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/Constants.h"
17 #include "llvm/Instructions.h"
18 #include "llvm/GlobalVariable.h"
19 #include "llvm/GlobalAlias.h"
20 #include "llvm/IntrinsicInst.h"
21 #include "llvm/LLVMContext.h"
22 #include "llvm/Operator.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/Support/GetElementPtrTypeIterator.h"
25 #include "llvm/Support/MathExtras.h"
26 #include "llvm/ADT/SmallPtrSet.h"
30 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
31 /// known to be either zero or one and return them in the KnownZero/KnownOne
32 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
34 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
35 /// we cannot optimize based on the assumption that it is zero without changing
36 /// it to be an explicit zero. If we don't change it to zero, other code could
37 /// optimized based on the contradictory assumption that it is non-zero.
38 /// Because instcombine aggressively folds operations with undef args anyway,
39 /// this won't lose us code quality.
41 /// This function is defined on values with integer type, values with pointer
42 /// type (but only if TD is non-null), and vectors of integers. In the case
43 /// where V is a vector, the mask, known zero, and known one values are the
44 /// same width as the vector element, and the bit is set only if it is true
45 /// for all of the elements in the vector.
46 void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
47 APInt &KnownZero, APInt &KnownOne,
48 const TargetData *TD, unsigned Depth) {
49 const unsigned MaxDepth = 6;
50 assert(V && "No Value?");
51 assert(Depth <= MaxDepth && "Limit Search Depth");
52 unsigned BitWidth = Mask.getBitWidth();
53 assert((V->getType()->isIntOrIntVectorTy() || V->getType()->isPointerTy())
54 && "Not integer or pointer type!");
56 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
57 (!V->getType()->isIntOrIntVectorTy() ||
58 V->getType()->getScalarSizeInBits() == BitWidth) &&
59 KnownZero.getBitWidth() == BitWidth &&
60 KnownOne.getBitWidth() == BitWidth &&
61 "V, Mask, KnownOne and KnownZero should have same BitWidth");
63 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
64 // We know all of the bits for a constant!
65 KnownOne = CI->getValue() & Mask;
66 KnownZero = ~KnownOne & Mask;
69 // Null and aggregate-zero are all-zeros.
70 if (isa<ConstantPointerNull>(V) ||
71 isa<ConstantAggregateZero>(V)) {
72 KnownOne.clearAllBits();
76 // Handle a constant vector by taking the intersection of the known bits of
78 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
79 KnownZero.setAllBits(); KnownOne.setAllBits();
80 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
81 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
82 ComputeMaskedBits(CV->getOperand(i), Mask, KnownZero2, KnownOne2,
84 KnownZero &= KnownZero2;
85 KnownOne &= KnownOne2;
89 // The address of an aligned GlobalValue has trailing zeros.
90 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
91 unsigned Align = GV->getAlignment();
92 if (Align == 0 && TD && GV->getType()->getElementType()->isSized()) {
93 const Type *ObjectType = GV->getType()->getElementType();
94 // If the object is defined in the current Module, we'll be giving
95 // it the preferred alignment. Otherwise, we have to assume that it
96 // may only have the minimum ABI alignment.
97 if (!GV->isDeclaration() && !GV->mayBeOverridden())
98 Align = TD->getPrefTypeAlignment(ObjectType);
100 Align = TD->getABITypeAlignment(ObjectType);
103 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
104 CountTrailingZeros_32(Align));
106 KnownZero.clearAllBits();
107 KnownOne.clearAllBits();
110 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
111 // the bits of its aliasee.
112 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
113 if (GA->mayBeOverridden()) {
114 KnownZero.clearAllBits(); KnownOne.clearAllBits();
116 ComputeMaskedBits(GA->getAliasee(), Mask, KnownZero, KnownOne,
122 KnownZero.clearAllBits(); KnownOne.clearAllBits(); // Start out not knowing anything.
124 if (Depth == MaxDepth || Mask == 0)
125 return; // Limit search depth.
127 Operator *I = dyn_cast<Operator>(V);
130 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
131 switch (I->getOpcode()) {
133 case Instruction::And: {
134 // If either the LHS or the RHS are Zero, the result is zero.
135 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
136 APInt Mask2(Mask & ~KnownZero);
137 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
139 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
140 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
142 // Output known-1 bits are only known if set in both the LHS & RHS.
143 KnownOne &= KnownOne2;
144 // Output known-0 are known to be clear if zero in either the LHS | RHS.
145 KnownZero |= KnownZero2;
148 case Instruction::Or: {
149 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
150 APInt Mask2(Mask & ~KnownOne);
151 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
153 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
154 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
156 // Output known-0 bits are only known if clear in both the LHS & RHS.
157 KnownZero &= KnownZero2;
158 // Output known-1 are known to be set if set in either the LHS | RHS.
159 KnownOne |= KnownOne2;
162 case Instruction::Xor: {
163 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
164 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
166 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
167 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
169 // Output known-0 bits are known if clear or set in both the LHS & RHS.
170 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
171 // Output known-1 are known to be set if set in only one of the LHS, RHS.
