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/Constants.h"
17 #include "llvm/Instructions.h"
18 #include "llvm/GlobalVariable.h"
19 #include "llvm/IntrinsicInst.h"
20 #include "llvm/Target/TargetData.h"
21 #include "llvm/Support/GetElementPtrTypeIterator.h"
22 #include "llvm/Support/MathExtras.h"
26 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
27 /// opcode value. Otherwise return UserOp1.
28 static unsigned getOpcode(const Value *V) {
29 if (const Instruction *I = dyn_cast<Instruction>(V))
30 return I->getOpcode();
31 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
32 return CE->getOpcode();
33 // Use UserOp1 to mean there's no opcode.
34 return Instruction::UserOp1;
38 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
39 /// known to be either zero or one and return them in the KnownZero/KnownOne
40 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
42 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
43 /// we cannot optimize based on the assumption that it is zero without changing
44 /// it to be an explicit zero. If we don't change it to zero, other code could
45 /// optimized based on the contradictory assumption that it is non-zero.
46 /// Because instcombine aggressively folds operations with undef args anyway,
47 /// this won't lose us code quality.
48 void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
49 APInt &KnownZero, APInt &KnownOne,
50 TargetData *TD, unsigned Depth) {
51 const unsigned MaxDepth = 6;
52 assert(V && "No Value?");
53 assert(Depth <= MaxDepth && "Limit Search Depth");
54 unsigned BitWidth = Mask.getBitWidth();
55 assert((V->getType()->isIntOrIntVector() || isa<PointerType>(V->getType())) &&
56 "Not integer or pointer type!");
58 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
59 (!V->getType()->isIntOrIntVector() ||
60 V->getType()->getScalarSizeInBits() == BitWidth) &&
61 KnownZero.getBitWidth() == BitWidth &&
62 KnownOne.getBitWidth() == BitWidth &&
63 "V, Mask, KnownOne and KnownZero should have same BitWidth");
65 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
66 // We know all of the bits for a constant!
67 KnownOne = CI->getValue() & Mask;
68 KnownZero = ~KnownOne & Mask;
71 // Null and aggregate-zero are all-zeros.
72 if (isa<ConstantPointerNull>(V) ||
73 isa<ConstantAggregateZero>(V)) {
78 // Handle a constant vector by taking the intersection of the known bits of
80 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
81 KnownZero.set(); KnownOne.set();
82 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
83 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
84 ComputeMaskedBits(CV->getOperand(i), Mask, KnownZero2, KnownOne2,
86 KnownZero &= KnownZero2;
87 KnownOne &= KnownOne2;
91 // The address of an aligned GlobalValue has trailing zeros.
92 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
93 unsigned Align = GV->getAlignment();
94 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
95 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
97 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
98 CountTrailingZeros_32(Align));
105 KnownZero.clear(); KnownOne.clear(); // Start out not knowing anything.
107 if (Depth == MaxDepth || Mask == 0)
108 return; // Limit search depth.
110 User *I = dyn_cast<User>(V);
113 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
114 switch (getOpcode(I)) {
116 case Instruction::And: {
117 // If either the LHS or the RHS are Zero, the result is zero.
118 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
119 APInt Mask2(Mask & ~KnownZero);
120 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
122 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
123 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
125 // Output known-1 bits are only known if set in both the LHS & RHS.
126 KnownOne &= KnownOne2;
127 // Output known-0 are known to be clear if zero in either the LHS | RHS.
128 KnownZero |= KnownZero2;
131 case Instruction::Or: {
132 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
133 APInt Mask2(Mask & ~KnownOne);
134 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
136 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
137 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
139 // Output known-0 bits are only known if clear in both the LHS & RHS.
140 KnownZero &= KnownZero2;
141 // Output known-1 are known to be set if set in either the LHS | RHS.
