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 && GV->getType()->getElementType()->isSized()) {
110 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
111 Type *ObjectType = GVar->getType()->getElementType();
112 // If the object is defined in the current Module, we'll be giving
113 // it the preferred alignment. Otherwise, we have to assume that it
114 // may only have the minimum ABI alignment.
115 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
116 Align = TD->getPreferredAlignment(GVar);
118 Align = TD->getABITypeAlignment(ObjectType);
122 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
123 CountTrailingZeros_32(Align));
125 KnownZero.clearAllBits();
126 KnownOne.clearAllBits();
129 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
130 // the bits of its aliasee.
131 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
132 if (GA->mayBeOverridden()) {
133 KnownZero.clearAllBits(); KnownOne.clearAllBits();
135 ComputeMaskedBits(GA->getAliasee(), Mask, KnownZero, KnownOne,
141 if (Argument *A = dyn_cast<Argument>(V)) {
142 // Get alignment information off byval arguments if specified in the IR.
143 if (A->hasByValAttr())
144 if (unsigned Align = A->getParamAlignment())
145 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
146 CountTrailingZeros_32(Align));
150 // Start out not knowing anything.
151 KnownZero.clearAllBits(); KnownOne.clearAllBits();
153 if (Depth == MaxDepth || Mask == 0)
154 return; // Limit search depth.
156 Operator *I = dyn_cast<Operator>(V);
159 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
160 switch (I->getOpcode()) {
162 case Instruction::And: {
163 // If either the LHS or the RHS are Zero, the result is zero.
164 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
165 APInt Mask2(Mask & ~KnownZero);
166 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
168 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
169 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
171 // Output known-1 bits are only known if set in both the LHS & RHS.
172 KnownOne &= KnownOne2;
173 // Output known-0 are known to be clear if zero in either the LHS | RHS.
174 KnownZero |= KnownZero2;
177 case Instruction::Or: {
178 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
179 APInt Mask2(Mask & ~KnownOne);
180 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
182 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
183 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
185 // Output known-0 bits are only known if clear in both the LHS & RHS.
186 KnownZero &= KnownZero2;
187 // Output known-1 are known to be set if set in either the LHS | RHS.
188 KnownOne |= KnownOne2;
191 case Instruction::Xor: {
192 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
193 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
195 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
196 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
198 // Output known-0 bits are known if clear or set in both the LHS & RHS.
199 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
200 // Output known-1 are known to be set if set in only one of the LHS, RHS.
201 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
202 KnownZero = KnownZeroOut;
205 case Instruction::Mul: {
206 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
207 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
208 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
210 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
211 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
213 bool isKnownNegative = false;
214 bool isKnownNonNegative = false;
215 // If the multiplication is known not to overflow, compute the sign bit.
216 if (Mask.isNegative() &&
217 cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap()) {
218 Value *Op1 = I->getOperand(1), *Op2 = I->getOperand(0);
220 // The product of a number with itself is non-negative.
221 isKnownNonNegative = true;
223 bool isKnownNonNegative1 = KnownZero.isNegative();
224 bool isKnownNonNegative2 = KnownZero2.isNegative();
225 bool isKnownNegative1 = KnownOne.isNegative();
226 bool isKnownNegative2 = KnownOne2.isNegative();
227 // The product of two numbers with the same sign is non-negative.
228 isKnownNonNegative = (isKnownNegative1 && isKnownNegative2) ||
229 (isKnownNonNegative1 && isKnownNonNegative2);
230 // The product of a negative number and a non-negative number is either
232 if (!isKnownNonNegative)
233 isKnownNegative = (isKnownNegative1 && isKnownNonNegative2 &&
234 isKnownNonZero(Op2, TD, Depth)) ||
235 (isKnownNegative2 && isKnownNonNegative1 &&
236 isKnownNonZero(Op1, TD, Depth));
240 // If low bits are zero in either operand, output low known-0 bits.
241 // Also compute a conserative estimate for high known-0 bits.
242 // More trickiness is possible, but this is sufficient for the
243 // interesting case of alignment computation.
244 KnownOne.clearAllBits();
245 unsigned TrailZ = KnownZero.countTrailingOnes() +
246 KnownZero2.countTrailingOnes();
247 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
248 KnownZero2.countLeadingOnes(),
249 BitWidth) - BitWidth;
251 TrailZ = std::min(TrailZ, BitWidth);
252 LeadZ = std::min(LeadZ, BitWidth);
253 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
254 APInt::getHighBitsSet(BitWidth, LeadZ);
257 // Only make use of no-wrap flags if we failed to compute the sign bit
258 // directly. This matters if the multiplication always overflows, in
259 // which case we prefer to follow the result of the direct computation,
260 // though as the program is invoking undefined behaviour we can choose
261 // whatever we like here.
262 if (isKnownNonNegative && !KnownOne.isNegative())
263 KnownZero.setBit(BitWidth - 1);
264 else if (isKnownNegative && !KnownZero.isNegative())
265 KnownOne.setBit(BitWidth - 1);
269 case Instruction::UDiv: {
270 // For the purposes of computing leading zeros we can conservatively
271 // treat a udiv as a logical right shift by the power of 2 known to
272 // be less than the denominator.
273 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
274 ComputeMaskedBits(I->getOperand(0),
275 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
276 unsigned LeadZ = KnownZero2.countLeadingOnes();
278 KnownOne2.clearAllBits();
279 KnownZero2.clearAllBits();
280 ComputeMaskedBits(I->getOperand(1),
281 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
282 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
283 if (RHSUnknownLeadingOnes != BitWidth)
284 LeadZ = std::min(BitWidth,
285 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
287 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
290 case Instruction::Select:
291 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
292 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
294 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
295 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
297 // Only known if known in both the LHS and RHS.
298 KnownOne &= KnownOne2;
299 KnownZero &= KnownZero2;
301 case Instruction::FPTrunc:
302 case Instruction::FPExt:
303 case Instruction::FPToUI:
304 case Instruction::FPToSI:
305 case Instruction::SIToFP:
306 case Instruction::UIToFP:
307 return; // Can't work with floating point.
