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?");
464 case Instruction::URem: {
465 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
466 APInt RA = Rem->getValue();
467 if (RA.isPowerOf2()) {
468 APInt LowBits = (RA - 1);
469 APInt Mask2 = LowBits & Mask;
470 KnownZero |= ~LowBits & Mask;
471 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
473 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
478 // Since the result is less than or equal to either operand, any leading
479 // zero bits in either operand must also exist in the result.
480 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
481 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
483 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
486 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
487 KnownZero2.countLeadingOnes());
488 KnownOne.clearAllBits();
489 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
493 case Instruction::Alloca: {
494 AllocaInst *AI = cast<AllocaInst>(V);
495 unsigned Align = AI->getAlignment();
496 if (Align == 0 && TD)
497 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
500 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
501 CountTrailingZeros_32(Align));
504 case Instruction::GetElementPtr: {
505 // Analyze all of the subscripts of this getelementptr instruction
506 // to determine if we can prove known low zero bits.
507 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
508 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
509 ComputeMaskedBits(I->getOperand(0), LocalMask,
510 LocalKnownZero, LocalKnownOne, TD, Depth+1);
511 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
513 gep_type_iterator GTI = gep_type_begin(I);
514 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
515 Value *Index = I->getOperand(i);
516 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
517 // Handle struct member offset arithmetic.
519 const StructLayout *SL = TD->getStructLayout(STy);
520 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
521 uint64_t Offset = SL->getElementOffset(Idx);
522 TrailZ = std::min(TrailZ,
523 CountTrailingZeros_64(Offset));
525 // Handle array index arithmetic.
526 const Type *IndexedTy = GTI.getIndexedType();
527 if (!IndexedTy->isSized()) return;
528 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
529 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
530 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
531 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
532 ComputeMaskedBits(Index, LocalMask,
533 LocalKnownZero, LocalKnownOne, TD, Depth+1);
534 TrailZ = std::min(TrailZ,
535 unsigned(CountTrailingZeros_64(TypeSize) +
536 LocalKnownZero.countTrailingOnes()));
540 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
543 case Instruction::PHI: {
544 PHINode *P = cast<PHINode>(I);
545 // Handle the case of a simple two-predecessor recurrence PHI.
546 // There's a lot more that could theoretically be done here, but
547 // this is sufficient to catch some interesting cases.
548 if (P->getNumIncomingValues() == 2) {
549 for (unsigned i = 0; i != 2; ++i) {
550 Value *L = P->getIncomingValue(i);
551 Value *R = P->getIncomingValue(!i);
552 Operator *LU = dyn_cast<Operator>(L);
555 unsigned Opcode = LU->getOpcode();
556 // Check for operations that have the property that if
557 // both their operands have low zero bits, the result
558 // will have low zero bits.
559 if (Opcode == Instruction::Add ||
560 Opcode == Instruction::Sub ||
561 Opcode == Instruction::And ||
562 Opcode == Instruction::Or ||
563 Opcode == Instruction::Mul) {
564 Value *LL = LU->getOperand(0);
565 Value *LR = LU->getOperand(1);
566 // Find a recurrence.
573 // Ok, we have a PHI of the form L op= R. Check for low
575 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
576 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
577 Mask2 = APInt::getLowBitsSet(BitWidth,
578 KnownZero2.countTrailingOnes());
580 // We need to take the minimum number of known bits
581 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
582 ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
585 APInt::getLowBitsSet(BitWidth,
586 std::min(KnownZero2.countTrailingOnes(),
587 KnownZero3.countTrailingOnes()));
593 // Otherwise take the unions of the known bit sets of the operands,
594 // taking conservative care to avoid excessive recursion.
595 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
596 KnownZero = APInt::getAllOnesValue(BitWidth);
597 KnownOne = APInt::getAllOnesValue(BitWidth);
598 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
599 // Skip direct self references.
600 if (P->getIncomingValue(i) == P) continue;
602 KnownZero2 = APInt(BitWidth, 0);
603 KnownOne2 = APInt(BitWidth, 0);
604 // Recurse, but cap the recursion to one level, because we don't
605 // want to waste time spinning around in loops.
606 ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
607 KnownZero2, KnownOne2, TD, MaxDepth-1);
608 KnownZero &= KnownZero2;
609 KnownOne &= KnownOne2;
610 // If all bits have been ruled out, there's no need to check
612 if (!KnownZero && !KnownOne)
618 case Instruction::Call:
619 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
620 switch (II->getIntrinsicID()) {
622 case Intrinsic::ctpop:
623 case Intrinsic::ctlz:
624 case Intrinsic::cttz: {
625 unsigned LowBits = Log2_32(BitWidth)+1;
626 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
635 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
636 /// one. Convenience wrapper around ComputeMaskedBits.
