1 //===- InstCombineCasts.cpp -----------------------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the visit functions for cast operations.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "InstCombine.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/IR/DataLayout.h"
17 #include "llvm/IR/PatternMatch.h"
18 #include "llvm/Target/TargetLibraryInfo.h"
19 using namespace llvm;
20 using namespace PatternMatch;
21
22 #define DEBUG_TYPE "instcombine"
23
24 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
25 /// expression. If so, decompose it, returning some value X, such that Val is
26 /// X*Scale+Offset.
27 ///
DecomposeSimpleLinearExpr(Value * Val,unsigned & Scale,uint64_t & Offset)28 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
29 uint64_t &Offset) {
30 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
31 Offset = CI->getZExtValue();
32 Scale = 0;
33 return ConstantInt::get(Val->getType(), 0);
34 }
35
36 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
37 // Cannot look past anything that might overflow.
38 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
39 if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) {
40 Scale = 1;
41 Offset = 0;
42 return Val;
43 }
44
45 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
46 if (I->getOpcode() == Instruction::Shl) {
47 // This is a value scaled by '1 << the shift amt'.
48 Scale = UINT64_C(1) << RHS->getZExtValue();
49 Offset = 0;
50 return I->getOperand(0);
51 }
52
53 if (I->getOpcode() == Instruction::Mul) {
54 // This value is scaled by 'RHS'.
55 Scale = RHS->getZExtValue();
56 Offset = 0;
57 return I->getOperand(0);
58 }
59
60 if (I->getOpcode() == Instruction::Add) {
61 // We have X+C. Check to see if we really have (X*C2)+C1,
62 // where C1 is divisible by C2.
63 unsigned SubScale;
64 Value *SubVal =
65 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
66 Offset += RHS->getZExtValue();
67 Scale = SubScale;
68 return SubVal;
69 }
70 }
71 }
72
73 // Otherwise, we can't look past this.
74 Scale = 1;
75 Offset = 0;
76 return Val;
77 }
78
79 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
80 /// try to eliminate the cast by moving the type information into the alloc.
PromoteCastOfAllocation(BitCastInst & CI,AllocaInst & AI)81 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
82 AllocaInst &AI) {
83 // This requires DataLayout to get the alloca alignment and size information.
84 if (!DL) return nullptr;
85
86 PointerType *PTy = cast<PointerType>(CI.getType());
87
88 BuilderTy AllocaBuilder(*Builder);
89 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
90
91 // Get the type really allocated and the type casted to.
92 Type *AllocElTy = AI.getAllocatedType();
93 Type *CastElTy = PTy->getElementType();
94 if (!AllocElTy->isSized() || !CastElTy->isSized()) return nullptr;
95
96 unsigned AllocElTyAlign = DL->getABITypeAlignment(AllocElTy);
97 unsigned CastElTyAlign = DL->getABITypeAlignment(CastElTy);
98 if (CastElTyAlign < AllocElTyAlign) return nullptr;
99
100 // If the allocation has multiple uses, only promote it if we are strictly
101 // increasing the alignment of the resultant allocation. If we keep it the
102 // same, we open the door to infinite loops of various kinds.
103 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return nullptr;
104
105 uint64_t AllocElTySize = DL->getTypeAllocSize(AllocElTy);
106 uint64_t CastElTySize = DL->getTypeAllocSize(CastElTy);
107 if (CastElTySize == 0 || AllocElTySize == 0) return nullptr;
108
109 // If the allocation has multiple uses, only promote it if we're not
110 // shrinking the amount of memory being allocated.
111 uint64_t AllocElTyStoreSize = DL->getTypeStoreSize(AllocElTy);
112 uint64_t CastElTyStoreSize = DL->getTypeStoreSize(CastElTy);
113 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr;
114
115 // See if we can satisfy the modulus by pulling a scale out of the array
116 // size argument.
117 unsigned ArraySizeScale;
118 uint64_t ArrayOffset;
119 Value *NumElements = // See if the array size is a decomposable linear expr.
120 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
121
122 // If we can now satisfy the modulus, by using a non-1 scale, we really can
123 // do the xform.
124 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
125 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return nullptr;
126
127 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
128 Value *Amt = nullptr;
129 if (Scale == 1) {
130 Amt = NumElements;
131 } else {
132 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
133 // Insert before the alloca, not before the cast.
134 Amt = AllocaBuilder.CreateMul(Amt, NumElements);
135 }
136
137 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
138 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
139 Offset, true);
140 Amt = AllocaBuilder.CreateAdd(Amt, Off);
141 }
142
143 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
144 New->setAlignment(AI.getAlignment());
145 New->takeName(&AI);
146 New->setUsedWithInAlloca(AI.isUsedWithInAlloca());
147
148 // If the allocation has multiple real uses, insert a cast and change all
149 // things that used it to use the new cast. This will also hack on CI, but it
150 // will die soon.
151 if (!AI.hasOneUse()) {
152 // New is the allocation instruction, pointer typed. AI is the original
153 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
154 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
155 ReplaceInstUsesWith(AI, NewCast);
156 }
157 return ReplaceInstUsesWith(CI, New);
158 }
159
160 /// EvaluateInDifferentType - Given an expression that
161 /// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
162 /// insert the code to evaluate the expression.
EvaluateInDifferentType(Value * V,Type * Ty,bool isSigned)163 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
164 bool isSigned) {
165 if (Constant *C = dyn_cast<Constant>(V)) {
166 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
167 // If we got a constantexpr back, try to simplify it with DL info.
168 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
169 C = ConstantFoldConstantExpression(CE, DL, TLI);
170 return C;
171 }
172
173 // Otherwise, it must be an instruction.
174 Instruction *I = cast<Instruction>(V);
175 Instruction *Res = nullptr;
176 unsigned Opc = I->getOpcode();
177 switch (Opc) {
178 case Instruction::Add:
179 case Instruction::Sub:
180 case Instruction::Mul:
181 case Instruction::And:
182 case Instruction::Or:
183 case Instruction::Xor:
184 case Instruction::AShr:
185 case Instruction::LShr:
186 case Instruction::Shl:
187 case Instruction::UDiv:
188 case Instruction::URem: {
189 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
190 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
191 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
192 break;
193 }
194 case Instruction::Trunc:
195 case Instruction::ZExt:
196 case Instruction::SExt:
197 // If the source type of the cast is the type we're trying for then we can
198 // just return the source. There's no need to insert it because it is not
199 // new.
200 if (I->getOperand(0)->getType() == Ty)
201 return I->getOperand(0);
202
203 // Otherwise, must be the same type of cast, so just reinsert a new one.
204 // This also handles the case of zext(trunc(x)) -> zext(x).
205 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
206 Opc == Instruction::SExt);
207 break;
208 case Instruction::Select: {
209 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
210 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
211 Res = SelectInst::Create(I->getOperand(0), True, False);
212 break;
213 }
214 case Instruction::PHI: {
215 PHINode *OPN = cast<PHINode>(I);
216 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
217 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
218 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
219 NPN->addIncoming(V, OPN->getIncomingBlock(i));
220 }
221 Res = NPN;
222 break;
223 }
224 default:
225 // TODO: Can handle more cases here.
226 llvm_unreachable("Unreachable!");
227 }
228
229 Res->takeName(I);
230 return InsertNewInstWith(Res, *I);
231 }
232
233
234 /// This function is a wrapper around CastInst::isEliminableCastPair. It
235 /// simply extracts arguments and returns what that function returns.