308 case Instruction::PtrToInt:
309 case Instruction::IntToPtr:
310 // We can't handle these if we don't know the pointer size.
312 // FALL THROUGH and handle them the same as zext/trunc.
313 case Instruction::ZExt:
314 case Instruction::Trunc: {
315 Type *SrcTy = I->getOperand(0)->getType();
317 unsigned SrcBitWidth;
318 // Note that we handle pointer operands here because of inttoptr/ptrtoint
319 // which fall through here.
320 if (SrcTy->isPointerTy())
321 SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
323 SrcBitWidth = SrcTy->getScalarSizeInBits();
325 APInt MaskIn = Mask.zextOrTrunc(SrcBitWidth);
326 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
327 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
328 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
330 KnownZero = KnownZero.zextOrTrunc(BitWidth);
331 KnownOne = KnownOne.zextOrTrunc(BitWidth);
332 // Any top bits are known to be zero.
333 if (BitWidth > SrcBitWidth)
334 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
337 case Instruction::BitCast: {
338 Type *SrcTy = I->getOperand(0)->getType();
339 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
340 // TODO: For now, not handling conversions like:
341 // (bitcast i64 %x to <2 x i32>)
342 !I->getType()->isVectorTy()) {
343 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
349 case Instruction::SExt: {
350 // Compute the bits in the result that are not present in the input.
351 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
353 APInt MaskIn = Mask.trunc(SrcBitWidth);
354 KnownZero = KnownZero.trunc(SrcBitWidth);
355 KnownOne = KnownOne.trunc(SrcBitWidth);
356 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
358 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
359 KnownZero = KnownZero.zext(BitWidth);
360 KnownOne = KnownOne.zext(BitWidth);
362 // If the sign bit of the input is known set or clear, then we know the
363 // top bits of the result.
364 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
365 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
366 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
367 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
370 case Instruction::Shl:
371 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
372 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
373 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
374 APInt Mask2(Mask.lshr(ShiftAmt));
375 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
377 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
378 KnownZero <<= ShiftAmt;
379 KnownOne <<= ShiftAmt;
380 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
384 case Instruction::LShr:
385 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
386 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
387 // Compute the new bits that are at the top now.
388 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
390 // Unsigned shift right.
391 APInt Mask2(Mask.shl(ShiftAmt));
392 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
394 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
395 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
396 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
397 // high bits known zero.
398 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
402 case Instruction::AShr:
403 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
404 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
405 // Compute the new bits that are at the top now.
406 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
408 // Signed shift right.
409 APInt Mask2(Mask.shl(ShiftAmt));
410 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
412 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
413 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
414 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
416 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
417 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
418 KnownZero |= HighBits;
419 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
420 KnownOne |= HighBits;
424 case Instruction::Sub: {
425 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
426 // We know that the top bits of C-X are clear if X contains less bits
427 // than C (i.e. no wrap-around can happen). For example, 20-X is
428 // positive if we can prove that X is >= 0 and < 16.
429 if (!CLHS->getValue().isNegative()) {
430 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
431 // NLZ can't be BitWidth with no sign bit
432 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
433 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
436 // If all of the MaskV bits are known to be zero, then we know the
437 // output top bits are zero, because we now know that the output is
439 if ((KnownZero2 & MaskV) == MaskV) {
440 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
441 // Top bits known zero.
442 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
448 case Instruction::Add: {
449 // If one of the operands has trailing zeros, then the bits that the
450 // other operand has in those bit positions will be preserved in the
451 // result. For an add, this works with either operand. For a subtract,
452 // this only works if the known zeros are in the right operand.
453 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
454 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
455 BitWidth - Mask.countLeadingZeros());
456 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
458 assert((LHSKnownZero & LHSKnownOne) == 0 &&
459 "Bits known to be one AND zero?");
460 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
462 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD,
464 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
465 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
467 // Determine which operand has more trailing zeros, and use that
468 // many bits from the other operand.
469 if (LHSKnownZeroOut > RHSKnownZeroOut) {
470 if (I->getOpcode() == Instruction::Add) {
471 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
472 KnownZero |= KnownZero2 & Mask;
473 KnownOne |= KnownOne2 & Mask;
475 // If the known zeros are in the left operand for a subtract,
476 // fall back to the minimum known zeros in both operands.
477 KnownZero |= APInt::getLowBitsSet(BitWidth,
478 std::min(LHSKnownZeroOut,
481 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
482 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
483 KnownZero |= LHSKnownZero & Mask;
484 KnownOne |= LHSKnownOne & Mask;
487 // Are we still trying to solve for the sign bit?
488 if (Mask.isNegative() && !KnownZero.isNegative() && !KnownOne.isNegative()){
489 OverflowingBinaryOperator *OBO = cast<OverflowingBinaryOperator>(I);
490 if (OBO->hasNoSignedWrap()) {
491 if (I->getOpcode() == Instruction::Add) {
492 // Adding two positive numbers can't wrap into negative
493 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
494 KnownZero |= APInt::getSignBit(BitWidth);
495 // and adding two negative numbers can't wrap into positive.
496 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
497 KnownOne |= APInt::getSignBit(BitWidth);
499 // Subtracting a negative number from a positive one can't wrap
500 if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
501 KnownZero |= APInt::getSignBit(BitWidth);
502 // neither can subtracting a positive number from a negative one.
503 else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
504 KnownOne |= APInt::getSignBit(BitWidth);
511 case Instruction::SRem:
512 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
513 APInt RA = Rem->getValue().abs();
514 if (RA.isPowerOf2()) {
515 APInt LowBits = RA - 1;
516 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
517 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
520 // The low bits of the first operand are unchanged by the srem.
521 KnownZero = KnownZero2 & LowBits;
522 KnownOne = KnownOne2 & LowBits;
524 // If the first operand is non-negative or has all low bits zero, then
525 // the upper bits are all zero.
526 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
527 KnownZero |= ~LowBits;
529 // If the first operand is negative and not all low bits are zero, then
530 // the upper bits are all one.
531 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
532 KnownOne |= ~LowBits;
537 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
541 // The sign bit is the LHS's sign bit, except when the result of the
542 // remainder is zero.
