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/ConstantRange.h"
20 #include "llvm/IR/Constants.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/GetElementPtrTypeIterator.h"
23 #include "llvm/IR/GlobalAlias.h"
24 #include "llvm/IR/GlobalVariable.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/LLVMContext.h"
28 #include "llvm/IR/Metadata.h"
29 #include "llvm/IR/Operator.h"
30 #include "llvm/IR/PatternMatch.h"
31 #include "llvm/Support/MathExtras.h"
34 using namespace llvm::PatternMatch;
36 const unsigned MaxDepth = 6;
38 /// getBitWidth - Returns the bitwidth of the given scalar or pointer type (if
39 /// unknown returns 0). For vector types, returns the element type's bitwidth.
40 static unsigned getBitWidth(Type *Ty, const DataLayout *TD) {
41 if (unsigned BitWidth = Ty->getScalarSizeInBits())
44 return TD ? TD->getPointerTypeSizeInBits(Ty) : 0;
47 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
48 APInt &KnownZero, APInt &KnownOne,
49 APInt &KnownZero2, APInt &KnownOne2,
50 const DataLayout *TD, unsigned Depth) {
52 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
53 // We know that the top bits of C-X are clear if X contains less bits
54 // than C (i.e. no wrap-around can happen). For example, 20-X is
55 // positive if we can prove that X is >= 0 and < 16.
56 if (!CLHS->getValue().isNegative()) {
57 unsigned BitWidth = KnownZero.getBitWidth();
58 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
59 // NLZ can't be BitWidth with no sign bit
60 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
61 llvm::computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1);
63 // If all of the MaskV bits are known to be zero, then we know the
64 // output top bits are zero, because we now know that the output is
66 if ((KnownZero2 & MaskV) == MaskV) {
67 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
68 // Top bits known zero.
69 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
75 unsigned BitWidth = KnownZero.getBitWidth();
77 // If one of the operands has trailing zeros, then the bits that the
78 // other operand has in those bit positions will be preserved in the
79 // result. For an add, this works with either operand. For a subtract,
80 // this only works if the known zeros are in the right operand.
81 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
82 llvm::computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, TD, Depth+1);
83 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
85 llvm::computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1);
86 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
88 // Determine which operand has more trailing zeros, and use that
89 // many bits from the other operand.
90 if (LHSKnownZeroOut > RHSKnownZeroOut) {
92 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
93 KnownZero |= KnownZero2 & Mask;
94 KnownOne |= KnownOne2 & Mask;
96 // If the known zeros are in the left operand for a subtract,
97 // fall back to the minimum known zeros in both operands.
98 KnownZero |= APInt::getLowBitsSet(BitWidth,
99 std::min(LHSKnownZeroOut,
102 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
103 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
104 KnownZero |= LHSKnownZero & Mask;
105 KnownOne |= LHSKnownOne & Mask;
108 // Are we still trying to solve for the sign bit?
109 if (!KnownZero.isNegative() && !KnownOne.isNegative()) {
112 // Adding two positive numbers can't wrap into negative
113 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
114 KnownZero |= APInt::getSignBit(BitWidth);
115 // and adding two negative numbers can't wrap into positive.
116 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
117 KnownOne |= APInt::getSignBit(BitWidth);
119 // Subtracting a negative number from a positive one can't wrap
120 if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
121 KnownZero |= APInt::getSignBit(BitWidth);
122 // neither can subtracting a positive number from a negative one.
123 else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
124 KnownOne |= APInt::getSignBit(BitWidth);
130 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
131 APInt &KnownZero, APInt &KnownOne,
132 APInt &KnownZero2, APInt &KnownOne2,
133 const DataLayout *TD, unsigned Depth) {
134 unsigned BitWidth = KnownZero.getBitWidth();
135 computeKnownBits(Op1, KnownZero, KnownOne, TD, Depth+1);
136 computeKnownBits(Op0, KnownZero2, KnownOne2, TD, Depth+1);
138 bool isKnownNegative = false;
139 bool isKnownNonNegative = false;
140 // If the multiplication is known not to overflow, compute the sign bit.
143 // The product of a number with itself is non-negative.
144 isKnownNonNegative = true;
146 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
147 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
148 bool isKnownNegativeOp1 = KnownOne.isNegative();
149 bool isKnownNegativeOp0 = KnownOne2.isNegative();
150 // The product of two numbers with the same sign is non-negative.
151 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
152 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
153 // The product of a negative number and a non-negative number is either
155 if (!isKnownNonNegative)
156 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
157 isKnownNonZero(Op0, TD, Depth)) ||
158 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
159 isKnownNonZero(Op1, TD, Depth));
163 // If low bits are zero in either operand, output low known-0 bits.
164 // Also compute a conserative estimate for high known-0 bits.
165 // More trickiness is possible, but this is sufficient for the
166 // interesting case of alignment computation.
167 KnownOne.clearAllBits();
168 unsigned TrailZ = KnownZero.countTrailingOnes() +
169 KnownZero2.countTrailingOnes();
170 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
171 KnownZero2.countLeadingOnes(),
172 BitWidth) - BitWidth;
174 TrailZ = std::min(TrailZ, BitWidth);
175 LeadZ = std::min(LeadZ, BitWidth);
176 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
177 APInt::getHighBitsSet(BitWidth, LeadZ);
179 // Only make use of no-wrap flags if we failed to compute the sign bit
180 // directly. This matters if the multiplication always overflows, in
181 // which case we prefer to follow the result of the direct computation,
182 // though as the program is invoking undefined behaviour we can choose
183 // whatever we like here.
184 if (isKnownNonNegative && !KnownOne.isNegative())
185 KnownZero.setBit(BitWidth - 1);
186 else if (isKnownNegative && !KnownZero.isNegative())
187 KnownOne.setBit(BitWidth - 1);
190 void llvm::computeKnownBitsLoad(const MDNode &Ranges, APInt &KnownZero) {
191 unsigned BitWidth = KnownZero.getBitWidth();
192 unsigned NumRanges = Ranges.getNumOperands() / 2;
193 assert(NumRanges >= 1);
195 // Use the high end of the ranges to find leading zeros.
196 unsigned MinLeadingZeros = BitWidth;
197 for (unsigned i = 0; i < NumRanges; ++i) {
198 ConstantInt *Lower = cast<ConstantInt>(Ranges.getOperand(2*i + 0));
199 ConstantInt *Upper = cast<ConstantInt>(Ranges.getOperand(2*i + 1));
200 ConstantRange Range(Lower->getValue(), Upper->getValue());
201 if (Range.isWrappedSet())
202 MinLeadingZeros = 0; // -1 has no zeros
203 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
204 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
207 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
210 /// Determine which bits of V are known to be either zero or one and return
211 /// them in the KnownZero/KnownOne bit sets.
213 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
214 /// we cannot optimize based on the assumption that it is zero without changing
215 /// it to be an explicit zero. If we don't change it to zero, other code could
216 /// optimized based on the contradictory assumption that it is non-zero.
217 /// Because instcombine aggressively folds operations with undef args anyway,
218 /// this won't lose us code quality.
