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/ADT/SmallPtrSet.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/MemoryBuiltins.h"
19 #include "llvm/IR/CallSite.h"
20 #include "llvm/IR/ConstantRange.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/GetElementPtrTypeIterator.h"
24 #include "llvm/IR/GlobalAlias.h"
25 #include "llvm/IR/GlobalVariable.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/Metadata.h"
30 #include "llvm/IR/Operator.h"
31 #include "llvm/IR/PatternMatch.h"
32 #include "llvm/Support/MathExtras.h"
35 using namespace llvm::PatternMatch;
37 const unsigned MaxDepth = 6;
39 /// getBitWidth - Returns the bitwidth of the given scalar or pointer type (if
40 /// unknown returns 0). For vector types, returns the element type's bitwidth.
41 static unsigned getBitWidth(Type *Ty, const DataLayout *TD) {
42 if (unsigned BitWidth = Ty->getScalarSizeInBits())
45 return TD ? TD->getPointerTypeSizeInBits(Ty) : 0;
48 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
49 APInt &KnownZero, APInt &KnownOne,
50 APInt &KnownZero2, APInt &KnownOne2,
51 const DataLayout *TD, unsigned Depth) {
53 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
54 // We know that the top bits of C-X are clear if X contains less bits
55 // than C (i.e. no wrap-around can happen). For example, 20-X is
56 // positive if we can prove that X is >= 0 and < 16.
57 if (!CLHS->getValue().isNegative()) {
58 unsigned BitWidth = KnownZero.getBitWidth();
59 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
60 // NLZ can't be BitWidth with no sign bit
61 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
62 llvm::computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1);
64 // If all of the MaskV bits are known to be zero, then we know the
65 // output top bits are zero, because we now know that the output is
67 if ((KnownZero2 & MaskV) == MaskV) {
68 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
69 // Top bits known zero.
70 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
76 unsigned BitWidth = KnownZero.getBitWidth();
78 // If one of the operands has trailing zeros, then the bits that the
79 // other operand has in those bit positions will be preserved in the
80 // result. For an add, this works with either operand. For a subtract,
81 // this only works if the known zeros are in the right operand.
82 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
83 llvm::computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, TD, Depth+1);
84 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
86 llvm::computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1);
87 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
89 // Determine which operand has more trailing zeros, and use that
90 // many bits from the other operand.
91 if (LHSKnownZeroOut > RHSKnownZeroOut) {
93 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
94 KnownZero |= KnownZero2 & Mask;
95 KnownOne |= KnownOne2 & Mask;
97 // If the known zeros are in the left operand for a subtract,
98 // fall back to the minimum known zeros in both operands.
99 KnownZero |= APInt::getLowBitsSet(BitWidth,
100 std::min(LHSKnownZeroOut,
103 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
104 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
105 KnownZero |= LHSKnownZero & Mask;
106 KnownOne |= LHSKnownOne & Mask;
109 // Are we still trying to solve for the sign bit?
110 if (!KnownZero.isNegative() && !KnownOne.isNegative()) {
113 // Adding two positive numbers can't wrap into negative
114 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
115 KnownZero |= APInt::getSignBit(BitWidth);
116 // and adding two negative numbers can't wrap into positive.
117 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
118 KnownOne |= APInt::getSignBit(BitWidth);
120 // Subtracting a negative number from a positive one can't wrap
121 if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
122 KnownZero |= APInt::getSignBit(BitWidth);
123 // neither can subtracting a positive number from a negative one.
124 else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
125 KnownOne |= APInt::getSignBit(BitWidth);
131 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
132 APInt &KnownZero, APInt &KnownOne,
133 APInt &KnownZero2, APInt &KnownOne2,
134 const DataLayout *TD, unsigned Depth) {
135 unsigned BitWidth = KnownZero.getBitWidth();
136 computeKnownBits(Op1, KnownZero, KnownOne, TD, Depth+1);
137 computeKnownBits(Op0, KnownZero2, KnownOne2, TD, Depth+1);
139 bool isKnownNegative = false;
140 bool isKnownNonNegative = false;
141 // If the multiplication is known not to overflow, compute the sign bit.
144 // The product of a number with itself is non-negative.
145 isKnownNonNegative = true;
147 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
148 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
149 bool isKnownNegativeOp1 = KnownOne.isNegative();
150 bool isKnownNegativeOp0 = KnownOne2.isNegative();
151 // The product of two numbers with the same sign is non-negative.
152 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
153 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
154 // The product of a negative number and a non-negative number is either
156 if (!isKnownNonNegative)
157 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
158 isKnownNonZero(Op0, TD, Depth)) ||
159 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
160 isKnownNonZero(Op1, TD, Depth));
164 // If low bits are zero in either operand, output low known-0 bits.
165 // Also compute a conserative estimate for high known-0 bits.
166 // More trickiness is possible, but this is sufficient for the
167 // interesting case of alignment computation.
168 KnownOne.clearAllBits();
169 unsigned TrailZ = KnownZero.countTrailingOnes() +
170 KnownZero2.countTrailingOnes();
171 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
172 KnownZero2.countLeadingOnes(),
173 BitWidth) - BitWidth;
175 TrailZ = std::min(TrailZ, BitWidth);
176 LeadZ = std::min(LeadZ, BitWidth);
177 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
178 APInt::getHighBitsSet(BitWidth, LeadZ);
180 // Only make use of no-wrap flags if we failed to compute the sign bit
181 // directly. This matters if the multiplication always overflows, in
182 // which case we prefer to follow the result of the direct computation,
183 // though as the program is invoking undefined behaviour we can choose
184 // whatever we like here.
185 if (isKnownNonNegative && !KnownOne.isNegative())
186 KnownZero.setBit(BitWidth - 1);
187 else if (isKnownNegative && !KnownZero.isNegative())
188 KnownOne.setBit(BitWidth - 1);
191 void llvm::computeKnownBitsLoad(const MDNode &Ranges, APInt &KnownZero) {
192 unsigned BitWidth = KnownZero.getBitWidth();
193 unsigned NumRanges = Ranges.getNumOperands() / 2;
194 assert(NumRanges >= 1);
196 // Use the high end of the ranges to find leading zeros.
197 unsigned MinLeadingZeros = BitWidth;
198 for (unsigned i = 0; i < NumRanges; ++i) {
199 ConstantInt *Lower = cast<ConstantInt>(Ranges.getOperand(2*i + 0));
200 ConstantInt *Upper = cast<ConstantInt>(Ranges.getOperand(2*i + 1));
201 ConstantRange Range(Lower->getValue(), Upper->getValue());
202 if (Range.isWrappedSet())
203 MinLeadingZeros = 0; // -1 has no zeros
204 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
205 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
208 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
211 /// Determine which bits of V are known to be either zero or one and return
212 /// them in the KnownZero/KnownOne bit sets.
214 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
215 /// we cannot optimize based on the assumption that it is zero without changing
216 /// it to be an explicit zero. If we don't change it to zero, other code could
217 /// optimized based on the contradictory assumption that it is non-zero.
218 /// Because instcombine aggressively folds operations with undef args anyway,
219 /// this won't lose us code quality.
