• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains routines that help analyze properties that chains of
11 // computations have.
12 //
13 //===----------------------------------------------------------------------===//
14 
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"
29 #include <cstring>
30 using namespace llvm;
31 using namespace llvm::PatternMatch;
32 
33 const unsigned MaxDepth = 6;
34 
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.
getBitWidth(Type * Ty,const TargetData * TD)37 static unsigned getBitWidth(Type *Ty, const TargetData *TD) {
38   if (unsigned BitWidth = Ty->getScalarSizeInBits())
39     return BitWidth;
40   assert(isa<PointerType>(Ty) && "Expected a pointer type!");
41   return TD ? TD->getPointerSizeInBits() : 0;
42 }
43 
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
47 /// processing.
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.
54 ///
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.
ComputeMaskedBits(Value * V,const APInt & Mask,APInt & KnownZero,APInt & KnownOne,const TargetData * TD,unsigned Depth)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!");
68   assert((!TD ||
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");
75 
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;
80     return;
81   }
82   // Null and aggregate-zero are all-zeros.
83   if (isa<ConstantPointerNull>(V) ||
84       isa<ConstantAggregateZero>(V)) {
85     KnownOne.clearAllBits();
86     KnownZero = Mask;
87     return;
88   }
89   // Handle a constant vector by taking the intersection of the known bits of
90   // each element.
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,
96                         TD, Depth);
97       KnownZero &= KnownZero2;
98       KnownOne &= KnownOne2;
99     }
100     return;
101   }
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       Type *ObjectType = GV->getType()->getElementType();
107       // If the object is defined in the current Module, we'll be giving
108       // it the preferred alignment. Otherwise, we have to assume that it
109       // may only have the minimum ABI alignment.
110       if (!GV->isDeclaration() && !GV->mayBeOverridden())
111         Align = TD->getPrefTypeAlignment(ObjectType);
112       else
113         Align = TD->getABITypeAlignment(ObjectType);
114     }
115     if (Align > 0)
116       KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
117                                               CountTrailingZeros_32(Align));
118     else
119       KnownZero.clearAllBits();
120     KnownOne.clearAllBits();
121     return;
122   }
123   // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
124   // the bits of its aliasee.
125   if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
126     if (GA->mayBeOverridden()) {
127       KnownZero.clearAllBits(); KnownOne.clearAllBits();
128     } else {
129       ComputeMaskedBits(GA->getAliasee(), Mask, KnownZero, KnownOne,
130                         TD, Depth+1);
131     }
132     return;
133   }
134 
135   if (Argument *A = dyn_cast<Argument>(V)) {
136     // Get alignment information off byval arguments if specified in the IR.
137     if (A->hasByValAttr())
138       if (unsigned Align = A->getParamAlignment())
139         KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
140                                                 CountTrailingZeros_32(Align));
141     return;
142   }
143 
144   // Start out not knowing anything.
145   KnownZero.clearAllBits(); KnownOne.clearAllBits();
146 
147   if (Depth == MaxDepth || Mask == 0)
148     return;  // Limit search depth.
149 
150   Operator *I = dyn_cast<Operator>(V);
151   if (!I) return;
152 
153   APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
154   switch (I->getOpcode()) {
155   default: break;
156   case Instruction::And: {
157     // If either the LHS or the RHS are Zero, the result is zero.
158     ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
159     APInt Mask2(Mask & ~KnownZero);
160     ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
161                       Depth+1);
162     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
163     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
164 
165     // Output known-1 bits are only known if set in both the LHS & RHS.
166     KnownOne &= KnownOne2;
167     // Output known-0 are known to be clear if zero in either the LHS | RHS.
168     KnownZero |= KnownZero2;
169     return;
170   }
171   case Instruction::Or: {
172     ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
173     APInt Mask2(Mask & ~KnownOne);
174     ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
175                       Depth+1);
176     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
177     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
178 
179     // Output known-0 bits are only known if clear in both the LHS & RHS.
180     KnownZero &= KnownZero2;
181     // Output known-1 are known to be set if set in either the LHS | RHS.
182     KnownOne |= KnownOne2;
183     return;
184   }
185   case Instruction::Xor: {
186     ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
187     ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
188                       Depth+1);
189     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
190     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
191 
192     // Output known-0 bits are known if clear or set in both the LHS & RHS.
193     APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
194     // Output known-1 are known to be set if set in only one of the LHS, RHS.
195     KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
196     KnownZero = KnownZeroOut;
197     return;
198   }
199   case Instruction::Mul: {
200     APInt Mask2 = APInt::getAllOnesValue(BitWidth);
201     ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
202     ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
203                       Depth+1);
204     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
205     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
206 
207     // If low bits are zero in either operand, output low known-0 bits.
208     // Also compute a conserative estimate for high known-0 bits.
209     // More trickiness is possible, but this is sufficient for the
210     // interesting case of alignment computation.
211     KnownOne.clearAllBits();
212     unsigned TrailZ = KnownZero.countTrailingOnes() +
213                       KnownZero2.countTrailingOnes();
214     unsigned LeadZ =  std::max(KnownZero.countLeadingOnes() +
215                                KnownZero2.countLeadingOnes(),
216                                BitWidth) - BitWidth;
217 
218     TrailZ = std::min(TrailZ, BitWidth);
219     LeadZ = std::min(LeadZ, BitWidth);
220     KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
221                 APInt::getHighBitsSet(BitWidth, LeadZ);
222     KnownZero &= Mask;
223     return;
224   }
225   case Instruction::UDiv: {
226     // For the purposes of computing leading zeros we can conservatively
227     // treat a udiv as a logical right shift by the power of 2 known to
228     // be less than the denominator.
229     APInt AllOnes = APInt::getAllOnesValue(BitWidth);
230     ComputeMaskedBits(I->getOperand(0),
231                       AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
232     unsigned LeadZ = KnownZero2.countLeadingOnes();
233 
234     KnownOne2.clearAllBits();
235     KnownZero2.clearAllBits();
236     ComputeMaskedBits(I->getOperand(1),
237                       AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
238     unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
239     if (RHSUnknownLeadingOnes != BitWidth)
240       LeadZ = std::min(BitWidth,
241                        LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
242 
243     KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
244     return;
245   }
246   case Instruction::Select:
247     ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
248     ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
249                       Depth+1);
250     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
251     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
252 
253     // Only known if known in both the LHS and RHS.
254     KnownOne &= KnownOne2;
255     KnownZero &= KnownZero2;
256     return;
257   case Instruction::FPTrunc:
258   case Instruction::FPExt:
259   case Instruction::FPToUI:
260   case Instruction::FPToSI:
261   case Instruction::SIToFP:
262   case Instruction::UIToFP:
263     return; // Can't work with floating point.
264   case Instruction::PtrToInt:
265   case Instruction::IntToPtr:
266     // We can't handle these if we don't know the pointer size.
267     if (!TD) return;
268     // FALL THROUGH and handle them the same as zext/trunc.
269   case Instruction::ZExt:
270   case Instruction::Trunc: {
271     Type *SrcTy = I->getOperand(0)->getType();
272 
273     unsigned SrcBitWidth;
274     // Note that we handle pointer operands here because of inttoptr/ptrtoint
275     // which fall through here.
276     if (SrcTy->isPointerTy())
277       SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
278     else
279       SrcBitWidth = SrcTy->getScalarSizeInBits();
280 
281     APInt MaskIn = Mask.zextOrTrunc(SrcBitWidth);
282     KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
283     KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
284     ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
285                       Depth+1);
286     KnownZero = KnownZero.zextOrTrunc(BitWidth);
287     KnownOne = KnownOne.zextOrTrunc(BitWidth);
288     // Any top bits are known to be zero.
