1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Constants.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/GlobalVariable.h"
20 #include "llvm/GlobalAlias.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/Operator.h"
24 #include "llvm/Target/TargetData.h"
25 #include "llvm/Support/GetElementPtrTypeIterator.h"
26 #include "llvm/Support/MathExtras.h"
27 #include "llvm/Support/PatternMatch.h"
28 #include "llvm/ADT/SmallPtrSet.h"
31 using namespace llvm::PatternMatch;
33 const unsigned MaxDepth = 6;
35 /// getBitWidth - Returns the bitwidth of the given scalar or pointer type (if
36 /// unknown returns 0). For vector types, returns the element type's bitwidth.
37 static unsigned getBitWidth(Type *Ty, const TargetData *TD) {
38 if (unsigned BitWidth = Ty->getScalarSizeInBits())
40 assert(isa<PointerType>(Ty) && "Expected a pointer type!");
41 return TD ? TD->getPointerSizeInBits() : 0;
44 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
45 /// known to be either zero or one and return them in the KnownZero/KnownOne
46 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
48 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
49 /// we cannot optimize based on the assumption that it is zero without changing
50 /// it to be an explicit zero. If we don't change it to zero, other code could
51 /// optimized based on the contradictory assumption that it is non-zero.
52 /// Because instcombine aggressively folds operations with undef args anyway,
53 /// this won't lose us code quality.
55 /// This function is defined on values with integer type, values with pointer
56 /// type (but only if TD is non-null), and vectors of integers. In the case
57 /// where V is a vector, the mask, known zero, and known one values are the
58 /// same width as the vector element, and the bit is set only if it is true
59 /// for all of the elements in the vector.
60 void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
61 APInt &KnownZero, APInt &KnownOne,
62 const TargetData *TD, unsigned Depth) {
63 assert(V && "No Value?");
64 assert(Depth <= MaxDepth && "Limit Search Depth");
65 unsigned BitWidth = Mask.getBitWidth();
66 assert((V->getType()->isIntOrIntVectorTy() ||
67 V->getType()->getScalarType()->isPointerTy()) &&
68 "Not integer or pointer type!");
70 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
71 (!V->getType()->isIntOrIntVectorTy() ||
72 V->getType()->getScalarSizeInBits() == BitWidth) &&
73 KnownZero.getBitWidth() == BitWidth &&
74 KnownOne.getBitWidth() == BitWidth &&
75 "V, Mask, KnownOne and KnownZero should have same BitWidth");
77 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
78 // We know all of the bits for a constant!
79 KnownOne = CI->getValue() & Mask;
80 KnownZero = ~KnownOne & Mask;
83 // Null and aggregate-zero are all-zeros.
84 if (isa<ConstantPointerNull>(V) ||
85 isa<ConstantAggregateZero>(V)) {
86 KnownOne.clearAllBits();
90 // Handle a constant vector by taking the intersection of the known bits of
92 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
93 KnownZero.setAllBits(); KnownOne.setAllBits();
94 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
95 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
96 ComputeMaskedBits(CV->getOperand(i), Mask, KnownZero2, KnownOne2,
98 KnownZero &= KnownZero2;
99 KnownOne &= KnownOne2;
103 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
104 // We know that CDS must be a vector of integers. Take the intersection of
106 KnownZero.setAllBits(); KnownOne.setAllBits();
107 APInt Elt(KnownZero.getBitWidth(), 0);
108 for (unsigned i = 0, e = CDS->getType()->getNumElements(); i != e; ++i) {
109 Elt = CDS->getElementAsInteger(i);
116 // The address of an aligned GlobalValue has trailing zeros.
117 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
118 unsigned Align = GV->getAlignment();
119 if (Align == 0 && TD && GV->getType()->getElementType()->isSized()) {
120 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
121 Type *ObjectType = GVar->getType()->getElementType();
122 // If the object is defined in the current Module, we'll be giving
123 // it the preferred alignment. Otherwise, we have to assume that it
124 // may only have the minimum ABI alignment.
125 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
126 Align = TD->getPreferredAlignment(GVar);
128 Align = TD->getABITypeAlignment(ObjectType);
132 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
133 CountTrailingZeros_32(Align));
135 KnownZero.clearAllBits();
136 KnownOne.clearAllBits();
139 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
140 // the bits of its aliasee.
141 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
142 if (GA->mayBeOverridden()) {
143 KnownZero.clearAllBits(); KnownOne.clearAllBits();
145 ComputeMaskedBits(GA->getAliasee(), Mask, KnownZero, KnownOne,
151 if (Argument *A = dyn_cast<Argument>(V)) {
152 // Get alignment information off byval arguments if specified in the IR.
153 if (A->hasByValAttr())
154 if (unsigned Align = A->getParamAlignment())
155 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
156 CountTrailingZeros_32(Align));
160 // Start out not knowing anything.
161 KnownZero.clearAllBits(); KnownOne.clearAllBits();
163 if (Depth == MaxDepth || Mask == 0)
164 return; // Limit search depth.
166 Operator *I = dyn_cast<Operator>(V);
169 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
170 switch (I->getOpcode()) {
172 case Instruction::And: {
173 // If either the LHS or the RHS are Zero, the result is zero.
174 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
175 APInt Mask2(Mask & ~KnownZero);
176 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
178 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
179 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
181 // Output known-1 bits are only known if set in both the LHS & RHS.
182 KnownOne &= KnownOne2;
183 // Output known-0 are known to be clear if zero in either the LHS | RHS.
184 KnownZero |= KnownZero2;
187 case Instruction::Or: {
188 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
189 APInt Mask2(Mask & ~KnownOne);
190 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
192 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
193 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
195 // Output known-0 bits are only known if clear in both the LHS & RHS.
196 KnownZero &= KnownZero2;
197 // Output known-1 are known to be set if set in either the LHS | RHS.
198 KnownOne |= KnownOne2;
201 case Instruction::Xor: {
202 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
203 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
205 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
206 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
208 // Output known-0 bits are known if clear or set in both the LHS & RHS.
209 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
210 // Output known-1 are known to be set if set in only one of the LHS, RHS.
211 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
212 KnownZero = KnownZeroOut;
215 case Instruction::Mul: {
216 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
217 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
218 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
220 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
221 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
223 bool isKnownNegative = false;
224 bool isKnownNonNegative = false;
225 // If the multiplication is known not to overflow, compute the sign bit.
226 if (Mask.isNegative() &&
227 cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap()) {
228 Value *Op1 = I->getOperand(1), *Op2 = I->getOperand(0);
230 // The product of a number with itself is non-negative.
231 isKnownNonNegative = true;
233 bool isKnownNonNegative1 = KnownZero.isNegative();
234 bool isKnownNonNegative2 = KnownZero2.isNegative();
235 bool isKnownNegative1 = KnownOne.isNegative();
236 bool isKnownNegative2 = KnownOne2.isNegative();
237 // The product of two numbers with the same sign is non-negative.