543 if (Mask.isNegative() && KnownZero.isNonNegative()) {
544 APInt Mask2 = APInt::getSignBit(BitWidth);
545 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
546 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
548 // If it's known zero, our sign bit is also zero.
549 if (LHSKnownZero.isNegative())
550 KnownZero |= LHSKnownZero;
554 case Instruction::URem: {
555 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
556 APInt RA = Rem->getValue();
557 if (RA.isPowerOf2()) {
558 APInt LowBits = (RA - 1);
559 APInt Mask2 = LowBits & Mask;
560 KnownZero |= ~LowBits & Mask;
561 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
563 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
568 // Since the result is less than or equal to either operand, any leading
569 // zero bits in either operand must also exist in the result.
570 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
571 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
573 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
576 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
577 KnownZero2.countLeadingOnes());
578 KnownOne.clearAllBits();
579 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
583 case Instruction::Alloca: {
584 AllocaInst *AI = cast<AllocaInst>(V);
585 unsigned Align = AI->getAlignment();
586 if (Align == 0 && TD)
587 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
590 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
591 CountTrailingZeros_32(Align));
594 case Instruction::GetElementPtr: {
595 // Analyze all of the subscripts of this getelementptr instruction
596 // to determine if we can prove known low zero bits.
597 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
598 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
599 ComputeMaskedBits(I->getOperand(0), LocalMask,
600 LocalKnownZero, LocalKnownOne, TD, Depth+1);
601 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
603 gep_type_iterator GTI = gep_type_begin(I);
604 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
605 Value *Index = I->getOperand(i);
606 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
607 // Handle struct member offset arithmetic.
609 const StructLayout *SL = TD->getStructLayout(STy);
610 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
611 uint64_t Offset = SL->getElementOffset(Idx);
612 TrailZ = std::min(TrailZ,
613 CountTrailingZeros_64(Offset));
615 // Handle array index arithmetic.
616 Type *IndexedTy = GTI.getIndexedType();
617 if (!IndexedTy->isSized()) return;
618 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
619 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
620 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
621 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
622 ComputeMaskedBits(Index, LocalMask,
623 LocalKnownZero, LocalKnownOne, TD, Depth+1);
624 TrailZ = std::min(TrailZ,
625 unsigned(CountTrailingZeros_64(TypeSize) +
626 LocalKnownZero.countTrailingOnes()));
630 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
633 case Instruction::PHI: {
634 PHINode *P = cast<PHINode>(I);
635 // Handle the case of a simple two-predecessor recurrence PHI.
636 // There's a lot more that could theoretically be done here, but
637 // this is sufficient to catch some interesting cases.
638 if (P->getNumIncomingValues() == 2) {
639 for (unsigned i = 0; i != 2; ++i) {
640 Value *L = P->getIncomingValue(i);
641 Value *R = P->getIncomingValue(!i);
642 Operator *LU = dyn_cast<Operator>(L);
645 unsigned Opcode = LU->getOpcode();
646 // Check for operations that have the property that if
647 // both their operands have low zero bits, the result
648 // will have low zero bits.
649 if (Opcode == Instruction::Add ||
650 Opcode == Instruction::Sub ||
651 Opcode == Instruction::And ||
652 Opcode == Instruction::Or ||
653 Opcode == Instruction::Mul) {
654 Value *LL = LU->getOperand(0);
655 Value *LR = LU->getOperand(1);
656 // Find a recurrence.
663 // Ok, we have a PHI of the form L op= R. Check for low
665 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
666 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
667 Mask2 = APInt::getLowBitsSet(BitWidth,
668 KnownZero2.countTrailingOnes());
670 // We need to take the minimum number of known bits
671 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
672 ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
675 APInt::getLowBitsSet(BitWidth,
676 std::min(KnownZero2.countTrailingOnes(),
677 KnownZero3.countTrailingOnes()));
683 // Unreachable blocks may have zero-operand PHI nodes.
684 if (P->getNumIncomingValues() == 0)
687 // Otherwise take the unions of the known bit sets of the operands,
688 // taking conservative care to avoid excessive recursion.
689 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
690 // Skip if every incoming value references to ourself.
691 if (P->hasConstantValue() == P)
696 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
697 // Skip direct self references.
698 if (P->getIncomingValue(i) == P) continue;
700 KnownZero2 = APInt(BitWidth, 0);
701 KnownOne2 = APInt(BitWidth, 0);
702 // Recurse, but cap the recursion to one level, because we don't
703 // want to waste time spinning around in loops.
704 ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
705 KnownZero2, KnownOne2, TD, MaxDepth-1);
706 KnownZero &= KnownZero2;
707 KnownOne &= KnownOne2;
708 // If all bits have been ruled out, there's no need to check
710 if (!KnownZero && !KnownOne)
716 case Instruction::Call:
717 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
718 switch (II->getIntrinsicID()) {
720 case Intrinsic::ctlz:
721 case Intrinsic::cttz: {
722 unsigned LowBits = Log2_32(BitWidth)+1;
723 // If this call is undefined for 0, the result will be less than 2^n.
724 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
726 KnownZero = Mask & APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
729 case Intrinsic::ctpop: {
730 unsigned LowBits = Log2_32(BitWidth)+1;
731 KnownZero = Mask & APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
734 case Intrinsic::x86_sse42_crc32_64_8:
735 case Intrinsic::x86_sse42_crc32_64_64:
736 KnownZero = Mask & APInt::getHighBitsSet(64, 32);
744 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
745 /// one. Convenience wrapper around ComputeMaskedBits.
746 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
747 const TargetData *TD, unsigned Depth) {
748 unsigned BitWidth = getBitWidth(V->getType(), TD);
754 APInt ZeroBits(BitWidth, 0);
755 APInt OneBits(BitWidth, 0);
756 ComputeMaskedBits(V, APInt::getSignBit(BitWidth), ZeroBits, OneBits, TD,
758 KnownOne = OneBits[BitWidth - 1];
759 KnownZero = ZeroBits[BitWidth - 1];
762 /// isPowerOfTwo - Return true if the given value is known to have exactly one
763 /// bit set when defined. For vectors return true if every element is known to
764 /// be a power of two when defined. Supports values with integer or pointer
765 /// types and vectors of integers.