220 /// This function is defined on values with integer type, values with pointer
221 /// type (but only if TD is non-null), and vectors of integers. In the case
222 /// where V is a vector, known zero, and known one values are the
223 /// same width as the vector element, and the bit is set only if it is true
224 /// for all of the elements in the vector.
225 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
226 const DataLayout *TD, unsigned Depth) {
227 assert(V && "No Value?");
228 assert(Depth <= MaxDepth && "Limit Search Depth");
229 unsigned BitWidth = KnownZero.getBitWidth();
231 assert((V->getType()->isIntOrIntVectorTy() ||
232 V->getType()->getScalarType()->isPointerTy()) &&
233 "Not integer or pointer type!");
235 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
236 (!V->getType()->isIntOrIntVectorTy() ||
237 V->getType()->getScalarSizeInBits() == BitWidth) &&
238 KnownZero.getBitWidth() == BitWidth &&
239 KnownOne.getBitWidth() == BitWidth &&
240 "V, KnownOne and KnownZero should have same BitWidth");
242 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
243 // We know all of the bits for a constant!
244 KnownOne = CI->getValue();
245 KnownZero = ~KnownOne;
248 // Null and aggregate-zero are all-zeros.
249 if (isa<ConstantPointerNull>(V) ||
250 isa<ConstantAggregateZero>(V)) {
251 KnownOne.clearAllBits();
252 KnownZero = APInt::getAllOnesValue(BitWidth);
255 // Handle a constant vector by taking the intersection of the known bits of
256 // each element. There is no real need to handle ConstantVector here, because
257 // we don't handle undef in any particularly useful way.
258 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
259 // We know that CDS must be a vector of integers. Take the intersection of
261 KnownZero.setAllBits(); KnownOne.setAllBits();
262 APInt Elt(KnownZero.getBitWidth(), 0);
263 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
264 Elt = CDS->getElementAsInteger(i);
271 // The address of an aligned GlobalValue has trailing zeros.
272 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
273 unsigned Align = GV->getAlignment();
274 if (Align == 0 && TD) {
275 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
276 Type *ObjectType = GVar->getType()->getElementType();
277 if (ObjectType->isSized()) {
278 // If the object is defined in the current Module, we'll be giving
279 // it the preferred alignment. Otherwise, we have to assume that it
280 // may only have the minimum ABI alignment.
281 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
282 Align = TD->getPreferredAlignment(GVar);
284 Align = TD->getABITypeAlignment(ObjectType);
289 KnownZero = APInt::getLowBitsSet(BitWidth,
290 countTrailingZeros(Align));
292 KnownZero.clearAllBits();
293 KnownOne.clearAllBits();
296 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
297 // the bits of its aliasee.
298 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
299 if (GA->mayBeOverridden()) {
300 KnownZero.clearAllBits(); KnownOne.clearAllBits();
302 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth+1);
307 if (Argument *A = dyn_cast<Argument>(V)) {
310 if (A->hasByValOrInAllocaAttr()) {
311 // Get alignment information off byval/inalloca arguments if specified in
313 Align = A->getParamAlignment();
314 } else if (TD && A->hasStructRetAttr()) {
315 // An sret parameter has at least the ABI alignment of the return type.
316 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
317 if (EltTy->isSized())
318 Align = TD->getABITypeAlignment(EltTy);
322 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
326 // Start out not knowing anything.
327 KnownZero.clearAllBits(); KnownOne.clearAllBits();
329 if (Depth == MaxDepth)
330 return; // Limit search depth.
332 Operator *I = dyn_cast<Operator>(V);
335 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
336 switch (I->getOpcode()) {
338 case Instruction::Load:
339 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
340 computeKnownBitsLoad(*MD, KnownZero);
342 case Instruction::And: {
343 // If either the LHS or the RHS are Zero, the result is zero.
344 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
345 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
347 // Output known-1 bits are only known if set in both the LHS & RHS.
348 KnownOne &= KnownOne2;
349 // Output known-0 are known to be clear if zero in either the LHS | RHS.
350 KnownZero |= KnownZero2;
353 case Instruction::Or: {
354 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
355 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
357 // Output known-0 bits are only known if clear in both the LHS & RHS.
358 KnownZero &= KnownZero2;
359 // Output known-1 are known to be set if set in either the LHS | RHS.
360 KnownOne |= KnownOne2;
363 case Instruction::Xor: {
364 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
365 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
367 // Output known-0 bits are known if clear or set in both the LHS & RHS.
368 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
369 // Output known-1 are known to be set if set in only one of the LHS, RHS.
370 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
371 KnownZero = KnownZeroOut;
374 case Instruction::Mul: {
375 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
376 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW,
377 KnownZero, KnownOne, KnownZero2, KnownOne2, TD, Depth);
380 case Instruction::UDiv: {
381 // For the purposes of computing leading zeros we can conservatively
382 // treat a udiv as a logical right shift by the power of 2 known to
383 // be less than the denominator.
384 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
385 unsigned LeadZ = KnownZero2.countLeadingOnes();
387 KnownOne2.clearAllBits();
388 KnownZero2.clearAllBits();
389 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
390 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
391 if (RHSUnknownLeadingOnes != BitWidth)
392 LeadZ = std::min(BitWidth,
393 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
395 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
398 case Instruction::Select:
399 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1);
400 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD,
403 // Only known if known in both the LHS and RHS.
404 KnownOne &= KnownOne2;
405 KnownZero &= KnownZero2;
407 case Instruction::FPTrunc:
408 case Instruction::FPExt:
409 case Instruction::FPToUI:
410 case Instruction::FPToSI:
411 case Instruction::SIToFP:
412 case Instruction::UIToFP:
413 break; // Can't work with floating point.
414 case Instruction::PtrToInt:
415 case Instruction::IntToPtr:
416 // We can't handle these if we don't know the pointer size.
418 // FALL THROUGH and handle them the same as zext/trunc.
419 case Instruction::ZExt:
420 case Instruction::Trunc: {
421 Type *SrcTy = I->getOperand(0)->getType();
423 unsigned SrcBitWidth;
424 // Note that we handle pointer operands here because of inttoptr/ptrtoint
425 // which fall through here.
427 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
429 SrcBitWidth = SrcTy->getScalarSizeInBits();
430 if (!SrcBitWidth) break;
433 assert(SrcBitWidth && "SrcBitWidth can't be zero");
434 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
435 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
436 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
437 KnownZero = KnownZero.zextOrTrunc(BitWidth);
438 KnownOne = KnownOne.zextOrTrunc(BitWidth);
439 // Any top bits are known to be zero.
440 if (BitWidth > SrcBitWidth)
441 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
444 case Instruction::BitCast: {
445 Type *SrcTy = I->getOperand(0)->getType();
446 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
447 // TODO: For now, not handling conversions like:
448 // (bitcast i64 %x to <2 x i32>)
449 !I->getType()->isVectorTy()) {
450 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
455 case Instruction::SExt: {
456 // Compute the bits in the result that are not present in the input.
457 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
459 KnownZero = KnownZero.trunc(SrcBitWidth);
460 KnownOne = KnownOne.trunc(SrcBitWidth);
461 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
462 KnownZero = KnownZero.zext(BitWidth);
463 KnownOne = KnownOne.zext(BitWidth);
465 // If the sign bit of the input is known set or clear, then we know the
466 // top bits of the result.