238 isKnownNonNegative = (isKnownNegative1 && isKnownNegative2) ||
239 (isKnownNonNegative1 && isKnownNonNegative2);
240 // The product of a negative number and a non-negative number is either
242 if (!isKnownNonNegative)
243 isKnownNegative = (isKnownNegative1 && isKnownNonNegative2 &&
244 isKnownNonZero(Op2, TD, Depth)) ||
245 (isKnownNegative2 && isKnownNonNegative1 &&
246 isKnownNonZero(Op1, TD, Depth));
250 // If low bits are zero in either operand, output low known-0 bits.
251 // Also compute a conserative estimate for high known-0 bits.
252 // More trickiness is possible, but this is sufficient for the
253 // interesting case of alignment computation.
254 KnownOne.clearAllBits();
255 unsigned TrailZ = KnownZero.countTrailingOnes() +
256 KnownZero2.countTrailingOnes();
257 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
258 KnownZero2.countLeadingOnes(),
259 BitWidth) - BitWidth;
261 TrailZ = std::min(TrailZ, BitWidth);
262 LeadZ = std::min(LeadZ, BitWidth);
263 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
264 APInt::getHighBitsSet(BitWidth, LeadZ);
267 // Only make use of no-wrap flags if we failed to compute the sign bit
268 // directly. This matters if the multiplication always overflows, in
269 // which case we prefer to follow the result of the direct computation,
270 // though as the program is invoking undefined behaviour we can choose
271 // whatever we like here.
272 if (isKnownNonNegative && !KnownOne.isNegative())
273 KnownZero.setBit(BitWidth - 1);
274 else if (isKnownNegative && !KnownZero.isNegative())
275 KnownOne.setBit(BitWidth - 1);
279 case Instruction::UDiv: {
280 // For the purposes of computing leading zeros we can conservatively
281 // treat a udiv as a logical right shift by the power of 2 known to
282 // be less than the denominator.
283 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
284 ComputeMaskedBits(I->getOperand(0),
285 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
286 unsigned LeadZ = KnownZero2.countLeadingOnes();
288 KnownOne2.clearAllBits();
289 KnownZero2.clearAllBits();
290 ComputeMaskedBits(I->getOperand(1),
291 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
292 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
293 if (RHSUnknownLeadingOnes != BitWidth)
294 LeadZ = std::min(BitWidth,
295 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
297 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
300 case Instruction::Select:
301 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
302 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
304 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
305 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
307 // Only known if known in both the LHS and RHS.
308 KnownOne &= KnownOne2;
309 KnownZero &= KnownZero2;
311 case Instruction::FPTrunc:
312 case Instruction::FPExt:
313 case Instruction::FPToUI:
314 case Instruction::FPToSI:
315 case Instruction::SIToFP:
316 case Instruction::UIToFP:
317 return; // Can't work with floating point.
318 case Instruction::PtrToInt:
319 case Instruction::IntToPtr:
320 // We can't handle these if we don't know the pointer size.
322 // FALL THROUGH and handle them the same as zext/trunc.
323 case Instruction::ZExt:
324 case Instruction::Trunc: {
325 Type *SrcTy = I->getOperand(0)->getType();
327 unsigned SrcBitWidth;
328 // Note that we handle pointer operands here because of inttoptr/ptrtoint
329 // which fall through here.
330 if (SrcTy->isPointerTy())
331 SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
333 SrcBitWidth = SrcTy->getScalarSizeInBits();
335 APInt MaskIn = Mask.zextOrTrunc(SrcBitWidth);
336 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
337 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
338 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
340 KnownZero = KnownZero.zextOrTrunc(BitWidth);
341 KnownOne = KnownOne.zextOrTrunc(BitWidth);
342 // Any top bits are known to be zero.
343 if (BitWidth > SrcBitWidth)
344 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
347 case Instruction::BitCast: {
348 Type *SrcTy = I->getOperand(0)->getType();
349 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
350 // TODO: For now, not handling conversions like:
351 // (bitcast i64 %x to <2 x i32>)
352 !I->getType()->isVectorTy()) {
353 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
359 case Instruction::SExt: {
360 // Compute the bits in the result that are not present in the input.
361 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
363 APInt MaskIn = Mask.trunc(SrcBitWidth);
364 KnownZero = KnownZero.trunc(SrcBitWidth);
365 KnownOne = KnownOne.trunc(SrcBitWidth);
366 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
368 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
369 KnownZero = KnownZero.zext(BitWidth);
370 KnownOne = KnownOne.zext(BitWidth);
372 // If the sign bit of the input is known set or clear, then we know the
373 // top bits of the result.
374 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
375 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
376 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
377 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
380 case Instruction::Shl:
381 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
382 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
383 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
384 APInt Mask2(Mask.lshr(ShiftAmt));
385 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
387 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
388 KnownZero <<= ShiftAmt;
389 KnownOne <<= ShiftAmt;
390 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
394 case Instruction::LShr:
395 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
396 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
397 // Compute the new bits that are at the top now.
398 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
400 // Unsigned shift right.
401 APInt Mask2(Mask.shl(ShiftAmt));
402 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
404 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
405 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
406 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
407 // high bits known zero.
408 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
412 case Instruction::AShr:
413 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
414 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
415 // Compute the new bits that are at the top now.
416 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
418 // Signed shift right.
419 APInt Mask2(Mask.shl(ShiftAmt));
420 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
422 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
423 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
424 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
426 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
427 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
428 KnownZero |= HighBits;
429 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
430 KnownOne |= HighBits;
434 case Instruction::Sub: {
435 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
436 // We know that the top bits of C-X are clear if X contains less bits
437 // than C (i.e. no wrap-around can happen). For example, 20-X is
438 // positive if we can prove that X is >= 0 and < 16.
439 if (!CLHS->getValue().isNegative()) {
440 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
441 // NLZ can't be BitWidth with no sign bit
442 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
443 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
446 // If all of the MaskV bits are known to be zero, then we know the
447 // output top bits are zero, because we now know that the output is
449 if ((KnownZero2 & MaskV) == MaskV) {
450 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
451 // Top bits known zero.
452 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
458 case Instruction::Add: {
459 // If one of the operands has trailing zeros, then the bits that the
460 // other operand has in those bit positions will be preserved in the
461 // result. For an add, this works with either operand. For a subtract,
462 // this only works if the known zeros are in the right operand.
463 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
464 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
465 BitWidth - Mask.countLeadingZeros());
466 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
468 assert((LHSKnownZero & LHSKnownOne) == 0 &&
469 "Bits known to be one AND zero?");
470 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
472 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD,
474 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
475 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
477 // Determine which operand has more trailing zeros, and use that
478 // many bits from the other operand.
479 if (LHSKnownZeroOut > RHSKnownZeroOut) {
480 if (I->getOpcode() == Instruction::Add) {
481 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
482 KnownZero |= KnownZero2 & Mask;
483 KnownOne |= KnownOne2 & Mask;
485 // If the known zeros are in the left operand for a subtract,
486 // fall back to the minimum known zeros in both operands.
487 KnownZero |= APInt::getLowBitsSet(BitWidth,
488 std::min(LHSKnownZeroOut,
491 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
492 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
493 KnownZero |= LHSKnownZero & Mask;
494 KnownOne |= LHSKnownOne & Mask;
497 // Are we still trying to solve for the sign bit?