142 KnownOne |= KnownOne2;
145 case Instruction::Xor: {
146 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
147 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
149 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
150 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
152 // Output known-0 bits are known if clear or set in both the LHS & RHS.
153 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
154 // Output known-1 are known to be set if set in only one of the LHS, RHS.
155 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
156 KnownZero = KnownZeroOut;
159 case Instruction::Mul: {
160 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
161 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
162 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
164 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
165 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
167 // If low bits are zero in either operand, output low known-0 bits.
168 // Also compute a conserative estimate for high known-0 bits.
169 // More trickiness is possible, but this is sufficient for the
170 // interesting case of alignment computation.
172 unsigned TrailZ = KnownZero.countTrailingOnes() +
173 KnownZero2.countTrailingOnes();
174 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
175 KnownZero2.countLeadingOnes(),
176 BitWidth) - BitWidth;
178 TrailZ = std::min(TrailZ, BitWidth);
179 LeadZ = std::min(LeadZ, BitWidth);
180 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
181 APInt::getHighBitsSet(BitWidth, LeadZ);
185 case Instruction::UDiv: {
186 // For the purposes of computing leading zeros we can conservatively
187 // treat a udiv as a logical right shift by the power of 2 known to
188 // be less than the denominator.
189 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
190 ComputeMaskedBits(I->getOperand(0),
191 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
192 unsigned LeadZ = KnownZero2.countLeadingOnes();
196 ComputeMaskedBits(I->getOperand(1),
197 AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
198 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
199 if (RHSUnknownLeadingOnes != BitWidth)
200 LeadZ = std::min(BitWidth,
201 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
203 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
206 case Instruction::Select:
207 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
208 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
210 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
211 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
213 // Only known if known in both the LHS and RHS.
214 KnownOne &= KnownOne2;
215 KnownZero &= KnownZero2;
217 case Instruction::FPTrunc:
218 case Instruction::FPExt:
219 case Instruction::FPToUI:
220 case Instruction::FPToSI:
221 case Instruction::SIToFP:
222 case Instruction::UIToFP:
223 return; // Can't work with floating point.
224 case Instruction::PtrToInt:
225 case Instruction::IntToPtr:
226 // We can't handle these if we don't know the pointer size.
228 // FALL THROUGH and handle them the same as zext/trunc.
229 case Instruction::ZExt:
230 case Instruction::Trunc: {
231 // Note that we handle pointer operands here because of inttoptr/ptrtoint
232 // which fall through here.
233 const Type *SrcTy = I->getOperand(0)->getType();
234 unsigned SrcBitWidth = TD ?
235 TD->getTypeSizeInBits(SrcTy) :
236 SrcTy->getScalarSizeInBits();
238 MaskIn.zextOrTrunc(SrcBitWidth);
239 KnownZero.zextOrTrunc(SrcBitWidth);
240 KnownOne.zextOrTrunc(SrcBitWidth);
241 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
243 KnownZero.zextOrTrunc(BitWidth);
244 KnownOne.zextOrTrunc(BitWidth);
245 // Any top bits are known to be zero.
246 if (BitWidth > SrcBitWidth)
247 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
250 case Instruction::BitCast: {
251 const Type *SrcTy = I->getOperand(0)->getType();
252 if (SrcTy->isInteger() || isa<PointerType>(SrcTy)) {
253 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
259 case Instruction::SExt: {
260 // Compute the bits in the result that are not present in the input.
261 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
262 unsigned SrcBitWidth = SrcTy->getBitWidth();
265 MaskIn.trunc(SrcBitWidth);
266 KnownZero.trunc(SrcBitWidth);
267 KnownOne.trunc(SrcBitWidth);
268 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
270 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
271 KnownZero.zext(BitWidth);
272 KnownOne.zext(BitWidth);
274 // If the sign bit of the input is known set or clear, then we know the
275 // top bits of the result.