467 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
468 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
469 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
470 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
473 case Instruction::Shl:
474 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
475 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
476 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
477 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
478 KnownZero <<= ShiftAmt;
479 KnownOne <<= ShiftAmt;
480 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
484 case Instruction::LShr:
485 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
486 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
487 // Compute the new bits that are at the top now.
488 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
490 // Unsigned shift right.
491 computeKnownBits(I->getOperand(0), KnownZero,KnownOne, TD, Depth+1);
492 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
493 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
494 // high bits known zero.
495 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
499 case Instruction::AShr:
500 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
501 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
502 // Compute the new bits that are at the top now.
503 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
505 // Signed shift right.
506 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
507 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
508 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
510 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
511 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
512 KnownZero |= HighBits;
513 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
514 KnownOne |= HighBits;
518 case Instruction::Sub: {
519 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
520 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
521 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
525 case Instruction::Add: {
526 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
527 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
528 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
532 case Instruction::SRem:
533 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
534 APInt RA = Rem->getValue().abs();
535 if (RA.isPowerOf2()) {
536 APInt LowBits = RA - 1;
537 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1);
539 // The low bits of the first operand are unchanged by the srem.
540 KnownZero = KnownZero2 & LowBits;
541 KnownOne = KnownOne2 & LowBits;
543 // If the first operand is non-negative or has all low bits zero, then
544 // the upper bits are all zero.
545 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
546 KnownZero |= ~LowBits;
548 // If the first operand is negative and not all low bits are zero, then
549 // the upper bits are all one.
550 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
551 KnownOne |= ~LowBits;
553 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
557 // The sign bit is the LHS's sign bit, except when the result of the
558 // remainder is zero.
559 if (KnownZero.isNonNegative()) {
560 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
561 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
563 // If it's known zero, our sign bit is also zero.
564 if (LHSKnownZero.isNegative())
565 KnownZero.setBit(BitWidth - 1);
569 case Instruction::URem: {
570 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
571 APInt RA = Rem->getValue();
572 if (RA.isPowerOf2()) {
573 APInt LowBits = (RA - 1);
574 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD,
576 KnownZero |= ~LowBits;
582 // Since the result is less than or equal to either operand, any leading
583 // zero bits in either operand must also exist in the result.
584 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
585 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1);
587 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
588 KnownZero2.countLeadingOnes());
589 KnownOne.clearAllBits();
590 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
594 case Instruction::Alloca: {
595 AllocaInst *AI = cast<AllocaInst>(V);
596 unsigned Align = AI->getAlignment();
597 if (Align == 0 && TD)
598 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
601 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(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 LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
608 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
610 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
612 gep_type_iterator GTI = gep_type_begin(I);
613 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
614 Value *Index = I->getOperand(i);
615 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
616 // Handle struct member offset arithmetic.
622 // Handle case when index is vector zeroinitializer
623 Constant *CIndex = cast<Constant>(Index);
624 if (CIndex->isZeroValue())
627 if (CIndex->getType()->isVectorTy())
628 Index = CIndex->getSplatValue();
630 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
631 const StructLayout *SL = TD->getStructLayout(STy);
632 uint64_t Offset = SL->getElementOffset(Idx);
633 TrailZ = std::min<unsigned>(TrailZ,
634 countTrailingZeros(Offset));
636 // Handle array index arithmetic.
637 Type *IndexedTy = GTI.getIndexedType();
638 if (!IndexedTy->isSized()) {
642 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
643 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
644 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
645 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1);
646 TrailZ = std::min(TrailZ,
647 unsigned(countTrailingZeros(TypeSize) +
648 LocalKnownZero.countTrailingOnes()));
652 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
655 case Instruction::PHI: {
656 PHINode *P = cast<PHINode>(I);
657 // Handle the case of a simple two-predecessor recurrence PHI.
658 // There's a lot more that could theoretically be done here, but
659 // this is sufficient to catch some interesting cases.
660 if (P->getNumIncomingValues() == 2) {
661 for (unsigned i = 0; i != 2; ++i) {
662 Value *L = P->getIncomingValue(i);
663 Value *R = P->getIncomingValue(!i);
664 Operator *LU = dyn_cast<Operator>(L);
667 unsigned Opcode = LU->getOpcode();
668 // Check for operations that have the property that if
669 // both their operands have low zero bits, the result
670 // will have low zero bits.
671 if (Opcode == Instruction::Add ||
672 Opcode == Instruction::Sub ||
673 Opcode == Instruction::And ||
674 Opcode == Instruction::Or ||
675 Opcode == Instruction::Mul) {
676 Value *LL = LU->getOperand(0);
677 Value *LR = LU->getOperand(1);
678 // Find a recurrence.
685 // Ok, we have a PHI of the form L op= R. Check for low
687 computeKnownBits(R, KnownZero2, KnownOne2, TD, Depth+1);
689 // We need to take the minimum number of known bits
690 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
691 computeKnownBits(L, KnownZero3, KnownOne3, TD, Depth+1);
693 KnownZero = APInt::getLowBitsSet(BitWidth,
694 std::min(KnownZero2.countTrailingOnes(),
695 KnownZero3.countTrailingOnes()));
701 // Unreachable blocks may have zero-operand PHI nodes.
702 if (P->getNumIncomingValues() == 0)
705 // Otherwise take the unions of the known bit sets of the operands,
706 // taking conservative care to avoid excessive recursion.
707 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
708 // Skip if every incoming value references to ourself.
709 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
712 KnownZero = APInt::getAllOnesValue(BitWidth);
713 KnownOne = APInt::getAllOnesValue(BitWidth);
714 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
715 // Skip direct self references.
716 if (P->getIncomingValue(i) == P) continue;
718 KnownZero2 = APInt(BitWidth, 0);
719 KnownOne2 = APInt(BitWidth, 0);
720 // Recurse, but cap the recursion to one level, because we don't
721 // want to waste time spinning around in loops.
722 computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
724 KnownZero &= KnownZero2;
725 KnownOne &= KnownOne2;
726 // If all bits have been ruled out, there's no need to check
728 if (!KnownZero && !KnownOne)
734 case Instruction::Call:
735 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
736 switch (II->getIntrinsicID()) {
738 case Intrinsic::ctlz:
739 case Intrinsic::cttz: {
740 unsigned LowBits = Log2_32(BitWidth)+1;
741 // If this call is undefined for 0, the result will be less than 2^n.