637 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
638 const TargetData *TD, unsigned Depth) {
639 unsigned BitWidth = getBitWidth(V->getType(), TD);
645 APInt ZeroBits(BitWidth, 0);
646 APInt OneBits(BitWidth, 0);
647 ComputeMaskedBits(V, APInt::getSignBit(BitWidth), ZeroBits, OneBits, TD,
649 KnownOne = OneBits[BitWidth - 1];
650 KnownZero = ZeroBits[BitWidth - 1];
653 /// isPowerOfTwo - Return true if the given value is known to have exactly one
654 /// bit set when defined. For vectors return true if every element is known to
655 /// be a power of two when defined. Supports values with integer or pointer
656 /// types and vectors of integers.
657 bool llvm::isPowerOfTwo(Value *V, const TargetData *TD, unsigned Depth) {
658 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
659 return CI->getValue().isPowerOf2();
660 // TODO: Handle vector constants.
662 // 1 << X is clearly a power of two if the one is not shifted off the end. If
663 // it is shifted off the end then the result is undefined.
664 if (match(V, m_Shl(m_One(), m_Value())))
667 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
668 // bottom. If it is shifted off the bottom then the result is undefined.
669 if (match(V, m_LShr(m_SignBit(), m_Value())))
672 // The remaining tests are all recursive, so bail out if we hit the limit.
673 if (Depth++ == MaxDepth)
676 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
677 return isPowerOfTwo(ZI->getOperand(0), TD, Depth);
679 if (SelectInst *SI = dyn_cast<SelectInst>(V))
680 return isPowerOfTwo(SI->getTrueValue(), TD, Depth) &&
681 isPowerOfTwo(SI->getFalseValue(), TD, Depth);
686 /// isKnownNonZero - Return true if the given value is known to be non-zero
687 /// when defined. For vectors return true if every element is known to be
688 /// non-zero when defined. Supports values with integer or pointer type and
689 /// vectors of integers.
690 bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
691 if (Constant *C = dyn_cast<Constant>(V)) {
692 if (C->isNullValue())
694 if (isa<ConstantInt>(C))
695 // Must be non-zero due to null test above.
697 // TODO: Handle vectors
701 // The remaining tests are all recursive, so bail out if we hit the limit.
702 if (Depth++ == MaxDepth)
705 unsigned BitWidth = getBitWidth(V->getType(), TD);
707 // X | Y != 0 if X != 0 or Y != 0.
708 Value *X = 0, *Y = 0;
709 if (match(V, m_Or(m_Value(X), m_Value(Y))))
710 return isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth);
712 // ext X != 0 if X != 0.
713 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
714 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth);
716 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
717 // if the lowest bit is shifted off the end.
718 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
719 APInt KnownZero(BitWidth, 0);
720 APInt KnownOne(BitWidth, 0);
721 ComputeMaskedBits(X, APInt(BitWidth, 1), KnownZero, KnownOne, TD, Depth);
725 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
726 // defined if the sign bit is shifted off the end.
727 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
728 bool XKnownNonNegative, XKnownNegative;
729 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
734 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
735 bool XKnownNonNegative, XKnownNegative;
736 bool YKnownNonNegative, YKnownNegative;
737 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth);
738 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth);
740 // If X and Y are both non-negative (as signed values) then their sum is not
741 // zero unless both X and Y are zero.
742 if (XKnownNonNegative && YKnownNonNegative)
743 if (isKnownNonZero(X, TD, Depth) || isKnownNonZero(Y, TD, Depth))
746 // If X and Y are both negative (as signed values) then their sum is not
747 // zero unless both X and Y equal INT_MIN.
748 if (BitWidth && XKnownNegative && YKnownNegative) {
749 APInt KnownZero(BitWidth, 0);
750 APInt KnownOne(BitWidth, 0);
751 APInt Mask = APInt::getSignedMaxValue(BitWidth);
752 // The sign bit of X is set. If some other bit is set then X is not equal
754 ComputeMaskedBits(X, Mask, KnownZero, KnownOne, TD, Depth);
755 if ((KnownOne & Mask) != 0)
757 // The sign bit of Y is set. If some other bit is set then Y is not equal
759 ComputeMaskedBits(Y, Mask, KnownZero, KnownOne, TD, Depth);
760 if ((KnownOne & Mask) != 0)
764 // The sum of a non-negative number and a power of two is not zero.
