1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Constants.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/GlobalVariable.h"
20 #include "llvm/GlobalAlias.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/Operator.h"
24 #include "llvm/Target/TargetData.h"
25 #include "llvm/Support/GetElementPtrTypeIterator.h"
26 #include "llvm/Support/MathExtras.h"
27 #include "llvm/Support/PatternMatch.h"
28 #include "llvm/ADT/SmallPtrSet.h"
31 using namespace llvm::PatternMatch;
33 const unsigned MaxDepth = 6;
35 /// getBitWidth - Returns the bitwidth of the given scalar or pointer type (if
36 /// unknown returns 0). For vector types, returns the element type's bitwidth.
37 static unsigned getBitWidth(Type *Ty, const TargetData *TD) {
38 if (unsigned BitWidth = Ty->getScalarSizeInBits())
40 assert(isa<PointerType>(Ty) && "Expected a pointer type!");
41 return TD ? TD->getPointerSizeInBits() : 0;
44 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
45 /// known to be either zero or one and return them in the KnownZero/KnownOne
46 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
48 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
49 /// we cannot optimize based on the assumption that it is zero without changing
50 /// it to be an explicit zero. If we don't change it to zero, other code could
51 /// optimized based on the contradictory assumption that it is non-zero.
52 /// Because instcombine aggressively folds operations with undef args anyway,
53 /// this won't lose us code quality.
55 /// This function is defined on values with integer type, values with pointer
56 /// type (but only if TD is non-null), and vectors of integers. In the case
57 /// where V is a vector, the mask, known zero, and known one values are the
58 /// same width as the vector element, and the bit is set only if it is true
59 /// for all of the elements in the vector.
60 void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
61 APInt &KnownZero, APInt &KnownOne,
62 const TargetData *TD, unsigned Depth) {
63 assert(V && "No Value?");
64 assert(Depth <= MaxDepth && "Limit Search Depth");
65 unsigned BitWidth = Mask.getBitWidth();
66 assert((V->getType()->isIntOrIntVectorTy() ||
67 V->getType()->getScalarType()->isPointerTy()) &&
68 "Not integer or pointer type!");
70 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
71 (!V->getType()->isIntOrIntVectorTy() ||
72 V->getType()->getScalarSizeInBits() == BitWidth) &&
73 KnownZero.getBitWidth() == BitWidth &&
74 KnownOne.getBitWidth() == BitWidth &&
75 "V, Mask, KnownOne and KnownZero should have same BitWidth");
77 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
78 // We know all of the bits for a constant!
79 KnownOne = CI->getValue() & Mask;
80 KnownZero = ~KnownOne & Mask;
83 // Null and aggregate-zero are all-zeros.
84 if (isa<ConstantPointerNull>(V) ||
85 isa<ConstantAggregateZero>(V)) {
86 KnownOne.clearAllBits();
90 // Handle a constant vector by taking the intersection of the known bits of
92 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
93 KnownZero.setAllBits(); KnownOne.setAllBits();
94 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
95 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
96 ComputeMaskedBits(CV->getOperand(i), Mask, KnownZero2, KnownOne2,
98 KnownZero &= KnownZero2;
99 KnownOne &= KnownOne2;
103 // The address of an aligned GlobalValue has trailing zeros.
104 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
105 unsigned Align = GV->getAlignment();
106 if (Align == 0 && TD && GV->getType()->getElementType()->isSized()) {
107 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
108 Type *ObjectType = GVar->getType()->getElementType();
109 // If the object is defined in the current Module, we'll be giving
110 // it the preferred alignment. Otherwise, we have to assume that it
111 // may only have the minimum ABI alignment.
112 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
113 Align = TD->getPreferredAlignment(GVar);
115 Align = TD->getABITypeAlignment(ObjectType);
119 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
120 CountTrailingZeros_32(Align));
122 KnownZero.clearAllBits();
123 KnownOne.clearAllBits();
126 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
127 // the bits of its aliasee.
128 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
129 if (GA->mayBeOverridden()) {
130 KnownZero.clearAllBits(); KnownOne.clearAllBits();
132 ComputeMaskedBits(GA->getAliasee(), Mask, KnownZero, KnownOne,
138 if (Argument *A = dyn_cast<Argument>(V)) {
139 // Get alignment information off byval arguments if specified in the IR.
140 if (A->hasByValAttr())
141 if (unsigned Align = A->getParamAlignment())
142 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
143 CountTrailingZeros_32(Align));
147 // Start out not knowing anything.
148 KnownZero.clearAllBits(); KnownOne.clearAllBits();
150 if (Depth == MaxDepth || Mask == 0)
151 return; // Limit search depth.
153 Operator *I = dyn_cast<Operator>(V);
156 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
157 switch (I->getOpcode()) {
159 case Instruction::And: {
160 // If either the LHS or the RHS are Zero, the result is zero.
161 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
162 APInt Mask2(Mask & ~KnownZero);
163 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
165 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
166 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
168 // Output known-1 bits are only known if set in both the LHS & RHS.
169 KnownOne &= KnownOne2;
170 // Output known-0 are known to be clear if zero in either the LHS | RHS.
171 KnownZero |= KnownZero2;
174 case Instruction::Or: {
175 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
176 APInt Mask2(Mask & ~KnownOne);
177 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
179 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
180 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
182 // Output known-0 bits are only known if clear in both the LHS & RHS.
183 KnownZero &= KnownZero2;
184 // Output known-1 are known to be set if set in either the LHS | RHS.
185 KnownOne |= KnownOne2;
188 case Instruction::Xor: {
189 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
190 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
192 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
193 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
195 // Output known-0 bits are known if clear or set in both the LHS & RHS.
196 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
197 // Output known-1 are known to be set if set in only one of the LHS, RHS.
198 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
199 KnownZero = KnownZeroOut;
202 case Instruction::Mul: {
203 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
204 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
205 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
207 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
208 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
210 bool isKnownNegative = false;
211 bool isKnownNonNegative = false;
212 // If the multiplication is known not to overflow, compute the sign bit.
213 if (Mask.isNegative() &&
214 cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap()) {
215 Value *Op1 = I->getOperand(1), *Op2 = I->getOperand(0);
217 // The product of a number with itself is non-negative.
218 isKnownNonNegative = true;
220 bool isKnownNonNegative1 = KnownZero.isNegative();
221 bool isKnownNonNegative2 = KnownZero2.isNegative();
222 bool isKnownNegative1 = KnownOne.isNegative();
223 bool isKnownNegative2 = KnownOne2.isNegative();
224 // The product of two numbers with the same sign is non-negative.
225 isKnownNonNegative = (isKnownNegative1 && isKnownNegative2) ||
226 (isKnownNonNegative1 && isKnownNonNegative2);
227 // The product of a negative number and a non-negative number is either
229 if (!isKnownNonNegative)
230 isKnownNegative = (isKnownNegative1 && isKnownNonNegative2 &&
231 isKnownNonZero(Op2, TD, Depth)) ||
232 (isKnownNegative2 && isKnownNonNegative1 &&
233 isKnownNonZero(Op1, TD, Depth));
237 // If low bits are zero in either operand, output low known-0 bits.