766 bool llvm::isPowerOfTwo(Value *V, const TargetData *TD, bool OrZero,
768 if (Constant *C = dyn_cast<Constant>(V)) {
769 if (C->isNullValue())
771 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
772 return CI->getValue().isPowerOf2();
773 // TODO: Handle vector constants.
776 // 1 << X is clearly a power of two if the one is not shifted off the end. If
777 // it is shifted off the end then the result is undefined.
778 if (match(V, m_Shl(m_One(), m_Value())))
781 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
782 // bottom. If it is shifted off the bottom then the result is undefined.
783 if (match(V, m_LShr(m_SignBit(), m_Value())))
786 // The remaining tests are all recursive, so bail out if we hit the limit.
787 if (Depth++ == MaxDepth)
790 Value *X = 0, *Y = 0;
791 // A shift of a power of two is a power of two or zero.
792 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
793 match(V, m_Shr(m_Value(X), m_Value()))))
794 return isPowerOfTwo(X, TD, /*OrZero*/true, Depth);
796 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
797 return isPowerOfTwo(ZI->getOperand(0), TD, OrZero, Depth);
799 if (SelectInst *SI = dyn_cast<SelectInst>(V))
800 return isPowerOfTwo(SI->getTrueValue(), TD, OrZero, Depth) &&
801 isPowerOfTwo(SI->getFalseValue(), TD, OrZero, Depth);
803 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
804 // A power of two and'd with anything is a power of two or zero.
805 if (isPowerOfTwo(X, TD, /*OrZero*/true, Depth) ||
806 isPowerOfTwo(Y, TD, /*OrZero*/true, Depth))
808 // X & (-X) is always a power of two or zero.
809 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
814 // An exact divide or right shift can only shift off zero bits, so the result
815 // is a power of two only if the first operand is a power of two and not
816 // copying a sign bit (sdiv int_min, 2).
817 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
818 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
819 return isPowerOfTwo(cast<Operator>(V)->getOperand(0), TD, OrZero, Depth);
825 /// isKnownNonZero - Return true if the given value is known to be non-zero
826 /// when defined. For vectors return true if every element is known to be
827 /// non-zero when defined. Supports values with integer or pointer type and
828 /// vectors of integers.
829 bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
830 if (Constant *C = dyn_cast<Constant>(V)) {
831 if (C->isNullValue())
833 if (isa<ConstantInt>(C))
834 // Must be non-zero due to null test above.
836 // TODO: Handle vectors
840 // The remaining tests are all recursive, so bail out if we hit the limit.
841 if (Depth++ >= MaxDepth)
844 unsigned BitWidth = getBitWidth(V->getType(), TD);
846 // X | Y != 0 if X != 0 or Y != 0.
847 Value *X = 0, *Y = 0;
848 if (match(V, m_Or(m_Value(X), m_Value(Y))))
849 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
851 // ext X != 0 if X != 0.
852 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
853 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
855 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
856 // if the lowest bit is shifted off the end.
857 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
858 // shl nuw can't remove any non-zero bits.
859 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
860 if (BO->hasNoUnsignedWrap())
861 return isKnownNonZero(X, TD, Depth);
863 APInt KnownZero(BitWidth, 0);
864 APInt KnownOne(BitWidth, 0);
865 ComputeMaskedBits(X, APInt(BitWidth, 1), KnownZero, KnownOne, TD, Depth);
869 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
870 // defined if the sign bit is shifted off the end.
871 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
872 // shr exact can only shift out zero bits.
873 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
875 return isKnownNonZero(X, TD, Depth);
877 bool XKnownNonNegative, XKnownNegative;
878 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
882 // div exact can only produce a zero if the dividend is zero.
883 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
884 return isKnownNonZero(X, TD, Depth);
887 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
888 bool XKnownNonNegative, XKnownNegative;
889 bool YKnownNonNegative, YKnownNegative;
890 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
891 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
893 // If X and Y are both non-negative (as signed values) then their sum is not
894 // zero unless both X and Y are zero.
895 if (XKnownNonNegative && YKnownNonNegative)
896 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
899 // If X and Y are both negative (as signed values) then their sum is not
900 // zero unless both X and Y equal INT_MIN.
901 if (BitWidth && XKnownNegative && YKnownNegative) {
902 APInt KnownZero(BitWidth, 0);
903 APInt KnownOne(BitWidth, 0);
904 APInt Mask = APInt::getSignedMaxValue(BitWidth);
905 // The sign bit of X is set. If some other bit is set then X is not equal
907 ComputeMaskedBits(X, Mask, KnownZero, KnownOne, TD, Depth);
908 if ((KnownOne & Mask) != 0)
910 // The sign bit of Y is set. If some other bit is set then Y is not equal
912 ComputeMaskedBits(Y, Mask, KnownZero, KnownOne, TD, Depth);
913 if ((KnownOne & Mask) != 0)
917 // The sum of a non-negative number and a power of two is not zero.
918 if (XKnownNonNegative && isPowerOfTwo(Y, TD, /*OrZero*/false, Depth))
920 if (YKnownNonNegative && isPowerOfTwo(X, TD, /*OrZero*/false, Depth))
924 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
925 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
926 // If X and Y are non-zero then so is X * Y as long as the multiplication
927 // does not overflow.
928 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
929 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
932 // (C ? X : Y) != 0 if X != 0 and Y != 0.
933 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
934 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
935 isKnownNonZero(SI->getFalseValue(), TD, Depth))
939 if (!BitWidth) return false;
940 APInt KnownZero(BitWidth, 0);
941 APInt KnownOne(BitWidth, 0);
942 ComputeMaskedBits(V, APInt::getAllOnesValue(BitWidth), KnownZero, KnownOne,
944 return KnownOne != 0;
947 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
948 /// this predicate to simplify operations downstream. Mask is known to be zero
949 /// for bits that V cannot have.