276 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
277 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
278 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
279 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
282 case Instruction::Shl:
283 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
284 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
285 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
286 APInt Mask2(Mask.lshr(ShiftAmt));
287 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
289 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
290 KnownZero <<= ShiftAmt;
291 KnownOne <<= ShiftAmt;
292 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
296 case Instruction::LShr:
297 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
298 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
299 // Compute the new bits that are at the top now.
300 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
302 // Unsigned shift right.
303 APInt Mask2(Mask.shl(ShiftAmt));
304 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
306 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
307 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
308 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
309 // high bits known zero.
310 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
314 case Instruction::AShr:
315 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
316 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
317 // Compute the new bits that are at the top now.
318 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
320 // Signed shift right.
321 APInt Mask2(Mask.shl(ShiftAmt));
322 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
324 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
325 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
326 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
328 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
329 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
330 KnownZero |= HighBits;
331 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
332 KnownOne |= HighBits;
336 case Instruction::Sub: {
337 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
338 // We know that the top bits of C-X are clear if X contains less bits
339 // than C (i.e. no wrap-around can happen). For example, 20-X is
340 // positive if we can prove that X is >= 0 and < 16.
341 if (!CLHS->getValue().isNegative()) {
342 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
343 // NLZ can't be BitWidth with no sign bit
344 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
345 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
348 // If all of the MaskV bits are known to be zero, then we know the
349 // output top bits are zero, because we now know that the output is
351 if ((KnownZero2 & MaskV) == MaskV) {
352 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
353 // Top bits known zero.
354 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
360 case Instruction::Add: {
361 // If one of the operands has trailing zeros, than the bits that the
362 // other operand has in those bit positions will be preserved in the
363 // result. For an add, this works with either operand. For a subtract,
364 // this only works if the known zeros are in the right operand.
365 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
366 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
367 BitWidth - Mask.countLeadingZeros());
368 ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
370 assert((LHSKnownZero & LHSKnownOne) == 0 &&
371 "Bits known to be one AND zero?");
372 unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
374 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD,
376 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
377 unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
379 // Determine which operand has more trailing zeros, and use that
380 // many bits from the other operand.
381 if (LHSKnownZeroOut > RHSKnownZeroOut) {
382 if (getOpcode(I) == Instruction::Add) {
383 APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
384 KnownZero |= KnownZero2 & Mask;
385 KnownOne |= KnownOne2 & Mask;
387 // If the known zeros are in the left operand for a subtract,
388 // fall back to the minimum known zeros in both operands.
389 KnownZero |= APInt::getLowBitsSet(BitWidth,
390 std::min(LHSKnownZeroOut,
393 } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
394 APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
395 KnownZero |= LHSKnownZero & Mask;
396 KnownOne |= LHSKnownOne & Mask;
400 case Instruction::SRem:
401 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
402 APInt RA = Rem->getValue();
403 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
404 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
405 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
406 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
409 // If the sign bit of the first operand is zero, the sign bit of
410 // the result is zero. If the first operand has no one bits below
411 // the second operand's single 1 bit, its sign will be zero.
412 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
413 KnownZero2 |= ~LowBits;
415 KnownZero |= KnownZero2 & Mask;
417 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
421 case Instruction::URem: {
422 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
423 APInt RA = Rem->getValue();
424 if (RA.isPowerOf2()) {
425 APInt LowBits = (RA - 1);
426 APInt Mask2 = LowBits & Mask;
427 KnownZero |= ~LowBits & Mask;
428 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
430 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
435 // Since the result is less than or equal to either operand, any leading
436 // zero bits in either operand must also exist in the result.
437 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
438 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
440 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
443 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
444 KnownZero2.countLeadingOnes());
446 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
450 case Instruction::Alloca:
451 case Instruction::Malloc: {
452 AllocationInst *AI = cast<AllocationInst>(V);
453 unsigned Align = AI->getAlignment();
454 if (Align == 0 && TD) {
455 if (isa<AllocaInst>(AI))
456 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
457 else if (isa<MallocInst>(AI)) {
458 // Malloc returns maximally aligned memory.