238 // Also compute a conserative estimate for high known-0 bits.
239 // More trickiness is possible, but this is sufficient for the
240 // interesting case of alignment computation.
241 KnownOne.clearAllBits();
242 unsigned TrailZ = KnownZero.countTrailingOnes() +
243 KnownZero2.countTrailingOnes();
244 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
245 KnownZero2.countLeadingOnes(),
246 BitWidth) - BitWidth;
248 TrailZ = std::min(TrailZ, BitWidth);
249 LeadZ = std::min(LeadZ, BitWidth);
250 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
251 APInt::getHighBitsSet(BitWidth, LeadZ);
254 // Only make use of no-wrap flags if we failed to compute the sign bit
255 // directly. This matters if the multiplication always overflows, in
256 // which case we prefer to follow the result of the direct computation,
257 // though as the program is invoking undefined behaviour we can choose
258 // whatever we like here.
259 if (isKnownNonNegative && !KnownOne.isNegative())
260 KnownZero.setBit(BitWidth - 1);
261 else if (isKnownNegative && !KnownZero.isNegative())
262 KnownOne.setBit(BitWidth - 1);
266 case Instruction::UDiv: {
267 // For the purposes of computing leading zeros we can conservatively
268 // treat a udiv as a logical right shift by the power of 2 known to
269 // be less than the denominator.
270 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
271 ComputeMaskedBits(I->getOperand(0),
272 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
273 unsigned LeadZ = KnownZero2.countLeadingOnes();
275 KnownOne2.clearAllBits();
276 KnownZero2.clearAllBits();
277 ComputeMaskedBits(I->getOperand(1),
278 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
279 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
280 if (RHSUnknownLeadingOnes != BitWidth)
281 LeadZ = std::min(BitWidth,
282 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
284 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
287 case Instruction::Select:
288 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
289 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
291 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
292 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
294 // Only known if known in both the LHS and RHS.
295 KnownOne &= KnownOne2;
296 KnownZero &= KnownZero2;
298 case Instruction::FPTrunc:
299 case Instruction::FPExt:
300 case Instruction::FPToUI:
301 case Instruction::FPToSI:
302 case Instruction::SIToFP:
303 case Instruction::UIToFP:
304 return; // Can't work with floating point.
305 case Instruction::PtrToInt:
306 case Instruction::IntToPtr:
307 // We can't handle these if we don't know the pointer size.
309 // FALL THROUGH and handle them the same as zext/trunc.
310 case Instruction::ZExt:
311 case Instruction::Trunc: {
312 Type *SrcTy = I->getOperand(0)->getType();
314 unsigned SrcBitWidth;
315 // Note that we handle pointer operands here because of inttoptr/ptrtoint
316 // which fall through here.
317 if (SrcTy->isPointerTy())
318 SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
320 SrcBitWidth = SrcTy->getScalarSizeInBits();
322 APInt MaskIn = Mask.zextOrTrunc(SrcBitWidth);
323 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
324 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
325 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
327 KnownZero = KnownZero.zextOrTrunc(BitWidth);
328 KnownOne = KnownOne.zextOrTrunc(BitWidth);
329 // Any top bits are known to be zero.
330 if (BitWidth > SrcBitWidth)
331 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
334 case Instruction::BitCast: {
335 Type *SrcTy = I->getOperand(0)->getType();
336 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
337 // TODO: For now, not handling conversions like:
338 // (bitcast i64 %x to <2 x i32>)
339 !I->getType()->isVectorTy()) {
340 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
346 case Instruction::SExt: {
347 // Compute the bits in the result that are not present in the input.
348 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
350 APInt MaskIn = Mask.trunc(SrcBitWidth);
351 KnownZero = KnownZero.trunc(SrcBitWidth);
352 KnownOne = KnownOne.trunc(SrcBitWidth);
353 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
355 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
356 KnownZero = KnownZero.zext(BitWidth);
357 KnownOne = KnownOne.zext(BitWidth);
359 // If the sign bit of the input is known set or clear, then we know the
360 // top bits of the result.
361 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
362 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
363 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
364 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
367 case Instruction::Shl:
368 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
369 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
370 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
371 APInt Mask2(Mask.lshr(ShiftAmt));
372 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
374 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
375 KnownZero <<= ShiftAmt;
376 KnownOne <<= ShiftAmt;
377 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
381 case Instruction::LShr:
382 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
383 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
384 // Compute the new bits that are at the top now.
385 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
387 // Unsigned shift right.
388 APInt Mask2(Mask.shl(ShiftAmt));
389 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
391 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
392 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
393 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
394 // high bits known zero.
395 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
399 case Instruction::AShr:
400 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
401 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
402 // Compute the new bits that are at the top now.
403 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
405 // Signed shift right.
406 APInt Mask2(Mask.shl(ShiftAmt));
407 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
409 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
410 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
411 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
413 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
414 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
415 KnownZero |= HighBits;
416 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
417 KnownOne |= HighBits;
421 case Instruction::Sub: {
422 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
423 // We know that the top bits of C-X are clear if X contains less bits
424 // than C (i.e. no wrap-around can happen). For example, 20-X is
425 // positive if we can prove that X is >= 0 and < 16.
426 if (!CLHS->getValue().isNegative()) {
427 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
428 // NLZ can't be BitWidth with no sign bit
429 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
430 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
433 // If all of the MaskV bits are known to be zero, then we know the
434 // output top bits are zero, because we now know that the output is
436 if ((KnownZero2 & MaskV) == MaskV) {
437 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
438 // Top bits known zero.
439 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
445 case Instruction::Add: {
446 // If one of the operands has trailing zeros, then the bits that the
447 // other operand has in those bit positions will be preserved in the
448 // result. For an add, this works with either operand. For a subtract,
449 // this only works if the known zeros are in the right operand.
450 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
451 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
452 BitWidth - Mask.countLeadingZeros());
453 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
455 assert((LHSKnownZero & LHSKnownOne) == 0 &&
456 "Bits known to be one AND zero?");
457 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
459 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD,
461 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
462 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
464 // Determine which operand has more trailing zeros, and use that
465 // many bits from the other operand.
466 if (LHSKnownZeroOut > RHSKnownZeroOut) {
467 if (I->getOpcode() == Instruction::Add) {
468 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
469 KnownZero |= KnownZero2 & Mask;
470 KnownOne |= KnownOne2 & Mask;
472 // If the known zeros are in the left operand for a subtract,
473 // fall back to the minimum known zeros in both operands.
474 KnownZero |= APInt::getLowBitsSet(BitWidth,
475 std::min(LHSKnownZeroOut,
478 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
479 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
480 KnownZero |= LHSKnownZero & Mask;
481 KnownOne |= LHSKnownOne & Mask;
484 // Are we still trying to solve for the sign bit?