221 /// This function is defined on values with integer type, values with pointer
222 /// type (but only if TD is non-null), and vectors of integers. In the case
223 /// where V is a vector, known zero, and known one values are the
224 /// same width as the vector element, and the bit is set only if it is true
225 /// for all of the elements in the vector.
226 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
227 const DataLayout *TD, unsigned Depth) {
228 assert(V && "No Value?");
229 assert(Depth <= MaxDepth && "Limit Search Depth");
230 unsigned BitWidth = KnownZero.getBitWidth();
232 assert((V->getType()->isIntOrIntVectorTy() ||
233 V->getType()->getScalarType()->isPointerTy()) &&
234 "Not integer or pointer type!");
236 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
237 (!V->getType()->isIntOrIntVectorTy() ||
238 V->getType()->getScalarSizeInBits() == BitWidth) &&
239 KnownZero.getBitWidth() == BitWidth &&
240 KnownOne.getBitWidth() == BitWidth &&
241 "V, KnownOne and KnownZero should have same BitWidth");
243 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
244 // We know all of the bits for a constant!
245 KnownOne = CI->getValue();
246 KnownZero = ~KnownOne;
249 // Null and aggregate-zero are all-zeros.
250 if (isa<ConstantPointerNull>(V) ||
251 isa<ConstantAggregateZero>(V)) {
252 KnownOne.clearAllBits();
253 KnownZero = APInt::getAllOnesValue(BitWidth);
256 // Handle a constant vector by taking the intersection of the known bits of
257 // each element. There is no real need to handle ConstantVector here, because
258 // we don't handle undef in any particularly useful way.
259 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
260 // We know that CDS must be a vector of integers. Take the intersection of
262 KnownZero.setAllBits(); KnownOne.setAllBits();
263 APInt Elt(KnownZero.getBitWidth(), 0);
264 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
265 Elt = CDS->getElementAsInteger(i);
272 // The address of an aligned GlobalValue has trailing zeros.
273 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
274 unsigned Align = GV->getAlignment();
275 if (Align == 0 && TD) {
276 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
277 Type *ObjectType = GVar->getType()->getElementType();
278 if (ObjectType->isSized()) {
279 // If the object is defined in the current Module, we'll be giving
280 // it the preferred alignment. Otherwise, we have to assume that it
281 // may only have the minimum ABI alignment.
282 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
283 Align = TD->getPreferredAlignment(GVar);
285 Align = TD->getABITypeAlignment(ObjectType);
290 KnownZero = APInt::getLowBitsSet(BitWidth,
291 countTrailingZeros(Align));
293 KnownZero.clearAllBits();
294 KnownOne.clearAllBits();
297 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
298 // the bits of its aliasee.
299 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
300 if (GA->mayBeOverridden()) {
301 KnownZero.clearAllBits(); KnownOne.clearAllBits();
303 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth+1);
308 if (Argument *A = dyn_cast<Argument>(V)) {
311 if (A->hasByValOrInAllocaAttr()) {
312 // Get alignment information off byval/inalloca arguments if specified in
314 Align = A->getParamAlignment();
315 } else if (TD && A->hasStructRetAttr()) {
316 // An sret parameter has at least the ABI alignment of the return type.
317 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
318 if (EltTy->isSized())
319 Align = TD->getABITypeAlignment(EltTy);
323 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
327 // Start out not knowing anything.
328 KnownZero.clearAllBits(); KnownOne.clearAllBits();
330 if (Depth == MaxDepth)
331 return; // Limit search depth.
333 Operator *I = dyn_cast<Operator>(V);
336 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
337 switch (I->getOpcode()) {
339 case Instruction::Load:
340 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
341 computeKnownBitsLoad(*MD, KnownZero);
343 case Instruction::And: {
344 // If either the LHS or the RHS are Zero, the result is zero.
345 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
346 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
348 // Output known-1 bits are only known if set in both the LHS & RHS.
349 KnownOne &= KnownOne2;
350 // Output known-0 are known to be clear if zero in either the LHS | RHS.
351 KnownZero |= KnownZero2;
354 case Instruction::Or: {
355 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
356 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
358 // Output known-0 bits are only known if clear in both the LHS & RHS.
359 KnownZero &= KnownZero2;
360 // Output known-1 are known to be set if set in either the LHS | RHS.
361 KnownOne |= KnownOne2;
364 case Instruction::Xor: {
365 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
366 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
368 // Output known-0 bits are known if clear or set in both the LHS & RHS.
369 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
370 // Output known-1 are known to be set if set in only one of the LHS, RHS.
371 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
372 KnownZero = KnownZeroOut;
375 case Instruction::Mul: {
376 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
377 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW,
378 KnownZero, KnownOne, KnownZero2, KnownOne2, TD, Depth);
381 case Instruction::UDiv: {
382 // For the purposes of computing leading zeros we can conservatively
383 // treat a udiv as a logical right shift by the power of 2 known to
384 // be less than the denominator.
385 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
386 unsigned LeadZ = KnownZero2.countLeadingOnes();
388 KnownOne2.clearAllBits();
389 KnownZero2.clearAllBits();
390 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
391 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
392 if (RHSUnknownLeadingOnes != BitWidth)
393 LeadZ = std::min(BitWidth,
394 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
396 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
399 case Instruction::Select:
400 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1);
401 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD,
404 // Only known if known in both the LHS and RHS.
405 KnownOne &= KnownOne2;
406 KnownZero &= KnownZero2;
408 case Instruction::FPTrunc:
409 case Instruction::FPExt:
410 case Instruction::FPToUI:
411 case Instruction::FPToSI:
412 case Instruction::SIToFP:
413 case Instruction::UIToFP:
414 break; // Can't work with floating point.
415 case Instruction::PtrToInt:
416 case Instruction::IntToPtr:
417 // We can't handle these if we don't know the pointer size.
419 // FALL THROUGH and handle them the same as zext/trunc.
420 case Instruction::ZExt:
421 case Instruction::Trunc: {
422 Type *SrcTy = I->getOperand(0)->getType();
424 unsigned SrcBitWidth;
425 // Note that we handle pointer operands here because of inttoptr/ptrtoint
426 // which fall through here.
428 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
430 SrcBitWidth = SrcTy->getScalarSizeInBits();
431 if (!SrcBitWidth) break;
434 assert(SrcBitWidth && "SrcBitWidth can't be zero");
435 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
436 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
437 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
438 KnownZero = KnownZero.zextOrTrunc(BitWidth);
439 KnownOne = KnownOne.zextOrTrunc(BitWidth);
440 // Any top bits are known to be zero.
441 if (BitWidth > SrcBitWidth)
442 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
445 case Instruction::BitCast: {
446 Type *SrcTy = I->getOperand(0)->getType();
447 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
448 // TODO: For now, not handling conversions like:
449 // (bitcast i64 %x to <2 x i32>)
450 !I->getType()->isVectorTy()) {
451 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
456 case Instruction::SExt: {
457 // Compute the bits in the result that are not present in the input.