172 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
173 KnownZero = KnownZeroOut;
176 case Instruction::Mul: {
177 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
178 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
179 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
181 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
182 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
184 // If low bits are zero in either operand, output low known-0 bits.
185 // Also compute a conserative estimate for high known-0 bits.
186 // More trickiness is possible, but this is sufficient for the
187 // interesting case of alignment computation.
188 KnownOne.clearAllBits();
189 unsigned TrailZ = KnownZero.countTrailingOnes() +
190 KnownZero2.countTrailingOnes();
191 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
192 KnownZero2.countLeadingOnes(),
193 BitWidth) - BitWidth;
195 TrailZ = std::min(TrailZ, BitWidth);
196 LeadZ = std::min(LeadZ, BitWidth);
197 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
198 APInt::getHighBitsSet(BitWidth, LeadZ);
202 case Instruction::UDiv: {
203 // For the purposes of computing leading zeros we can conservatively
204 // treat a udiv as a logical right shift by the power of 2 known to
205 // be less than the denominator.
206 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
207 ComputeMaskedBits(I->getOperand(0),
208 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
209 unsigned LeadZ = KnownZero2.countLeadingOnes();
211 KnownOne2.clearAllBits();
212 KnownZero2.clearAllBits();
213 ComputeMaskedBits(I->getOperand(1),
214 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
215 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
216 if (RHSUnknownLeadingOnes != BitWidth)
217 LeadZ = std::min(BitWidth,
218 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
220 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
223 case Instruction::Select:
224 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
225 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
227 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
228 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
230 // Only known if known in both the LHS and RHS.
231 KnownOne &= KnownOne2;
232 KnownZero &= KnownZero2;
234 case Instruction::FPTrunc:
235 case Instruction::FPExt:
236 case Instruction::FPToUI:
237 case Instruction::FPToSI:
238 case Instruction::SIToFP:
239 case Instruction::UIToFP:
240 return; // Can't work with floating point.
241 case Instruction::PtrToInt:
242 case Instruction::IntToPtr:
243 // We can't handle these if we don't know the pointer size.
245 // FALL THROUGH and handle them the same as zext/trunc.
246 case Instruction::ZExt:
247 case Instruction::Trunc: {
248 const Type *SrcTy = I->getOperand(0)->getType();
250 unsigned SrcBitWidth;
251 // Note that we handle pointer operands here because of inttoptr/ptrtoint
252 // which fall through here.
253 if (SrcTy->isPointerTy())
254 SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
256 SrcBitWidth = SrcTy->getScalarSizeInBits();
258 APInt MaskIn = Mask.zextOrTrunc(SrcBitWidth);
259 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
260 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
261 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
263 KnownZero = KnownZero.zextOrTrunc(BitWidth);
264 KnownOne = KnownOne.zextOrTrunc(BitWidth);
265 // Any top bits are known to be zero.
266 if (BitWidth > SrcBitWidth)
267 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
270 case Instruction::BitCast: {
271 const Type *SrcTy = I->getOperand(0)->getType();
272 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
273 // TODO: For now, not handling conversions like:
274 // (bitcast i64 %x to <2 x i32>)
275 !I->getType()->isVectorTy()) {
276 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
282 case Instruction::SExt: {
283 // Compute the bits in the result that are not present in the input.
284 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
286 APInt MaskIn = Mask.trunc(SrcBitWidth);
287 KnownZero = KnownZero.trunc(SrcBitWidth);
288 KnownOne = KnownOne.trunc(SrcBitWidth);
289 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
291 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
292 KnownZero = KnownZero.zext(BitWidth);
293 KnownOne = KnownOne.zext(BitWidth);
295 // If the sign bit of the input is known set or clear, then we know the
296 // top bits of the result.
297 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
298 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
299 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
300 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
303 case Instruction::Shl:
304 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
305 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
306 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
307 APInt Mask2(Mask.lshr(ShiftAmt));
308 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
310 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
311 KnownZero <<= ShiftAmt;
312 KnownOne <<= ShiftAmt;
313 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
317 case Instruction::LShr:
318 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
319 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
320 // Compute the new bits that are at the top now.
321 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
323 // Unsigned shift right.
324 APInt Mask2(Mask.shl(ShiftAmt));
325 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
327 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
328 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
329 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
330 // high bits known zero.
331 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
335 case Instruction::AShr:
336 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
337 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
338 // Compute the new bits that are at the top now.
339 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
341 // Signed shift right.
342 APInt Mask2(Mask.shl(ShiftAmt));
343 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
345 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
346 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
347 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
349 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
350 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
351 KnownZero |= HighBits;
352 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
353 KnownOne |= HighBits;
357 case Instruction::Sub: {
358 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
359 // We know that the top bits of C-X are clear if X contains less bits
360 // than C (i.e. no wrap-around can happen). For example, 20-X is
361 // positive if we can prove that X is >= 0 and < 16.
362 if (!CLHS->getValue().isNegative()) {
363 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
364 // NLZ can't be BitWidth with no sign bit
365 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
366 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
369 // If all of the MaskV bits are known to be zero, then we know the
370 // output top bits are zero, because we now know that the output is
372 if ((KnownZero2 & MaskV) == MaskV) {
373 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
374 // Top bits known zero.