236 static Instruction::CastOps
isEliminableCastPair(const CastInst * CI,unsigned opcode,Type * DstTy,const DataLayout * DL)237 isEliminableCastPair(
238 const CastInst *CI, ///< The first cast instruction
239 unsigned opcode, ///< The opcode of the second cast instruction
240 Type *DstTy, ///< The target type for the second cast instruction
241 const DataLayout *DL ///< The target data for pointer size
242 ) {
243
244 Type *SrcTy = CI->getOperand(0)->getType(); // A from above
245 Type *MidTy = CI->getType(); // B from above
246
247 // Get the opcodes of the two Cast instructions
248 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
249 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
250 Type *SrcIntPtrTy = DL && SrcTy->isPtrOrPtrVectorTy() ?
251 DL->getIntPtrType(SrcTy) : nullptr;
252 Type *MidIntPtrTy = DL && MidTy->isPtrOrPtrVectorTy() ?
253 DL->getIntPtrType(MidTy) : nullptr;
254 Type *DstIntPtrTy = DL && DstTy->isPtrOrPtrVectorTy() ?
255 DL->getIntPtrType(DstTy) : nullptr;
256 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
257 DstTy, SrcIntPtrTy, MidIntPtrTy,
258 DstIntPtrTy);
259
260 // We don't want to form an inttoptr or ptrtoint that converts to an integer
261 // type that differs from the pointer size.
262 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
263 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
264 Res = 0;
265
266 return Instruction::CastOps(Res);
267 }
268
269 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
270 /// results in any code being generated and is interesting to optimize out. If
271 /// the cast can be eliminated by some other simple transformation, we prefer
272 /// to do the simplification first.
ShouldOptimizeCast(Instruction::CastOps opc,const Value * V,Type * Ty)273 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
274 Type *Ty) {
275 // Noop casts and casts of constants should be eliminated trivially.
276 if (V->getType() == Ty || isa<Constant>(V)) return false;
277
278 // If this is another cast that can be eliminated, we prefer to have it
279 // eliminated.
280 if (const CastInst *CI = dyn_cast<CastInst>(V))
281 if (isEliminableCastPair(CI, opc, Ty, DL))
282 return false;
283
284 // If this is a vector sext from a compare, then we don't want to break the
285 // idiom where each element of the extended vector is either zero or all ones.
286 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
287 return false;
288
289 return true;
290 }
291
292
293 /// @brief Implement the transforms common to all CastInst visitors.
commonCastTransforms(CastInst & CI)294 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
295 Value *Src = CI.getOperand(0);
296
297 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
298 // eliminate it now.
299 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
300 if (Instruction::CastOps opc =
301 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), DL)) {
302 // The first cast (CSrc) is eliminable so we need to fix up or replace
303 // the second cast (CI). CSrc will then have a good chance of being dead.
304 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
305 }
306 }
307
308 // If we are casting a select then fold the cast into the select
309 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
310 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
311 return NV;
312
313 // If we are casting a PHI then fold the cast into the PHI
314 if (isa<PHINode>(Src)) {
315 // We don't do this if this would create a PHI node with an illegal type if
316 // it is currently legal.
317 if (!Src->getType()->isIntegerTy() ||
318 !CI.getType()->isIntegerTy() ||
319 ShouldChangeType(CI.getType(), Src->getType()))
320 if (Instruction *NV = FoldOpIntoPhi(CI))
321 return NV;
322 }
323
324 return nullptr;
325 }
326
327 /// CanEvaluateTruncated - Return true if we can evaluate the specified
328 /// expression tree as type Ty instead of its larger type, and arrive with the
329 /// same value. This is used by code that tries to eliminate truncates.
330 ///
331 /// Ty will always be a type smaller than V. We should return true if trunc(V)
332 /// can be computed by computing V in the smaller type. If V is an instruction,
333 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
334 /// makes sense if x and y can be efficiently truncated.
335 ///
336 /// This function works on both vectors and scalars.
337 ///
CanEvaluateTruncated(Value * V,Type * Ty,InstCombiner & IC,Instruction * CxtI)338 static bool CanEvaluateTruncated(Value *V, Type *Ty, InstCombiner &IC,
339 Instruction *CxtI) {
340 // We can always evaluate constants in another type.
341 if (isa<Constant>(V))
342 return true;
343
344 Instruction *I = dyn_cast<Instruction>(V);
345 if (!I) return false;
346
347 Type *OrigTy = V->getType();
348
349 // If this is an extension from the dest type, we can eliminate it, even if it
350 // has multiple uses.
351 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
352 I->getOperand(0)->getType() == Ty)
353 return true;
354
355 // We can't extend or shrink something that has multiple uses: doing so would
356 // require duplicating the instruction in general, which isn't profitable.
357 if (!I->hasOneUse()) return false;
358
359 unsigned Opc = I->getOpcode();
360 switch (Opc) {
361 case Instruction::Add:
362 case Instruction::Sub:
363 case Instruction::Mul:
364 case Instruction::And:
365 case Instruction::Or:
366 case Instruction::Xor:
367 // These operators can all arbitrarily be extended or truncated.
368 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
369 CanEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
370
371 case Instruction::UDiv:
372 case Instruction::URem: {
373 // UDiv and URem can be truncated if all the truncated bits are zero.
374 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
375 uint32_t BitWidth = Ty->getScalarSizeInBits();
376 if (BitWidth < OrigBitWidth) {
377 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
378 if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) &&
379 IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) {
380 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
381 CanEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
382 }
383 }
384 break;
385 }
386 case Instruction::Shl:
387 // If we are truncating the result of this SHL, and if it's a shift of a
388 // constant amount, we can always perform a SHL in a smaller type.
389 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
390 uint32_t BitWidth = Ty->getScalarSizeInBits();
391 if (CI->getLimitedValue(BitWidth) < BitWidth)
392 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
393 }
394 break;
395 case Instruction::LShr:
396 // If this is a truncate of a logical shr, we can truncate it to a smaller
397 // lshr iff we know that the bits we would otherwise be shifting in are
398 // already zeros.
399 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
400 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
401 uint32_t BitWidth = Ty->getScalarSizeInBits();
402 if (IC.MaskedValueIsZero(I->getOperand(0),
403 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth), 0, CxtI) &&
404 CI->getLimitedValue(BitWidth) < BitWidth) {
405 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
406 }
407 }
408 break;
409 case Instruction::Trunc:
410 // trunc(trunc(x)) -> trunc(x)
411 return true;
412 case Instruction::ZExt:
413 case Instruction::SExt:
414 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
415 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
416 return true;
417 case Instruction::Select: {
418 SelectInst *SI = cast<SelectInst>(I);
419 return CanEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) &&
420 CanEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI);
421 }
422 case Instruction::PHI: {
423 // We can change a phi if we can change all operands. Note that we never
424 // get into trouble with cyclic PHIs here because we only consider
425 // instructions with a single use.
426 PHINode *PN = cast<PHINode>(I);
427 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
428 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty, IC, CxtI))
429 return false;
430 return true;
431 }
432 default:
433 // TODO: Can handle more cases here.
434 break;
435 }
436
437 return false;
438 }
439
visitTrunc(TruncInst & CI)440 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
441 if (Instruction *Result = commonCastTransforms(CI))
442 return Result;
443
444 // See if we can simplify any instructions used by the input whose sole
445 // purpose is to compute bits we don't care about.
446 if (SimplifyDemandedInstructionBits(CI))
447 return &CI;
448
449 Value *Src = CI.getOperand(0);
450 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
451
452 // Attempt to truncate the entire input expression tree to the destination
453 // type. Only do this if the dest type is a simple type, don't convert the
454 // expression tree to something weird like i93 unless the source is also
455 // strange.
456 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
457 CanEvaluateTruncated(Src, DestTy, *this, &CI)) {
458
459 // If this cast is a truncate, evaluting in a different type always
460 // eliminates the cast, so it is always a win.
461 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
462 " to avoid cast: " << CI << '\n');
463 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
464 assert(Res->getType() == DestTy);
465 return ReplaceInstUsesWith(CI, Res);
466 }
467
468 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
469 if (DestTy->getScalarSizeInBits() == 1) {
470 Constant *One = ConstantInt::get(Src->getType(), 1);
471 Src = Builder->CreateAnd(Src, One);
472 Value *Zero = Constant::getNullValue(Src->getType());
473 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
474 }
475
476 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
477 Value *A = nullptr; ConstantInt *Cst = nullptr;
478 if (Src->hasOneUse() &&
479 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
480 // We have three types to worry about here, the type of A, the source of
481 // the truncate (MidSize), and the destination of the truncate. We know that
482 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
483 // between ASize and ResultSize.