765 if (XKnownNonNegative && isPowerOfTwo(Y, TD, Depth))
767 if (YKnownNonNegative && isPowerOfTwo(X, TD, Depth))
770 // (C ? X : Y) != 0 if X != 0 and Y != 0.
771 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
772 if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&
773 isKnownNonZero(SI->getFalseValue(), TD, Depth))
777 if (!BitWidth) return false;
778 APInt KnownZero(BitWidth, 0);
779 APInt KnownOne(BitWidth, 0);
780 ComputeMaskedBits(V, APInt::getAllOnesValue(BitWidth), KnownZero, KnownOne,
782 return KnownOne != 0;
785 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
786 /// this predicate to simplify operations downstream. Mask is known to be zero
787 /// for bits that V cannot have.
789 /// This function is defined on values with integer type, values with pointer
790 /// type (but only if TD is non-null), and vectors of integers. In the case
791 /// where V is a vector, the mask, known zero, and known one values are the
792 /// same width as the vector element, and the bit is set only if it is true
793 /// for all of the elements in the vector.
794 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
795 const TargetData *TD, unsigned Depth) {
796 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
797 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
798 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
799 return (KnownZero & Mask) == Mask;
804 /// ComputeNumSignBits - Return the number of times the sign bit of the
805 /// register is replicated into the other bits. We know that at least 1 bit
806 /// is always equal to the sign bit (itself), but other cases can give us
807 /// information. For example, immediately after an "ashr X, 2", we know that
808 /// the top 3 bits are all equal to each other, so we return 3.
810 /// 'Op' must have a scalar integer type.
812 unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
814 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
815 "ComputeNumSignBits requires a TargetData object to operate "
816 "on non-integer values!");
817 const Type *Ty = V->getType();
818 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
819 Ty->getScalarSizeInBits();
821 unsigned FirstAnswer = 1;
823 // Note that ConstantInt is handled by the general ComputeMaskedBits case
827 return 1; // Limit search depth.
829 Operator *U = dyn_cast<Operator>(V);
830 switch (Operator::getOpcode(V)) {
832 case Instruction::SExt:
833 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
834 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
836 case Instruction::AShr:
837 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
838 // ashr X, C -> adds C sign bits.
839 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
840 Tmp += C->getZExtValue();
841 if (Tmp > TyBits) Tmp = TyBits;
843 // vector ashr X, <C, C, C, C> -> adds C sign bits
844 if (ConstantVector *C = dyn_cast<ConstantVector>(U->getOperand(1))) {
845 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) {
846 Tmp += CI->getZExtValue();
847 if (Tmp > TyBits) Tmp = TyBits;
851 case Instruction::Shl:
852 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
853 // shl destroys sign bits.
854 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
855 if (C->getZExtValue() >= TyBits || // Bad shift.
856 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
857 return Tmp - C->getZExtValue();
860 case Instruction::And:
861 case Instruction::Or:
862 case Instruction::Xor: // NOT is handled here.
863 // Logical binary ops preserve the number of sign bits at the worst.
864 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
866 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
867 FirstAnswer = std::min(Tmp, Tmp2);
868 // We computed what we know about the sign bits as our first
869 // answer. Now proceed to the generic code that uses
870 // ComputeMaskedBits, and pick whichever answer is better.
874 case Instruction::Select:
875 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
876 if (Tmp == 1) return 1; // Early out.
877 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
878 return std::min(Tmp, Tmp2);
880 case Instruction::Add:
881 // Add can have at most one carry bit. Thus we know that the output
882 // is, at worst, one more bit than the inputs.
883 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
884 if (Tmp == 1) return 1; // Early out.
886 // Special case decrementing a value (ADD X, -1):
887 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
888 if (CRHS->isAllOnesValue()) {
889 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
890 APInt Mask = APInt::getAllOnesValue(TyBits);
891 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
894 // If the input is known to be 0 or 1, the output is 0/-1, which is all
896 if ((KnownZero | APInt(TyBits, 1)) == Mask)
899 // If we are subtracting one from a positive number, there is no carry
900 // out of the result.
901 if (KnownZero.isNegative())
905 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
906 if (Tmp2 == 1) return 1;
907 return std::min(Tmp, Tmp2)-1;
909 case Instruction::Sub:
910 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
911 if (Tmp2 == 1) return 1;
914 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
915 if (CLHS->isNullValue()) {
916 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
917 APInt Mask = APInt::getAllOnesValue(TyBits);
918 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
920 // If the input is known to be 0 or 1, the output is 0/-1, which is all
922 if ((KnownZero | APInt(TyBits, 1)) == Mask)
925 // If the input is known to be positive (the sign bit is known clear),
926 // the output of the NEG has the same number of sign bits as the input.