459 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
462 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
465 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
470 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
471 CountTrailingZeros_32(Align));
474 case Instruction::GetElementPtr: {
475 // Analyze all of the subscripts of this getelementptr instruction
476 // to determine if we can prove known low zero bits.
477 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
478 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
479 ComputeMaskedBits(I->getOperand(0), LocalMask,
480 LocalKnownZero, LocalKnownOne, TD, Depth+1);
481 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
483 gep_type_iterator GTI = gep_type_begin(I);
484 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
485 Value *Index = I->getOperand(i);
486 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
487 // Handle struct member offset arithmetic.
489 const StructLayout *SL = TD->getStructLayout(STy);
490 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
491 uint64_t Offset = SL->getElementOffset(Idx);
492 TrailZ = std::min(TrailZ,
493 CountTrailingZeros_64(Offset));
495 // Handle array index arithmetic.
496 const Type *IndexedTy = GTI.getIndexedType();
497 if (!IndexedTy->isSized()) return;
498 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
499 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
500 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
501 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
502 ComputeMaskedBits(Index, LocalMask,
503 LocalKnownZero, LocalKnownOne, TD, Depth+1);
504 TrailZ = std::min(TrailZ,
505 unsigned(CountTrailingZeros_64(TypeSize) +
506 LocalKnownZero.countTrailingOnes()));
510 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
513 case Instruction::PHI: {
514 PHINode *P = cast<PHINode>(I);
515 // Handle the case of a simple two-predecessor recurrence PHI.
516 // There's a lot more that could theoretically be done here, but
517 // this is sufficient to catch some interesting cases.
518 if (P->getNumIncomingValues() == 2) {
519 for (unsigned i = 0; i != 2; ++i) {
520 Value *L = P->getIncomingValue(i);
521 Value *R = P->getIncomingValue(!i);
522 User *LU = dyn_cast<User>(L);
525 unsigned Opcode = getOpcode(LU);
526 // Check for operations that have the property that if
527 // both their operands have low zero bits, the result
528 // will have low zero bits.
529 if (Opcode == Instruction::Add ||
530 Opcode == Instruction::Sub ||
531 Opcode == Instruction::And ||
532 Opcode == Instruction::Or ||
533 Opcode == Instruction::Mul) {
534 Value *LL = LU->getOperand(0);
535 Value *LR = LU->getOperand(1);
536 // Find a recurrence.
543 // Ok, we have a PHI of the form L op= R. Check for low
545 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
546 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
547 Mask2 = APInt::getLowBitsSet(BitWidth,
548 KnownZero2.countTrailingOnes());
550 // We need to take the minimum number of known bits
551 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
552 ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
555 APInt::getLowBitsSet(BitWidth,
556 std::min(KnownZero2.countTrailingOnes(),
557 KnownZero3.countTrailingOnes()));
563 // Otherwise take the unions of the known bit sets of the operands,
564 // taking conservative care to avoid excessive recursion.
565 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
566 KnownZero = APInt::getAllOnesValue(BitWidth);
567 KnownOne = APInt::getAllOnesValue(BitWidth);
568 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
569 // Skip direct self references.
570 if (P->getIncomingValue(i) == P) continue;
572 KnownZero2 = APInt(BitWidth, 0);
573 KnownOne2 = APInt(BitWidth, 0);
574 // Recurse, but cap the recursion to one level, because we don't
575 // want to waste time spinning around in loops.
576 ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
577 KnownZero2, KnownOne2, TD, MaxDepth-1);
578 KnownZero &= KnownZero2;
579 KnownOne &= KnownOne2;
580 // If all bits have been ruled out, there's no need to check
582 if (!KnownZero && !KnownOne)
588 case Instruction::Call:
589 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
590 switch (II->getIntrinsicID()) {
592 case Intrinsic::ctpop:
593 case Intrinsic::ctlz:
594 case Intrinsic::cttz: {
595 unsigned LowBits = Log2_32(BitWidth)+1;
596 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
605 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
606 /// this predicate to simplify operations downstream. Mask is known to be zero
607 /// for bits that V cannot have.