484 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
485
486 // If the shift amount is larger than the size of A, then the result is
487 // known to be zero because all the input bits got shifted out.
488 if (Cst->getZExtValue() >= ASize)
489 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
490
491 // Since we're doing an lshr and a zero extend, and know that the shift
492 // amount is smaller than ASize, it is always safe to do the shift in A's
493 // type, then zero extend or truncate to the result.
494 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
495 Shift->takeName(Src);
496 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
497 }
498
499 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
500 // type isn't non-native.
501 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
502 ShouldChangeType(Src->getType(), CI.getType()) &&
503 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
504 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
505 return BinaryOperator::CreateAnd(NewTrunc,
506 ConstantExpr::getTrunc(Cst, CI.getType()));
507 }
508
509 return nullptr;
510 }
511
512 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
513 /// in order to eliminate the icmp.
transformZExtICmp(ICmpInst * ICI,Instruction & CI,bool DoXform)514 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
515 bool DoXform) {
516 // If we are just checking for a icmp eq of a single bit and zext'ing it
517 // to an integer, then shift the bit to the appropriate place and then
518 // cast to integer to avoid the comparison.
519 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
520 const APInt &Op1CV = Op1C->getValue();
521
522 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
523 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
524 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
525 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
526 if (!DoXform) return ICI;
527
528 Value *In = ICI->getOperand(0);
529 Value *Sh = ConstantInt::get(In->getType(),
530 In->getType()->getScalarSizeInBits()-1);
531 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
532 if (In->getType() != CI.getType())
533 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
534
535 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
536 Constant *One = ConstantInt::get(In->getType(), 1);
537 In = Builder->CreateXor(In, One, In->getName()+".not");
538 }
539
540 return ReplaceInstUsesWith(CI, In);
541 }
542
543 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
544 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
545 // zext (X == 1) to i32 --> X iff X has only the low bit set.
546 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
547 // zext (X != 0) to i32 --> X iff X has only the low bit set.
548 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
549 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
550 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
551 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
552 // This only works for EQ and NE
553 ICI->isEquality()) {
554 // If Op1C some other power of two, convert:
555 uint32_t BitWidth = Op1C->getType()->getBitWidth();
556 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
557 computeKnownBits(ICI->getOperand(0), KnownZero, KnownOne, 0, &CI);
558
559 APInt KnownZeroMask(~KnownZero);
560 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
561 if (!DoXform) return ICI;
562
563 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
564 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
565 // (X&4) == 2 --> false
566 // (X&4) != 2 --> true
567 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
568 isNE);
569 Res = ConstantExpr::getZExt(Res, CI.getType());
570 return ReplaceInstUsesWith(CI, Res);
571 }
572
573 uint32_t ShiftAmt = KnownZeroMask.logBase2();
574 Value *In = ICI->getOperand(0);
575 if (ShiftAmt) {
576 // Perform a logical shr by shiftamt.
577 // Insert the shift to put the result in the low bit.
578 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
579 In->getName()+".lobit");
580 }
581
582 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
583 Constant *One = ConstantInt::get(In->getType(), 1);
584 In = Builder->CreateXor(In, One);
585 }
586
587 if (CI.getType() == In->getType())
588 return ReplaceInstUsesWith(CI, In);
589 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
590 }
591 }
592 }
593
594 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
595 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
596 // may lead to additional simplifications.
597 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
598 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
599 uint32_t BitWidth = ITy->getBitWidth();
600 Value *LHS = ICI->getOperand(0);
601 Value *RHS = ICI->getOperand(1);
602
603 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
604 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
605 computeKnownBits(LHS, KnownZeroLHS, KnownOneLHS, 0, &CI);
606 computeKnownBits(RHS, KnownZeroRHS, KnownOneRHS, 0, &CI);
607
608 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
609 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
610 APInt UnknownBit = ~KnownBits;
611 if (UnknownBit.countPopulation() == 1) {
612 if (!DoXform) return ICI;
613
614 Value *Result = Builder->CreateXor(LHS, RHS);
615
616 // Mask off any bits that are set and won't be shifted away.
617 if (KnownOneLHS.uge(UnknownBit))
618 Result = Builder->CreateAnd(Result,
619 ConstantInt::get(ITy, UnknownBit));
620
621 // Shift the bit we're testing down to the lsb.
622 Result = Builder->CreateLShr(
623 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
624
625 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
626 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
627 Result->takeName(ICI);
628 return ReplaceInstUsesWith(CI, Result);
629 }
630 }
631 }
632 }
633
634 return nullptr;
635 }
636
637 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
638 /// specified wider type and produce the same low bits. If not, return false.
639 ///
640 /// If this function returns true, it can also return a non-zero number of bits
641 /// (in BitsToClear) which indicates that the value it computes is correct for
642 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
643 /// out. For example, to promote something like:
644 ///
645 /// %B = trunc i64 %A to i32
646 /// %C = lshr i32 %B, 8
647 /// %E = zext i32 %C to i64
648 ///
649 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
650 /// set to 8 to indicate that the promoted value needs to have bits 24-31
651 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
652 /// clear the top bits anyway, doing this has no extra cost.
653 ///
654 /// This function works on both vectors and scalars.
CanEvaluateZExtd(Value * V,Type * Ty,unsigned & BitsToClear,InstCombiner & IC,Instruction * CxtI)655 static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
656 InstCombiner &IC, Instruction *CxtI) {
657 BitsToClear = 0;
658 if (isa<Constant>(V))
659 return true;
660
661 Instruction *I = dyn_cast<Instruction>(V);
662 if (!I) return false;
663
664 // If the input is a truncate from the destination type, we can trivially
665 // eliminate it.
666 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
667 return true;
668
669 // We can't extend or shrink something that has multiple uses: doing so would
670 // require duplicating the instruction in general, which isn't profitable.
671 if (!I->hasOneUse()) return false;
672
673 unsigned Opc = I->getOpcode(), Tmp;
674 switch (Opc) {
675 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
676 case Instruction::SExt: // zext(sext(x)) -> sext(x).
677 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
678 return true;
679 case Instruction::And:
680 case Instruction::Or:
681 case Instruction::Xor:
682 case Instruction::Add:
683 case Instruction::Sub:
684 case Instruction::Mul:
685 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
686 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
687 return false;
688 // These can all be promoted if neither operand has 'bits to clear'.
689 if (BitsToClear == 0 && Tmp == 0)
690 return true;
691
692 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
693 // other side, BitsToClear is ok.
694 if (Tmp == 0 &&
695 (Opc == Instruction::And || Opc == Instruction::Or ||
696 Opc == Instruction::Xor)) {
697 // We use MaskedValueIsZero here for generality, but the case we care
698 // about the most is constant RHS.
699 unsigned VSize = V->getType()->getScalarSizeInBits();
700 if (IC.MaskedValueIsZero(I->getOperand(1),
701 APInt::getHighBitsSet(VSize, BitsToClear),
702 0, CxtI))
703 return true;
704 }
705
706 // Otherwise, we don't know how to analyze this BitsToClear case yet.
707 return false;
708
709 case Instruction::Shl:
710 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
711 // upper bits we can reduce BitsToClear by the shift amount.
712 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
713 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
714 return false;
715 uint64_t ShiftAmt = Amt->getZExtValue();
716 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
717 return true;
718 }
719 return false;
720 case Instruction::LShr:
721 // We can promote lshr(x, cst) if we can promote x. This requires the
722 // ultimate 'and' to clear out the high zero bits we're clearing out though.
723 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
724 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
725 return false;
726 BitsToClear += Amt->getZExtValue();
727 if (BitsToClear > V->getType()->getScalarSizeInBits())
728 BitsToClear = V->getType()->getScalarSizeInBits();
729 return true;
730 }
731 // Cannot promote variable LSHR.
732 return false;
733 case Instruction::Select:
734 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
735 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
736 // TODO: If important, we could handle the case when the BitsToClear are
737 // known zero in the disagreeing side.