927 if (KnownZero.isNegative())
930 // Otherwise, we treat this like a SUB.
933 // Sub can have at most one carry bit. Thus we know that the output
934 // is, at worst, one more bit than the inputs.
935 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
936 if (Tmp == 1) return 1; // Early out.
937 return std::min(Tmp, Tmp2)-1;
939 case Instruction::PHI: {
940 PHINode *PN = cast<PHINode>(U);
941 // Don't analyze large in-degree PHIs.
942 if (PN->getNumIncomingValues() > 4) break;
944 // Take the minimum of all incoming values. This can't infinitely loop
945 // because of our depth threshold.
946 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
947 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
948 if (Tmp == 1) return Tmp;
950 ComputeNumSignBits(PN->getIncomingValue(i), TD, Depth+1));
955 case Instruction::Trunc:
956 // FIXME: it's tricky to do anything useful for this, but it is an important
957 // case for targets like X86.
961 // Finally, if we can prove that the top bits of the result are 0's or 1's,
962 // use this information.
963 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
964 APInt Mask = APInt::getAllOnesValue(TyBits);
965 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
967 if (KnownZero.isNegative()) { // sign bit is 0
969 } else if (KnownOne.isNegative()) { // sign bit is 1;
976 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
977 // the number of identical bits in the top of the input value.
979 Mask <<= Mask.getBitWidth()-TyBits;
980 // Return # leading zeros. We use 'min' here in case Val was zero before
981 // shifting. We don't want to return '64' as for an i32 "0".
982 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
985 /// ComputeMultiple - This function computes the integer multiple of Base that
986 /// equals V. If successful, it returns true and returns the multiple in
987 /// Multiple. If unsuccessful, it returns false. It looks
988 /// through SExt instructions only if LookThroughSExt is true.
989 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
990 bool LookThroughSExt, unsigned Depth) {
991 const unsigned MaxDepth = 6;
993 assert(V && "No Value?");
994 assert(Depth <= MaxDepth && "Limit Search Depth");
995 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
997 const Type *T = V->getType();
999 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1009 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1010 Constant *BaseVal = ConstantInt::get(T, Base);
1011 if (CO && CO == BaseVal) {
1013 Multiple = ConstantInt::get(T, 1);
1017 if (CI && CI->getZExtValue() % Base == 0) {
1018 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1022 if (Depth == MaxDepth) return false; // Limit search depth.
1024 Operator *I = dyn_cast<Operator>(V);
1025 if (!I) return false;
1027 switch (I->getOpcode()) {
1029 case Instruction::SExt:
1030 if (!LookThroughSExt) return false;
1031 // otherwise fall through to ZExt
1032 case Instruction::ZExt:
1033 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1034 LookThroughSExt, Depth+1);
1035 case Instruction::Shl:
1036 case Instruction::Mul: {
1037 Value *Op0 = I->getOperand(0);
1038 Value *Op1 = I->getOperand(1);
1040 if (I->getOpcode() == Instruction::Shl) {
1041 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1042 if (!Op1CI) return false;
1043 // Turn Op0 << Op1 into Op0 * 2^Op1
1044 APInt Op1Int = Op1CI->getValue();
1045 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1046 APInt API(Op1Int.getBitWidth(), 0);
1047 API.setBit(BitToSet);
1048 Op1 = ConstantInt::get(V->getContext(), API);
1052 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1053 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1054 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1055 if (Op1C->getType()->getPrimitiveSizeInBits() <
1056 MulC->getType()->getPrimitiveSizeInBits())
1057 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1058 if (Op1C->getType()->getPrimitiveSizeInBits() >
1059 MulC->getType()->getPrimitiveSizeInBits())
1060 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1062 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1063 Multiple = ConstantExpr::getMul(MulC, Op1C);
1067 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1068 if (Mul0CI->getValue() == 1) {
1069 // V == Base * Op1, so return Op1
1076 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1077 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1078 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1079 if (Op0C->getType()->getPrimitiveSizeInBits() <
1080 MulC->getType()->getPrimitiveSizeInBits())
1081 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1082 if (Op0C->getType()->getPrimitiveSizeInBits() >
1083 MulC->getType()->getPrimitiveSizeInBits())
1084 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1086 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1087 Multiple = ConstantExpr::getMul(MulC, Op0C);
1091 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1092 if (Mul1CI->getValue() == 1) {
1093 // V == Base * Op0, so return Op0
1101 // We could not determine if V is a multiple of Base.
1105 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1106 /// value is never equal to -0.0.