289     if (BitWidth > SrcBitWidth)
290       KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
291     return;
292   }
293   case Instruction::BitCast: {
294     Type *SrcTy = I->getOperand(0)->getType();
295     if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
296         // TODO: For now, not handling conversions like:
297         // (bitcast i64 %x to <2 x i32>)
298         !I->getType()->isVectorTy()) {
299       ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
300                         Depth+1);
301       return;
302     }
303     break;
304   }
305   case Instruction::SExt: {
306     // Compute the bits in the result that are not present in the input.
307     unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
308 
309     APInt MaskIn = Mask.trunc(SrcBitWidth);
310     KnownZero = KnownZero.trunc(SrcBitWidth);
311     KnownOne = KnownOne.trunc(SrcBitWidth);
312     ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
313                       Depth+1);
314     assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
315     KnownZero = KnownZero.zext(BitWidth);
316     KnownOne = KnownOne.zext(BitWidth);
317 
318     // If the sign bit of the input is known set or clear, then we know the
319     // top bits of the result.
320     if (KnownZero[SrcBitWidth-1])             // Input sign bit known zero
321       KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
322     else if (KnownOne[SrcBitWidth-1])           // Input sign bit known set
323       KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
324     return;
325   }
326   case Instruction::Shl:
327     // (shl X, C1) & C2 == 0   iff   (X & C2 >>u C1) == 0
328     if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
329       uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
330       APInt Mask2(Mask.lshr(ShiftAmt));
331       ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
332                         Depth+1);
333       assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
334       KnownZero <<= ShiftAmt;
335       KnownOne  <<= ShiftAmt;
336       KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
337       return;
338     }
339     break;
340   case Instruction::LShr:
341     // (ushr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
342     if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
343       // Compute the new bits that are at the top now.
344       uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
345 
346       // Unsigned shift right.
347       APInt Mask2(Mask.shl(ShiftAmt));
348       ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
349                         Depth+1);
350       assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
351       KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
352       KnownOne  = APIntOps::lshr(KnownOne, ShiftAmt);
353       // high bits known zero.
354       KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
355       return;
356     }
357     break;
358   case Instruction::AShr:
359     // (ashr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
360     if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
361       // Compute the new bits that are at the top now.
362       uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
363 
364       // Signed shift right.
365       APInt Mask2(Mask.shl(ShiftAmt));
366       ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
367                         Depth+1);
368       assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
369       KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
370       KnownOne  = APIntOps::lshr(KnownOne, ShiftAmt);
371 
372       APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
373       if (KnownZero[BitWidth-ShiftAmt-1])    // New bits are known zero.
374         KnownZero |= HighBits;
375       else if (KnownOne[BitWidth-ShiftAmt-1])  // New bits are known one.
376         KnownOne |= HighBits;
377       return;
378     }
379     break;
380   case Instruction::Sub: {
381     if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
382       // We know that the top bits of C-X are clear if X contains less bits
383       // than C (i.e. no wrap-around can happen).  For example, 20-X is
384       // positive if we can prove that X is >= 0 and < 16.
385       if (!CLHS->getValue().isNegative()) {
386         unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
387         // NLZ can't be BitWidth with no sign bit
388         APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
389         ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
390                           TD, Depth+1);
391 
392         // If all of the MaskV bits are known to be zero, then we know the
393         // output top bits are zero, because we now know that the output is
394         // from [0-C].
395         if ((KnownZero2 & MaskV) == MaskV) {
396           unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
397           // Top bits known zero.
398           KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
399         }
400       }
401     }
402   }
403   // fall through
404   case Instruction::Add: {
405     // If one of the operands has trailing zeros, then the bits that the
406     // other operand has in those bit positions will be preserved in the
407     // result. For an add, this works with either operand. For a subtract,
408     // this only works if the known zeros are in the right operand.
409     APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
410     APInt Mask2 = APInt::getLowBitsSet(BitWidth,
411                                        BitWidth - Mask.countLeadingZeros());
412     ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
413                       Depth+1);
414     assert((LHSKnownZero & LHSKnownOne) == 0 &&
415            "Bits known to be one AND zero?");
416     unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
417 
418     ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD,
419                       Depth+1);
420     assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
421     unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
422 
423     // Determine which operand has more trailing zeros, and use that
424     // many bits from the other operand.
425     if (LHSKnownZeroOut > RHSKnownZeroOut) {
426       if (I->getOpcode() == Instruction::Add) {
427         APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
428         KnownZero |= KnownZero2 & Mask;
429         KnownOne  |= KnownOne2 & Mask;
430       } else {
431         // If the known zeros are in the left operand for a subtract,
432         // fall back to the minimum known zeros in both operands.
433         KnownZero |= APInt::getLowBitsSet(BitWidth,
434                                           std::min(LHSKnownZeroOut,
435                                                    RHSKnownZeroOut));
436       }
437     } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
438       APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
439       KnownZero |= LHSKnownZero & Mask;
440       KnownOne  |= LHSKnownOne & Mask;
441     }
442 
443     // Are we still trying to solve for the sign bit?
444     if (Mask.isNegative() && !KnownZero.isNegative() && !KnownOne.isNegative()){
445       OverflowingBinaryOperator *OBO = cast<OverflowingBinaryOperator>(I);
446       if (OBO->hasNoSignedWrap()) {
447         if (I->getOpcode() == Instruction::Add) {
448           // Adding two positive numbers can't wrap into negative
449           if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
450             KnownZero |= APInt::getSignBit(BitWidth);
451           // and adding two negative numbers can't wrap into positive.
452           else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
453             KnownOne |= APInt::getSignBit(BitWidth);
454         } else {
455           // Subtracting a negative number from a positive one can't wrap
456           if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
457             KnownZero |= APInt::getSignBit(BitWidth);
458           // neither can subtracting a positive number from a negative one.
459           else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
460             KnownOne |= APInt::getSignBit(BitWidth);
461         }
462       }
463     }
464 
465     return;
466   }
467   case Instruction::SRem:
468     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
469       APInt RA = Rem->getValue().abs();
470       if (RA.isPowerOf2()) {
471         APInt LowBits = RA - 1;
472         APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
473         ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
474                           Depth+1);
475 
476         // The low bits of the first operand are unchanged by the srem.
477         KnownZero = KnownZero2 & LowBits;
478         KnownOne = KnownOne2 & LowBits;
479 
480         // If the first operand is non-negative or has all low bits zero, then
481         // the upper bits are all zero.
482         if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
483           KnownZero |= ~LowBits;
484 
485         // If the first operand is negative and not all low bits are zero, then
486         // the upper bits are all one.
487         if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
488           KnownOne |= ~LowBits;
489 
490         KnownZero &= Mask;
491         KnownOne &= Mask;
492 
493         assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
494       }
495     }
496 
497     // The sign bit is the LHS's sign bit, except when the result of the
498     // remainder is zero.
499     if (Mask.isNegative() && KnownZero.isNonNegative()) {
500       APInt Mask2 = APInt::getSignBit(BitWidth);
501       APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
502       ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
503                         Depth+1);
504       // If it's known zero, our sign bit is also zero.
505       if (LHSKnownZero.isNegative())
506         KnownZero |= LHSKnownZero;
507     }
508 
509     break;
510   case Instruction::URem: {
511     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
512       APInt RA = Rem->getValue();
513       if (RA.isPowerOf2()) {
514         APInt LowBits = (RA - 1);
515         APInt Mask2 = LowBits & Mask;
516         KnownZero |= ~LowBits & Mask;
517         ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
518                           Depth+1);
519         assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
520         break;
521       }
522     }
523 
524     // Since the result is less than or equal to either operand, any leading
525     // zero bits in either operand must also exist in the result.