738 Tmp != BitsToClear)
739 return false;
740 return true;
741
742 case Instruction::PHI: {
743 // We can change a phi if we can change all operands. Note that we never
744 // get into trouble with cyclic PHIs here because we only consider
745 // instructions with a single use.
746 PHINode *PN = cast<PHINode>(I);
747 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
748 return false;
749 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
750 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
751 // TODO: If important, we could handle the case when the BitsToClear
752 // are known zero in the disagreeing input.
753 Tmp != BitsToClear)
754 return false;
755 return true;
756 }
757 default:
758 // TODO: Can handle more cases here.
759 return false;
760 }
761 }
762
visitZExt(ZExtInst & CI)763 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
764 // If this zero extend is only used by a truncate, let the truncate be
765 // eliminated before we try to optimize this zext.
766 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
767 return nullptr;
768
769 // If one of the common conversion will work, do it.
770 if (Instruction *Result = commonCastTransforms(CI))
771 return Result;
772
773 // See if we can simplify any instructions used by the input whose sole
774 // purpose is to compute bits we don't care about.
775 if (SimplifyDemandedInstructionBits(CI))
776 return &CI;
777
778 Value *Src = CI.getOperand(0);
779 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
780
781 // Attempt to extend the entire input expression tree to the destination
782 // type. Only do this if the dest type is a simple type, don't convert the
783 // expression tree to something weird like i93 unless the source is also
784 // strange.
785 unsigned BitsToClear;
786 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
787 CanEvaluateZExtd(Src, DestTy, BitsToClear, *this, &CI)) {
788 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
789 "Unreasonable BitsToClear");
790
791 // Okay, we can transform this! Insert the new expression now.
792 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
793 " to avoid zero extend: " << CI);
794 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
795 assert(Res->getType() == DestTy);
796
797 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
798 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
799
800 // If the high bits are already filled with zeros, just replace this
801 // cast with the result.
802 if (MaskedValueIsZero(Res,
803 APInt::getHighBitsSet(DestBitSize,
804 DestBitSize-SrcBitsKept),
805 0, &CI))
806 return ReplaceInstUsesWith(CI, Res);
807
808 // We need to emit an AND to clear the high bits.
809 Constant *C = ConstantInt::get(Res->getType(),
810 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
811 return BinaryOperator::CreateAnd(Res, C);
812 }
813
814 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
815 // types and if the sizes are just right we can convert this into a logical
816 // 'and' which will be much cheaper than the pair of casts.
817 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
818 // TODO: Subsume this into EvaluateInDifferentType.
819
820 // Get the sizes of the types involved. We know that the intermediate type
821 // will be smaller than A or C, but don't know the relation between A and C.
822 Value *A = CSrc->getOperand(0);
823 unsigned SrcSize = A->getType()->getScalarSizeInBits();
824 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
825 unsigned DstSize = CI.getType()->getScalarSizeInBits();
826 // If we're actually extending zero bits, then if
827 // SrcSize < DstSize: zext(a & mask)
828 // SrcSize == DstSize: a & mask
829 // SrcSize > DstSize: trunc(a) & mask
830 if (SrcSize < DstSize) {
831 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
832 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
833 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
834 return new ZExtInst(And, CI.getType());
835 }
836
837 if (SrcSize == DstSize) {
838 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
839 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
840 AndValue));
841 }
842 if (SrcSize > DstSize) {
843 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
844 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
845 return BinaryOperator::CreateAnd(Trunc,
846 ConstantInt::get(Trunc->getType(),
847 AndValue));
848 }
849 }
850
851 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
852 return transformZExtICmp(ICI, CI);
853
854 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
855 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
856 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
857 // of the (zext icmp) will be transformed.
858 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
859 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
860 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
861 (transformZExtICmp(LHS, CI, false) ||
862 transformZExtICmp(RHS, CI, false))) {
863 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
864 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
865 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
866 }
867 }
868
869 // zext(trunc(X) & C) -> (X & zext(C)).
870 Constant *C;
871 Value *X;
872 if (SrcI &&
873 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
874 X->getType() == CI.getType())
875 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType()));
876
877 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
878 Value *And;
879 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
880 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
881 X->getType() == CI.getType()) {
882 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
883 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC);
884 }
885
886 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
887 if (SrcI && SrcI->hasOneUse() &&
888 SrcI->getType()->getScalarType()->isIntegerTy(1) &&
889 match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa<CmpInst>(X))) {
890 Value *New = Builder->CreateZExt(X, CI.getType());
891 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
892 }
893
894 return nullptr;
895 }
896
897 /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations
898 /// in order to eliminate the icmp.
transformSExtICmp(ICmpInst * ICI,Instruction & CI)899 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
900 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
901 ICmpInst::Predicate Pred = ICI->getPredicate();
902
903 // Don't bother if Op1 isn't of vector or integer type.
904 if (!Op1->getType()->isIntOrIntVectorTy())
905 return nullptr;
906
907 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
908 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
909 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
910 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) ||
911 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
912
913 Value *Sh = ConstantInt::get(Op0->getType(),
914 Op0->getType()->getScalarSizeInBits()-1);
915 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
916 if (In->getType() != CI.getType())
917 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
918
919 if (Pred == ICmpInst::ICMP_SGT)
920 In = Builder->CreateNot(In, In->getName()+".not");
921 return ReplaceInstUsesWith(CI, In);
922 }
923 }
924
925 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
926 // If we know that only one bit of the LHS of the icmp can be set and we
927 // have an equality comparison with zero or a power of 2, we can transform
928 // the icmp and sext into bitwise/integer operations.
929 if (ICI->hasOneUse() &&
930 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
931 unsigned BitWidth = Op1C->getType()->getBitWidth();
932 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
933 computeKnownBits(Op0, KnownZero, KnownOne, 0, &CI);
934
935 APInt KnownZeroMask(~KnownZero);
936 if (KnownZeroMask.isPowerOf2()) {
937 Value *In = ICI->getOperand(0);
938
939 // If the icmp tests for a known zero bit we can constant fold it.
940 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
941 Value *V = Pred == ICmpInst::ICMP_NE ?
942 ConstantInt::getAllOnesValue(CI.getType()) :
943 ConstantInt::getNullValue(CI.getType());
944 return ReplaceInstUsesWith(CI, V);
945 }
946
947 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
948 // sext ((x & 2^n) == 0) -> (x >> n) - 1
949 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
950 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
951 // Perform a right shift to place the desired bit in the LSB.
952 if (ShiftAmt)
953 In = Builder->CreateLShr(In,
954 ConstantInt::get(In->getType(), ShiftAmt));
955
956 // At this point "In" is either 1 or 0. Subtract 1 to turn
957 // {1, 0} -> {0, -1}.
958 In = Builder->CreateAdd(In,
959 ConstantInt::getAllOnesValue(In->getType()),
960 "sext");
961 } else {
962 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
963 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
964 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
965 // Perform a left shift to place the desired bit in the MSB.
966 if (ShiftAmt)
967 In = Builder->CreateShl(In,
968 ConstantInt::get(In->getType(), ShiftAmt));
969
970 // Distribute the bit over the whole bit width.
971 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
972 BitWidth - 1), "sext");
973 }
974
975 if (CI.getType() == In->getType())
976 return ReplaceInstUsesWith(CI, In);
977 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
978 }
979 }
980 }
981
982 return nullptr;
983 }
984
985 /// CanEvaluateSExtd - Return true if we can take the specified value
986 /// and return it as type Ty without inserting any new casts and without
987 /// changing the value of the common low bits. This is used by code that tries
988 /// to promote integer operations to a wider types will allow us to eliminate
989 /// the extension.
990 ///
991 /// This function works on both vectors and scalars.
992 ///
CanEvaluateSExtd(Value * V,Type * Ty)993 static bool CanEvaluateSExtd(Value *V, Type *Ty) {
994 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
995 "Can't sign extend type to a smaller type");
996 // If this is a constant, it can be trivially promoted.
