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(const 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() || V->getType()->isPointerTy())
67 && "Not integer or pointer type!");
69 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
70 (!V->getType()->isIntOrIntVectorTy() ||
71 V->getType()->getScalarSizeInBits() == BitWidth) &&
72 KnownZero.getBitWidth() == BitWidth &&
73 KnownOne.getBitWidth() == BitWidth &&
74 "V, Mask, KnownOne and KnownZero should have same BitWidth");
76 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
77 // We know all of the bits for a constant!
78 KnownOne = CI->getValue() & Mask;
79 KnownZero = ~KnownOne & Mask;
82 // Null and aggregate-zero are all-zeros.
83 if (isa<ConstantPointerNull>(V) ||
84 isa<ConstantAggregateZero>(V)) {
85 KnownOne.clearAllBits();
89 // Handle a constant vector by taking the intersection of the known bits of
91 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
92 KnownZero.setAllBits(); KnownOne.setAllBits();
93 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
94 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
95 ComputeMaskedBits(CV->getOperand(i), Mask, KnownZero2, KnownOne2,
97 KnownZero &= KnownZero2;
98 KnownOne &= KnownOne2;
102 // The address of an aligned GlobalValue has trailing zeros.
103 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
104 unsigned Align = GV->getAlignment();
105 if (Align == 0 && TD && GV->getType()->getElementType()->isSized()) {
106 const Type *ObjectType = GV->getType()->getElementType();
107 // If the object is defined in the current Module, we'll be giving
108 // it the preferred alignment. Otherwise, we have to assume that it
109 // may only have the minimum ABI alignment.
110 if (!GV->isDeclaration() && !GV->mayBeOverridden())
111 Align = TD->getPrefTypeAlignment(ObjectType);
113 Align = TD->getABITypeAlignment(ObjectType);
116 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
117 CountTrailingZeros_32(Align));
119 KnownZero.clearAllBits();
120 KnownOne.clearAllBits();
123 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
124 // the bits of its aliasee.
125 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
126 if (GA->mayBeOverridden()) {
127 KnownZero.clearAllBits(); KnownOne.clearAllBits();
129 ComputeMaskedBits(GA->getAliasee(), Mask, KnownZero, KnownOne,
135 KnownZero.clearAllBits(); KnownOne.clearAllBits(); // Start out not knowing anything.
137 if (Depth == MaxDepth || Mask == 0)
138 return; // Limit search depth.
140 Operator *I = dyn_cast<Operator>(V);
143 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
144 switch (I->getOpcode()) {
146 case Instruction::And: {
147 // If either the LHS or the RHS are Zero, the result is zero.
148 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
149 APInt Mask2(Mask & ~KnownZero);
150 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
152 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
153 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
155 // Output known-1 bits are only known if set in both the LHS & RHS.
156 KnownOne &= KnownOne2;
157 // Output known-0 are known to be clear if zero in either the LHS | RHS.
158 KnownZero |= KnownZero2;
161 case Instruction::Or: {
162 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
163 APInt Mask2(Mask & ~KnownOne);
164 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
166 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
167 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
169 // Output known-0 bits are only known if clear in both the LHS & RHS.
170 KnownZero &= KnownZero2;
171 // Output known-1 are known to be set if set in either the LHS | RHS.
172 KnownOne |= KnownOne2;
175 case Instruction::Xor: {
176 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
177 ComputeMaskedBits(I->getOperand(0), Mask, 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 known if clear or set in both the LHS & RHS.
183 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
184 // Output known-1 are known to be set if set in only one of the LHS, RHS.
185 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
186 KnownZero = KnownZeroOut;
189 case Instruction::Mul: {
190 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
191 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
192 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
194 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
195 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
197 // If low bits are zero in either operand, output low known-0 bits.
198 // Also compute a conserative estimate for high known-0 bits.
199 // More trickiness is possible, but this is sufficient for the
200 // interesting case of alignment computation.
201 KnownOne.clearAllBits();
202 unsigned TrailZ = KnownZero.countTrailingOnes() +
203 KnownZero2.countTrailingOnes();
204 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
205 KnownZero2.countLeadingOnes(),
206 BitWidth) - BitWidth;
208 TrailZ = std::min(TrailZ, BitWidth);
209 LeadZ = std::min(LeadZ, BitWidth);
210 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
211 APInt::getHighBitsSet(BitWidth, LeadZ);
215 case Instruction::UDiv: {
216 // For the purposes of computing leading zeros we can conservatively
217 // treat a udiv as a logical right shift by the power of 2 known to
218 // be less than the denominator.
219 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
220 ComputeMaskedBits(I->getOperand(0),
221 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
222 unsigned LeadZ = KnownZero2.countLeadingOnes();
224 KnownOne2.clearAllBits();
225 KnownZero2.clearAllBits();
226 ComputeMaskedBits(I->getOperand(1),
227 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
228 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
229 if (RHSUnknownLeadingOnes != BitWidth)
230 LeadZ = std::min(BitWidth,
231 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
233 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
236 case Instruction::Select:
237 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
238 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
240 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
241 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
243 // Only known if known in both the LHS and RHS.
244 KnownOne &= KnownOne2;
245 KnownZero &= KnownZero2;
247 case Instruction::FPTrunc:
248 case Instruction::FPExt:
249 case Instruction::FPToUI:
250 case Instruction::FPToSI:
251 case Instruction::SIToFP:
252 case Instruction::UIToFP:
253 return; // Can't work with floating point.
254 case Instruction::PtrToInt:
255 case Instruction::IntToPtr:
256 // We can't handle these if we don't know the pointer size.
258 // FALL THROUGH and handle them the same as zext/trunc.
259 case Instruction::ZExt:
260 case Instruction::Trunc: {
261 const Type *SrcTy = I->getOperand(0)->getType();
263 unsigned SrcBitWidth;
264 // Note that we handle pointer operands here because of inttoptr/ptrtoint
265 // which fall through here.
