1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Constants.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/GlobalVariable.h"
20 #include "llvm/GlobalAlias.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/Operator.h"
24 #include "llvm/Target/TargetData.h"
25 #include "llvm/Support/GetElementPtrTypeIterator.h"
26 #include "llvm/Support/MathExtras.h"
27 #include "llvm/Support/PatternMatch.h"
28 #include "llvm/ADT/SmallPtrSet.h"
31 using namespace llvm::PatternMatch;
33 const unsigned MaxDepth = 6;
35 /// getBitWidth - Returns the bitwidth of the given scalar or pointer type (if
36 /// unknown returns 0). For vector types, returns the element type's bitwidth.
37 static unsigned getBitWidth(Type *Ty, const TargetData *TD) {
38 if (unsigned BitWidth = Ty->getScalarSizeInBits())
40 assert(isa<PointerType>(Ty) && "Expected a pointer type!");
41 return TD ? TD->getPointerSizeInBits() : 0;
44 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
45 /// known to be either zero or one and return them in the KnownZero/KnownOne
46 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
48 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
49 /// we cannot optimize based on the assumption that it is zero without changing
50 /// it to be an explicit zero. If we don't change it to zero, other code could
51 /// optimized based on the contradictory assumption that it is non-zero.
52 /// Because instcombine aggressively folds operations with undef args anyway,
53 /// this won't lose us code quality.
55 /// This function is defined on values with integer type, values with pointer
56 /// type (but only if TD is non-null), and vectors of integers. In the case
57 /// where V is a vector, the mask, known zero, and known one values are the
58 /// same width as the vector element, and the bit is set only if it is true
59 /// for all of the elements in the vector.
60 void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
61 APInt &KnownZero, APInt &KnownOne,
62 const TargetData *TD, unsigned Depth) {
63 assert(V && "No Value?");
64 assert(Depth <= MaxDepth && "Limit Search Depth");
65 unsigned BitWidth = Mask.getBitWidth();
66 assert((V->getType()->isIntOrIntVectorTy() ||
67 V->getType()->getScalarType()->isPointerTy()) &&
68 "Not integer or pointer type!");
70 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
71 (!V->getType()->isIntOrIntVectorTy() ||
72 V->getType()->getScalarSizeInBits() == BitWidth) &&
73 KnownZero.getBitWidth() == BitWidth &&
74 KnownOne.getBitWidth() == BitWidth &&
75 "V, Mask, KnownOne and KnownZero should have same BitWidth");
77 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
78 // We know all of the bits for a constant!
79 KnownOne = CI->getValue() & Mask;
80 KnownZero = ~KnownOne & Mask;
83 // Null and aggregate-zero are all-zeros.
84 if (isa<ConstantPointerNull>(V) ||
85 isa<ConstantAggregateZero>(V)) {
86 KnownOne.clearAllBits();
90 // Handle a constant vector by taking the intersection of the known bits of
91 // each element. There is no real need to handle ConstantVector here, because
92 // we don't handle undef in any particularly useful way.
93 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
94 // We know that CDS must be a vector of integers. Take the intersection of
96 KnownZero.setAllBits(); KnownOne.setAllBits();
97 APInt Elt(KnownZero.getBitWidth(), 0);
98 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
99 Elt = CDS->getElementAsInteger(i);
106 // The address of an aligned GlobalValue has trailing zeros.
107 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
108 unsigned Align = GV->getAlignment();
109 if (Align == 0 && TD) {
110 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
111 Type *ObjectType = GVar->getType()->getElementType();
112 if (ObjectType->isSized()) {
113 // If the object is defined in the current Module, we'll be giving
114 // it the preferred alignment. Otherwise, we have to assume that it
115 // may only have the minimum ABI alignment.
116 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
117 Align = TD->getPreferredAlignment(GVar);
119 Align = TD->getABITypeAlignment(ObjectType);
124 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
125 CountTrailingZeros_32(Align));
127 KnownZero.clearAllBits();
128 KnownOne.clearAllBits();
131 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
132 // the bits of its aliasee.
133 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
134 if (GA->mayBeOverridden()) {
135 KnownZero.clearAllBits(); KnownOne.clearAllBits();
137 ComputeMaskedBits(GA->getAliasee(), Mask, KnownZero, KnownOne,
143 if (Argument *A = dyn_cast<Argument>(V)) {
144 // Get alignment information off byval arguments if specified in the IR.
145 if (A->hasByValAttr())
146 if (unsigned Align = A->getParamAlignment())
147 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
148 CountTrailingZeros_32(Align));
152 // Start out not knowing anything.
153 KnownZero.clearAllBits(); KnownOne.clearAllBits();
155 if (Depth == MaxDepth || Mask == 0)
156 return; // Limit search depth.
158 Operator *I = dyn_cast<Operator>(V);
161 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
162 switch (I->getOpcode()) {
164 case Instruction::And: {
165 // If either the LHS or the RHS are Zero, the result is zero.
166 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
167 APInt Mask2(Mask & ~KnownZero);
168 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
170 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
171 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
173 // Output known-1 bits are only known if set in both the LHS & RHS.
174 KnownOne &= KnownOne2;
175 // Output known-0 are known to be clear if zero in either the LHS | RHS.
176 KnownZero |= KnownZero2;
179 case Instruction::Or: {
180 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
181 APInt Mask2(Mask & ~KnownOne);
182 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
184 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
185 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
187 // Output known-0 bits are only known if clear in both the LHS & RHS.
188 KnownZero &= KnownZero2;
189 // Output known-1 are known to be set if set in either the LHS | RHS.
190 KnownOne |= KnownOne2;
193 case Instruction::Xor: {
194 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
195 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
197 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
198 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
200 // Output known-0 bits are known if clear or set in both the LHS & RHS.
201 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
202 // Output known-1 are known to be set if set in only one of the LHS, RHS.
203 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
204 KnownZero = KnownZeroOut;
207 case Instruction::Mul: {
208 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
209 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
210 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
212 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
213 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
215 bool isKnownNegative = false;
216 bool isKnownNonNegative = false;
217 // If the multiplication is known not to overflow, compute the sign bit.
218 if (Mask.isNegative() &&
219 cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap()) {
220 Value *Op1 = I->getOperand(1), *Op2 = I->getOperand(0);
222 // The product of a number with itself is non-negative.
223 isKnownNonNegative = true;
225 bool isKnownNonNegative1 = KnownZero.isNegative();
226 bool isKnownNonNegative2 = KnownZero2.isNegative();
227 bool isKnownNegative1 = KnownOne.isNegative();
228 bool isKnownNegative2 = KnownOne2.isNegative();
229 // The product of two numbers with the same sign is non-negative.