375 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
381 case Instruction::Add: {
382 // If one of the operands has trailing zeros, then the bits that the
383 // other operand has in those bit positions will be preserved in the
384 // result. For an add, this works with either operand. For a subtract,
385 // this only works if the known zeros are in the right operand.
386 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
387 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
388 BitWidth - Mask.countLeadingZeros());
389 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
391 assert((LHSKnownZero & LHSKnownOne) == 0 &&
392 "Bits known to be one AND zero?");
393 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
395 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD,
397 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
398 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
400 // Determine which operand has more trailing zeros, and use that
401 // many bits from the other operand.
402 if (LHSKnownZeroOut > RHSKnownZeroOut) {
403 if (I->getOpcode() == Instruction::Add) {
404 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
405 KnownZero |= KnownZero2 & Mask;
406 KnownOne |= KnownOne2 & Mask;
408 // If the known zeros are in the left operand for a subtract,
409 // fall back to the minimum known zeros in both operands.
410 KnownZero |= APInt::getLowBitsSet(BitWidth,
411 std::min(LHSKnownZeroOut,
414 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
415 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
416 KnownZero |= LHSKnownZero & Mask;
417 KnownOne |= LHSKnownOne & Mask;
421 case Instruction::SRem:
422 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
423 APInt RA = Rem->getValue().abs();
424 if (RA.isPowerOf2()) {
425 APInt LowBits = RA - 1;
426 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
427 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
430 // The low bits of the first operand are unchanged by the srem.
431 KnownZero = KnownZero2 & LowBits;
432 KnownOne = KnownOne2 & LowBits;
434 // If the first operand is non-negative or has all low bits zero, then
435 // the upper bits are all zero.
436 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
437 KnownZero |= ~LowBits;
439 // If the first operand is negative and not all low bits are zero, then
440 // the upper bits are all one.
441 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
442 KnownOne |= ~LowBits;
447 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
451 case Instruction::URem: {
452 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
453 APInt RA = Rem->getValue();
454 if (RA.isPowerOf2()) {
455 APInt LowBits = (RA - 1);
456 APInt Mask2 = LowBits & Mask;
457 KnownZero |= ~LowBits & Mask;
458 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
460 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
465 // Since the result is less than or equal to either operand, any leading
466 // zero bits in either operand must also exist in the result.
467 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
468 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
470 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
473 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
474 KnownZero2.countLeadingOnes());
475 KnownOne.clearAllBits();
476 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
480 case Instruction::Alloca: {
481 AllocaInst *AI = cast<AllocaInst>(V);
482 unsigned Align = AI->getAlignment();
483 if (Align == 0 && TD)
484 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
487 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
488 CountTrailingZeros_32(Align));
491 case Instruction::GetElementPtr: {
492 // Analyze all of the subscripts of this getelementptr instruction
493 // to determine if we can prove known low zero bits.
494 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
495 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
496 ComputeMaskedBits(I->getOperand(0), LocalMask,
497 LocalKnownZero, LocalKnownOne, TD, Depth+1);
498 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
500 gep_type_iterator GTI = gep_type_begin(I);
501 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
502 Value *Index = I->getOperand(i);
503 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
504 // Handle struct member offset arithmetic.
506 const StructLayout *SL = TD->getStructLayout(STy);
507 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
508 uint64_t Offset = SL->getElementOffset(Idx);
509 TrailZ = std::min(TrailZ,
510 CountTrailingZeros_64(Offset));
512 // Handle array index arithmetic.
513 const Type *IndexedTy = GTI.getIndexedType();
514 if (!IndexedTy->isSized()) return;
515 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
516 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
517 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
518 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
519 ComputeMaskedBits(Index, LocalMask,
520 LocalKnownZero, LocalKnownOne, TD, Depth+1);
521 TrailZ = std::min(TrailZ,
522 unsigned(CountTrailingZeros_64(TypeSize) +
523 LocalKnownZero.countTrailingOnes()));
527 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
530 case Instruction::PHI: {
531 PHINode *P = cast<PHINode>(I);
532 // Handle the case of a simple two-predecessor recurrence PHI.
533 // There's a lot more that could theoretically be done here, but
534 // this is sufficient to catch some interesting cases.
535 if (P->getNumIncomingValues() == 2) {
536 for (unsigned i = 0; i != 2; ++i) {
537 Value *L = P->getIncomingValue(i);
538 Value *R = P->getIncomingValue(!i);
539 Operator *LU = dyn_cast<Operator>(L);
542 unsigned Opcode = LU->getOpcode();
543 // Check for operations that have the property that if
544 // both their operands have low zero bits, the result
545 // will have low zero bits.
546 if (Opcode == Instruction::Add ||
547 Opcode == Instruction::Sub ||
548 Opcode == Instruction::And ||
549 Opcode == Instruction::Or ||
550 Opcode == Instruction::Mul) {
551 Value *LL = LU->getOperand(0);
552 Value *LR = LU->getOperand(1);
553 // Find a recurrence.