951 /// This function is defined on values with integer type, values with pointer
952 /// type (but only if TD is non-null), and vectors of integers. In the case
953 /// where V is a vector, the mask, known zero, and known one values are the
954 /// same width as the vector element, and the bit is set only if it is true
955 /// for all of the elements in the vector.
956 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
957 const TargetData *TD, unsigned Depth) {
958 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
959 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
960 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
961 return (KnownZero & Mask) == Mask;
966 /// ComputeNumSignBits - Return the number of times the sign bit of the
967 /// register is replicated into the other bits. We know that at least 1 bit
968 /// is always equal to the sign bit (itself), but other cases can give us
969 /// information. For example, immediately after an "ashr X, 2", we know that
970 /// the top 3 bits are all equal to each other, so we return 3.
972 /// 'Op' must have a scalar integer type.
974 unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
976 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
977 "ComputeNumSignBits requires a TargetData object to operate "
978 "on non-integer values!");
979 Type *Ty = V->getType();
980 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
981 Ty->getScalarSizeInBits();
983 unsigned FirstAnswer = 1;
985 // Note that ConstantInt is handled by the general ComputeMaskedBits case
989 return 1; // Limit search depth.
991 Operator *U = dyn_cast<Operator>(V);
992 switch (Operator::getOpcode(V)) {
994 case Instruction::SExt:
995 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
996 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
998 case Instruction::AShr: {
999 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1000 // ashr X, C -> adds C sign bits. Vectors too.
1002 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1003 Tmp += ShAmt->getZExtValue();
1004 if (Tmp > TyBits) Tmp = TyBits;
1008 case Instruction::Shl: {
1010 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1011 // shl destroys sign bits.
1012 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1013 Tmp2 = ShAmt->getZExtValue();
1014 if (Tmp2 >= TyBits || // Bad shift.
1015 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1020 case Instruction::And:
1021 case Instruction::Or:
1022 case Instruction::Xor: // NOT is handled here.
1023 // Logical binary ops preserve the number of sign bits at the worst.
1024 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1026 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1027 FirstAnswer = std::min(Tmp, Tmp2);
1028 // We computed what we know about the sign bits as our first
1029 // answer. Now proceed to the generic code that uses
1030 // ComputeMaskedBits, and pick whichever answer is better.
1034 case Instruction::Select:
1035 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1036 if (Tmp == 1) return 1; // Early out.
1037 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1038 return std::min(Tmp, Tmp2);
1040 case Instruction::Add:
1041 // Add can have at most one carry bit. Thus we know that the output
1042 // is, at worst, one more bit than the inputs.
1043 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1044 if (Tmp == 1) return 1; // Early out.
1046 // Special case decrementing a value (ADD X, -1):
1047 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1048 if (CRHS->isAllOnesValue()) {
1049 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1050 APInt Mask = APInt::getAllOnesValue(TyBits);
1051 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
1054 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1056 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1059 // If we are subtracting one from a positive number, there is no carry
1060 // out of the result.
1061 if (KnownZero.isNegative())
1065 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1066 if (Tmp2 == 1) return 1;
1067 return std::min(Tmp, Tmp2)-1;
1069 case Instruction::Sub:
1070 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1071 if (Tmp2 == 1) return 1;
1074 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1075 if (CLHS->isNullValue()) {
1076 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1077 APInt Mask = APInt::getAllOnesValue(TyBits);
1078 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
1080 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1082 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1085 // If the input is known to be positive (the sign bit is known clear),
1086 // the output of the NEG has the same number of sign bits as the input.
1087 if (KnownZero.isNegative())
1090 // Otherwise, we treat this like a SUB.
1093 // Sub can have at most one carry bit. Thus we know that the output
1094 // is, at worst, one more bit than the inputs.
1095 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1096 if (Tmp == 1) return 1; // Early out.
1097 return std::min(Tmp, Tmp2)-1;
1099 case Instruction::PHI: {
1100 PHINode *PN = cast<PHINode>(U);
1101 // Don't analyze large in-degree PHIs.
1102 if (PN->getNumIncomingValues() > 4) break;
1104 // Take the minimum of all incoming values. This can't infinitely loop
1105 // because of our depth threshold.
1106 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1107 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1108 if (Tmp == 1) return Tmp;
1110 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1115 case Instruction::Trunc:
1116 // FIXME: it's tricky to do anything useful for this, but it is an important
1117 // case for targets like X86.
1121 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1122 // use this information.
1123 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1124 APInt Mask = APInt::getAllOnesValue(TyBits);
1125 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
1127 if (KnownZero.isNegative()) { // sign bit is 0
1129 } else if (KnownOne.isNegative()) { // sign bit is 1;
1136 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1137 // the number of identical bits in the top of the input value.
1139 Mask <<= Mask.getBitWidth()-TyBits;
1140 // Return # leading zeros. We use 'min' here in case Val was zero before
1141 // shifting. We don't want to return '64' as for an i32 "0".
1142 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1145 /// ComputeMultiple - This function computes the integer multiple of Base that
1146 /// equals V. If successful, it returns true and returns the multiple in
1147 /// Multiple. If unsuccessful, it returns false. It looks
1148 /// through SExt instructions only if LookThroughSExt is true.
1149 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1150 bool LookThroughSExt, unsigned Depth) {
1151 const unsigned MaxDepth = 6;
1153 assert(V && "No Value?");
1154 assert(Depth <= MaxDepth && "Limit Search Depth");
1155 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1157 Type *T = V->getType();
1159 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1169 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1170 Constant *BaseVal = ConstantInt::get(T, Base);
1171 if (CO && CO == BaseVal) {
1173 Multiple = ConstantInt::get(T, 1);
1177 if (CI && CI->getZExtValue() % Base == 0) {
1178 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1182 if (Depth == MaxDepth) return false; // Limit search depth.