230 isKnownNonNegative = (isKnownNegative1 && isKnownNegative2) ||
231 (isKnownNonNegative1 && isKnownNonNegative2);
232 // The product of a negative number and a non-negative number is either
234 if (!isKnownNonNegative)
235 isKnownNegative = (isKnownNegative1 && isKnownNonNegative2 &&
236 isKnownNonZero(Op2, TD, Depth)) ||
237 (isKnownNegative2 && isKnownNonNegative1 &&
238 isKnownNonZero(Op1, TD, Depth));
242 // If low bits are zero in either operand, output low known-0 bits.
243 // Also compute a conserative estimate for high known-0 bits.
244 // More trickiness is possible, but this is sufficient for the
245 // interesting case of alignment computation.
246 KnownOne.clearAllBits();
247 unsigned TrailZ = KnownZero.countTrailingOnes() +
248 KnownZero2.countTrailingOnes();
249 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
250 KnownZero2.countLeadingOnes(),
251 BitWidth) - BitWidth;
253 TrailZ = std::min(TrailZ, BitWidth);
254 LeadZ = std::min(LeadZ, BitWidth);
255 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
256 APInt::getHighBitsSet(BitWidth, LeadZ);
259 // Only make use of no-wrap flags if we failed to compute the sign bit
260 // directly. This matters if the multiplication always overflows, in
261 // which case we prefer to follow the result of the direct computation,
262 // though as the program is invoking undefined behaviour we can choose
263 // whatever we like here.
264 if (isKnownNonNegative && !KnownOne.isNegative())
265 KnownZero.setBit(BitWidth - 1);
266 else if (isKnownNegative && !KnownZero.isNegative())
267 KnownOne.setBit(BitWidth - 1);
271 case Instruction::UDiv: {
272 // For the purposes of computing leading zeros we can conservatively
273 // treat a udiv as a logical right shift by the power of 2 known to
274 // be less than the denominator.
275 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
276 ComputeMaskedBits(I->getOperand(0),
277 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
278 unsigned LeadZ = KnownZero2.countLeadingOnes();
280 KnownOne2.clearAllBits();
281 KnownZero2.clearAllBits();
282 ComputeMaskedBits(I->getOperand(1),
283 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
284 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
285 if (RHSUnknownLeadingOnes != BitWidth)
286 LeadZ = std::min(BitWidth,
287 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
289 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
292 case Instruction::Select:
293 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
294 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
296 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
297 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
299 // Only known if known in both the LHS and RHS.
300 KnownOne &= KnownOne2;
301 KnownZero &= KnownZero2;
303 case Instruction::FPTrunc:
304 case Instruction::FPExt:
305 case Instruction::FPToUI:
306 case Instruction::FPToSI:
307 case Instruction::SIToFP:
308 case Instruction::UIToFP:
309 return; // Can't work with floating point.
310 case Instruction::PtrToInt:
311 case Instruction::IntToPtr:
312 // We can't handle these if we don't know the pointer size.
314 // FALL THROUGH and handle them the same as zext/trunc.
315 case Instruction::ZExt:
316 case Instruction::Trunc: {
317 Type *SrcTy = I->getOperand(0)->getType();
319 unsigned SrcBitWidth;
320 // Note that we handle pointer operands here because of inttoptr/ptrtoint
321 // which fall through here.
322 if (SrcTy->isPointerTy())
323 SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
325 SrcBitWidth = SrcTy->getScalarSizeInBits();
327 APInt MaskIn = Mask.zextOrTrunc(SrcBitWidth);
328 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
329 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
330 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
332 KnownZero = KnownZero.zextOrTrunc(BitWidth);
333 KnownOne = KnownOne.zextOrTrunc(BitWidth);
334 // Any top bits are known to be zero.
335 if (BitWidth > SrcBitWidth)
336 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
339 case Instruction::BitCast: {
340 Type *SrcTy = I->getOperand(0)->getType();
341 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
342 // TODO: For now, not handling conversions like:
343 // (bitcast i64 %x to <2 x i32>)
344 !I->getType()->isVectorTy()) {
345 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
351 case Instruction::SExt: {
352 // Compute the bits in the result that are not present in the input.
353 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
355 APInt MaskIn = Mask.trunc(SrcBitWidth);
356 KnownZero = KnownZero.trunc(SrcBitWidth);
357 KnownOne = KnownOne.trunc(SrcBitWidth);
358 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
360 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
361 KnownZero = KnownZero.zext(BitWidth);
362 KnownOne = KnownOne.zext(BitWidth);
364 // If the sign bit of the input is known set or clear, then we know the
365 // top bits of the result.
366 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
367 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
368 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
369 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
372 case Instruction::Shl:
373 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
374 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
375 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
376 APInt Mask2(Mask.lshr(ShiftAmt));
377 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
379 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
380 KnownZero <<= ShiftAmt;
381 KnownOne <<= ShiftAmt;
382 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
386 case Instruction::LShr:
387 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
388 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
389 // Compute the new bits that are at the top now.
390 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
392 // Unsigned shift right.
393 APInt Mask2(Mask.shl(ShiftAmt));
394 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
396 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
397 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
398 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
399 // high bits known zero.
400 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
404 case Instruction::AShr:
405 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
406 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
407 // Compute the new bits that are at the top now.
408 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
410 // Signed shift right.
411 APInt Mask2(Mask.shl(ShiftAmt));
412 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
414 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
415 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
416 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
418 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
419 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
420 KnownZero |= HighBits;
421 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
422 KnownOne |= HighBits;
426 case Instruction::Sub: {
427 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
428 // We know that the top bits of C-X are clear if X contains less bits
429 // than C (i.e. no wrap-around can happen). For example, 20-X is
430 // positive if we can prove that X is >= 0 and < 16.
431 if (!CLHS->getValue().isNegative()) {
432 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
433 // NLZ can't be BitWidth with no sign bit
434 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
435 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
438 // If all of the MaskV bits are known to be zero, then we know the
439 // output top bits are zero, because we now know that the output is
441 if ((KnownZero2 & MaskV) == MaskV) {
442 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
443 // Top bits known zero.
444 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
450 case Instruction::Add: {
451 // If one of the operands has trailing zeros, then the bits that the
452 // other operand has in those bit positions will be preserved in the
453 // result. For an add, this works with either operand. For a subtract,
454 // this only works if the known zeros are in the right operand.