560 // Ok, we have a PHI of the form L op= R. Check for low
562 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
563 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
564 Mask2 = APInt::getLowBitsSet(BitWidth,
565 KnownZero2.countTrailingOnes());
567 // We need to take the minimum number of known bits
568 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
569 ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
572 APInt::getLowBitsSet(BitWidth,
573 std::min(KnownZero2.countTrailingOnes(),
574 KnownZero3.countTrailingOnes()));
580 // Otherwise take the unions of the known bit sets of the operands,
581 // taking conservative care to avoid excessive recursion.
582 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
583 KnownZero = APInt::getAllOnesValue(BitWidth);
584 KnownOne = APInt::getAllOnesValue(BitWidth);
585 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
586 // Skip direct self references.
587 if (P->getIncomingValue(i) == P) continue;
589 KnownZero2 = APInt(BitWidth, 0);
590 KnownOne2 = APInt(BitWidth, 0);
591 // Recurse, but cap the recursion to one level, because we don't
592 // want to waste time spinning around in loops.
593 ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
594 KnownZero2, KnownOne2, TD, MaxDepth-1);
595 KnownZero &= KnownZero2;
596 KnownOne &= KnownOne2;
597 // If all bits have been ruled out, there's no need to check
599 if (!KnownZero && !KnownOne)
605 case Instruction::Call:
606 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
607 switch (II->getIntrinsicID()) {
609 case Intrinsic::ctpop:
610 case Intrinsic::ctlz:
611 case Intrinsic::cttz: {
612 unsigned LowBits = Log2_32(BitWidth)+1;
613 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
622 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
623 /// this predicate to simplify operations downstream. Mask is known to be zero
624 /// for bits that V cannot have.
626 /// This function is defined on values with integer type, values with pointer
627 /// type (but only if TD is non-null), and vectors of integers. In the case
628 /// where V is a vector, the mask, known zero, and known one values are the
629 /// same width as the vector element, and the bit is set only if it is true
630 /// for all of the elements in the vector.
631 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
632 const TargetData *TD, unsigned Depth) {
633 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
634 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
635 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
636 return (KnownZero & Mask) == Mask;
641 /// ComputeNumSignBits - Return the number of times the sign bit of the
642 /// register is replicated into the other bits. We know that at least 1 bit
643 /// is always equal to the sign bit (itself), but other cases can give us
644 /// information. For example, immediately after an "ashr X, 2", we know that
645 /// the top 3 bits are all equal to each other, so we return 3.
647 /// 'Op' must have a scalar integer type.
649 unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
651 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
652 "ComputeNumSignBits requires a TargetData object to operate "
653 "on non-integer values!");
654 const Type *Ty = V->getType();
655 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
656 Ty->getScalarSizeInBits();
658 unsigned FirstAnswer = 1;
660 // Note that ConstantInt is handled by the general ComputeMaskedBits case
664 return 1; // Limit search depth.
666 Operator *U = dyn_cast<Operator>(V);
667 switch (Operator::getOpcode(V)) {
669 case Instruction::SExt:
670 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
671 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
673 case Instruction::AShr:
674 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
675 // ashr X, C -> adds C sign bits.
676 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
677 Tmp += C->getZExtValue();
678 if (Tmp > TyBits) Tmp = TyBits;
681 case Instruction::Shl:
682 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
683 // shl destroys sign bits.
684 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
685 if (C->getZExtValue() >= TyBits || // Bad shift.
686 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
687 return Tmp - C->getZExtValue();
690 case Instruction::And:
691 case Instruction::Or:
692 case Instruction::Xor: // NOT is handled here.
693 // Logical binary ops preserve the number of sign bits at the worst.
694 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
696 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
697 FirstAnswer = std::min(Tmp, Tmp2);
698 // We computed what we know about the sign bits as our first
699 // answer. Now proceed to the generic code that uses
700 // ComputeMaskedBits, and pick whichever answer is better.
704 case Instruction::Select:
705 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
706 if (Tmp == 1) return 1; // Early out.
707 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
708 return std::min(Tmp, Tmp2);
710 case Instruction::Add:
711 // Add can have at most one carry bit. Thus we know that the output
712 // is, at worst, one more bit than the inputs.
713 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
714 if (Tmp == 1) return 1; // Early out.
716 // Special case decrementing a value (ADD X, -1):
717 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
718 if (CRHS->isAllOnesValue()) {
719 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
720 APInt Mask = APInt::getAllOnesValue(TyBits);
721 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
724 // If the input is known to be 0 or 1, the output is 0/-1, which is all
726 if ((KnownZero | APInt(TyBits, 1)) == Mask)
729 // If we are subtracting one from a positive number, there is no carry
730 // out of the result.