742 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
744 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
747 case Intrinsic::ctpop: {
748 unsigned LowBits = Log2_32(BitWidth)+1;
749 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
752 case Intrinsic::x86_sse42_crc32_64_64:
753 KnownZero = APInt::getHighBitsSet(64, 32);
758 case Instruction::ExtractValue:
759 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
760 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
761 if (EVI->getNumIndices() != 1) break;
762 if (EVI->getIndices()[0] == 0) {
763 switch (II->getIntrinsicID()) {
765 case Intrinsic::uadd_with_overflow:
766 case Intrinsic::sadd_with_overflow:
767 computeKnownBitsAddSub(true, II->getArgOperand(0),
768 II->getArgOperand(1), false, KnownZero,
769 KnownOne, KnownZero2, KnownOne2, TD, Depth);
771 case Intrinsic::usub_with_overflow:
772 case Intrinsic::ssub_with_overflow:
773 computeKnownBitsAddSub(false, II->getArgOperand(0),
774 II->getArgOperand(1), false, KnownZero,
775 KnownOne, KnownZero2, KnownOne2, TD, Depth);
777 case Intrinsic::umul_with_overflow:
778 case Intrinsic::smul_with_overflow:
779 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1),
780 false, KnownZero, KnownOne,
781 KnownZero2, KnownOne2, TD, Depth);
788 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
791 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
792 /// one. Convenience wrapper around computeKnownBits.
793 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
794 const DataLayout *TD, unsigned Depth) {
795 unsigned BitWidth = getBitWidth(V->getType(), TD);
801 APInt ZeroBits(BitWidth, 0);
802 APInt OneBits(BitWidth, 0);
803 computeKnownBits(V, ZeroBits, OneBits, TD, Depth);
804 KnownOne = OneBits[BitWidth - 1];
805 KnownZero = ZeroBits[BitWidth - 1];
808 /// isKnownToBeAPowerOfTwo - Return true if the given value is known to have exactly one
809 /// bit set when defined. For vectors return true if every element is known to
810 /// be a power of two when defined. Supports values with integer or pointer
811 /// types and vectors of integers.
812 bool llvm::isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth) {
813 if (Constant *C = dyn_cast<Constant>(V)) {
814 if (C->isNullValue())
816 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
817 return CI->getValue().isPowerOf2();
818 // TODO: Handle vector constants.
821 // 1 << X is clearly a power of two if the one is not shifted off the end. If
822 // it is shifted off the end then the result is undefined.
823 if (match(V, m_Shl(m_One(), m_Value())))
826 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
827 // bottom. If it is shifted off the bottom then the result is undefined.
828 if (match(V, m_LShr(m_SignBit(), m_Value())))
831 // The remaining tests are all recursive, so bail out if we hit the limit.
832 if (Depth++ == MaxDepth)
835 Value *X = nullptr, *Y = nullptr;
836 // A shift of a power of two is a power of two or zero.
837 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
838 match(V, m_Shr(m_Value(X), m_Value()))))
839 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth);
841 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
842 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth);
844 if (SelectInst *SI = dyn_cast<SelectInst>(V))
845 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth) &&
846 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth);
848 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
849 // A power of two and'd with anything is a power of two or zero.
850 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth) ||
851 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth))
853 // X & (-X) is always a power of two or zero.
854 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
859 // Adding a power-of-two or zero to the same power-of-two or zero yields
860 // either the original power-of-two, a larger power-of-two or zero.
861 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
862 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
863 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
864 if (match(X, m_And(m_Specific(Y), m_Value())) ||
865 match(X, m_And(m_Value(), m_Specific(Y))))
866 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth))
868 if (match(Y, m_And(m_Specific(X), m_Value())) ||
869 match(Y, m_And(m_Value(), m_Specific(X))))
870 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth))
873 unsigned BitWidth = V->getType()->getScalarSizeInBits();
874 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
875 computeKnownBits(X, LHSZeroBits, LHSOneBits, nullptr, Depth);
877 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
878 computeKnownBits(Y, RHSZeroBits, RHSOneBits, nullptr, Depth);
879 // If i8 V is a power of two or zero:
880 // ZeroBits: 1 1 1 0 1 1 1 1
881 // ~ZeroBits: 0 0 0 1 0 0 0 0
882 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
883 // If OrZero isn't set, we cannot give back a zero result.
884 // Make sure either the LHS or RHS has a bit set.
885 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
890 // An exact divide or right shift can only shift off zero bits, so the result
891 // is a power of two only if the first operand is a power of two and not
892 // copying a sign bit (sdiv int_min, 2).
893 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
894 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
895 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, Depth);
901 /// \brief Test whether a GEP's result is known to be non-null.
903 /// Uses properties inherent in a GEP to try to determine whether it is known
906 /// Currently this routine does not support vector GEPs.
907 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
909 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
912 // FIXME: Support vector-GEPs.
913 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
915 // If the base pointer is non-null, we cannot walk to a null address with an
916 // inbounds GEP in address space zero.
917 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth))
920 // Past this, if we don't have DataLayout, we can't do much.
924 // Walk the GEP operands and see if any operand introduces a non-zero offset.
925 // If so, then the GEP cannot produce a null pointer, as doing so would
926 // inherently violate the inbounds contract within address space zero.
927 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
929 // Struct types are easy -- they must always be indexed by a constant.
930 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
931 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
932 unsigned ElementIdx = OpC->getZExtValue();
933 const StructLayout *SL = DL->getStructLayout(STy);
934 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
935 if (ElementOffset > 0)
940 // If we have a zero-sized type, the index doesn't matter. Keep looping.
941 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
944 // Fast path the constant operand case both for efficiency and so we don't
945 // increment Depth when just zipping down an all-constant GEP.
946 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
952 // We post-increment Depth here because while isKnownNonZero increments it
953 // as well, when we pop back up that increment won't persist. We don't want
954 // to recurse 10k times just because we have 10k GEP operands. We don't
955 // bail completely out because we want to handle constant GEPs regardless
957 if (Depth++ >= MaxDepth)
960 if (isKnownNonZero(GTI.getOperand(), DL, Depth))
967 /// isKnownNonZero - Return true if the given value is known to be non-zero
968 /// when defined. For vectors return true if every element is known to be
969 /// non-zero when defined. Supports values with integer or pointer type and
970 /// vectors of integers.
971 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth) {
972 if (Constant *C = dyn_cast<Constant>(V)) {
973 if (C->isNullValue())
975 if (isa<ConstantInt>(C))
976 // Must be non-zero due to null test above.
978 // TODO: Handle vectors
982 // The remaining tests are all recursive, so bail out if we hit the limit.
983 if (Depth++ >= MaxDepth)
986 // Check for pointer simplifications.
987 if (V->getType()->isPointerTy()) {
988 if (isKnownNonNull(V))
990 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
991 if (isGEPKnownNonNull(GEP, TD, Depth))
995 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
997 // X | Y != 0 if X != 0 or Y != 0.
998 Value *X = nullptr, *Y = nullptr;
999 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1000 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
1002 // ext X != 0 if X != 0.
1003 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1004 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
1006 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1007 // if the lowest bit is shifted off the end.
1008 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1009 // shl nuw can't remove any non-zero bits.
1010 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1011 if (BO->hasNoUnsignedWrap())
1012 return isKnownNonZero(X, TD, Depth);
1014 APInt KnownZero(BitWidth, 0);
1015 APInt KnownOne(BitWidth, 0);
1016 computeKnownBits(X, KnownZero, KnownOne, TD, Depth);
1020 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1021 // defined if the sign bit is shifted off the end.
1022 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1023 // shr exact can only shift out zero bits.