526     APInt AllOnes = APInt::getAllOnesValue(BitWidth);
527     ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
528                       TD, Depth+1);
529     ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
530                       TD, Depth+1);
531 
532     unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
533                                 KnownZero2.countLeadingOnes());
534     KnownOne.clearAllBits();
535     KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
536     break;
537   }
538 
539   case Instruction::Alloca: {
540     AllocaInst *AI = cast<AllocaInst>(V);
541     unsigned Align = AI->getAlignment();
542     if (Align == 0 && TD)
543       Align = TD->getABITypeAlignment(AI->getType()->getElementType());
544 
545     if (Align > 0)
546       KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
547                                               CountTrailingZeros_32(Align));
548     break;
549   }
550   case Instruction::GetElementPtr: {
551     // Analyze all of the subscripts of this getelementptr instruction
552     // to determine if we can prove known low zero bits.
553     APInt LocalMask = APInt::getAllOnesValue(BitWidth);
554     APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
555     ComputeMaskedBits(I->getOperand(0), LocalMask,
556                       LocalKnownZero, LocalKnownOne, TD, Depth+1);
557     unsigned TrailZ = LocalKnownZero.countTrailingOnes();
558 
559     gep_type_iterator GTI = gep_type_begin(I);
560     for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
561       Value *Index = I->getOperand(i);
562       if (StructType *STy = dyn_cast<StructType>(*GTI)) {
563         // Handle struct member offset arithmetic.
564         if (!TD) return;
565         const StructLayout *SL = TD->getStructLayout(STy);
566         unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
567         uint64_t Offset = SL->getElementOffset(Idx);
568         TrailZ = std::min(TrailZ,
569                           CountTrailingZeros_64(Offset));
570       } else {
571         // Handle array index arithmetic.
572         Type *IndexedTy = GTI.getIndexedType();
573         if (!IndexedTy->isSized()) return;
574         unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
575         uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
576         LocalMask = APInt::getAllOnesValue(GEPOpiBits);
577         LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
578         ComputeMaskedBits(Index, LocalMask,
579                           LocalKnownZero, LocalKnownOne, TD, Depth+1);
580         TrailZ = std::min(TrailZ,
581                           unsigned(CountTrailingZeros_64(TypeSize) +
582                                    LocalKnownZero.countTrailingOnes()));
583       }
584     }
585 
586     KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
587     break;
588   }
589   case Instruction::PHI: {
590     PHINode *P = cast<PHINode>(I);
591     // Handle the case of a simple two-predecessor recurrence PHI.
592     // There's a lot more that could theoretically be done here, but
593     // this is sufficient to catch some interesting cases.
594     if (P->getNumIncomingValues() == 2) {
595       for (unsigned i = 0; i != 2; ++i) {
596         Value *L = P->getIncomingValue(i);
597         Value *R = P->getIncomingValue(!i);
598         Operator *LU = dyn_cast<Operator>(L);
599         if (!LU)
600           continue;
601         unsigned Opcode = LU->getOpcode();
602         // Check for operations that have the property that if
603         // both their operands have low zero bits, the result
604         // will have low zero bits.
605         if (Opcode == Instruction::Add ||
606             Opcode == Instruction::Sub ||
607             Opcode == Instruction::And ||
608             Opcode == Instruction::Or ||
609             Opcode == Instruction::Mul) {
610           Value *LL = LU->getOperand(0);
611           Value *LR = LU->getOperand(1);
612           // Find a recurrence.
613           if (LL == I)
614             L = LR;
615           else if (LR == I)
616             L = LL;
617           else
618             break;
619           // Ok, we have a PHI of the form L op= R. Check for low
620           // zero bits.
621           APInt Mask2 = APInt::getAllOnesValue(BitWidth);
622           ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
623           Mask2 = APInt::getLowBitsSet(BitWidth,
624                                        KnownZero2.countTrailingOnes());
625 
626           // We need to take the minimum number of known bits
627           APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
628           ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
629 
630           KnownZero = Mask &
631                       APInt::getLowBitsSet(BitWidth,
632                                            std::min(KnownZero2.countTrailingOnes(),
633                                                     KnownZero3.countTrailingOnes()));
634           break;
635         }
636       }
637     }
638 
639     // Unreachable blocks may have zero-operand PHI nodes.
640     if (P->getNumIncomingValues() == 0)
641       return;
642 
643     // Otherwise take the unions of the known bit sets of the operands,
644     // taking conservative care to avoid excessive recursion.
645     if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
646       // Skip if every incoming value references to ourself.
647       if (P->hasConstantValue() == P)
648         break;
649 
650       KnownZero = APInt::getAllOnesValue(BitWidth);
651       KnownOne = APInt::getAllOnesValue(BitWidth);
652       for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
653         // Skip direct self references.
654         if (P->getIncomingValue(i) == P) continue;
655 
656         KnownZero2 = APInt(BitWidth, 0);
657         KnownOne2 = APInt(BitWidth, 0);
658         // Recurse, but cap the recursion to one level, because we don't
659         // want to waste time spinning around in loops.
660         ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
661                           KnownZero2, KnownOne2, TD, MaxDepth-1);
662         KnownZero &= KnownZero2;
663         KnownOne &= KnownOne2;
664         // If all bits have been ruled out, there's no need to check
665         // more operands.
666         if (!KnownZero && !KnownOne)
667           break;
668       }
669     }
670     break;
671   }
672   case Instruction::Call:
673     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
674       switch (II->getIntrinsicID()) {
675       default: break;
676       case Intrinsic::ctpop:
677       case Intrinsic::ctlz:
678       case Intrinsic::cttz: {
679         unsigned LowBits = Log2_32(BitWidth)+1;
680         KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
681         break;
682       }
683       case Intrinsic::x86_sse42_crc32_64_8:
684       case Intrinsic::x86_sse42_crc32_64_64:
685         KnownZero = APInt::getHighBitsSet(64, 32);
686         break;
687       }
688     }
689     break;
690   }
691 }
692 
693 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
694 /// one.  Convenience wrapper around ComputeMaskedBits.
ComputeSignBit(Value * V,bool & KnownZero,bool & KnownOne,const TargetData * TD,unsigned Depth)695 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
696                           const TargetData *TD, unsigned Depth) {
697   unsigned BitWidth = getBitWidth(V->getType(), TD);
698   if (!BitWidth) {
699     KnownZero = false;
700     KnownOne = false;
701     return;
702   }
703   APInt ZeroBits(BitWidth, 0);
704   APInt OneBits(BitWidth, 0);
705   ComputeMaskedBits(V, APInt::getSignBit(BitWidth), ZeroBits, OneBits, TD,
706                     Depth);
707   KnownOne = OneBits[BitWidth - 1];
708   KnownZero = ZeroBits[BitWidth - 1];
709 }
710 
711 /// isPowerOfTwo - Return true if the given value is known to have exactly one
712 /// bit set when defined. For vectors return true if every element is known to
713 /// be a power of two when defined.  Supports values with integer or pointer
714 /// types and vectors of integers.
isPowerOfTwo(Value * V,const TargetData * TD,unsigned Depth)715 bool llvm::isPowerOfTwo(Value *V, const TargetData *TD, unsigned Depth) {
716   if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
717     return CI->getValue().isPowerOf2();
718   // TODO: Handle vector constants.
719 
720   // 1 << X is clearly a power of two if the one is not shifted off the end.  If
721   // it is shifted off the end then the result is undefined.
722   if (match(V, m_Shl(m_One(), m_Value())))
723     return true;
724 
725   // (signbit) >>l X is clearly a power of two if the one is not shifted off the
726   // bottom.  If it is shifted off the bottom then the result is undefined.
727   if (match(V, m_LShr(m_SignBit(), m_Value())))
728     return true;
729 
730   // The remaining tests are all recursive, so bail out if we hit the limit.
731   if (Depth++ == MaxDepth)
732     return false;
733 
734   if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
735     return isPowerOfTwo(ZI->getOperand(0), TD, Depth);
736 
737   if (SelectInst *SI = dyn_cast<SelectInst>(V))
738     return isPowerOfTwo(SI->getTrueValue(), TD, Depth) &&
739       isPowerOfTwo(SI->getFalseValue(), TD, Depth);
740 
741   // An exact divide or right shift can only shift off zero bits, so the result
742   // is a power of two only if the first operand is a power of two and not
743   // copying a sign bit (sdiv int_min, 2).