731 if (KnownZero.isNegative())
735 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
736 if (Tmp2 == 1) return 1;
737 return std::min(Tmp, Tmp2)-1;
739 case Instruction::Sub:
740 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
741 if (Tmp2 == 1) return 1;
744 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
745 if (CLHS->isNullValue()) {
746 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
747 APInt Mask = APInt::getAllOnesValue(TyBits);
748 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
750 // If the input is known to be 0 or 1, the output is 0/-1, which is all
752 if ((KnownZero | APInt(TyBits, 1)) == Mask)
755 // If the input is known to be positive (the sign bit is known clear),
756 // the output of the NEG has the same number of sign bits as the input.
757 if (KnownZero.isNegative())
760 // Otherwise, we treat this like a SUB.
763 // Sub can have at most one carry bit. Thus we know that the output
764 // is, at worst, one more bit than the inputs.
765 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
766 if (Tmp == 1) return 1; // Early out.
767 return std::min(Tmp, Tmp2)-1;
769 case Instruction::PHI: {
770 PHINode *PN = cast<PHINode>(U);
771 // Don't analyze large in-degree PHIs.
772 if (PN->getNumIncomingValues() > 4) break;
774 // Take the minimum of all incoming values. This can't infinitely loop
775 // because of our depth threshold.
776 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
777 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
778 if (Tmp == 1) return Tmp;
780 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
785 case Instruction::Trunc:
786 // FIXME: it's tricky to do anything useful for this, but it is an important
787 // case for targets like X86.
791 // Finally, if we can prove that the top bits of the result are 0's or 1's,
792 // use this information.
793 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
794 APInt Mask = APInt::getAllOnesValue(TyBits);
795 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
797 if (KnownZero.isNegative()) { // sign bit is 0
799 } else if (KnownOne.isNegative()) { // sign bit is 1;
806 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
807 // the number of identical bits in the top of the input value.
809 Mask <<= Mask.getBitWidth()-TyBits;
810 // Return # leading zeros. We use 'min' here in case Val was zero before
811 // shifting. We don't want to return '64' as for an i32 "0".
812 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
815 /// ComputeMultiple - This function computes the integer multiple of Base that
816 /// equals V. If successful, it returns true and returns the multiple in
817 /// Multiple. If unsuccessful, it returns false. It looks
818 /// through SExt instructions only if LookThroughSExt is true.
819 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
820 bool LookThroughSExt, unsigned Depth) {
821 const unsigned MaxDepth = 6;
823 assert(V && "No Value?");
824 assert(Depth <= MaxDepth && "Limit Search Depth");
825 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
827 const Type *T = V->getType();
829 ConstantInt *CI = dyn_cast<ConstantInt>(V);
839 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
840 Constant *BaseVal = ConstantInt::get(T, Base);
841 if (CO && CO == BaseVal) {
843 Multiple = ConstantInt::get(T, 1);
847 if (CI && CI->getZExtValue() % Base == 0) {
848 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
852 if (Depth == MaxDepth) return false; // Limit search depth.
854 Operator *I = dyn_cast<Operator>(V);
855 if (!I) return false;
857 switch (I->getOpcode()) {
859 case Instruction::SExt:
860 if (!LookThroughSExt) return false;
861 // otherwise fall through to ZExt
862 case Instruction::ZExt:
863 return ComputeMultiple(I->getOperand(0), Base, Multiple,
864 LookThroughSExt, Depth+1);
865 case Instruction::Shl:
866 case Instruction::Mul: {
867 Value *Op0 = I->getOperand(0);
868 Value *Op1 = I->getOperand(1);
870 if (I->getOpcode() == Instruction::Shl) {
871 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
872 if (!Op1CI) return false;
873 // Turn Op0 << Op1 into Op0 * 2^Op1
874 APInt Op1Int = Op1CI->getValue();
875 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
876 APInt API(Op1Int.getBitWidth(), 0);
877 API.setBit(BitToSet);
878 Op1 = ConstantInt::get(V->getContext(), API);
882 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
883 if (Constant *Op1C = dyn_cast<Constant>(Op1))
884 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
885 if (Op1C->getType()->getPrimitiveSizeInBits() <
886 MulC->getType()->getPrimitiveSizeInBits())
887 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
888 if (Op1C->getType()->getPrimitiveSizeInBits() >
889 MulC->getType()->getPrimitiveSizeInBits())
890 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
892 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
893 Multiple = ConstantExpr::getMul(MulC, Op1C);
897 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
898 if (Mul0CI->getValue() == 1) {
899 // V == Base * Op1, so return Op1
906 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
907 if (Constant *Op0C = dyn_cast<Constant>(Op0))
908 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
909 if (Op0C->getType()->getPrimitiveSizeInBits() <
910 MulC->getType()->getPrimitiveSizeInBits())
911 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
912 if (Op0C->getType()->getPrimitiveSizeInBits() >
913 MulC->getType()->getPrimitiveSizeInBits())
914 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
916 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
917 Multiple = ConstantExpr::getMul(MulC, Op0C);
921 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
922 if (Mul1CI->getValue() == 1) {
923 // V == Base * Op0, so return Op0
931 // We could not determine if V is a multiple of Base.
935 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
936 /// value is never equal to -0.0.
938 /// NOTE: this function will need to be revisited when we support non-default
941 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
942 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
943 return !CFP->getValueAPF().isNegZero();
946 return 1; // Limit search depth.