266 if (SrcTy->isPointerTy())
267 SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
269 SrcBitWidth = SrcTy->getScalarSizeInBits();
271 APInt MaskIn = Mask.zextOrTrunc(SrcBitWidth);
272 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
273 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
274 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
276 KnownZero = KnownZero.zextOrTrunc(BitWidth);
277 KnownOne = KnownOne.zextOrTrunc(BitWidth);
278 // Any top bits are known to be zero.
279 if (BitWidth > SrcBitWidth)
280 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
283 case Instruction::BitCast: {
284 const Type *SrcTy = I->getOperand(0)->getType();
285 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
286 // TODO: For now, not handling conversions like:
287 // (bitcast i64 %x to <2 x i32>)
288 !I->getType()->isVectorTy()) {
289 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
295 case Instruction::SExt: {
296 // Compute the bits in the result that are not present in the input.
297 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
299 APInt MaskIn = Mask.trunc(SrcBitWidth);
300 KnownZero = KnownZero.trunc(SrcBitWidth);
301 KnownOne = KnownOne.trunc(SrcBitWidth);
302 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
304 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
305 KnownZero = KnownZero.zext(BitWidth);
306 KnownOne = KnownOne.zext(BitWidth);
308 // If the sign bit of the input is known set or clear, then we know the
309 // top bits of the result.
310 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
311 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
312 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
313 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
316 case Instruction::Shl:
317 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
318 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
319 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
320 APInt Mask2(Mask.lshr(ShiftAmt));
321 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
323 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
324 KnownZero <<= ShiftAmt;
325 KnownOne <<= ShiftAmt;
326 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
330 case Instruction::LShr:
331 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
332 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
333 // Compute the new bits that are at the top now.
334 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
336 // Unsigned shift right.
337 APInt Mask2(Mask.shl(ShiftAmt));
338 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
340 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
341 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
342 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
343 // high bits known zero.
344 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
348 case Instruction::AShr:
349 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
350 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
351 // Compute the new bits that are at the top now.
352 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
354 // Signed shift right.
355 APInt Mask2(Mask.shl(ShiftAmt));
356 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
358 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
359 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
360 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
362 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
363 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
364 KnownZero |= HighBits;
365 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
366 KnownOne |= HighBits;
370 case Instruction::Sub: {
371 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
372 // We know that the top bits of C-X are clear if X contains less bits
373 // than C (i.e. no wrap-around can happen). For example, 20-X is
374 // positive if we can prove that X is >= 0 and < 16.
375 if (!CLHS->getValue().isNegative()) {
376 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
377 // NLZ can't be BitWidth with no sign bit
378 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
379 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
382 // If all of the MaskV bits are known to be zero, then we know the
383 // output top bits are zero, because we now know that the output is
385 if ((KnownZero2 & MaskV) == MaskV) {
386 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
387 // Top bits known zero.
388 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
394 case Instruction::Add: {
395 // If one of the operands has trailing zeros, then the bits that the
396 // other operand has in those bit positions will be preserved in the
397 // result. For an add, this works with either operand. For a subtract,
398 // this only works if the known zeros are in the right operand.
399 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
400 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
401 BitWidth - Mask.countLeadingZeros());
402 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
404 assert((LHSKnownZero & LHSKnownOne) == 0 &&
405 "Bits known to be one AND zero?");
406 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
408 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD,
410 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
411 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
413 // Determine which operand has more trailing zeros, and use that
414 // many bits from the other operand.
415 if (LHSKnownZeroOut > RHSKnownZeroOut) {
416 if (I->getOpcode() == Instruction::Add) {
417 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
418 KnownZero |= KnownZero2 & Mask;
419 KnownOne |= KnownOne2 & Mask;
421 // If the known zeros are in the left operand for a subtract,
422 // fall back to the minimum known zeros in both operands.
423 KnownZero |= APInt::getLowBitsSet(BitWidth,
424 std::min(LHSKnownZeroOut,
427 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
428 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
429 KnownZero |= LHSKnownZero & Mask;
430 KnownOne |= LHSKnownOne & Mask;
434 case Instruction::SRem:
435 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
436 APInt RA = Rem->getValue().abs();
437 if (RA.isPowerOf2()) {
438 APInt LowBits = RA - 1;
439 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
440 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
443 // The low bits of the first operand are unchanged by the srem.
444 KnownZero = KnownZero2 & LowBits;
445 KnownOne = KnownOne2 & LowBits;
447 // If the first operand is non-negative or has all low bits zero, then
448 // the upper bits are all zero.
449 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
450 KnownZero |= ~LowBits;
452 // If the first operand is negative and not all low bits are zero, then
453 // the upper bits are all one.
454 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
455 KnownOne |= ~LowBits;
460 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
463 if (Mask.isNegative()) { // We're looking for the sign bit.
464 APInt Mask2 = APInt::getSignBit(BitWidth);
467 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
469 if (KnownOne2[BitWidth-1])
471 if (KnownZero2[BitWidth-1])
473 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
476 case Instruction::URem: {
477 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
478 APInt RA = Rem->getValue();
479 if (RA.isPowerOf2()) {
480 APInt LowBits = (RA - 1);
481 APInt Mask2 = LowBits & Mask;
482 KnownZero |= ~LowBits & Mask;
483 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
485 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
490 // Since the result is less than or equal to either operand, any leading
491 // zero bits in either operand must also exist in the result.
492 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
493 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
495 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
498 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
499 KnownZero2.countLeadingOnes());
500 KnownOne.clearAllBits();
501 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
505 case Instruction::Alloca: {
506 AllocaInst *AI = cast<AllocaInst>(V);
507 unsigned Align = AI->getAlignment();
508 if (Align == 0 && TD)
509 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
512 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
513 CountTrailingZeros_32(Align));
516 case Instruction::GetElementPtr: {
517 // Analyze all of the subscripts of this getelementptr instruction
518 // to determine if we can prove known low zero bits.