498 if (Mask.isNegative() && !KnownZero.isNegative() && !KnownOne.isNegative()){
499 OverflowingBinaryOperator *OBO = cast<OverflowingBinaryOperator>(I);
500 if (OBO->hasNoSignedWrap()) {
501 if (I->getOpcode() == Instruction::Add) {
502 // Adding two positive numbers can't wrap into negative
503 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
504 KnownZero |= APInt::getSignBit(BitWidth);
505 // and adding two negative numbers can't wrap into positive.
506 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
507 KnownOne |= APInt::getSignBit(BitWidth);
509 // Subtracting a negative number from a positive one can't wrap
510 if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
511 KnownZero |= APInt::getSignBit(BitWidth);
512 // neither can subtracting a positive number from a negative one.
513 else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
514 KnownOne |= APInt::getSignBit(BitWidth);
521 case Instruction::SRem:
522 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
523 APInt RA = Rem->getValue().abs();
524 if (RA.isPowerOf2()) {
525 APInt LowBits = RA - 1;
526 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
527 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
530 // The low bits of the first operand are unchanged by the srem.
531 KnownZero = KnownZero2 & LowBits;
532 KnownOne = KnownOne2 & LowBits;
534 // If the first operand is non-negative or has all low bits zero, then
535 // the upper bits are all zero.
536 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
537 KnownZero |= ~LowBits;
539 // If the first operand is negative and not all low bits are zero, then
540 // the upper bits are all one.
541 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
542 KnownOne |= ~LowBits;
547 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
551 // The sign bit is the LHS's sign bit, except when the result of the
552 // remainder is zero.
553 if (Mask.isNegative() && KnownZero.isNonNegative()) {
554 APInt Mask2 = APInt::getSignBit(BitWidth);
555 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
556 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
558 // If it's known zero, our sign bit is also zero.
559 if (LHSKnownZero.isNegative())
560 KnownZero |= LHSKnownZero;
564 case Instruction::URem: {
565 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
566 APInt RA = Rem->getValue();
567 if (RA.isPowerOf2()) {
568 APInt LowBits = (RA - 1);
569 APInt Mask2 = LowBits & Mask;
570 KnownZero |= ~LowBits & Mask;
571 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
573 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
578 // Since the result is less than or equal to either operand, any leading
579 // zero bits in either operand must also exist in the result.
580 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
581 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
583 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
586 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
587 KnownZero2.countLeadingOnes());
588 KnownOne.clearAllBits();
589 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
593 case Instruction::Alloca: {
594 AllocaInst *AI = cast<AllocaInst>(V);
595 unsigned Align = AI->getAlignment();
596 if (Align == 0 && TD)
597 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
600 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
601 CountTrailingZeros_32(Align));
604 case Instruction::GetElementPtr: {
605 // Analyze all of the subscripts of this getelementptr instruction
606 // to determine if we can prove known low zero bits.
607 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
608 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
609 ComputeMaskedBits(I->getOperand(0), LocalMask,
610 LocalKnownZero, LocalKnownOne, TD, Depth+1);
611 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
613 gep_type_iterator GTI = gep_type_begin(I);
614 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
615 Value *Index = I->getOperand(i);
616 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
617 // Handle struct member offset arithmetic.
619 const StructLayout *SL = TD->getStructLayout(STy);
620 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
621 uint64_t Offset = SL->getElementOffset(Idx);
622 TrailZ = std::min(TrailZ,
623 CountTrailingZeros_64(Offset));
625 // Handle array index arithmetic.
626 Type *IndexedTy = GTI.getIndexedType();
627 if (!IndexedTy->isSized()) return;
628 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
629 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
630 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
631 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
632 ComputeMaskedBits(Index, LocalMask,
633 LocalKnownZero, LocalKnownOne, TD, Depth+1);
634 TrailZ = std::min(TrailZ,
635 unsigned(CountTrailingZeros_64(TypeSize) +
636 LocalKnownZero.countTrailingOnes()));
640 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
643 case Instruction::PHI: {
644 PHINode *P = cast<PHINode>(I);
645 // Handle the case of a simple two-predecessor recurrence PHI.
646 // There's a lot more that could theoretically be done here, but
647 // this is sufficient to catch some interesting cases.
648 if (P->getNumIncomingValues() == 2) {
649 for (unsigned i = 0; i != 2; ++i) {
650 Value *L = P->getIncomingValue(i);
651 Value *R = P->getIncomingValue(!i);
652 Operator *LU = dyn_cast<Operator>(L);
655 unsigned Opcode = LU->getOpcode();
656 // Check for operations that have the property that if
657 // both their operands have low zero bits, the result
658 // will have low zero bits.
659 if (Opcode == Instruction::Add ||
660 Opcode == Instruction::Sub ||
661 Opcode == Instruction::And ||
662 Opcode == Instruction::Or ||
663 Opcode == Instruction::Mul) {
664 Value *LL = LU->getOperand(0);
665 Value *LR = LU->getOperand(1);
666 // Find a recurrence.
673 // Ok, we have a PHI of the form L op= R. Check for low
675 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
676 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
677 Mask2 = APInt::getLowBitsSet(BitWidth,
678 KnownZero2.countTrailingOnes());
680 // We need to take the minimum number of known bits
681 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
682 ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
685 APInt::getLowBitsSet(BitWidth,
686 std::min(KnownZero2.countTrailingOnes(),
687 KnownZero3.countTrailingOnes()));
693 // Unreachable blocks may have zero-operand PHI nodes.
694 if (P->getNumIncomingValues() == 0)
697 // Otherwise take the unions of the known bit sets of the operands,
698 // taking conservative care to avoid excessive recursion.
699 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
700 // Skip if every incoming value references to ourself.
701 if (P->hasConstantValue() == P)
704 KnownZero = APInt::getAllOnesValue(BitWidth);
705 KnownOne = APInt::getAllOnesValue(BitWidth);
706 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
707 // Skip direct self references.
708 if (P->getIncomingValue(i) == P) continue;
710 KnownZero2 = APInt(BitWidth, 0);
711 KnownOne2 = APInt(BitWidth, 0);
712 // Recurse, but cap the recursion to one level, because we don't
713 // want to waste time spinning around in loops.
714 ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
715 KnownZero2, KnownOne2, TD, MaxDepth-1);
716 KnownZero &= KnownZero2;
717 KnownOne &= KnownOne2;
718 // If all bits have been ruled out, there's no need to check
720 if (!KnownZero && !KnownOne)
726 case Instruction::Call:
727 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
728 switch (II->getIntrinsicID()) {
730 case Intrinsic::ctlz:
731 case Intrinsic::cttz: {
732 unsigned LowBits = Log2_32(BitWidth)+1;
733 // If this call is undefined for 0, the result will be less than 2^n.