455 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
456 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
457 BitWidth - Mask.countLeadingZeros());
458 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
460 assert((LHSKnownZero & LHSKnownOne) == 0 &&
461 "Bits known to be one AND zero?");
462 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
464 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD,
466 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
467 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
469 // Determine which operand has more trailing zeros, and use that
470 // many bits from the other operand.
471 if (LHSKnownZeroOut > RHSKnownZeroOut) {
472 if (I->getOpcode() == Instruction::Add) {
473 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
474 KnownZero |= KnownZero2 & Mask;
475 KnownOne |= KnownOne2 & Mask;
477 // If the known zeros are in the left operand for a subtract,
478 // fall back to the minimum known zeros in both operands.
479 KnownZero |= APInt::getLowBitsSet(BitWidth,
480 std::min(LHSKnownZeroOut,
483 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
484 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
485 KnownZero |= LHSKnownZero & Mask;
486 KnownOne |= LHSKnownOne & Mask;
489 // Are we still trying to solve for the sign bit?
490 if (Mask.isNegative() && !KnownZero.isNegative() && !KnownOne.isNegative()){
491 OverflowingBinaryOperator *OBO = cast<OverflowingBinaryOperator>(I);
492 if (OBO->hasNoSignedWrap()) {
493 if (I->getOpcode() == Instruction::Add) {
494 // Adding two positive numbers can't wrap into negative
495 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
496 KnownZero |= APInt::getSignBit(BitWidth);
497 // and adding two negative numbers can't wrap into positive.
498 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
499 KnownOne |= APInt::getSignBit(BitWidth);
501 // Subtracting a negative number from a positive one can't wrap
502 if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
503 KnownZero |= APInt::getSignBit(BitWidth);
504 // neither can subtracting a positive number from a negative one.
505 else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
506 KnownOne |= APInt::getSignBit(BitWidth);
513 case Instruction::SRem:
514 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
515 APInt RA = Rem->getValue().abs();
516 if (RA.isPowerOf2()) {
517 APInt LowBits = RA - 1;
518 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
519 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
522 // The low bits of the first operand are unchanged by the srem.
523 KnownZero = KnownZero2 & LowBits;
524 KnownOne = KnownOne2 & LowBits;
526 // If the first operand is non-negative or has all low bits zero, then
527 // the upper bits are all zero.
528 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
529 KnownZero |= ~LowBits;
531 // If the first operand is negative and not all low bits are zero, then
532 // the upper bits are all one.
533 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
534 KnownOne |= ~LowBits;
539 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
543 // The sign bit is the LHS's sign bit, except when the result of the
544 // remainder is zero.
545 if (Mask.isNegative() && KnownZero.isNonNegative()) {
546 APInt Mask2 = APInt::getSignBit(BitWidth);
547 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
548 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
550 // If it's known zero, our sign bit is also zero.
551 if (LHSKnownZero.isNegative())
552 KnownZero |= LHSKnownZero;
556 case Instruction::URem: {
557 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
558 APInt RA = Rem->getValue();
559 if (RA.isPowerOf2()) {
560 APInt LowBits = (RA - 1);
561 APInt Mask2 = LowBits & Mask;
562 KnownZero |= ~LowBits & Mask;
563 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
565 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
570 // Since the result is less than or equal to either operand, any leading
571 // zero bits in either operand must also exist in the result.
572 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
573 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
575 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
578 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
579 KnownZero2.countLeadingOnes());
580 KnownOne.clearAllBits();
581 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
585 case Instruction::Alloca: {
586 AllocaInst *AI = cast<AllocaInst>(V);
587 unsigned Align = AI->getAlignment();
588 if (Align == 0 && TD)
589 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
592 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
593 CountTrailingZeros_32(Align));
596 case Instruction::GetElementPtr: {
597 // Analyze all of the subscripts of this getelementptr instruction
598 // to determine if we can prove known low zero bits.
599 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
600 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
601 ComputeMaskedBits(I->getOperand(0), LocalMask,
602 LocalKnownZero, LocalKnownOne, TD, Depth+1);
603 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
605 gep_type_iterator GTI = gep_type_begin(I);
606 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
607 Value *Index = I->getOperand(i);
608 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
609 // Handle struct member offset arithmetic.
611 const StructLayout *SL = TD->getStructLayout(STy);
612 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
613 uint64_t Offset = SL->getElementOffset(Idx);
614 TrailZ = std::min(TrailZ,
615 CountTrailingZeros_64(Offset));
617 // Handle array index arithmetic.
618 Type *IndexedTy = GTI.getIndexedType();
619 if (!IndexedTy->isSized()) return;
620 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
621 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
622 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
623 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
624 ComputeMaskedBits(Index, LocalMask,
625 LocalKnownZero, LocalKnownOne, TD, Depth+1);
626 TrailZ = std::min(TrailZ,
627 unsigned(CountTrailingZeros_64(TypeSize) +
628 LocalKnownZero.countTrailingOnes()));
632 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
635 case Instruction::PHI: {
636 PHINode *P = cast<PHINode>(I);
637 // Handle the case of a simple two-predecessor recurrence PHI.
638 // There's a lot more that could theoretically be done here, but
639 // this is sufficient to catch some interesting cases.
640 if (P->getNumIncomingValues() == 2) {
641 for (unsigned i = 0; i != 2; ++i) {
642 Value *L = P->getIncomingValue(i);
643 Value *R = P->getIncomingValue(!i);
644 Operator *LU = dyn_cast<Operator>(L);
647 unsigned Opcode = LU->getOpcode();
648 // Check for operations that have the property that if
649 // both their operands have low zero bits, the result
650 // will have low zero bits.
651 if (Opcode == Instruction::Add ||
652 Opcode == Instruction::Sub ||
653 Opcode == Instruction::And ||
654 Opcode == Instruction::Or ||
655 Opcode == Instruction::Mul) {
656 Value *LL = LU->getOperand(0);
657 Value *LR = LU->getOperand(1);
658 // Find a recurrence.
665 // Ok, we have a PHI of the form L op= R. Check for low
667 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
668 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
669 Mask2 = APInt::getLowBitsSet(BitWidth,
670 KnownZero2.countTrailingOnes());
672 // We need to take the minimum number of known bits
673 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
674 ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
677 APInt::getLowBitsSet(BitWidth,
678 std::min(KnownZero2.countTrailingOnes(),
679 KnownZero3.countTrailingOnes()));
685 // Unreachable blocks may have zero-operand PHI nodes.