1024 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1026 return isKnownNonZero(X, TD, Depth);
1028 bool XKnownNonNegative, XKnownNegative;
1029 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
1033 // div exact can only produce a zero if the dividend is zero.
1034 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1035 return isKnownNonZero(X, TD, Depth);
1038 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1039 bool XKnownNonNegative, XKnownNegative;
1040 bool YKnownNonNegative, YKnownNegative;
1041 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
1042 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
1044 // If X and Y are both non-negative (as signed values) then their sum is not
1045 // zero unless both X and Y are zero.
1046 if (XKnownNonNegative && YKnownNonNegative)
1047 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
1050 // If X and Y are both negative (as signed values) then their sum is not
1051 // zero unless both X and Y equal INT_MIN.
1052 if (BitWidth && XKnownNegative && YKnownNegative) {
1053 APInt KnownZero(BitWidth, 0);
1054 APInt KnownOne(BitWidth, 0);
1055 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1056 // The sign bit of X is set. If some other bit is set then X is not equal
1058 computeKnownBits(X, KnownZero, KnownOne, TD, Depth);
1059 if ((KnownOne & Mask) != 0)
1061 // The sign bit of Y is set. If some other bit is set then Y is not equal
1063 computeKnownBits(Y, KnownZero, KnownOne, TD, Depth);
1064 if ((KnownOne & Mask) != 0)
1068 // The sum of a non-negative number and a power of two is not zero.
1069 if (XKnownNonNegative && isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth))
1071 if (YKnownNonNegative && isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth))
1075 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1076 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1077 // If X and Y are non-zero then so is X * Y as long as the multiplication
1078 // does not overflow.
1079 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1080 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
1083 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1084 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1085 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
1086 isKnownNonZero(SI->getFalseValue(), TD, Depth))
1090 if (!BitWidth) return false;
1091 APInt KnownZero(BitWidth, 0);
1092 APInt KnownOne(BitWidth, 0);
1093 computeKnownBits(V, KnownZero, KnownOne, TD, Depth);
1094 return KnownOne != 0;
1097 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1098 /// this predicate to simplify operations downstream. Mask is known to be zero
1099 /// for bits that V cannot have.
1101 /// This function is defined on values with integer type, values with pointer
1102 /// type (but only if TD is non-null), and vectors of integers. In the case
1103 /// where V is a vector, the mask, known zero, and known one values are the
1104 /// same width as the vector element, and the bit is set only if it is true
1105 /// for all of the elements in the vector.
1106 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
1107 const DataLayout *TD, unsigned Depth) {
1108 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1109 computeKnownBits(V, KnownZero, KnownOne, TD, Depth);
1110 return (KnownZero & Mask) == Mask;
1115 /// ComputeNumSignBits - Return the number of times the sign bit of the
1116 /// register is replicated into the other bits. We know that at least 1 bit
1117 /// is always equal to the sign bit (itself), but other cases can give us
1118 /// information. For example, immediately after an "ashr X, 2", we know that
1119 /// the top 3 bits are all equal to each other, so we return 3.
1121 /// 'Op' must have a scalar integer type.
1123 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
1125 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1126 "ComputeNumSignBits requires a DataLayout object to operate "
1127 "on non-integer values!");
1128 Type *Ty = V->getType();
1129 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1130 Ty->getScalarSizeInBits();
1132 unsigned FirstAnswer = 1;
1134 // Note that ConstantInt is handled by the general computeKnownBits case
1138 return 1; // Limit search depth.
1140 Operator *U = dyn_cast<Operator>(V);
1141 switch (Operator::getOpcode(V)) {
1143 case Instruction::SExt:
1144 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1145 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
1147 case Instruction::AShr: {
1148 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1149 // ashr X, C -> adds C sign bits. Vectors too.
1151 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1152 Tmp += ShAmt->getZExtValue();
1153 if (Tmp > TyBits) Tmp = TyBits;
1157 case Instruction::Shl: {
1159 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1160 // shl destroys sign bits.
1161 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1162 Tmp2 = ShAmt->getZExtValue();
1163 if (Tmp2 >= TyBits || // Bad shift.
1164 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1169 case Instruction::And:
1170 case Instruction::Or:
1171 case Instruction::Xor: // NOT is handled here.
1172 // Logical binary ops preserve the number of sign bits at the worst.
1173 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1175 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1176 FirstAnswer = std::min(Tmp, Tmp2);
1177 // We computed what we know about the sign bits as our first
1178 // answer. Now proceed to the generic code that uses
1179 // computeKnownBits, and pick whichever answer is better.
1183 case Instruction::Select:
1184 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1185 if (Tmp == 1) return 1; // Early out.
1186 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1187 return std::min(Tmp, Tmp2);
1189 case Instruction::Add:
1190 // Add can have at most one carry bit. Thus we know that the output
1191 // is, at worst, one more bit than the inputs.
1192 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1193 if (Tmp == 1) return 1; // Early out.
1195 // Special case decrementing a value (ADD X, -1):
1196 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1197 if (CRHS->isAllOnesValue()) {
1198 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1199 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1);
1201 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1203 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1206 // If we are subtracting one from a positive number, there is no carry
1207 // out of the result.
1208 if (KnownZero.isNegative())
1212 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1213 if (Tmp2 == 1) return 1;
1214 return std::min(Tmp, Tmp2)-1;
1216 case Instruction::Sub:
1217 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1218 if (Tmp2 == 1) return 1;
1221 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1222 if (CLHS->isNullValue()) {
1223 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1224 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1);
1225 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1227 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1230 // If the input is known to be positive (the sign bit is known clear),
1231 // the output of the NEG has the same number of sign bits as the input.
1232 if (KnownZero.isNegative())
1235 // Otherwise, we treat this like a SUB.
1238 // Sub can have at most one carry bit. Thus we know that the output
1239 // is, at worst, one more bit than the inputs.
1240 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1241 if (Tmp == 1) return 1; // Early out.
1242 return std::min(Tmp, Tmp2)-1;
1244 case Instruction::PHI: {
1245 PHINode *PN = cast<PHINode>(U);
1246 // Don't analyze large in-degree PHIs.
1247 if (PN->getNumIncomingValues() > 4) break;
1249 // Take the minimum of all incoming values. This can't infinitely loop
1250 // because of our depth threshold.
1251 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1252 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1253 if (Tmp == 1) return Tmp;
1255 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1260 case Instruction::Trunc:
1261 // FIXME: it's tricky to do anything useful for this, but it is an important
1262 // case for targets like X86.
1266 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1267 // use this information.
1268 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1270 computeKnownBits(V, KnownZero, KnownOne, TD, Depth);
1272 if (KnownZero.isNegative()) { // sign bit is 0
1274 } else if (KnownOne.isNegative()) { // sign bit is 1;
1281 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1282 // the number of identical bits in the top of the input value.
1284 Mask <<= Mask.getBitWidth()-TyBits;
1285 // Return # leading zeros. We use 'min' here in case Val was zero before
1286 // shifting. We don't want to return '64' as for an i32 "0".
1287 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1290 /// ComputeMultiple - This function computes the integer multiple of Base that
1291 /// equals V. If successful, it returns true and returns the multiple in
1292 /// Multiple. If unsuccessful, it returns false. It looks
1293 /// through SExt instructions only if LookThroughSExt is true.