734 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
736 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
739 case Intrinsic::ctpop: {
740 unsigned LowBits = Log2_32(BitWidth)+1;
741 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
744 case Intrinsic::x86_sse42_crc32_64_8:
745 case Intrinsic::x86_sse42_crc32_64_64:
746 KnownZero = APInt::getHighBitsSet(64, 32);
754 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
755 /// one. Convenience wrapper around ComputeMaskedBits.
756 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
757 const TargetData *TD, unsigned Depth) {
758 unsigned BitWidth = getBitWidth(V->getType(), TD);
764 APInt ZeroBits(BitWidth, 0);
765 APInt OneBits(BitWidth, 0);
766 ComputeMaskedBits(V, APInt::getSignBit(BitWidth), ZeroBits, OneBits, TD,
768 KnownOne = OneBits[BitWidth - 1];
769 KnownZero = ZeroBits[BitWidth - 1];
772 /// isPowerOfTwo - Return true if the given value is known to have exactly one
773 /// bit set when defined. For vectors return true if every element is known to
774 /// be a power of two when defined. Supports values with integer or pointer
775 /// types and vectors of integers.
776 bool llvm::isPowerOfTwo(Value *V, const TargetData *TD, bool OrZero,
778 if (Constant *C = dyn_cast<Constant>(V)) {
779 if (C->isNullValue())
781 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
782 return CI->getValue().isPowerOf2();
783 // TODO: Handle vector constants.
786 // 1 << X is clearly a power of two if the one is not shifted off the end. If
787 // it is shifted off the end then the result is undefined.
788 if (match(V, m_Shl(m_One(), m_Value())))
791 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
792 // bottom. If it is shifted off the bottom then the result is undefined.
793 if (match(V, m_LShr(m_SignBit(), m_Value())))
796 // The remaining tests are all recursive, so bail out if we hit the limit.
797 if (Depth++ == MaxDepth)
800 Value *X = 0, *Y = 0;
801 // A shift of a power of two is a power of two or zero.
802 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
803 match(V, m_Shr(m_Value(X), m_Value()))))
804 return isPowerOfTwo(X, TD, /*OrZero*/true, Depth);
806 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
807 return isPowerOfTwo(ZI->getOperand(0), TD, OrZero, Depth);
809 if (SelectInst *SI = dyn_cast<SelectInst>(V))
810 return isPowerOfTwo(SI->getTrueValue(), TD, OrZero, Depth) &&
811 isPowerOfTwo(SI->getFalseValue(), TD, OrZero, Depth);
813 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
814 // A power of two and'd with anything is a power of two or zero.
815 if (isPowerOfTwo(X, TD, /*OrZero*/true, Depth) ||
816 isPowerOfTwo(Y, TD, /*OrZero*/true, Depth))
818 // X & (-X) is always a power of two or zero.
819 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
824 // An exact divide or right shift can only shift off zero bits, so the result
825 // is a power of two only if the first operand is a power of two and not
826 // copying a sign bit (sdiv int_min, 2).
827 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
828 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
829 return isPowerOfTwo(cast<Operator>(V)->getOperand(0), TD, OrZero, Depth);
835 /// isKnownNonZero - Return true if the given value is known to be non-zero
836 /// when defined. For vectors return true if every element is known to be
837 /// non-zero when defined. Supports values with integer or pointer type and
838 /// vectors of integers.
839 bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
840 if (Constant *C = dyn_cast<Constant>(V)) {
841 if (C->isNullValue())
843 if (isa<ConstantInt>(C))
844 // Must be non-zero due to null test above.
846 // TODO: Handle vectors
850 // The remaining tests are all recursive, so bail out if we hit the limit.
851 if (Depth++ >= MaxDepth)
854 unsigned BitWidth = getBitWidth(V->getType(), TD);
856 // X | Y != 0 if X != 0 or Y != 0.
857 Value *X = 0, *Y = 0;
858 if (match(V, m_Or(m_Value(X), m_Value(Y))))
859 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
861 // ext X != 0 if X != 0.
862 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
863 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
865 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
866 // if the lowest bit is shifted off the end.
867 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
868 // shl nuw can't remove any non-zero bits.
869 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
870 if (BO->hasNoUnsignedWrap())
871 return isKnownNonZero(X, TD, Depth);
873 APInt KnownZero(BitWidth, 0);
874 APInt KnownOne(BitWidth, 0);
875 ComputeMaskedBits(X, APInt(BitWidth, 1), KnownZero, KnownOne, TD, Depth);
879 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
880 // defined if the sign bit is shifted off the end.
881 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
882 // shr exact can only shift out zero bits.
883 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
885 return isKnownNonZero(X, TD, Depth);
887 bool XKnownNonNegative, XKnownNegative;
888 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
892 // div exact can only produce a zero if the dividend is zero.
893 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
894 return isKnownNonZero(X, TD, Depth);
897 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
898 bool XKnownNonNegative, XKnownNegative;
899 bool YKnownNonNegative, YKnownNegative;
900 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
901 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
903 // If X and Y are both non-negative (as signed values) then their sum is not
904 // zero unless both X and Y are zero.
905 if (XKnownNonNegative && YKnownNonNegative)
906 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
909 // If X and Y are both negative (as signed values) then their sum is not
910 // zero unless both X and Y equal INT_MIN.
911 if (BitWidth && XKnownNegative && YKnownNegative) {
912 APInt KnownZero(BitWidth, 0);
913 APInt KnownOne(BitWidth, 0);
914 APInt Mask = APInt::getSignedMaxValue(BitWidth);
915 // The sign bit of X is set. If some other bit is set then X is not equal
917 ComputeMaskedBits(X, Mask, KnownZero, KnownOne, TD, Depth);
918 if ((KnownOne & Mask) != 0)
920 // The sign bit of Y is set. If some other bit is set then Y is not equal
922 ComputeMaskedBits(Y, Mask, KnownZero, KnownOne, TD, Depth);
923 if ((KnownOne & Mask) != 0)
927 // The sum of a non-negative number and a power of two is not zero.
928 if (XKnownNonNegative && isPowerOfTwo(Y, TD, /*OrZero*/false, Depth))
930 if (YKnownNonNegative && isPowerOfTwo(X, TD, /*OrZero*/false, Depth))
934 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
935 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
936 // If X and Y are non-zero then so is X * Y as long as the multiplication
937 // does not overflow.
938 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
939 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
942 // (C ? X : Y) != 0 if X != 0 and Y != 0.
943 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
944 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
945 isKnownNonZero(SI->getFalseValue(), TD, Depth))
949 if (!BitWidth) return false;
950 APInt KnownZero(BitWidth, 0);
951 APInt KnownOne(BitWidth, 0);
952 ComputeMaskedBits(V, APInt::getAllOnesValue(BitWidth), KnownZero, KnownOne,
954 return KnownOne != 0;
957 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
958 /// this predicate to simplify operations downstream. Mask is known to be zero
959 /// for bits that V cannot have.
961 /// This function is defined on values with integer type, values with pointer
962 /// type (but only if TD is non-null), and vectors of integers. In the case
963 /// where V is a vector, the mask, known zero, and known one values are the
964 /// same width as the vector element, and the bit is set only if it is true
965 /// for all of the elements in the vector.