686 if (P->getNumIncomingValues() == 0)
689 // Otherwise take the unions of the known bit sets of the operands,
690 // taking conservative care to avoid excessive recursion.
691 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
692 // Skip if every incoming value references to ourself.
693 if (P->hasConstantValue() == P)
698 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
699 // Skip direct self references.
700 if (P->getIncomingValue(i) == P) continue;
702 KnownZero2 = APInt(BitWidth, 0);
703 KnownOne2 = APInt(BitWidth, 0);
704 // Recurse, but cap the recursion to one level, because we don't
705 // want to waste time spinning around in loops.
706 ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
707 KnownZero2, KnownOne2, TD, MaxDepth-1);
708 KnownZero &= KnownZero2;
709 KnownOne &= KnownOne2;
710 // If all bits have been ruled out, there's no need to check
712 if (!KnownZero && !KnownOne)
718 case Instruction::Call:
719 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
720 switch (II->getIntrinsicID()) {
722 case Intrinsic::ctlz:
723 case Intrinsic::cttz: {
724 unsigned LowBits = Log2_32(BitWidth)+1;
725 // If this call is undefined for 0, the result will be less than 2^n.
726 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
728 KnownZero = Mask & APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
731 case Intrinsic::ctpop: {
732 unsigned LowBits = Log2_32(BitWidth)+1;
733 KnownZero = Mask & APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
736 case Intrinsic::x86_sse42_crc32_64_8:
737 case Intrinsic::x86_sse42_crc32_64_64:
738 KnownZero = Mask & APInt::getHighBitsSet(64, 32);
746 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
747 /// one. Convenience wrapper around ComputeMaskedBits.
748 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
749 const TargetData *TD, unsigned Depth) {
750 unsigned BitWidth = getBitWidth(V->getType(), TD);
756 APInt ZeroBits(BitWidth, 0);
757 APInt OneBits(BitWidth, 0);
758 ComputeMaskedBits(V, APInt::getSignBit(BitWidth), ZeroBits, OneBits, TD,
760 KnownOne = OneBits[BitWidth - 1];
761 KnownZero = ZeroBits[BitWidth - 1];
764 /// isPowerOfTwo - Return true if the given value is known to have exactly one
765 /// bit set when defined. For vectors return true if every element is known to
766 /// be a power of two when defined. Supports values with integer or pointer
767 /// types and vectors of integers.
768 bool llvm::isPowerOfTwo(Value *V, const TargetData *TD, bool OrZero,
770 if (Constant *C = dyn_cast<Constant>(V)) {
771 if (C->isNullValue())
773 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
774 return CI->getValue().isPowerOf2();
775 // TODO: Handle vector constants.
778 // 1 << X is clearly a power of two if the one is not shifted off the end. If
779 // it is shifted off the end then the result is undefined.
780 if (match(V, m_Shl(m_One(), m_Value())))
783 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
784 // bottom. If it is shifted off the bottom then the result is undefined.
785 if (match(V, m_LShr(m_SignBit(), m_Value())))
788 // The remaining tests are all recursive, so bail out if we hit the limit.
789 if (Depth++ == MaxDepth)
792 Value *X = 0, *Y = 0;
793 // A shift of a power of two is a power of two or zero.
794 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
795 match(V, m_Shr(m_Value(X), m_Value()))))
796 return isPowerOfTwo(X, TD, /*OrZero*/true, Depth);
798 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
799 return isPowerOfTwo(ZI->getOperand(0), TD, OrZero, Depth);
801 if (SelectInst *SI = dyn_cast<SelectInst>(V))
802 return isPowerOfTwo(SI->getTrueValue(), TD, OrZero, Depth) &&
803 isPowerOfTwo(SI->getFalseValue(), TD, OrZero, Depth);
805 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
806 // A power of two and'd with anything is a power of two or zero.
807 if (isPowerOfTwo(X, TD, /*OrZero*/true, Depth) ||
808 isPowerOfTwo(Y, TD, /*OrZero*/true, Depth))
810 // X & (-X) is always a power of two or zero.
811 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
816 // An exact divide or right shift can only shift off zero bits, so the result
817 // is a power of two only if the first operand is a power of two and not
818 // copying a sign bit (sdiv int_min, 2).
819 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
820 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
821 return isPowerOfTwo(cast<Operator>(V)->getOperand(0), TD, OrZero, Depth);
827 /// isKnownNonZero - Return true if the given value is known to be non-zero
828 /// when defined. For vectors return true if every element is known to be
829 /// non-zero when defined. Supports values with integer or pointer type and
830 /// vectors of integers.
831 bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
832 if (Constant *C = dyn_cast<Constant>(V)) {
833 if (C->isNullValue())
835 if (isa<ConstantInt>(C))
836 // Must be non-zero due to null test above.
838 // TODO: Handle vectors
842 // The remaining tests are all recursive, so bail out if we hit the limit.
843 if (Depth++ >= MaxDepth)
846 unsigned BitWidth = getBitWidth(V->getType(), TD);
848 // X | Y != 0 if X != 0 or Y != 0.
849 Value *X = 0, *Y = 0;
850 if (match(V, m_Or(m_Value(X), m_Value(Y))))
851 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
853 // ext X != 0 if X != 0.
854 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
855 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
857 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
858 // if the lowest bit is shifted off the end.
859 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
860 // shl nuw can't remove any non-zero bits.
861 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
862 if (BO->hasNoUnsignedWrap())
863 return isKnownNonZero(X, TD, Depth);
865 APInt KnownZero(BitWidth, 0);
866 APInt KnownOne(BitWidth, 0);
867 ComputeMaskedBits(X, APInt(BitWidth, 1), KnownZero, KnownOne, TD, Depth);
871 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
872 // defined if the sign bit is shifted off the end.
873 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
874 // shr exact can only shift out zero bits.
875 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
877 return isKnownNonZero(X, TD, Depth);
879 bool XKnownNonNegative, XKnownNegative;
880 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
884 // div exact can only produce a zero if the dividend is zero.
885 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
886 return isKnownNonZero(X, TD, Depth);
889 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
890 bool XKnownNonNegative, XKnownNegative;
891 bool YKnownNonNegative, YKnownNegative;
892 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
893 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
895 // If X and Y are both non-negative (as signed values) then their sum is not
896 // zero unless both X and Y are zero.
897 if (XKnownNonNegative && YKnownNonNegative)
898 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
901 // If X and Y are both negative (as signed values) then their sum is not
902 // zero unless both X and Y equal INT_MIN.