519 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
520 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
521 ComputeMaskedBits(I->getOperand(0), LocalMask,
522 LocalKnownZero, LocalKnownOne, TD, Depth+1);
523 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
525 gep_type_iterator GTI = gep_type_begin(I);
526 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
527 Value *Index = I->getOperand(i);
528 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
529 // Handle struct member offset arithmetic.
531 const StructLayout *SL = TD->getStructLayout(STy);
532 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
533 uint64_t Offset = SL->getElementOffset(Idx);
534 TrailZ = std::min(TrailZ,
535 CountTrailingZeros_64(Offset));
537 // Handle array index arithmetic.
538 const Type *IndexedTy = GTI.getIndexedType();
539 if (!IndexedTy->isSized()) return;
540 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
541 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
542 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
543 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
544 ComputeMaskedBits(Index, LocalMask,
545 LocalKnownZero, LocalKnownOne, TD, Depth+1);
546 TrailZ = std::min(TrailZ,
547 unsigned(CountTrailingZeros_64(TypeSize) +
548 LocalKnownZero.countTrailingOnes()));
552 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
555 case Instruction::PHI: {
556 PHINode *P = cast<PHINode>(I);
557 // Handle the case of a simple two-predecessor recurrence PHI.
558 // There's a lot more that could theoretically be done here, but
559 // this is sufficient to catch some interesting cases.
560 if (P->getNumIncomingValues() == 2) {
561 for (unsigned i = 0; i != 2; ++i) {
562 Value *L = P->getIncomingValue(i);
563 Value *R = P->getIncomingValue(!i);
564 Operator *LU = dyn_cast<Operator>(L);
567 unsigned Opcode = LU->getOpcode();
568 // Check for operations that have the property that if
569 // both their operands have low zero bits, the result
570 // will have low zero bits.
571 if (Opcode == Instruction::Add ||
572 Opcode == Instruction::Sub ||
573 Opcode == Instruction::And ||
574 Opcode == Instruction::Or ||
575 Opcode == Instruction::Mul) {
576 Value *LL = LU->getOperand(0);
577 Value *LR = LU->getOperand(1);
578 // Find a recurrence.
585 // Ok, we have a PHI of the form L op= R. Check for low
587 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
588 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
589 Mask2 = APInt::getLowBitsSet(BitWidth,
590 KnownZero2.countTrailingOnes());
592 // We need to take the minimum number of known bits
593 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
594 ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
597 APInt::getLowBitsSet(BitWidth,
598 std::min(KnownZero2.countTrailingOnes(),
599 KnownZero3.countTrailingOnes()));
605 // Unreachable blocks may have zero-operand PHI nodes.
606 if (P->getNumIncomingValues() == 0)
609 // Otherwise take the unions of the known bit sets of the operands,
610 // taking conservative care to avoid excessive recursion.
611 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
612 KnownZero = APInt::getAllOnesValue(BitWidth);
613 KnownOne = APInt::getAllOnesValue(BitWidth);
614 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
615 // Skip direct self references.
616 if (P->getIncomingValue(i) == P) continue;
618 KnownZero2 = APInt(BitWidth, 0);
619 KnownOne2 = APInt(BitWidth, 0);
620 // Recurse, but cap the recursion to one level, because we don't
621 // want to waste time spinning around in loops.
622 ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
623 KnownZero2, KnownOne2, TD, MaxDepth-1);
624 KnownZero &= KnownZero2;
625 KnownOne &= KnownOne2;
626 // If all bits have been ruled out, there's no need to check
628 if (!KnownZero && !KnownOne)
634 case Instruction::Call:
635 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
636 switch (II->getIntrinsicID()) {
638 case Intrinsic::ctpop:
639 case Intrinsic::ctlz:
640 case Intrinsic::cttz: {
641 unsigned LowBits = Log2_32(BitWidth)+1;
642 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
651 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
652 /// one. Convenience wrapper around ComputeMaskedBits.
653 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
654 const TargetData *TD, unsigned Depth) {
655 unsigned BitWidth = getBitWidth(V->getType(), TD);
661 APInt ZeroBits(BitWidth, 0);
662 APInt OneBits(BitWidth, 0);
663 ComputeMaskedBits(V, APInt::getSignBit(BitWidth), ZeroBits, OneBits, TD,
665 KnownOne = OneBits[BitWidth - 1];
666 KnownZero = ZeroBits[BitWidth - 1];
669 /// isPowerOfTwo - Return true if the given value is known to have exactly one
670 /// bit set when defined. For vectors return true if every element is known to
671 /// be a power of two when defined. Supports values with integer or pointer
672 /// types and vectors of integers.
673 bool llvm::isPowerOfTwo(Value *V, const TargetData *TD, unsigned Depth) {
674 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
675 return CI->getValue().isPowerOf2();
676 // TODO: Handle vector constants.
678 // 1 << X is clearly a power of two if the one is not shifted off the end. If
679 // it is shifted off the end then the result is undefined.
680 if (match(V, m_Shl(m_One(), m_Value())))
683 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
684 // bottom. If it is shifted off the bottom then the result is undefined.
685 if (match(V, m_LShr(m_SignBit(), m_Value())))
688 // The remaining tests are all recursive, so bail out if we hit the limit.
689 if (Depth++ == MaxDepth)
692 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
693 return isPowerOfTwo(ZI->getOperand(0), TD, Depth);
695 if (SelectInst *SI = dyn_cast<SelectInst>(V))
696 return isPowerOfTwo(SI->getTrueValue(), TD, Depth) &&
697 isPowerOfTwo(SI->getFalseValue(), TD, Depth);
702 /// isKnownNonZero - Return true if the given value is known to be non-zero
703 /// when defined. For vectors return true if every element is known to be
704 /// non-zero when defined. Supports values with integer or pointer type and
705 /// vectors of integers.