485 if (Mask.isNegative() && !KnownZero.isNegative() && !KnownOne.isNegative()){
486 OverflowingBinaryOperator *OBO = cast<OverflowingBinaryOperator>(I);
487 if (OBO->hasNoSignedWrap()) {
488 if (I->getOpcode() == Instruction::Add) {
489 // Adding two positive numbers can't wrap into negative
490 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
491 KnownZero |= APInt::getSignBit(BitWidth);
492 // and adding two negative numbers can't wrap into positive.
493 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
494 KnownOne |= APInt::getSignBit(BitWidth);
496 // Subtracting a negative number from a positive one can't wrap
497 if (LHSKnownZero.isNegative() && KnownOne2.isNegative())
498 KnownZero |= APInt::getSignBit(BitWidth);
499 // neither can subtracting a positive number from a negative one.
500 else if (LHSKnownOne.isNegative() && KnownZero2.isNegative())
501 KnownOne |= APInt::getSignBit(BitWidth);
508 case Instruction::SRem:
509 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
510 APInt RA = Rem->getValue().abs();
511 if (RA.isPowerOf2()) {
512 APInt LowBits = RA - 1;
513 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
514 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
517 // The low bits of the first operand are unchanged by the srem.
518 KnownZero = KnownZero2 & LowBits;
519 KnownOne = KnownOne2 & LowBits;
521 // If the first operand is non-negative or has all low bits zero, then
522 // the upper bits are all zero.
523 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
524 KnownZero |= ~LowBits;
526 // If the first operand is negative and not all low bits are zero, then
527 // the upper bits are all one.
528 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
529 KnownOne |= ~LowBits;
534 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
538 // The sign bit is the LHS's sign bit, except when the result of the
539 // remainder is zero.
540 if (Mask.isNegative() && KnownZero.isNonNegative()) {
541 APInt Mask2 = APInt::getSignBit(BitWidth);
542 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
543 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
545 // If it's known zero, our sign bit is also zero.
546 if (LHSKnownZero.isNegative())
547 KnownZero |= LHSKnownZero;
551 case Instruction::URem: {
552 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
553 APInt RA = Rem->getValue();
554 if (RA.isPowerOf2()) {
555 APInt LowBits = (RA - 1);
556 APInt Mask2 = LowBits & Mask;
557 KnownZero |= ~LowBits & Mask;
558 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
560 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
565 // Since the result is less than or equal to either operand, any leading
566 // zero bits in either operand must also exist in the result.
567 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
568 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
570 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
573 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
574 KnownZero2.countLeadingOnes());
575 KnownOne.clearAllBits();
576 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
580 case Instruction::Alloca: {
581 AllocaInst *AI = cast<AllocaInst>(V);
582 unsigned Align = AI->getAlignment();
583 if (Align == 0 && TD)
584 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
587 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
588 CountTrailingZeros_32(Align));
591 case Instruction::GetElementPtr: {
592 // Analyze all of the subscripts of this getelementptr instruction
593 // to determine if we can prove known low zero bits.
594 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
595 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
596 ComputeMaskedBits(I->getOperand(0), LocalMask,
597 LocalKnownZero, LocalKnownOne, TD, Depth+1);
598 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
600 gep_type_iterator GTI = gep_type_begin(I);
601 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
602 Value *Index = I->getOperand(i);
603 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
604 // Handle struct member offset arithmetic.
606 const StructLayout *SL = TD->getStructLayout(STy);
607 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
608 uint64_t Offset = SL->getElementOffset(Idx);
609 TrailZ = std::min(TrailZ,
610 CountTrailingZeros_64(Offset));
612 // Handle array index arithmetic.
613 Type *IndexedTy = GTI.getIndexedType();
614 if (!IndexedTy->isSized()) return;
615 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
616 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
617 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
618 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
619 ComputeMaskedBits(Index, LocalMask,
620 LocalKnownZero, LocalKnownOne, TD, Depth+1);
621 TrailZ = std::min(TrailZ,
622 unsigned(CountTrailingZeros_64(TypeSize) +
623 LocalKnownZero.countTrailingOnes()));
627 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
630 case Instruction::PHI: {
631 PHINode *P = cast<PHINode>(I);
632 // Handle the case of a simple two-predecessor recurrence PHI.
633 // There's a lot more that could theoretically be done here, but
634 // this is sufficient to catch some interesting cases.
635 if (P->getNumIncomingValues() == 2) {
636 for (unsigned i = 0; i != 2; ++i) {
637 Value *L = P->getIncomingValue(i);
638 Value *R = P->getIncomingValue(!i);
639 Operator *LU = dyn_cast<Operator>(L);
642 unsigned Opcode = LU->getOpcode();
643 // Check for operations that have the property that if
644 // both their operands have low zero bits, the result
645 // will have low zero bits.
646 if (Opcode == Instruction::Add ||
647 Opcode == Instruction::Sub ||
648 Opcode == Instruction::And ||
649 Opcode == Instruction::Or ||
650 Opcode == Instruction::Mul) {
651 Value *LL = LU->getOperand(0);
652 Value *LR = LU->getOperand(1);
653 // Find a recurrence.
660 // Ok, we have a PHI of the form L op= R. Check for low
662 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
663 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
664 Mask2 = APInt::getLowBitsSet(BitWidth,
665 KnownZero2.countTrailingOnes());
667 // We need to take the minimum number of known bits
668 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
669 ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
672 APInt::getLowBitsSet(BitWidth,
673 std::min(KnownZero2.countTrailingOnes(),
674 KnownZero3.countTrailingOnes()));
680 // Unreachable blocks may have zero-operand PHI nodes.
681 if (P->getNumIncomingValues() == 0)
684 // Otherwise take the unions of the known bit sets of the operands,
685 // taking conservative care to avoid excessive recursion.
686 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
687 // Skip if every incoming value references to ourself.
688 if (P->hasConstantValue() == P)
691 KnownZero = APInt::getAllOnesValue(BitWidth);
692 KnownOne = APInt::getAllOnesValue(BitWidth);
693 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
694 // Skip direct self references.
695 if (P->getIncomingValue(i) == P) continue;
697 KnownZero2 = APInt(BitWidth, 0);
698 KnownOne2 = APInt(BitWidth, 0);
699 // Recurse, but cap the recursion to one level, because we don't
700 // want to waste time spinning around in loops.
701 ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
702 KnownZero2, KnownOne2, TD, MaxDepth-1);
703 KnownZero &= KnownZero2;
704 KnownOne &= KnownOne2;
705 // If all bits have been ruled out, there's no need to check
707 if (!KnownZero && !KnownOne)
713 case Instruction::Call:
714 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
715 switch (II->getIntrinsicID()) {
717 case Intrinsic::ctlz:
718 case Intrinsic::cttz: {
719 unsigned LowBits = Log2_32(BitWidth)+1;
720 // If this call is undefined for 0, the result will be less than 2^n.
721 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
723 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
726 case Intrinsic::ctpop: {
727 unsigned LowBits = Log2_32(BitWidth)+1;
728 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
731 case Intrinsic::x86_sse42_crc32_64_8:
732 case Intrinsic::x86_sse42_crc32_64_64:
733 KnownZero = APInt::getHighBitsSet(64, 32);
741 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
742 /// one. Convenience wrapper around ComputeMaskedBits.