608 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
609 TargetData *TD, unsigned Depth) {
610 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
611 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
612 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
613 return (KnownZero & Mask) == Mask;
618 /// ComputeNumSignBits - Return the number of times the sign bit of the
619 /// register is replicated into the other bits. We know that at least 1 bit
620 /// is always equal to the sign bit (itself), but other cases can give us
621 /// information. For example, immediately after an "ashr X, 2", we know that
622 /// the top 3 bits are all equal to each other, so we return 3.
624 /// 'Op' must have a scalar integer type.
626 unsigned llvm::ComputeNumSignBits(Value *V, TargetData *TD, unsigned Depth) {
627 assert((TD || V->getType()->isIntOrIntVector()) &&
628 "ComputeNumSignBits requires a TargetData object to operate "
629 "on non-integer values!");
630 const Type *Ty = V->getType();
631 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
632 Ty->getScalarSizeInBits();
634 unsigned FirstAnswer = 1;
636 // Note that ConstantInt is handled by the general ComputeMaskedBits case
640 return 1; // Limit search depth.
642 User *U = dyn_cast<User>(V);
643 switch (getOpcode(V)) {
645 case Instruction::SExt:
646 Tmp = TyBits-cast<IntegerType>(U->getOperand(0)->getType())->getBitWidth();
647 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
649 case Instruction::AShr:
650 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
651 // ashr X, C -> adds C sign bits.
652 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
653 Tmp += C->getZExtValue();
654 if (Tmp > TyBits) Tmp = TyBits;
657 case Instruction::Shl:
658 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
659 // shl destroys sign bits.
660 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
661 if (C->getZExtValue() >= TyBits || // Bad shift.
662 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
663 return Tmp - C->getZExtValue();
666 case Instruction::And:
667 case Instruction::Or:
668 case Instruction::Xor: // NOT is handled here.
669 // Logical binary ops preserve the number of sign bits at the worst.
670 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
672 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
673 FirstAnswer = std::min(Tmp, Tmp2);
674 // We computed what we know about the sign bits as our first
675 // answer. Now proceed to the generic code that uses
676 // ComputeMaskedBits, and pick whichever answer is better.
680 case Instruction::Select:
681 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
682 if (Tmp == 1) return 1; // Early out.
683 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
684 return std::min(Tmp, Tmp2);
686 case Instruction::Add:
687 // Add can have at most one carry bit. Thus we know that the output
688 // is, at worst, one more bit than the inputs.
689 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
690 if (Tmp == 1) return 1; // Early out.
692 // Special case decrementing a value (ADD X, -1):
693 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
694 if (CRHS->isAllOnesValue()) {
695 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
696 APInt Mask = APInt::getAllOnesValue(TyBits);
697 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
700 // If the input is known to be 0 or 1, the output is 0/-1, which is all
702 if ((KnownZero | APInt(TyBits, 1)) == Mask)
705 // If we are subtracting one from a positive number, there is no carry
706 // out of the result.
707 if (KnownZero.isNegative())
711 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
712 if (Tmp2 == 1) return 1;
713 return std::min(Tmp, Tmp2)-1;
716 case Instruction::Sub:
717 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
718 if (Tmp2 == 1) return 1;
721 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
722 if (CLHS->isNullValue()) {
723 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
724 APInt Mask = APInt::getAllOnesValue(TyBits);
725 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
727 // If the input is known to be 0 or 1, the output is 0/-1, which is all
729 if ((KnownZero | APInt(TyBits, 1)) == Mask)
732 // If the input is known to be positive (the sign bit is known clear),
733 // the output of the NEG has the same number of sign bits as the input.