997 if (isa<Constant>(V))
998 return true;
999
1000 Instruction *I = dyn_cast<Instruction>(V);
1001 if (!I) return false;
1002
1003 // If this is a truncate from the dest type, we can trivially eliminate it.
1004 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1005 return true;
1006
1007 // We can't extend or shrink something that has multiple uses: doing so would
1008 // require duplicating the instruction in general, which isn't profitable.
1009 if (!I->hasOneUse()) return false;
1010
1011 switch (I->getOpcode()) {
1012 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1013 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1014 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1015 return true;
1016 case Instruction::And:
1017 case Instruction::Or:
1018 case Instruction::Xor:
1019 case Instruction::Add:
1020 case Instruction::Sub:
1021 case Instruction::Mul:
1022 // These operators can all arbitrarily be extended if their inputs can.
1023 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
1024 CanEvaluateSExtd(I->getOperand(1), Ty);
1025
1026 //case Instruction::Shl: TODO
1027 //case Instruction::LShr: TODO
1028
1029 case Instruction::Select:
1030 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
1031 CanEvaluateSExtd(I->getOperand(2), Ty);
1032
1033 case Instruction::PHI: {
1034 // We can change a phi if we can change all operands. Note that we never
1035 // get into trouble with cyclic PHIs here because we only consider
1036 // instructions with a single use.
1037 PHINode *PN = cast<PHINode>(I);
1038 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1039 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
1040 return true;
1041 }
1042 default:
1043 // TODO: Can handle more cases here.
1044 break;
1045 }
1046
1047 return false;
1048 }
1049
visitSExt(SExtInst & CI)1050 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1051 // If this sign extend is only used by a truncate, let the truncate be
1052 // eliminated before we try to optimize this sext.
1053 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
1054 return nullptr;
1055
1056 if (Instruction *I = commonCastTransforms(CI))
1057 return I;
1058
1059 // See if we can simplify any instructions used by the input whose sole
1060 // purpose is to compute bits we don't care about.
1061 if (SimplifyDemandedInstructionBits(CI))
1062 return &CI;
1063
1064 Value *Src = CI.getOperand(0);
1065 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1066
1067 // Attempt to extend the entire input expression tree to the destination
1068 // type. Only do this if the dest type is a simple type, don't convert the
1069 // expression tree to something weird like i93 unless the source is also
1070 // strange.
1071 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1072 CanEvaluateSExtd(Src, DestTy)) {
1073 // Okay, we can transform this! Insert the new expression now.
1074 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1075 " to avoid sign extend: " << CI);
1076 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1077 assert(Res->getType() == DestTy);
1078
1079 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1080 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1081
1082 // If the high bits are already filled with sign bit, just replace this
1083 // cast with the result.
1084 if (ComputeNumSignBits(Res, 0, &CI) > DestBitSize - SrcBitSize)
1085 return ReplaceInstUsesWith(CI, Res);
1086
1087 // We need to emit a shl + ashr to do the sign extend.
1088 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1089 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1090 ShAmt);
1091 }
1092
1093 // If this input is a trunc from our destination, then turn sext(trunc(x))
1094 // into shifts.
1095 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1096 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1097 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1098 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1099
1100 // We need to emit a shl + ashr to do the sign extend.
1101 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1102 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1103 return BinaryOperator::CreateAShr(Res, ShAmt);
1104 }
1105
1106 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1107 return transformSExtICmp(ICI, CI);
1108
1109 // If the input is a shl/ashr pair of a same constant, then this is a sign
1110 // extension from a smaller value. If we could trust arbitrary bitwidth
1111 // integers, we could turn this into a truncate to the smaller bit and then
1112 // use a sext for the whole extension. Since we don't, look deeper and check
1113 // for a truncate. If the source and dest are the same type, eliminate the
1114 // trunc and extend and just do shifts. For example, turn:
1115 // %a = trunc i32 %i to i8
1116 // %b = shl i8 %a, 6
1117 // %c = ashr i8 %b, 6
1118 // %d = sext i8 %c to i32
1119 // into:
1120 // %a = shl i32 %i, 30
1121 // %d = ashr i32 %a, 30
1122 Value *A = nullptr;
1123 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1124 ConstantInt *BA = nullptr, *CA = nullptr;
1125 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1126 m_ConstantInt(CA))) &&
1127 BA == CA && A->getType() == CI.getType()) {
1128 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1129 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1130 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1131 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1132 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1133 return BinaryOperator::CreateAShr(A, ShAmtV);
1134 }
1135
1136 return nullptr;
1137 }
1138
1139
1140 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1141 /// in the specified FP type without changing its value.
FitsInFPType(ConstantFP * CFP,const fltSemantics & Sem)1142 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1143 bool losesInfo;
1144 APFloat F = CFP->getValueAPF();
1145 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1146 if (!losesInfo)
1147 return ConstantFP::get(CFP->getContext(), F);
1148 return nullptr;
1149 }
1150
1151 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1152 /// through it until we get the source value.
LookThroughFPExtensions(Value * V)1153 static Value *LookThroughFPExtensions(Value *V) {
1154 if (Instruction *I = dyn_cast<Instruction>(V))
1155 if (I->getOpcode() == Instruction::FPExt)
1156 return LookThroughFPExtensions(I->getOperand(0));
1157
1158 // If this value is a constant, return the constant in the smallest FP type
1159 // that can accurately represent it. This allows us to turn
1160 // (float)((double)X+2.0) into x+2.0f.
1161 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1162 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1163 return V; // No constant folding of this.
1164 // See if the value can be truncated to half and then reextended.
1165 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf))
1166 return V;
1167 // See if the value can be truncated to float and then reextended.
1168 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1169 return V;
1170 if (CFP->getType()->isDoubleTy())
1171 return V; // Won't shrink.
1172 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1173 return V;
1174 // Don't try to shrink to various long double types.
1175 }
1176
1177 return V;
1178 }
1179
visitFPTrunc(FPTruncInst & CI)1180 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1181 if (Instruction *I = commonCastTransforms(CI))
1182 return I;
1183 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1184 // simpilify this expression to avoid one or more of the trunc/extend
1185 // operations if we can do so without changing the numerical results.
1186 //
1187 // The exact manner in which the widths of the operands interact to limit
1188 // what we can and cannot do safely varies from operation to operation, and
1189 // is explained below in the various case statements.
1190 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1191 if (OpI && OpI->hasOneUse()) {
1192 Value *LHSOrig = LookThroughFPExtensions(OpI->getOperand(0));
1193 Value *RHSOrig = LookThroughFPExtensions(OpI->getOperand(1));
1194 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1195 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1196 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1197 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1198 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1199 switch (OpI->getOpcode()) {
1200 default: break;
1201 case Instruction::FAdd:
1202 case Instruction::FSub:
1203 // For addition and subtraction, the infinitely precise result can
1204 // essentially be arbitrarily wide; proving that double rounding
1205 // will not occur because the result of OpI is exact (as we will for
1206 // FMul, for example) is hopeless. However, we *can* nonetheless
1207 // frequently know that double rounding cannot occur (or that it is
1208 // innocuous) by taking advantage of the specific structure of
1209 // infinitely-precise results that admit double rounding.
1210 //
1211 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1212 // to represent both sources, we can guarantee that the double
1213 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1214 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1215 // for proof of this fact).
1216 //
1217 // Note: Figueroa does not consider the case where DstFormat !=
1218 // SrcFormat. It's possible (likely even!) that this analysis
1219 // could be tightened for those cases, but they are rare (the main
1220 // case of interest here is (float)((double)float + float)).
1221 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1222 if (LHSOrig->getType() != CI.getType())
1223 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1224 if (RHSOrig->getType() != CI.getType())
1225 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1226 Instruction *RI =
1227 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1228 RI->copyFastMathFlags(OpI);
1229 return RI;
1230 }
1231 break;
1232 case Instruction::FMul:
1233 // For multiplication, the infinitely precise result has at most
1234 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1235 // that such a value can be exactly represented, then no double
1236 // rounding can possibly occur; we can safely perform the operation
1237 // in the destination format if it can represent both sources.