1108 /// NOTE: this function will need to be revisited when we support non-default
1111 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1112 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1113 return !CFP->getValueAPF().isNegZero();
1116 return 1; // Limit search depth.
1118 const Operator *I = dyn_cast<Operator>(V);
1119 if (I == 0) return false;
1121 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1122 if (I->getOpcode() == Instruction::FAdd &&
1123 isa<ConstantFP>(I->getOperand(1)) &&
1124 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1127 // sitofp and uitofp turn into +0.0 for zero.
1128 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1131 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1132 // sqrt(-0.0) = -0.0, no other negative results are possible.
1133 if (II->getIntrinsicID() == Intrinsic::sqrt)
1134 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1136 if (const CallInst *CI = dyn_cast<CallInst>(I))
1137 if (const Function *F = CI->getCalledFunction()) {
1138 if (F->isDeclaration()) {
1140 if (F->getName() == "abs") return true;
1141 // fabs[lf](x) != -0.0
1142 if (F->getName() == "fabs") return true;
1143 if (F->getName() == "fabsf") return true;
1144 if (F->getName() == "fabsl") return true;
1145 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1146 F->getName() == "sqrtl")
1147 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1154 /// isBytewiseValue - If the specified value can be set by repeating the same
1155 /// byte in memory, return the i8 value that it is represented with. This is
1156 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1157 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1158 /// byte store (e.g. i16 0x1234), return null.
1159 Value *llvm::isBytewiseValue(Value *V) {
1160 // All byte-wide stores are splatable, even of arbitrary variables.
1161 if (V->getType()->isIntegerTy(8)) return V;
1163 // Constant float and double values can be handled as integer values if the
1164 // corresponding integer value is "byteable". An important case is 0.0.
1165 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1166 if (CFP->getType()->isFloatTy())
1167 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1168 if (CFP->getType()->isDoubleTy())
1169 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1170 // Don't handle long double formats, which have strange constraints.
1173 // We can handle constant integers that are power of two in size and a
1174 // multiple of 8 bits.
1175 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1176 unsigned Width = CI->getBitWidth();
1177 if (isPowerOf2_32(Width) && Width > 8) {
1178 // We can handle this value if the recursive binary decomposition is the
1179 // same at all levels.
1180 APInt Val = CI->getValue();
1182 while (Val.getBitWidth() != 8) {
1183 unsigned NextWidth = Val.getBitWidth()/2;
1184 Val2 = Val.lshr(NextWidth);
1185 Val2 = Val2.trunc(Val.getBitWidth()/2);
1186 Val = Val.trunc(Val.getBitWidth()/2);
1188 // If the top/bottom halves aren't the same, reject it.
1192 return ConstantInt::get(V->getContext(), Val);
1196 // A ConstantArray is splatable if all its members are equal and also
1198 if (ConstantArray *CA = dyn_cast<ConstantArray>(V)) {
1199 if (CA->getNumOperands() == 0)
1202 Value *Val = isBytewiseValue(CA->getOperand(0));
1206 for (unsigned I = 1, E = CA->getNumOperands(); I != E; ++I)
1207 if (CA->getOperand(I-1) != CA->getOperand(I))
1213 // Conceptually, we could handle things like:
1214 // %a = zext i8 %X to i16
1215 // %b = shl i16 %a, 8
1216 // %c = or i16 %a, %b
1217 // but until there is an example that actually needs this, it doesn't seem
1218 // worth worrying about.
1223 // This is the recursive version of BuildSubAggregate. It takes a few different
1224 // arguments. Idxs is the index within the nested struct From that we are
1225 // looking at now (which is of type IndexedType). IdxSkip is the number of
1226 // indices from Idxs that should be left out when inserting into the resulting
1227 // struct. To is the result struct built so far, new insertvalue instructions
1229 static Value *BuildSubAggregate(Value *From, Value* To, const Type *IndexedType,
1230 SmallVector<unsigned, 10> &Idxs,
1232 Instruction *InsertBefore) {
1233 const llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
1235 // Save the original To argument so we can modify it
1237 // General case, the type indexed by Idxs is a struct
1238 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1239 // Process each struct element recursively
1242 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1246 // Couldn't find any inserted value for this index? Cleanup
1247 while (PrevTo != OrigTo) {
1248 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1249 PrevTo = Del->getAggregateOperand();
1250 Del->eraseFromParent();
1252 // Stop processing elements
1256 // If we succesfully found a value for each of our subaggregates
1260 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1261 // the struct's elements had a value that was inserted directly. In the latter
1262 // case, perhaps we can't determine each of the subelements individually, but
1263 // we might be able to find the complete struct somewhere.