903 if (BitWidth && XKnownNegative && YKnownNegative) {
904 APInt KnownZero(BitWidth, 0);
905 APInt KnownOne(BitWidth, 0);
906 APInt Mask = APInt::getSignedMaxValue(BitWidth);
907 // The sign bit of X is set. If some other bit is set then X is not equal
909 ComputeMaskedBits(X, Mask, KnownZero, KnownOne, TD, Depth);
910 if ((KnownOne & Mask) != 0)
912 // The sign bit of Y is set. If some other bit is set then Y is not equal
914 ComputeMaskedBits(Y, Mask, KnownZero, KnownOne, TD, Depth);
915 if ((KnownOne & Mask) != 0)
919 // The sum of a non-negative number and a power of two is not zero.
920 if (XKnownNonNegative && isPowerOfTwo(Y, TD, /*OrZero*/false, Depth))
922 if (YKnownNonNegative && isPowerOfTwo(X, TD, /*OrZero*/false, Depth))
926 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
927 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
928 // If X and Y are non-zero then so is X * Y as long as the multiplication
929 // does not overflow.
930 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
931 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
934 // (C ? X : Y) != 0 if X != 0 and Y != 0.
935 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
936 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
937 isKnownNonZero(SI->getFalseValue(), TD, Depth))
941 if (!BitWidth) return false;
942 APInt KnownZero(BitWidth, 0);
943 APInt KnownOne(BitWidth, 0);
944 ComputeMaskedBits(V, APInt::getAllOnesValue(BitWidth), KnownZero, KnownOne,
946 return KnownOne != 0;
949 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
950 /// this predicate to simplify operations downstream. Mask is known to be zero
951 /// for bits that V cannot have.
953 /// This function is defined on values with integer type, values with pointer
954 /// type (but only if TD is non-null), and vectors of integers. In the case
955 /// where V is a vector, the mask, known zero, and known one values are the
956 /// same width as the vector element, and the bit is set only if it is true
957 /// for all of the elements in the vector.
958 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
959 const TargetData *TD, unsigned Depth) {
960 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
961 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
962 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
963 return (KnownZero & Mask) == Mask;
968 /// ComputeNumSignBits - Return the number of times the sign bit of the
969 /// register is replicated into the other bits. We know that at least 1 bit
970 /// is always equal to the sign bit (itself), but other cases can give us
971 /// information. For example, immediately after an "ashr X, 2", we know that
972 /// the top 3 bits are all equal to each other, so we return 3.
974 /// 'Op' must have a scalar integer type.
976 unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
978 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
979 "ComputeNumSignBits requires a TargetData object to operate "
980 "on non-integer values!");
981 Type *Ty = V->getType();
982 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
983 Ty->getScalarSizeInBits();
985 unsigned FirstAnswer = 1;
987 // Note that ConstantInt is handled by the general ComputeMaskedBits case
991 return 1; // Limit search depth.
993 Operator *U = dyn_cast<Operator>(V);
994 switch (Operator::getOpcode(V)) {
996 case Instruction::SExt:
997 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
998 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
1000 case Instruction::AShr: {
1001 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1002 // ashr X, C -> adds C sign bits. Vectors too.
1004 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1005 Tmp += ShAmt->getZExtValue();
1006 if (Tmp > TyBits) Tmp = TyBits;
1010 case Instruction::Shl: {
1012 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1013 // shl destroys sign bits.
1014 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1015 Tmp2 = ShAmt->getZExtValue();
1016 if (Tmp2 >= TyBits || // Bad shift.
1017 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1022 case Instruction::And:
1023 case Instruction::Or:
1024 case Instruction::Xor: // NOT is handled here.
1025 // Logical binary ops preserve the number of sign bits at the worst.
1026 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1028 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1029 FirstAnswer = std::min(Tmp, Tmp2);
1030 // We computed what we know about the sign bits as our first
1031 // answer. Now proceed to the generic code that uses
1032 // ComputeMaskedBits, and pick whichever answer is better.
1036 case Instruction::Select:
1037 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1038 if (Tmp == 1) return 1; // Early out.
1039 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1040 return std::min(Tmp, Tmp2);
1042 case Instruction::Add:
1043 // Add can have at most one carry bit. Thus we know that the output
1044 // is, at worst, one more bit than the inputs.
1045 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1046 if (Tmp == 1) return 1; // Early out.
1048 // Special case decrementing a value (ADD X, -1):
1049 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1050 if (CRHS->isAllOnesValue()) {
1051 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1052 APInt Mask = APInt::getAllOnesValue(TyBits);
1053 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
1056 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1058 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1061 // If we are subtracting one from a positive number, there is no carry
1062 // out of the result.
1063 if (KnownZero.isNegative())
1067 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1068 if (Tmp2 == 1) return 1;
1069 return std::min(Tmp, Tmp2)-1;
1071 case Instruction::Sub:
1072 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1073 if (Tmp2 == 1) return 1;
1076 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1077 if (CLHS->isNullValue()) {
1078 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1079 APInt Mask = APInt::getAllOnesValue(TyBits);
1080 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
1082 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1084 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1087 // If the input is known to be positive (the sign bit is known clear),
1088 // the output of the NEG has the same number of sign bits as the input.
1089 if (KnownZero.isNegative())
1092 // Otherwise, we treat this like a SUB.
1095 // Sub can have at most one carry bit. Thus we know that the output
1096 // is, at worst, one more bit than the inputs.
1097 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1098 if (Tmp == 1) return 1; // Early out.
1099 return std::min(Tmp, Tmp2)-1;
1101 case Instruction::PHI: {
1102 PHINode *PN = cast<PHINode>(U);
1103 // Don't analyze large in-degree PHIs.
1104 if (PN->getNumIncomingValues() > 4) break;
1106 // Take the minimum of all incoming values. This can't infinitely loop
1107 // because of our depth threshold.
1108 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1109 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1110 if (Tmp == 1) return Tmp;
1112 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1117 case Instruction::Trunc:
1118 // FIXME: it's tricky to do anything useful for this, but it is an important
1119 // case for targets like X86.
1123 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1124 // use this information.
1125 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1126 APInt Mask = APInt::getAllOnesValue(TyBits);
1127 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
1129 if (KnownZero.isNegative()) { // sign bit is 0
1131 } else if (KnownOne.isNegative()) { // sign bit is 1;
1138 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1139 // the number of identical bits in the top of the input value.
1141 Mask <<= Mask.getBitWidth()-TyBits;
1142 // Return # leading zeros. We use 'min' here in case Val was zero before
1143 // shifting. We don't want to return '64' as for an i32 "0".