744   if (match(V, m_LShr(m_Value(), m_Value())) ||
745       match(V, m_UDiv(m_Value(), m_Value()))) {
746     PossiblyExactOperator *PEO = cast<PossiblyExactOperator>(V);
747     if (PEO->isExact())
748       return isPowerOfTwo(PEO->getOperand(0), TD, Depth);
749   }
750 
751   return false;
752 }
753 
754 /// isKnownNonZero - Return true if the given value is known to be non-zero
755 /// when defined.  For vectors return true if every element is known to be
756 /// non-zero when defined.  Supports values with integer or pointer type and
757 /// vectors of integers.
isKnownNonZero(Value * V,const TargetData * TD,unsigned Depth)758 bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
759   if (Constant *C = dyn_cast<Constant>(V)) {
760     if (C->isNullValue())
761       return false;
762     if (isa<ConstantInt>(C))
763       // Must be non-zero due to null test above.
764       return true;
765     // TODO: Handle vectors
766     return false;
767   }
768 
769   // The remaining tests are all recursive, so bail out if we hit the limit.
770   if (Depth++ == MaxDepth)
771     return false;
772 
773   unsigned BitWidth = getBitWidth(V->getType(), TD);
774 
775   // X | Y != 0 if X != 0 or Y != 0.
776   Value *X = 0, *Y = 0;
777   if (match(V, m_Or(m_Value(X), m_Value(Y))))
778     return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
779 
780   // ext X != 0 if X != 0.
781   if (isa<SExtInst>(V) || isa<ZExtInst>(V))
782     return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
783 
784   // shl X, Y != 0 if X is odd.  Note that the value of the shift is undefined
785   // if the lowest bit is shifted off the end.
786   if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
787     // shl nuw can't remove any non-zero bits.
788     BinaryOperator *BO = cast<BinaryOperator>(V);
789     if (BO->hasNoUnsignedWrap())
790       return isKnownNonZero(X, TD, Depth);
791 
792     APInt KnownZero(BitWidth, 0);
793     APInt KnownOne(BitWidth, 0);
794     ComputeMaskedBits(X, APInt(BitWidth, 1), KnownZero, KnownOne, TD, Depth);
795     if (KnownOne[0])
796       return true;
797   }
798   // shr X, Y != 0 if X is negative.  Note that the value of the shift is not
799   // defined if the sign bit is shifted off the end.
800   else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
801     // shr exact can only shift out zero bits.
802     BinaryOperator *BO = cast<BinaryOperator>(V);
803     if (BO->isExact())
804       return isKnownNonZero(X, TD, Depth);
805 
806     bool XKnownNonNegative, XKnownNegative;
807     ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
808     if (XKnownNegative)
809       return true;
810   }
811   // div exact can only produce a zero if the dividend is zero.
812   else if (match(V, m_IDiv(m_Value(X), m_Value()))) {
813     BinaryOperator *BO = cast<BinaryOperator>(V);
814     if (BO->isExact())
815       return isKnownNonZero(X, TD, Depth);
816   }
817   // X + Y.
818   else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
819     bool XKnownNonNegative, XKnownNegative;
820     bool YKnownNonNegative, YKnownNegative;
821     ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
822     ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
823 
824     // If X and Y are both non-negative (as signed values) then their sum is not
825     // zero unless both X and Y are zero.
826     if (XKnownNonNegative && YKnownNonNegative)
827       if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
828         return true;
829 
830     // If X and Y are both negative (as signed values) then their sum is not
831     // zero unless both X and Y equal INT_MIN.
832     if (BitWidth && XKnownNegative && YKnownNegative) {
833       APInt KnownZero(BitWidth, 0);
834       APInt KnownOne(BitWidth, 0);
835       APInt Mask = APInt::getSignedMaxValue(BitWidth);
836       // The sign bit of X is set.  If some other bit is set then X is not equal
837       // to INT_MIN.
838       ComputeMaskedBits(X, Mask, KnownZero, KnownOne, TD, Depth);
839       if ((KnownOne & Mask) != 0)
840         return true;
841       // The sign bit of Y is set.  If some other bit is set then Y is not equal
842       // to INT_MIN.
843       ComputeMaskedBits(Y, Mask, KnownZero, KnownOne, TD, Depth);
844       if ((KnownOne & Mask) != 0)
845         return true;
846     }
847 
848     // The sum of a non-negative number and a power of two is not zero.
849     if (XKnownNonNegative && isPowerOfTwo(Y, TD, Depth))
850       return true;
851     if (YKnownNonNegative && isPowerOfTwo(X, TD, Depth))
852       return true;
853   }
854   // (C ? X : Y) != 0 if X != 0 and Y != 0.
855   else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
856     if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
857         isKnownNonZero(SI->getFalseValue(), TD, Depth))
858       return true;
859   }
860 
861   if (!BitWidth) return false;
862   APInt KnownZero(BitWidth, 0);
863   APInt KnownOne(BitWidth, 0);
864   ComputeMaskedBits(V, APInt::getAllOnesValue(BitWidth), KnownZero, KnownOne,
865                     TD, Depth);
866   return KnownOne != 0;
867 }
868 
869 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero.  We use
870 /// this predicate to simplify operations downstream.  Mask is known to be zero
871 /// for bits that V cannot have.
872 ///
873 /// This function is defined on values with integer type, values with pointer
874 /// type (but only if TD is non-null), and vectors of integers.  In the case
875 /// where V is a vector, the mask, known zero, and known one values are the
876 /// same width as the vector element, and the bit is set only if it is true
877 /// for all of the elements in the vector.
MaskedValueIsZero(Value * V,const APInt & Mask,const TargetData * TD,unsigned Depth)878 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
879                              const TargetData *TD, unsigned Depth) {
880   APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
881   ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
882   assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
883   return (KnownZero & Mask) == Mask;
884 }
885 
886 
887 
888 /// ComputeNumSignBits - Return the number of times the sign bit of the
889 /// register is replicated into the other bits.  We know that at least 1 bit
890 /// is always equal to the sign bit (itself), but other cases can give us
891 /// information.  For example, immediately after an "ashr X, 2", we know that
892 /// the top 3 bits are all equal to each other, so we return 3.
893 ///
894 /// 'Op' must have a scalar integer type.
895 ///
ComputeNumSignBits(Value * V,const TargetData * TD,unsigned Depth)896 unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
897                                   unsigned Depth) {
898   assert((TD || V->getType()->isIntOrIntVectorTy()) &&
899          "ComputeNumSignBits requires a TargetData object to operate "
900          "on non-integer values!");
901   Type *Ty = V->getType();
902   unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
903                          Ty->getScalarSizeInBits();
904   unsigned Tmp, Tmp2;
905   unsigned FirstAnswer = 1;
906 
907   // Note that ConstantInt is handled by the general ComputeMaskedBits case
908   // below.
909 
910   if (Depth == 6)
911     return 1;  // Limit search depth.
912 
913   Operator *U = dyn_cast<Operator>(V);
914   switch (Operator::getOpcode(V)) {
915   default: break;
916   case Instruction::SExt:
917     Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
918     return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
919 
920   case Instruction::AShr:
921     Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
922     // ashr X, C   -> adds C sign bits.
923     if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
924       Tmp += C->getZExtValue();
925       if (Tmp > TyBits) Tmp = TyBits;
926     }
927     // vector ashr X, <C, C, C, C>  -> adds C sign bits
928     if (ConstantVector *C = dyn_cast<ConstantVector>(U->getOperand(1))) {
929       if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) {
930         Tmp += CI->getZExtValue();
931         if (Tmp > TyBits) Tmp = TyBits;
932       }
933     }
934     return Tmp;
935   case Instruction::Shl:
936     if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
937       // shl destroys sign bits.
938       Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
939       if (C->getZExtValue() >= TyBits ||      // Bad shift.