966 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
967 const TargetData *TD, unsigned Depth) {
968 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
969 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
970 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
971 return (KnownZero & Mask) == Mask;
976 /// ComputeNumSignBits - Return the number of times the sign bit of the
977 /// register is replicated into the other bits. We know that at least 1 bit
978 /// is always equal to the sign bit (itself), but other cases can give us
979 /// information. For example, immediately after an "ashr X, 2", we know that
980 /// the top 3 bits are all equal to each other, so we return 3.
982 /// 'Op' must have a scalar integer type.
984 unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
986 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
987 "ComputeNumSignBits requires a TargetData object to operate "
988 "on non-integer values!");
989 Type *Ty = V->getType();
990 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
991 Ty->getScalarSizeInBits();
993 unsigned FirstAnswer = 1;
995 // Note that ConstantInt is handled by the general ComputeMaskedBits case
999 return 1; // Limit search depth.
1001 Operator *U = dyn_cast<Operator>(V);
1002 switch (Operator::getOpcode(V)) {
1004 case Instruction::SExt:
1005 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1006 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
1008 case Instruction::AShr:
1009 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1010 // ashr X, C -> adds C sign bits.
1011 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
1012 Tmp += C->getZExtValue();
1013 if (Tmp > TyBits) Tmp = TyBits;
1015 // vector ashr X, <C, C, C, C> -> adds C sign bits
1016 if (ConstantVector *C = dyn_cast<ConstantVector>(U->getOperand(1))) {
1017 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) {
1018 Tmp += CI->getZExtValue();
1019 if (Tmp > TyBits) Tmp = TyBits;
1023 case Instruction::Shl:
1024 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
1025 // shl destroys sign bits.
1026 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1027 if (C->getZExtValue() >= TyBits || // Bad shift.
1028 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
1029 return Tmp - C->getZExtValue();
1032 case Instruction::And:
1033 case Instruction::Or:
1034 case Instruction::Xor: // NOT is handled here.
1035 // Logical binary ops preserve the number of sign bits at the worst.
1036 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1038 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1039 FirstAnswer = std::min(Tmp, Tmp2);
1040 // We computed what we know about the sign bits as our first
1041 // answer. Now proceed to the generic code that uses
1042 // ComputeMaskedBits, and pick whichever answer is better.
1046 case Instruction::Select:
1047 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1048 if (Tmp == 1) return 1; // Early out.
1049 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1050 return std::min(Tmp, Tmp2);
1052 case Instruction::Add:
1053 // Add can have at most one carry bit. Thus we know that the output
1054 // is, at worst, one more bit than the inputs.
1055 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1056 if (Tmp == 1) return 1; // Early out.
1058 // Special case decrementing a value (ADD X, -1):
1059 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1060 if (CRHS->isAllOnesValue()) {
1061 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1062 APInt Mask = APInt::getAllOnesValue(TyBits);
1063 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
1066 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1068 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1071 // If we are subtracting one from a positive number, there is no carry
1072 // out of the result.
1073 if (KnownZero.isNegative())
1077 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1078 if (Tmp2 == 1) return 1;
1079 return std::min(Tmp, Tmp2)-1;
1081 case Instruction::Sub:
1082 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1083 if (Tmp2 == 1) return 1;
1086 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1087 if (CLHS->isNullValue()) {
1088 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1089 APInt Mask = APInt::getAllOnesValue(TyBits);
1090 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
1092 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1094 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1097 // If the input is known to be positive (the sign bit is known clear),
1098 // the output of the NEG has the same number of sign bits as the input.
1099 if (KnownZero.isNegative())
1102 // Otherwise, we treat this like a SUB.
1105 // Sub can have at most one carry bit. Thus we know that the output
1106 // is, at worst, one more bit than the inputs.
1107 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1108 if (Tmp == 1) return 1; // Early out.
1109 return std::min(Tmp, Tmp2)-1;
1111 case Instruction::PHI: {
1112 PHINode *PN = cast<PHINode>(U);
1113 // Don't analyze large in-degree PHIs.
1114 if (PN->getNumIncomingValues() > 4) break;
1116 // Take the minimum of all incoming values. This can't infinitely loop
1117 // because of our depth threshold.
1118 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1119 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1120 if (Tmp == 1) return Tmp;
1122 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1127 case Instruction::Trunc:
1128 // FIXME: it's tricky to do anything useful for this, but it is an important
1129 // case for targets like X86.
1133 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1134 // use this information.
1135 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1136 APInt Mask = APInt::getAllOnesValue(TyBits);
1137 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
1139 if (KnownZero.isNegative()) { // sign bit is 0
1141 } else if (KnownOne.isNegative()) { // sign bit is 1;
1148 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1149 // the number of identical bits in the top of the input value.
1151 Mask <<= Mask.getBitWidth()-TyBits;
1152 // Return # leading zeros. We use 'min' here in case Val was zero before
1153 // shifting. We don't want to return '64' as for an i32 "0".
1154 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1157 /// ComputeMultiple - This function computes the integer multiple of Base that
1158 /// equals V. If successful, it returns true and returns the multiple in
1159 /// Multiple. If unsuccessful, it returns false. It looks
1160 /// through SExt instructions only if LookThroughSExt is true.
1161 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1162 bool LookThroughSExt, unsigned Depth) {
1163 const unsigned MaxDepth = 6;
1165 assert(V && "No Value?");
1166 assert(Depth <= MaxDepth && "Limit Search Depth");
1167 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1169 Type *T = V->getType();
1171 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1181 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1182 Constant *BaseVal = ConstantInt::get(T, Base);
1183 if (CO && CO == BaseVal) {
1185 Multiple = ConstantInt::get(T, 1);
1189 if (CI && CI->getZExtValue() % Base == 0) {
1190 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1194 if (Depth == MaxDepth) return false; // Limit search depth.
1196 Operator *I = dyn_cast<Operator>(V);
1197 if (!I) return false;
1199 switch (I->getOpcode()) {
1201 case Instruction::SExt:
1202 if (!LookThroughSExt) return false;
1203 // otherwise fall through to ZExt
1204 case Instruction::ZExt:
1205 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1206 LookThroughSExt, Depth+1);
1207 case Instruction::Shl:
1208 case Instruction::Mul: {
1209 Value *Op0 = I->getOperand(0);
1210 Value *Op1 = I->getOperand(1);
1212 if (I->getOpcode() == Instruction::Shl) {
1213 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1214 if (!Op1CI) return false;
1215 // Turn Op0 << Op1 into Op0 * 2^Op1
1216 APInt Op1Int = Op1CI->getValue();
1217 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1218 APInt API(Op1Int.getBitWidth(), 0);
1219 API.setBit(BitToSet);
1220 Op1 = ConstantInt::get(V->getContext(), API);
1224 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1225 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1226 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1227 if (Op1C->getType()->getPrimitiveSizeInBits() <
1228 MulC->getType()->getPrimitiveSizeInBits())
1229 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1230 if (Op1C->getType()->getPrimitiveSizeInBits() >
1231 MulC->getType()->getPrimitiveSizeInBits())
1232 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1234 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1235 Multiple = ConstantExpr::getMul(MulC, Op1C);
1239 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1240 if (Mul0CI->getValue() == 1) {
1241 // V == Base * Op1, so return Op1
1248 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1249 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1250 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1251 if (Op0C->getType()->getPrimitiveSizeInBits() <
1252 MulC->getType()->getPrimitiveSizeInBits())
1253 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1254 if (Op0C->getType()->getPrimitiveSizeInBits() >
1255 MulC->getType()->getPrimitiveSizeInBits())
1256 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1258 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1259 Multiple = ConstantExpr::getMul(MulC, Op0C);
1263 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1264 if (Mul1CI->getValue() == 1) {
1265 // V == Base * Op0, so return Op0
1273 // We could not determine if V is a multiple of Base.