1294 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1295 bool LookThroughSExt, unsigned Depth) {
1296 const unsigned MaxDepth = 6;
1298 assert(V && "No Value?");
1299 assert(Depth <= MaxDepth && "Limit Search Depth");
1300 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1302 Type *T = V->getType();
1304 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1314 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1315 Constant *BaseVal = ConstantInt::get(T, Base);
1316 if (CO && CO == BaseVal) {
1318 Multiple = ConstantInt::get(T, 1);
1322 if (CI && CI->getZExtValue() % Base == 0) {
1323 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1327 if (Depth == MaxDepth) return false; // Limit search depth.
1329 Operator *I = dyn_cast<Operator>(V);
1330 if (!I) return false;
1332 switch (I->getOpcode()) {
1334 case Instruction::SExt:
1335 if (!LookThroughSExt) return false;
1336 // otherwise fall through to ZExt
1337 case Instruction::ZExt:
1338 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1339 LookThroughSExt, Depth+1);
1340 case Instruction::Shl:
1341 case Instruction::Mul: {
1342 Value *Op0 = I->getOperand(0);
1343 Value *Op1 = I->getOperand(1);
1345 if (I->getOpcode() == Instruction::Shl) {
1346 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1347 if (!Op1CI) return false;
1348 // Turn Op0 << Op1 into Op0 * 2^Op1
1349 APInt Op1Int = Op1CI->getValue();
1350 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1351 APInt API(Op1Int.getBitWidth(), 0);
1352 API.setBit(BitToSet);
1353 Op1 = ConstantInt::get(V->getContext(), API);
1356 Value *Mul0 = nullptr;
1357 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1358 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1359 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1360 if (Op1C->getType()->getPrimitiveSizeInBits() <
1361 MulC->getType()->getPrimitiveSizeInBits())
1362 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1363 if (Op1C->getType()->getPrimitiveSizeInBits() >
1364 MulC->getType()->getPrimitiveSizeInBits())
1365 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1367 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1368 Multiple = ConstantExpr::getMul(MulC, Op1C);
1372 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1373 if (Mul0CI->getValue() == 1) {
1374 // V == Base * Op1, so return Op1
1380 Value *Mul1 = nullptr;
1381 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1382 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1383 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1384 if (Op0C->getType()->getPrimitiveSizeInBits() <
1385 MulC->getType()->getPrimitiveSizeInBits())
1386 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1387 if (Op0C->getType()->getPrimitiveSizeInBits() >
1388 MulC->getType()->getPrimitiveSizeInBits())
1389 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1391 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1392 Multiple = ConstantExpr::getMul(MulC, Op0C);
1396 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1397 if (Mul1CI->getValue() == 1) {
1398 // V == Base * Op0, so return Op0
1406 // We could not determine if V is a multiple of Base.
1410 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1411 /// value is never equal to -0.0.
1413 /// NOTE: this function will need to be revisited when we support non-default
1416 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1417 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1418 return !CFP->getValueAPF().isNegZero();
1421 return 1; // Limit search depth.
1423 const Operator *I = dyn_cast<Operator>(V);
1424 if (!I) return false;
1426 // Check if the nsz fast-math flag is set
1427 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
1428 if (FPO->hasNoSignedZeros())
1431 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1432 if (I->getOpcode() == Instruction::FAdd)
1433 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
1434 if (CFP->isNullValue())
1437 // sitofp and uitofp turn into +0.0 for zero.
1438 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1441 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1442 // sqrt(-0.0) = -0.0, no other negative results are possible.
1443 if (II->getIntrinsicID() == Intrinsic::sqrt)
1444 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1446 if (const CallInst *CI = dyn_cast<CallInst>(I))
1447 if (const Function *F = CI->getCalledFunction()) {
1448 if (F->isDeclaration()) {
1450 if (F->getName() == "abs") return true;
1451 // fabs[lf](x) != -0.0
1452 if (F->getName() == "fabs") return true;
1453 if (F->getName() == "fabsf") return true;
1454 if (F->getName() == "fabsl") return true;
1455 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1456 F->getName() == "sqrtl")
1457 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1464 /// isBytewiseValue - If the specified value can be set by repeating the same
1465 /// byte in memory, return the i8 value that it is represented with. This is
1466 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1467 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1468 /// byte store (e.g. i16 0x1234), return null.
1469 Value *llvm::isBytewiseValue(Value *V) {
1470 // All byte-wide stores are splatable, even of arbitrary variables.
1471 if (V->getType()->isIntegerTy(8)) return V;
1473 // Handle 'null' ConstantArrayZero etc.
1474 if (Constant *C = dyn_cast<Constant>(V))
1475 if (C->isNullValue())
1476 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1478 // Constant float and double values can be handled as integer values if the
1479 // corresponding integer value is "byteable". An important case is 0.0.
1480 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1481 if (CFP->getType()->isFloatTy())
1482 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1483 if (CFP->getType()->isDoubleTy())
1484 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1485 // Don't handle long double formats, which have strange constraints.
1488 // We can handle constant integers that are power of two in size and a
1489 // multiple of 8 bits.
1490 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1491 unsigned Width = CI->getBitWidth();
1492 if (isPowerOf2_32(Width) && Width > 8) {
1493 // We can handle this value if the recursive binary decomposition is the
1494 // same at all levels.
1495 APInt Val = CI->getValue();
1497 while (Val.getBitWidth() != 8) {
1498 unsigned NextWidth = Val.getBitWidth()/2;
1499 Val2 = Val.lshr(NextWidth);
1500 Val2 = Val2.trunc(Val.getBitWidth()/2);
1501 Val = Val.trunc(Val.getBitWidth()/2);
1503 // If the top/bottom halves aren't the same, reject it.
1507 return ConstantInt::get(V->getContext(), Val);
1511 // A ConstantDataArray/Vector is splatable if all its members are equal and
1513 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1514 Value *Elt = CA->getElementAsConstant(0);
1515 Value *Val = isBytewiseValue(Elt);
1519 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1520 if (CA->getElementAsConstant(I) != Elt)
1526 // Conceptually, we could handle things like:
1527 // %a = zext i8 %X to i16
1528 // %b = shl i16 %a, 8
1529 // %c = or i16 %a, %b
1530 // but until there is an example that actually needs this, it doesn't seem
1531 // worth worrying about.
1536 // This is the recursive version of BuildSubAggregate. It takes a few different
1537 // arguments. Idxs is the index within the nested struct From that we are
1538 // looking at now (which is of type IndexedType). IdxSkip is the number of
1539 // indices from Idxs that should be left out when inserting into the resulting
1540 // struct. To is the result struct built so far, new insertvalue instructions
1542 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1543 SmallVectorImpl<unsigned> &Idxs,
1545 Instruction *InsertBefore) {
1546 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
1548 // Save the original To argument so we can modify it
1550 // General case, the type indexed by Idxs is a struct
1551 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1552 // Process each struct element recursively
1555 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1559 // Couldn't find any inserted value for this index? Cleanup
1560 while (PrevTo != OrigTo) {
1561 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1562 PrevTo = Del->getAggregateOperand();
1563 Del->eraseFromParent();
1565 // Stop processing elements
1569 // If we successfully found a value for each of our subaggregates
1573 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1574 // the struct's elements had a value that was inserted directly. In the latter
1575 // case, perhaps we can't determine each of the subelements individually, but
1576 // we might be able to find the complete struct somewhere.