706 bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
707 if (Constant *C = dyn_cast<Constant>(V)) {
708 if (C->isNullValue())
710 if (isa<ConstantInt>(C))
711 // Must be non-zero due to null test above.
713 // TODO: Handle vectors
717 // The remaining tests are all recursive, so bail out if we hit the limit.
718 if (Depth++ == MaxDepth)
721 unsigned BitWidth = getBitWidth(V->getType(), TD);
723 // X | Y != 0 if X != 0 or Y != 0.
724 Value *X = 0, *Y = 0;
725 if (match(V, m_Or(m_Value(X), m_Value(Y))))
726 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
728 // ext X != 0 if X != 0.
729 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
730 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
732 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
733 // if the lowest bit is shifted off the end.
734 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
735 APInt KnownZero(BitWidth, 0);
736 APInt KnownOne(BitWidth, 0);
737 ComputeMaskedBits(X, APInt(BitWidth, 1), KnownZero, KnownOne, TD, Depth);
741 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
742 // defined if the sign bit is shifted off the end.
743 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
744 bool XKnownNonNegative, XKnownNegative;
745 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
750 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
751 bool XKnownNonNegative, XKnownNegative;
752 bool YKnownNonNegative, YKnownNegative;
753 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
754 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
756 // If X and Y are both non-negative (as signed values) then their sum is not
757 // zero unless both X and Y are zero.
758 if (XKnownNonNegative && YKnownNonNegative)
759 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
762 // If X and Y are both negative (as signed values) then their sum is not
763 // zero unless both X and Y equal INT_MIN.
764 if (BitWidth && XKnownNegative && YKnownNegative) {
765 APInt KnownZero(BitWidth, 0);
766 APInt KnownOne(BitWidth, 0);
767 APInt Mask = APInt::getSignedMaxValue(BitWidth);
768 // The sign bit of X is set. If some other bit is set then X is not equal
770 ComputeMaskedBits(X, Mask, KnownZero, KnownOne, TD, Depth);
771 if ((KnownOne & Mask) != 0)
773 // The sign bit of Y is set. If some other bit is set then Y is not equal
775 ComputeMaskedBits(Y, Mask, KnownZero, KnownOne, TD, Depth);
776 if ((KnownOne & Mask) != 0)
780 // The sum of a non-negative number and a power of two is not zero.
781 if (XKnownNonNegative && isPowerOfTwo(Y, TD, Depth))
783 if (YKnownNonNegative && isPowerOfTwo(X, TD, Depth))
786 // (C ? X : Y) != 0 if X != 0 and Y != 0.
787 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
788 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
789 isKnownNonZero(SI->getFalseValue(), TD, Depth))
793 if (!BitWidth) return false;
794 APInt KnownZero(BitWidth, 0);
795 APInt KnownOne(BitWidth, 0);
796 ComputeMaskedBits(V, APInt::getAllOnesValue(BitWidth), KnownZero, KnownOne,
798 return KnownOne != 0;
801 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
802 /// this predicate to simplify operations downstream. Mask is known to be zero
803 /// for bits that V cannot have.
805 /// This function is defined on values with integer type, values with pointer
806 /// type (but only if TD is non-null), and vectors of integers. In the case
807 /// where V is a vector, the mask, known zero, and known one values are the
808 /// same width as the vector element, and the bit is set only if it is true
809 /// for all of the elements in the vector.
810 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
811 const TargetData *TD, unsigned Depth) {
812 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
813 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
814 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
815 return (KnownZero & Mask) == Mask;
820 /// ComputeNumSignBits - Return the number of times the sign bit of the
821 /// register is replicated into the other bits. We know that at least 1 bit
822 /// is always equal to the sign bit (itself), but other cases can give us
823 /// information. For example, immediately after an "ashr X, 2", we know that
824 /// the top 3 bits are all equal to each other, so we return 3.
826 /// 'Op' must have a scalar integer type.
828 unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
830 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
831 "ComputeNumSignBits requires a TargetData object to operate "
832 "on non-integer values!");
833 const Type *Ty = V->getType();
834 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
835 Ty->getScalarSizeInBits();
837 unsigned FirstAnswer = 1;
839 // Note that ConstantInt is handled by the general ComputeMaskedBits case
843 return 1; // Limit search depth.
845 Operator *U = dyn_cast<Operator>(V);
846 switch (Operator::getOpcode(V)) {
848 case Instruction::SExt:
849 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
850 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
852 case Instruction::AShr:
853 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
854 // ashr X, C -> adds C sign bits.
855 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
856 Tmp += C->getZExtValue();
857 if (Tmp > TyBits) Tmp = TyBits;
859 // vector ashr X, <C, C, C, C> -> adds C sign bits
860 if (ConstantVector *C = dyn_cast<ConstantVector>(U->getOperand(1))) {
861 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) {
862 Tmp += CI->getZExtValue();
863 if (Tmp > TyBits) Tmp = TyBits;
867 case Instruction::Shl:
868 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
869 // shl destroys sign bits.
870 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
871 if (C->getZExtValue() >= TyBits || // Bad shift.
872 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
873 return Tmp - C->getZExtValue();
876 case Instruction::And:
877 case Instruction::Or:
878 case Instruction::Xor: // NOT is handled here.
879 // Logical binary ops preserve the number of sign bits at the worst.
880 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
882 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
883 FirstAnswer = std::min(Tmp, Tmp2);
884 // We computed what we know about the sign bits as our first
885 // answer. Now proceed to the generic code that uses
886 // ComputeMaskedBits, and pick whichever answer is better.
890 case Instruction::Select:
891 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
892 if (Tmp == 1) return 1; // Early out.
893 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
894 return std::min(Tmp, Tmp2);
896 case Instruction::Add:
897 // Add can have at most one carry bit. Thus we know that the output
898 // is, at worst, one more bit than the inputs.
899 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
900 if (Tmp == 1) return 1; // Early out.