1277 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1278 /// value is never equal to -0.0.
1280 /// NOTE: this function will need to be revisited when we support non-default
1283 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1284 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1285 return !CFP->getValueAPF().isNegZero();
1288 return 1; // Limit search depth.
1290 const Operator *I = dyn_cast<Operator>(V);
1291 if (I == 0) return false;
1293 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1294 if (I->getOpcode() == Instruction::FAdd &&
1295 isa<ConstantFP>(I->getOperand(1)) &&
1296 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1299 // sitofp and uitofp turn into +0.0 for zero.
1300 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1303 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1304 // sqrt(-0.0) = -0.0, no other negative results are possible.
1305 if (II->getIntrinsicID() == Intrinsic::sqrt)
1306 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1308 if (const CallInst *CI = dyn_cast<CallInst>(I))
1309 if (const Function *F = CI->getCalledFunction()) {
1310 if (F->isDeclaration()) {
1312 if (F->getName() == "abs") return true;
1313 // fabs[lf](x) != -0.0
1314 if (F->getName() == "fabs") return true;
1315 if (F->getName() == "fabsf") return true;
1316 if (F->getName() == "fabsl") return true;
1317 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1318 F->getName() == "sqrtl")
1319 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1326 /// isBytewiseValue - If the specified value can be set by repeating the same
1327 /// byte in memory, return the i8 value that it is represented with. This is
1328 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1329 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1330 /// byte store (e.g. i16 0x1234), return null.
1331 Value *llvm::isBytewiseValue(Value *V) {
1332 // All byte-wide stores are splatable, even of arbitrary variables.
1333 if (V->getType()->isIntegerTy(8)) return V;
1335 // Handle 'null' ConstantArrayZero etc.
1336 if (Constant *C = dyn_cast<Constant>(V))
1337 if (C->isNullValue())
1338 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1340 // Constant float and double values can be handled as integer values if the
1341 // corresponding integer value is "byteable". An important case is 0.0.
1342 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1343 if (CFP->getType()->isFloatTy())
1344 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1345 if (CFP->getType()->isDoubleTy())
1346 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1347 // Don't handle long double formats, which have strange constraints.
1350 // We can handle constant integers that are power of two in size and a
1351 // multiple of 8 bits.
1352 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1353 unsigned Width = CI->getBitWidth();
1354 if (isPowerOf2_32(Width) && Width > 8) {
1355 // We can handle this value if the recursive binary decomposition is the
1356 // same at all levels.
1357 APInt Val = CI->getValue();
1359 while (Val.getBitWidth() != 8) {
1360 unsigned NextWidth = Val.getBitWidth()/2;
1361 Val2 = Val.lshr(NextWidth);
1362 Val2 = Val2.trunc(Val.getBitWidth()/2);
1363 Val = Val.trunc(Val.getBitWidth()/2);
1365 // If the top/bottom halves aren't the same, reject it.
1369 return ConstantInt::get(V->getContext(), Val);
1373 // A ConstantArray is splatable if all its members are equal and also
1375 if (ConstantArray *CA = dyn_cast<ConstantArray>(V)) {
1376 if (CA->getNumOperands() == 0)
1379 Value *Val = isBytewiseValue(CA->getOperand(0));
1383 for (unsigned I = 1, E = CA->getNumOperands(); I != E; ++I)
1384 if (CA->getOperand(I-1) != CA->getOperand(I))
1390 // FIXME: Vector types (e.g., <4 x i32> <i32 -1, i32 -1, i32 -1, i32 -1>).
1392 // Conceptually, we could handle things like:
1393 // %a = zext i8 %X to i16
1394 // %b = shl i16 %a, 8
1395 // %c = or i16 %a, %b
1396 // but until there is an example that actually needs this, it doesn't seem
1397 // worth worrying about.
1402 // This is the recursive version of BuildSubAggregate. It takes a few different
1403 // arguments. Idxs is the index within the nested struct From that we are
1404 // looking at now (which is of type IndexedType). IdxSkip is the number of
1405 // indices from Idxs that should be left out when inserting into the resulting
1406 // struct. To is the result struct built so far, new insertvalue instructions
1408 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1409 SmallVector<unsigned, 10> &Idxs,
1411 Instruction *InsertBefore) {
1412 llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1414 // Save the original To argument so we can modify it
1416 // General case, the type indexed by Idxs is a struct
1417 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1418 // Process each struct element recursively
1421 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1425 // Couldn't find any inserted value for this index? Cleanup
1426 while (PrevTo != OrigTo) {
1427 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1428 PrevTo = Del->getAggregateOperand();
1429 Del->eraseFromParent();
1431 // Stop processing elements
1435 // If we successfully found a value for each of our subaggregates
1439 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1440 // the struct's elements had a value that was inserted directly. In the latter
1441 // case, perhaps we can't determine each of the subelements individually, but
1442 // we might be able to find the complete struct somewhere.
1444 // Find the value that is at that particular spot
1445 Value *V = FindInsertedValue(From, Idxs);
1450 // Insert the value in the new (sub) aggregrate
1451 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1452 "tmp", InsertBefore);
1455 // This helper takes a nested struct and extracts a part of it (which is again a
1456 // struct) into a new value. For example, given the struct:
1457 // { a, { b, { c, d }, e } }
1458 // and the indices "1, 1" this returns
1461 // It does this by inserting an insertvalue for each element in the resulting
1462 // struct, as opposed to just inserting a single struct. This will only work if
1463 // each of the elements of the substruct are known (ie, inserted into From by an
1464 // insertvalue instruction somewhere).
1466 // All inserted insertvalue instructions are inserted before InsertBefore
1467 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1468 Instruction *InsertBefore) {
1469 assert(InsertBefore && "Must have someplace to insert!");
1470 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1472 Value *To = UndefValue::get(IndexedType);
1473 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1474 unsigned IdxSkip = Idxs.size();
1476 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1479 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1480 /// the scalar value indexed is already around as a register, for example if it
1481 /// were inserted directly into the aggregrate.
1483 /// If InsertBefore is not null, this function will duplicate (modified)
1484 /// insertvalues when a part of a nested struct is extracted.
1485 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1486 Instruction *InsertBefore) {
1487 // Nothing to index? Just return V then (this is useful at the end of our
1489 if (idx_range.empty())
1491 // We have indices, so V should have an indexable type.