940           C->getZExtValue() >= Tmp) break;    // Shifted all sign bits out.
941       return Tmp - C->getZExtValue();
942     }
943     break;
944   case Instruction::And:
945   case Instruction::Or:
946   case Instruction::Xor:    // NOT is handled here.
947     // Logical binary ops preserve the number of sign bits at the worst.
948     Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
949     if (Tmp != 1) {
950       Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
951       FirstAnswer = std::min(Tmp, Tmp2);
952       // We computed what we know about the sign bits as our first
953       // answer. Now proceed to the generic code that uses
954       // ComputeMaskedBits, and pick whichever answer is better.
955     }
956     break;
957 
958   case Instruction::Select:
959     Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
960     if (Tmp == 1) return 1;  // Early out.
961     Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
962     return std::min(Tmp, Tmp2);
963 
964   case Instruction::Add:
965     // Add can have at most one carry bit.  Thus we know that the output
966     // is, at worst, one more bit than the inputs.
967     Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
968     if (Tmp == 1) return 1;  // Early out.
969 
970     // Special case decrementing a value (ADD X, -1):
971     if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
972       if (CRHS->isAllOnesValue()) {
973         APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
974         APInt Mask = APInt::getAllOnesValue(TyBits);
975         ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
976                           Depth+1);
977 
978         // If the input is known to be 0 or 1, the output is 0/-1, which is all
979         // sign bits set.
980         if ((KnownZero | APInt(TyBits, 1)) == Mask)
981           return TyBits;
982 
983         // If we are subtracting one from a positive number, there is no carry
984         // out of the result.
985         if (KnownZero.isNegative())
986           return Tmp;
987       }
988 
989     Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
990     if (Tmp2 == 1) return 1;
991     return std::min(Tmp, Tmp2)-1;
992 
993   case Instruction::Sub:
994     Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
995     if (Tmp2 == 1) return 1;
996 
997     // Handle NEG.
998     if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
999       if (CLHS->isNullValue()) {
1000         APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1001         APInt Mask = APInt::getAllOnesValue(TyBits);
1002         ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
1003                           TD, Depth+1);
1004         // If the input is known to be 0 or 1, the output is 0/-1, which is all
1005         // sign bits set.
1006         if ((KnownZero | APInt(TyBits, 1)) == Mask)
1007           return TyBits;
1008 
1009         // If the input is known to be positive (the sign bit is known clear),
1010         // the output of the NEG has the same number of sign bits as the input.
1011         if (KnownZero.isNegative())
1012           return Tmp2;
1013 
1014         // Otherwise, we treat this like a SUB.
1015       }
1016 
1017     // Sub can have at most one carry bit.  Thus we know that the output
1018     // is, at worst, one more bit than the inputs.
1019     Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1020     if (Tmp == 1) return 1;  // Early out.
1021     return std::min(Tmp, Tmp2)-1;
1022 
1023   case Instruction::PHI: {
1024     PHINode *PN = cast<PHINode>(U);
1025     // Don't analyze large in-degree PHIs.
1026     if (PN->getNumIncomingValues() > 4) break;
1027 
1028     // Take the minimum of all incoming values.  This can't infinitely loop
1029     // because of our depth threshold.
1030     Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1031     for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1032       if (Tmp == 1) return Tmp;
1033       Tmp = std::min(Tmp,
1034                      ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1035     }
1036     return Tmp;
1037   }
1038 
1039   case Instruction::Trunc:
1040     // FIXME: it's tricky to do anything useful for this, but it is an important
1041     // case for targets like X86.
1042     break;
1043   }
1044 
1045   // Finally, if we can prove that the top bits of the result are 0's or 1's,
1046   // use this information.
1047   APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1048   APInt Mask = APInt::getAllOnesValue(TyBits);
1049   ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
1050 
1051   if (KnownZero.isNegative()) {        // sign bit is 0
1052     Mask = KnownZero;
1053   } else if (KnownOne.isNegative()) {  // sign bit is 1;
1054     Mask = KnownOne;
1055   } else {
1056     // Nothing known.
1057     return FirstAnswer;
1058   }
1059 
1060   // Okay, we know that the sign bit in Mask is set.  Use CLZ to determine
1061   // the number of identical bits in the top of the input value.
1062   Mask = ~Mask;
1063   Mask <<= Mask.getBitWidth()-TyBits;
1064   // Return # leading zeros.  We use 'min' here in case Val was zero before
1065   // shifting.  We don't want to return '64' as for an i32 "0".
1066   return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1067 }
1068 
1069 /// ComputeMultiple - This function computes the integer multiple of Base that
1070 /// equals V.  If successful, it returns true and returns the multiple in
1071 /// Multiple.  If unsuccessful, it returns false. It looks
1072 /// through SExt instructions only if LookThroughSExt is true.
ComputeMultiple(Value * V,unsigned Base,Value * & Multiple,bool LookThroughSExt,unsigned Depth)1073 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1074                            bool LookThroughSExt, unsigned Depth) {
1075   const unsigned MaxDepth = 6;
1076 
1077   assert(V && "No Value?");
1078   assert(Depth <= MaxDepth && "Limit Search Depth");
1079   assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1080 
1081   Type *T = V->getType();
1082 
1083   ConstantInt *CI = dyn_cast<ConstantInt>(V);
1084 
1085   if (Base == 0)
1086     return false;
1087 
1088   if (Base == 1) {
1089     Multiple = V;
1090     return true;
1091   }
1092 
1093   ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1094   Constant *BaseVal = ConstantInt::get(T, Base);
1095   if (CO && CO == BaseVal) {
1096     // Multiple is 1.
1097     Multiple = ConstantInt::get(T, 1);
1098     return true;
1099   }
1100 
1101   if (CI && CI->getZExtValue() % Base == 0) {
1102     Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1103     return true;
1104   }
1105 
1106   if (Depth == MaxDepth) return false;  // Limit search depth.
1107 
1108   Operator *I = dyn_cast<Operator>(V);
1109   if (!I) return false;
1110 
1111   switch (I->getOpcode()) {
1112   default: break;
1113   case Instruction::SExt:
1114     if (!LookThroughSExt) return false;
1115     // otherwise fall through to ZExt
1116   case Instruction::ZExt:
1117     return ComputeMultiple(I->getOperand(0), Base, Multiple,
1118                            LookThroughSExt, Depth+1);
1119   case Instruction::Shl:
1120   case Instruction::Mul: {
1121     Value *Op0 = I->getOperand(0);
1122     Value *Op1 = I->getOperand(1);
1123 
1124     if (I->getOpcode() == Instruction::Shl) {
1125       ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1126       if (!Op1CI) return false;
1127       // Turn Op0 << Op1 into Op0 * 2^Op1
1128       APInt Op1Int = Op1CI->getValue();
1129       uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1130       APInt API(Op1Int.getBitWidth(), 0);
1131       API.setBit(BitToSet);
1132       Op1 = ConstantInt::get(V->getContext(), API);
1133     }
1134 
1135     Value *Mul0 = NULL;
1136     if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1137       if (Constant *Op1C = dyn_cast<Constant>(Op1))
1138         if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1139           if (Op1C->getType()->getPrimitiveSizeInBits() <
1140               MulC->getType()->getPrimitiveSizeInBits())
1141             Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1142           if (Op1C->getType()->getPrimitiveSizeInBits() >
1143               MulC->getType()->getPrimitiveSizeInBits())
1144             MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1145 
1146           // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1147           Multiple = ConstantExpr::getMul(MulC, Op1C);
1148           return true;
1149         }
1150 
1151       if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1152         if (Mul0CI->getValue() == 1) {
1153           // V == Base * Op1, so return Op1
1154           Multiple = Op1;
1155           return true;
1156         }
1157     }
1158 
1159     Value *Mul1 = NULL;
1160     if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1161       if (Constant *Op0C = dyn_cast<Constant>(Op0))
1162         if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1163           if (Op0C->getType()->getPrimitiveSizeInBits() <
1164               MulC->getType()->getPrimitiveSizeInBits())
1165             Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1166           if (Op0C->getType()->getPrimitiveSizeInBits() >
1167               MulC->getType()->getPrimitiveSizeInBits())
1168             MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1169 
1170           // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1171           Multiple = ConstantExpr::getMul(MulC, Op0C);
1172           return true;
1173         }
1174 
1175       if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1176         if (Mul1CI->getValue() == 1) {
1177           // V == Base * Op0, so return Op0
1178           Multiple = Op0;
1179           return true;
1180         }
1181     }
1182   }
1183   }
1184 
1185   // We could not determine if V is a multiple of Base.