1265 // Find the value that is at that particular spot
1266 Value *V = FindInsertedValue(From, Idxs.begin(), Idxs.end());
1271 // Insert the value in the new (sub) aggregrate
1272 return llvm::InsertValueInst::Create(To, V, Idxs.begin() + IdxSkip,
1273 Idxs.end(), "tmp", InsertBefore);
1276 // This helper takes a nested struct and extracts a part of it (which is again a
1277 // struct) into a new value. For example, given the struct:
1278 // { a, { b, { c, d }, e } }
1279 // and the indices "1, 1" this returns
1282 // It does this by inserting an insertvalue for each element in the resulting
1283 // struct, as opposed to just inserting a single struct. This will only work if
1284 // each of the elements of the substruct are known (ie, inserted into From by an
1285 // insertvalue instruction somewhere).
1287 // All inserted insertvalue instructions are inserted before InsertBefore
1288 static Value *BuildSubAggregate(Value *From, const unsigned *idx_begin,
1289 const unsigned *idx_end,
1290 Instruction *InsertBefore) {
1291 assert(InsertBefore && "Must have someplace to insert!");
1292 const Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1295 Value *To = UndefValue::get(IndexedType);
1296 SmallVector<unsigned, 10> Idxs(idx_begin, idx_end);
1297 unsigned IdxSkip = Idxs.size();
1299 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1302 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1303 /// the scalar value indexed is already around as a register, for example if it
1304 /// were inserted directly into the aggregrate.
1306 /// If InsertBefore is not null, this function will duplicate (modified)
1307 /// insertvalues when a part of a nested struct is extracted.
1308 Value *llvm::FindInsertedValue(Value *V, const unsigned *idx_begin,
1309 const unsigned *idx_end, Instruction *InsertBefore) {
1310 // Nothing to index? Just return V then (this is useful at the end of our
1312 if (idx_begin == idx_end)
1314 // We have indices, so V should have an indexable type
1315 assert((V->getType()->isStructTy() || V->getType()->isArrayTy())
1316 && "Not looking at a struct or array?");
1317 assert(ExtractValueInst::getIndexedType(V->getType(), idx_begin, idx_end)
1318 && "Invalid indices for type?");
1319 const CompositeType *PTy = cast<CompositeType>(V->getType());
1321 if (isa<UndefValue>(V))
1322 return UndefValue::get(ExtractValueInst::getIndexedType(PTy,
1325 else if (isa<ConstantAggregateZero>(V))
1326 return Constant::getNullValue(ExtractValueInst::getIndexedType(PTy,
1329 else if (Constant *C = dyn_cast<Constant>(V)) {
1330 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C))
1331 // Recursively process this constant
1332 return FindInsertedValue(C->getOperand(*idx_begin), idx_begin + 1,
1333 idx_end, InsertBefore);
1334 } else if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1335 // Loop the indices for the insertvalue instruction in parallel with the
1336 // requested indices
1337 const unsigned *req_idx = idx_begin;
1338 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1339 i != e; ++i, ++req_idx) {
1340 if (req_idx == idx_end) {
1342 // The requested index identifies a part of a nested aggregate. Handle
1343 // this specially. For example,
1344 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1345 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1346 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1347 // This can be changed into
1348 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1349 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1350 // which allows the unused 0,0 element from the nested struct to be
1352 return BuildSubAggregate(V, idx_begin, req_idx, InsertBefore);
1354 // We can't handle this without inserting insertvalues
1358 // This insert value inserts something else than what we are looking for.
1359 // See if the (aggregrate) value inserted into has the value we are
1360 // looking for, then.
1362 return FindInsertedValue(I->getAggregateOperand(), idx_begin, idx_end,
1365 // If we end up here, the indices of the insertvalue match with those
1366 // requested (though possibly only partially). Now we recursively look at
1367 // the inserted value, passing any remaining indices.
1368 return FindInsertedValue(I->getInsertedValueOperand(), req_idx, idx_end,
1370 } else if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
1371 // If we're extracting a value from an aggregrate that was extracted from
1372 // something else, we can extract from that something else directly instead.
1373 // However, we will need to chain I's indices with the requested indices.
1375 // Calculate the number of indices required
1376 unsigned size = I->getNumIndices() + (idx_end - idx_begin);
1377 // Allocate some space to put the new indices in
1378 SmallVector<unsigned, 5> Idxs;
1380 // Add indices from the extract value instruction
1381 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1385 // Add requested indices
1386 for (const unsigned *i = idx_begin, *e = idx_end; i != e; ++i)
1389 assert(Idxs.size() == size
1390 && "Number of indices added not correct?");
1392 return FindInsertedValue(I->getAggregateOperand(), Idxs.begin(), Idxs.end(),
1395 // Otherwise, we don't know (such as, extracting from a function return value
1396 // or load instruction)
1400 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
1401 /// it can be expressed as a base pointer plus a constant offset. Return the
1402 /// base and offset to the caller.