1238 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1239 if (LHSOrig->getType() != CI.getType())
1240 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1241 if (RHSOrig->getType() != CI.getType())
1242 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1243 Instruction *RI =
1244 BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1245 RI->copyFastMathFlags(OpI);
1246 return RI;
1247 }
1248 break;
1249 case Instruction::FDiv:
1250 // For division, we use again use the bound from Figueroa's
1251 // dissertation. I am entirely certain that this bound can be
1252 // tightened in the unbalanced operand case by an analysis based on
1253 // the diophantine rational approximation bound, but the well-known
1254 // condition used here is a good conservative first pass.
1255 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1256 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1257 if (LHSOrig->getType() != CI.getType())
1258 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1259 if (RHSOrig->getType() != CI.getType())
1260 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1261 Instruction *RI =
1262 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1263 RI->copyFastMathFlags(OpI);
1264 return RI;
1265 }
1266 break;
1267 case Instruction::FRem:
1268 // Remainder is straightforward. Remainder is always exact, so the
1269 // type of OpI doesn't enter into things at all. We simply evaluate
1270 // in whichever source type is larger, then convert to the
1271 // destination type.
1272 if (SrcWidth == OpWidth)
1273 break;
1274 if (LHSWidth < SrcWidth)
1275 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
1276 else if (RHSWidth <= SrcWidth)
1277 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
1278 if (LHSOrig != OpI->getOperand(0) || RHSOrig != OpI->getOperand(1)) {
1279 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
1280 if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
1281 RI->copyFastMathFlags(OpI);
1282 return CastInst::CreateFPCast(ExactResult, CI.getType());
1283 }
1284 }
1285
1286 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1287 if (BinaryOperator::isFNeg(OpI)) {
1288 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1289 CI.getType());
1290 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc);
1291 RI->copyFastMathFlags(OpI);
1292 return RI;
1293 }
1294 }
1295
1296 // (fptrunc (select cond, R1, Cst)) -->
1297 // (select cond, (fptrunc R1), (fptrunc Cst))
1298 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1299 if (SI &&
1300 (isa<ConstantFP>(SI->getOperand(1)) ||
1301 isa<ConstantFP>(SI->getOperand(2)))) {
1302 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1),
1303 CI.getType());
1304 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
1305 CI.getType());
1306 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1307 }
1308
1309 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1310 if (II) {
1311 switch (II->getIntrinsicID()) {
1312 default: break;
1313 case Intrinsic::fabs: {
1314 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1315 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
1316 CI.getType());
1317 Type *IntrinsicType[] = { CI.getType() };
1318 Function *Overload =
1319 Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(),
1320 II->getIntrinsicID(), IntrinsicType);
1321
1322 Value *Args[] = { InnerTrunc };
1323 return CallInst::Create(Overload, Args, II->getName());
1324 }
1325 }
1326 }
1327
1328 return nullptr;
1329 }
1330
visitFPExt(CastInst & CI)1331 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1332 return commonCastTransforms(CI);
1333 }
1334
visitFPToUI(FPToUIInst & FI)1335 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1336 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1337 if (!OpI)
1338 return commonCastTransforms(FI);
1339
1340 // fptoui(uitofp(X)) --> X
1341 // fptoui(sitofp(X)) --> X
1342 // This is safe if the intermediate type has enough bits in its mantissa to
1343 // accurately represent all values of X. For example, do not do this with
1344 // i64->float->i64. This is also safe for sitofp case, because any negative
1345 // 'X' value would cause an undefined result for the fptoui.
1346 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1347 OpI->getOperand(0)->getType() == FI.getType() &&
1348 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1349 OpI->getType()->getFPMantissaWidth())
1350 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1351
1352 return commonCastTransforms(FI);
1353 }
1354
visitFPToSI(FPToSIInst & FI)1355 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1356 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1357 if (!OpI)
1358 return commonCastTransforms(FI);
1359
1360 // fptosi(sitofp(X)) --> X
1361 // fptosi(uitofp(X)) --> X
1362 // This is safe if the intermediate type has enough bits in its mantissa to
1363 // accurately represent all values of X. For example, do not do this with
1364 // i64->float->i64. This is also safe for sitofp case, because any negative
1365 // 'X' value would cause an undefined result for the fptoui.
1366 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1367 OpI->getOperand(0)->getType() == FI.getType() &&
1368 (int)FI.getType()->getScalarSizeInBits() <=
1369 OpI->getType()->getFPMantissaWidth())
1370 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1371
1372 return commonCastTransforms(FI);
1373 }
1374
visitUIToFP(CastInst & CI)1375 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1376 return commonCastTransforms(CI);
1377 }
1378
visitSIToFP(CastInst & CI)1379 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1380 return commonCastTransforms(CI);
1381 }
1382
visitIntToPtr(IntToPtrInst & CI)1383 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1384 // If the source integer type is not the intptr_t type for this target, do a
1385 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1386 // cast to be exposed to other transforms.
1387
1388 if (DL) {
1389 unsigned AS = CI.getAddressSpace();
1390 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1391 DL->getPointerSizeInBits(AS)) {
1392 Type *Ty = DL->getIntPtrType(CI.getContext(), AS);
1393 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1394 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1395
1396 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1397 return new IntToPtrInst(P, CI.getType());
1398 }
1399 }
1400
1401 if (Instruction *I = commonCastTransforms(CI))
1402 return I;
1403
1404 return nullptr;
1405 }
1406
1407 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
commonPointerCastTransforms(CastInst & CI)1408 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1409 Value *Src = CI.getOperand(0);
1410
1411 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1412 // If casting the result of a getelementptr instruction with no offset, turn
1413 // this into a cast of the original pointer!
1414 if (GEP->hasAllZeroIndices() &&
1415 // If CI is an addrspacecast and GEP changes the poiner type, merging
1416 // GEP into CI would undo canonicalizing addrspacecast with different
1417 // pointer types, causing infinite loops.
1418 (!isa<AddrSpaceCastInst>(CI) ||
1419 GEP->getType() == GEP->getPointerOperand()->getType())) {
1420 // Changing the cast operand is usually not a good idea but it is safe
1421 // here because the pointer operand is being replaced with another
1422 // pointer operand so the opcode doesn't need to change.
1423 Worklist.Add(GEP);
1424 CI.setOperand(0, GEP->getOperand(0));
1425 return &CI;
1426 }
1427
1428 if (!DL)
1429 return commonCastTransforms(CI);
1430
1431 // If the GEP has a single use, and the base pointer is a bitcast, and the
1432 // GEP computes a constant offset, see if we can convert these three
1433 // instructions into fewer. This typically happens with unions and other
1434 // non-type-safe code.
1435 unsigned AS = GEP->getPointerAddressSpace();
1436 unsigned OffsetBits = DL->getPointerSizeInBits(AS);
1437 APInt Offset(OffsetBits, 0);
1438 BitCastInst *BCI = dyn_cast<BitCastInst>(GEP->getOperand(0));
1439 if (GEP->hasOneUse() &&
1440 BCI &&
1441 GEP->accumulateConstantOffset(*DL, Offset)) {
1442 // Get the base pointer input of the bitcast, and the type it points to.
1443 Value *OrigBase = BCI->getOperand(0);
1444 SmallVector<Value*, 8> NewIndices;
1445 if (FindElementAtOffset(OrigBase->getType(),
1446 Offset.getSExtValue(),
1447 NewIndices)) {
1448 // If we were able to index down into an element, create the GEP
1449 // and bitcast the result. This eliminates one bitcast, potentially
1450 // two.
1451 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1452 Builder->CreateInBoundsGEP(OrigBase, NewIndices) :
1453 Builder->CreateGEP(OrigBase, NewIndices);
1454 NGEP->takeName(GEP);
1455
1456 if (isa<BitCastInst>(CI))
1457 return new BitCastInst(NGEP, CI.getType());
1458 assert(isa<PtrToIntInst>(CI));
1459 return new PtrToIntInst(NGEP, CI.getType());
1460 }
1461 }
1462 }
1463
1464 return commonCastTransforms(CI);
1465 }
1466
visitPtrToInt(PtrToIntInst & CI)1467 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1468 // If the destination integer type is not the intptr_t type for this target,
1469 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1470 // to be exposed to other transforms.