743 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
744 const TargetData *TD, unsigned Depth) {
745 unsigned BitWidth = getBitWidth(V->getType(), TD);
751 APInt ZeroBits(BitWidth, 0);
752 APInt OneBits(BitWidth, 0);
753 ComputeMaskedBits(V, APInt::getSignBit(BitWidth), ZeroBits, OneBits, TD,
755 KnownOne = OneBits[BitWidth - 1];
756 KnownZero = ZeroBits[BitWidth - 1];
759 /// isPowerOfTwo - Return true if the given value is known to have exactly one
760 /// bit set when defined. For vectors return true if every element is known to
761 /// be a power of two when defined. Supports values with integer or pointer
762 /// types and vectors of integers.
763 bool llvm::isPowerOfTwo(Value *V, const TargetData *TD, bool OrZero,
765 if (Constant *C = dyn_cast<Constant>(V)) {
766 if (C->isNullValue())
768 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
769 return CI->getValue().isPowerOf2();
770 // TODO: Handle vector constants.
773 // 1 << X is clearly a power of two if the one is not shifted off the end. If
774 // it is shifted off the end then the result is undefined.
775 if (match(V, m_Shl(m_One(), m_Value())))
778 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
779 // bottom. If it is shifted off the bottom then the result is undefined.
780 if (match(V, m_LShr(m_SignBit(), m_Value())))
783 // The remaining tests are all recursive, so bail out if we hit the limit.
784 if (Depth++ == MaxDepth)
787 Value *X = 0, *Y = 0;
788 // A shift of a power of two is a power of two or zero.
789 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
790 match(V, m_Shr(m_Value(X), m_Value()))))
791 return isPowerOfTwo(X, TD, /*OrZero*/true, Depth);
793 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
794 return isPowerOfTwo(ZI->getOperand(0), TD, OrZero, Depth);
796 if (SelectInst *SI = dyn_cast<SelectInst>(V))
797 return isPowerOfTwo(SI->getTrueValue(), TD, OrZero, Depth) &&
798 isPowerOfTwo(SI->getFalseValue(), TD, OrZero, Depth);
800 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
801 // A power of two and'd with anything is a power of two or zero.
802 if (isPowerOfTwo(X, TD, /*OrZero*/true, Depth) ||
803 isPowerOfTwo(Y, TD, /*OrZero*/true, Depth))
805 // X & (-X) is always a power of two or zero.
806 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
811 // An exact divide or right shift can only shift off zero bits, so the result
812 // is a power of two only if the first operand is a power of two and not
813 // copying a sign bit (sdiv int_min, 2).
814 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
815 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
816 return isPowerOfTwo(cast<Operator>(V)->getOperand(0), TD, OrZero, Depth);
822 /// isKnownNonZero - Return true if the given value is known to be non-zero
823 /// when defined. For vectors return true if every element is known to be
824 /// non-zero when defined. Supports values with integer or pointer type and
825 /// vectors of integers.
826 bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
827 if (Constant *C = dyn_cast<Constant>(V)) {
828 if (C->isNullValue())
830 if (isa<ConstantInt>(C))
831 // Must be non-zero due to null test above.
833 // TODO: Handle vectors
837 // The remaining tests are all recursive, so bail out if we hit the limit.
838 if (Depth++ >= MaxDepth)
841 unsigned BitWidth = getBitWidth(V->getType(), TD);
843 // X | Y != 0 if X != 0 or Y != 0.
844 Value *X = 0, *Y = 0;
845 if (match(V, m_Or(m_Value(X), m_Value(Y))))
846 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
848 // ext X != 0 if X != 0.
849 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
850 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
852 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
853 // if the lowest bit is shifted off the end.
854 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
855 // shl nuw can't remove any non-zero bits.
856 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
857 if (BO->hasNoUnsignedWrap())
858 return isKnownNonZero(X, TD, Depth);
860 APInt KnownZero(BitWidth, 0);
861 APInt KnownOne(BitWidth, 0);
862 ComputeMaskedBits(X, APInt(BitWidth, 1), KnownZero, KnownOne, TD, Depth);
866 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
867 // defined if the sign bit is shifted off the end.
868 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
869 // shr exact can only shift out zero bits.
870 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
872 return isKnownNonZero(X, TD, Depth);
874 bool XKnownNonNegative, XKnownNegative;
875 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
879 // div exact can only produce a zero if the dividend is zero.
880 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
881 return isKnownNonZero(X, TD, Depth);
884 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
885 bool XKnownNonNegative, XKnownNegative;
886 bool YKnownNonNegative, YKnownNegative;
887 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
888 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
890 // If X and Y are both non-negative (as signed values) then their sum is not
891 // zero unless both X and Y are zero.
892 if (XKnownNonNegative && YKnownNonNegative)
893 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
896 // If X and Y are both negative (as signed values) then their sum is not
897 // zero unless both X and Y equal INT_MIN.
898 if (BitWidth && XKnownNegative && YKnownNegative) {
899 APInt KnownZero(BitWidth, 0);
900 APInt KnownOne(BitWidth, 0);
901 APInt Mask = APInt::getSignedMaxValue(BitWidth);
902 // The sign bit of X is set. If some other bit is set then X is not equal
904 ComputeMaskedBits(X, Mask, KnownZero, KnownOne, TD, Depth);
905 if ((KnownOne & Mask) != 0)
907 // The sign bit of Y is set. If some other bit is set then Y is not equal
909 ComputeMaskedBits(Y, Mask, KnownZero, KnownOne, TD, Depth);
910 if ((KnownOne & Mask) != 0)
914 // The sum of a non-negative number and a power of two is not zero.
915 if (XKnownNonNegative && isPowerOfTwo(Y, TD, /*OrZero*/false, Depth))
917 if (YKnownNonNegative && isPowerOfTwo(X, TD, /*OrZero*/false, Depth))
921 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
922 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
923 // If X and Y are non-zero then so is X * Y as long as the multiplication
924 // does not overflow.
925 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
926 isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
929 // (C ? X : Y) != 0 if X != 0 and Y != 0.
930 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
931 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
932 isKnownNonZero(SI->getFalseValue(), TD, Depth))
936 if (!BitWidth) return false;
937 APInt KnownZero(BitWidth, 0);
938 APInt KnownOne(BitWidth, 0);
939 ComputeMaskedBits(V, APInt::getAllOnesValue(BitWidth), KnownZero, KnownOne,
941 return KnownOne != 0;
944 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
945 /// this predicate to simplify operations downstream. Mask is known to be zero
946 /// for bits that V cannot have.
948 /// This function is defined on values with integer type, values with pointer
949 /// type (but only if TD is non-null), and vectors of integers. In the case
950 /// where V is a vector, the mask, known zero, and known one values are the
951 /// same width as the vector element, and the bit is set only if it is true
952 /// for all of the elements in the vector.
953 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
954 const TargetData *TD, unsigned Depth) {
955 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
956 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
957 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
958 return (KnownZero & Mask) == Mask;
963 /// ComputeNumSignBits - Return the number of times the sign bit of the
964 /// register is replicated into the other bits. We know that at least 1 bit
965 /// is always equal to the sign bit (itself), but other cases can give us
966 /// information. For example, immediately after an "ashr X, 2", we know that
967 /// the top 3 bits are all equal to each other, so we return 3.