1144 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1147 /// ComputeMultiple - This function computes the integer multiple of Base that
1148 /// equals V. If successful, it returns true and returns the multiple in
1149 /// Multiple. If unsuccessful, it returns false. It looks
1150 /// through SExt instructions only if LookThroughSExt is true.
1151 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1152 bool LookThroughSExt, unsigned Depth) {
1153 const unsigned MaxDepth = 6;
1155 assert(V && "No Value?");
1156 assert(Depth <= MaxDepth && "Limit Search Depth");
1157 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1159 Type *T = V->getType();
1161 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1171 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1172 Constant *BaseVal = ConstantInt::get(T, Base);
1173 if (CO && CO == BaseVal) {
1175 Multiple = ConstantInt::get(T, 1);
1179 if (CI && CI->getZExtValue() % Base == 0) {
1180 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1184 if (Depth == MaxDepth) return false; // Limit search depth.
1186 Operator *I = dyn_cast<Operator>(V);
1187 if (!I) return false;
1189 switch (I->getOpcode()) {
1191 case Instruction::SExt:
1192 if (!LookThroughSExt) return false;
1193 // otherwise fall through to ZExt
1194 case Instruction::ZExt:
1195 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1196 LookThroughSExt, Depth+1);
1197 case Instruction::Shl:
1198 case Instruction::Mul: {
1199 Value *Op0 = I->getOperand(0);
1200 Value *Op1 = I->getOperand(1);
1202 if (I->getOpcode() == Instruction::Shl) {
1203 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1204 if (!Op1CI) return false;
1205 // Turn Op0 << Op1 into Op0 * 2^Op1
1206 APInt Op1Int = Op1CI->getValue();
1207 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1208 APInt API(Op1Int.getBitWidth(), 0);
1209 API.setBit(BitToSet);
1210 Op1 = ConstantInt::get(V->getContext(), API);
1214 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1215 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1216 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1217 if (Op1C->getType()->getPrimitiveSizeInBits() <
1218 MulC->getType()->getPrimitiveSizeInBits())
1219 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1220 if (Op1C->getType()->getPrimitiveSizeInBits() >
1221 MulC->getType()->getPrimitiveSizeInBits())
1222 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1224 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1225 Multiple = ConstantExpr::getMul(MulC, Op1C);
1229 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1230 if (Mul0CI->getValue() == 1) {
1231 // V == Base * Op1, so return Op1
1238 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1239 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1240 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1241 if (Op0C->getType()->getPrimitiveSizeInBits() <
1242 MulC->getType()->getPrimitiveSizeInBits())
1243 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1244 if (Op0C->getType()->getPrimitiveSizeInBits() >
1245 MulC->getType()->getPrimitiveSizeInBits())
1246 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1248 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1249 Multiple = ConstantExpr::getMul(MulC, Op0C);
1253 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1254 if (Mul1CI->getValue() == 1) {
1255 // V == Base * Op0, so return Op0
1263 // We could not determine if V is a multiple of Base.
1267 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1268 /// value is never equal to -0.0.
1270 /// NOTE: this function will need to be revisited when we support non-default
1273 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1274 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1275 return !CFP->getValueAPF().isNegZero();
1278 return 1; // Limit search depth.
1280 const Operator *I = dyn_cast<Operator>(V);
1281 if (I == 0) return false;
1283 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1284 if (I->getOpcode() == Instruction::FAdd &&
1285 isa<ConstantFP>(I->getOperand(1)) &&
1286 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1289 // sitofp and uitofp turn into +0.0 for zero.
1290 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1293 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1294 // sqrt(-0.0) = -0.0, no other negative results are possible.
1295 if (II->getIntrinsicID() == Intrinsic::sqrt)
1296 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1298 if (const CallInst *CI = dyn_cast<CallInst>(I))
1299 if (const Function *F = CI->getCalledFunction()) {
1300 if (F->isDeclaration()) {
1302 if (F->getName() == "abs") return true;
1303 // fabs[lf](x) != -0.0
1304 if (F->getName() == "fabs") return true;
1305 if (F->getName() == "fabsf") return true;
1306 if (F->getName() == "fabsl") return true;
1307 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1308 F->getName() == "sqrtl")
1309 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1316 /// isBytewiseValue - If the specified value can be set by repeating the same
1317 /// byte in memory, return the i8 value that it is represented with. This is
1318 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1319 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1320 /// byte store (e.g. i16 0x1234), return null.
1321 Value *llvm::isBytewiseValue(Value *V) {
1322 // All byte-wide stores are splatable, even of arbitrary variables.
1323 if (V->getType()->isIntegerTy(8)) return V;
1325 // Handle 'null' ConstantArrayZero etc.
1326 if (Constant *C = dyn_cast<Constant>(V))
1327 if (C->isNullValue())
1328 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1330 // Constant float and double values can be handled as integer values if the
1331 // corresponding integer value is "byteable". An important case is 0.0.
1332 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1333 if (CFP->getType()->isFloatTy())
1334 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1335 if (CFP->getType()->isDoubleTy())
1336 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1337 // Don't handle long double formats, which have strange constraints.
1340 // We can handle constant integers that are power of two in size and a
1341 // multiple of 8 bits.
1342 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1343 unsigned Width = CI->getBitWidth();
1344 if (isPowerOf2_32(Width) && Width > 8) {
1345 // We can handle this value if the recursive binary decomposition is the
1346 // same at all levels.
1347 APInt Val = CI->getValue();
1349 while (Val.getBitWidth() != 8) {
1350 unsigned NextWidth = Val.getBitWidth()/2;
1351 Val2 = Val.lshr(NextWidth);
1352 Val2 = Val2.trunc(Val.getBitWidth()/2);
1353 Val = Val.trunc(Val.getBitWidth()/2);
1355 // If the top/bottom halves aren't the same, reject it.
1359 return ConstantInt::get(V->getContext(), Val);
1363 // A ConstantDataArray/Vector is splatable if all its members are equal and
1365 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1366 Value *Elt = CA->getElementAsConstant(0);
1367 Value *Val = isBytewiseValue(Elt);
1371 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1372 if (CA->getElementAsConstant(I) != Elt)
1378 // Conceptually, we could handle things like:
1379 // %a = zext i8 %X to i16
1380 // %b = shl i16 %a, 8
1381 // %c = or i16 %a, %b
1382 // but until there is an example that actually needs this, it doesn't seem
1383 // worth worrying about.