734 if (KnownZero.isNegative())
737 // Otherwise, we treat this like a SUB.
740 // Sub can have at most one carry bit. Thus we know that the output
741 // is, at worst, one more bit than the inputs.
742 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
743 if (Tmp == 1) return 1; // Early out.
744 return std::min(Tmp, Tmp2)-1;
746 case Instruction::Trunc:
747 // FIXME: it's tricky to do anything useful for this, but it is an important
748 // case for targets like X86.
752 // Finally, if we can prove that the top bits of the result are 0's or 1's,
753 // use this information.
754 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
755 APInt Mask = APInt::getAllOnesValue(TyBits);
756 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
758 if (KnownZero.isNegative()) { // sign bit is 0
760 } else if (KnownOne.isNegative()) { // sign bit is 1;
767 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
768 // the number of identical bits in the top of the input value.
770 Mask <<= Mask.getBitWidth()-TyBits;
771 // Return # leading zeros. We use 'min' here in case Val was zero before
772 // shifting. We don't want to return '64' as for an i32 "0".
773 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
776 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
777 /// value is never equal to -0.0.
779 /// NOTE: this function will need to be revisited when we support non-default
782 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
783 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
784 return !CFP->getValueAPF().isNegZero();
787 return 1; // Limit search depth.
789 const Instruction *I = dyn_cast<Instruction>(V);
790 if (I == 0) return false;
792 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
793 if (I->getOpcode() == Instruction::FAdd &&
794 isa<ConstantFP>(I->getOperand(1)) &&
795 cast<ConstantFP>(I->getOperand(1))->isNullValue())
798 // sitofp and uitofp turn into +0.0 for zero.
799 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
802 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
803 // sqrt(-0.0) = -0.0, no other negative results are possible.
804 if (II->getIntrinsicID() == Intrinsic::sqrt)
805 return CannotBeNegativeZero(II->getOperand(1), Depth+1);
807 if (const CallInst *CI = dyn_cast<CallInst>(I))
808 if (const Function *F = CI->getCalledFunction()) {
809 if (F->isDeclaration()) {
810 switch (F->getNameLen()) {
811 case 3: // abs(x) != -0.0
812 if (!strcmp(F->getNameStart(), "abs")) return true;
814 case 4: // abs[lf](x) != -0.0
815 if (!strcmp(F->getNameStart(), "absf")) return true;
816 if (!strcmp(F->getNameStart(), "absl")) return true;
825 // This is the recursive version of BuildSubAggregate. It takes a few different
826 // arguments. Idxs is the index within the nested struct From that we are
827 // looking at now (which is of type IndexedType). IdxSkip is the number of
828 // indices from Idxs that should be left out when inserting into the resulting
829 // struct. To is the result struct built so far, new insertvalue instructions
831 Value *BuildSubAggregate(Value *From, Value* To, const Type *IndexedType,
832 SmallVector<unsigned, 10> &Idxs,
834 Instruction *InsertBefore) {
835 const llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
837 // Save the original To argument so we can modify it
839 // General case, the type indexed by Idxs is a struct
840 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
841 // Process each struct element recursively
844 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
848 // Couldn't find any inserted value for this index? Cleanup
849 while (PrevTo != OrigTo) {
850 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
851 PrevTo = Del->getAggregateOperand();
852 Del->eraseFromParent();
854 // Stop processing elements
858 // If we succesfully found a value for each of our subaggregates
862 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
863 // the struct's elements had a value that was inserted directly. In the latter
864 // case, perhaps we can't determine each of the subelements individually, but
865 // we might be able to find the complete struct somewhere.