969 /// 'Op' must have a scalar integer type.
971 unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
973 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
974 "ComputeNumSignBits requires a TargetData object to operate "
975 "on non-integer values!");
976 Type *Ty = V->getType();
977 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
978 Ty->getScalarSizeInBits();
980 unsigned FirstAnswer = 1;
982 // Note that ConstantInt is handled by the general ComputeMaskedBits case
986 return 1; // Limit search depth.
988 Operator *U = dyn_cast<Operator>(V);
989 switch (Operator::getOpcode(V)) {
991 case Instruction::SExt:
992 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
993 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
995 case Instruction::AShr:
996 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
997 // ashr X, C -> adds C sign bits.
998 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
999 Tmp += C->getZExtValue();
1000 if (Tmp > TyBits) Tmp = TyBits;
1002 // vector ashr X, <C, C, C, C> -> adds C sign bits
1003 if (ConstantVector *C = dyn_cast<ConstantVector>(U->getOperand(1))) {
1004 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) {
1005 Tmp += CI->getZExtValue();
1006 if (Tmp > TyBits) Tmp = TyBits;
1010 case Instruction::Shl:
1011 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
1012 // shl destroys sign bits.
1013 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1014 if (C->getZExtValue() >= TyBits || // Bad shift.
1015 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
1016 return Tmp - C->getZExtValue();
1019 case Instruction::And:
1020 case Instruction::Or:
1021 case Instruction::Xor: // NOT is handled here.
1022 // Logical binary ops preserve the number of sign bits at the worst.
1023 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1025 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1026 FirstAnswer = std::min(Tmp, Tmp2);
1027 // We computed what we know about the sign bits as our first
1028 // answer. Now proceed to the generic code that uses
1029 // ComputeMaskedBits, and pick whichever answer is better.
1033 case Instruction::Select:
1034 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1035 if (Tmp == 1) return 1; // Early out.
1036 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
1037 return std::min(Tmp, Tmp2);
1039 case Instruction::Add:
1040 // Add can have at most one carry bit. Thus we know that the output
1041 // is, at worst, one more bit than the inputs.
1042 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1043 if (Tmp == 1) return 1; // Early out.
1045 // Special case decrementing a value (ADD X, -1):
1046 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1047 if (CRHS->isAllOnesValue()) {
1048 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1049 APInt Mask = APInt::getAllOnesValue(TyBits);
1050 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
1053 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1055 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1058 // If we are subtracting one from a positive number, there is no carry
1059 // out of the result.
1060 if (KnownZero.isNegative())
1064 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1065 if (Tmp2 == 1) return 1;
1066 return std::min(Tmp, Tmp2)-1;
1068 case Instruction::Sub:
1069 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
1070 if (Tmp2 == 1) return 1;
1073 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1074 if (CLHS->isNullValue()) {
1075 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1076 APInt Mask = APInt::getAllOnesValue(TyBits);
1077 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
1079 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1081 if ((KnownZero | APInt(TyBits, 1)) == Mask)
1084 // If the input is known to be positive (the sign bit is known clear),
1085 // the output of the NEG has the same number of sign bits as the input.
1086 if (KnownZero.isNegative())
1089 // Otherwise, we treat this like a SUB.
1092 // Sub can have at most one carry bit. Thus we know that the output
1093 // is, at worst, one more bit than the inputs.
1094 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
1095 if (Tmp == 1) return 1; // Early out.
1096 return std::min(Tmp, Tmp2)-1;
1098 case Instruction::PHI: {
1099 PHINode *PN = cast<PHINode>(U);
1100 // Don't analyze large in-degree PHIs.
1101 if (PN->getNumIncomingValues() > 4) break;
1103 // Take the minimum of all incoming values. This can't infinitely loop
1104 // because of our depth threshold.
1105 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
1106 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1107 if (Tmp == 1) return Tmp;
1109 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
1114 case Instruction::Trunc:
1115 // FIXME: it's tricky to do anything useful for this, but it is an important
1116 // case for targets like X86.
1120 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1121 // use this information.
1122 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1123 APInt Mask = APInt::getAllOnesValue(TyBits);
1124 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
1126 if (KnownZero.isNegative()) { // sign bit is 0
1128 } else if (KnownOne.isNegative()) { // sign bit is 1;
1135 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1136 // the number of identical bits in the top of the input value.
1138 Mask <<= Mask.getBitWidth()-TyBits;
1139 // Return # leading zeros. We use 'min' here in case Val was zero before
1140 // shifting. We don't want to return '64' as for an i32 "0".
1141 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1144 /// ComputeMultiple - This function computes the integer multiple of Base that
1145 /// equals V. If successful, it returns true and returns the multiple in
1146 /// Multiple. If unsuccessful, it returns false. It looks
1147 /// through SExt instructions only if LookThroughSExt is true.
1148 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1149 bool LookThroughSExt, unsigned Depth) {
1150 const unsigned MaxDepth = 6;
1152 assert(V && "No Value?");
1153 assert(Depth <= MaxDepth && "Limit Search Depth");
1154 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1156 Type *T = V->getType();
1158 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1168 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1169 Constant *BaseVal = ConstantInt::get(T, Base);
1170 if (CO && CO == BaseVal) {
1172 Multiple = ConstantInt::get(T, 1);
1176 if (CI && CI->getZExtValue() % Base == 0) {
1177 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1181 if (Depth == MaxDepth) return false; // Limit search depth.
1183 Operator *I = dyn_cast<Operator>(V);
1184 if (!I) return false;
1186 switch (I->getOpcode()) {
1188 case Instruction::SExt:
1189 if (!LookThroughSExt) return false;
1190 // otherwise fall through to ZExt
1191 case Instruction::ZExt:
1192 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1193 LookThroughSExt, Depth+1);
1194 case Instruction::Shl:
1195 case Instruction::Mul: {
1196 Value *Op0 = I->getOperand(0);
1197 Value *Op1 = I->getOperand(1);
1199 if (I->getOpcode() == Instruction::Shl) {
1200 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1201 if (!Op1CI) return false;
1202 // Turn Op0 << Op1 into Op0 * 2^Op1
1203 APInt Op1Int = Op1CI->getValue();
1204 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1205 APInt API(Op1Int.getBitWidth(), 0);
1206 API.setBit(BitToSet);
1207 Op1 = ConstantInt::get(V->getContext(), API);
1211 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1212 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1213 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1214 if (Op1C->getType()->getPrimitiveSizeInBits() <
1215 MulC->getType()->getPrimitiveSizeInBits())
1216 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1217 if (Op1C->getType()->getPrimitiveSizeInBits() >
1218 MulC->getType()->getPrimitiveSizeInBits())
1219 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1221 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1222 Multiple = ConstantExpr::getMul(MulC, Op1C);
1226 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1227 if (Mul0CI->getValue() == 1) {
1228 // V == Base * Op1, so return Op1
1235 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1236 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1237 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1238 if (Op0C->getType()->getPrimitiveSizeInBits() <
1239 MulC->getType()->getPrimitiveSizeInBits())
1240 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1241 if (Op0C->getType()->getPrimitiveSizeInBits() >
1242 MulC->getType()->getPrimitiveSizeInBits())
1243 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1245 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1246 Multiple = ConstantExpr::getMul(MulC, Op0C);
1250 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1251 if (Mul1CI->getValue() == 1) {
1252 // V == Base * Op0, so return Op0
1260 // We could not determine if V is a multiple of Base.