1186   return false;
1187 }
1188 
1189 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1190 /// value is never equal to -0.0.
1191 ///
1192 /// NOTE: this function will need to be revisited when we support non-default
1193 /// rounding modes!
1194 ///
CannotBeNegativeZero(const Value * V,unsigned Depth)1195 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1196   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1197     return !CFP->getValueAPF().isNegZero();
1198 
1199   if (Depth == 6)
1200     return 1;  // Limit search depth.
1201 
1202   const Operator *I = dyn_cast<Operator>(V);
1203   if (I == 0) return false;
1204 
1205   // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1206   if (I->getOpcode() == Instruction::FAdd &&
1207       isa<ConstantFP>(I->getOperand(1)) &&
1208       cast<ConstantFP>(I->getOperand(1))->isNullValue())
1209     return true;
1210 
1211   // sitofp and uitofp turn into +0.0 for zero.
1212   if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1213     return true;
1214 
1215   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1216     // sqrt(-0.0) = -0.0, no other negative results are possible.
1217     if (II->getIntrinsicID() == Intrinsic::sqrt)
1218       return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1219 
1220   if (const CallInst *CI = dyn_cast<CallInst>(I))
1221     if (const Function *F = CI->getCalledFunction()) {
1222       if (F->isDeclaration()) {
1223         // abs(x) != -0.0
1224         if (F->getName() == "abs") return true;
1225         // fabs[lf](x) != -0.0
1226         if (F->getName() == "fabs") return true;
1227         if (F->getName() == "fabsf") return true;
1228         if (F->getName() == "fabsl") return true;
1229         if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1230             F->getName() == "sqrtl")
1231           return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1232       }
1233     }
1234 
1235   return false;
1236 }
1237 
1238 /// isBytewiseValue - If the specified value can be set by repeating the same
1239 /// byte in memory, return the i8 value that it is represented with.  This is
1240 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1241 /// i16 0xF0F0, double 0.0 etc.  If the value can't be handled with a repeated
1242 /// byte store (e.g. i16 0x1234), return null.
isBytewiseValue(Value * V)1243 Value *llvm::isBytewiseValue(Value *V) {
1244   // All byte-wide stores are splatable, even of arbitrary variables.
1245   if (V->getType()->isIntegerTy(8)) return V;
1246 
1247   // Handle 'null' ConstantArrayZero etc.
1248   if (Constant *C = dyn_cast<Constant>(V))
1249     if (C->isNullValue())
1250       return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1251 
1252   // Constant float and double values can be handled as integer values if the
1253   // corresponding integer value is "byteable".  An important case is 0.0.
1254   if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1255     if (CFP->getType()->isFloatTy())
1256       V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1257     if (CFP->getType()->isDoubleTy())
1258       V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1259     // Don't handle long double formats, which have strange constraints.
1260   }
1261 
1262   // We can handle constant integers that are power of two in size and a
1263   // multiple of 8 bits.
1264   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1265     unsigned Width = CI->getBitWidth();
1266     if (isPowerOf2_32(Width) && Width > 8) {
1267       // We can handle this value if the recursive binary decomposition is the
1268       // same at all levels.
1269       APInt Val = CI->getValue();
1270       APInt Val2;
1271       while (Val.getBitWidth() != 8) {
1272         unsigned NextWidth = Val.getBitWidth()/2;
1273         Val2  = Val.lshr(NextWidth);
1274         Val2 = Val2.trunc(Val.getBitWidth()/2);
1275         Val = Val.trunc(Val.getBitWidth()/2);
1276 
1277         // If the top/bottom halves aren't the same, reject it.
1278         if (Val != Val2)
1279           return 0;
1280       }
1281       return ConstantInt::get(V->getContext(), Val);
1282     }
1283   }
1284 
1285   // A ConstantArray is splatable if all its members are equal and also
1286   // splatable.
1287   if (ConstantArray *CA = dyn_cast<ConstantArray>(V)) {
1288     if (CA->getNumOperands() == 0)
1289       return 0;
1290 
1291     Value *Val = isBytewiseValue(CA->getOperand(0));
1292     if (!Val)
1293       return 0;
1294 
1295     for (unsigned I = 1, E = CA->getNumOperands(); I != E; ++I)
1296       if (CA->getOperand(I-1) != CA->getOperand(I))
1297         return 0;
1298 
1299     return Val;
1300   }
1301 
1302   // Conceptually, we could handle things like:
1303   //   %a = zext i8 %X to i16
1304   //   %b = shl i16 %a, 8
1305   //   %c = or i16 %a, %b
1306   // but until there is an example that actually needs this, it doesn't seem
1307   // worth worrying about.
1308   return 0;
1309 }
1310 
1311 
1312 // This is the recursive version of BuildSubAggregate. It takes a few different
1313 // arguments. Idxs is the index within the nested struct From that we are
1314 // looking at now (which is of type IndexedType). IdxSkip is the number of
1315 // indices from Idxs that should be left out when inserting into the resulting
1316 // struct. To is the result struct built so far, new insertvalue instructions
1317 // build on that.
BuildSubAggregate(Value * From,Value * To,Type * IndexedType,SmallVector<unsigned,10> & Idxs,unsigned IdxSkip,Instruction * InsertBefore)1318 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1319                                 SmallVector<unsigned, 10> &Idxs,
1320                                 unsigned IdxSkip,
1321                                 Instruction *InsertBefore) {
1322   llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1323   if (STy) {
1324     // Save the original To argument so we can modify it
1325     Value *OrigTo = To;
1326     // General case, the type indexed by Idxs is a struct
1327     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1328       // Process each struct element recursively
1329       Idxs.push_back(i);
1330       Value *PrevTo = To;
1331       To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1332                              InsertBefore);
1333       Idxs.pop_back();
1334       if (!To) {
1335         // Couldn't find any inserted value for this index? Cleanup
1336         while (PrevTo != OrigTo) {
1337           InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1338           PrevTo = Del->getAggregateOperand();
1339           Del->eraseFromParent();
1340         }
1341         // Stop processing elements
1342         break;
1343       }
1344     }
1345     // If we successfully found a value for each of our subaggregates
1346     if (To)
1347       return To;
1348   }
1349   // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1350   // the struct's elements had a value that was inserted directly. In the latter
1351   // case, perhaps we can't determine each of the subelements individually, but
1352   // we might be able to find the complete struct somewhere.
1353 
1354   // Find the value that is at that particular spot
1355   Value *V = FindInsertedValue(From, Idxs);
1356 
1357   if (!V)
1358     return NULL;
1359 
1360   // Insert the value in the new (sub) aggregrate
1361   return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1362                                        "tmp", InsertBefore);
1363 }
1364 
1365 // This helper takes a nested struct and extracts a part of it (which is again a
1366 // struct) into a new value. For example, given the struct:
1367 // { a, { b, { c, d }, e } }
1368 // and the indices "1, 1" this returns
1369 // { c, d }.
1370 //
1371 // It does this by inserting an insertvalue for each element in the resulting
1372 // struct, as opposed to just inserting a single struct. This will only work if
1373 // each of the elements of the substruct are known (ie, inserted into From by an
1374 // insertvalue instruction somewhere).