1184 Operator *I = dyn_cast<Operator>(V);
1185 if (!I) return false;
1187 switch (I->getOpcode()) {
1189 case Instruction::SExt:
1190 if (!LookThroughSExt) return false;
1191 // otherwise fall through to ZExt
1192 case Instruction::ZExt:
1193 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1194 LookThroughSExt, Depth+1);
1195 case Instruction::Shl:
1196 case Instruction::Mul: {
1197 Value *Op0 = I->getOperand(0);
1198 Value *Op1 = I->getOperand(1);
1200 if (I->getOpcode() == Instruction::Shl) {
1201 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1202 if (!Op1CI) return false;
1203 // Turn Op0 << Op1 into Op0 * 2^Op1
1204 APInt Op1Int = Op1CI->getValue();
1205 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1206 APInt API(Op1Int.getBitWidth(), 0);
1207 API.setBit(BitToSet);
1208 Op1 = ConstantInt::get(V->getContext(), API);
1212 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1213 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1214 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1215 if (Op1C->getType()->getPrimitiveSizeInBits() <
1216 MulC->getType()->getPrimitiveSizeInBits())
1217 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1218 if (Op1C->getType()->getPrimitiveSizeInBits() >
1219 MulC->getType()->getPrimitiveSizeInBits())
1220 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1222 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1223 Multiple = ConstantExpr::getMul(MulC, Op1C);
1227 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1228 if (Mul0CI->getValue() == 1) {
1229 // V == Base * Op1, so return Op1
1236 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1237 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1238 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1239 if (Op0C->getType()->getPrimitiveSizeInBits() <
1240 MulC->getType()->getPrimitiveSizeInBits())
1241 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1242 if (Op0C->getType()->getPrimitiveSizeInBits() >
1243 MulC->getType()->getPrimitiveSizeInBits())
1244 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1246 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1247 Multiple = ConstantExpr::getMul(MulC, Op0C);
1251 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1252 if (Mul1CI->getValue() == 1) {
1253 // V == Base * Op0, so return Op0
1261 // We could not determine if V is a multiple of Base.
1265 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1266 /// value is never equal to -0.0.
1268 /// NOTE: this function will need to be revisited when we support non-default
1271 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1272 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1273 return !CFP->getValueAPF().isNegZero();
1276 return 1; // Limit search depth.
1278 const Operator *I = dyn_cast<Operator>(V);
1279 if (I == 0) return false;
1281 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1282 if (I->getOpcode() == Instruction::FAdd &&
1283 isa<ConstantFP>(I->getOperand(1)) &&
1284 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1287 // sitofp and uitofp turn into +0.0 for zero.
1288 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1291 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1292 // sqrt(-0.0) = -0.0, no other negative results are possible.
1293 if (II->getIntrinsicID() == Intrinsic::sqrt)
1294 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1296 if (const CallInst *CI = dyn_cast<CallInst>(I))
1297 if (const Function *F = CI->getCalledFunction()) {
1298 if (F->isDeclaration()) {
1300 if (F->getName() == "abs") return true;
1301 // fabs[lf](x) != -0.0
1302 if (F->getName() == "fabs") return true;
1303 if (F->getName() == "fabsf") return true;
1304 if (F->getName() == "fabsl") return true;
1305 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1306 F->getName() == "sqrtl")
1307 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1314 /// isBytewiseValue - If the specified value can be set by repeating the same
1315 /// byte in memory, return the i8 value that it is represented with. This is
1316 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1317 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1318 /// byte store (e.g. i16 0x1234), return null.
1319 Value *llvm::isBytewiseValue(Value *V) {
1320 // All byte-wide stores are splatable, even of arbitrary variables.
1321 if (V->getType()->isIntegerTy(8)) return V;
1323 // Handle 'null' ConstantArrayZero etc.
1324 if (Constant *C = dyn_cast<Constant>(V))
1325 if (C->isNullValue())
1326 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1328 // Constant float and double values can be handled as integer values if the
1329 // corresponding integer value is "byteable". An important case is 0.0.
1330 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1331 if (CFP->getType()->isFloatTy())
1332 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1333 if (CFP->getType()->isDoubleTy())
1334 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1335 // Don't handle long double formats, which have strange constraints.
1338 // We can handle constant integers that are power of two in size and a
1339 // multiple of 8 bits.
1340 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1341 unsigned Width = CI->getBitWidth();
1342 if (isPowerOf2_32(Width) && Width > 8) {
1343 // We can handle this value if the recursive binary decomposition is the
1344 // same at all levels.
1345 APInt Val = CI->getValue();
1347 while (Val.getBitWidth() != 8) {
1348 unsigned NextWidth = Val.getBitWidth()/2;
1349 Val2 = Val.lshr(NextWidth);
1350 Val2 = Val2.trunc(Val.getBitWidth()/2);
1351 Val = Val.trunc(Val.getBitWidth()/2);
1353 // If the top/bottom halves aren't the same, reject it.
1357 return ConstantInt::get(V->getContext(), Val);
1361 // A ConstantDataArray/Vector is splatable if all its members are equal and
1363 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1364 Value *Elt = CA->getElementAsConstant(0);
1365 Value *Val = isBytewiseValue(Elt);
1369 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1370 if (CA->getElementAsConstant(I) != Elt)
1376 // Conceptually, we could handle things like:
1377 // %a = zext i8 %X to i16
1378 // %b = shl i16 %a, 8
1379 // %c = or i16 %a, %b
1380 // but until there is an example that actually needs this, it doesn't seem
1381 // worth worrying about.
1386 // This is the recursive version of BuildSubAggregate. It takes a few different
1387 // arguments. Idxs is the index within the nested struct From that we are
1388 // looking at now (which is of type IndexedType). IdxSkip is the number of
1389 // indices from Idxs that should be left out when inserting into the resulting
1390 // struct. To is the result struct built so far, new insertvalue instructions
1392 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1393 SmallVector<unsigned, 10> &Idxs,
1395 Instruction *InsertBefore) {
1396 llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1398 // Save the original To argument so we can modify it
1400 // General case, the type indexed by Idxs is a struct
1401 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1402 // Process each struct element recursively
1405 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1409 // Couldn't find any inserted value for this index? Cleanup
1410 while (PrevTo != OrigTo) {
1411 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1412 PrevTo = Del->getAggregateOperand();
1413 Del->eraseFromParent();
1415 // Stop processing elements
1419 // If we successfully found a value for each of our subaggregates
1423 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1424 // the struct's elements had a value that was inserted directly. In the latter
1425 // case, perhaps we can't determine each of the subelements individually, but
1426 // we might be able to find the complete struct somewhere.