948 const Operator *I = dyn_cast<Operator>(V);
949 if (I == 0) return false;
951 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
952 if (I->getOpcode() == Instruction::FAdd &&
953 isa<ConstantFP>(I->getOperand(1)) &&
954 cast<ConstantFP>(I->getOperand(1))->isNullValue())
957 // sitofp and uitofp turn into +0.0 for zero.
958 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
961 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
962 // sqrt(-0.0) = -0.0, no other negative results are possible.
963 if (II->getIntrinsicID() == Intrinsic::sqrt)
964 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
966 if (const CallInst *CI = dyn_cast<CallInst>(I))
967 if (const Function *F = CI->getCalledFunction()) {
968 if (F->isDeclaration()) {
970 if (F->getName() == "abs") return true;
971 // fabs[lf](x) != -0.0
972 if (F->getName() == "fabs") return true;
973 if (F->getName() == "fabsf") return true;
974 if (F->getName() == "fabsl") return true;
975 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
976 F->getName() == "sqrtl")
977 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
984 // This is the recursive version of BuildSubAggregate. It takes a few different
985 // arguments. Idxs is the index within the nested struct From that we are
986 // looking at now (which is of type IndexedType). IdxSkip is the number of
987 // indices from Idxs that should be left out when inserting into the resulting
988 // struct. To is the result struct built so far, new insertvalue instructions
990 static Value *BuildSubAggregate(Value *From, Value* To, const Type *IndexedType,
991 SmallVector<unsigned, 10> &Idxs,
993 Instruction *InsertBefore) {
994 const llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
996 // Save the original To argument so we can modify it
998 // General case, the type indexed by Idxs is a struct
999 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1000 // Process each struct element recursively
1003 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1007 // Couldn't find any inserted value for this index? Cleanup
1008 while (PrevTo != OrigTo) {
1009 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1010 PrevTo = Del->getAggregateOperand();
1011 Del->eraseFromParent();
1013 // Stop processing elements
1017 // If we succesfully found a value for each of our subaggregates
1021 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1022 // the struct's elements had a value that was inserted directly. In the latter
1023 // case, perhaps we can't determine each of the subelements individually, but
1024 // we might be able to find the complete struct somewhere.
1026 // Find the value that is at that particular spot
1027 Value *V = FindInsertedValue(From, Idxs.begin(), Idxs.end());
1032 // Insert the value in the new (sub) aggregrate
1033 return llvm::InsertValueInst::Create(To, V, Idxs.begin() + IdxSkip,
1034 Idxs.end(), "tmp", InsertBefore);
1037 // This helper takes a nested struct and extracts a part of it (which is again a
1038 // struct) into a new value. For example, given the struct:
1039 // { a, { b, { c, d }, e } }
1040 // and the indices "1, 1" this returns
1043 // It does this by inserting an insertvalue for each element in the resulting
1044 // struct, as opposed to just inserting a single struct. This will only work if
1045 // each of the elements of the substruct are known (ie, inserted into From by an
1046 // insertvalue instruction somewhere).
1048 // All inserted insertvalue instructions are inserted before InsertBefore
1049 static Value *BuildSubAggregate(Value *From, const unsigned *idx_begin,
1050 const unsigned *idx_end,
1051 Instruction *InsertBefore) {
1052 assert(InsertBefore && "Must have someplace to insert!");
1053 const Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1056 Value *To = UndefValue::get(IndexedType);
1057 SmallVector<unsigned, 10> Idxs(idx_begin, idx_end);
1058 unsigned IdxSkip = Idxs.size();
1060 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1063 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1064 /// the scalar value indexed is already around as a register, for example if it
1065 /// were inserted directly into the aggregrate.
1067 /// If InsertBefore is not null, this function will duplicate (modified)
1068 /// insertvalues when a part of a nested struct is extracted.
1069 Value *llvm::FindInsertedValue(Value *V, const unsigned *idx_begin,
1070 const unsigned *idx_end, Instruction *InsertBefore) {
1071 // Nothing to index? Just return V then (this is useful at the end of our
1073 if (idx_begin == idx_end)
1075 // We have indices, so V should have an indexable type
1076 assert((V->getType()->isStructTy() || V->getType()->isArrayTy())
1077 && "Not looking at a struct or array?");
1078 assert(ExtractValueInst::getIndexedType(V->getType(), idx_begin, idx_end)
1079 && "Invalid indices for type?");
1080 const CompositeType *PTy = cast<CompositeType>(V->getType());
1082 if (isa<UndefValue>(V))
1083 return UndefValue::get(ExtractValueInst::getIndexedType(PTy,
1086 else if (isa<ConstantAggregateZero>(V))
1087 return Constant::getNullValue(ExtractValueInst::getIndexedType(PTy,
1090 else if (Constant *C = dyn_cast<Constant>(V)) {
1091 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C))
1092 // Recursively process this constant
1093 return FindInsertedValue(C->getOperand(*idx_begin), idx_begin + 1,
1094 idx_end, InsertBefore);
1095 } else if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1096 // Loop the indices for the insertvalue instruction in parallel with the
1097 // requested indices
1098 const unsigned *req_idx = idx_begin;
1099 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1100 i != e; ++i, ++req_idx) {
1101 if (req_idx == idx_end) {
1103 // The requested index identifies a part of a nested aggregate. Handle
1104 // this specially. For example,
1105 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1106 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1107 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1108 // This can be changed into
1109 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1110 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1111 // which allows the unused 0,0 element from the nested struct to be
1113 return BuildSubAggregate(V, idx_begin, req_idx, InsertBefore);
1115 // We can't handle this without inserting insertvalues
1119 // This insert value inserts something else than what we are looking for.