867 // Find the value that is at that particular spot
868 Value *V = FindInsertedValue(From, Idxs.begin(), Idxs.end());
873 // Insert the value in the new (sub) aggregrate
874 return llvm::InsertValueInst::Create(To, V, Idxs.begin() + IdxSkip,
875 Idxs.end(), "tmp", InsertBefore);
878 // This helper takes a nested struct and extracts a part of it (which is again a
879 // struct) into a new value. For example, given the struct:
880 // { a, { b, { c, d }, e } }
881 // and the indices "1, 1" this returns
884 // It does this by inserting an insertvalue for each element in the resulting
885 // struct, as opposed to just inserting a single struct. This will only work if
886 // each of the elements of the substruct are known (ie, inserted into From by an
887 // insertvalue instruction somewhere).
889 // All inserted insertvalue instructions are inserted before InsertBefore
890 Value *BuildSubAggregate(Value *From, const unsigned *idx_begin,
891 const unsigned *idx_end, Instruction *InsertBefore) {
892 assert(InsertBefore && "Must have someplace to insert!");
893 const Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
896 Value *To = UndefValue::get(IndexedType);
897 SmallVector<unsigned, 10> Idxs(idx_begin, idx_end);
898 unsigned IdxSkip = Idxs.size();
900 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
903 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
904 /// the scalar value indexed is already around as a register, for example if it
905 /// were inserted directly into the aggregrate.
907 /// If InsertBefore is not null, this function will duplicate (modified)
908 /// insertvalues when a part of a nested struct is extracted.
909 Value *llvm::FindInsertedValue(Value *V, const unsigned *idx_begin,
910 const unsigned *idx_end, Instruction *InsertBefore) {
911 // Nothing to index? Just return V then (this is useful at the end of our
913 if (idx_begin == idx_end)
915 // We have indices, so V should have an indexable type
916 assert((isa<StructType>(V->getType()) || isa<ArrayType>(V->getType()))
917 && "Not looking at a struct or array?");
918 assert(ExtractValueInst::getIndexedType(V->getType(), idx_begin, idx_end)
919 && "Invalid indices for type?");
920 const CompositeType *PTy = cast<CompositeType>(V->getType());
922 if (isa<UndefValue>(V))
923 return UndefValue::get(ExtractValueInst::getIndexedType(PTy,
926 else if (isa<ConstantAggregateZero>(V))
927 return Constant::getNullValue(ExtractValueInst::getIndexedType(PTy,
930 else if (Constant *C = dyn_cast<Constant>(V)) {
931 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C))
932 // Recursively process this constant
933 return FindInsertedValue(C->getOperand(*idx_begin), idx_begin + 1, idx_end,
935 } else if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
936 // Loop the indices for the insertvalue instruction in parallel with the
938 const unsigned *req_idx = idx_begin;
939 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
940 i != e; ++i, ++req_idx) {
941 if (req_idx == idx_end) {
943 // The requested index identifies a part of a nested aggregate. Handle
944 // this specially. For example,
945 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
946 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
947 // %C = extractvalue {i32, { i32, i32 } } %B, 1
948 // This can be changed into
949 // %A = insertvalue {i32, i32 } undef, i32 10, 0
950 // %C = insertvalue {i32, i32 } %A, i32 11, 1
951 // which allows the unused 0,0 element from the nested struct to be
953 return BuildSubAggregate(V, idx_begin, req_idx, InsertBefore);
955 // We can't handle this without inserting insertvalues
959 // This insert value inserts something else than what we are looking for.
960 // See if the (aggregrate) value inserted into has the value we are
961 // looking for, then.
963 return FindInsertedValue(I->getAggregateOperand(), idx_begin, idx_end,
966 // If we end up here, the indices of the insertvalue match with those
967 // requested (though possibly only partially). Now we recursively look at
968 // the inserted value, passing any remaining indices.
969 return FindInsertedValue(I->getInsertedValueOperand(), req_idx, idx_end,
971 } else if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
972 // If we're extracting a value from an aggregrate that was extracted from
973 // something else, we can extract from that something else directly instead.