1403 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
1404 const TargetData &TD) {
1405 Operator *PtrOp = dyn_cast<Operator>(Ptr);
1406 if (PtrOp == 0) return Ptr;
1408 // Just look through bitcasts.
1409 if (PtrOp->getOpcode() == Instruction::BitCast)
1410 return GetPointerBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
1412 // If this is a GEP with constant indices, we can look through it.
1413 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
1414 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
1416 gep_type_iterator GTI = gep_type_begin(GEP);
1417 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
1419 ConstantInt *OpC = cast<ConstantInt>(*I);
1420 if (OpC->isZero()) continue;
1422 // Handle a struct and array indices which add their offset to the pointer.
1423 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
1424 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1426 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
1427 Offset += OpC->getSExtValue()*Size;
1431 // Re-sign extend from the pointer size if needed to get overflow edge cases
1433 unsigned PtrSize = TD.getPointerSizeInBits();
1435 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1437 return GetPointerBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1441 /// GetConstantStringInfo - This function computes the length of a
1442 /// null-terminated C string pointed to by V. If successful, it returns true
1443 /// and returns the string in Str. If unsuccessful, it returns false.
1444 bool llvm::GetConstantStringInfo(const Value *V, std::string &Str,
1447 // If V is NULL then return false;
1448 if (V == NULL) return false;
1450 // Look through bitcast instructions.
1451 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1452 return GetConstantStringInfo(BCI->getOperand(0), Str, Offset, StopAtNul);
1454 // If the value is not a GEP instruction nor a constant expression with a
1455 // GEP instruction, then return false because ConstantArray can't occur
1457 const User *GEP = 0;
1458 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1460 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1461 if (CE->getOpcode() == Instruction::BitCast)
1462 return GetConstantStringInfo(CE->getOperand(0), Str, Offset, StopAtNul);
1463 if (CE->getOpcode() != Instruction::GetElementPtr)
1469 // Make sure the GEP has exactly three arguments.
1470 if (GEP->getNumOperands() != 3)
1473 // Make sure the index-ee is a pointer to array of i8.
1474 const PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
1475 const ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
1476 if (AT == 0 || !AT->getElementType()->isIntegerTy(8))
1479 // Check to make sure that the first operand of the GEP is an integer and
1480 // has value 0 so that we are sure we're indexing into the initializer.
1481 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
1482 if (FirstIdx == 0 || !FirstIdx->isZero())
1485 // If the second index isn't a ConstantInt, then this is a variable index
1486 // into the array. If this occurs, we can't say anything meaningful about
1488 uint64_t StartIdx = 0;
1489 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1490 StartIdx = CI->getZExtValue();
1493 return GetConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset,
1497 // The GEP instruction, constant or instruction, must reference a global
1498 // variable that is a constant and is initialized. The referenced constant
1499 // initializer is the array that we'll use for optimization.
1500 const GlobalVariable* GV = dyn_cast<GlobalVariable>(V);
1501 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
1503 const Constant *GlobalInit = GV->getInitializer();
1505 // Handle the ConstantAggregateZero case
1506 if (isa<ConstantAggregateZero>(GlobalInit)) {
1507 // This is a degenerate case. The initializer is constant zero so the
1508 // length of the string must be zero.
1513 // Must be a Constant Array
1514 const ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1515 if (Array == 0 || !Array->getType()->getElementType()->isIntegerTy(8))
1518 // Get the number of elements in the array
1519 uint64_t NumElts = Array->getType()->getNumElements();
1521 if (Offset > NumElts)
1524 // Traverse the constant array from 'Offset' which is the place the GEP refers
1526 Str.reserve(NumElts-Offset);
1527 for (unsigned i = Offset; i != NumElts; ++i) {
1528 const Constant *Elt = Array->getOperand(i);
1529 const ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1530 if (!CI) // This array isn't suitable, non-int initializer.
1532 if (StopAtNul && CI->isZero())
1533 return true; // we found end of string, success!
1534 Str += (char)CI->getZExtValue();
1537 // The array isn't null terminated, but maybe this is a memcpy, not a strcpy.
1541 // These next two are very similar to the above, but also look through PHI
1543 // TODO: See if we can integrate these two together.