1264 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1265 /// value is never equal to -0.0.
1267 /// NOTE: this function will need to be revisited when we support non-default
1270 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1271 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1272 return !CFP->getValueAPF().isNegZero();
1275 return 1; // Limit search depth.
1277 const Operator *I = dyn_cast<Operator>(V);
1278 if (I == 0) return false;
1280 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1281 if (I->getOpcode() == Instruction::FAdd &&
1282 isa<ConstantFP>(I->getOperand(1)) &&
1283 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1286 // sitofp and uitofp turn into +0.0 for zero.
1287 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1290 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1291 // sqrt(-0.0) = -0.0, no other negative results are possible.
1292 if (II->getIntrinsicID() == Intrinsic::sqrt)
1293 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1295 if (const CallInst *CI = dyn_cast<CallInst>(I))
1296 if (const Function *F = CI->getCalledFunction()) {
1297 if (F->isDeclaration()) {
1299 if (F->getName() == "abs") return true;
1300 // fabs[lf](x) != -0.0
1301 if (F->getName() == "fabs") return true;
1302 if (F->getName() == "fabsf") return true;
1303 if (F->getName() == "fabsl") return true;
1304 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1305 F->getName() == "sqrtl")
1306 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1313 /// isBytewiseValue - If the specified value can be set by repeating the same
1314 /// byte in memory, return the i8 value that it is represented with. This is
1315 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1316 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1317 /// byte store (e.g. i16 0x1234), return null.
1318 Value *llvm::isBytewiseValue(Value *V) {
1319 // All byte-wide stores are splatable, even of arbitrary variables.
1320 if (V->getType()->isIntegerTy(8)) return V;
1322 // Handle 'null' ConstantArrayZero etc.
1323 if (Constant *C = dyn_cast<Constant>(V))
1324 if (C->isNullValue())
1325 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1327 // Constant float and double values can be handled as integer values if the
1328 // corresponding integer value is "byteable". An important case is 0.0.
1329 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1330 if (CFP->getType()->isFloatTy())
1331 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1332 if (CFP->getType()->isDoubleTy())
1333 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1334 // Don't handle long double formats, which have strange constraints.
1337 // We can handle constant integers that are power of two in size and a
1338 // multiple of 8 bits.
1339 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1340 unsigned Width = CI->getBitWidth();
1341 if (isPowerOf2_32(Width) && Width > 8) {
1342 // We can handle this value if the recursive binary decomposition is the
1343 // same at all levels.
1344 APInt Val = CI->getValue();
1346 while (Val.getBitWidth() != 8) {
1347 unsigned NextWidth = Val.getBitWidth()/2;
1348 Val2 = Val.lshr(NextWidth);
1349 Val2 = Val2.trunc(Val.getBitWidth()/2);
1350 Val = Val.trunc(Val.getBitWidth()/2);
1352 // If the top/bottom halves aren't the same, reject it.
1356 return ConstantInt::get(V->getContext(), Val);
1360 // A ConstantArray is splatable if all its members are equal and also
1362 if (ConstantArray *CA = dyn_cast<ConstantArray>(V)) {
1363 if (CA->getNumOperands() == 0)
1366 Value *Val = isBytewiseValue(CA->getOperand(0));
1370 for (unsigned I = 1, E = CA->getNumOperands(); I != E; ++I)
1371 if (CA->getOperand(I-1) != CA->getOperand(I))
1377 // FIXME: Vector types (e.g., <4 x i32> <i32 -1, i32 -1, i32 -1, i32 -1>).
1379 // Conceptually, we could handle things like:
1380 // %a = zext i8 %X to i16
1381 // %b = shl i16 %a, 8
1382 // %c = or i16 %a, %b
1383 // but until there is an example that actually needs this, it doesn't seem
1384 // worth worrying about.
1389 // This is the recursive version of BuildSubAggregate. It takes a few different
1390 // arguments. Idxs is the index within the nested struct From that we are
1391 // looking at now (which is of type IndexedType). IdxSkip is the number of
1392 // indices from Idxs that should be left out when inserting into the resulting
1393 // struct. To is the result struct built so far, new insertvalue instructions
1395 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1396 SmallVector<unsigned, 10> &Idxs,
1398 Instruction *InsertBefore) {
1399 llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1401 // Save the original To argument so we can modify it
1403 // General case, the type indexed by Idxs is a struct
1404 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1405 // Process each struct element recursively
1408 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1412 // Couldn't find any inserted value for this index? Cleanup
1413 while (PrevTo != OrigTo) {
1414 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1415 PrevTo = Del->getAggregateOperand();
1416 Del->eraseFromParent();
1418 // Stop processing elements
1422 // If we successfully found a value for each of our subaggregates
1426 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1427 // the struct's elements had a value that was inserted directly. In the latter
1428 // case, perhaps we can't determine each of the subelements individually, but
1429 // we might be able to find the complete struct somewhere.
1431 // Find the value that is at that particular spot
1432 Value *V = FindInsertedValue(From, Idxs);
1437 // Insert the value in the new (sub) aggregrate
1438 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1439 "tmp", InsertBefore);
1442 // This helper takes a nested struct and extracts a part of it (which is again a
1443 // struct) into a new value. For example, given the struct:
1444 // { a, { b, { c, d }, e } }
1445 // and the indices "1, 1" this returns
1448 // It does this by inserting an insertvalue for each element in the resulting
1449 // struct, as opposed to just inserting a single struct. This will only work if
1450 // each of the elements of the substruct are known (ie, inserted into From by an
1451 // insertvalue instruction somewhere).
1453 // All inserted insertvalue instructions are inserted before InsertBefore
1454 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1455 Instruction *InsertBefore) {
1456 assert(InsertBefore && "Must have someplace to insert!");
1457 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1459 Value *To = UndefValue::get(IndexedType);
1460 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1461 unsigned IdxSkip = Idxs.size();
1463 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1466 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1467 /// the scalar value indexed is already around as a register, for example if it
1468 /// were inserted directly into the aggregrate.
1470 /// If InsertBefore is not null, this function will duplicate (modified)
1471 /// insertvalues when a part of a nested struct is extracted.