458 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
460 KnownZero = KnownZero.trunc(SrcBitWidth);
461 KnownOne = KnownOne.trunc(SrcBitWidth);
462 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
463 KnownZero = KnownZero.zext(BitWidth);
464 KnownOne = KnownOne.zext(BitWidth);
466 // If the sign bit of the input is known set or clear, then we know the
467 // top bits of the result.
468 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
469 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
470 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
471 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
474 case Instruction::Shl:
475 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
476 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
477 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
478 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
479 KnownZero <<= ShiftAmt;
480 KnownOne <<= ShiftAmt;
481 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
485 case Instruction::LShr:
486 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
487 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
488 // Compute the new bits that are at the top now.
489 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
491 // Unsigned shift right.
492 computeKnownBits(I->getOperand(0), KnownZero,KnownOne, TD, Depth+1);
493 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
494 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
495 // high bits known zero.
496 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
500 case Instruction::AShr:
501 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
502 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
503 // Compute the new bits that are at the top now.
504 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
506 // Signed shift right.
507 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
508 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
509 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
511 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
512 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
513 KnownZero |= HighBits;
514 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
515 KnownOne |= HighBits;
519 case Instruction::Sub: {
520 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
521 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
522 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
526 case Instruction::Add: {
527 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
528 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
529 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
533 case Instruction::SRem:
534 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
535 APInt RA = Rem->getValue().abs();
536 if (RA.isPowerOf2()) {
537 APInt LowBits = RA - 1;
538 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
540 // The low bits of the first operand are unchanged by the srem.
541 KnownZero = KnownZero2 & LowBits;
542 KnownOne = KnownOne2 & LowBits;
544 // If the first operand is non-negative or has all low bits zero, then
545 // the upper bits are all zero.
546 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
547 KnownZero |= ~LowBits;
549 // If the first operand is negative and not all low bits are zero, then
550 // the upper bits are all one.
551 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
552 KnownOne |= ~LowBits;
554 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
558 // The sign bit is the LHS's sign bit, except when the result of the
559 // remainder is zero.
560 if (KnownZero.isNonNegative()) {
561 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
562 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
564 // If it's known zero, our sign bit is also zero.
565 if (LHSKnownZero.isNegative())
566 KnownZero.setBit(BitWidth - 1);
570 case Instruction::URem: {
571 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
572 APInt RA = Rem->getValue();
573 if (RA.isPowerOf2()) {
574 APInt LowBits = (RA - 1);
575 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD,
577 KnownZero |= ~LowBits;
583 // Since the result is less than or equal to either operand, any leading
584 // zero bits in either operand must also exist in the result.
585 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
586 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
588 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
589 KnownZero2.countLeadingOnes());
590 KnownOne.clearAllBits();
591 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
595 case Instruction::Alloca: {
596 AllocaInst *AI = cast<AllocaInst>(V);
597 unsigned Align = AI->getAlignment();
598 if (Align == 0 && TD)
599 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
602 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
605 case Instruction::GetElementPtr: {
606 // Analyze all of the subscripts of this getelementptr instruction
607 // to determine if we can prove known low zero bits.
608 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
609 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
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.
623 // Handle case when index is vector zeroinitializer
624 Constant *CIndex = cast<Constant>(Index);
625 if (CIndex->isZeroValue())
628 if (CIndex->getType()->isVectorTy())
629 Index = CIndex->getSplatValue();
631 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
632 const StructLayout *SL = TD->getStructLayout(STy);
633 uint64_t Offset = SL->getElementOffset(Idx);
634 TrailZ = std::min<unsigned>(TrailZ,
635 countTrailingZeros(Offset));
637 // Handle array index arithmetic.
638 Type *IndexedTy = GTI.getIndexedType();
639 if (!IndexedTy->isSized()) {
643 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
644 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
645 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
646 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1);
647 TrailZ = std::min(TrailZ,
648 unsigned(countTrailingZeros(TypeSize) +
649 LocalKnownZero.countTrailingOnes()));
653 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
656 case Instruction::PHI: {
657 PHINode *P = cast<PHINode>(I);
658 // Handle the case of a simple two-predecessor recurrence PHI.
659 // There's a lot more that could theoretically be done here, but
660 // this is sufficient to catch some interesting cases.
661 if (P->getNumIncomingValues() == 2) {
662 for (unsigned i = 0; i != 2; ++i) {
663 Value *L = P->getIncomingValue(i);
664 Value *R = P->getIncomingValue(!i);
665 Operator *LU = dyn_cast<Operator>(L);
668 unsigned Opcode = LU->getOpcode();
669 // Check for operations that have the property that if
670 // both their operands have low zero bits, the result
671 // will have low zero bits.
672 if (Opcode == Instruction::Add ||
673 Opcode == Instruction::Sub ||
674 Opcode == Instruction::And ||
675 Opcode == Instruction::Or ||
676 Opcode == Instruction::Mul) {
677 Value *LL = LU->getOperand(0);
678 Value *LR = LU->getOperand(1);
679 // Find a recurrence.
686 // Ok, we have a PHI of the form L op= R. Check for low
688 computeKnownBits(R, KnownZero2, KnownOne2, TD, Depth+1);
690 // We need to take the minimum number of known bits
691 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
692 computeKnownBits(L, KnownZero3, KnownOne3, TD, Depth+1);
694 KnownZero = APInt::getLowBitsSet(BitWidth,
695 std::min(KnownZero2.countTrailingOnes(),
696 KnownZero3.countTrailingOnes()));
702 // Unreachable blocks may have zero-operand PHI nodes.
703 if (P->getNumIncomingValues() == 0)
706 // Otherwise take the unions of the known bit sets of the operands,
707 // taking conservative care to avoid excessive recursion.
708 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
709 // Skip if every incoming value references to ourself.
710 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
713 KnownZero = APInt::getAllOnesValue(BitWidth);
714 KnownOne = APInt::getAllOnesValue(BitWidth);
715 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
716 // Skip direct self references.
717 if (P->getIncomingValue(i) == P) continue;
719 KnownZero2 = APInt(BitWidth, 0);
720 KnownOne2 = APInt(BitWidth, 0);
721 // Recurse, but cap the recursion to one level, because we don't
722 // want to waste time spinning around in loops.
723 computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
725 KnownZero &= KnownZero2;
726 KnownOne &= KnownOne2;
727 // If all bits have been ruled out, there's no need to check
729 if (!KnownZero && !KnownOne)
735 case Instruction::Call:
736 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
737 switch (II->getIntrinsicID()) {
739 case Intrinsic::ctlz:
740 case Intrinsic::cttz: {
741 unsigned LowBits = Log2_32(BitWidth)+1;
742 // If this call is undefined for 0, the result will be less than 2^n.