1388 // This is the recursive version of BuildSubAggregate. It takes a few different
1389 // arguments. Idxs is the index within the nested struct From that we are
1390 // looking at now (which is of type IndexedType). IdxSkip is the number of
1391 // indices from Idxs that should be left out when inserting into the resulting
1392 // struct. To is the result struct built so far, new insertvalue instructions
1394 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1395 SmallVector<unsigned, 10> &Idxs,
1397 Instruction *InsertBefore) {
1398 llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1400 // Save the original To argument so we can modify it
1402 // General case, the type indexed by Idxs is a struct
1403 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1404 // Process each struct element recursively
1407 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1411 // Couldn't find any inserted value for this index? Cleanup
1412 while (PrevTo != OrigTo) {
1413 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1414 PrevTo = Del->getAggregateOperand();
1415 Del->eraseFromParent();
1417 // Stop processing elements
1421 // If we successfully found a value for each of our subaggregates
1425 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1426 // the struct's elements had a value that was inserted directly. In the latter
1427 // case, perhaps we can't determine each of the subelements individually, but
1428 // we might be able to find the complete struct somewhere.
1430 // Find the value that is at that particular spot
1431 Value *V = FindInsertedValue(From, Idxs);
1436 // Insert the value in the new (sub) aggregrate
1437 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1438 "tmp", InsertBefore);
1441 // This helper takes a nested struct and extracts a part of it (which is again a
1442 // struct) into a new value. For example, given the struct:
1443 // { a, { b, { c, d }, e } }
1444 // and the indices "1, 1" this returns
1447 // It does this by inserting an insertvalue for each element in the resulting
1448 // struct, as opposed to just inserting a single struct. This will only work if
1449 // each of the elements of the substruct are known (ie, inserted into From by an
1450 // insertvalue instruction somewhere).
1452 // All inserted insertvalue instructions are inserted before InsertBefore
1453 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1454 Instruction *InsertBefore) {
1455 assert(InsertBefore && "Must have someplace to insert!");
1456 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1458 Value *To = UndefValue::get(IndexedType);
1459 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1460 unsigned IdxSkip = Idxs.size();
1462 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1465 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1466 /// the scalar value indexed is already around as a register, for example if it
1467 /// were inserted directly into the aggregrate.
1469 /// If InsertBefore is not null, this function will duplicate (modified)
1470 /// insertvalues when a part of a nested struct is extracted.
1471 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1472 Instruction *InsertBefore) {
1473 // Nothing to index? Just return V then (this is useful at the end of our
1475 if (idx_range.empty())
1477 // We have indices, so V should have an indexable type.
1478 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1479 "Not looking at a struct or array?");
1480 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1481 "Invalid indices for type?");
1483 if (Constant *C = dyn_cast<Constant>(V)) {
1484 C = C->getAggregateElement(idx_range[0]);
1485 if (C == 0) return 0;
1486 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1489 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1490 // Loop the indices for the insertvalue instruction in parallel with the
1491 // requested indices
1492 const unsigned *req_idx = idx_range.begin();
1493 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1494 i != e; ++i, ++req_idx) {
1495 if (req_idx == idx_range.end()) {
1496 // We can't handle this without inserting insertvalues
1500 // The requested index identifies a part of a nested aggregate. Handle
1501 // this specially. For example,
1502 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1503 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1504 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1505 // This can be changed into
1506 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1507 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1508 // which allows the unused 0,0 element from the nested struct to be
1510 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1514 // This insert value inserts something else than what we are looking for.
1515 // See if the (aggregrate) value inserted into has the value we are
1516 // looking for, then.
1518 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1521 // If we end up here, the indices of the insertvalue match with those
1522 // requested (though possibly only partially). Now we recursively look at
1523 // the inserted value, passing any remaining indices.
1524 return FindInsertedValue(I->getInsertedValueOperand(),
1525 makeArrayRef(req_idx, idx_range.end()),
1529 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1530 // If we're extracting a value from an aggregrate that was extracted from
1531 // something else, we can extract from that something else directly instead.
1532 // However, we will need to chain I's indices with the requested indices.
1534 // Calculate the number of indices required
1535 unsigned size = I->getNumIndices() + idx_range.size();
1536 // Allocate some space to put the new indices in
1537 SmallVector<unsigned, 5> Idxs;
1539 // Add indices from the extract value instruction
1540 Idxs.append(I->idx_begin(), I->idx_end());
1542 // Add requested indices
1543 Idxs.append(idx_range.begin(), idx_range.end());
1545 assert(Idxs.size() == size
1546 && "Number of indices added not correct?");
1548 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1550 // Otherwise, we don't know (such as, extracting from a function return value
1551 // or load instruction)
1555 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1556 /// it can be expressed as a base pointer plus a constant offset. Return the
1557 /// base and offset to the caller.
1558 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1559 const TargetData &TD) {
1560 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1561 if (PtrOp == 0 || Ptr->getType()->isVectorTy())
1564 // Just look through bitcasts.
1565 if (PtrOp->getOpcode() == Instruction::BitCast)
1566 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1568 // If this is a GEP with constant indices, we can look through it.
1569 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1570 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1572 gep_type_iterator GTI = gep_type_begin(GEP);
1573 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1575 ConstantInt *OpC = cast<ConstantInt>(*I);
1576 if (OpC->isZero()) continue;
1578 // Handle a struct and array indices which add their offset to the pointer.
1579 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1580 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1582 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1583 Offset += OpC->getSExtValue()*Size;
1587 // Re-sign extend from the pointer size if needed to get overflow edge cases
1589 unsigned PtrSize = TD.getPointerSizeInBits();
1591 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1593 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1597 /// getConstantStringInfo - This function computes the length of a
1598 /// null-terminated C string pointed to by V. If successful, it returns true
1599 /// and returns the string in Str. If unsuccessful, it returns false.
1600 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
1601 uint64_t Offset, bool TrimAtNul) {
1604 // Look through bitcast instructions and geps.
1605 V = V->stripPointerCasts();
1607 // If the value is a GEP instructionor constant expression, treat it as an
1609 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1610 // Make sure the GEP has exactly three arguments.
1611 if (GEP->getNumOperands() != 3)
1614 // Make sure the index-ee is a pointer to array of i8.
1615 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1616 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1617 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1620 // Check to make sure that the first operand of the GEP is an integer and
1621 // has value 0 so that we are sure we're indexing into the initializer.
1622 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1623 if (FirstIdx == 0 || !FirstIdx->isZero())
1626 // If the second index isn't a ConstantInt, then this is a variable index
1627 // into the array. If this occurs, we can't say anything meaningful about
1629 uint64_t StartIdx = 0;
1630 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1631 StartIdx = CI->getZExtValue();
1634 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
1637 // The GEP instruction, constant or instruction, must reference a global
1638 // variable that is a constant and is initialized. The referenced constant
1639 // initializer is the array that we'll use for optimization.