1471
1472 if (!DL)
1473 return commonPointerCastTransforms(CI);
1474
1475 Type *Ty = CI.getType();
1476 unsigned AS = CI.getPointerAddressSpace();
1477
1478 if (Ty->getScalarSizeInBits() == DL->getPointerSizeInBits(AS))
1479 return commonPointerCastTransforms(CI);
1480
1481 Type *PtrTy = DL->getIntPtrType(CI.getContext(), AS);
1482 if (Ty->isVectorTy()) // Handle vectors of pointers.
1483 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1484
1485 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy);
1486 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1487 }
1488
1489 /// OptimizeVectorResize - This input value (which is known to have vector type)
1490 /// is being zero extended or truncated to the specified vector type. Try to
1491 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1492 ///
1493 /// The source and destination vector types may have different element types.
OptimizeVectorResize(Value * InVal,VectorType * DestTy,InstCombiner & IC)1494 static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy,
1495 InstCombiner &IC) {
1496 // We can only do this optimization if the output is a multiple of the input
1497 // element size, or the input is a multiple of the output element size.
1498 // Convert the input type to have the same element type as the output.
1499 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1500
1501 if (SrcTy->getElementType() != DestTy->getElementType()) {
1502 // The input types don't need to be identical, but for now they must be the
1503 // same size. There is no specific reason we couldn't handle things like
1504 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1505 // there yet.
1506 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1507 DestTy->getElementType()->getPrimitiveSizeInBits())
1508 return nullptr;
1509
1510 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1511 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1512 }
1513
1514 // Now that the element types match, get the shuffle mask and RHS of the
1515 // shuffle to use, which depends on whether we're increasing or decreasing the
1516 // size of the input.
1517 SmallVector<uint32_t, 16> ShuffleMask;
1518 Value *V2;
1519
1520 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1521 // If we're shrinking the number of elements, just shuffle in the low
1522 // elements from the input and use undef as the second shuffle input.
1523 V2 = UndefValue::get(SrcTy);
1524 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1525 ShuffleMask.push_back(i);
1526
1527 } else {
1528 // If we're increasing the number of elements, shuffle in all of the
1529 // elements from InVal and fill the rest of the result elements with zeros
1530 // from a constant zero.
1531 V2 = Constant::getNullValue(SrcTy);
1532 unsigned SrcElts = SrcTy->getNumElements();
1533 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1534 ShuffleMask.push_back(i);
1535
1536 // The excess elements reference the first element of the zero input.
1537 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1538 ShuffleMask.push_back(SrcElts);
1539 }
1540
1541 return new ShuffleVectorInst(InVal, V2,
1542 ConstantDataVector::get(V2->getContext(),
1543 ShuffleMask));
1544 }
1545
isMultipleOfTypeSize(unsigned Value,Type * Ty)1546 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1547 return Value % Ty->getPrimitiveSizeInBits() == 0;
1548 }
1549
getTypeSizeIndex(unsigned Value,Type * Ty)1550 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1551 return Value / Ty->getPrimitiveSizeInBits();
1552 }
1553
1554 /// CollectInsertionElements - V is a value which is inserted into a vector of
1555 /// VecEltTy. Look through the value to see if we can decompose it into
1556 /// insertions into the vector. See the example in the comment for
1557 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1558 /// The type of V is always a non-zero multiple of VecEltTy's size.
1559 /// Shift is the number of bits between the lsb of V and the lsb of
1560 /// the vector.
1561 ///
1562 /// This returns false if the pattern can't be matched or true if it can,
1563 /// filling in Elements with the elements found here.
CollectInsertionElements(Value * V,unsigned Shift,SmallVectorImpl<Value * > & Elements,Type * VecEltTy,InstCombiner & IC)1564 static bool CollectInsertionElements(Value *V, unsigned Shift,
1565 SmallVectorImpl<Value*> &Elements,
1566 Type *VecEltTy, InstCombiner &IC) {
1567 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1568 "Shift should be a multiple of the element type size");
1569
1570 // Undef values never contribute useful bits to the result.
1571 if (isa<UndefValue>(V)) return true;
1572
1573 // If we got down to a value of the right type, we win, try inserting into the
1574 // right element.
1575 if (V->getType() == VecEltTy) {
1576 // Inserting null doesn't actually insert any elements.
1577 if (Constant *C = dyn_cast<Constant>(V))
1578 if (C->isNullValue())
1579 return true;
1580
1581 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1582 if (IC.getDataLayout()->isBigEndian())
1583 ElementIndex = Elements.size() - ElementIndex - 1;
1584
1585 // Fail if multiple elements are inserted into this slot.
1586 if (Elements[ElementIndex])
1587 return false;
1588
1589 Elements[ElementIndex] = V;
1590 return true;
1591 }
1592
1593 if (Constant *C = dyn_cast<Constant>(V)) {
1594 // Figure out the # elements this provides, and bitcast it or slice it up
1595 // as required.
1596 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1597 VecEltTy);
1598 // If the constant is the size of a vector element, we just need to bitcast
1599 // it to the right type so it gets properly inserted.
1600 if (NumElts == 1)
1601 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1602 Shift, Elements, VecEltTy, IC);
1603
1604 // Okay, this is a constant that covers multiple elements. Slice it up into
1605 // pieces and insert each element-sized piece into the vector.
1606 if (!isa<IntegerType>(C->getType()))
1607 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1608 C->getType()->getPrimitiveSizeInBits()));
1609 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1610 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1611
1612 for (unsigned i = 0; i != NumElts; ++i) {
1613 unsigned ShiftI = Shift+i*ElementSize;
1614 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1615 ShiftI));
1616 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1617 if (!CollectInsertionElements(Piece, ShiftI, Elements, VecEltTy, IC))
1618 return false;
1619 }
1620 return true;
1621 }
1622
1623 if (!V->hasOneUse()) return false;
1624
1625 Instruction *I = dyn_cast<Instruction>(V);
1626 if (!I) return false;
1627 switch (I->getOpcode()) {
1628 default: return false; // Unhandled case.
1629 case Instruction::BitCast:
1630 return CollectInsertionElements(I->getOperand(0), Shift,
1631 Elements, VecEltTy, IC);
1632 case Instruction::ZExt:
1633 if (!isMultipleOfTypeSize(
1634 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1635 VecEltTy))
1636 return false;
1637 return CollectInsertionElements(I->getOperand(0), Shift,
1638 Elements, VecEltTy, IC);
1639 case Instruction::Or:
1640 return CollectInsertionElements(I->getOperand(0), Shift,
1641 Elements, VecEltTy, IC) &&
1642 CollectInsertionElements(I->getOperand(1), Shift,
1643 Elements, VecEltTy, IC);
1644 case Instruction::Shl: {
1645 // Must be shifting by a constant that is a multiple of the element size.
1646 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1647 if (!CI) return false;
1648 Shift += CI->getZExtValue();
1649 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1650 return CollectInsertionElements(I->getOperand(0), Shift,
1651 Elements, VecEltTy, IC);
1652 }
1653
1654 }
1655 }
1656
1657
1658 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1659 /// may be doing shifts and ors to assemble the elements of the vector manually.
1660 /// Try to rip the code out and replace it with insertelements. This is to
1661 /// optimize code like this:
1662 ///
1663 /// %tmp37 = bitcast float %inc to i32
1664 /// %tmp38 = zext i32 %tmp37 to i64
1665 /// %tmp31 = bitcast float %inc5 to i32
1666 /// %tmp32 = zext i32 %tmp31 to i64
1667 /// %tmp33 = shl i64 %tmp32, 32
1668 /// %ins35 = or i64 %tmp33, %tmp38
1669 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1670 ///
1671 /// Into two insertelements that do "buildvector{%inc, %inc5}".