743 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
745 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
748 case Intrinsic::ctpop: {
749 unsigned LowBits = Log2_32(BitWidth)+1;
750 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
753 case Intrinsic::x86_sse42_crc32_64_64:
754 KnownZero = APInt::getHighBitsSet(64, 32);
759 case Instruction::ExtractValue:
760 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
761 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
762 if (EVI->getNumIndices() != 1) break;
763 if (EVI->getIndices()[0] == 0) {
764 switch (II->getIntrinsicID()) {
766 case Intrinsic::uadd_with_overflow:
767 case Intrinsic::sadd_with_overflow:
768 computeKnownBitsAddSub(true, II->getArgOperand(0),
769 II->getArgOperand(1), false, KnownZero,
770 KnownOne, KnownZero2, KnownOne2, TD, Depth);
772 case Intrinsic::usub_with_overflow:
773 case Intrinsic::ssub_with_overflow:
774 computeKnownBitsAddSub(false, II->getArgOperand(0),
775 II->getArgOperand(1), false, KnownZero,
776 KnownOne, KnownZero2, KnownOne2, TD, Depth);
778 case Intrinsic::umul_with_overflow:
779 case Intrinsic::smul_with_overflow:
780 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1),
781 false, KnownZero, KnownOne,
782 KnownZero2, KnownOne2, TD, Depth);
789 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
792 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
793 /// one. Convenience wrapper around computeKnownBits.
794 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
795 const DataLayout *TD, unsigned Depth) {
796 unsigned BitWidth = getBitWidth(V->getType(), TD);
802 APInt ZeroBits(BitWidth, 0);
803 APInt OneBits(BitWidth, 0);
804 computeKnownBits(V, ZeroBits, OneBits, TD, Depth);
805 KnownOne = OneBits[BitWidth - 1];
806 KnownZero = ZeroBits[BitWidth - 1];
809 /// isKnownToBeAPowerOfTwo - Return true if the given value is known to have exactly one
810 /// bit set when defined. For vectors return true if every element is known to
811 /// be a power of two when defined. Supports values with integer or pointer
812 /// types and vectors of integers.
813 bool llvm::isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth) {
814 if (Constant *C = dyn_cast<Constant>(V)) {
815 if (C->isNullValue())
817 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
818 return CI->getValue().isPowerOf2();
819 // TODO: Handle vector constants.
822 // 1 << X is clearly a power of two if the one is not shifted off the end. If
823 // it is shifted off the end then the result is undefined.
824 if (match(V, m_Shl(m_One(), m_Value())))
827 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
828 // bottom. If it is shifted off the bottom then the result is undefined.
829 if (match(V, m_LShr(m_SignBit(), m_Value())))
832 // The remaining tests are all recursive, so bail out if we hit the limit.
833 if (Depth++ == MaxDepth)
836 Value *X = nullptr, *Y = nullptr;
837 // A shift of a power of two is a power of two or zero.
838 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
839 match(V, m_Shr(m_Value(X), m_Value()))))
840 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth);
842 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
843 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth);
845 if (SelectInst *SI = dyn_cast<SelectInst>(V))
846 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth) &&
847 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth);
849 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
850 // A power of two and'd with anything is a power of two or zero.
851 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth) ||
852 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth))
854 // X & (-X) is always a power of two or zero.
855 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
860 // Adding a power-of-two or zero to the same power-of-two or zero yields
861 // either the original power-of-two, a larger power-of-two or zero.
862 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
863 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
864 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
865 if (match(X, m_And(m_Specific(Y), m_Value())) ||
866 match(X, m_And(m_Value(), m_Specific(Y))))
867 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth))
869 if (match(Y, m_And(m_Specific(X), m_Value())) ||
870 match(Y, m_And(m_Value(), m_Specific(X))))
871 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth))
874 unsigned BitWidth = V->getType()->getScalarSizeInBits();
875 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
876 computeKnownBits(X, LHSZeroBits, LHSOneBits, nullptr, Depth);
878 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
879 computeKnownBits(Y, RHSZeroBits, RHSOneBits, nullptr, Depth);
880 // If i8 V is a power of two or zero:
881 // ZeroBits: 1 1 1 0 1 1 1 1
882 // ~ZeroBits: 0 0 0 1 0 0 0 0
883 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
884 // If OrZero isn't set, we cannot give back a zero result.
885 // Make sure either the LHS or RHS has a bit set.
886 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
891 // An exact divide or right shift can only shift off zero bits, so the result
892 // is a power of two only if the first operand is a power of two and not
893 // copying a sign bit (sdiv int_min, 2).
894 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
895 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
896 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, Depth);
902 /// \brief Test whether a GEP's result is known to be non-null.
904 /// Uses properties inherent in a GEP to try to determine whether it is known
907 /// Currently this routine does not support vector GEPs.
908 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
910 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
913 // FIXME: Support vector-GEPs.
914 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
916 // If the base pointer is non-null, we cannot walk to a null address with an
917 // inbounds GEP in address space zero.
918 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth))
921 // Past this, if we don't have DataLayout, we can't do much.
925 // Walk the GEP operands and see if any operand introduces a non-zero offset.
926 // If so, then the GEP cannot produce a null pointer, as doing so would
927 // inherently violate the inbounds contract within address space zero.
928 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
930 // Struct types are easy -- they must always be indexed by a constant.
931 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
932 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
933 unsigned ElementIdx = OpC->getZExtValue();
934 const StructLayout *SL = DL->getStructLayout(STy);
935 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
936 if (ElementOffset > 0)
941 // If we have a zero-sized type, the index doesn't matter. Keep looping.
942 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
945 // Fast path the constant operand case both for efficiency and so we don't
946 // increment Depth when just zipping down an all-constant GEP.
947 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
953 // We post-increment Depth here because while isKnownNonZero increments it
954 // as well, when we pop back up that increment won't persist. We don't want
955 // to recurse 10k times just because we have 10k GEP operands. We don't
956 // bail completely out because we want to handle constant GEPs regardless
958 if (Depth++ >= MaxDepth)
961 if (isKnownNonZero(GTI.getOperand(), DL, Depth))
968 /// isKnownNonZero - Return true if the given value is known to be non-zero
969 /// when defined. For vectors return true if every element is known to be
970 /// non-zero when defined. Supports values with integer or pointer type and
971 /// vectors of integers.
972 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth) {
973 if (Constant *C = dyn_cast<Constant>(V)) {
974 if (C->isNullValue())
976 if (isa<ConstantInt>(C))
977 // Must be non-zero due to null test above.
979 // TODO: Handle vectors
983 // The remaining tests are all recursive, so bail out if we hit the limit.
984 if (Depth++ >= MaxDepth)
987 // Check for pointer simplifications.
988 if (V->getType()->isPointerTy()) {
989 if (isKnownNonNull(V))
991 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
992 if (isGEPKnownNonNull(GEP, TD, Depth))
996 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
998 // X | Y != 0 if X != 0 or Y != 0.
999 Value *X = nullptr, *Y = nullptr;
1000 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1001 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
1003 // ext X != 0 if X != 0.
1004 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1005 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
1007 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1008 // if the lowest bit is shifted off the end.
1009 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1010 // shl nuw can't remove any non-zero bits.
1011 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1012 if (BO->hasNoUnsignedWrap())
1013 return isKnownNonZero(X, TD, Depth);
1015 APInt KnownZero(BitWidth, 0);
1016 APInt KnownOne(BitWidth, 0);
1017 computeKnownBits(X, KnownZero, KnownOne, TD, Depth);
1021 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1022 // defined if the sign bit is shifted off the end.
1023 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1024 // shr exact can only shift out zero bits.
1025 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1027 return isKnownNonZero(X, TD, Depth);
1029 bool XKnownNonNegative, XKnownNegative;
1030 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
1034 // div exact can only produce a zero if the dividend is zero.