1428 // Find the value that is at that particular spot
1429 Value *V = FindInsertedValue(From, Idxs);
1434 // Insert the value in the new (sub) aggregrate
1435 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1436 "tmp", InsertBefore);
1439 // This helper takes a nested struct and extracts a part of it (which is again a
1440 // struct) into a new value. For example, given the struct:
1441 // { a, { b, { c, d }, e } }
1442 // and the indices "1, 1" this returns
1445 // It does this by inserting an insertvalue for each element in the resulting
1446 // struct, as opposed to just inserting a single struct. This will only work if
1447 // each of the elements of the substruct are known (ie, inserted into From by an
1448 // insertvalue instruction somewhere).
1450 // All inserted insertvalue instructions are inserted before InsertBefore
1451 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1452 Instruction *InsertBefore) {
1453 assert(InsertBefore && "Must have someplace to insert!");
1454 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1456 Value *To = UndefValue::get(IndexedType);
1457 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1458 unsigned IdxSkip = Idxs.size();
1460 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1463 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1464 /// the scalar value indexed is already around as a register, for example if it
1465 /// were inserted directly into the aggregrate.
1467 /// If InsertBefore is not null, this function will duplicate (modified)
1468 /// insertvalues when a part of a nested struct is extracted.
1469 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1470 Instruction *InsertBefore) {
1471 // Nothing to index? Just return V then (this is useful at the end of our
1473 if (idx_range.empty())
1475 // We have indices, so V should have an indexable type.
1476 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1477 "Not looking at a struct or array?");
1478 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1479 "Invalid indices for type?");
1481 if (Constant *C = dyn_cast<Constant>(V)) {
1482 C = C->getAggregateElement(idx_range[0]);
1483 if (C == 0) return 0;
1484 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1487 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1488 // Loop the indices for the insertvalue instruction in parallel with the
1489 // requested indices
1490 const unsigned *req_idx = idx_range.begin();
1491 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1492 i != e; ++i, ++req_idx) {
1493 if (req_idx == idx_range.end()) {
1494 // We can't handle this without inserting insertvalues
1498 // The requested index identifies a part of a nested aggregate. Handle
1499 // this specially. For example,
1500 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1501 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1502 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1503 // This can be changed into
1504 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1505 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1506 // which allows the unused 0,0 element from the nested struct to be
1508 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1512 // This insert value inserts something else than what we are looking for.
1513 // See if the (aggregrate) value inserted into has the value we are
1514 // looking for, then.
1516 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1519 // If we end up here, the indices of the insertvalue match with those
1520 // requested (though possibly only partially). Now we recursively look at
1521 // the inserted value, passing any remaining indices.
1522 return FindInsertedValue(I->getInsertedValueOperand(),
1523 makeArrayRef(req_idx, idx_range.end()),
1527 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1528 // If we're extracting a value from an aggregrate that was extracted from
1529 // something else, we can extract from that something else directly instead.
1530 // However, we will need to chain I's indices with the requested indices.
1532 // Calculate the number of indices required
1533 unsigned size = I->getNumIndices() + idx_range.size();
1534 // Allocate some space to put the new indices in
1535 SmallVector<unsigned, 5> Idxs;
1537 // Add indices from the extract value instruction
1538 Idxs.append(I->idx_begin(), I->idx_end());
1540 // Add requested indices
1541 Idxs.append(idx_range.begin(), idx_range.end());
1543 assert(Idxs.size() == size
1544 && "Number of indices added not correct?");
1546 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1548 // Otherwise, we don't know (such as, extracting from a function return value
1549 // or load instruction)
1553 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1554 /// it can be expressed as a base pointer plus a constant offset. Return the
1555 /// base and offset to the caller.
1556 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1557 const TargetData &TD) {
1558 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1559 if (PtrOp == 0 || Ptr->getType()->isVectorTy())
1562 // Just look through bitcasts.
1563 if (PtrOp->getOpcode() == Instruction::BitCast)
1564 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1566 // If this is a GEP with constant indices, we can look through it.
1567 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1568 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1570 gep_type_iterator GTI = gep_type_begin(GEP);
1571 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1573 ConstantInt *OpC = cast<ConstantInt>(*I);
1574 if (OpC->isZero()) continue;
1576 // Handle a struct and array indices which add their offset to the pointer.
1577 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1578 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1580 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1581 Offset += OpC->getSExtValue()*Size;
1585 // Re-sign extend from the pointer size if needed to get overflow edge cases
1587 unsigned PtrSize = TD.getPointerSizeInBits();
1589 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1591 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1595 /// getConstantStringInfo - This function computes the length of a
1596 /// null-terminated C string pointed to by V. If successful, it returns true
1597 /// and returns the string in Str. If unsuccessful, it returns false.
1598 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
1599 uint64_t Offset, bool TrimAtNul) {
1602 // Look through bitcast instructions and geps.
1603 V = V->stripPointerCasts();
1605 // If the value is a GEP instructionor constant expression, treat it as an
1607 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1608 // Make sure the GEP has exactly three arguments.
1609 if (GEP->getNumOperands() != 3)
1612 // Make sure the index-ee is a pointer to array of i8.
1613 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1614 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1615 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1618 // Check to make sure that the first operand of the GEP is an integer and
1619 // has value 0 so that we are sure we're indexing into the initializer.
1620 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1621 if (FirstIdx == 0 || !FirstIdx->isZero())
1624 // If the second index isn't a ConstantInt, then this is a variable index
1625 // into the array. If this occurs, we can't say anything meaningful about
1627 uint64_t StartIdx = 0;
1628 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1629 StartIdx = CI->getZExtValue();
1632 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
1635 // The GEP instruction, constant or instruction, must reference a global
1636 // variable that is a constant and is initialized. The referenced constant
1637 // initializer is the array that we'll use for optimization.
1638 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
1639 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1642 // Handle the all-zeros case
1643 if (GV->getInitializer()->isNullValue()) {
1644 // This is a degenerate case. The initializer is constant zero so the
1645 // length of the string must be zero.
