getOperand(i);
1682 const ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1683 if (!CI) // This array isn't suitable, non-int initializer.
1685 if (StopAtNul && CI->isZero())
1686 return true; // we found end of string, success!
1687 Str += (char)CI->getZExtValue();
1690 // The array isn't null terminated, but maybe this is a memcpy, not a strcpy.
1694 // These next two are very similar to the above, but also look through PHI
1696 // TODO: See if we can integrate these two together.
1698 /// GetStringLengthH - If we can compute the length of the string pointed to by
1699 /// the specified pointer, return 'len+1'. If we can't, return 0.
1700 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1701 // Look through noop bitcast instructions.
1702 if (BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1703 return GetStringLengthH(BCI->getOperand(0), PHIs);
1705 // If this is a PHI node, there are two cases: either we have already seen it
1707 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1708 if (!PHIs.insert(PN))
1709 return ~0ULL; // already in the set.
1711 // If it was new, see if all the input strings are the same length.
1712 uint64_t LenSoFar = ~0ULL;
1713 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1714 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1715 if (Len == 0) return 0; // Unknown length -> unknown.
1717 if (Len == ~0ULL) continue;
1719 if (Len != LenSoFar && LenSoFar != ~0ULL)
1720 return 0; // Disagree -> unknown.
1724 // Success, all agree.
1728 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1729 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1730 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1731 if (Len1 == 0) return 0;
1732 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1733 if (Len2 == 0) return 0;
1734 if (Len1 == ~0ULL) return Len2;
1735 if (Len2 == ~0ULL) return Len1;
1736 if (Len1 != Len2) return 0;
1740 // As a special-case, "@string = constant i8 0" is also a string with zero
1741 // length, not wrapped in a bitcast or GEP.
1742 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
1743 if (GV->isConstant() && GV->hasDefinitiveInitializer())
1744 if (GV->getInitializer()->isNullValue()) return 1;
1748 // If the value is not a GEP instruction nor a constant expression with a
1749 // GEP instruction, then return unknown.
1751 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1753 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1754 if (CE->getOpcode() != Instruction::GetElementPtr)
1761 // Make sure the GEP has exactly three arguments.
1762 if (GEP->getNumOperands() != 3)
1765 // Check to make sure that the first operand of the GEP is an integer and
1766 // has value 0 so that we are sure we're indexing into the initializer.
1767 if (ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(1))) {
1773 // If the second index isn't a ConstantInt, then this is a variable index
1774 // into the array. If this occurs, we can't say anything meaningful about
1776 uint64_t StartIdx = 0;
1777 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1778 StartIdx = CI->getZExtValue();
1782 // The GEP instruction, constant or instruction, must reference a global
1783 // variable that is a constant and is initialized. The referenced constant
1784 // initializer is the array that we'll use for optimization.
1785 GlobalVariable* GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
1786 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
1787 GV->mayBeOverridden())
1789 Constant *GlobalInit = GV->getInitializer();
1791 // Handle the ConstantAggregateZero case, which is a degenerate case. The
1792 // initializer is constant zero so the length of the string must be zero.
1793 if (isa<ConstantAggregateZero>(GlobalInit))
1794 return 1; // Len = 0 offset by 1.
1796 // Must be a Constant Array
1797 ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1798 if (!Array || !Array->getType()->getElementType()->isIntegerTy(8))
1801 // Get the number of elements in the array
1802 uint64_t NumElts = Array->getType()->getNumElements();
1804 // Traverse the constant array from StartIdx (derived above) which is
1805 // the place the GEP refers to in the array.
1806 for (unsigned i = StartIdx; i != NumElts; ++i) {
1807 Constant *Elt = Array->getOperand(i);
1808 ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1809 if (!CI) // This array isn't suitable, non-int initializer.
1812 return i-StartIdx+1; // We found end of string, success!
1815 return 0; // The array isn't null terminated, conservatively return 'unknown'.
1818 /// GetStringLength - If we can compute the length of the string pointed to by
1819 /// the specified pointer, return 'len+1'. If we can't, return 0.
1820 uint64_t llvm::GetStringLength(Value *V) {
1821 if (!V->getType()->isPointerTy()) return 0;
1823 SmallPtrSet<PHINode*, 32> PHIs;
1824 uint64_t Len = GetStringLengthH(V, PHIs);
1825 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1826 // an empty string as a length.
1827 return Len == ~0ULL ? 1 : Len;
1831 llvm::GetUnderlyingObject(Value *V, const TargetData *TD, unsigned MaxLookup) {
1832 if (!V->getType()->isPointerTy())
1834 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1835 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1836 V = GEP->getPointerOperand();
1837 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1838 V = cast<Operator>(V)->getOperand(0);
1839 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1840 if (GA->mayBeOverridden())
1842 V = GA->getAliasee();
1844 // See if InstructionSimplify knows any relevant tricks.
1845 if (Instruction *I = dyn_cast<Instruction>(V))
1846 // TODO: Acquire a DominatorTree and use it.
1847 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1854 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1859 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1860 /// are lifetime markers.
1862 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1863 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1865 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1866 if (!II) return false;
1868 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1869 II->getIntrinsicID() != Intrinsic::lifetime_end)