1035 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1036 return isKnownNonZero(X, TD, Depth);
1039 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1040 bool XKnownNonNegative, XKnownNegative;
1041 bool YKnownNonNegative, YKnownNegative;
1042 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
1043 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
1045 // If X and Y are both non-negative (as signed values) then their sum is not
1046 // zero unless both X and Y are zero.
1047 if (XKnownNonNegative && YKnownNonNegative)
1048 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
1051 // If X and Y are both negative (as signed values) then their sum is not
1052 // zero unless both X and Y equal INT_MIN.
1053 if (BitWidth && XKnownNegative && YKnownNegative) {
1054 APInt KnownZero(BitWidth, 0);
1055 APInt KnownOne(BitWidth, 0);
1056 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1057 // The sign bit of X is set. If some other bit is set then X is not equal
1059 computeKnownBits(X, KnownZero, KnownOne, TD, Depth);
1060 if ((KnownOne & Mask) != 0)
1062 // The sign bit of Y is set. If some other bit is set then Y is not equal
1064 computeKnownBits(Y, KnownZero, KnownOne, TD, Depth);
1065 if ((KnownOne & Mask) != 0)
1069 // The sum of a non-negative number and a power of two is not zero.
1070 if (XKnownNonNegative && isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth))
1072 if (YKnownNonNegative && isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth))
1076 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1077 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1078 // If X and Y are non-zero then so is X * Y as long as the multiplication
1079 // does not overflow.
1080 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1081 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
1084 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1085 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1086 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
1087 isKnownNonZero(SI->getFalseValue(), TD, Depth))
1091 if (!BitWidth) return false;
1092 APInt KnownZero(BitWidth, 0);
1093 APInt KnownOne(BitWidth, 0);
1094 computeKnownBits(V, KnownZero, KnownOne, TD, Depth);
1095 return KnownOne != 0;
1098 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1099 /// this predicate to simplify operations downstream. Mask is known to be zero
1100 /// for bits that V cannot have.
1102 /// This function is defined on values with integer type, values with pointer
1103 /// type (but only if TD is non-null), and vectors of integers. In the case
1104 /// where V is a vector, the mask, known zero, and known one values are the
1105 /// same width as the vector element, and the bit is set only if it is true
1106 /// for all of the elements in the vector.
1107 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
1108 const DataLayout *TD, unsigned Depth) {
1109 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1110 computeKnownBits(V, KnownZero, KnownOne, TD, Depth);
1111 return (KnownZero & Mask) == Mask;
1116 /// ComputeNumSignBits - Return the number of times the sign bit of the
1117 /// register is replicated into the other bits. We know that at least 1 bit
1118 /// is always equal to the sign bit (itself), but other cases can give us
1119 /// information. For example, immediately after an "ashr X, 2", we know that
1120 /// the top 3 bits are all equal to each other, so we return 3.
1122 /// 'Op' must have a scalar integer type.
1124 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
1126 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1127 "ComputeNumSignBits requires a DataLayout object to operate "
1128 "on non-integer values!");
1129 Type *Ty = V->getType();
1130 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1131 Ty->getScalarSizeInBits();
1133 unsigned FirstAnswer = 1;
1135 // Note that ConstantInt is handled by the general computeKnownBits case
1139 return 1; // Limit search depth.
1141 Operator *U = dyn_cast<Operator>(V);
1142 switch (Operator::getOpcode(V)) {
1144 case Instruction::SExt:
1145 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1146 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
1148 case Instruction::AShr: {
1149 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1150 // ashr X, C -> adds C sign bits. Vectors too.
1152 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1153 Tmp += ShAmt->getZExtValue();
1154 if (Tmp > TyBits) Tmp = TyBits;
1158 case Instruction::Shl: {
1160 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1161 // shl destroys sign bits.
1162 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1163 Tmp2 = ShAmt->getZExtValue();
1164 if (Tmp2 >= TyBits || // Bad shift.
1165 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1170 case Instruction::And:
1171 case Instruction::Or:
1172 case Instruction::Xor: // NOT is handled here.
1173 // Logical binary ops preserve the number of sign bits at the worst.
1174 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1176 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1177 FirstAnswer = std::min(Tmp, Tmp2);
1178 // We computed what we know about the sign bits as our first
1179 // answer. Now proceed to the generic code that uses
1180 // computeKnownBits, and pick whichever answer is better.
1184 case Instruction::Select:
1185 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1186 if (Tmp == 1) return 1; // Early out.
1187 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1188 return std::min(Tmp, Tmp2);
1190 case Instruction::Add:
1191 // Add can have at most one carry bit. Thus we know that the output
1192 // is, at worst, one more bit than the inputs.
1193 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1194 if (Tmp == 1) return 1; // Early out.
1196 // Special case decrementing a value (ADD X, -1):
1197 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1198 if (CRHS->isAllOnesValue()) {
1199 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1200 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
1202 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1204 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1207 // If we are subtracting one from a positive number, there is no carry
1208 // out of the result.
1209 if (KnownZero.isNegative())
1213 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1214 if (Tmp2 == 1) return 1;
1215 return std::min(Tmp, Tmp2)-1;
1217 case Instruction::Sub:
1218 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1219 if (Tmp2 == 1) return 1;
1222 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1223 if (CLHS->isNullValue()) {
1224 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1225 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
1226 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1228 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1231 // If the input is known to be positive (the sign bit is known clear),
1232 // the output of the NEG has the same number of sign bits as the input.
1233 if (KnownZero.isNegative())
1236 // Otherwise, we treat this like a SUB.
1239 // Sub can have at most one carry bit. Thus we know that the output
1240 // is, at worst, one more bit than the inputs.
1241 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1242 if (Tmp == 1) return 1; // Early out.
1243 return std::min(Tmp, Tmp2)-1;
1245 case Instruction::PHI: {
1246 PHINode *PN = cast<PHINode>(U);
1247 // Don't analyze large in-degree PHIs.
1248 if (PN->getNumIncomingValues() > 4) break;
1250 // Take the minimum of all incoming values. This can't infinitely loop
1251 // because of our depth threshold.
1252 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1253 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1254 if (Tmp == 1) return Tmp;
1256 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1261 case Instruction::Trunc:
1262 // FIXME: it's tricky to do anything useful for this, but it is an important
1263 // case for targets like X86.
1267 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1268 // use this information.
1269 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1271 computeKnownBits(V, KnownZero, KnownOne, TD, Depth);
1273 if (KnownZero.isNegative()) { // sign bit is 0
1275 } else if (KnownOne.isNegative()) { // sign bit is 1;
1282 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1283 // the number of identical bits in the top of the input value.
1285 Mask <<= Mask.getBitWidth()-TyBits;
1286 // Return # leading zeros. We use 'min' here in case Val was zero before
1287 // shifting. We don't want to return '64' as for an i32 "0".
1288 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1291 /// ComputeMultiple - This function computes the integer multiple of Base that
1292 /// equals V. If successful, it returns true and returns the multiple in
1293 /// Multiple. If unsuccessful, it returns false. It looks
1294 /// through SExt instructions only if LookThroughSExt is true.