902 // Special case decrementing a value (ADD X, -1):
903 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
904 if (CRHS->isAllOnesValue()) {
905 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
906 APInt Mask = APInt::getAllOnesValue(TyBits);
907 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
910 // If the input is known to be 0 or 1, the output is 0/-1, which is all
912 if ((KnownZero | APInt(TyBits, 1)) == Mask)
915 // If we are subtracting one from a positive number, there is no carry
916 // out of the result.
917 if (KnownZero.isNegative())
921 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
922 if (Tmp2 == 1) return 1;
923 return std::min(Tmp, Tmp2)-1;
925 case Instruction::Sub:
926 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
927 if (Tmp2 == 1) return 1;
930 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
931 if (CLHS->isNullValue()) {
932 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
933 APInt Mask = APInt::getAllOnesValue(TyBits);
934 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
936 // If the input is known to be 0 or 1, the output is 0/-1, which is all
938 if ((KnownZero | APInt(TyBits, 1)) == Mask)
941 // If the input is known to be positive (the sign bit is known clear),
942 // the output of the NEG has the same number of sign bits as the input.
943 if (KnownZero.isNegative())
946 // Otherwise, we treat this like a SUB.
949 // Sub can have at most one carry bit. Thus we know that the output
950 // is, at worst, one more bit than the inputs.
951 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
952 if (Tmp == 1) return 1; // Early out.
953 return std::min(Tmp, Tmp2)-1;
955 case Instruction::PHI: {
956 PHINode *PN = cast<PHINode>(U);
957 // Don't analyze large in-degree PHIs.
958 if (PN->getNumIncomingValues() > 4) break;
960 // Take the minimum of all incoming values. This can't infinitely loop
961 // because of our depth threshold.
962 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
963 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
964 if (Tmp == 1) return Tmp;
966 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
971 case Instruction::Trunc:
972 // FIXME: it's tricky to do anything useful for this, but it is an important
973 // case for targets like X86.
977 // Finally, if we can prove that the top bits of the result are 0's or 1's,
978 // use this information.
979 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
980 APInt Mask = APInt::getAllOnesValue(TyBits);
981 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
983 if (KnownZero.isNegative()) { // sign bit is 0
985 } else if (KnownOne.isNegative()) { // sign bit is 1;
992 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
993 // the number of identical bits in the top of the input value.
995 Mask <<= Mask.getBitWidth()-TyBits;
996 // Return # leading zeros. We use 'min' here in case Val was zero before
997 // shifting. We don't want to return '64' as for an i32 "0".
998 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1001 /// ComputeMultiple - This function computes the integer multiple of Base that
1002 /// equals V. If successful, it returns true and returns the multiple in
1003 /// Multiple. If unsuccessful, it returns false. It looks
1004 /// through SExt instructions only if LookThroughSExt is true.
1005 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1006 bool LookThroughSExt, unsigned Depth) {
1007 const unsigned MaxDepth = 6;
1009 assert(V && "No Value?");
1010 assert(Depth <= MaxDepth && "Limit Search Depth");
1011 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1013 const Type *T = V->getType();
1015 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1025 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1026 Constant *BaseVal = ConstantInt::get(T, Base);
1027 if (CO && CO == BaseVal) {
1029 Multiple = ConstantInt::get(T, 1);
1033 if (CI && CI->getZExtValue() % Base == 0) {
1034 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1038 if (Depth == MaxDepth) return false; // Limit search depth.
1040 Operator *I = dyn_cast<Operator>(V);
1041 if (!I) return false;
1043 switch (I->getOpcode()) {
1045 case Instruction::SExt:
1046 if (!LookThroughSExt) return false;
1047 // otherwise fall through to ZExt
1048 case Instruction::ZExt:
1049 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1050 LookThroughSExt, Depth+1);
1051 case Instruction::Shl:
1052 case Instruction::Mul: {
1053 Value *Op0 = I->getOperand(0);
1054 Value *Op1 = I->getOperand(1);
1056 if (I->getOpcode() == Instruction::Shl) {
1057 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1058 if (!Op1CI) return false;
1059 // Turn Op0 << Op1 into Op0 * 2^Op1
1060 APInt Op1Int = Op1CI->getValue();
1061 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1062 APInt API(Op1Int.getBitWidth(), 0);
1063 API.setBit(BitToSet);
1064 Op1 = ConstantInt::get(V->getContext(), API);
1068 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1069 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1070 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1071 if (Op1C->getType()->getPrimitiveSizeInBits() <
1072 MulC->getType()->getPrimitiveSizeInBits())
1073 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1074 if (Op1C->getType()->getPrimitiveSizeInBits() >
1075 MulC->getType()->getPrimitiveSizeInBits())
1076 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1078 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1079 Multiple = ConstantExpr::getMul(MulC, Op1C);
1083 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1084 if (Mul0CI->getValue() == 1) {
1085 // V == Base * Op1, so return Op1
1092 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1093 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1094 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1095 if (Op0C->getType()->getPrimitiveSizeInBits() <
1096 MulC->getType()->getPrimitiveSizeInBits())
1097 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1098 if (Op0C->getType()->getPrimitiveSizeInBits() >
1099 MulC->getType()->getPrimitiveSizeInBits())
1100 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1102 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1103 Multiple = ConstantExpr::getMul(MulC, Op0C);
1107 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1108 if (Mul1CI->getValue() == 1) {
1109 // V == Base * Op0, so return Op0
1117 // We could not determine if V is a multiple of Base.
1121 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1122 /// value is never equal to -0.0.
1124 /// NOTE: this function will need to be revisited when we support non-default
1127 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1128 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1129 return !CFP->getValueAPF().isNegZero();
1132 return 1; // Limit search depth.