1492 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1493 "Not looking at a struct or array?");
1494 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1495 "Invalid indices for type?");
1496 CompositeType *PTy = cast<CompositeType>(V->getType());
1498 if (isa<UndefValue>(V))
1499 return UndefValue::get(ExtractValueInst::getIndexedType(PTy, idx_range));
1500 if (isa<ConstantAggregateZero>(V))
1501 return Constant::getNullValue(ExtractValueInst::getIndexedType(PTy,
1503 if (isa<ConstantArray>(V) || isa<ConstantStruct>(V))
1504 // Recursively process this constant
1505 return FindInsertedValue(cast<Constant>(V)->getOperand(idx_range[0]),
1506 idx_range.slice(1), InsertBefore);
1507 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V))
1508 return CDS->getElementAsConstant(idx_range[0]);
1510 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1511 // Loop the indices for the insertvalue instruction in parallel with the
1512 // requested indices
1513 const unsigned *req_idx = idx_range.begin();
1514 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1515 i != e; ++i, ++req_idx) {
1516 if (req_idx == idx_range.end()) {
1517 // We can't handle this without inserting insertvalues
1521 // The requested index identifies a part of a nested aggregate. Handle
1522 // this specially. For example,
1523 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1524 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1525 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1526 // This can be changed into
1527 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1528 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1529 // which allows the unused 0,0 element from the nested struct to be
1531 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1535 // This insert value inserts something else than what we are looking for.
1536 // See if the (aggregrate) value inserted into has the value we are
1537 // looking for, then.
1539 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1542 // If we end up here, the indices of the insertvalue match with those
1543 // requested (though possibly only partially). Now we recursively look at
1544 // the inserted value, passing any remaining indices.
1545 return FindInsertedValue(I->getInsertedValueOperand(),
1546 makeArrayRef(req_idx, idx_range.end()),
1550 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1551 // If we're extracting a value from an aggregrate that was extracted from
1552 // something else, we can extract from that something else directly instead.
1553 // However, we will need to chain I's indices with the requested indices.
1555 // Calculate the number of indices required
1556 unsigned size = I->getNumIndices() + idx_range.size();
1557 // Allocate some space to put the new indices in
1558 SmallVector<unsigned, 5> Idxs;
1560 // Add indices from the extract value instruction
1561 Idxs.append(I->idx_begin(), I->idx_end());
1563 // Add requested indices
1564 Idxs.append(idx_range.begin(), idx_range.end());
1566 assert(Idxs.size() == size
1567 && "Number of indices added not correct?");
1569 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1571 // Otherwise, we don't know (such as, extracting from a function return value
1572 // or load instruction)
1576 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1577 /// it can be expressed as a base pointer plus a constant offset. Return the
1578 /// base and offset to the caller.
1579 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1580 const TargetData &TD) {
1581 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1582 if (PtrOp == 0 || Ptr->getType()->isVectorTy())
1585 // Just look through bitcasts.
1586 if (PtrOp->getOpcode() == Instruction::BitCast)
1587 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1589 // If this is a GEP with constant indices, we can look through it.
1590 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1591 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1593 gep_type_iterator GTI = gep_type_begin(GEP);
1594 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1596 ConstantInt *OpC = cast<ConstantInt>(*I);
1597 if (OpC->isZero()) continue;
1599 // Handle a struct and array indices which add their offset to the pointer.
1600 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1601 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1603 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1604 Offset += OpC->getSExtValue()*Size;
1608 // Re-sign extend from the pointer size if needed to get overflow edge cases
1610 unsigned PtrSize = TD.getPointerSizeInBits();
1612 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1614 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1618 /// GetConstantStringInfo - This function computes the length of a
1619 /// null-terminated C string pointed to by V. If successful, it returns true
1620 /// and returns the string in Str. If unsuccessful, it returns false.
1621 bool llvm::GetConstantStringInfo(const Value *V, std::string &Str,
1622 uint64_t Offset, bool StopAtNul) {
1623 // If V is NULL then return false;
1624 if (V == NULL) return false;
1626 // Look through bitcast instructions.
1627 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1628 return GetConstantStringInfo(BCI->getOperand(0), Str, Offset, StopAtNul);
1630 // If the value is not a GEP instruction nor a constant expression with a
1631 // GEP instruction, then return false because ConstantArray can't occur
1633 const User *GEP = 0;
1634 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1636 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1637 if (CE->getOpcode() == Instruction::BitCast)
1638 return GetConstantStringInfo(CE->getOperand(0), Str, Offset, StopAtNul);
1639 if (CE->getOpcode() != Instruction::GetElementPtr)
1645 // Make sure the GEP has exactly three arguments.
1646 if (GEP->getNumOperands() != 3)
1649 // Make sure the index-ee is a pointer to array of i8.
1650 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1651 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1652 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1655 // Check to make sure that the first operand of the GEP is an integer and
1656 // has value 0 so that we are sure we're indexing into the initializer.
1657 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1658 if (FirstIdx == 0 || !FirstIdx->isZero())
1661 // If the second index isn't a ConstantInt, then this is a variable index
1662 // into the array. If this occurs, we can't say anything meaningful about
1664 uint64_t StartIdx = 0;
1665 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1666 StartIdx = CI->getZExtValue();
1669 return GetConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset,
1673 // The GEP instruction, constant or instruction, must reference a global
1674 // variable that is a constant and is initialized. The referenced constant
1675 // initializer is the array that we'll use for optimization.
1676 const GlobalVariable* GV = dyn_cast<GlobalVariable>(V);
1677 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1679 const Constant *GlobalInit = GV->getInitializer();
1681 // Handle the all-zeros case
1682 if (GlobalInit->isNullValue()) {
1683 // This is a degenerate case. The initializer is constant zero so the
1684 // length of the string must be zero.
1689 // Must be a Constant Array
1690 const ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1691 if (Array == 0 || !Array->getType()->getElementType()->isIntegerTy(8))
1694 // Get the number of elements in the array
1695 uint64_t NumElts = Array->getType()->getNumElements();
1697 if (Offset > NumElts)
1700 // Traverse the constant array from 'Offset' which is the place the GEP refers
1702 Str.reserve(NumElts-Offset);
1703 for (unsigned i = Offset; i != NumElts; ++i) {
1704 const Constant *Elt = Array->getOperand(i);
1705 const ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1706 if (!CI) // This array isn't suitable, non-int initializer.
1708 if (StopAtNul && CI->isZero())
1709 return true; // we found end of string, success!
1710 Str += (char)CI->getZExtValue();
1713 // The array isn't null terminated, but maybe this is a memcpy, not a strcpy.
1717 // These next two are very similar to the above, but also look through PHI
1719 // TODO: See if we can integrate these two together.
1721 /// GetStringLengthH - If we can compute the length of the string pointed to by
1722 /// the specified pointer, return 'len+1'. If we can't, return 0.
1723 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1724 // Look through noop bitcast instructions.
1725 if (BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1726 return GetStringLengthH(BCI->getOperand(0), PHIs);
1728 // If this is a PHI node, there are two cases: either we have already seen it
1730 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1731 if (!PHIs.insert(PN))
1732 return ~0ULL; // already in the set.