1295 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1296 bool LookThroughSExt, unsigned Depth) {
1297 const unsigned MaxDepth = 6;
1299 assert(V && "No Value?");
1300 assert(Depth <= MaxDepth && "Limit Search Depth");
1301 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1303 Type *T = V->getType();
1305 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1315 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1316 Constant *BaseVal = ConstantInt::get(T, Base);
1317 if (CO && CO == BaseVal) {
1319 Multiple = ConstantInt::get(T, 1);
1323 if (CI && CI->getZExtValue() % Base == 0) {
1324 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1328 if (Depth == MaxDepth) return false; // Limit search depth.
1330 Operator *I = dyn_cast<Operator>(V);
1331 if (!I) return false;
1333 switch (I->getOpcode()) {
1335 case Instruction::SExt:
1336 if (!LookThroughSExt) return false;
1337 // otherwise fall through to ZExt
1338 case Instruction::ZExt:
1339 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1340 LookThroughSExt, Depth+1);
1341 case Instruction::Shl:
1342 case Instruction::Mul: {
1343 Value *Op0 = I->getOperand(0);
1344 Value *Op1 = I->getOperand(1);
1346 if (I->getOpcode() == Instruction::Shl) {
1347 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1348 if (!Op1CI) return false;
1349 // Turn Op0 << Op1 into Op0 * 2^Op1
1350 APInt Op1Int = Op1CI->getValue();
1351 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1352 APInt API(Op1Int.getBitWidth(), 0);
1353 API.setBit(BitToSet);
1354 Op1 = ConstantInt::get(V->getContext(), API);
1357 Value *Mul0 = nullptr;
1358 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1359 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1360 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1361 if (Op1C->getType()->getPrimitiveSizeInBits() <
1362 MulC->getType()->getPrimitiveSizeInBits())
1363 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1364 if (Op1C->getType()->getPrimitiveSizeInBits() >
1365 MulC->getType()->getPrimitiveSizeInBits())
1366 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1368 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1369 Multiple = ConstantExpr::getMul(MulC, Op1C);
1373 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1374 if (Mul0CI->getValue() == 1) {
1375 // V == Base * Op1, so return Op1
1381 Value *Mul1 = nullptr;
1382 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1383 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1384 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1385 if (Op0C->getType()->getPrimitiveSizeInBits() <
1386 MulC->getType()->getPrimitiveSizeInBits())
1387 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1388 if (Op0C->getType()->getPrimitiveSizeInBits() >
1389 MulC->getType()->getPrimitiveSizeInBits())
1390 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1392 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1393 Multiple = ConstantExpr::getMul(MulC, Op0C);
1397 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1398 if (Mul1CI->getValue() == 1) {
1399 // V == Base * Op0, so return Op0
1407 // We could not determine if V is a multiple of Base.
1411 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1412 /// value is never equal to -0.0.
1414 /// NOTE: this function will need to be revisited when we support non-default
1417 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1418 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1419 return !CFP->getValueAPF().isNegZero();
1422 return 1; // Limit search depth.
1424 const Operator *I = dyn_cast<Operator>(V);
1425 if (!I) return false;
1427 // Check if the nsz fast-math flag is set
1428 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
1429 if (FPO->hasNoSignedZeros())
1432 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1433 if (I->getOpcode() == Instruction::FAdd)
1434 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
1435 if (CFP->isNullValue())
1438 // sitofp and uitofp turn into +0.0 for zero.
1439 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1442 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1443 // sqrt(-0.0) = -0.0, no other negative results are possible.
1444 if (II->getIntrinsicID() == Intrinsic::sqrt)
1445 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1447 if (const CallInst *CI = dyn_cast<CallInst>(I))
1448 if (const Function *F = CI->getCalledFunction()) {
1449 if (F->isDeclaration()) {
1451 if (F->getName() == "abs") return true;
1452 // fabs[lf](x) != -0.0
1453 if (F->getName() == "fabs") return true;
1454 if (F->getName() == "fabsf") return true;
1455 if (F->getName() == "fabsl") return true;
1456 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1457 F->getName() == "sqrtl")
1458 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1465 /// isBytewiseValue - If the specified value can be set by repeating the same
1466 /// byte in memory, return the i8 value that it is represented with. This is
1467 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1468 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1469 /// byte store (e.g. i16 0x1234), return null.
1470 Value *llvm::isBytewiseValue(Value *V) {
1471 // All byte-wide stores are splatable, even of arbitrary variables.
1472 if (V->getType()->isIntegerTy(8)) return V;
1474 // Handle 'null' ConstantArrayZero etc.
1475 if (Constant *C = dyn_cast<Constant>(V))
1476 if (C->isNullValue())
1477 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1479 // Constant float and double values can be handled as integer values if the
1480 // corresponding integer value is "byteable". An important case is 0.0.
1481 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1482 if (CFP->getType()->isFloatTy())
1483 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1484 if (CFP->getType()->isDoubleTy())
1485 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1486 // Don't handle long double formats, which have strange constraints.
1489 // We can handle constant integers that are power of two in size and a
1490 // multiple of 8 bits.
1491 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1492 unsigned Width = CI->getBitWidth();
1493 if (isPowerOf2_32(Width) && Width > 8) {
1494 // We can handle this value if the recursive binary decomposition is the
1495 // same at all levels.
1496 APInt Val = CI->getValue();
1498 while (Val.getBitWidth() != 8) {
1499 unsigned NextWidth = Val.getBitWidth()/2;
1500 Val2 = Val.lshr(NextWidth);
1501 Val2 = Val2.trunc(Val.getBitWidth()/2);
1502 Val = Val.trunc(Val.getBitWidth()/2);
1504 // If the top/bottom halves aren't the same, reject it.
1508 return ConstantInt::get(V->getContext(), Val);
1512 // A ConstantDataArray/Vector is splatable if all its members are equal and
1514 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1515 Value *Elt = CA->getElementAsConstant(0);
1516 Value *Val = isBytewiseValue(Elt);
1520 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1521 if (CA->getElementAsConstant(I) != Elt)
1527 // Conceptually, we could handle things like:
1528 // %a = zext i8 %X to i16
1529 // %b = shl i16 %a, 8
1530 // %c = or i16 %a, %b
1531 // but until there is an example that actually needs this, it doesn't seem
1532 // worth worrying about.
1537 // This is the recursive version of BuildSubAggregate. It takes a few different
1538 // arguments. Idxs is the index within the nested struct From that we are
1539 // looking at now (which is of type IndexedType). IdxSkip is the number of
1540 // indices from Idxs that should be left out when inserting into the resulting
1541 // struct. To is the result struct built so far, new insertvalue instructions
1543 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1544 SmallVectorImpl<unsigned> &Idxs,
1546 Instruction *InsertBefore) {
1547 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
1549 // Save the original To argument so we can modify it
1551 // General case, the type indexed by Idxs is a struct
1552 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1553 // Process each struct element recursively
1556 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1560 // Couldn't find any inserted value for this index? Cleanup
1561 while (PrevTo != OrigTo) {
1562 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1563 PrevTo = Del->getAggregateOperand();
1564 Del->eraseFromParent();
1566 // Stop processing elements
1570 // If we successfully found a value for each of our subaggregates
1574 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1575 // the struct's elements had a value that was inserted directly. In the latter
1576 // case, perhaps we can't determine each of the subelements individually, but
1577 // we might be able to find the complete struct somewhere.