1472 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1473 Instruction *InsertBefore) {
1474 // Nothing to index? Just return V then (this is useful at the end of our
1476 if (idx_range.empty())
1478 // We have indices, so V should have an indexable type
1479 assert((V->getType()->isStructTy() || V->getType()->isArrayTy())
1480 && "Not looking at a struct or array?");
1481 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range)
1482 && "Invalid indices for type?");
1483 CompositeType *PTy = cast<CompositeType>(V->getType());
1485 if (isa<UndefValue>(V))
1486 return UndefValue::get(ExtractValueInst::getIndexedType(PTy,
1488 else if (isa<ConstantAggregateZero>(V))
1489 return Constant::getNullValue(ExtractValueInst::getIndexedType(PTy,
1491 else if (Constant *C = dyn_cast<Constant>(V)) {
1492 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C))
1493 // Recursively process this constant
1494 return FindInsertedValue(C->getOperand(idx_range[0]), idx_range.slice(1),
1496 } else if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1497 // Loop the indices for the insertvalue instruction in parallel with the
1498 // requested indices
1499 const unsigned *req_idx = idx_range.begin();
1500 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1501 i != e; ++i, ++req_idx) {
1502 if (req_idx == idx_range.end()) {
1504 // The requested index identifies a part of a nested aggregate. Handle
1505 // this specially. For example,
1506 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1507 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1508 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1509 // This can be changed into
1510 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1511 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1512 // which allows the unused 0,0 element from the nested struct to be
1514 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1517 // We can't handle this without inserting insertvalues
1521 // This insert value inserts something else than what we are looking for.
1522 // See if the (aggregrate) value inserted into has the value we are
1523 // looking for, then.
1525 return FindInsertedValue(I->getAggregateOperand(), idx_range,
1528 // If we end up here, the indices of the insertvalue match with those
1529 // requested (though possibly only partially). Now we recursively look at
1530 // the inserted value, passing any remaining indices.
1531 return FindInsertedValue(I->getInsertedValueOperand(),
1532 makeArrayRef(req_idx, idx_range.end()),
1534 } else if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1535 // If we're extracting a value from an aggregrate that was extracted from
1536 // something else, we can extract from that something else directly instead.
1537 // However, we will need to chain I's indices with the requested indices.
1539 // Calculate the number of indices required
1540 unsigned size = I->getNumIndices() + idx_range.size();
1541 // Allocate some space to put the new indices in
1542 SmallVector<unsigned, 5> Idxs;
1544 // Add indices from the extract value instruction
1545 Idxs.append(I->idx_begin(), I->idx_end());
1547 // Add requested indices
1548 Idxs.append(idx_range.begin(), idx_range.end());
1550 assert(Idxs.size() == size
1551 && "Number of indices added not correct?");
1553 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
1555 // Otherwise, we don't know (such as, extracting from a function return value
1556 // or load instruction)
1560 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1561 /// it can be expressed as a base pointer plus a constant offset. Return the
1562 /// base and offset to the caller.
1563 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1564 const TargetData &TD) {
1565 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1566 if (PtrOp == 0 || Ptr->getType()->isVectorTy())
1569 // Just look through bitcasts.
1570 if (PtrOp->getOpcode() == Instruction::BitCast)
1571 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1573 // If this is a GEP with constant indices, we can look through it.
1574 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1575 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1577 gep_type_iterator GTI = gep_type_begin(GEP);
1578 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1580 ConstantInt *OpC = cast<ConstantInt>(*I);
1581 if (OpC->isZero()) continue;
1583 // Handle a struct and array indices which add their offset to the pointer.
1584 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1585 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1587 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1588 Offset += OpC->getSExtValue()*Size;
1592 // Re-sign extend from the pointer size if needed to get overflow edge cases
1594 unsigned PtrSize = TD.getPointerSizeInBits();
1596 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1598 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1602 /// GetConstantStringInfo - This function computes the length of a
1603 /// null-terminated C string pointed to by V. If successful, it returns true
1604 /// and returns the string in Str. If unsuccessful, it returns false.
1605 bool llvm::GetConstantStringInfo(const Value *V, std::string &Str,
1606 uint64_t Offset, bool StopAtNul) {
1607 // If V is NULL then return false;
1608 if (V == NULL) return false;
1610 // Look through bitcast instructions.
1611 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1612 return GetConstantStringInfo(BCI->getOperand(0), Str, Offset, StopAtNul);
1614 // If the value is not a GEP instruction nor a constant expression with a
1615 // GEP instruction, then return false because ConstantArray can't occur
1617 const User *GEP = 0;
1618 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1620 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1621 if (CE->getOpcode() == Instruction::BitCast)
1622 return GetConstantStringInfo(CE->getOperand(0), Str, Offset, StopAtNul);
1623 if (CE->getOpcode() != Instruction::GetElementPtr)
1629 // Make sure the GEP has exactly three arguments.
1630 if (GEP->getNumOperands() != 3)
1633 // Make sure the index-ee is a pointer to array of i8.
1634 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1635 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1636 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1639 // Check to make sure that the first operand of the GEP is an integer and
1640 // has value 0 so that we are sure we're indexing into the initializer.
1641 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1642 if (FirstIdx == 0 || !FirstIdx->isZero())
1645 // If the second index isn't a ConstantInt, then this is a variable index
1646 // into the array. If this occurs, we can't say anything meaningful about
1648 uint64_t StartIdx = 0;
1649 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1650 StartIdx = CI->getZExtValue();
1653 return GetConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset,
1657 // The GEP instruction, constant or instruction, must reference a global
1658 // variable that is a constant and is initialized. The referenced constant
1659 // initializer is the array that we'll use for optimization.
1660 const GlobalVariable* GV = dyn_cast<GlobalVariable>(V);
1661 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1663 const Constant *GlobalInit = GV->getInitializer();
1665 // Handle the all-zeros case
1666 if (GlobalInit->isNullValue()) {
1667 // This is a degenerate case. The initializer is constant zero so the
1668 // length of the string must be zero.
1673 // Must be a Constant Array
1674 const ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1675 if (Array == 0 || !Array->getType()->getElementType()->isIntegerTy(8))
1678 // Get the number of elements in the array
1679 uint64_t NumElts = Array->getType()->getNumElements();
1681 if (Offset > NumElts)
1684 // Traverse the constant array from 'Offset' which is the place the GEP refers
1686 Str.reserve(NumElts-Offset);
1687 for (unsigned i = Offset; i != NumElts; ++i) {
1688 const Constant *Elt = Array->getOperand(i);
1689 const ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1690 if (!CI) // This array isn't suitable, non-int initializer.
1692 if (StopAtNul && CI->isZero())
1693 return true; // we found end of string, success!
1694 Str += (char)CI->getZExtValue();
1697 // The array isn't null terminated, but maybe this is a memcpy, not a strcpy.
1701 // These next two are very similar to the above, but also look through PHI
1703 // TODO: See if we can integrate these two together.
1705 /// GetStringLengthH - If we can compute the length of the string pointed to by
1706 /// the specified pointer, return 'len+1'. If we can't, return 0.
1707 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1708 // Look through noop bitcast instructions.
1709 if (BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1710 return GetStringLengthH(BCI->getOperand(0), PHIs);
1712 // If this is a PHI node, there are two cases: either we have already seen it
1714 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1715 if (!PHIs.insert(PN))
1716 return ~0ULL; // already in the set.
1718 // If it was new, see if all the input strings are the same length.