1375 //
1376 // All inserted insertvalue instructions are inserted before InsertBefore
BuildSubAggregate(Value * From,ArrayRef<unsigned> idx_range,Instruction * InsertBefore)1377 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1378                                 Instruction *InsertBefore) {
1379   assert(InsertBefore && "Must have someplace to insert!");
1380   Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1381                                                              idx_range);
1382   Value *To = UndefValue::get(IndexedType);
1383   SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1384   unsigned IdxSkip = Idxs.size();
1385 
1386   return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1387 }
1388 
1389 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1390 /// the scalar value indexed is already around as a register, for example if it
1391 /// were inserted directly into the aggregrate.
1392 ///
1393 /// If InsertBefore is not null, this function will duplicate (modified)
1394 /// insertvalues when a part of a nested struct is extracted.
FindInsertedValue(Value * V,ArrayRef<unsigned> idx_range,Instruction * InsertBefore)1395 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1396                                Instruction *InsertBefore) {
1397   // Nothing to index? Just return V then (this is useful at the end of our
1398   // recursion)
1399   if (idx_range.empty())
1400     return V;
1401   // We have indices, so V should have an indexable type
1402   assert((V->getType()->isStructTy() || V->getType()->isArrayTy())
1403          && "Not looking at a struct or array?");
1404   assert(ExtractValueInst::getIndexedType(V->getType(), idx_range)
1405          && "Invalid indices for type?");
1406   CompositeType *PTy = cast<CompositeType>(V->getType());
1407 
1408   if (isa<UndefValue>(V))
1409     return UndefValue::get(ExtractValueInst::getIndexedType(PTy,
1410                                                               idx_range));
1411   else if (isa<ConstantAggregateZero>(V))
1412     return Constant::getNullValue(ExtractValueInst::getIndexedType(PTy,
1413                                                                   idx_range));
1414   else if (Constant *C = dyn_cast<Constant>(V)) {
1415     if (isa<ConstantArray>(C) || isa<ConstantStruct>(C))
1416       // Recursively process this constant
1417       return FindInsertedValue(C->getOperand(idx_range[0]), idx_range.slice(1),
1418                                InsertBefore);
1419   } else if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1420     // Loop the indices for the insertvalue instruction in parallel with the
1421     // requested indices
1422     const unsigned *req_idx = idx_range.begin();
1423     for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1424          i != e; ++i, ++req_idx) {
1425       if (req_idx == idx_range.end()) {
1426         if (InsertBefore)
1427           // The requested index identifies a part of a nested aggregate. Handle
1428           // this specially. For example,
1429           // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1430           // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1431           // %C = extractvalue {i32, { i32, i32 } } %B, 1
1432           // This can be changed into
1433           // %A = insertvalue {i32, i32 } undef, i32 10, 0
1434           // %C = insertvalue {i32, i32 } %A, i32 11, 1
1435           // which allows the unused 0,0 element from the nested struct to be
1436           // removed.
1437           return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1438                                    InsertBefore);
1439         else
1440           // We can't handle this without inserting insertvalues
1441           return 0;
1442       }
1443 
1444       // This insert value inserts something else than what we are looking for.
1445       // See if the (aggregrate) value inserted into has the value we are
1446       // looking for, then.
1447       if (*req_idx != *i)
1448         return FindInsertedValue(I->getAggregateOperand(), idx_range,
1449                                  InsertBefore);
1450     }
1451     // If we end up here, the indices of the insertvalue match with those
1452     // requested (though possibly only partially). Now we recursively look at
1453     // the inserted value, passing any remaining indices.
1454     return FindInsertedValue(I->getInsertedValueOperand(),
1455                              makeArrayRef(req_idx, idx_range.end()),
1456                              InsertBefore);
1457   } else if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1458     // If we're extracting a value from an aggregrate that was extracted from
1459     // something else, we can extract from that something else directly instead.
1460     // However, we will need to chain I's indices with the requested indices.
1461 
1462     // Calculate the number of indices required
1463     unsigned size = I->getNumIndices() + idx_range.size();
1464     // Allocate some space to put the new indices in
1465     SmallVector<unsigned, 5> Idxs;
1466     Idxs.reserve(size);
1467     // Add indices from the extract value instruction
1468     Idxs.append(I->idx_begin(), I->idx_end());
1469 
1470     // Add requested indices
1471     Idxs.append(idx_range.begin(), idx_range.end());
1472 
1473     assert(Idxs.size() == size
1474            && "Number of indices added not correct?");
1475 
1476     return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1477   }
1478   // Otherwise, we don't know (such as, extracting from a function return value
1479   // or load instruction)
1480   return 0;
1481 }
1482 
1483 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1484 /// it can be expressed as a base pointer plus a constant offset.  Return the
1485 /// base and offset to the caller.
GetPointerBaseWithConstantOffset(Value * Ptr,int64_t & Offset,const TargetData & TD)1486 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1487                                               const TargetData &TD) {
1488   Operator *PtrOp = dyn_cast<Operator>(Ptr);
1489   if (PtrOp == 0) return Ptr;
1490 
1491   // Just look through bitcasts.
1492   if (PtrOp->getOpcode() == Instruction::BitCast)
1493     return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1494 
1495   // If this is a GEP with constant indices, we can look through it.
1496   GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1497   if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1498 
1499   gep_type_iterator GTI = gep_type_begin(GEP);
1500   for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1501        ++I, ++GTI) {
1502     ConstantInt *OpC = cast<ConstantInt>(*I);
1503     if (OpC->isZero()) continue;
1504 
1505     // Handle a struct and array indices which add their offset to the pointer.
1506     if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1507       Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1508     } else {
1509       uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1510       Offset += OpC->getSExtValue()*Size;
1511     }
1512   }
1513 
1514   // Re-sign extend from the pointer size if needed to get overflow edge cases
1515   // right.
1516   unsigned PtrSize = TD.getPointerSizeInBits();
1517   if (PtrSize < 64)
1518     Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1519 
1520   return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1521 }
1522 
1523 
1524 /// GetConstantStringInfo - This function computes the length of a
1525 /// null-terminated C string pointed to by V.  If successful, it returns true
1526 /// and returns the string in Str.  If unsuccessful, it returns false.
GetConstantStringInfo(const Value * V,std::string & Str,uint64_t Offset,bool StopAtNul)1527 bool llvm::GetConstantStringInfo(const Value *V, std::string &Str,
1528                                  uint64_t Offset,
1529                                  bool StopAtNul) {
1530   // If V is NULL then return false;
1531   if (V == NULL) return false;
1532 
1533   // Look through bitcast instructions.
1534   if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1535     return GetConstantStringInfo(BCI->getOperand(0), Str, Offset, StopAtNul);
1536 
1537   // If the value is not a GEP instruction nor a constant expression with a
1538   // GEP instruction, then return false because ConstantArray can't occur
1539   // any other way
1540   const User *GEP = 0;
1541   if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1542     GEP = GEPI;
1543   } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1544     if (CE->getOpcode() == Instruction::BitCast)
1545       return GetConstantStringInfo(CE->getOperand(0), Str, Offset, StopAtNul);
1546     if (CE->getOpcode() != Instruction::GetElementPtr)
1547       return false;
1548     GEP = CE;
1549   }
1550 
1551   if (GEP) {
1552     // Make sure the GEP has exactly three arguments.
1553     if (GEP->getNumOperands() != 3)
1554       return false;
1555 
1556     // Make sure the index-ee is a pointer to array of i8.
1557     PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1558     ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1559     if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1560       return false;
1561 
1562     // Check to make sure that the first operand of the GEP is an integer and
1563     // has value 0 so that we are sure we're indexing into the initializer.
1564     const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1565     if (FirstIdx == 0 || !FirstIdx->isZero())
1566       return false;
1567 
1568     // If the second index isn't a ConstantInt, then this is a variable index
1569     // into the array.  If this occurs, we can't say anything meaningful about
1570     // the string.