1578 // Find the value that is at that particular spot
1579 Value *V = FindInsertedValue(From, Idxs);
1584 // Insert the value in the new (sub) aggregrate
1585 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1586 "tmp", InsertBefore);
1589 // This helper takes a nested struct and extracts a part of it (which is again a
1590 // struct) into a new value. For example, given the struct:
1591 // { a, { b, { c, d }, e } }
1592 // and the indices "1, 1" this returns
1595 // It does this by inserting an insertvalue for each element in the resulting
1596 // struct, as opposed to just inserting a single struct. This will only work if
1597 // each of the elements of the substruct are known (ie, inserted into From by an
1598 // insertvalue instruction somewhere).
1600 // All inserted insertvalue instructions are inserted before InsertBefore
1601 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1602 Instruction *InsertBefore) {
1603 assert(InsertBefore && "Must have someplace to insert!");
1604 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1606 Value *To = UndefValue::get(IndexedType);
1607 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1608 unsigned IdxSkip = Idxs.size();
1610 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1613 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1614 /// the scalar value indexed is already around as a register, for example if it
1615 /// were inserted directly into the aggregrate.
1617 /// If InsertBefore is not null, this function will duplicate (modified)
1618 /// insertvalues when a part of a nested struct is extracted.
1619 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1620 Instruction *InsertBefore) {
1621 // Nothing to index? Just return V then (this is useful at the end of our
1623 if (idx_range.empty())
1625 // We have indices, so V should have an indexable type.
1626 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1627 "Not looking at a struct or array?");
1628 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1629 "Invalid indices for type?");
1631 if (Constant *C = dyn_cast<Constant>(V)) {
1632 C = C->getAggregateElement(idx_range[0]);
1633 if (!C) return nullptr;
1634 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1637 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1638 // Loop the indices for the insertvalue instruction in parallel with the
1639 // requested indices
1640 const unsigned *req_idx = idx_range.begin();
1641 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1642 i != e; ++i, ++req_idx) {
1643 if (req_idx == idx_range.end()) {
1644 // We can't handle this without inserting insertvalues
1648 // The requested index identifies a part of a nested aggregate. Handle
1649 // this specially. For example,
1650 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1651 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1652 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1653 // This can be changed into
1654 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1655 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1656 // which allows the unused 0,0 element from the nested struct to be
1658 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1662 // This insert value inserts something else than what we are looking for.
1663 // See if the (aggregrate) value inserted into has the value we are
1664 // looking for, then.
1666 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1669 // If we end up here, the indices of the insertvalue match with those
1670 // requested (though possibly only partially). Now we recursively look at
1671 // the inserted value, passing any remaining indices.
1672 return FindInsertedValue(I->getInsertedValueOperand(),
1673 makeArrayRef(req_idx, idx_range.end()),
1677 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1678 // If we're extracting a value from an aggregrate that was extracted from
1679 // something else, we can extract from that something else directly instead.
1680 // However, we will need to chain I's indices with the requested indices.
1682 // Calculate the number of indices required
1683 unsigned size = I->getNumIndices() + idx_range.size();
1684 // Allocate some space to put the new indices in
1685 SmallVector<unsigned, 5> Idxs;
1687 // Add indices from the extract value instruction
1688 Idxs.append(I->idx_begin(), I->idx_end());
1690 // Add requested indices
1691 Idxs.append(idx_range.begin(), idx_range.end());
1693 assert(Idxs.size() == size
1694 && "Number of indices added not correct?");
1696 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1698 // Otherwise, we don't know (such as, extracting from a function return value
1699 // or load instruction)
1703 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1704 /// it can be expressed as a base pointer plus a constant offset. Return the
1705 /// base and offset to the caller.
1706 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1707 const DataLayout *DL) {
1708 // Without DataLayout, conservatively assume 64-bit offsets, which is
1709 // the widest we support.
1710 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(Ptr->getType()) : 64;
1711 APInt ByteOffset(BitWidth, 0);
1713 if (Ptr->getType()->isVectorTy())
1716 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1718 APInt GEPOffset(BitWidth, 0);
1719 if (!GEP->accumulateConstantOffset(*DL, GEPOffset))
1722 ByteOffset += GEPOffset;
1725 Ptr = GEP->getPointerOperand();
1726 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1727 Ptr = cast<Operator>(Ptr)->getOperand(0);
1728 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1729 if (GA->mayBeOverridden())
1731 Ptr = GA->getAliasee();
1736 Offset = ByteOffset.getSExtValue();
1741 /// getConstantStringInfo - This function computes the length of a
1742 /// null-terminated C string pointed to by V. If successful, it returns true
1743 /// and returns the string in Str. If unsuccessful, it returns false.
1744 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
1745 uint64_t Offset, bool TrimAtNul) {
1748 // Look through bitcast instructions and geps.
1749 V = V->stripPointerCasts();
1751 // If the value is a GEP instructionor constant expression, treat it as an
1753 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1754 // Make sure the GEP has exactly three arguments.
1755 if (GEP->getNumOperands() != 3)
1758 // Make sure the index-ee is a pointer to array of i8.
1759 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1760 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1761 if (!AT || !AT->getElementType()->isIntegerTy(8))
1764 // Check to make sure that the first operand of the GEP is an integer and
1765 // has value 0 so that we are sure we're indexing into the initializer.
1766 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1767 if (!FirstIdx || !FirstIdx->isZero())
1770 // If the second index isn't a ConstantInt, then this is a variable index
1771 // into the array. If this occurs, we can't say anything meaningful about
1773 uint64_t StartIdx = 0;
1774 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1775 StartIdx = CI->getZExtValue();
1778 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
1781 // The GEP instruction, constant or instruction, must reference a global
1782 // variable that is a constant and is initialized. The referenced constant
1783 // initializer is the array that we'll use for optimization.
1784 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
1785 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1788 // Handle the all-zeros case
1789 if (GV->getInitializer()->isNullValue()) {
1790 // This is a degenerate case. The initializer is constant zero so the
1791 // length of the string must be zero.
1796 // Must be a Constant Array
1797 const ConstantDataArray *Array =
1798 dyn_cast<ConstantDataArray>(GV->getInitializer());
1799 if (!Array || !Array->isString())
1802 // Get the number of elements in the array
1803 uint64_t NumElts = Array->getType()->getArrayNumElements();
1805 // Start out with the entire array in the StringRef.
1806 Str = Array->getAsString();
1808 if (Offset > NumElts)
1811 // Skip over 'offset' bytes.
1812 Str = Str.substr(Offset);
1815 // Trim off the \0 and anything after it. If the array is not nul
1816 // terminated, we just return the whole end of string. The client may know
1817 // some other way that the string is length-bound.
1818 Str = Str.substr(0, Str.find('\0'));
1823 // These next two are very similar to the above, but also look through PHI
1825 // TODO: See if we can integrate these two together.