974 // However, we will need to chain I's indices with the requested indices.
976 // Calculate the number of indices required
977 unsigned size = I->getNumIndices() + (idx_end - idx_begin);
978 // Allocate some space to put the new indices in
979 SmallVector<unsigned, 5> Idxs;
981 // Add indices from the extract value instruction
982 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
986 // Add requested indices
987 for (const unsigned *i = idx_begin, *e = idx_end; i != e; ++i)
990 assert(Idxs.size() == size
991 && "Number of indices added not correct?");
993 return FindInsertedValue(I->getAggregateOperand(), Idxs.begin(), Idxs.end(),
996 // Otherwise, we don't know (such as, extracting from a function return value
997 // or load instruction)
1001 /// GetConstantStringInfo - This function computes the length of a
1002 /// null-terminated C string pointed to by V. If successful, it returns true
1003 /// and returns the string in Str. If unsuccessful, it returns false.
1004 bool llvm::GetConstantStringInfo(Value *V, std::string &Str, uint64_t Offset,
1006 // If V is NULL then return false;
1007 if (V == NULL) return false;
1009 // Look through bitcast instructions.
1010 if (BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1011 return GetConstantStringInfo(BCI->getOperand(0), Str, Offset, StopAtNul);
1013 // If the value is not a GEP instruction nor a constant expression with a
1014 // GEP instruction, then return false because ConstantArray can't occur
1017 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1019 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1020 if (CE->getOpcode() == Instruction::BitCast)
1021 return GetConstantStringInfo(CE->getOperand(0), Str, Offset, StopAtNul);
1022 if (CE->getOpcode() != Instruction::GetElementPtr)
1028 // Make sure the GEP has exactly three arguments.
1029 if (GEP->getNumOperands() != 3)
1032 // Make sure the index-ee is a pointer to array of i8.
1033 const PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1034 const ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1035 if (AT == 0 || AT->getElementType() != Type::Int8Ty)
1038 // Check to make sure that the first operand of the GEP is an integer and
1039 // has value 0 so that we are sure we're indexing into the initializer.
1040 ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1041 if (FirstIdx == 0 || !FirstIdx->isZero())
1044 // If the second index isn't a ConstantInt, then this is a variable index
1045 // into the array. If this occurs, we can't say anything meaningful about
1047 uint64_t StartIdx = 0;
1048 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1049 StartIdx = CI->getZExtValue();
1052 return GetConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset,
1056 // The GEP instruction, constant or instruction, must reference a global
1057 // variable that is a constant and is initialized. The referenced constant
1058 // initializer is the array that we'll use for optimization.
1059 GlobalVariable* GV = dyn_cast<GlobalVariable>(V);
1060 if (!GV || !GV->isConstant() || !GV->hasInitializer())
1062 Constant *GlobalInit = GV->getInitializer();
1064 // Handle the ConstantAggregateZero case
1065 if (isa<ConstantAggregateZero>(GlobalInit)) {
1066 // This is a degenerate case. The initializer is constant zero so the
1067 // length of the string must be zero.
1072 // Must be a Constant Array
1073 ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1074 if (Array == 0 || Array->getType()->getElementType() != Type::Int8Ty)
1077 // Get the number of elements in the array
1078 uint64_t NumElts = Array->getType()->getNumElements();
1080 if (Offset > NumElts)
1083 // Traverse the constant array from 'Offset' which is the place the GEP refers
1085 Str.reserve(NumElts-Offset);
1086 for (unsigned i = Offset; i != NumElts; ++i) {
1087 Constant *Elt = Array->getOperand(i);
1088 ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1089 if (!CI) // This array isn't suitable, non-int initializer.
1091 if (StopAtNul && CI->isZero())
1092 return true; // we found end of string, success!
1093 Str += (char)CI->getZExtValue();
1096 // The array isn't null terminated, but maybe this is a memcpy, not a strcpy.