1579 // Find the value that is at that particular spot
1580 Value *V = FindInsertedValue(From, Idxs);
1585 // Insert the value in the new (sub) aggregrate
1586 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1587 "tmp", InsertBefore);
1590 // This helper takes a nested struct and extracts a part of it (which is again a
1591 // struct) into a new value. For example, given the struct:
1592 // { a, { b, { c, d }, e } }
1593 // and the indices "1, 1" this returns
1596 // It does this by inserting an insertvalue for each element in the resulting
1597 // struct, as opposed to just inserting a single struct. This will only work if
1598 // each of the elements of the substruct are known (ie, inserted into From by an
1599 // insertvalue instruction somewhere).
1601 // All inserted insertvalue instructions are inserted before InsertBefore
1602 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1603 Instruction *InsertBefore) {
1604 assert(InsertBefore && "Must have someplace to insert!");
1605 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1607 Value *To = UndefValue::get(IndexedType);
1608 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1609 unsigned IdxSkip = Idxs.size();
1611 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1614 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1615 /// the scalar value indexed is already around as a register, for example if it
1616 /// were inserted directly into the aggregrate.
1618 /// If InsertBefore is not null, this function will duplicate (modified)
1619 /// insertvalues when a part of a nested struct is extracted.
1620 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1621 Instruction *InsertBefore) {
1622 // Nothing to index? Just return V then (this is useful at the end of our
1624 if (idx_range.empty())
1626 // We have indices, so V should have an indexable type.
1627 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1628 "Not looking at a struct or array?");
1629 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1630 "Invalid indices for type?");
1632 if (Constant *C = dyn_cast<Constant>(V)) {
1633 C = C->getAggregateElement(idx_range[0]);
1634 if (!C) return nullptr;
1635 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1638 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1639 // Loop the indices for the insertvalue instruction in parallel with the
1640 // requested indices
1641 const unsigned *req_idx = idx_range.begin();
1642 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1643 i != e; ++i, ++req_idx) {
1644 if (req_idx == idx_range.end()) {
1645 // We can't handle this without inserting insertvalues
1649 // The requested index identifies a part of a nested aggregate. Handle
1650 // this specially. For example,
1651 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1652 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1653 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1654 // This can be changed into
1655 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1656 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1657 // which allows the unused 0,0 element from the nested struct to be
1659 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1663 // This insert value inserts something else than what we are looking for.
1664 // See if the (aggregrate) value inserted into has the value we are
1665 // looking for, then.
1667 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1670 // If we end up here, the indices of the insertvalue match with those
1671 // requested (though possibly only partially). Now we recursively look at
1672 // the inserted value, passing any remaining indices.
1673 return FindInsertedValue(I->getInsertedValueOperand(),
1674 makeArrayRef(req_idx, idx_range.end()),
1678 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1679 // If we're extracting a value from an aggregrate that was extracted from
1680 // something else, we can extract from that something else directly instead.
1681 // However, we will need to chain I's indices with the requested indices.
1683 // Calculate the number of indices required
1684 unsigned size = I->getNumIndices() + idx_range.size();
1685 // Allocate some space to put the new indices in
1686 SmallVector<unsigned, 5> Idxs;
1688 // Add indices from the extract value instruction
1689 Idxs.append(I->idx_begin(), I->idx_end());
1691 // Add requested indices
1692 Idxs.append(idx_range.begin(), idx_range.end());
1694 assert(Idxs.size() == size
1695 && "Number of indices added not correct?");
1697 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1699 // Otherwise, we don't know (such as, extracting from a function return value
1700 // or load instruction)
1704 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1705 /// it can be expressed as a base pointer plus a constant offset. Return the
1706 /// base and offset to the caller.
1707 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1708 const DataLayout *DL) {
1709 // Without DataLayout, conservatively assume 64-bit offsets, which is
1710 // the widest we support.
1711 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(Ptr->getType()) : 64;
1712 APInt ByteOffset(BitWidth, 0);
1714 if (Ptr->getType()->isVectorTy())
1717 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1719 APInt GEPOffset(BitWidth, 0);
1720 if (!GEP->accumulateConstantOffset(*DL, GEPOffset))
1723 ByteOffset += GEPOffset;
1726 Ptr = GEP->getPointerOperand();
1727 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1728 Ptr = cast<Operator>(Ptr)->getOperand(0);
1729 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1730 if (GA->mayBeOverridden())
1732 Ptr = GA->getAliasee();
1737 Offset = ByteOffset.getSExtValue();
1742 /// getConstantStringInfo - This function computes the length of a
1743 /// null-terminated C string pointed to by V. If successful, it returns true
1744 /// and returns the string in Str. If unsuccessful, it returns false.
1745 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
1746 uint64_t Offset, bool TrimAtNul) {
1749 // Look through bitcast instructions and geps.
1750 V = V->stripPointerCasts();
1752 // If the value is a GEP instructionor constant expression, treat it as an
1754 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1755 // Make sure the GEP has exactly three arguments.
1756 if (GEP->getNumOperands() != 3)
1759 // Make sure the index-ee is a pointer to array of i8.
1760 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1761 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1762 if (!AT || !AT->getElementType()->isIntegerTy(8))
1765 // Check to make sure that the first operand of the GEP is an integer and
1766 // has value 0 so that we are sure we're indexing into the initializer.
1767 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1768 if (!FirstIdx || !FirstIdx->isZero())
1771 // If the second index isn't a ConstantInt, then this is a variable index
1772 // into the array. If this occurs, we can't say anything meaningful about
1774 uint64_t StartIdx = 0;
1775 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1776 StartIdx = CI->getZExtValue();
1779 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
1782 // The GEP instruction, constant or instruction, must reference a global
1783 // variable that is a constant and is initialized. The referenced constant
1784 // initializer is the array that we'll use for optimization.
1785 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
1786 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1789 // Handle the all-zeros case
1790 if (GV->getInitializer()->isNullValue()) {
1791 // This is a degenerate case. The initializer is constant zero so the
1792 // length of the string must be zero.
1797 // Must be a Constant Array
1798 const ConstantDataArray *Array =
1799 dyn_cast<ConstantDataArray>(GV->getInitializer());
1800 if (!Array || !Array->isString())
1803 // Get the number of elements in the array
1804 uint64_t NumElts = Array->getType()->getArrayNumElements();
1806 // Start out with the entire array in the StringRef.
1807 Str = Array->getAsString();
1809 if (Offset > NumElts)
1812 // Skip over 'offset' bytes.
1813 Str = Str.substr(Offset);
1816 // Trim off the \0 and anything after it. If the array is not nul
1817 // terminated, we just return the whole end of string. The client may know
1818 // some other way that the string is length-bound.
1819 Str = Str.substr(0, Str.find('\0'));
1824 // These next two are very similar to the above, but also look through PHI
1826 // TODO: See if we can integrate these two together.