1545 /// GetStringLengthH - If we can compute the length of the string pointed to by
1546 /// the specified pointer, return 'len+1'. If we can't, return 0.
1547 static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) {
1548 // Look through noop bitcast instructions.
1549 if (BitCastInst *BCI = dyn_cast<BitCastInst>(V))
1550 return GetStringLengthH(BCI->getOperand(0), PHIs);
1552 // If this is a PHI node, there are two cases: either we have already seen it
1554 if (PHINode *PN = dyn_cast<PHINode>(V)) {
1555 if (!PHIs.insert(PN))
1556 return ~0ULL; // already in the set.
1558 // If it was new, see if all the input strings are the same length.
1559 uint64_t LenSoFar = ~0ULL;
1560 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1561 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
1562 if (Len == 0) return 0; // Unknown length -> unknown.
1564 if (Len == ~0ULL) continue;
1566 if (Len != LenSoFar && LenSoFar != ~0ULL)
1567 return 0; // Disagree -> unknown.
1571 // Success, all agree.
1575 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
1576 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1577 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
1578 if (Len1 == 0) return 0;
1579 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
1580 if (Len2 == 0) return 0;
1581 if (Len1 == ~0ULL) return Len2;
1582 if (Len2 == ~0ULL) return Len1;
1583 if (Len1 != Len2) return 0;
1587 // If the value is not a GEP instruction nor a constant expression with a
1588 // GEP instruction, then return unknown.
1590 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
1592 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
1593 if (CE->getOpcode() != Instruction::GetElementPtr)
1600 // Make sure the GEP has exactly three arguments.
1601 if (GEP->getNumOperands() != 3)
1604 // Check to make sure that the first operand of the GEP is an integer and
1605 // has value 0 so that we are sure we're indexing into the initializer.
1606 if (ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(1))) {
1612 // If the second index isn't a ConstantInt, then this is a variable index
1613 // into the array. If this occurs, we can't say anything meaningful about
1615 uint64_t StartIdx = 0;
1616 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1617 StartIdx = CI->getZExtValue();
1621 // The GEP instruction, constant or instruction, must reference a global
1622 // variable that is a constant and is initialized. The referenced constant
1623 // initializer is the array that we'll use for optimization.
1624 GlobalVariable* GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
1625 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
1626 GV->mayBeOverridden())
1628 Constant *GlobalInit = GV->getInitializer();
1630 // Handle the ConstantAggregateZero case, which is a degenerate case. The
1631 // initializer is constant zero so the length of the string must be zero.
1632 if (isa<ConstantAggregateZero>(GlobalInit))
1633 return 1; // Len = 0 offset by 1.
1635 // Must be a Constant Array
1636 ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
1637 if (!Array || !Array->getType()->getElementType()->isIntegerTy(8))
1640 // Get the number of elements in the array
1641 uint64_t NumElts = Array->getType()->getNumElements();
1643 // Traverse the constant array from StartIdx (derived above) which is
1644 // the place the GEP refers to in the array.
1645 for (unsigned i = StartIdx; i != NumElts; ++i) {
1646 Constant *Elt = Array->getOperand(i);
1647 ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
1648 if (!CI) // This array isn't suitable, non-int initializer.
1651 return i-StartIdx+1; // We found end of string, success!
1654 return 0; // The array isn't null terminated, conservatively return 'unknown'.
1657 /// GetStringLength - If we can compute the length of the string pointed to by
1658 /// the specified pointer, return 'len+1'. If we can't, return 0.
1659 uint64_t llvm::GetStringLength(Value *V) {
1660 if (!V->getType()->isPointerTy()) return 0;
1662 SmallPtrSet<PHINode*, 32> PHIs;
1663 uint64_t Len = GetStringLengthH(V, PHIs);
1664 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
1665 // an empty string as a length.
1666 return Len == ~0ULL ? 1 : Len;
1670 llvm::GetUnderlyingObject(Value *V, const TargetData *TD, unsigned MaxLookup) {
1671 if (!V->getType()->isPointerTy())
1673 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
1674 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1675 V = GEP->getPointerOperand();
1676 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1677 V = cast<Operator>(V)->getOperand(0);
1678 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1679 if (GA->mayBeOverridden())
1681 V = GA->getAliasee();
1683 // See if InstructionSimplify knows any relevant tricks.
1684 if (Instruction *I = dyn_cast<Instruction>(V))
1685 // TODO: Aquire a DominatorTree and use it.
1686 if (Value *Simplified = SimplifyInstruction(I, TD, 0)) {
1693 assert(V->getType()->isPointerTy() && "Unexpected operand type!");