1640 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
1641 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1644 // Handle the all-zeros case
1645 if (GV->getInitializer()->isNullValue()) {
1646 // This is a degenerate case. The initializer is constant zero so the
1647 // length of the string must be zero.
1652 // Must be a Constant Array
1653 const ConstantDataArray *Array =
1654 dyn_cast<ConstantDataArray>(GV->getInitializer());
1655 if (Array == 0 || !Array->isString())
1658 // Get the number of elements in the array
1659 uint64_t NumElts = Array->getType()->getArrayNumElements();
1661 // Start out with the entire array in the StringRef.
1662 Str = Array->getAsString();
1664 if (Offset > NumElts)
1667 // Skip over 'offset' bytes.
1668 Str = Str.substr(Offset);
1671 // Trim off the \0 and anything after it. If the array is not nul
1672 // terminated, we just return the whole end of string. The client may know
1673 // some other way that the string is length-bound.
1674 Str = Str.substr(0, Str.find('\0'));
1679 // These next two are very similar to the above, but also look through PHI
1681 // TODO: See if we can integrate these two together.
1683 /// GetStringLengthH - If we can compute the length of the string pointed to by
1684 /// the specified pointer, return 'len+1'. If we can't, return 0.
1685 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1686 // Look through noop bitcast instructions.
1687 V = V->stripPointerCasts();
1689 // If this is a PHI node, there are two cases: either we have already seen it
1691 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1692 if (!PHIs.insert(PN))
1693 return ~0ULL; // already in the set.
1695 // If it was new, see if all the input strings are the same length.
1696 uint64_t LenSoFar = ~0ULL;
1697 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1698 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1699 if (Len == 0) return 0; // Unknown length -> unknown.
1701 if (Len == ~0ULL) continue;
1703 if (Len != LenSoFar && LenSoFar != ~0ULL)
1704 return 0; // Disagree -> unknown.
1708 // Success, all agree.
1712 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1713 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1714 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1715 if (Len1 == 0) return 0;
1716 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1717 if (Len2 == 0) return 0;
1718 if (Len1 == ~0ULL) return Len2;
1719 if (Len2 == ~0ULL) return Len1;
1720 if (Len1 != Len2) return 0;
1724 // Otherwise, see if we can read the string.
1726 if (!getConstantStringInfo(V, StrData))
1729 return StrData.size()+1;
1732 /// GetStringLength - If we can compute the length of the string pointed to by
1733 /// the specified pointer, return 'len+1'. If we can't, return 0.
1734 uint64_t llvm::GetStringLength(Value *V) {
1735 if (!V->getType()->isPointerTy()) return 0;
1737 SmallPtrSet<PHINode*, 32> PHIs;
1738 uint64_t Len = GetStringLengthH(V, PHIs);
1739 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1740 // an empty string as a length.
1741 return Len == ~0ULL ? 1 : Len;
1745 llvm::GetUnderlyingObject(Value *V, const TargetData *TD, unsigned MaxLookup) {
1746 if (!V->getType()->isPointerTy())
1748 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1749 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1750 V = GEP->getPointerOperand();
1751 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1752 V = cast<Operator>(V)->getOperand(0);
1753 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1754 if (GA->mayBeOverridden())
1756 V = GA->getAliasee();
1758 // See if InstructionSimplify knows any relevant tricks.
1759 if (Instruction *I = dyn_cast<Instruction>(V))
1760 // TODO: Acquire a DominatorTree and use it.
1761 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1768 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1773 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1774 /// are lifetime markers.
1776 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1777 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1779 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1780 if (!II) return false;
1782 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1783 II->getIntrinsicID() != Intrinsic::lifetime_end)
1789 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1790 const TargetData *TD) {
1791 const Operator *Inst = dyn_cast<Operator>(V);
1795 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1796 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1800 switch (Inst->getOpcode()) {
1803 case Instruction::UDiv:
1804 case Instruction::URem:
1805 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1806 return isKnownNonZero(Inst->getOperand(1), TD);
1807 case Instruction::SDiv:
1808 case Instruction::SRem: {
1809 Value *Op = Inst->getOperand(1);
1810 // x / y is undefined if y == 0
1811 if (!isKnownNonZero(Op, TD))
1813 // x / y might be undefined if y == -1
1814 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1817 APInt KnownZero(BitWidth, 0);
1818 APInt KnownOne(BitWidth, 0);
1819 ComputeMaskedBits(Op, APInt::getAllOnesValue(BitWidth),
1820 KnownZero, KnownOne, TD);
1823 case Instruction::Load: {
1824 const LoadInst *LI = cast<LoadInst>(Inst);
1825 if (!LI->isUnordered())
1827 return LI->getPointerOperand()->isDereferenceablePointer();
1829 case Instruction::Call: {
1830 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1831 switch (II->getIntrinsicID()) {
1832 case Intrinsic::bswap:
1833 case Intrinsic::ctlz:
1834 case Intrinsic::ctpop:
1835 case Intrinsic::cttz:
1836 case Intrinsic::objectsize:
1837 case Intrinsic::sadd_with_overflow:
1838 case Intrinsic::smul_with_overflow:
1839 case Intrinsic::ssub_with_overflow:
1840 case Intrinsic::uadd_with_overflow:
1841 case Intrinsic::umul_with_overflow:
1842 case Intrinsic::usub_with_overflow:
1844 // TODO: some fp intrinsics are marked as having the same error handling
1845 // as libm. They're safe to speculate when they won't error.
1846 // TODO: are convert_{from,to}_fp16 safe?
1847 // TODO: can we list target-specific intrinsics here?
1851 return false; // The called function could have undefined behavior or
1852 // side-effects, even if marked readnone nounwind.
1854 case Instruction::VAArg:
1855 case Instruction::Alloca:
1856 case Instruction::Invoke:
1857 case Instruction::PHI:
1858 case Instruction::Store:
1859 case Instruction::Ret:
1860 case Instruction::Br:
1861 case Instruction::IndirectBr:
1862 case Instruction::Switch:
1863 case Instruction::Unreachable:
1864 case Instruction::Fence:
1865 case Instruction::LandingPad:
1866 case Instruction::AtomicRMW:
1867 case Instruction::AtomicCmpXchg:
1868 case Instruction::Resume:
1869 return false; // Misc instructions which have effects