1719 uint64_t LenSoFar = ~0ULL;
1720 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1721 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1722 if (Len == 0) return 0; // Unknown length -> unknown.
1724 if (Len == ~0ULL) continue;
1726 if (Len != LenSoFar && LenSoFar != ~0ULL)
1727 return 0; // Disagree -> unknown.
1731 // Success, all agree.
1735 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1736 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1737 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1738 if (Len1 == 0) return 0;
1739 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1740 if (Len2 == 0) return 0;
1741 if (Len1 == ~0ULL) return Len2;
1742 if (Len2 == ~0ULL) return Len1;
1743 if (Len1 != Len2) return 0;
1747 // As a special-case, "@string = constant i8 0" is also a string with zero
1748 // length, not wrapped in a bitcast or GEP.
1749 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
1750 if (GV->isConstant() && GV->hasDefinitiveInitializer())
1751 if (GV->getInitializer()->isNullValue()) return 1;
1755 // If the value is not a GEP instruction nor a constant expression with a
1756 // GEP instruction, then return unknown.
1758 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1760 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1761 if (CE->getOpcode() != Instruction::GetElementPtr)
1768 // Make sure the GEP has exactly three arguments.
1769 if (GEP->getNumOperands() != 3)
1772 // Check to make sure that the first operand of the GEP is an integer and
1773 // has value 0 so that we are sure we're indexing into the initializer.
1774 if (ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(1))) {
1780 // If the second index isn't a ConstantInt, then this is a variable index
1781 // into the array. If this occurs, we can't say anything meaningful about
1783 uint64_t StartIdx = 0;
1784 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1785 StartIdx = CI->getZExtValue();
1789 // The GEP instruction, constant or instruction, must reference a global
1790 // variable that is a constant and is initialized. The referenced constant
1791 // initializer is the array that we'll use for optimization.
1792 GlobalVariable* GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
1793 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
1794 GV->mayBeOverridden())
1796 Constant *GlobalInit = GV->getInitializer();
1798 // Handle the ConstantAggregateZero case, which is a degenerate case. The
1799 // initializer is constant zero so the length of the string must be zero.
1800 if (isa<ConstantAggregateZero>(GlobalInit))
1801 return 1; // Len = 0 offset by 1.
1803 // Must be a Constant Array
1804 ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1805 if (!Array || !Array->getType()->getElementType()->isIntegerTy(8))
1808 // Get the number of elements in the array
1809 uint64_t NumElts = Array->getType()->getNumElements();
1811 // Traverse the constant array from StartIdx (derived above) which is
1812 // the place the GEP refers to in the array.
1813 for (unsigned i = StartIdx; i != NumElts; ++i) {
1814 Constant *Elt = Array->getOperand(i);
1815 ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1816 if (!CI) // This array isn't suitable, non-int initializer.
1819 return i-StartIdx+1; // We found end of string, success!
1822 return 0; // The array isn't null terminated, conservatively return 'unknown'.
1825 /// GetStringLength - If we can compute the length of the string pointed to by
1826 /// the specified pointer, return 'len+1'. If we can't, return 0.
1827 uint64_t llvm::GetStringLength(Value *V) {
1828 if (!V->getType()->isPointerTy()) return 0;
1830 SmallPtrSet<PHINode*, 32> PHIs;
1831 uint64_t Len = GetStringLengthH(V, PHIs);
1832 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1833 // an empty string as a length.
1834 return Len == ~0ULL ? 1 : Len;
1838 llvm::GetUnderlyingObject(Value *V, const TargetData *TD, unsigned MaxLookup) {
1839 if (!V->getType()->isPointerTy())
1841 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1842 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1843 V = GEP->getPointerOperand();
1844 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1845 V = cast<Operator>(V)->getOperand(0);
1846 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1847 if (GA->mayBeOverridden())
1849 V = GA->getAliasee();
1851 // See if InstructionSimplify knows any relevant tricks.
1852 if (Instruction *I = dyn_cast<Instruction>(V))
1853 // TODO: Acquire a DominatorTree and use it.
1854 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1861 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1866 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
1867 /// are lifetime markers.
1869 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
1870 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1872 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
1873 if (!II) return false;
1875 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1876 II->getIntrinsicID() != Intrinsic::lifetime_end)
1882 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
1883 const TargetData *TD) {
1884 const Operator *Inst = dyn_cast<Operator>(V);
1888 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
1889 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
1893 switch (Inst->getOpcode()) {
1896 case Instruction::UDiv:
1897 case Instruction::URem:
1898 // x / y is undefined if y == 0, but calcuations like x / 3 are safe.
1899 return isKnownNonZero(Inst->getOperand(1), TD);
1900 case Instruction::SDiv:
1901 case Instruction::SRem: {
1902 Value *Op = Inst->getOperand(1);
1903 // x / y is undefined if y == 0
1904 if (!isKnownNonZero(Op, TD))
1906 // x / y might be undefined if y == -1
1907 unsigned BitWidth = getBitWidth(Op->getType(), TD);
1910 APInt KnownZero(BitWidth, 0);
1911 APInt KnownOne(BitWidth, 0);
1912 ComputeMaskedBits(Op, APInt::getAllOnesValue(BitWidth),
1913 KnownZero, KnownOne, TD);
1916 case Instruction::Load: {
1917 const LoadInst *LI = cast<LoadInst>(Inst);
1918 if (!LI->isUnordered())
1920 return LI->getPointerOperand()->isDereferenceablePointer();
1922 case Instruction::Call: {
1923 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1924 switch (II->getIntrinsicID()) {
1925 case Intrinsic::bswap:
1926 case Intrinsic::ctlz:
1927 case Intrinsic::ctpop:
1928 case Intrinsic::cttz:
1929 case Intrinsic::objectsize:
1930 case Intrinsic::sadd_with_overflow:
1931 case Intrinsic::smul_with_overflow:
1932 case Intrinsic::ssub_with_overflow:
1933 case Intrinsic::uadd_with_overflow:
1934 case Intrinsic::umul_with_overflow:
1935 case Intrinsic::usub_with_overflow:
1937 // TODO: some fp intrinsics are marked as having the same error handling
1938 // as libm. They're safe to speculate when they won't error.
1939 // TODO: are convert_{from,to}_fp16 safe?
1940 // TODO: can we list target-specific intrinsics here?
1944 return false; // The called function could have undefined behavior or
1945 // side-effects, even if marked readnone nounwind.
1947 case Instruction::VAArg:
1948 case Instruction::Alloca:
1949 case Instruction::Invoke:
1950 case Instruction::PHI:
1951 case Instruction::Store:
1952 case Instruction::Ret:
1953 case Instruction::Br:
1954 case Instruction::IndirectBr:
1955 case Instruction::Switch:
1956 case Instruction::Unwind:
1957 case Instruction::Unreachable:
1958 case Instruction::Fence:
1959 case Instruction::LandingPad:
1960 case Instruction::AtomicRMW:
1961 case Instruction::AtomicCmpXchg:
1962 case Instruction::Resume:
1963 return false; // Misc instructions which have effects