1828 /// GetStringLengthH - If we can compute the length of the string pointed to by
1829 /// the specified pointer, return 'len+1'. If we can't, return 0.
1830 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1831 // Look through noop bitcast instructions.
1832 V = V->stripPointerCasts();
1834 // If this is a PHI node, there are two cases: either we have already seen it
1836 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1837 if (!PHIs.insert(PN))
1838 return ~0ULL; // already in the set.
1840 // If it was new, see if all the input strings are the same length.
1841 uint64_t LenSoFar = ~0ULL;
1842 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1843 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1844 if (Len == 0) return 0; // Unknown length -> unknown.
1846 if (Len == ~0ULL) continue;
1848 if (Len != LenSoFar && LenSoFar != ~0ULL)
1849 return 0; // Disagree -> unknown.
1853 // Success, all agree.
1857 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1858 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1859 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1860 if (Len1 == 0) return 0;
1861 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1862 if (Len2 == 0) return 0;
1863 if (Len1 == ~0ULL) return Len2;
1864 if (Len2 == ~0ULL) return Len1;
1865 if (Len1 != Len2) return 0;
1869 // Otherwise, see if we can read the string.
1871 if (!getConstantStringInfo(V, StrData))
1874 return StrData.size()+1;
1877 /// GetStringLength - If we can compute the length of the string pointed to by
1878 /// the specified pointer, return 'len+1'. If we can't, return 0.
1879 uint64_t llvm::GetStringLength(Value *V) {
1880 if (!V->getType()->isPointerTy()) return 0;
1882 SmallPtrSet<PHINode*, 32> PHIs;
1883 uint64_t Len = GetStringLengthH(V, PHIs);
1884 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1885 // an empty string as a length.
1886 return Len == ~0ULL ? 1 : Len;
1890 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
1891 if (!V->getType()->isPointerTy())
1893 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1894 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1895 V = GEP->getPointerOperand();
1896 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1897 V = cast<Operator>(V)->getOperand(0);
1898 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1899 if (GA->mayBeOverridden())
1901 V = GA->getAliasee();
1903 // See if InstructionSimplify knows any relevant tricks.
1904 if (Instruction *I = dyn_cast<Instruction>(V))
1905 // TODO: Acquire a DominatorTree and use it.
1906 if (Value *Simplified = SimplifyInstruction(I, TD, nullptr)) {
1913 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1919 llvm::GetUnderlyingObjects(Value *V,
1920 SmallVectorImpl<Value *> &Objects,
1921 const DataLayout *TD,
1922 unsigned MaxLookup) {
1923 SmallPtrSet<Value *, 4> Visited;
1924 SmallVector<Value *, 4> Worklist;
1925 Worklist.push_back(V);
1927 Value *P = Worklist.pop_back_val();
1928 P = GetUnderlyingObject(P, TD, MaxLookup);
1930 if (!Visited.insert(P))
1933 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
1934 Worklist.push_back(SI->getTrueValue());
1935 Worklist.push_back(SI->getFalseValue());
1939 if (PHINode *PN = dyn_cast<PHINode>(P)) {
1940 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1941 Worklist.push_back(PN->getIncomingValue(i));
1945 Objects.push_back(P);
1946 } while (!Worklist.empty());
1949 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1950 /// are lifetime markers.
1952 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1953 for (const User *U : V->users()) {
1954 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1955 if (!II) return false;
1957 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1958 II->getIntrinsicID() != Intrinsic::lifetime_end)
1964 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1965 const DataLayout *TD) {
1966 const Operator *Inst = dyn_cast<Operator>(V);
1970 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1971 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1975 switch (Inst->getOpcode()) {
1978 case Instruction::UDiv:
1979 case Instruction::URem:
1980 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1981 return isKnownNonZero(Inst->getOperand(1), TD);
1982 case Instruction::SDiv:
1983 case Instruction::SRem: {
1984 Value *Op = Inst->getOperand(1);
1985 // x / y is undefined if y == 0
1986 if (!isKnownNonZero(Op, TD))
1988 // x / y might be undefined if y == -1
1989 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1992 APInt KnownZero(BitWidth, 0);
1993 APInt KnownOne(BitWidth, 0);
1994 computeKnownBits(Op, KnownZero, KnownOne, TD);
1997 case Instruction::Load: {
1998 const LoadInst *LI = cast<LoadInst>(Inst);
1999 if (!LI->isUnordered() ||
2000 // Speculative load may create a race that did not exist in the source.
2001 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2003 return LI->getPointerOperand()->isDereferenceablePointer();
2005 case Instruction::Call: {
2006 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2007 switch (II->getIntrinsicID()) {
2008 // These synthetic intrinsics have no side-effects, and just mark
2009 // information about their operands.
2010 // FIXME: There are other no-op synthetic instructions that potentially
2011 // should be considered at least *safe* to speculate...
2012 case Intrinsic::dbg_declare:
2013 case Intrinsic::dbg_value:
2016 case Intrinsic::bswap:
2017 case Intrinsic::ctlz:
2018 case Intrinsic::ctpop:
2019 case Intrinsic::cttz:
2020 case Intrinsic::objectsize:
2021 case Intrinsic::sadd_with_overflow:
2022 case Intrinsic::smul_with_overflow:
2023 case Intrinsic::ssub_with_overflow:
2024 case Intrinsic::uadd_with_overflow:
2025 case Intrinsic::umul_with_overflow:
2026 case Intrinsic::usub_with_overflow:
2028 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2029 // errno like libm sqrt would.
2030 case Intrinsic::sqrt:
2031 case Intrinsic::fma:
2032 case Intrinsic::fmuladd:
2034 // TODO: some fp intrinsics are marked as having the same error handling
2035 // as libm. They're safe to speculate when they won't error.
2036 // TODO: are convert_{from,to}_fp16 safe?
2037 // TODO: can we list target-specific intrinsics here?
2041 return false; // The called function could have undefined behavior or
2042 // side-effects, even if marked readnone nounwind.
2044 case Instruction::VAArg:
2045 case Instruction::Alloca:
2046 case Instruction::Invoke:
2047 case Instruction::PHI:
2048 case Instruction::Store:
2049 case Instruction::Ret:
2050 case Instruction::Br:
2051 case Instruction::IndirectBr:
2052 case Instruction::Switch:
2053 case Instruction::Unreachable:
2054 case Instruction::Fence:
2055 case Instruction::LandingPad:
2056 case Instruction::AtomicRMW:
2057 case Instruction::AtomicCmpXchg:
2058 case Instruction::Resume:
2059 return false; // Misc instructions which have effects
2063 /// isKnownNonNull - Return true if we know that the specified value is never
2065 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2066 // Alloca never returns null, malloc might.
2067 if (isa<AllocaInst>(V)) return true;
2069 // A byval, inalloca, or nonnull argument is never null.
2070 if (const Argument *A = dyn_cast<Argument>(V))
2071 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
2073 // Global values are not null unless extern weak.
2074 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2075 return !GV->hasExternalWeakLinkage();
2077 if (ImmutableCallSite CS = V)
2078 if (CS.paramHasAttr(0, Attribute::NonNull))
2081 // operator new never returns null.
2082 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))