1134 const Operator *I = dyn_cast<Operator>(V);
1135 if (I == 0) return false;
1137 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1138 if (I->getOpcode() == Instruction::FAdd &&
1139 isa<ConstantFP>(I->getOperand(1)) &&
1140 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1143 // sitofp and uitofp turn into +0.0 for zero.
1144 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1147 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1148 // sqrt(-0.0) = -0.0, no other negative results are possible.
1149 if (II->getIntrinsicID() == Intrinsic::sqrt)
1150 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1152 if (const CallInst *CI = dyn_cast<CallInst>(I))
1153 if (const Function *F = CI->getCalledFunction()) {
1154 if (F->isDeclaration()) {
1156 if (F->getName() == "abs") return true;
1157 // fabs[lf](x) != -0.0
1158 if (F->getName() == "fabs") return true;
1159 if (F->getName() == "fabsf") return true;
1160 if (F->getName() == "fabsl") return true;
1161 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1162 F->getName() == "sqrtl")
1163 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1170 /// isBytewiseValue - If the specified value can be set by repeating the same
1171 /// byte in memory, return the i8 value that it is represented with. This is
1172 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1173 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1174 /// byte store (e.g. i16 0x1234), return null.
1175 Value *llvm::isBytewiseValue(Value *V) {
1176 // All byte-wide stores are splatable, even of arbitrary variables.
1177 if (V->getType()->isIntegerTy(8)) return V;
1179 // Handle 'null' ConstantArrayZero etc.
1180 if (Constant *C = dyn_cast<Constant>(V))
1181 if (C->isNullValue())
1182 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1184 // Constant float and double values can be handled as integer values if the
1185 // corresponding integer value is "byteable". An important case is 0.0.
1186 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1187 if (CFP->getType()->isFloatTy())
1188 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1189 if (CFP->getType()->isDoubleTy())
1190 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1191 // Don't handle long double formats, which have strange constraints.
1194 // We can handle constant integers that are power of two in size and a
1195 // multiple of 8 bits.
1196 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1197 unsigned Width = CI->getBitWidth();
1198 if (isPowerOf2_32(Width) && Width > 8) {
1199 // We can handle this value if the recursive binary decomposition is the
1200 // same at all levels.
1201 APInt Val = CI->getValue();
1203 while (Val.getBitWidth() != 8) {
1204 unsigned NextWidth = Val.getBitWidth()/2;
1205 Val2 = Val.lshr(NextWidth);
1206 Val2 = Val2.trunc(Val.getBitWidth()/2);
1207 Val = Val.trunc(Val.getBitWidth()/2);
1209 // If the top/bottom halves aren't the same, reject it.
1213 return ConstantInt::get(V->getContext(), Val);
1217 // A ConstantArray is splatable if all its members are equal and also
1219 if (ConstantArray *CA = dyn_cast<ConstantArray>(V)) {
1220 if (CA->getNumOperands() == 0)
1223 Value *Val = isBytewiseValue(CA->getOperand(0));
1227 for (unsigned I = 1, E = CA->getNumOperands(); I != E; ++I)
1228 if (CA->getOperand(I-1) != CA->getOperand(I))
1234 // Conceptually, we could handle things like:
1235 // %a = zext i8 %X to i16
1236 // %b = shl i16 %a, 8
1237 // %c = or i16 %a, %b
1238 // but until there is an example that actually needs this, it doesn't seem
1239 // worth worrying about.
1244 // This is the recursive version of BuildSubAggregate. It takes a few different
1245 // arguments. Idxs is the index within the nested struct From that we are
1246 // looking at now (which is of type IndexedType). IdxSkip is the number of
1247 // indices from Idxs that should be left out when inserting into the resulting
1248 // struct. To is the result struct built so far, new insertvalue instructions
1250 static Value *BuildSubAggregate(Value *From, Value* To, const Type *IndexedType,
1251 SmallVector<unsigned, 10> &Idxs,
1253 Instruction *InsertBefore) {
1254 const llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1256 // Save the original To argument so we can modify it
1258 // General case, the type indexed by Idxs is a struct
1259 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1260 // Process each struct element recursively
1263 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1267 // Couldn't find any inserted value for this index? Cleanup
1268 while (PrevTo != OrigTo) {
1269 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1270 PrevTo = Del->getAggregateOperand();
1271 Del->eraseFromParent();
1273 // Stop processing elements
1277 // If we succesfully found a value for each of our subaggregates
1281 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1282 // the struct's elements had a value that was inserted directly. In the latter
1283 // case, perhaps we can't determine each of the subelements individually, but
1284 // we might be able to find the complete struct somewhere.
1286 // Find the value that is at that particular spot
1287 Value *V = FindInsertedValue(From, Idxs.begin(), Idxs.end());
1292 // Insert the value in the new (sub) aggregrate
1293 return llvm::InsertValueInst::Create(To, V, Idxs.begin() + IdxSkip,
1294 Idxs.end(), "tmp", InsertBefore);
1297 // This helper takes a nested struct and extracts a part of it (which is again a
1298 // struct) into a new value. For example, given the struct:
1299 // { a, { b, { c, d }, e } }
1300 // and the indices "1, 1" this returns
1303 // It does this by inserting an insertvalue for each element in the resulting
1304 // struct, as opposed to just inserting a single struct. This will only work if
1305 // each of the elements of the substruct are known (ie, inserted into From by an
1306 // insertvalue instruction somewhere).
1308 // All inserted insertvalue instructions are inserted before InsertBefore
1309 static Value *BuildSubAggregate(Value *From, const unsigned *idx_begin,
1310 const unsigned *idx_end,
1311 Instruction *InsertBefore) {
1312 assert(InsertBefore && "Must have someplace to insert!");
1313 const Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1316 Value *To = UndefValue::get(IndexedType);
1317 SmallVector<unsigned, 10> Idxs(idx_begin, idx_end);
1318 unsigned IdxSkip = Idxs.size();
1320 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1323 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1324 /// the scalar value indexed is already around as a register, for example if it
1325 /// were inserted directly into the aggregrate.