1571     uint64_t StartIdx = 0;
1572     if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1573       StartIdx = CI->getZExtValue();
1574     else
1575       return false;
1576     return GetConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset,
1577                                  StopAtNul);
1578   }
1579 
1580   // The GEP instruction, constant or instruction, must reference a global
1581   // variable that is a constant and is initialized. The referenced constant
1582   // initializer is the array that we'll use for optimization.
1583   const GlobalVariable* GV = dyn_cast<GlobalVariable>(V);
1584   if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1585     return false;
1586   const Constant *GlobalInit = GV->getInitializer();
1587 
1588   // Handle the ConstantAggregateZero case
1589   if (isa<ConstantAggregateZero>(GlobalInit)) {
1590     // This is a degenerate case. The initializer is constant zero so the
1591     // length of the string must be zero.
1592     Str.clear();
1593     return true;
1594   }
1595 
1596   // Must be a Constant Array
1597   const ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1598   if (Array == 0 || !Array->getType()->getElementType()->isIntegerTy(8))
1599     return false;
1600 
1601   // Get the number of elements in the array
1602   uint64_t NumElts = Array->getType()->getNumElements();
1603 
1604   if (Offset > NumElts)
1605     return false;
1606 
1607   // Traverse the constant array from 'Offset' which is the place the GEP refers
1608   // to in the array.
1609   Str.reserve(NumElts-Offset);
1610   for (unsigned i = Offset; i != NumElts; ++i) {
1611     const Constant *Elt = Array->getOperand(i);
1612     const ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1613     if (!CI) // This array isn't suitable, non-int initializer.
1614       return false;
1615     if (StopAtNul && CI->isZero())
1616       return true; // we found end of string, success!
1617     Str += (char)CI->getZExtValue();
1618   }
1619 
1620   // The array isn't null terminated, but maybe this is a memcpy, not a strcpy.
1621   return true;
1622 }
1623 
1624 // These next two are very similar to the above, but also look through PHI
1625 // nodes.
1626 // TODO: See if we can integrate these two together.
1627 
1628 /// GetStringLengthH - If we can compute the length of the string pointed to by
1629 /// the specified pointer, return 'len+1'.  If we can't, return 0.
GetStringLengthH(Value * V,SmallPtrSet<PHINode *,32> & PHIs)1630 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1631   // Look through noop bitcast instructions.
1632   if (BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1633     return GetStringLengthH(BCI->getOperand(0), PHIs);
1634 
1635   // If this is a PHI node, there are two cases: either we have already seen it
1636   // or we haven't.
1637   if (PHINode *PN = dyn_cast<PHINode>(V)) {
1638     if (!PHIs.insert(PN))
1639       return ~0ULL;  // already in the set.
1640 
1641     // If it was new, see if all the input strings are the same length.
1642     uint64_t LenSoFar = ~0ULL;
1643     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1644       uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1645       if (Len == 0) return 0; // Unknown length -> unknown.
1646 
1647       if (Len == ~0ULL) continue;
1648 
1649       if (Len != LenSoFar && LenSoFar != ~0ULL)
1650         return 0;    // Disagree -> unknown.
1651       LenSoFar = Len;
1652     }
1653 
1654     // Success, all agree.
1655     return LenSoFar;
1656   }
1657 
1658   // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1659   if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1660     uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1661     if (Len1 == 0) return 0;
1662     uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1663     if (Len2 == 0) return 0;
1664     if (Len1 == ~0ULL) return Len2;
1665     if (Len2 == ~0ULL) return Len1;
1666     if (Len1 != Len2) return 0;
1667     return Len1;
1668   }
1669 
1670   // If the value is not a GEP instruction nor a constant expression with a
1671   // GEP instruction, then return unknown.
1672   User *GEP = 0;
1673   if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1674     GEP = GEPI;
1675   } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1676     if (CE->getOpcode() != Instruction::GetElementPtr)
1677       return 0;
1678     GEP = CE;
1679   } else {
1680     return 0;
1681   }
1682 
1683   // Make sure the GEP has exactly three arguments.
1684   if (GEP->getNumOperands() != 3)
1685     return 0;
1686 
1687   // Check to make sure that the first operand of the GEP is an integer and
1688   // has value 0 so that we are sure we're indexing into the initializer.
1689   if (ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(1))) {
1690     if (!Idx->isZero())
1691       return 0;
1692   } else
1693     return 0;
1694 
1695   // If the second index isn't a ConstantInt, then this is a variable index
1696   // into the array.  If this occurs, we can't say anything meaningful about
1697   // the string.
1698   uint64_t StartIdx = 0;
1699   if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1700     StartIdx = CI->getZExtValue();
1701   else
1702     return 0;
1703 
1704   // The GEP instruction, constant or instruction, must reference a global
1705   // variable that is a constant and is initialized. The referenced constant
1706   // initializer is the array that we'll use for optimization.
1707   GlobalVariable* GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
1708   if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
1709       GV->mayBeOverridden())
1710     return 0;
1711   Constant *GlobalInit = GV->getInitializer();
1712 
1713   // Handle the ConstantAggregateZero case, which is a degenerate case. The
1714   // initializer is constant zero so the length of the string must be zero.
1715   if (isa<ConstantAggregateZero>(GlobalInit))
1716     return 1;  // Len = 0 offset by 1.
1717 
1718   // Must be a Constant Array
1719   ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1720   if (!Array || !Array->getType()->getElementType()->isIntegerTy(8))
1721     return false;
1722 
1723   // Get the number of elements in the array
1724   uint64_t NumElts = Array->getType()->getNumElements();
1725 
1726   // Traverse the constant array from StartIdx (derived above) which is
1727   // the place the GEP refers to in the array.
1728   for (unsigned i = StartIdx; i != NumElts; ++i) {
1729     Constant *Elt = Array->getOperand(i);
1730     ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1731     if (!CI) // This array isn't suitable, non-int initializer.
1732       return 0;
1733     if (CI->isZero())
1734       return i-StartIdx+1; // We found end of string, success!
1735   }
1736 
1737   return 0; // The array isn't null terminated, conservatively return 'unknown'.
1738 }
1739 
1740 /// GetStringLength - If we can compute the length of the string pointed to by
1741 /// the specified pointer, return 'len+1'.  If we can't, return 0.
GetStringLength(Value * V)1742 uint64_t llvm::GetStringLength(Value *V) {
1743   if (!V->getType()->isPointerTy()) return 0;
1744 
1745   SmallPtrSet<PHINode*, 32> PHIs;
1746   uint64_t Len = GetStringLengthH(V, PHIs);
1747   // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1748   // an empty string as a length.
1749   return Len == ~0ULL ? 1 : Len;
1750 }
1751 
1752 Value *
GetUnderlyingObject(Value * V,const TargetData * TD,unsigned MaxLookup)1753 llvm::GetUnderlyingObject(Value *V, const TargetData *TD, unsigned MaxLookup) {
1754   if (!V->getType()->isPointerTy())
1755     return V;
1756   for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1757     if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1758       V = GEP->getPointerOperand();
1759     } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1760       V = cast<Operator>(V)->getOperand(0);
1761     } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1762       if (GA->mayBeOverridden())
1763         return V;
1764       V = GA->getAliasee();
1765     } else {
1766       // See if InstructionSimplify knows any relevant tricks.
1767       if (Instruction *I = dyn_cast<Instruction>(V))
1768         // TODO: Acquire a DominatorTree and use it.
1769         if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1770           V = Simplified;
1771           continue;
1772         }
1773 
1774       return V;
1775     }
1776     assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1777   }
1778   return V;
1779 }
1780 
1781 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1782 /// are lifetime markers.
1783 ///
onlyUsedByLifetimeMarkers(const Value * V)1784 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1785   for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1786        UI != UE; ++UI) {
1787     const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1788     if (!II) return false;
1789 
1790     if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1791         II->getIntrinsicID() != Intrinsic::lifetime_end)
1792       return false;
1793   }
1794   return true;
1795 }
1796