1827 /// GetStringLengthH - If we can compute the length of the string pointed to by
1828 /// the specified pointer, return 'len+1'. If we can't, return 0.
1829 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1830 // Look through noop bitcast instructions.
1831 V = V->stripPointerCasts();
1833 // If this is a PHI node, there are two cases: either we have already seen it
1835 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1836 if (!PHIs.insert(PN))
1837 return ~0ULL; // already in the set.
1839 // If it was new, see if all the input strings are the same length.
1840 uint64_t LenSoFar = ~0ULL;
1841 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1842 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1843 if (Len == 0) return 0; // Unknown length -> unknown.
1845 if (Len == ~0ULL) continue;
1847 if (Len != LenSoFar && LenSoFar != ~0ULL)
1848 return 0; // Disagree -> unknown.
1852 // Success, all agree.
1856 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1857 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1858 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1859 if (Len1 == 0) return 0;
1860 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1861 if (Len2 == 0) return 0;
1862 if (Len1 == ~0ULL) return Len2;
1863 if (Len2 == ~0ULL) return Len1;
1864 if (Len1 != Len2) return 0;
1868 // Otherwise, see if we can read the string.
1870 if (!getConstantStringInfo(V, StrData))
1873 return StrData.size()+1;
1876 /// GetStringLength - If we can compute the length of the string pointed to by
1877 /// the specified pointer, return 'len+1'. If we can't, return 0.
1878 uint64_t llvm::GetStringLength(Value *V) {
1879 if (!V->getType()->isPointerTy()) return 0;
1881 SmallPtrSet<PHINode*, 32> PHIs;
1882 uint64_t Len = GetStringLengthH(V, PHIs);
1883 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1884 // an empty string as a length.
1885 return Len == ~0ULL ? 1 : Len;
1889 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
1890 if (!V->getType()->isPointerTy())
1892 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1893 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1894 V = GEP->getPointerOperand();
1895 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1896 V = cast<Operator>(V)->getOperand(0);
1897 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1898 if (GA->mayBeOverridden())
1900 V = GA->getAliasee();
1902 // See if InstructionSimplify knows any relevant tricks.
1903 if (Instruction *I = dyn_cast<Instruction>(V))
1904 // TODO: Acquire a DominatorTree and use it.
1905 if (Value *Simplified = SimplifyInstruction(I, TD, nullptr)) {
1912 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1918 llvm::GetUnderlyingObjects(Value *V,
1919 SmallVectorImpl<Value *> &Objects,
1920 const DataLayout *TD,
1921 unsigned MaxLookup) {
1922 SmallPtrSet<Value *, 4> Visited;
1923 SmallVector<Value *, 4> Worklist;
1924 Worklist.push_back(V);
1926 Value *P = Worklist.pop_back_val();
1927 P = GetUnderlyingObject(P, TD, MaxLookup);
1929 if (!Visited.insert(P))
1932 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
1933 Worklist.push_back(SI->getTrueValue());
1934 Worklist.push_back(SI->getFalseValue());
1938 if (PHINode *PN = dyn_cast<PHINode>(P)) {
1939 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1940 Worklist.push_back(PN->getIncomingValue(i));
1944 Objects.push_back(P);
1945 } while (!Worklist.empty());
1948 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1949 /// are lifetime markers.
1951 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1952 for (const User *U : V->users()) {
1953 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1954 if (!II) return false;
1956 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1957 II->getIntrinsicID() != Intrinsic::lifetime_end)
1963 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1964 const DataLayout *TD) {
1965 const Operator *Inst = dyn_cast<Operator>(V);
1969 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1970 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1974 switch (Inst->getOpcode()) {
1977 case Instruction::UDiv:
1978 case Instruction::URem:
1979 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1980 return isKnownNonZero(Inst->getOperand(1), TD);
1981 case Instruction::SDiv:
1982 case Instruction::SRem: {
1983 Value *Op = Inst->getOperand(1);
1984 // x / y is undefined if y == 0
1985 if (!isKnownNonZero(Op, TD))
1987 // x / y might be undefined if y == -1
1988 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1991 APInt KnownZero(BitWidth, 0);
1992 APInt KnownOne(BitWidth, 0);
1993 computeKnownBits(Op, KnownZero, KnownOne, TD);
1996 case Instruction::Load: {
1997 const LoadInst *LI = cast<LoadInst>(Inst);
1998 if (!LI->isUnordered() ||
1999 // Speculative load may create a race that did not exist in the source.
2000 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2002 return LI->getPointerOperand()->isDereferenceablePointer();
2004 case Instruction::Call: {
2005 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2006 switch (II->getIntrinsicID()) {
2007 // These synthetic intrinsics have no side-effects, and just mark
2008 // information about their operands.
2009 // FIXME: There are other no-op synthetic instructions that potentially
2010 // should be considered at least *safe* to speculate...
2011 case Intrinsic::dbg_declare:
2012 case Intrinsic::dbg_value:
2015 case Intrinsic::bswap:
2016 case Intrinsic::ctlz:
2017 case Intrinsic::ctpop:
2018 case Intrinsic::cttz:
2019 case Intrinsic::objectsize:
2020 case Intrinsic::sadd_with_overflow:
2021 case Intrinsic::smul_with_overflow:
2022 case Intrinsic::ssub_with_overflow:
2023 case Intrinsic::uadd_with_overflow:
2024 case Intrinsic::umul_with_overflow:
2025 case Intrinsic::usub_with_overflow:
2027 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2028 // errno like libm sqrt would.
2029 case Intrinsic::sqrt:
2030 case Intrinsic::fma:
2031 case Intrinsic::fmuladd:
2033 // TODO: some fp intrinsics are marked as having the same error handling
2034 // as libm. They're safe to speculate when they won't error.
2035 // TODO: are convert_{from,to}_fp16 safe?
2036 // TODO: can we list target-specific intrinsics here?
2040 return false; // The called function could have undefined behavior or
2041 // side-effects, even if marked readnone nounwind.
2043 case Instruction::VAArg:
2044 case Instruction::Alloca:
2045 case Instruction::Invoke:
2046 case Instruction::PHI:
2047 case Instruction::Store:
2048 case Instruction::Ret:
2049 case Instruction::Br:
2050 case Instruction::IndirectBr:
2051 case Instruction::Switch:
2052 case Instruction::Unreachable:
2053 case Instruction::Fence:
2054 case Instruction::LandingPad:
2055 case Instruction::AtomicRMW:
2056 case Instruction::AtomicCmpXchg:
2057 case Instruction::Resume:
2058 return false; // Misc instructions which have effects
2062 /// isKnownNonNull - Return true if we know that the specified value is never
2064 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2065 // Alloca never returns null, malloc might.
2066 if (isa<AllocaInst>(V)) return true;
2068 // A byval or inalloca argument is never null.
2069 if (const Argument *A = dyn_cast<Argument>(V))
2070 return A->hasByValOrInAllocaAttr();
2072 // Global values are not null unless extern weak.
2073 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2074 return !GV->hasExternalWeakLinkage();
2076 // operator new never returns null.
2077 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))