1327 /// If InsertBefore is not null, this function will duplicate (modified)
1328 /// insertvalues when a part of a nested struct is extracted.
1329 Value *llvm::FindInsertedValue(Value *V, const unsigned *idx_begin,
1330 const unsigned *idx_end, Instruction *InsertBefore) {
1331 // Nothing to index? Just return V then (this is useful at the end of our
1333 if (idx_begin == idx_end)
1335 // We have indices, so V should have an indexable type
1336 assert((V->getType()->isStructTy() || V->getType()->isArrayTy())
1337 && "Not looking at a struct or array?");
1338 assert(ExtractValueInst::getIndexedType(V->getType(), idx_begin, idx_end)
1339 && "Invalid indices for type?");
1340 const CompositeType *PTy = cast<CompositeType>(V->getType());
1342 if (isa<UndefValue>(V))
1343 return UndefValue::get(ExtractValueInst::getIndexedType(PTy,
1346 else if (isa<ConstantAggregateZero>(V))
1347 return Constant::getNullValue(ExtractValueInst::getIndexedType(PTy,
1350 else if (Constant *C = dyn_cast<Constant>(V)) {
1351 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C))
1352 // Recursively process this constant
1353 return FindInsertedValue(C->getOperand(*idx_begin), idx_begin + 1,
1354 idx_end, InsertBefore);
1355 } else if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1356 // Loop the indices for the insertvalue instruction in parallel with the
1357 // requested indices
1358 const unsigned *req_idx = idx_begin;
1359 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1360 i != e; ++i, ++req_idx) {
1361 if (req_idx == idx_end) {
1363 // The requested index identifies a part of a nested aggregate. Handle
1364 // this specially. For example,
1365 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1366 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1367 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1368 // This can be changed into
1369 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1370 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1371 // which allows the unused 0,0 element from the nested struct to be
1373 return BuildSubAggregate(V, idx_begin, req_idx, InsertBefore);
1375 // We can't handle this without inserting insertvalues
1379 // This insert value inserts something else than what we are looking for.
1380 // See if the (aggregrate) value inserted into has the value we are
1381 // looking for, then.
1383 return FindInsertedValue(I->getAggregateOperand(), idx_begin, idx_end,
1386 // If we end up here, the indices of the insertvalue match with those
1387 // requested (though possibly only partially). Now we recursively look at
1388 // the inserted value, passing any remaining indices.
1389 return FindInsertedValue(I->getInsertedValueOperand(), req_idx, idx_end,
1391 } else if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1392 // If we're extracting a value from an aggregrate that was extracted from
1393 // something else, we can extract from that something else directly instead.
1394 // However, we will need to chain I's indices with the requested indices.
1396 // Calculate the number of indices required
1397 unsigned size = I->getNumIndices() + (idx_end - idx_begin);
1398 // Allocate some space to put the new indices in
1399 SmallVector<unsigned, 5> Idxs;
1401 // Add indices from the extract value instruction
1402 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1406 // Add requested indices
1407 for (const unsigned *i = idx_begin, *e = idx_end; i != e; ++i)
1410 assert(Idxs.size() == size
1411 && "Number of indices added not correct?");
1413 return FindInsertedValue(I->getAggregateOperand(), Idxs.begin(), Idxs.end(),
1416 // Otherwise, we don't know (such as, extracting from a function return value
1417 // or load instruction)
1421 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1422 /// it can be expressed as a base pointer plus a constant offset. Return the
1423 /// base and offset to the caller.
1424 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1425 const TargetData &TD) {
1426 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1427 if (PtrOp == 0) return Ptr;
1429 // Just look through bitcasts.
1430 if (PtrOp->getOpcode() == Instruction::BitCast)
1431 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1433 // If this is a GEP with constant indices, we can look through it.
1434 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1435 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1437 gep_type_iterator GTI = gep_type_begin(GEP);
1438 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1440 ConstantInt *OpC = cast<ConstantInt>(*I);
1441 if (OpC->isZero()) continue;
1443 // Handle a struct and array indices which add their offset to the pointer.
1444 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
1445 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1447 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1448 Offset += OpC->getSExtValue()*Size;
1452 // Re-sign extend from the pointer size if needed to get overflow edge cases
1454 unsigned PtrSize = TD.getPointerSizeInBits();
1456 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1458 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1462 /// GetConstantStringInfo - This function computes the length of a
1463 /// null-terminated C string pointed to by V. If successful, it returns true
1464 /// and returns the string in Str. If unsuccessful, it returns false.
1465 bool llvm::GetConstantStringInfo(const Value *V, std::string &Str,
1468 // If V is NULL then return false;
1469 if (V == NULL) return false;
1471 // Look through bitcast instructions.
1472 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1473 return GetConstantStringInfo(BCI->getOperand(0), Str, Offset, StopAtNul);
1475 // If the value is not a GEP instruction nor a constant expression with a
1476 // GEP instruction, then return false because ConstantArray can't occur
1478 const User *GEP = 0;
1479 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1481 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1482 if (CE->getOpcode() == Instruction::BitCast)
1483 return GetConstantStringInfo(CE->getOperand(0), Str, Offset, StopAtNul);
1484 if (CE->getOpcode() != Instruction::GetElementPtr)
1490 // Make sure the GEP has exactly three arguments.
1491 if (GEP->getNumOperands() != 3)
1494 // Make sure the index-ee is a pointer to array of i8.
1495 const PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1496 const ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1497 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1500 // Check to make sure that the first operand of the GEP is an integer and
1501 // has value 0 so that we are sure we're indexing into the initializer.
1502 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1503 if (FirstIdx == 0 || !FirstIdx->isZero())
1506 // If the second index isn't a ConstantInt, then this is a variable index
1507 // into the array. If this occurs, we can't say anything meaningful about
1509 uint64_t StartIdx = 0;
1510 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1511 StartIdx = CI->getZExtValue();
1514 return GetConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset,
1518 // The GEP instruction, constant or instruction, must reference a global
1519 // variable that is a constant and is initialized. The referenced constant
1520 // initializer is the array that we'll use for optimization.