1734 // If it was new, see if all the input strings are the same length.
1735 uint64_t LenSoFar = ~0ULL;
1736 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1737 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1738 if (Len == 0) return 0; // Unknown length -> unknown.
1740 if (Len == ~0ULL) continue;
1742 if (Len != LenSoFar && LenSoFar != ~0ULL)
1743 return 0; // Disagree -> unknown.
1747 // Success, all agree.
1751 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1752 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1753 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1754 if (Len1 == 0) return 0;
1755 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1756 if (Len2 == 0) return 0;
1757 if (Len1 == ~0ULL) return Len2;
1758 if (Len2 == ~0ULL) return Len1;
1759 if (Len1 != Len2) return 0;
1763 // As a special-case, "@string = constant i8 0" is also a string with zero
1764 // length, not wrapped in a bitcast or GEP.
1765 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
1766 if (GV->isConstant() && GV->hasDefinitiveInitializer())
1767 if (GV->getInitializer()->isNullValue()) return 1;
1771 // If the value is not a GEP instruction nor a constant expression with a
1772 // GEP instruction, then return unknown.
1774 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1776 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1777 if (CE->getOpcode() != Instruction::GetElementPtr)
1784 // Make sure the GEP has exactly three arguments.
1785 if (GEP->getNumOperands() != 3)
1788 // Check to make sure that the first operand of the GEP is an integer and
1789 // has value 0 so that we are sure we're indexing into the initializer.
1790 if (ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(1))) {
1796 // If the second index isn't a ConstantInt, then this is a variable index
1797 // into the array. If this occurs, we can't say anything meaningful about
1799 uint64_t StartIdx = 0;
1800 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1801 StartIdx = CI->getZExtValue();
1805 // The GEP instruction, constant or instruction, must reference a global
1806 // variable that is a constant and is initialized. The referenced constant
1807 // initializer is the array that we'll use for optimization.
1808 GlobalVariable* GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
1809 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
1810 GV->mayBeOverridden())
1812 Constant *GlobalInit = GV->getInitializer();
1814 // Handle the ConstantAggregateZero case, which is a degenerate case. The
1815 // initializer is constant zero so the length of the string must be zero.
1816 if (isa<ConstantAggregateZero>(GlobalInit))
1817 return 1; // Len = 0 offset by 1.
1819 // Must be a Constant Array
1820 ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1821 if (!Array || !Array->getType()->getElementType()->isIntegerTy(8))
1824 // Get the number of elements in the array
1825 uint64_t NumElts = Array->getType()->getNumElements();
1827 // Traverse the constant array from StartIdx (derived above) which is
1828 // the place the GEP refers to in the array.
1829 for (unsigned i = StartIdx; i != NumElts; ++i) {
1830 Constant *Elt = Array->getOperand(i);
1831 ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1832 if (!CI) // This array isn't suitable, non-int initializer.
1835 return i-StartIdx+1; // We found end of string, success!
1838 return 0; // The array isn't null terminated, conservatively return 'unknown'.
1841 /// GetStringLength - If we can compute the length of the string pointed to by
1842 /// the specified pointer, return 'len+1'. If we can't, return 0.
1843 uint64_t llvm::GetStringLength(Value *V) {
1844 if (!V->getType()->isPointerTy()) return 0;
1846 SmallPtrSet<PHINode*, 32> PHIs;
1847 uint64_t Len = GetStringLengthH(V, PHIs);
1848 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1849 // an empty string as a length.
1850 return Len == ~0ULL ? 1 : Len;
1854 llvm::GetUnderlyingObject(Value *V, const TargetData *TD, unsigned MaxLookup) {
1855 if (!V->getType()->isPointerTy())
1857 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1858 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1859 V = GEP->getPointerOperand();
1860 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1861 V = cast<Operator>(V)->getOperand(0);
1862 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1863 if (GA->mayBeOverridden())
1865 V = GA->getAliasee();
1867 // See if InstructionSimplify knows any relevant tricks.
1868 if (Instruction *I = dyn_cast<Instruction>(V))
1869 // TODO: Acquire a DominatorTree and use it.
1870 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1877 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1882 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1883 /// are lifetime markers.
1885 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1886 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1888 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1889 if (!II) return false;
1891 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1892 II->getIntrinsicID() != Intrinsic::lifetime_end)
1898 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1899 const TargetData *TD) {
1900 const Operator *Inst = dyn_cast<Operator>(V);
1904 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1905 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1909 switch (Inst->getOpcode()) {
1912 case Instruction::UDiv:
1913 case Instruction::URem:
1914 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1915 return isKnownNonZero(Inst->getOperand(1), TD);
1916 case Instruction::SDiv:
1917 case Instruction::SRem: {
1918 Value *Op = Inst->getOperand(1);
1919 // x / y is undefined if y == 0
1920 if (!isKnownNonZero(Op, TD))
1922 // x / y might be undefined if y == -1
1923 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1926 APInt KnownZero(BitWidth, 0);
1927 APInt KnownOne(BitWidth, 0);
1928 ComputeMaskedBits(Op, APInt::getAllOnesValue(BitWidth),
1929 KnownZero, KnownOne, TD);
1932 case Instruction::Load: {
1933 const LoadInst *LI = cast<LoadInst>(Inst);
1934 if (!LI->isUnordered())
1936 return LI->getPointerOperand()->isDereferenceablePointer();
1938 case Instruction::Call: {
1939 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1940 switch (II->getIntrinsicID()) {
1941 case Intrinsic::bswap:
1942 case Intrinsic::ctlz:
1943 case Intrinsic::ctpop:
1944 case Intrinsic::cttz:
1945 case Intrinsic::objectsize:
1946 case Intrinsic::sadd_with_overflow:
1947 case Intrinsic::smul_with_overflow:
1948 case Intrinsic::ssub_with_overflow:
1949 case Intrinsic::uadd_with_overflow:
1950 case Intrinsic::umul_with_overflow:
1951 case Intrinsic::usub_with_overflow:
1953 // TODO: some fp intrinsics are marked as having the same error handling
1954 // as libm. They're safe to speculate when they won't error.
1955 // TODO: are convert_{from,to}_fp16 safe?
1956 // TODO: can we list target-specific intrinsics here?
1960 return false; // The called function could have undefined behavior or
1961 // side-effects, even if marked readnone nounwind.
1963 case Instruction::VAArg:
1964 case Instruction::Alloca:
1965 case Instruction::Invoke:
1966 case Instruction::PHI:
1967 case Instruction::Store:
1968 case Instruction::Ret:
1969 case Instruction::Br:
1970 case Instruction::IndirectBr:
1971 case Instruction::Switch:
1972 case Instruction::Unwind:
1973 case Instruction::Unreachable:
1974 case Instruction::Fence:
1975 case Instruction::LandingPad:
1976 case Instruction::AtomicRMW:
1977 case Instruction::AtomicCmpXchg:
1978 case Instruction::Resume:
1979 return false; // Misc instructions which have effects