OptimizeIntegerToVectorInsertions(BitCastInst & CI,InstCombiner & IC)1672 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1673 InstCombiner &IC) {
1674 // We need to know the target byte order to perform this optimization.
1675 if (!IC.getDataLayout()) return nullptr;
1676
1677 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1678 Value *IntInput = CI.getOperand(0);
1679
1680 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1681 if (!CollectInsertionElements(IntInput, 0, Elements,
1682 DestVecTy->getElementType(), IC))
1683 return nullptr;
1684
1685 // If we succeeded, we know that all of the element are specified by Elements
1686 // or are zero if Elements has a null entry. Recast this as a set of
1687 // insertions.
1688 Value *Result = Constant::getNullValue(CI.getType());
1689 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1690 if (!Elements[i]) continue; // Unset element.
1691
1692 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1693 IC.Builder->getInt32(i));
1694 }
1695
1696 return Result;
1697 }
1698
1699
1700 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1701 /// bitcast. The various long double bitcasts can't get in here.
OptimizeIntToFloatBitCast(BitCastInst & CI,InstCombiner & IC)1702 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1703 // We need to know the target byte order to perform this optimization.
1704 if (!IC.getDataLayout()) return nullptr;
1705
1706 Value *Src = CI.getOperand(0);
1707 Type *DestTy = CI.getType();
1708
1709 // If this is a bitcast from int to float, check to see if the int is an
1710 // extraction from a vector.
1711 Value *VecInput = nullptr;
1712 // bitcast(trunc(bitcast(somevector)))
1713 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1714 isa<VectorType>(VecInput->getType())) {
1715 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1716 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1717
1718 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1719 // If the element type of the vector doesn't match the result type,
1720 // bitcast it to be a vector type we can extract from.
1721 if (VecTy->getElementType() != DestTy) {
1722 VecTy = VectorType::get(DestTy,
1723 VecTy->getPrimitiveSizeInBits() / DestWidth);
1724 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1725 }
1726
1727 unsigned Elt = 0;
1728 if (IC.getDataLayout()->isBigEndian())
1729 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1;
1730 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1731 }
1732 }
1733
1734 // bitcast(trunc(lshr(bitcast(somevector), cst))
1735 ConstantInt *ShAmt = nullptr;
1736 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1737 m_ConstantInt(ShAmt)))) &&
1738 isa<VectorType>(VecInput->getType())) {
1739 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1740 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1741 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1742 ShAmt->getZExtValue() % DestWidth == 0) {
1743 // If the element type of the vector doesn't match the result type,
1744 // bitcast it to be a vector type we can extract from.
1745 if (VecTy->getElementType() != DestTy) {
1746 VecTy = VectorType::get(DestTy,
1747 VecTy->getPrimitiveSizeInBits() / DestWidth);
1748 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1749 }
1750
1751 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1752 if (IC.getDataLayout()->isBigEndian())
1753 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1 - Elt;
1754 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1755 }
1756 }
1757 return nullptr;
1758 }
1759
visitBitCast(BitCastInst & CI)1760 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1761 // If the operands are integer typed then apply the integer transforms,
1762 // otherwise just apply the common ones.
1763 Value *Src = CI.getOperand(0);
1764 Type *SrcTy = Src->getType();
1765 Type *DestTy = CI.getType();
1766
1767 // Get rid of casts from one type to the same type. These are useless and can
1768 // be replaced by the operand.
1769 if (DestTy == Src->getType())
1770 return ReplaceInstUsesWith(CI, Src);
1771
1772 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1773 PointerType *SrcPTy = cast<PointerType>(SrcTy);
1774 Type *DstElTy = DstPTy->getElementType();
1775 Type *SrcElTy = SrcPTy->getElementType();
1776
1777 // If we are casting a alloca to a pointer to a type of the same
1778 // size, rewrite the allocation instruction to allocate the "right" type.
1779 // There is no need to modify malloc calls because it is their bitcast that
1780 // needs to be cleaned up.
1781 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1782 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1783 return V;
1784
1785 // If the source and destination are pointers, and this cast is equivalent
1786 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1787 // This can enhance SROA and other transforms that want type-safe pointers.
1788 Constant *ZeroUInt =
1789 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1790 unsigned NumZeros = 0;
1791 while (SrcElTy != DstElTy &&
1792 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1793 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1794 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1795 ++NumZeros;
1796 }
1797
1798 // If we found a path from the src to dest, create the getelementptr now.
1799 if (SrcElTy == DstElTy) {
1800 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1801 return GetElementPtrInst::CreateInBounds(Src, Idxs);
1802 }
1803 }
1804
1805 // Try to optimize int -> float bitcasts.
1806 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1807 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1808 return I;
1809
1810 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1811 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1812 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1813 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1814 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1815 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1816 }
1817
1818 if (isa<IntegerType>(SrcTy)) {
1819 // If this is a cast from an integer to vector, check to see if the input
1820 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1821 // the casts with a shuffle and (potentially) a bitcast.
1822 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1823 CastInst *SrcCast = cast<CastInst>(Src);
1824 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1825 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1826 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1827 cast<VectorType>(DestTy), *this))
1828 return I;
1829 }
1830
1831 // If the input is an 'or' instruction, we may be doing shifts and ors to
1832 // assemble the elements of the vector manually. Try to rip the code out
1833 // and replace it with insertelements.
1834 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1835 return ReplaceInstUsesWith(CI, V);
1836 }
1837 }
1838
1839 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1840 if (SrcVTy->getNumElements() == 1) {
1841 // If our destination is not a vector, then make this a straight
1842 // scalar-scalar cast.
1843 if (!DestTy->isVectorTy()) {
1844 Value *Elem =
1845 Builder->CreateExtractElement(Src,
1846 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1847 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1848 }
1849
1850 // Otherwise, see if our source is an insert. If so, then use the scalar
1851 // component directly.
1852 if (InsertElementInst *IEI =
1853 dyn_cast<InsertElementInst>(CI.getOperand(0)))
1854 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
1855 DestTy);
1856 }
1857 }
1858
1859 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1860 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1861 // a bitcast to a vector with the same # elts.
1862 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1863 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
1864 SVI->getType()->getNumElements() ==
1865 SVI->getOperand(0)->getType()->getVectorNumElements()) {
1866 BitCastInst *Tmp;
1867 // If either of the operands is a cast from CI.getType(), then
1868 // evaluating the shuffle in the casted destination's type will allow
1869 // us to eliminate at least one cast.
1870 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1871 Tmp->getOperand(0)->getType() == DestTy) ||
1872 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1873 Tmp->getOperand(0)->getType() == DestTy)) {
1874 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1875 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1876 // Return a new shuffle vector. Use the same element ID's, as we
1877 // know the vector types match #elts.
1878 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1879 }
1880 }
1881 }
1882
1883 if (SrcTy->isPointerTy())
1884 return commonPointerCastTransforms(CI);
1885 return commonCastTransforms(CI);
1886 }
1887
visitAddrSpaceCast(AddrSpaceCastInst & CI)1888 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
1889 // If the destination pointer element type is not the same as the source's
1890 // first do a bitcast to the destination type, and then the addrspacecast.
1891 // This allows the cast to be exposed to other transforms.
1892 Value *Src = CI.getOperand(0);
1893 PointerType *SrcTy = cast<PointerType>(Src->getType()->getScalarType());
1894 PointerType *DestTy = cast<PointerType>(CI.getType()->getScalarType());
1895
1896 Type *DestElemTy = DestTy->getElementType();
1897 if (SrcTy->getElementType() != DestElemTy) {
1898 Type *MidTy = PointerType::get(DestElemTy, SrcTy->getAddressSpace());
1899 if (VectorType *VT = dyn_cast<VectorType>(CI.getType())) {
1900 // Handle vectors of pointers.
1901 MidTy = VectorType::get(MidTy, VT->getNumElements());
1902 }
1903
1904 Value *NewBitCast = Builder->CreateBitCast(Src, MidTy);
1905 return new AddrSpaceCastInst(NewBitCast, CI.getType());
1906 }
1907
1908 return commonPointerCastTransforms(CI);
1909 }
1910