1521 const GlobalVariable* GV = dyn_cast<GlobalVariable>(V);
1522 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1524 const Constant *GlobalInit = GV->getInitializer();
1526 // Handle the ConstantAggregateZero case
1527 if (isa<ConstantAggregateZero>(GlobalInit)) {
1528 // This is a degenerate case. The initializer is constant zero so the
1529 // length of the string must be zero.
1534 // Must be a Constant Array
1535 const ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1536 if (Array == 0 || !Array->getType()->getElementType()->isIntegerTy(8))
1539 // Get the number of elements in the array
1540 uint64_t NumElts = Array->getType()->getNumElements();
1542 if (Offset > NumElts)
1545 // Traverse the constant array from 'Offset' which is the place the GEP refers
1547 Str.reserve(NumElts-Offset);
1548 for (unsigned i = Offset; i != NumElts; ++i) {
1549 const Constant *Elt = Array->getOperand(i);
1550 const ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1551 if (!CI) // This array isn't suitable, non-int initializer.
1553 if (StopAtNul && CI->isZero())
1554 return true; // we found end of string, success!
1555 Str += (char)CI->getZExtValue();
1558 // The array isn't null terminated, but maybe this is a memcpy, not a strcpy.
1562 // These next two are very similar to the above, but also look through PHI
1564 // TODO: See if we can integrate these two together.
1566 /// GetStringLengthH - If we can compute the length of the string pointed to by
1567 /// the specified pointer, return 'len+1'. If we can't, return 0.
1568 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1569 // Look through noop bitcast instructions.
1570 if (BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1571 return GetStringLengthH(BCI->getOperand(0), PHIs);
1573 // If this is a PHI node, there are two cases: either we have already seen it
1575 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1576 if (!PHIs.insert(PN))
1577 return ~0ULL; // already in the set.
1579 // If it was new, see if all the input strings are the same length.
1580 uint64_t LenSoFar = ~0ULL;
1581 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1582 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1583 if (Len == 0) return 0; // Unknown length -> unknown.
1585 if (Len == ~0ULL) continue;
1587 if (Len != LenSoFar && LenSoFar != ~0ULL)
1588 return 0; // Disagree -> unknown.
1592 // Success, all agree.
1596 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1597 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1598 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1599 if (Len1 == 0) return 0;
1600 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1601 if (Len2 == 0) return 0;
1602 if (Len1 == ~0ULL) return Len2;
1603 if (Len2 == ~0ULL) return Len1;
1604 if (Len1 != Len2) return 0;
1608 // If the value is not a GEP instruction nor a constant expression with a
1609 // GEP instruction, then return unknown.
1611 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1613 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1614 if (CE->getOpcode() != Instruction::GetElementPtr)
1621 // Make sure the GEP has exactly three arguments.
1622 if (GEP->getNumOperands() != 3)
1625 // Check to make sure that the first operand of the GEP is an integer and
1626 // has value 0 so that we are sure we're indexing into the initializer.
1627 if (ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(1))) {
1633 // If the second index isn't a ConstantInt, then this is a variable index
1634 // into the array. If this occurs, we can't say anything meaningful about
1636 uint64_t StartIdx = 0;
1637 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1638 StartIdx = CI->getZExtValue();
1642 // The GEP instruction, constant or instruction, must reference a global
1643 // variable that is a constant and is initialized. The referenced constant
1644 // initializer is the array that we'll use for optimization.
1645 GlobalVariable* GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
1646 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
1647 GV->mayBeOverridden())
1649 Constant *GlobalInit = GV->getInitializer();
1651 // Handle the ConstantAggregateZero case, which is a degenerate case. The
1652 // initializer is constant zero so the length of the string must be zero.
1653 if (isa<ConstantAggregateZero>(GlobalInit))
1654 return 1; // Len = 0 offset by 1.
1656 // Must be a Constant Array
1657 ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1658 if (!Array || !Array->getType()->getElementType()->isIntegerTy(8))
1661 // Get the number of elements in the array
1662 uint64_t NumElts = Array->getType()->getNumElements();
1664 // Traverse the constant array from StartIdx (derived above) which is
1665 // the place the GEP refers to in the array.
1666 for (unsigned i = StartIdx; i != NumElts; ++i) {
1667 Constant *Elt = Array->getOperand(i);
1668 ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1669 if (!CI) // This array isn't suitable, non-int initializer.
1672 return i-StartIdx+1; // We found end of string, success!
1675 return 0; // The array isn't null terminated, conservatively return 'unknown'.
1678 /// GetStringLength - If we can compute the length of the string pointed to by
1679 /// the specified pointer, return 'len+1'. If we can't, return 0.
1680 uint64_t llvm::GetStringLength(Value *V) {
1681 if (!V->getType()->isPointerTy()) return 0;
1683 SmallPtrSet<PHINode*, 32> PHIs;
1684 uint64_t Len = GetStringLengthH(V, PHIs);
1685 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1686 // an empty string as a length.
1687 return Len == ~0ULL ? 1 : Len;
1691 llvm::GetUnderlyingObject(Value *V, const TargetData *TD, unsigned MaxLookup) {
1692 if (!V->getType()->isPointerTy())
1694 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1695 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1696 V = GEP->getPointerOperand();
1697 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1698 V = cast<Operator>(V)->getOperand(0);
1699 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1700 if (GA->mayBeOverridden())
1702 V = GA->getAliasee();
1704 // See if InstructionSimplify knows any relevant tricks.
1705 if (Instruction *I = dyn_cast<Instruction>(V))
1706 // TODO: Aquire a DominatorTree and use it.
1707 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1714 assert(V->getType()->isPointerTy() && "Unexpected operand type!");