1120 // See if the (aggregrate) value inserted into has the value we are
1121 // looking for, then.
1123 return FindInsertedValue(I->getAggregateOperand(), idx_begin, idx_end,
1126 // If we end up here, the indices of the insertvalue match with those
1127 // requested (though possibly only partially). Now we recursively look at
1128 // the inserted value, passing any remaining indices.
1129 return FindInsertedValue(I->getInsertedValueOperand(), req_idx, idx_end,
1131 } else if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1132 // If we're extracting a value from an aggregrate that was extracted from
1133 // something else, we can extract from that something else directly instead.
1134 // However, we will need to chain I's indices with the requested indices.
1136 // Calculate the number of indices required
1137 unsigned size = I->getNumIndices() + (idx_end - idx_begin);
1138 // Allocate some space to put the new indices in
1139 SmallVector<unsigned, 5> Idxs;
1141 // Add indices from the extract value instruction
1142 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1146 // Add requested indices
1147 for (const unsigned *i = idx_begin, *e = idx_end; i != e; ++i)
1150 assert(Idxs.size() == size
1151 && "Number of indices added not correct?");
1153 return FindInsertedValue(I->getAggregateOperand(), Idxs.begin(), Idxs.end(),
1156 // Otherwise, we don't know (such as, extracting from a function return value
1157 // or load instruction)
1161 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1162 /// it can be expressed as a base pointer plus a constant offset. Return the
1163 /// base and offset to the caller.
1164 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1165 const TargetData &TD) {
1166 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1167 if (PtrOp == 0) return Ptr;
1169 // Just look through bitcasts.
1170 if (PtrOp->getOpcode() == Instruction::BitCast)
1171 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1173 // If this is a GEP with constant indices, we can look through it.
1174 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1175 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1177 gep_type_iterator GTI = gep_type_begin(GEP);
1178 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1180 ConstantInt *OpC = cast<ConstantInt>(*I);
1181 if (OpC->isZero()) continue;
1183 // Handle a struct and array indices which add their offset to the pointer.
1184 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
1185 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1187 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1188 Offset += OpC->getSExtValue()*Size;
1192 // Re-sign extend from the pointer size if needed to get overflow edge cases
1194 unsigned PtrSize = TD.getPointerSizeInBits();
1196 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1198 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1202 /// GetConstantStringInfo - This function computes the length of a
1203 /// null-terminated C string pointed to by V. If successful, it returns true
1204 /// and returns the string in Str. If unsuccessful, it returns false.
1205 bool llvm::GetConstantStringInfo(const Value *V, std::string &Str,
1208 // If V is NULL then return false;
1209 if (V == NULL) return false;
1211 // Look through bitcast instructions.
1212 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1213 return GetConstantStringInfo(BCI->getOperand(0), Str, Offset, StopAtNul);
1215 // If the value is not a GEP instruction nor a constant expression with a
1216 // GEP instruction, then return false because ConstantArray can't occur
1218 const User *GEP = 0;
1219 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1221 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1222 if (CE->getOpcode() == Instruction::BitCast)
1223 return GetConstantStringInfo(CE->getOperand(0), Str, Offset, StopAtNul);
1224 if (CE->getOpcode() != Instruction::GetElementPtr)
1230 // Make sure the GEP has exactly three arguments.
1231 if (GEP->getNumOperands() != 3)
1234 // Make sure the index-ee is a pointer to array of i8.
1235 const PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1236 const ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1237 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1240 // Check to make sure that the first operand of the GEP is an integer and
1241 // has value 0 so that we are sure we're indexing into the initializer.
1242 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1243 if (FirstIdx == 0 || !FirstIdx->isZero())
1246 // If the second index isn't a ConstantInt, then this is a variable index
1247 // into the array. If this occurs, we can't say anything meaningful about
1249 uint64_t StartIdx = 0;
1250 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1251 StartIdx = CI->getZExtValue();
1254 return GetConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset,
1258 // The GEP instruction, constant or instruction, must reference a global
1259 // variable that is a constant and is initialized. The referenced constant
1260 // initializer is the array that we'll use for optimization.
1261 const GlobalVariable* GV = dyn_cast<GlobalVariable>(V);
1262 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1264 const Constant *GlobalInit = GV->getInitializer();
1266 // Handle the ConstantAggregateZero case
1267 if (isa<ConstantAggregateZero>(GlobalInit)) {
1268 // This is a degenerate case. The initializer is constant zero so the
1269 // length of the string must be zero.