1650 // Must be a Constant Array
1651 const ConstantDataArray *Array =
1652 dyn_cast<ConstantDataArray>(GV->getInitializer());
1653 if (Array == 0 || !Array->isString())
1656 // Get the number of elements in the array
1657 uint64_t NumElts = Array->getType()->getArrayNumElements();
1659 // Start out with the entire array in the StringRef.
1660 Str = Array->getAsString();
1662 if (Offset > NumElts)
1665 // Skip over 'offset' bytes.
1666 Str = Str.substr(Offset);
1669 // Trim off the \0 and anything after it. If the array is not nul
1670 // terminated, we just return the whole end of string. The client may know
1671 // some other way that the string is length-bound.
1672 Str = Str.substr(0, Str.find('\0'));
1677 // These next two are very similar to the above, but also look through PHI
1679 // TODO: See if we can integrate these two together.
1681 /// GetStringLengthH - If we can compute the length of the string pointed to by
1682 /// the specified pointer, return 'len+1'. If we can't, return 0.
1683 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1684 // Look through noop bitcast instructions.
1685 V = V->stripPointerCasts();
1687 // If this is a PHI node, there are two cases: either we have already seen it
1689 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1690 if (!PHIs.insert(PN))
1691 return ~0ULL; // already in the set.
1693 // If it was new, see if all the input strings are the same length.
1694 uint64_t LenSoFar = ~0ULL;
1695 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1696 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1697 if (Len == 0) return 0; // Unknown length -> unknown.
1699 if (Len == ~0ULL) continue;
1701 if (Len != LenSoFar && LenSoFar != ~0ULL)
1702 return 0; // Disagree -> unknown.
1706 // Success, all agree.
1710 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1711 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1712 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1713 if (Len1 == 0) return 0;
1714 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1715 if (Len2 == 0) return 0;
1716 if (Len1 == ~0ULL) return Len2;
1717 if (Len2 == ~0ULL) return Len1;
1718 if (Len1 != Len2) return 0;
1722 // Otherwise, see if we can read the string.
1724 if (!getConstantStringInfo(V, StrData))
1727 return StrData.size()+1;
1730 /// GetStringLength - If we can compute the length of the string pointed to by
1731 /// the specified pointer, return 'len+1'. If we can't, return 0.
1732 uint64_t llvm::GetStringLength(Value *V) {
1733 if (!V->getType()->isPointerTy()) return 0;
1735 SmallPtrSet<PHINode*, 32> PHIs;
1736 uint64_t Len = GetStringLengthH(V, PHIs);
1737 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1738 // an empty string as a length.
1739 return Len == ~0ULL ? 1 : Len;
1743 llvm::GetUnderlyingObject(Value *V, const TargetData *TD, unsigned MaxLookup) {
1744 if (!V->getType()->isPointerTy())
1746 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1747 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1748 V = GEP->getPointerOperand();
1749 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1750 V = cast<Operator>(V)->getOperand(0);
1751 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1752 if (GA->mayBeOverridden())
1754 V = GA->getAliasee();
1756 // See if InstructionSimplify knows any relevant tricks.
1757 if (Instruction *I = dyn_cast<Instruction>(V))
1758 // TODO: Acquire a DominatorTree and use it.
1759 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1766 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1771 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1772 /// are lifetime markers.
1774 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1775 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1777 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1778 if (!II) return false;
1780 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1781 II->getIntrinsicID() != Intrinsic::lifetime_end)
1787 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1788 const TargetData *TD) {
1789 const Operator *Inst = dyn_cast<Operator>(V);
1793 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1794 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1798 switch (Inst->getOpcode()) {
1801 case Instruction::UDiv:
1802 case Instruction::URem:
1803 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1804 return isKnownNonZero(Inst->getOperand(1), TD);
1805 case Instruction::SDiv:
1806 case Instruction::SRem: {
1807 Value *Op = Inst->getOperand(1);
1808 // x / y is undefined if y == 0
1809 if (!isKnownNonZero(Op, TD))
1811 // x / y might be undefined if y == -1
1812 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1815 APInt KnownZero(BitWidth, 0);
1816 APInt KnownOne(BitWidth, 0);
1817 ComputeMaskedBits(Op, APInt::getAllOnesValue(BitWidth),
1818 KnownZero, KnownOne, TD);
1821 case Instruction::Load: {
1822 const LoadInst *LI = cast<LoadInst>(Inst);
1823 if (!LI->isUnordered())
1825 return LI->getPointerOperand()->isDereferenceablePointer();
1827 case Instruction::Call: {
1828 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1829 switch (II->getIntrinsicID()) {
1830 case Intrinsic::bswap:
1831 case Intrinsic::ctlz:
1832 case Intrinsic::ctpop:
1833 case Intrinsic::cttz:
1834 case Intrinsic::objectsize:
1835 case Intrinsic::sadd_with_overflow:
1836 case Intrinsic::smul_with_overflow:
1837 case Intrinsic::ssub_with_overflow:
1838 case Intrinsic::uadd_with_overflow:
1839 case Intrinsic::umul_with_overflow:
1840 case Intrinsic::usub_with_overflow:
1842 // TODO: some fp intrinsics are marked as having the same error handling
1843 // as libm. They're safe to speculate when they won't error.
1844 // TODO: are convert_{from,to}_fp16 safe?
1845 // TODO: can we list target-specific intrinsics here?
1849 return false; // The called function could have undefined behavior or
1850 // side-effects, even if marked readnone nounwind.
1852 case Instruction::VAArg:
1853 case Instruction::Alloca:
1854 case Instruction::Invoke:
1855 case Instruction::PHI:
1856 case Instruction::Store:
1857 case Instruction::Ret:
1858 case Instruction::Br:
1859 case Instruction::IndirectBr:
1860 case Instruction::Switch:
1861 case Instruction::Unreachable:
1862 case Instruction::Fence:
1863 case Instruction::LandingPad:
1864 case Instruction::AtomicRMW:
1865 case Instruction::AtomicCmpXchg:
1866 case Instruction::Resume:
1867 return false; // Misc instructions which have effects