1274 // Must be a Constant Array
1275 const ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1276 if (Array == 0 || !Array->getType()->getElementType()->isIntegerTy(8))
1279 // Get the number of elements in the array
1280 uint64_t NumElts = Array->getType()->getNumElements();
1282 if (Offset > NumElts)
1285 // Traverse the constant array from 'Offset' which is the place the GEP refers
1287 Str.reserve(NumElts-Offset);
1288 for (unsigned i = Offset; i != NumElts; ++i) {
1289 const Constant *Elt = Array->getOperand(i);
1290 const ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1291 if (!CI) // This array isn't suitable, non-int initializer.
1293 if (StopAtNul && CI->isZero())
1294 return true; // we found end of string, success!
1295 Str += (char)CI->getZExtValue();
1298 // The array isn't null terminated, but maybe this is a memcpy, not a strcpy.
1302 // These next two are very similar to the above, but also look through PHI
1304 // TODO: See if we can integrate these two together.
1306 /// GetStringLengthH - If we can compute the length of the string pointed to by
1307 /// the specified pointer, return 'len+1'. If we can't, return 0.
1308 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1309 // Look through noop bitcast instructions.
1310 if (BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1311 return GetStringLengthH(BCI->getOperand(0), PHIs);
1313 // If this is a PHI node, there are two cases: either we have already seen it
1315 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1316 if (!PHIs.insert(PN))
1317 return ~0ULL; // already in the set.
1319 // If it was new, see if all the input strings are the same length.
1320 uint64_t LenSoFar = ~0ULL;
1321 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1322 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1323 if (Len == 0) return 0; // Unknown length -> unknown.
1325 if (Len == ~0ULL) continue;
1327 if (Len != LenSoFar && LenSoFar != ~0ULL)
1328 return 0; // Disagree -> unknown.
1332 // Success, all agree.
1336 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1337 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1338 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1339 if (Len1 == 0) return 0;
1340 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1341 if (Len2 == 0) return 0;
1342 if (Len1 == ~0ULL) return Len2;
1343 if (Len2 == ~0ULL) return Len1;
1344 if (Len1 != Len2) return 0;
1348 // If the value is not a GEP instruction nor a constant expression with a
1349 // GEP instruction, then return unknown.
1351 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1353 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1354 if (CE->getOpcode() != Instruction::GetElementPtr)
1361 // Make sure the GEP has exactly three arguments.
1362 if (GEP->getNumOperands() != 3)
1365 // Check to make sure that the first operand of the GEP is an integer and
1366 // has value 0 so that we are sure we're indexing into the initializer.
1367 if (ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(1))) {
1373 // If the second index isn't a ConstantInt, then this is a variable index
1374 // into the array. If this occurs, we can't say anything meaningful about
1376 uint64_t StartIdx = 0;
1377 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1378 StartIdx = CI->getZExtValue();
1382 // The GEP instruction, constant or instruction, must reference a global
1383 // variable that is a constant and is initialized. The referenced constant
1384 // initializer is the array that we'll use for optimization.
1385 GlobalVariable* GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
1386 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
1387 GV->mayBeOverridden())
1389 Constant *GlobalInit = GV->getInitializer();
1391 // Handle the ConstantAggregateZero case, which is a degenerate case. The
1392 // initializer is constant zero so the length of the string must be zero.
1393 if (isa<ConstantAggregateZero>(GlobalInit))
1394 return 1; // Len = 0 offset by 1.
1396 // Must be a Constant Array
1397 ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1398 if (!Array || !Array->getType()->getElementType()->isIntegerTy(8))
1401 // Get the number of elements in the array
1402 uint64_t NumElts = Array->getType()->getNumElements();
1404 // Traverse the constant array from StartIdx (derived above) which is
1405 // the place the GEP refers to in the array.
1406 for (unsigned i = StartIdx; i != NumElts; ++i) {
1407 Constant *Elt = Array->getOperand(i);
1408 ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1409 if (!CI) // This array isn't suitable, non-int initializer.
1412 return i-StartIdx+1; // We found end of string, success!
1415 return 0; // The array isn't null terminated, conservatively return 'unknown'.
1418 /// GetStringLength - If we can compute the length of the string pointed to by
1419 /// the specified pointer, return 'len+1'. If we can't, return 0.
1420 uint64_t llvm::GetStringLength(Value *V) {
1421 if (!V->getType()->isPointerTy()) return 0;
1423 SmallPtrSet<PHINode*, 32> PHIs;
1424 uint64_t Len = GetStringLengthH(V, PHIs);
1425 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1426 // an empty string as a length.
1427 return Len == ~0ULL ? 1 : Len;
1430 Value *llvm::GetUnderlyingObject(Value *V, unsigned MaxLookup) {
1431 if (!V->getType()->isPointerTy())
1433 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1434 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1435 V = GEP->getPointerOperand();
1436 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1437 V = cast<Operator>(V)->getOperand(0);
1438 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1439 if (GA->mayBeOverridden())
1441 V = GA->getAliasee();
1445 assert(V->getType()->isPointerTy() && "Unexpected operand type!");